1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625
7626
7627
7628
7629
7630
7631
7632
7633
7634
7635
7636
7637
7638
7639
7640
7641
7642
7643
7644
7645
7646
7647
7648
7649
7650
7651
7652
7653
7654
7655
7656
7657
7658
7659
7660
7661
7662
7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
7676
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
7747
7748
7749
7750
7751
7752
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
7769
7770
7771
7772
7773
7774
7775
7776
7777
7778
7779
7780
7781
7782
7783
7784
7785
7786
7787
7788
7789
7790
7791
7792
7793
7794
7795
7796
7797
7798
7799
7800
7801
7802
7803
7804
7805
7806
7807
7808
7809
7810
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
7846
7847
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860
7861
7862
7863
7864
7865
7866
7867
7868
7869
7870
7871
7872
7873
7874
7875
7876
7877
7878
7879
7880
7881
7882
7883
7884
7885
7886
7887
7888
7889
7890
7891
7892
7893
7894
7895
7896
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927
7928
7929
7930
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944
7945
7946
7947
7948
7949
7950
7951
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
7976
7977
7978
7979
7980
7981
7982
7983
7984
7985
7986
7987
7988
7989
7990
7991
7992
7993
7994
7995
7996
7997
7998
7999
8000
8001
8002
8003
8004
8005
8006
8007
8008
8009
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
8026
8027
8028
8029
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
8046
8047
8048
8049
8050
8051
8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
8077
8078
8079
8080
8081
8082
8083
8084
8085
8086
8087
8088
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
8113
8114
8115
8116
8117
8118
8119
8120
8121
8122
8123
8124
8125
8126
8127
8128
8129
8130
8131
8132
8133
8134
8135
8136
8137
8138
8139
8140
8141
8142
8143
8144
8145
8146
8147
8148
8149
8150
8151
8152
8153
8154
8155
8156
8157
8158
8159
8160
8161
8162
8163
8164
8165
8166
8167
8168
8169
8170
8171
8172
8173
8174
8175
8176
8177
8178
8179
8180
8181
8182
8183
8184
8185
8186
8187
8188
8189
8190
8191
8192
8193
8194
8195
8196
8197
8198
8199
8200
8201
8202
8203
8204
8205
8206
8207
8208
8209
8210
8211
8212
8213
8214
8215
8216
8217
8218
8219
8220
8221
8222
8223
8224
8225
8226
8227
8228
8229
8230
8231
8232
8233
8234
8235
8236
8237
8238
8239
8240
8241
8242
8243
8244
8245
8246
8247
8248
8249
8250
8251
8252
8253
8254
8255
8256
8257
8258
8259
8260
8261
8262
8263
8264
8265
8266
8267
8268
8269
8270
8271
8272
8273
8274
8275
8276
8277
8278
8279
8280
8281
8282
8283
8284
8285
8286
8287
8288
8289
8290
8291
8292
8293
8294
8295
8296
8297
8298
8299
8300
8301
8302
8303
8304
8305
8306
8307
8308
8309
8310
8311
8312
8313
8314
8315
8316
8317
8318
8319
8320
8321
8322
8323
8324
8325
8326
8327
8328
8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339
8340
8341
8342
8343
8344
8345
8346
8347
8348
8349
8350
8351
8352
8353
8354
8355
8356
8357
8358
8359
8360
8361
8362
8363
8364
8365
8366
8367
8368
8369
8370
8371
8372
8373
8374
8375
8376
8377
8378
8379
8380
8381
8382
8383
8384
8385
8386
8387
8388
8389
8390
8391
8392
8393
8394
8395
8396
8397
8398
8399
8400
8401
8402
8403
8404
8405
8406
8407
8408
8409
8410
8411
8412
8413
8414
8415
8416
8417
8418
8419
8420
8421
8422
8423
8424
8425
8426
8427
8428
8429
8430
8431
8432
8433
8434
8435
8436
8437
8438
8439
8440
8441
8442
8443
8444
8445
8446
8447
8448
8449
8450
8451
8452
8453
8454
8455
8456
8457
8458
8459
8460
8461
8462
8463
8464
8465
8466
8467
8468
8469
8470
8471
8472
8473
8474
8475
8476
8477
8478
8479
8480
8481
8482
8483
8484
8485
8486
8487
8488
8489
8490
8491
8492
8493
8494
8495
8496
8497
8498
8499
8500
8501
8502
8503
8504
8505
8506
8507
8508
8509
8510
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525
8526
8527
8528
8529
8530
8531
8532
8533
8534
8535
8536
8537
8538
8539
8540
8541
8542
8543
8544
8545
8546
8547
8548
8549
8550
8551
8552
8553
8554
8555
8556
8557
8558
8559
8560
8561
8562
8563
8564
8565
8566
8567
8568
8569
8570
8571
8572
8573
8574
8575
8576
8577
8578
8579
8580
8581
8582
8583
8584
8585
8586
8587
8588
8589
8590
8591
8592
8593
8594
8595
8596
8597
8598
8599
8600
8601
8602
8603
8604
8605
8606
8607
8608
8609
8610
8611
8612
8613
8614
8615
8616
8617
8618
8619
8620
8621
8622
8623
8624
8625
8626
8627
8628
8629
8630
8631
8632
8633
8634
8635
8636
8637
8638
8639
8640
8641
8642
8643
8644
8645
8646
8647
8648
8649
8650
8651
8652
8653
8654
8655
8656
8657
8658
8659
8660
8661
8662
8663
8664
8665
8666
8667
8668
8669
8670
8671
8672
8673
8674
8675
8676
8677
8678
8679
8680
8681
8682
8683
8684
8685
8686
8687
8688
8689
8690
8691
8692
8693
8694
8695
8696
8697
8698
8699
8700
8701
8702
8703
8704
8705
8706
8707
8708
8709
8710
8711
8712
8713
8714
8715
8716
8717
8718
8719
8720
8721
8722
8723
8724
8725
8726
8727
8728
8729
8730
8731
8732
8733
8734
8735
8736
8737
8738
8739
8740
8741
8742
8743
8744
8745
8746
8747
8748
8749
8750
8751
8752
8753
8754
8755
8756
8757
8758
8759
8760
8761
8762
8763
8764
8765
8766
8767
8768
8769
8770
8771
8772
8773
8774
8775
8776
8777
8778
8779
8780
8781
8782
8783
8784
8785
8786
8787
8788
8789
8790
8791
8792
8793
8794
8795
8796
8797
8798
8799
8800
8801
8802
8803
8804
8805
8806
8807
8808
8809
8810
8811
8812
8813
8814
8815
8816
8817
8818
8819
8820
8821
8822
8823
8824
8825
8826
8827
8828
8829
8830
8831
8832
8833
8834
8835
8836
8837
8838
8839
8840
8841
8842
8843
8844
8845
8846
8847
8848
8849
8850
8851
8852
8853
8854
8855
8856
8857
8858
8859
8860
8861
8862
8863
8864
8865
8866
8867
8868
8869
8870
8871
8872
8873
8874
8875
8876
8877
8878
8879
8880
8881
8882
8883
8884
8885
8886
8887
8888
8889
8890
8891
8892
8893
8894
8895
8896
8897
8898
8899
8900
8901
8902
8903
8904
8905
8906
8907
8908
8909
8910
8911
8912
8913
8914
8915
8916
8917
8918
8919
8920
8921
8922
8923
8924
8925
8926
8927
8928
8929
8930
8931
8932
8933
8934
8935
8936
8937
8938
8939
8940
8941
8942
8943
8944
8945
8946
8947
8948
8949
8950
8951
8952
8953
8954
8955
8956
8957
8958
8959
8960
8961
8962
8963
8964
8965
8966
8967
8968
8969
8970
8971
8972
8973
8974
8975
8976
8977
8978
8979
8980
8981
8982
8983
8984
8985
8986
8987
8988
8989
8990
8991
8992
8993
8994
8995
8996
8997
8998
8999
9000
9001
9002
9003
9004
9005
9006
9007
9008
9009
9010
9011
9012
9013
9014
9015
9016
9017
9018
9019
9020
9021
9022
9023
9024
9025
9026
9027
9028
9029
9030
9031
9032
9033
9034
9035
9036
9037
9038
9039
9040
9041
9042
9043
9044
9045
9046
9047
9048
9049
9050
9051
9052
9053
9054
9055
9056
9057
9058
9059
9060
9061
9062
9063
9064
9065
9066
9067
9068
9069
9070
9071
9072
9073
9074
9075
9076
9077
9078
9079
9080
9081
9082
9083
9084
9085
9086
9087
9088
9089
9090
9091
9092
9093
9094
9095
9096
9097
9098
9099
9100
9101
9102
9103
9104
9105
9106
9107
9108
9109
9110
9111
9112
9113
9114
9115
9116
9117
9118
9119
9120
9121
9122
9123
9124
9125
9126
9127
9128
9129
9130
9131
9132
9133
9134
9135
9136
9137
9138
9139
9140
9141
9142
9143
9144
9145
9146
9147
9148
9149
9150
9151
9152
9153
9154
9155
9156
9157
9158
9159
9160
9161
9162
9163
9164
9165
9166
9167
9168
9169
9170
9171
9172
9173
9174
9175
9176
9177
9178
9179
9180
9181
9182
9183
9184
9185
9186
9187
9188
9189
9190
9191
9192
9193
9194
9195
9196
9197
9198
9199
9200
9201
9202
9203
9204
9205
9206
9207
9208
9209
9210
9211
9212
9213
9214
9215
9216
9217
9218
9219
9220
9221
9222
9223
9224
9225
9226
9227
9228
9229
9230
9231
9232
9233
9234
9235
9236
9237
9238
9239
9240
9241
9242
9243
9244
9245
9246
9247
9248
9249
9250
9251
9252
9253
9254
9255
9256
9257
9258
9259
9260
9261
9262
9263
9264
9265
9266
9267
9268
9269
9270
9271
9272
9273
9274
9275
9276
9277
9278
9279
9280
9281
9282
9283
9284
9285
9286
9287
9288
9289
9290
9291
9292
9293
9294
9295
9296
9297
9298
9299
9300
9301
9302
9303
9304
9305
9306
9307
9308
9309
9310
9311
9312
9313
9314
9315
9316
9317
9318
9319
9320
9321
9322
9323
9324
9325
9326
9327
9328
9329
9330
9331
9332
9333
9334
9335
9336
9337
9338
9339
9340
9341
9342
9343
9344
9345
9346
9347
9348
9349
9350
9351
9352
9353
9354
9355
9356
9357
9358
9359
9360
9361
9362
9363
9364
9365
9366
9367
9368
9369
9370
9371
9372
9373
9374
9375
9376
9377
9378
9379
9380
9381
9382
9383
9384
9385
9386
9387
9388
9389
9390
9391
9392
9393
9394
9395
9396
9397
9398
9399
9400
9401
9402
9403
9404
9405
9406
9407
9408
9409
9410
9411
9412
9413
9414
9415
9416
9417
9418
9419
9420
9421
9422
9423
9424
9425
9426
9427
9428
9429
9430
9431
9432
9433
9434
9435
9436
9437
9438
9439
9440
9441
9442
9443
9444
9445
9446
9447
9448
9449
9450
9451
9452
9453
9454
9455
9456
9457
9458
9459
9460
9461
9462
9463
9464
9465
9466
9467
9468
9469
9470
9471
9472
9473
9474
9475
9476
9477
9478
9479
9480
9481
9482
9483
9484
9485
9486
9487
9488
9489
9490
9491
9492
9493
9494
9495
9496
9497
9498
9499
9500
9501
9502
9503
9504
9505
9506
9507
9508
9509
9510
9511
9512
9513
9514
9515
9516
9517
9518
9519
9520
9521
9522
9523
9524
9525
9526
9527
9528
9529
9530
9531
9532
9533
9534
9535
9536
9537
9538
9539
9540
9541
9542
9543
9544
9545
9546
9547
9548
9549
9550
9551
9552
9553
9554
9555
9556
9557
9558
9559
9560
9561
9562
9563
9564
9565
9566
9567
9568
9569
9570
9571
9572
9573
9574
9575
9576
9577
9578
9579
9580
9581
9582
9583
9584
9585
9586
9587
9588
9589
9590
9591
9592
9593
9594
9595
9596
9597
9598
9599
9600
9601
9602
9603
9604
9605
9606
9607
9608
9609
9610
9611
9612
9613
9614
9615
9616
9617
9618
9619
9620
9621
9622
9623
9624
9625
9626
9627
9628
9629
9630
9631
9632
9633
9634
9635
9636
9637
9638
9639
9640
9641
9642
9643
9644
9645
9646
9647
9648
9649
9650
9651
9652
9653
9654
9655
9656
9657
9658
9659
9660
9661
9662
9663
9664
9665
9666
9667
9668
9669
9670
9671
9672
9673
9674
9675
9676
9677
9678
9679
9680
9681
9682
9683
9684
9685
9686
9687
9688
9689
9690
9691
9692
9693
9694
9695
9696
9697
9698
9699
9700
9701
9702
9703
9704
9705
9706
9707
9708
9709
9710
9711
9712
9713
9714
9715
9716
9717
9718
9719
9720
9721
9722
9723
9724
9725
9726
9727
9728
9729
9730
9731
9732
9733
9734
9735
9736
9737
9738
9739
9740
9741
9742
9743
9744
9745
9746
9747
9748
9749
9750
9751
9752
9753
9754
9755
9756
9757
9758
9759
9760
9761
9762
9763
9764
9765
9766
9767
9768
9769
9770
9771
9772
9773
9774
9775
9776
9777
9778
9779
9780
9781
9782
9783
9784
9785
9786
9787
9788
9789
9790
9791
9792
9793
9794
9795
9796
9797
9798
9799
9800
9801
9802
9803
9804
9805
9806
9807
9808
9809
9810
9811
9812
9813
9814
9815
9816
9817
9818
9819
9820
9821
9822
9823
9824
9825
9826
9827
9828
9829
9830
9831
9832
9833
9834
9835
9836
9837
9838
9839
9840
9841
9842
9843
9844
9845
9846
9847
9848
9849
9850
9851
9852
9853
9854
9855
9856
9857
9858
9859
9860
9861
9862
9863
9864
9865
9866
9867
9868
9869
9870
9871
9872
9873
9874
9875
9876
9877
9878
9879
9880
9881
9882
9883
9884
9885
9886
9887
9888
9889
9890
9891
9892
9893
9894
9895
9896
9897
9898
9899
9900
9901
9902
9903
9904
9905
9906
9907
9908
9909
9910
9911
9912
9913
9914
9915
9916
9917
9918
9919
9920
9921
9922
9923
9924
9925
9926
9927
9928
9929
9930
9931
9932
9933
9934
9935
9936
9937
9938
9939
9940
9941
9942
9943
9944
9945
9946
9947
9948
9949
9950
9951
9952
9953
9954
9955
9956
9957
9958
9959
9960
9961
9962
9963
9964
9965
9966
9967
9968
9969
9970
9971
9972
9973
9974
9975
9976
9977
9978
9979
9980
9981
9982
9983
9984
9985
9986
9987
9988
9989
9990
9991
9992
9993
9994
9995
9996
9997
9998
9999
10000
10001
10002
10003
10004
10005
10006
10007
10008
10009
10010
10011
10012
10013
10014
10015
10016
10017
10018
10019
10020
10021
10022
10023
10024
10025
10026
10027
10028
10029
10030
10031
10032
10033
10034
10035
10036
10037
10038
10039
10040
10041
10042
10043
10044
10045
10046
10047
10048
10049
10050
10051
10052
10053
10054
10055
10056
10057
10058
10059
10060
10061
10062
10063
10064
10065
10066
10067
10068
10069
10070
10071
10072
10073
10074
10075
10076
10077
10078
10079
10080
10081
10082
10083
10084
10085
10086
10087
10088
10089
10090
10091
10092
10093
10094
10095
10096
10097
10098
10099
10100
10101
10102
10103
10104
10105
10106
10107
10108
10109
10110
10111
10112
10113
10114
10115
10116
10117
10118
10119
10120
10121
10122
10123
10124
10125
10126
10127
10128
10129
10130
10131
10132
10133
10134
10135
10136
10137
10138
10139
10140
10141
10142
10143
10144
10145
10146
10147
10148
10149
10150
10151
10152
10153
10154
10155
10156
10157
10158
10159
10160
10161
10162
10163
10164
10165
10166
10167
10168
10169
10170
10171
10172
10173
10174
10175
10176
10177
10178
10179
10180
10181
10182
10183
10184
10185
10186
10187
10188
10189
10190
10191
10192
10193
10194
10195
10196
10197
10198
10199
10200
10201
10202
10203
10204
10205
10206
10207
10208
10209
10210
10211
10212
10213
10214
10215
10216
10217
10218
10219
10220
10221
10222
10223
10224
10225
10226
10227
10228
10229
10230
10231
10232
10233
10234
10235
10236
10237
10238
10239
10240
10241
10242
10243
10244
10245
10246
10247
10248
10249
10250
10251
10252
10253
10254
10255
10256
10257
10258
10259
10260
10261
10262
10263
10264
10265
10266
10267
10268
10269
10270
10271
10272
10273
10274
10275
10276
10277
10278
10279
10280
10281
10282
10283
10284
10285
10286
10287
10288
10289
10290
10291
10292
10293
10294
10295
10296
10297
10298
10299
10300
10301
10302
10303
10304
10305
10306
10307
10308
10309
10310
10311
10312
10313
10314
10315
10316
10317
10318
10319
10320
10321
10322
10323
10324
10325
10326
10327
10328
10329
10330
10331
10332
10333
10334
10335
10336
10337
10338
10339
10340
10341
10342
10343
10344
10345
10346
10347
10348
10349
10350
10351
10352
10353
10354
10355
10356
10357
10358
10359
10360
10361
10362
10363
10364
10365
10366
10367
10368
10369
10370
10371
10372
10373
10374
10375
10376
10377
10378
10379
10380
10381
10382
10383
10384
10385
10386
10387
10388
10389
10390
10391
10392
10393
10394
10395
10396
10397
10398
10399
10400
10401
10402
10403
10404
10405
10406
10407
10408
10409
10410
10411
10412
10413
10414
10415
10416
10417
10418
10419
10420
10421
10422
10423
10424
10425
10426
10427
10428
10429
10430
10431
10432
10433
10434
10435
10436
10437
10438
10439
10440
10441
10442
10443
10444
10445
10446
10447
10448
10449
10450
10451
10452
10453
10454
10455
10456
10457
10458
10459
10460
10461
10462
10463
10464
10465
10466
10467
10468
10469
10470
10471
10472
10473
10474
10475
10476
10477
10478
10479
10480
10481
10482
10483
10484
10485
10486
10487
10488
10489
10490
10491
10492
10493
10494
10495
10496
10497
10498
10499
10500
10501
10502
10503
10504
10505
10506
10507
10508
10509
10510
10511
10512
10513
10514
10515
10516
10517
10518
10519
10520
10521
10522
10523
10524
10525
10526
10527
10528
10529
10530
10531
10532
10533
10534
10535
10536
10537
10538
10539
10540
10541
10542
10543
10544
10545
10546
10547
10548
10549
10550
10551
10552
10553
10554
10555
10556
10557
10558
10559
10560
10561
10562
10563
10564
10565
10566
10567
10568
10569
10570
10571
10572
10573
10574
10575
10576
10577
10578
10579
10580
10581
10582
10583
10584
10585
10586
10587
10588
10589
10590
10591
10592
10593
10594
10595
10596
10597
10598
10599
10600
10601
10602
10603
10604
10605
10606
10607
10608
10609
10610
10611
10612
10613
10614
10615
10616
10617
10618
10619
10620
10621
10622
10623
10624
10625
10626
10627
10628
10629
10630
10631
10632
10633
10634
10635
10636
10637
10638
10639
10640
10641
10642
10643
10644
10645
10646
10647
10648
10649
10650
10651
10652
10653
10654
10655
10656
10657
10658
10659
10660
10661
10662
10663
10664
10665
10666
10667
10668
10669
10670
10671
10672
10673
10674
10675
10676
10677
10678
10679
10680
10681
10682
10683
10684
10685
10686
10687
10688
10689
10690
10691
10692
10693
10694
10695
10696
10697
10698
10699
10700
10701
10702
10703
10704
10705
10706
10707
10708
10709
10710
10711
10712
10713
10714
10715
10716
10717
10718
10719
10720
10721
10722
10723
10724
10725
10726
10727
10728
10729
10730
10731
10732
10733
10734
10735
10736
10737
10738
10739
10740
10741
10742
10743
10744
10745
10746
10747
10748
10749
10750
10751
10752
10753
10754
10755
10756
10757
10758
10759
10760
10761
10762
10763
10764
10765
10766
10767
10768
10769
10770
10771
10772
10773
10774
10775
10776
10777
10778
10779
10780
10781
10782
10783
10784
10785
10786
10787
10788
10789
10790
10791
10792
10793
10794
10795
10796
10797
10798
10799
10800
10801
10802
10803
10804
10805
10806
10807
10808
10809
10810
10811
10812
10813
10814
10815
10816
10817
10818
10819
10820
10821
10822
10823
10824
10825
10826
10827
10828
10829
10830
10831
10832
10833
10834
10835
10836
10837
10838
10839
10840
10841
10842
10843
10844
10845
10846
10847
10848
10849
10850
10851
10852
10853
10854
10855
10856
10857
10858
10859
10860
10861
10862
10863
10864
10865
10866
10867
10868
10869
10870
10871
10872
10873
10874
10875
10876
10877
10878
10879
10880
10881
10882
10883
10884
10885
10886
10887
10888
10889
10890
10891
10892
10893
10894
10895
10896
10897
10898
10899
10900
10901
10902
10903
10904
10905
10906
10907
10908
10909
10910
10911
10912
10913
10914
10915
10916
10917
10918
10919
10920
10921
10922
10923
10924
10925
10926
10927
10928
10929
10930
10931
10932
10933
10934
10935
10936
10937
10938
10939
10940
10941
10942
10943
10944
10945
10946
10947
10948
10949
10950
10951
10952
10953
10954
10955
10956
10957
10958
10959
10960
10961
10962
10963
10964
10965
10966
10967
10968
10969
10970
10971
10972
10973
10974
10975
10976
10977
10978
10979
10980
10981
10982
10983
10984
10985
10986
10987
10988
10989
10990
10991
10992
10993
10994
10995
10996
10997
10998
10999
11000
11001
11002
11003
11004
11005
11006
11007
11008
11009
11010
11011
11012
11013
11014
11015
11016
11017
11018
11019
11020
11021
11022
11023
11024
11025
11026
11027
11028
11029
11030
11031
11032
11033
11034
11035
11036
11037
11038
11039
11040
11041
11042
11043
11044
11045
11046
11047
11048
11049
11050
11051
11052
11053
11054
11055
11056
11057
11058
11059
11060
11061
11062
11063
11064
11065
11066
11067
11068
11069
11070
11071
11072
11073
11074
11075
11076
11077
11078
11079
11080
11081
11082
11083
11084
11085
11086
11087
11088
11089
11090
11091
11092
11093
11094
11095
11096
11097
11098
11099
11100
11101
11102
11103
11104
11105
11106
11107
11108
11109
11110
11111
11112
11113
11114
11115
11116
11117
11118
11119
11120
11121
11122
11123
11124
11125
11126
11127
11128
11129
11130
11131
11132
11133
11134
11135
11136
11137
11138
11139
11140
11141
11142
11143
11144
11145
11146
11147
11148
11149
11150
11151
11152
11153
11154
11155
11156
11157
11158
11159
11160
11161
11162
11163
11164
11165
11166
11167
11168
11169
11170
11171
11172
11173
11174
11175
11176
11177
11178
11179
11180
11181
11182
11183
11184
11185
11186
11187
11188
11189
11190
11191
11192
11193
11194
11195
11196
11197
11198
11199
11200
11201
11202
11203
11204
11205
11206
11207
11208
11209
11210
11211
11212
11213
11214
11215
11216
11217
11218
11219
11220
11221
11222
11223
11224
11225
11226
11227
11228
11229
11230
11231
11232
11233
11234
11235
11236
11237
11238
11239
11240
11241
11242
11243
11244
11245
11246
11247
11248
11249
11250
11251
11252
11253
11254
11255
11256
11257
11258
11259
11260
11261
11262
11263
11264
11265
11266
11267
11268
11269
11270
11271
11272
11273
11274
11275
11276
11277
11278
11279
11280
11281
11282
11283
11284
11285
11286
11287
11288
11289
11290
11291
11292
11293
11294
11295
11296
11297
11298
11299
11300
11301
11302
11303
11304
11305
11306
11307
11308
11309
11310
11311
11312
11313
11314
11315
11316
11317
11318
11319
11320
11321
11322
11323
11324
11325
11326
11327
11328
11329
11330
11331
11332
11333
11334
11335
11336
11337
11338
11339
11340
11341
11342
11343
11344
11345
11346
11347
11348
11349
11350
11351
11352
11353
11354
11355
11356
11357
11358
11359
11360
11361
11362
11363
11364
11365
11366
11367
11368
11369
11370
11371
11372
11373
11374
11375
11376
11377
11378
11379
11380
11381
11382
11383
11384
11385
11386
11387
11388
11389
11390
11391
11392
11393
11394
11395
11396
11397
11398
11399
11400
11401
11402
11403
11404
11405
11406
11407
11408
11409
11410
11411
11412
11413
11414
11415
11416
11417
11418
11419
11420
11421
11422
11423
11424
11425
11426
11427
11428
11429
11430
11431
11432
11433
11434
11435
11436
11437
11438
11439
11440
11441
11442
11443
11444
11445
11446
11447
11448
11449
11450
11451
11452
11453
11454
11455
11456
11457
11458
11459
11460
11461
11462
11463
11464
11465
11466
11467
11468
11469
11470
11471
11472
11473
11474
11475
11476
11477
11478
11479
11480
11481
11482
11483
11484
11485
11486
11487
11488
11489
11490
11491
11492
11493
11494
11495
11496
11497
11498
11499
11500
11501
11502
11503
11504
11505
11506
11507
11508
11509
11510
11511
11512
11513
11514
11515
11516
11517
11518
11519
11520
11521
11522
11523
11524
11525
11526
11527
11528
11529
11530
11531
11532
11533
11534
11535
11536
11537
11538
11539
11540
11541
11542
11543
11544
11545
11546
11547
11548
11549
11550
11551
11552
11553
11554
11555
11556
11557
11558
11559
11560
11561
11562
11563
11564
11565
11566
11567
11568
11569
11570
11571
11572
11573
11574
11575
11576
11577
11578
11579
11580
11581
11582
11583
11584
11585
11586
11587
11588
11589
11590
11591
11592
11593
11594
11595
11596
11597
11598
11599
11600
11601
11602
11603
11604
11605
11606
11607
11608
11609
11610
11611
11612
11613
11614
11615
11616
11617
11618
11619
11620
11621
11622
11623
11624
11625
11626
11627
11628
11629
11630
11631
11632
11633
11634
11635
11636
11637
11638
11639
11640
11641
11642
11643
11644
11645
11646
11647
11648
11649
11650
11651
11652
11653
11654
11655
11656
11657
11658
11659
11660
11661
11662
11663
11664
11665
11666
11667
11668
11669
11670
11671
11672
11673
11674
11675
11676
11677
11678
11679
11680
11681
11682
11683
11684
11685
11686
11687
11688
11689
11690
11691
11692
11693
11694
11695
11696
11697
11698
11699
11700
11701
11702
11703
11704
11705
11706
11707
11708
11709
11710
11711
11712
11713
11714
11715
11716
11717
11718
11719
11720
11721
11722
11723
11724
11725
11726
11727
11728
11729
11730
11731
11732
11733
11734
11735
11736
11737
11738
11739
11740
11741
11742
11743
11744
11745
11746
11747
11748
11749
11750
11751
11752
11753
11754
11755
11756
11757
11758
11759
11760
11761
11762
11763
11764
11765
11766
11767
11768
11769
11770
11771
11772
11773
11774
11775
11776
11777
11778
11779
11780
11781
11782
11783
11784
11785
11786
11787
11788
11789
11790
11791
11792
11793
11794
11795
11796
11797
11798
11799
11800
11801
11802
11803
11804
11805
11806
11807
11808
11809
11810
11811
11812
11813
11814
11815
11816
11817
11818
11819
11820
11821
11822
11823
11824
11825
11826
11827
11828
11829
11830
11831
11832
11833
11834
11835
11836
11837
11838
11839
11840
11841
11842
11843
11844
11845
11846
11847
11848
11849
11850
11851
11852
11853
11854
11855
11856
11857
11858
11859
11860
11861
11862
11863
11864
11865
11866
11867
11868
11869
11870
11871
11872
11873
11874
11875
11876
11877
11878
11879
11880
11881
11882
11883
11884
11885
11886
11887
11888
11889
11890
11891
11892
11893
11894
11895
11896
11897
11898
11899
11900
11901
11902
11903
11904
11905
11906
11907
11908
11909
11910
11911
11912
11913
11914
11915
11916
11917
11918
11919
11920
11921
11922
11923
11924
11925
11926
11927
11928
11929
11930
11931
11932
11933
11934
11935
11936
11937
11938
11939
11940
11941
11942
11943
11944
11945
11946
11947
11948
11949
11950
11951
11952
11953
11954
11955
11956
11957
11958
11959
11960
11961
11962
11963
11964
11965
11966
11967
11968
11969
11970
11971
11972
11973
11974
11975
11976
11977
11978
11979
11980
11981
11982
11983
11984
11985
11986
11987
11988
11989
11990
11991
11992
11993
11994
11995
11996
11997
11998
11999
12000
12001
12002
12003
12004
12005
12006
12007
12008
12009
12010
12011
12012
12013
12014
12015
12016
12017
12018
12019
12020
12021
12022
12023
12024
12025
12026
12027
12028
12029
12030
12031
12032
12033
12034
12035
12036
12037
12038
12039
12040
12041
12042
12043
12044
12045
12046
12047
12048
12049
12050
12051
12052
12053
12054
12055
12056
12057
12058
12059
12060
12061
12062
12063
12064
12065
12066
12067
12068
12069
12070
12071
12072
12073
12074
12075
12076
12077
12078
12079
12080
12081
12082
12083
12084
12085
12086
12087
12088
12089
12090
12091
12092
12093
12094
12095
12096
12097
12098
12099
12100
12101
12102
12103
12104
12105
12106
12107
12108
12109
12110
12111
12112
12113
12114
12115
12116
12117
12118
12119
12120
12121
12122
12123
12124
12125
12126
12127
12128
12129
12130
12131
12132
12133
12134
12135
12136
12137
12138
12139
12140
12141
12142
12143
12144
12145
12146
12147
12148
12149
12150
12151
12152
12153
12154
12155
12156
12157
12158
12159
12160
12161
12162
12163
12164
12165
12166
12167
12168
12169
12170
12171
12172
12173
12174
12175
12176
12177
12178
12179
12180
12181
12182
12183
12184
12185
12186
12187
12188
12189
12190
12191
12192
12193
12194
12195
12196
12197
12198
12199
12200
12201
12202
12203
12204
12205
12206
12207
12208
12209
12210
12211
12212
12213
12214
12215
12216
12217
12218
12219
12220
12221
12222
12223
12224
12225
12226
12227
12228
12229
12230
12231
12232
12233
12234
12235
12236
12237
12238
12239
12240
12241
12242
12243
12244
12245
12246
12247
12248
12249
12250
12251
12252
12253
12254
12255
12256
12257
12258
12259
12260
12261
12262
12263
12264
12265
12266
12267
12268
12269
12270
12271
12272
12273
12274
12275
12276
12277
12278
12279
12280
12281
12282
12283
12284
12285
12286
12287
12288
12289
12290
12291
12292
12293
12294
12295
12296
12297
12298
12299
12300
12301
12302
12303
12304
12305
12306
12307
12308
12309
12310
12311
12312
12313
12314
12315
12316
12317
12318
12319
12320
12321
12322
12323
12324
12325
12326
12327
12328
12329
12330
12331
12332
12333
12334
12335
12336
12337
12338
12339
12340
12341
12342
12343
12344
12345
12346
12347
12348
12349
12350
12351
12352
12353
12354
12355
12356
12357
12358
12359
12360
12361
12362
12363
12364
12365
12366
12367
12368
12369
12370
12371
12372
12373
12374
12375
12376
12377
12378
12379
12380
12381
12382
12383
12384
12385
12386
12387
12388
12389
12390
12391
12392
12393
12394
12395
12396
12397
12398
12399
12400
12401
12402
12403
12404
12405
12406
12407
12408
12409
12410
12411
12412
12413
12414
12415
12416
12417
12418
12419
12420
12421
12422
12423
12424
12425
12426
12427
12428
12429
12430
12431
12432
12433
12434
12435
12436
12437
12438
12439
12440
12441
12442
12443
12444
12445
12446
12447
12448
12449
12450
12451
12452
12453
12454
12455
12456
12457
12458
12459
12460
12461
12462
12463
12464
12465
12466
12467
12468
12469
12470
12471
12472
12473
12474
12475
12476
12477
12478
12479
12480
12481
12482
12483
12484
12485
12486
12487
12488
12489
12490
12491
12492
12493
12494
12495
12496
12497
12498
12499
12500
12501
12502
12503
12504
12505
12506
12507
12508
12509
12510
12511
12512
12513
12514
12515
12516
12517
12518
12519
12520
12521
12522
12523
12524
12525
12526
12527
12528
12529
12530
12531
12532
12533
12534
12535
12536
12537
12538
12539
12540
12541
12542
12543
12544
12545
12546
12547
12548
12549
12550
12551
12552
12553
12554
12555
12556
12557
12558
12559
12560
12561
12562
12563
12564
12565
12566
12567
12568
12569
12570
12571
12572
12573
12574
12575
12576
12577
12578
12579
12580
12581
12582
12583
12584
12585
12586
12587
12588
12589
12590
12591
12592
12593
12594
12595
12596
12597
12598
12599
12600
12601
12602
12603
12604
12605
12606
12607
12608
12609
12610
12611
12612
12613
12614
12615
12616
12617
12618
12619
12620
12621
12622
12623
12624
12625
12626
12627
12628
12629
12630
12631
12632
12633
12634
12635
12636
12637
12638
12639
12640
12641
12642
12643
12644
12645
12646
12647
12648
12649
12650
12651
12652
12653
12654
12655
12656
12657
12658
12659
12660
12661
12662
12663
12664
12665
12666
12667
12668
12669
12670
12671
12672
12673
12674
12675
12676
12677
12678
12679
12680
12681
12682
12683
12684
12685
12686
12687
12688
12689
12690
12691
12692
12693
12694
12695
12696
12697
12698
12699
12700
12701
12702
12703
12704
12705
12706
12707
12708
12709
12710
12711
12712
12713
12714
12715
12716
12717
12718
12719
12720
12721
12722
12723
12724
12725
12726
12727
12728
12729
12730
12731
12732
12733
12734
12735
12736
12737
12738
12739
12740
12741
12742
12743
12744
12745
12746
12747
12748
12749
12750
12751
12752
12753
12754
12755
12756
12757
12758
12759
12760
12761
12762
12763
12764
12765
12766
12767
12768
12769
12770
12771
12772
12773
12774
12775
12776
12777
12778
12779
12780
12781
12782
12783
12784
12785
12786
12787
12788
12789
12790
12791
12792
12793
12794
12795
12796
12797
12798
12799
12800
12801
12802
12803
12804
12805
12806
12807
12808
12809
12810
12811
12812
12813
12814
12815
12816
12817
12818
12819
12820
12821
12822
12823
12824
12825
12826
12827
12828
12829
12830
12831
12832
12833
12834
12835
12836
12837
12838
12839
12840
12841
12842
12843
12844
12845
12846
12847
12848
12849
12850
12851
12852
12853
12854
12855
12856
12857
12858
12859
12860
12861
12862
12863
12864
12865
12866
12867
12868
12869
12870
12871
12872
12873
12874
12875
12876
12877
12878
12879
12880
12881
12882
12883
12884
12885
12886
12887
12888
12889
12890
12891
12892
12893
12894
12895
12896
12897
12898
12899
12900
12901
12902
12903
12904
12905
12906
12907
12908
12909
12910
12911
12912
12913
12914
12915
12916
12917
12918
12919
12920
12921
12922
12923
12924
12925
12926
12927
12928
12929
12930
12931
12932
12933
12934
12935
12936
12937
12938
12939
12940
12941
12942
12943
12944
12945
12946
12947
12948
12949
12950
12951
12952
12953
12954
12955
12956
12957
12958
12959
12960
12961
12962
12963
12964
12965
12966
12967
12968
12969
12970
12971
12972
12973
12974
12975
12976
12977
12978
12979
12980
12981
12982
12983
12984
12985
12986
12987
12988
12989
12990
12991
12992
12993
12994
12995
12996
12997
12998
12999
13000
13001
13002
13003
13004
13005
13006
13007
13008
13009
13010
13011
13012
13013
13014
13015
13016
13017
13018
13019
13020
13021
13022
13023
13024
13025
13026
13027
13028
13029
13030
13031
13032
13033
13034
13035
13036
13037
13038
13039
13040
13041
13042
13043
13044
13045
13046
13047
13048
13049
13050
13051
13052
13053
13054
13055
13056
13057
13058
13059
13060
13061
13062
13063
13064
13065
13066
13067
13068
13069
13070
13071
13072
13073
13074
13075
13076
13077
13078
13079
13080
13081
13082
13083
13084
13085
13086
13087
13088
13089
13090
13091
13092
13093
13094
13095
13096
13097
13098
13099
13100
13101
13102
13103
13104
13105
13106
13107
13108
13109
13110
13111
13112
13113
13114
13115
13116
13117
13118
13119
13120
13121
13122
13123
13124
13125
13126
13127
13128
13129
13130
13131
13132
13133
13134
13135
13136
13137
13138
13139
13140
13141
13142
13143
13144
13145
13146
13147
13148
13149
13150
13151
13152
13153
13154
13155
13156
13157
13158
13159
13160
13161
13162
13163
13164
13165
13166
13167
13168
13169
13170
13171
13172
13173
13174
13175
13176
13177
13178
13179
13180
13181
13182
13183
13184
13185
13186
13187
13188
13189
13190
13191
13192
13193
13194
13195
13196
13197
13198
13199
13200
13201
13202
13203
13204
13205
13206
13207
13208
13209
13210
13211
13212
13213
13214
13215
13216
13217
13218
13219
13220
13221
13222
13223
13224
13225
13226
13227
13228
13229
13230
13231
13232
13233
13234
13235
13236
13237
13238
13239
13240
13241
13242
13243
13244
13245
13246
13247
13248
13249
13250
13251
13252
13253
13254
13255
13256
13257
13258
13259
13260
13261
13262
13263
13264
13265
13266
13267
13268
13269
13270
13271
13272
13273
13274
13275
13276
13277
13278
13279
13280
13281
13282
13283
13284
13285
13286
13287
13288
13289
13290
13291
13292
13293
13294
13295
13296
13297
13298
13299
13300
13301
13302
13303
13304
13305
13306
13307
13308
13309
13310
13311
13312
13313
13314
13315
13316
13317
13318
13319
13320
13321
13322
13323
13324
13325
13326
13327
13328
13329
13330
13331
13332
13333
13334
13335
13336
13337
13338
13339
13340
13341
13342
13343
13344
13345
13346
13347
13348
13349
13350
13351
13352
13353
13354
13355
13356
13357
13358
13359
13360
13361
13362
13363
13364
13365
13366
13367
13368
13369
13370
13371
13372
13373
13374
13375
13376
13377
13378
13379
13380
13381
13382
13383
13384
13385
13386
13387
13388
13389
13390
13391
13392
13393
13394
13395
13396
13397
13398
13399
13400
13401
13402
13403
13404
13405
13406
13407
13408
13409
13410
13411
13412
13413
13414
13415
13416
13417
13418
13419
13420
13421
13422
13423
13424
13425
13426
13427
13428
13429
13430
13431
13432
13433
13434
13435
13436
13437
13438
13439
13440
13441
13442
13443
13444
13445
13446
13447
13448
13449
13450
13451
13452
13453
13454
13455
13456
13457
13458
13459
13460
13461
13462
13463
13464
13465
13466
13467
13468
13469
13470
13471
13472
13473
13474
13475
13476
13477
13478
13479
13480
13481
13482
13483
13484
13485
13486
13487
13488
13489
13490
13491
13492
13493
13494
13495
13496
13497
13498
13499
13500
13501
13502
13503
13504
13505
13506
13507
13508
13509
13510
13511
13512
13513
13514
13515
13516
13517
13518
13519
13520
13521
13522
13523
13524
13525
13526
13527
13528
13529
13530
13531
13532
13533
13534
13535
13536
13537
13538
13539
13540
13541
13542
13543
13544
13545
13546
13547
13548
13549
13550
13551
13552
13553
13554
13555
13556
13557
13558
13559
13560
13561
13562
13563
13564
13565
13566
13567
13568
13569
13570
13571
13572
13573
13574
13575
13576
13577
13578
13579
13580
13581
13582
13583
13584
13585
13586
13587
13588
13589
13590
13591
13592
13593
13594
13595
13596
13597
13598
13599
13600
13601
13602
13603
13604
13605
13606
13607
13608
13609
13610
13611
13612
13613
13614
13615
13616
13617
13618
13619
13620
13621
13622
13623
13624
13625
13626
13627
13628
13629
13630
13631
13632
13633
13634
13635
13636
13637
13638
13639
13640
13641
13642
13643
13644
13645
13646
13647
13648
13649
13650
13651
13652
13653
13654
13655
13656
13657
13658
13659
13660
13661
13662
13663
13664
13665
13666
13667
13668
13669
13670
13671
13672
13673
13674
13675
13676
13677
13678
13679
13680
13681
13682
13683
13684
13685
13686
13687
13688
13689
13690
13691
13692
13693
13694
13695
13696
13697
13698
13699
13700
13701
13702
13703
13704
13705
13706
13707
13708
13709
13710
13711
13712
13713
13714
13715
13716
13717
13718
13719
13720
13721
13722
13723
13724
13725
13726
13727
13728
13729
13730
13731
13732
13733
13734
13735
13736
13737
13738
13739
13740
13741
13742
13743
13744
13745
13746
13747
13748
13749
13750
13751
13752
13753
13754
13755
13756
13757
13758
13759
13760
13761
13762
13763
13764
13765
13766
13767
13768
13769
13770
13771
13772
13773
13774
13775
13776
13777
13778
13779
13780
13781
13782
13783
13784
13785
13786
13787
13788
13789
13790
13791
13792
13793
13794
13795
13796
13797
13798
13799
13800
13801
13802
13803
13804
13805
13806
13807
13808
13809
13810
13811
13812
13813
13814
13815
13816
13817
13818
13819
13820
13821
13822
13823
13824
13825
13826
13827
13828
13829
13830
13831
13832
13833
13834
13835
13836
13837
13838
13839
13840
13841
13842
13843
13844
13845
13846
13847
13848
13849
13850
13851
13852
13853
13854
13855
13856
13857
13858
13859
13860
13861
13862
13863
13864
13865
13866
13867
13868
13869
13870
13871
13872
13873
13874
13875
13876
13877
13878
13879
13880
13881
13882
13883
13884
13885
13886
13887
13888
13889
13890
13891
13892
13893
13894
13895
13896
13897
13898
13899
13900
13901
13902
13903
13904
13905
13906
13907
13908
13909
13910
13911
13912
13913
13914
13915
13916
13917
13918
13919
13920
13921
13922
13923
13924
13925
13926
13927
13928
13929
13930
13931
13932
13933
13934
13935
13936
13937
13938
13939
13940
13941
13942
13943
13944
13945
13946
13947
13948
13949
13950
13951
13952
13953
13954
13955
13956
13957
13958
13959
13960
13961
13962
13963
13964
13965
13966
13967
13968
13969
13970
13971
13972
13973
13974
13975
13976
13977
13978
13979
13980
13981
13982
13983
13984
13985
13986
13987
13988
13989
13990
13991
13992
13993
13994
13995
13996
13997
13998
13999
14000
14001
14002
14003
14004
14005
14006
14007
14008
14009
14010
14011
14012
14013
14014
14015
14016
14017
14018
14019
14020
14021
14022
14023
14024
14025
14026
14027
14028
14029
14030
14031
14032
14033
14034
14035
14036
14037
14038
14039
14040
14041
14042
14043
14044
14045
14046
14047
14048
14049
14050
14051
14052
14053
14054
14055
14056
14057
14058
14059
14060
14061
14062
14063
14064
14065
14066
14067
14068
14069
14070
14071
14072
14073
14074
14075
14076
14077
14078
14079
14080
14081
14082
14083
14084
14085
14086
14087
14088
14089
14090
14091
14092
14093
14094
14095
14096
14097
14098
14099
14100
14101
14102
14103
14104
14105
14106
14107
14108
14109
14110
14111
14112
14113
14114
14115
14116
14117
14118
14119
14120
14121
14122
14123
14124
14125
14126
14127
14128
14129
14130
14131
14132
14133
14134
14135
14136
14137
14138
14139
14140
14141
14142
14143
14144
14145
14146
14147
14148
14149
14150
14151
14152
14153
14154
14155
14156
14157
14158
14159
14160
14161
14162
14163
14164
14165
14166
14167
14168
14169
14170
14171
14172
14173
14174
14175
14176
14177
14178
14179
14180
14181
14182
14183
14184
14185
14186
14187
14188
14189
14190
14191
14192
14193
14194
14195
14196
14197
14198
14199
14200
14201
14202
14203
14204
14205
14206
14207
14208
14209
14210
14211
14212
14213
14214
14215
14216
14217
14218
14219
14220
14221
14222
14223
14224
14225
14226
14227
14228
14229
14230
14231
14232
14233
14234
14235
14236
14237
14238
14239
14240
14241
14242
14243
14244
14245
14246
14247
14248
14249
14250
14251
14252
14253
14254
14255
14256
14257
14258
14259
14260
14261
14262
14263
14264
14265
14266
14267
14268
14269
14270
14271
14272
14273
14274
14275
14276
14277
14278
14279
14280
14281
14282
14283
14284
14285
14286
14287
14288
14289
14290
14291
14292
14293
14294
14295
14296
14297
14298
14299
14300
14301
14302
14303
14304
14305
14306
14307
14308
14309
14310
14311
14312
14313
14314
14315
14316
14317
14318
14319
14320
14321
14322
14323
14324
14325
14326
14327
14328
14329
14330
14331
14332
14333
14334
14335
14336
14337
14338
14339
14340
14341
14342
14343
14344
14345
14346
14347
14348
14349
14350
14351
14352
14353
14354
14355
14356
14357
14358
14359
14360
14361
14362
14363
14364
14365
14366
14367
14368
14369
14370
14371
14372
14373
14374
14375
14376
14377
14378
14379
14380
14381
14382
14383
14384
14385
14386
14387
14388
14389
14390
14391
14392
14393
14394
14395
14396
14397
14398
14399
14400
14401
14402
14403
14404
14405
14406
14407
14408
14409
14410
14411
14412
14413
14414
14415
14416
14417
14418
14419
14420
14421
14422
14423
14424
14425
14426
14427
14428
14429
14430
14431
14432
14433
14434
14435
14436
14437
14438
14439
14440
14441
14442
14443
14444
14445
14446
14447
14448
14449
14450
14451
14452
14453
14454
14455
14456
14457
14458
14459
14460
14461
14462
14463
14464
14465
14466
14467
14468
14469
14470
14471
14472
14473
14474
14475
14476
14477
14478
14479
14480
14481
14482
14483
14484
14485
14486
14487
14488
14489
14490
14491
14492
14493
14494
14495
14496
14497
14498
14499
14500
14501
14502
14503
14504
14505
14506
14507
14508
14509
14510
14511
14512
14513
14514
14515
14516
14517
14518
14519
14520
14521
14522
14523
14524
14525
14526
14527
14528
14529
14530
14531
14532
14533
14534
14535
14536
14537
14538
14539
14540
14541
14542
14543
14544
14545
14546
14547
14548
14549
14550
14551
14552
14553
14554
14555
14556
14557
14558
14559
14560
14561
14562
14563
14564
14565
14566
14567
14568
14569
14570
14571
14572
14573
14574
14575
14576
14577
14578
14579
14580
14581
14582
14583
14584
14585
14586
14587
14588
14589
14590
14591
14592
14593
14594
14595
14596
14597
14598
14599
14600
14601
14602
14603
14604
14605
14606
14607
14608
14609
14610
14611
14612
14613
14614
14615
14616
14617
14618
14619
14620
14621
14622
14623
14624
14625
14626
14627
14628
14629
14630
14631
14632
14633
14634
14635
14636
14637
14638
14639
14640
14641
14642
14643
14644
14645
14646
14647
14648
14649
14650
14651
14652
14653
14654
14655
14656
14657
14658
14659
14660
14661
14662
14663
14664
14665
14666
14667
14668
14669
14670
14671
14672
14673
14674
14675
14676
14677
14678
14679
14680
14681
14682
14683
14684
14685
14686
14687
14688
14689
14690
14691
14692
14693
14694
14695
14696
14697
14698
14699
14700
14701
14702
14703
14704
14705
14706
14707
14708
14709
14710
14711
14712
14713
14714
14715
14716
14717
14718
14719
14720
14721
14722
14723
14724
14725
14726
14727
14728
14729
14730
14731
14732
14733
14734
14735
14736
14737
14738
14739
14740
14741
14742
14743
14744
14745
14746
14747
14748
14749
14750
14751
14752
14753
14754
14755
14756
14757
14758
14759
14760
14761
14762
14763
14764
14765
14766
14767
14768
14769
14770
14771
14772
14773
14774
14775
14776
14777
14778
14779
14780
14781
14782
14783
14784
14785
14786
14787
14788
14789
14790
14791
14792
14793
14794
14795
14796
14797
14798
14799
14800
14801
14802
14803
14804
14805
14806
14807
14808
14809
14810
14811
14812
14813
14814
14815
14816
14817
14818
14819
14820
14821
14822
14823
14824
14825
14826
14827
14828
14829
14830
14831
14832
14833
14834
14835
14836
14837
14838
14839
14840
14841
14842
14843
14844
14845
14846
14847
14848
14849
14850
14851
14852
14853
14854
14855
14856
14857
14858
14859
14860
14861
14862
14863
14864
14865
14866
14867
14868
14869
14870
14871
14872
14873
14874
14875
14876
14877
14878
14879
14880
14881
14882
14883
14884
14885
14886
14887
14888
14889
14890
14891
14892
14893
14894
14895
14896
14897
14898
14899
14900
14901
14902
14903
14904
14905
14906
14907
14908
14909
14910
14911
14912
14913
14914
14915
14916
14917
14918
14919
14920
14921
14922
14923
14924
14925
14926
14927
14928
14929
14930
14931
14932
14933
14934
14935
14936
14937
14938
14939
14940
14941
14942
14943
14944
14945
14946
14947
14948
14949
14950
14951
14952
14953
14954
14955
14956
14957
14958
14959
14960
14961
14962
14963
14964
14965
14966
14967
14968
14969
14970
14971
14972
14973
14974
14975
14976
14977
14978
14979
14980
14981
14982
14983
14984
14985
14986
14987
14988
14989
14990
14991
14992
14993
14994
14995
14996
14997
14998
14999
15000
15001
15002
15003
15004
15005
15006
15007
15008
15009
15010
15011
15012
15013
15014
15015
15016
15017
15018
15019
15020
15021
15022
15023
15024
15025
15026
15027
15028
15029
15030
15031
15032
15033
15034
15035
15036
15037
15038
15039
15040
15041
15042
15043
15044
15045
15046
15047
15048
15049
15050
15051
15052
15053
15054
15055
15056
15057
15058
15059
15060
15061
15062
15063
15064
15065
15066
15067
15068
15069
15070
15071
15072
15073
15074
15075
15076
15077
15078
15079
15080
15081
15082
15083
15084
15085
15086
15087
15088
15089
15090
15091
15092
15093
15094
15095
15096
15097
15098
15099
15100
15101
15102
15103
15104
15105
15106
15107
15108
15109
15110
15111
15112
15113
15114
15115
15116
15117
15118
15119
15120
15121
15122
15123
15124
15125
15126
15127
15128
15129
15130
15131
15132
15133
15134
15135
15136
15137
15138
15139
15140
15141
15142
15143
15144
15145
15146
15147
15148
15149
15150
15151
15152
15153
15154
15155
15156
15157
15158
15159
15160
15161
15162
15163
15164
15165
15166
15167
15168
15169
15170
15171
15172
15173
15174
15175
15176
15177
15178
15179
15180
15181
15182
15183
15184
15185
15186
15187
15188
15189
15190
15191
15192
15193
15194
15195
15196
15197
15198
15199
15200
15201
15202
15203
15204
15205
15206
15207
15208
15209
15210
15211
15212
15213
15214
15215
15216
15217
15218
15219
15220
15221
15222
15223
15224
15225
15226
15227
15228
15229
15230
15231
15232
15233
15234
15235
15236
15237
15238
15239
15240
15241
15242
15243
15244
15245
15246
15247
15248
15249
15250
15251
15252
15253
15254
15255
15256
15257
15258
15259
15260
15261
15262
15263
15264
15265
15266
15267
15268
15269
15270
15271
15272
15273
15274
15275
15276
15277
15278
15279
15280
15281
15282
15283
15284
15285
15286
15287
15288
15289
15290
15291
15292
15293
15294
15295
15296
15297
15298
15299
15300
15301
15302
15303
15304
15305
15306
15307
15308
15309
15310
15311
15312
15313
15314
15315
15316
15317
15318
15319
15320
15321
15322
15323
15324
15325
15326
15327
15328
15329
15330
15331
15332
15333
15334
15335
15336
15337
15338
15339
15340
15341
15342
15343
15344
15345
15346
15347
15348
15349
15350
15351
15352
15353
15354
15355
15356
15357
15358
15359
15360
15361
15362
15363
15364
15365
15366
15367
15368
15369
15370
15371
15372
15373
15374
15375
15376
15377
15378
15379
15380
15381
15382
15383
15384
15385
15386
15387
15388
15389
15390
15391
15392
15393
15394
15395
15396
15397
15398
15399
15400
15401
15402
15403
15404
15405
15406
15407
15408
15409
15410
15411
15412
15413
15414
15415
15416
15417
15418
15419
15420
15421
15422
15423
15424
15425
15426
15427
15428
15429
15430
15431
15432
15433
15434
15435
15436
15437
15438
15439
15440
15441
15442
15443
15444
15445
15446
15447
15448
15449
15450
15451
15452
15453
15454
15455
15456
15457
15458
15459
15460
15461
15462
15463
15464
15465
15466
15467
15468
15469
15470
15471
15472
15473
15474
15475
15476
15477
15478
15479
15480
15481
15482
15483
15484
15485
15486
15487
15488
15489
15490
15491
15492
15493
15494
15495
15496
15497
15498
15499
15500
15501
15502
15503
15504
15505
15506
15507
15508
15509
15510
15511
15512
15513
15514
15515
15516
15517
15518
15519
15520
15521
15522
15523
15524
15525
15526
15527
15528
15529
15530
15531
15532
15533
15534
15535
15536
15537
15538
15539
15540
15541
15542
15543
15544
15545
15546
15547
15548
15549
15550
15551
15552
15553
15554
15555
15556
15557
15558
15559
15560
15561
15562
15563
15564
15565
15566
15567
15568
15569
15570
15571
15572
15573
15574
15575
15576
15577
15578
15579
15580
15581
15582
15583
15584
15585
15586
15587
15588
15589
15590
15591
15592
15593
15594
15595
15596
15597
15598
15599
15600
15601
15602
15603
15604
15605
15606
15607
15608
15609
15610
15611
15612
15613
15614
15615
15616
15617
15618
15619
15620
15621
15622
15623
15624
15625
15626
15627
15628
15629
15630
15631
15632
15633
15634
15635
15636
15637
15638
15639
15640
15641
15642
15643
15644
15645
15646
15647
15648
15649
15650
15651
15652
15653
15654
15655
15656
15657
15658
15659
15660
15661
15662
15663
15664
15665
15666
15667
15668
15669
15670
15671
15672
15673
15674
15675
15676
15677
15678
15679
15680
15681
15682
15683
15684
15685
15686
15687
15688
15689
15690
15691
15692
15693
15694
15695
15696
15697
15698
15699
15700
15701
15702
15703
15704
15705
15706
15707
15708
15709
15710
15711
15712
15713
15714
15715
15716
15717
15718
15719
15720
15721
15722
15723
15724
15725
15726
15727
15728
15729
15730
15731
15732
15733
15734
15735
15736
15737
15738
15739
15740
15741
15742
15743
15744
15745
15746
15747
15748
15749
15750
15751
15752
15753
15754
15755
15756
15757
15758
15759
15760
15761
15762
15763
15764
15765
15766
15767
15768
15769
15770
15771
15772
15773
15774
15775
15776
15777
15778
15779
15780
15781
15782
15783
15784
15785
15786
15787
15788
15789
15790
15791
15792
15793
15794
15795
15796
15797
15798
15799
15800
15801
15802
15803
15804
15805
15806
15807
15808
15809
15810
15811
15812
15813
15814
15815
15816
15817
15818
15819
15820
15821
15822
15823
15824
15825
15826
15827
15828
15829
15830
15831
15832
15833
15834
15835
15836
15837
15838
15839
15840
15841
15842
15843
15844
15845
15846
15847
15848
15849
15850
15851
15852
15853
15854
15855
15856
15857
15858
15859
15860
15861
15862
15863
15864
15865
15866
15867
15868
15869
15870
15871
15872
15873
15874
15875
15876
15877
15878
15879
15880
15881
15882
15883
15884
15885
15886
15887
15888
15889
15890
15891
15892
15893
15894
15895
15896
15897
15898
15899
15900
15901
15902
15903
15904
15905
15906
15907
15908
15909
15910
15911
15912
15913
15914
15915
15916
15917
15918
15919
15920
15921
15922
15923
15924
15925
15926
15927
15928
15929
15930
15931
15932
15933
15934
15935
15936
15937
15938
15939
15940
15941
15942
15943
15944
15945
15946
15947
15948
15949
15950
15951
15952
15953
15954
15955
15956
15957
15958
15959
15960
15961
15962
15963
15964
15965
15966
15967
15968
15969
15970
15971
15972
15973
15974
15975
15976
15977
15978
15979
15980
15981
15982
15983
15984
15985
15986
15987
15988
15989
15990
15991
15992
15993
15994
15995
15996
15997
15998
15999
16000
16001
16002
16003
16004
16005
16006
16007
16008
16009
16010
16011
16012
16013
16014
16015
16016
16017
16018
16019
16020
16021
16022
16023
16024
16025
16026
16027
16028
16029
16030
16031
16032
16033
16034
16035
16036
16037
16038
16039
16040
16041
16042
16043
16044
16045
16046
16047
16048
16049
16050
16051
16052
16053
16054
16055
16056
16057
16058
16059
16060
16061
16062
16063
16064
16065
16066
16067
16068
16069
16070
16071
16072
16073
16074
16075
16076
16077
16078
16079
16080
16081
16082
16083
16084
16085
16086
16087
16088
16089
16090
16091
16092
16093
16094
16095
16096
16097
16098
16099
16100
16101
16102
16103
16104
16105
16106
16107
16108
16109
16110
16111
16112
16113
16114
16115
16116
16117
16118
16119
16120
16121
16122
16123
16124
16125
16126
16127
16128
16129
16130
16131
16132
16133
16134
16135
16136
16137
16138
16139
16140
16141
16142
16143
16144
16145
16146
16147
16148
16149
16150
16151
16152
16153
16154
16155
16156
16157
16158
16159
16160
16161
16162
16163
16164
16165
16166
16167
16168
16169
16170
16171
16172
16173
16174
16175
16176
16177
16178
16179
16180
16181
16182
16183
16184
16185
16186
16187
16188
16189
16190
16191
16192
16193
16194
16195
16196
16197
16198
16199
16200
16201
16202
16203
16204
16205
16206
16207
16208
16209
16210
16211
16212
16213
16214
16215
16216
16217
16218
16219
16220
16221
16222
16223
16224
16225
16226
16227
16228
16229
16230
16231
16232
16233
16234
16235
16236
16237
16238
16239
16240
16241
16242
16243
16244
16245
16246
16247
16248
16249
16250
16251
16252
16253
16254
16255
16256
16257
16258
16259
16260
16261
16262
16263
16264
16265
16266
16267
16268
16269
16270
16271
16272
16273
16274
16275
16276
16277
16278
16279
16280
16281
16282
16283
16284
16285
16286
16287
16288
16289
16290
16291
16292
16293
16294
16295
16296
16297
16298
16299
16300
16301
16302
16303
16304
16305
16306
16307
16308
16309
16310
16311
16312
16313
16314
16315
16316
16317
16318
16319
16320
16321
16322
16323
16324
16325
16326
16327
16328
16329
16330
16331
16332
16333
16334
16335
16336
16337
16338
16339
16340
16341
16342
16343
16344
16345
16346
16347
16348
16349
16350
16351
16352
16353
16354
16355
16356
16357
16358
16359
16360
16361
16362
16363
16364
16365
16366
16367
16368
16369
16370
16371
16372
16373
16374
16375
16376
16377
16378
16379
16380
16381
16382
16383
16384
16385
16386
16387
16388
16389
16390
16391
16392
16393
16394
16395
16396
16397
16398
16399
16400
16401
16402
16403
16404
16405
16406
16407
16408
16409
16410
16411
16412
16413
16414
16415
16416
16417
16418
16419
16420
16421
16422
16423
16424
16425
16426
16427
16428
16429
16430
16431
16432
16433
16434
16435
16436
16437
16438
16439
16440
16441
16442
16443
16444
16445
16446
16447
16448
16449
16450
16451
16452
16453
16454
16455
16456
16457
16458
16459
16460
16461
16462
16463
16464
16465
16466
16467
16468
16469
16470
16471
16472
16473
16474
16475
16476
16477
16478
16479
16480
16481
16482
16483
16484
16485
16486
16487
16488
16489
16490
16491
16492
16493
16494
16495
16496
16497
16498
16499
16500
16501
16502
16503
16504
16505
16506
16507
16508
16509
16510
16511
16512
16513
16514
16515
16516
16517
16518
16519
16520
16521
16522
16523
16524
16525
16526
16527
16528
16529
16530
16531
16532
16533
16534
16535
16536
16537
16538
16539
16540
16541
16542
16543
16544
16545
16546
16547
16548
16549
16550
16551
16552
16553
16554
16555
16556
16557
16558
16559
16560
16561
16562
16563
16564
16565
16566
16567
16568
16569
16570
16571
16572
16573
16574
16575
16576
16577
16578
16579
16580
16581
16582
16583
16584
16585
16586
16587
16588
16589
16590
16591
16592
16593
16594
16595
16596
16597
16598
16599
16600
16601
16602
16603
16604
16605
16606
16607
16608
16609
16610
16611
16612
16613
16614
16615
16616
16617
16618
16619
16620
16621
16622
16623
16624
16625
16626
16627
16628
16629
16630
16631
16632
16633
16634
16635
16636
16637
16638
16639
16640
16641
16642
16643
16644
16645
16646
16647
16648
16649
16650
16651
16652
16653
16654
16655
16656
16657
16658
16659
16660
16661
16662
16663
16664
16665
16666
16667
16668
16669
16670
16671
16672
16673
16674
16675
16676
16677
16678
16679
16680
16681
16682
16683
16684
16685
16686
16687
16688
16689
16690
16691
16692
16693
16694
16695
16696
16697
16698
16699
16700
16701
16702
16703
16704
16705
16706
16707
16708
16709
16710
16711
16712
16713
16714
16715
16716
16717
16718
16719
16720
16721
16722
16723
16724
16725
16726
16727
16728
16729
16730
16731
16732
16733
16734
16735
16736
16737
16738
16739
16740
16741
16742
16743
16744
16745
16746
16747
16748
16749
16750
16751
16752
16753
16754
16755
16756
16757
16758
16759
16760
16761
16762
16763
16764
16765
16766
16767
16768
16769
16770
16771
16772
16773
16774
16775
16776
16777
16778
16779
16780
16781
16782
16783
16784
16785
16786
16787
16788
16789
16790
16791
16792
16793
16794
16795
16796
16797
16798
16799
16800
16801
16802
16803
16804
16805
16806
16807
16808
16809
16810
16811
16812
16813
16814
16815
16816
16817
16818
16819
16820
16821
16822
16823
16824
16825
16826
16827
16828
16829
16830
16831
16832
16833
16834
16835
16836
16837
16838
16839
16840
16841
16842
16843
16844
16845
16846
16847
16848
16849
16850
16851
16852
16853
16854
16855
16856
16857
16858
16859
16860
16861
16862
16863
16864
16865
16866
16867
16868
16869
16870
16871
16872
16873
16874
16875
16876
16877
16878
16879
16880
16881
16882
16883
16884
16885
16886
16887
16888
16889
16890
16891
16892
16893
16894
16895
16896
16897
16898
16899
16900
16901
16902
16903
16904
16905
16906
16907
16908
16909
16910
16911
16912
16913
16914
16915
16916
16917
16918
16919
16920
16921
16922
16923
16924
16925
16926
16927
16928
16929
16930
16931
16932
16933
16934
16935
16936
16937
16938
16939
16940
16941
16942
16943
16944
16945
16946
16947
16948
16949
16950
16951
16952
16953
16954
16955
16956
16957
16958
16959
16960
16961
16962
16963
16964
16965
16966
16967
16968
16969
16970
16971
16972
16973
16974
16975
16976
16977
16978
16979
16980
16981
16982
16983
16984
16985
16986
16987
16988
16989
16990
16991
16992
16993
16994
16995
16996
16997
16998
16999
17000
17001
17002
17003
17004
17005
17006
17007
17008
17009
17010
17011
17012
17013
17014
17015
17016
17017
17018
17019
17020
17021
17022
17023
17024
17025
17026
17027
17028
17029
17030
17031
17032
17033
17034
17035
17036
17037
17038
17039
17040
17041
17042
17043
17044
17045
17046
17047
17048
17049
17050
17051
17052
17053
17054
17055
17056
17057
17058
17059
17060
17061
17062
17063
17064
17065
17066
17067
17068
17069
17070
17071
17072
17073
17074
17075
17076
17077
17078
17079
17080
17081
17082
17083
17084
17085
17086
17087
17088
17089
17090
17091
17092
17093
17094
17095
17096
17097
17098
17099
17100
17101
17102
17103
17104
17105
17106
17107
17108
17109
17110
17111
17112
17113
17114
17115
17116
17117
17118
17119
17120
17121
17122
17123
17124
17125
17126
17127
17128
17129
17130
17131
17132
17133
17134
17135
17136
17137
17138
17139
17140
17141
17142
17143
17144
17145
17146
17147
17148
17149
17150
17151
17152
17153
17154
17155
17156
17157
17158
17159
17160
17161
17162
17163
17164
17165
17166
17167
17168
17169
17170
17171
17172
17173
17174
17175
17176
17177
17178
17179
17180
17181
17182
17183
17184
17185
17186
17187
17188
17189
17190
17191
17192
17193
17194
17195
17196
17197
17198
17199
17200
17201
17202
17203
17204
17205
17206
17207
17208
17209
17210
17211
17212
17213
17214
17215
17216
17217
17218
17219
17220
17221
17222
17223
17224
17225
17226
17227
17228
17229
17230
17231
17232
17233
17234
17235
17236
17237
17238
17239
17240
17241
17242
17243
17244
17245
17246
17247
17248
17249
17250
17251
17252
17253
17254
17255
17256
17257
17258
17259
17260
17261
17262
17263
17264
17265
17266
17267
17268
17269
17270
17271
17272
17273
17274
17275
17276
17277
17278
17279
17280
17281
17282
17283
17284
17285
17286
17287
17288
17289
17290
17291
17292
17293
17294
17295
17296
17297
17298
17299
17300
17301
17302
17303
17304
17305
17306
17307
17308
17309
17310
17311
17312
17313
17314
17315
17316
17317
17318
17319
17320
17321
17322
17323
17324
17325
17326
17327
17328
17329
17330
17331
17332
17333
17334
17335
17336
17337
17338
17339
17340
17341
17342
17343
17344
17345
17346
17347
17348
17349
17350
17351
17352
17353
17354
17355
17356
17357
17358
17359
17360
17361
17362
17363
17364
17365
17366
17367
17368
17369
17370
17371
17372
17373
17374
17375
17376
17377
17378
17379
17380
17381
17382
17383
17384
17385
17386
17387
17388
17389
17390
17391
17392
17393
17394
17395
17396
17397
17398
17399
17400
17401
17402
17403
17404
17405
17406
17407
17408
17409
17410
17411
17412
17413
17414
17415
17416
17417
17418
17419
17420
17421
17422
17423
17424
17425
17426
17427
17428
17429
17430
17431
17432
17433
17434
17435
17436
17437
17438
17439
17440
17441
17442
17443
17444
17445
17446
17447
17448
17449
17450
17451
17452
17453
17454
17455
17456
17457
17458
17459
17460
17461
17462
17463
17464
17465
17466
17467
17468
17469
17470
17471
17472
17473
17474
17475
17476
17477
17478
17479
17480
17481
17482
17483
17484
17485
17486
17487
17488
17489
17490
17491
17492
17493
17494
17495
17496
17497
17498
17499
17500
17501
17502
17503
17504
17505
17506
17507
17508
17509
17510
17511
17512
17513
17514
17515
17516
17517
17518
17519
17520
17521
17522
17523
17524
17525
17526
17527
17528
17529
17530
17531
17532
17533
17534
17535
17536
17537
17538
17539
17540
17541
17542
17543
17544
17545
17546
17547
17548
17549
17550
17551
17552
17553
17554
17555
17556
17557
17558
17559
17560
17561
17562
17563
17564
17565
17566
17567
17568
17569
17570
17571
17572
17573
17574
17575
17576
17577
17578
17579
17580
17581
17582
17583
17584
17585
17586
17587
17588
17589
17590
17591
17592
17593
17594
17595
17596
17597
17598
17599
17600
17601
17602
17603
17604
17605
17606
17607
17608
17609
17610
17611
17612
17613
17614
17615
17616
17617
17618
17619
17620
17621
17622
17623
17624
17625
17626
17627
17628
17629
17630
17631
17632
17633
17634
17635
17636
17637
17638
17639
17640
17641
17642
17643
17644
17645
17646
17647
17648
17649
17650
17651
17652
17653
17654
17655
17656
17657
17658
17659
17660
17661
17662
17663
17664
17665
17666
17667
17668
17669
17670
17671
17672
17673
17674
17675
17676
17677
17678
17679
17680
17681
17682
17683
17684
17685
17686
17687
17688
17689
17690
17691
17692
17693
17694
17695
17696
17697
17698
17699
17700
17701
17702
17703
17704
17705
17706
17707
17708
17709
17710
17711
17712
17713
17714
17715
17716
17717
17718
17719
17720
17721
17722
17723
17724
17725
17726
17727
17728
17729
17730
17731
17732
17733
17734
17735
17736
17737
17738
17739
17740
17741
17742
17743
17744
17745
17746
17747
17748
17749
17750
17751
17752
17753
17754
17755
17756
17757
17758
17759
17760
17761
17762
17763
17764
17765
17766
17767
17768
17769
17770
17771
17772
17773
17774
17775
17776
17777
17778
17779
17780
17781
17782
17783
17784
17785
17786
17787
17788
17789
17790
17791
17792
17793
17794
17795
17796
17797
17798
17799
17800
17801
17802
17803
17804
17805
17806
17807
17808
17809
17810
17811
17812
17813
17814
17815
17816
17817
17818
17819
17820
17821
17822
17823
17824
17825
17826
17827
17828
17829
17830
17831
17832
17833
17834
17835
17836
17837
17838
17839
17840
17841
17842
17843
17844
17845
17846
17847
17848
17849
17850
17851
17852
17853
17854
17855
17856
17857
17858
17859
17860
17861
17862
17863
17864
17865
17866
17867
17868
17869
17870
17871
17872
17873
17874
17875
17876
17877
17878
17879
17880
17881
17882
17883
17884
17885
17886
17887
17888
17889
17890
17891
17892
17893
17894
17895
17896
17897
17898
17899
17900
17901
17902
17903
17904
17905
17906
17907
17908
17909
17910
17911
17912
17913
17914
17915
17916
17917
17918
17919
17920
17921
17922
17923
17924
17925
17926
17927
17928
17929
17930
17931
17932
17933
17934
17935
17936
17937
17938
17939
17940
17941
17942
17943
17944
17945
17946
17947
17948
17949
17950
17951
17952
17953
17954
17955
17956
17957
17958
17959
17960
17961
17962
17963
17964
17965
17966
17967
17968
17969
17970
17971
17972
17973
17974
17975
17976
17977
17978
17979
17980
17981
17982
17983
17984
17985
17986
17987
17988
17989
17990
17991
17992
17993
17994
17995
17996
17997
17998
17999
18000
18001
18002
18003
18004
18005
18006
18007
18008
18009
18010
18011
18012
18013
18014
18015
18016
18017
18018
18019
18020
18021
18022
18023
18024
18025
18026
18027
18028
18029
18030
18031
18032
18033
18034
18035
18036
18037
18038
18039
18040
18041
18042
18043
18044
18045
18046
18047
18048
18049
18050
18051
18052
18053
18054
18055
18056
18057
18058
18059
18060
18061
18062
18063
18064
18065
18066
18067
18068
18069
18070
18071
18072
18073
18074
18075
18076
18077
18078
18079
18080
18081
18082
18083
18084
18085
18086
18087
18088
18089
18090
18091
18092
18093
18094
18095
18096
18097
18098
18099
18100
18101
18102
18103
18104
18105
18106
18107
18108
18109
18110
18111
18112
18113
18114
18115
18116
18117
18118
18119
18120
18121
18122
18123
18124
18125
18126
18127
18128
18129
18130
18131
18132
18133
18134
18135
18136
18137
18138
18139
18140
18141
18142
18143
18144
18145
18146
18147
18148
18149
18150
18151
18152
18153
18154
18155
18156
18157
18158
18159
18160
18161
18162
18163
18164
18165
18166
18167
18168
18169
18170
18171
18172
18173
18174
18175
18176
18177
18178
18179
18180
18181
18182
18183
18184
18185
18186
18187
18188
18189
18190
18191
18192
18193
18194
18195
18196
18197
18198
18199
18200
18201
18202
18203
18204
18205
18206
18207
18208
18209
18210
18211
18212
18213
18214
18215
18216
18217
18218
18219
18220
18221
18222
18223
18224
18225
18226
18227
18228
18229
18230
18231
18232
18233
18234
18235
18236
18237
18238
18239
18240
18241
18242
18243
18244
18245
18246
18247
18248
18249
18250
18251
18252
18253
18254
18255
18256
18257
18258
18259
18260
18261
18262
18263
18264
18265
18266
18267
18268
18269
18270
18271
18272
18273
18274
18275
18276
18277
18278
18279
18280
18281
18282
18283
18284
18285
18286
18287
18288
18289
18290
18291
18292
18293
18294
18295
18296
18297
18298
18299
18300
18301
18302
18303
18304
18305
18306
18307
18308
18309
18310
18311
18312
18313
18314
18315
18316
18317
18318
18319
18320
18321
18322
18323
18324
18325
18326
18327
18328
18329
18330
18331
18332
18333
18334
18335
18336
18337
18338
18339
18340
18341
18342
18343
18344
18345
18346
18347
18348
18349
18350
18351
18352
18353
18354
18355
18356
18357
18358
18359
18360
18361
18362
18363
18364
18365
18366
18367
18368
18369
18370
18371
18372
18373
18374
18375
18376
18377
18378
18379
18380
18381
18382
18383
18384
18385
18386
18387
18388
18389
18390
18391
18392
18393
18394
18395
18396
18397
18398
18399
18400
18401
18402
18403
18404
18405
18406
18407
18408
18409
18410
18411
18412
18413
18414
18415
18416
18417
18418
18419
18420
18421
18422
18423
18424
18425
18426
18427
18428
18429
18430
18431
18432
18433
18434
18435
18436
18437
18438
18439
18440
18441
18442
18443
18444
18445
18446
18447
18448
18449
18450
18451
18452
18453
18454
18455
18456
18457
18458
18459
18460
18461
18462
18463
18464
18465
18466
18467
18468
18469
18470
18471
18472
18473
18474
18475
18476
18477
18478
18479
18480
18481
18482
18483
18484
18485
18486
18487
18488
18489
18490
18491
18492
18493
18494
18495
18496
18497
18498
18499
18500
18501
18502
18503
18504
18505
18506
18507
18508
18509
18510
18511
18512
18513
18514
18515
18516
18517
18518
18519
18520
18521
18522
18523
18524
18525
18526
18527
18528
18529
18530
18531
18532
18533
18534
18535
18536
18537
18538
18539
18540
18541
18542
18543
18544
18545
18546
18547
18548
18549
18550
18551
18552
18553
18554
18555
18556
18557
18558
18559
18560
18561
18562
18563
18564
18565
18566
18567
18568
18569
18570
18571
18572
18573
18574
18575
18576
18577
18578
18579
18580
18581
18582
18583
18584
18585
18586
18587
18588
18589
18590
18591
18592
18593
18594
18595
18596
18597
18598
18599
18600
18601
18602
18603
18604
18605
18606
18607
18608
18609
18610
18611
18612
18613
18614
18615
18616
18617
18618
18619
18620
18621
18622
18623
18624
18625
18626
18627
18628
18629
18630
18631
18632
18633
18634
18635
18636
18637
18638
18639
18640
18641
18642
18643
18644
18645
18646
18647
18648
18649
18650
18651
18652
18653
18654
18655
18656
18657
18658
18659
18660
18661
18662
18663
18664
18665
18666
18667
18668
18669
18670
18671
18672
18673
18674
18675
18676
18677
18678
18679
18680
18681
18682
18683
18684
18685
18686
18687
18688
18689
18690
18691
18692
18693
18694
18695
18696
18697
18698
18699
18700
18701
18702
18703
18704
18705
18706
18707
18708
18709
18710
18711
18712
18713
18714
18715
18716
18717
18718
18719
18720
18721
18722
18723
18724
18725
18726
18727
18728
18729
18730
18731
18732
18733
18734
18735
18736
18737
18738
18739
18740
18741
18742
18743
18744
18745
18746
18747
18748
18749
18750
18751
18752
18753
18754
18755
18756
18757
18758
18759
18760
18761
18762
18763
18764
18765
18766
18767
18768
18769
18770
18771
18772
18773
18774
18775
18776
18777
18778
18779
18780
18781
18782
18783
18784
18785
18786
18787
18788
18789
18790
18791
18792
18793
18794
18795
18796
18797
18798
18799
18800
18801
18802
18803
18804
18805
18806
18807
18808
18809
18810
18811
18812
18813
18814
18815
18816
18817
18818
18819
18820
18821
18822
18823
18824
18825
18826
18827
18828
18829
18830
18831
18832
18833
18834
18835
18836
18837
18838
18839
18840
18841
18842
18843
18844
18845
18846
18847
18848
18849
18850
18851
18852
18853
18854
18855
18856
18857
18858
18859
18860
18861
18862
18863
18864
18865
18866
18867
18868
18869
18870
18871
18872
18873
18874
18875
18876
18877
18878
18879
18880
18881
18882
18883
18884
18885
18886
18887
18888
18889
18890
18891
18892
18893
18894
18895
18896
18897
18898
18899
18900
18901
18902
18903
18904
18905
18906
18907
18908
18909
18910
18911
18912
18913
18914
18915
18916
18917
18918
18919
18920
18921
18922
18923
18924
18925
18926
18927
18928
18929
18930
18931
18932
18933
18934
18935
18936
18937
18938
18939
18940
18941
18942
18943
18944
18945
18946
18947
18948
18949
18950
18951
18952
18953
18954
18955
18956
18957
18958
18959
18960
18961
18962
18963
18964
18965
18966
18967
18968
18969
18970
18971
18972
18973
18974
18975
18976
18977
18978
18979
18980
18981
18982
18983
18984
18985
18986
18987
18988
18989
18990
18991
18992
18993
18994
18995
18996
18997
18998
18999
19000
19001
19002
19003
19004
19005
19006
19007
19008
19009
19010
19011
19012
19013
19014
19015
19016
19017
19018
19019
19020
19021
19022
19023
19024
19025
19026
19027
19028
19029
19030
19031
19032
19033
19034
19035
19036
19037
19038
19039
19040
19041
19042
19043
19044
19045
19046
19047
19048
19049
19050
19051
19052
19053
19054
19055
19056
19057
19058
19059
19060
19061
19062
19063
19064
19065
19066
19067
19068
19069
19070
19071
19072
19073
19074
19075
19076
19077
19078
19079
19080
19081
19082
19083
19084
19085
19086
19087
19088
19089
19090
19091
19092
19093
19094
19095
19096
19097
19098
19099
19100
19101
19102
19103
19104
19105
19106
19107
19108
19109
19110
19111
19112
19113
19114
19115
19116
19117
19118
19119
19120
19121
19122
19123
19124
19125
19126
19127
19128
19129
19130
19131
19132
19133
19134
19135
19136
19137
19138
19139
19140
19141
19142
19143
19144
19145
19146
19147
19148
19149
19150
19151
19152
19153
19154
19155
19156
19157
19158
19159
19160
19161
19162
19163
19164
19165
19166
19167
19168
19169
19170
19171
19172
19173
19174
19175
19176
19177
19178
19179
19180
19181
19182
19183
19184
19185
19186
19187
19188
19189
19190
19191
19192
19193
19194
19195
19196
19197
19198
19199
19200
19201
19202
19203
19204
19205
19206
19207
19208
19209
19210
19211
19212
19213
19214
19215
19216
19217
19218
19219
19220
19221
19222
19223
19224
19225
19226
19227
19228
19229
19230
19231
19232
19233
19234
19235
19236
19237
19238
19239
19240
19241
19242
19243
19244
19245
19246
19247
19248
19249
19250
19251
19252
19253
19254
19255
19256
19257
19258
19259
19260
19261
19262
19263
19264
19265
19266
19267
19268
19269
19270
19271
19272
19273
19274
19275
19276
19277
19278
19279
19280
19281
19282
19283
19284
19285
19286
19287
19288
19289
19290
19291
19292
19293
19294
19295
19296
19297
19298
19299
19300
19301
19302
19303
19304
19305
19306
19307
19308
19309
19310
19311
19312
19313
19314
19315
19316
19317
19318
19319
19320
19321
19322
19323
19324
19325
19326
19327
19328
19329
19330
19331
19332
19333
19334
19335
19336
19337
19338
19339
19340
19341
19342
19343
19344
19345
19346
19347
19348
19349
19350
19351
19352
19353
19354
19355
19356
19357
19358
19359
19360
19361
19362
19363
19364
19365
19366
19367
19368
19369
19370
19371
19372
19373
19374
19375
19376
19377
19378
19379
19380
19381
19382
19383
19384
19385
19386
19387
19388
19389
19390
19391
19392
19393
19394
19395
19396
19397
19398
19399
19400
19401
19402
19403
19404
19405
19406
19407
19408
19409
19410
19411
19412
19413
19414
19415
19416
19417
19418
19419
19420
19421
19422
19423
19424
19425
19426
19427
19428
19429
19430
19431
19432
19433
19434
19435
19436
19437
19438
19439
19440
19441
19442
19443
19444
19445
19446
19447
19448
19449
19450
19451
19452
19453
19454
19455
19456
19457
19458
19459
19460
19461
19462
19463
19464
19465
19466
19467
19468
19469
19470
19471
19472
19473
19474
19475
19476
19477
19478
19479
19480
19481
19482
19483
19484
19485
19486
19487
19488
19489
19490
19491
19492
19493
19494
19495
19496
19497
19498
19499
19500
19501
19502
19503
19504
19505
19506
19507
19508
19509
19510
19511
19512
19513
19514
19515
19516
19517
19518
19519
19520
19521
19522
19523
19524
19525
19526
19527
19528
19529
19530
19531
19532
19533
19534
19535
19536
19537
19538
19539
19540
19541
19542
19543
19544
19545
19546
19547
19548
19549
19550
19551
19552
19553
19554
19555
19556
19557
19558
19559
19560
19561
19562
19563
19564
19565
19566
19567
19568
19569
19570
19571
19572
19573
19574
19575
19576
19577
19578
19579
19580
19581
19582
19583
19584
19585
19586
19587
19588
19589
19590
19591
19592
19593
19594
19595
19596
19597
19598
19599
19600
19601
19602
19603
19604
19605
19606
19607
19608
19609
19610
19611
19612
19613
19614
19615
19616
19617
19618
19619
19620
19621
19622
19623
19624
19625
19626
19627
19628
19629
19630
19631
19632
19633
19634
19635
19636
19637
19638
19639
19640
19641
19642
19643
19644
19645
19646
19647
19648
19649
19650
19651
19652
19653
19654
19655
19656
19657
19658
19659
19660
19661
19662
19663
19664
19665
19666
19667
19668
19669
19670
19671
19672
19673
19674
19675
19676
19677
19678
19679
19680
19681
19682
19683
19684
19685
19686
19687
19688
19689
19690
19691
19692
19693
19694
19695
19696
19697
19698
19699
19700
19701
19702
19703
19704
19705
19706
19707
19708
19709
19710
19711
19712
19713
19714
19715
19716
19717
19718
19719
19720
19721
19722
19723
19724
19725
19726
19727
19728
19729
19730
19731
19732
19733
19734
19735
19736
19737
19738
19739
19740
19741
19742
19743
19744
19745
19746
19747
19748
19749
19750
19751
19752
19753
19754
19755
19756
19757
19758
19759
19760
19761
19762
19763
19764
19765
19766
19767
19768
19769
19770
19771
19772
19773
19774
19775
19776
19777
19778
19779
19780
19781
19782
19783
19784
19785
19786
19787
19788
19789
19790
19791
19792
19793
19794
19795
19796
19797
19798
19799
19800
19801
19802
19803
19804
19805
19806
19807
19808
19809
19810
19811
19812
19813
19814
19815
19816
19817
19818
19819
19820
19821
19822
19823
19824
19825
19826
19827
19828
19829
19830
19831
19832
19833
19834
19835
19836
19837
19838
19839
19840
19841
19842
19843
19844
19845
19846
19847
19848
19849
19850
19851
19852
19853
19854
19855
19856
19857
19858
19859
19860
19861
19862
19863
19864
19865
19866
19867
19868
19869
19870
19871
19872
19873
19874
19875
19876
19877
19878
19879
19880
19881
19882
19883
19884
19885
19886
19887
19888
19889
19890
19891
19892
19893
19894
19895
19896
19897
19898
19899
19900
19901
19902
19903
19904
19905
19906
19907
19908
19909
19910
19911
19912
19913
19914
19915
19916
19917
19918
19919
19920
19921
19922
19923
19924
19925
19926
19927
19928
19929
19930
19931
19932
19933
19934
19935
19936
19937
19938
19939
19940
19941
19942
19943
19944
19945
19946
19947
19948
19949
19950
19951
19952
19953
19954
19955
19956
19957
19958
19959
19960
19961
19962
19963
19964
19965
19966
19967
19968
19969
19970
19971
19972
19973
19974
19975
19976
19977
19978
19979
19980
19981
19982
19983
19984
19985
19986
19987
19988
19989
19990
19991
19992
19993
19994
19995
19996
19997
19998
19999
20000
20001
20002
20003
20004
20005
20006
20007
20008
20009
20010
20011
20012
20013
20014
20015
20016
20017
20018
20019
20020
20021
20022
20023
20024
20025
20026
20027
20028
20029
20030
20031
20032
20033
20034
20035
20036
20037
20038
20039
20040
20041
20042
20043
20044
20045
20046
20047
20048
20049
20050
20051
20052
20053
20054
20055
20056
20057
20058
20059
20060
20061
20062
20063
20064
20065
20066
20067
20068
20069
20070
20071
20072
20073
20074
20075
20076
20077
20078
20079
20080
20081
20082
20083
20084
20085
20086
20087
20088
20089
20090
20091
20092
20093
20094
20095
20096
20097
20098
20099
20100
20101
20102
20103
20104
20105
20106
20107
20108
20109
20110
20111
20112
20113
20114
20115
20116
20117
20118
20119
20120
20121
20122
20123
20124
20125
20126
20127
20128
20129
20130
20131
20132
20133
20134
20135
20136
20137
20138
20139
20140
20141
20142
20143
20144
20145
20146
20147
20148
20149
20150
20151
20152
20153
20154
20155
20156
20157
20158
20159
20160
20161
20162
20163
20164
20165
20166
20167
20168
20169
20170
20171
20172
20173
20174
20175
20176
20177
20178
20179
20180
20181
20182
20183
20184
20185
20186
20187
20188
20189
20190
20191
20192
20193
20194
20195
20196
20197
20198
20199
20200
20201
20202
20203
20204
20205
20206
20207
20208
20209
20210
20211
20212
20213
20214
20215
20216
20217
20218
20219
20220
20221
20222
20223
20224
20225
20226
20227
20228
20229
20230
20231
20232
20233
20234
20235
20236
20237
20238
20239
20240
20241
20242
20243
20244
20245
20246
20247
20248
20249
20250
20251
20252
20253
20254
20255
20256
20257
20258
20259
20260
20261
20262
20263
20264
20265
20266
20267
20268
20269
20270
20271
20272
20273
20274
20275
20276
20277
20278
20279
20280
20281
20282
20283
20284
20285
20286
20287
20288
20289
20290
20291
20292
20293
20294
20295
20296
20297
20298
20299
20300
20301
20302
20303
20304
20305
20306
20307
20308
20309
20310
20311
20312
20313
20314
20315
20316
20317
20318
20319
20320
20321
20322
20323
20324
20325
20326
20327
20328
20329
20330
20331
20332
20333
20334
20335
20336
20337
20338
20339
20340
20341
20342
20343
20344
20345
20346
20347
20348
20349
20350
20351
20352
20353
20354
20355
20356
20357
20358
20359
20360
20361
20362
20363
20364
20365
20366
20367
20368
20369
20370
20371
20372
20373
20374
20375
20376
20377
20378
20379
20380
20381
20382
20383
20384
20385
20386
20387
20388
20389
20390
20391
20392
20393
20394
20395
20396
20397
20398
20399
20400
20401
20402
20403
20404
20405
20406
20407
20408
20409
20410
20411
20412
20413
20414
20415
20416
20417
20418
20419
20420
20421
20422
20423
20424
20425
20426
20427
20428
20429
20430
20431
20432
20433
20434
20435
20436
20437
20438
20439
20440
20441
20442
20443
20444
20445
20446
20447
20448
20449
20450
20451
20452
20453
20454
20455
20456
20457
20458
20459
20460
20461
20462
20463
20464
20465
20466
20467
20468
20469
20470
20471
20472
20473
20474
20475
20476
20477
20478
20479
20480
20481
20482
20483
20484
20485
20486
20487
20488
20489
20490
20491
20492
20493
20494
20495
20496
20497
20498
20499
20500
20501
20502
20503
20504
20505
20506
20507
20508
20509
20510
20511
20512
20513
20514
20515
20516
20517
20518
20519
20520
20521
20522
20523
20524
20525
20526
20527
20528
20529
20530
20531
20532
20533
20534
20535
20536
20537
20538
20539
20540
20541
20542
20543
20544
20545
20546
20547
20548
20549
20550
20551
20552
20553
20554
20555
20556
20557
20558
20559
20560
20561
20562
20563
20564
20565
20566
20567
20568
20569
20570
20571
20572
20573
20574
20575
20576
20577
20578
20579
20580
20581
20582
20583
20584
20585
20586
20587
20588
20589
20590
20591
20592
20593
20594
20595
20596
20597
20598
20599
20600
20601
20602
20603
20604
20605
20606
20607
20608
20609
20610
20611
20612
20613
20614
20615
20616
20617
20618
20619
20620
20621
20622
20623
20624
20625
20626
20627
20628
20629
20630
20631
20632
20633
20634
20635
20636
20637
20638
20639
20640
20641
20642
20643
20644
20645
20646
20647
20648
20649
20650
20651
20652
20653
20654
20655
20656
20657
20658
20659
20660
20661
20662
20663
20664
20665
20666
20667
20668
20669
20670
20671
20672
20673
20674
20675
20676
20677
20678
20679
20680
20681
20682
20683
20684
20685
20686
20687
20688
20689
20690
20691
20692
20693
20694
20695
20696
20697
20698
20699
20700
20701
20702
20703
20704
20705
20706
20707
20708
20709
20710
20711
20712
20713
20714
20715
20716
20717
20718
20719
20720
20721
20722
20723
20724
20725
20726
20727
20728
20729
20730
20731
20732
20733
20734
20735
20736
20737
20738
20739
20740
20741
20742
20743
20744
20745
20746
20747
20748
20749
20750
20751
20752
20753
20754
20755
20756
20757
20758
20759
20760
20761
20762
20763
20764
20765
20766
20767
20768
20769
20770
20771
20772
20773
20774
20775
20776
20777
20778
20779
20780
20781
20782
20783
20784
20785
20786
20787
20788
20789
20790
20791
20792
20793
20794
20795
20796
20797
20798
20799
20800
20801
20802
20803
20804
20805
20806
20807
20808
20809
20810
20811
20812
20813
20814
20815
20816
20817
20818
20819
20820
20821
20822
20823
20824
20825
20826
20827
20828
20829
20830
20831
20832
20833
20834
20835
20836
20837
20838
20839
20840
20841
20842
20843
20844
20845
20846
20847
20848
20849
20850
20851
20852
20853
20854
20855
20856
20857
20858
20859
20860
20861
20862
20863
20864
20865
20866
20867
20868
20869
20870
20871
20872
20873
20874
20875
20876
20877
20878
20879
20880
20881
20882
20883
20884
20885
20886
20887
20888
20889
20890
20891
20892
20893
20894
20895
20896
20897
20898
20899
20900
20901
20902
20903
20904
20905
20906
20907
20908
20909
20910
20911
20912
20913
20914
20915
20916
20917
20918
20919
20920
20921
20922
20923
20924
20925
20926
20927
20928
20929
20930
20931
20932
20933
20934
20935
20936
20937
20938
20939
20940
20941
20942
20943
20944
20945
20946
20947
20948
20949
20950
20951
20952
20953
20954
20955
20956
20957
20958
20959
20960
20961
20962
20963
20964
20965
20966
20967
20968
20969
20970
20971
20972
20973
20974
20975
20976
20977
20978
20979
20980
20981
20982
20983
20984
20985
20986
20987
20988
20989
20990
20991
20992
20993
20994
20995
20996
20997
20998
20999
21000
21001
21002
21003
21004
21005
21006
21007
21008
21009
21010
21011
21012
21013
21014
21015
21016
21017
21018
21019
21020
21021
21022
21023
21024
21025
21026
21027
21028
21029
21030
21031
21032
21033
21034
21035
21036
21037
21038
21039
21040
21041
21042
21043
21044
21045
21046
21047
21048
21049
21050
21051
21052
21053
21054
21055
21056
21057
21058
21059
21060
21061
21062
21063
21064
21065
21066
21067
21068
21069
21070
21071
21072
21073
21074
21075
21076
21077
21078
21079
21080
21081
21082
21083
21084
21085
21086
21087
21088
21089
21090
21091
21092
21093
21094
21095
21096
21097
21098
21099
21100
21101
21102
21103
21104
21105
21106
21107
21108
21109
21110
21111
21112
21113
21114
21115
21116
21117
21118
21119
21120
21121
21122
21123
21124
21125
21126
21127
21128
21129
21130
21131
21132
21133
21134
21135
21136
21137
21138
21139
21140
21141
21142
21143
21144
21145
21146
21147
21148
21149
21150
21151
21152
21153
21154
21155
21156
21157
21158
21159
21160
21161
21162
21163
21164
21165
21166
21167
21168
21169
21170
21171
21172
21173
21174
21175
21176
21177
21178
21179
21180
21181
21182
21183
21184
21185
21186
21187
21188
21189
21190
21191
21192
21193
21194
21195
21196
21197
21198
21199
21200
21201
21202
21203
21204
21205
21206
21207
21208
21209
21210
21211
21212
21213
21214
21215
21216
21217
21218
21219
21220
21221
21222
21223
21224
21225
21226
21227
21228
21229
21230
21231
21232
21233
21234
21235
21236
21237
21238
21239
21240
21241
21242
21243
21244
21245
21246
21247
21248
21249
21250
21251
21252
21253
21254
21255
21256
21257
21258
21259
21260
21261
21262
21263
21264
21265
21266
21267
21268
21269
21270
21271
21272
21273
21274
21275
21276
21277
21278
21279
21280
21281
21282
21283
21284
21285
21286
21287
21288
21289
21290
21291
21292
21293
21294
21295
21296
21297
21298
21299
21300
21301
21302
21303
21304
21305
21306
21307
21308
21309
21310
21311
21312
21313
21314
21315
21316
21317
21318
21319
21320
21321
21322
21323
21324
21325
21326
21327
21328
21329
21330
21331
21332
21333
21334
21335
21336
21337
21338
21339
21340
21341
21342
21343
21344
21345
21346
21347
21348
21349
21350
21351
21352
21353
21354
21355
21356
21357
21358
21359
21360
21361
21362
21363
21364
21365
21366
21367
21368
21369
21370
21371
21372
21373
21374
21375
21376
21377
21378
21379
21380
21381
21382
21383
21384
21385
21386
21387
21388
21389
21390
21391
21392
21393
21394
21395
21396
21397
21398
21399
21400
21401
21402
21403
21404
21405
21406
21407
21408
21409
21410
21411
21412
21413
21414
21415
21416
21417
21418
21419
21420
21421
21422
21423
21424
21425
21426
21427
21428
21429
21430
21431
21432
21433
21434
21435
21436
21437
21438
21439
21440
21441
21442
21443
21444
21445
21446
21447
21448
21449
21450
21451
21452
21453
21454
21455
21456
21457
21458
21459
21460
21461
21462
|
//===-- RISCVISelLowering.cpp - RISC-V DAG Lowering Implementation -------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file defines the interfaces that RISC-V uses to lower LLVM code into a
// selection DAG.
//
//===----------------------------------------------------------------------===//
#include "RISCVISelLowering.h"
#include "MCTargetDesc/RISCVMatInt.h"
#include "RISCV.h"
#include "RISCVMachineFunctionInfo.h"
#include "RISCVRegisterInfo.h"
#include "RISCVSubtarget.h"
#include "RISCVTargetMachine.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAGAddressAnalysis.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/DiagnosticPrinter.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicsRISCV.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/InstructionCost.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <optional>
using namespace llvm;
#define DEBUG_TYPE "riscv-lower"
STATISTIC(NumTailCalls, "Number of tail calls");
static cl::opt<unsigned> ExtensionMaxWebSize(
DEBUG_TYPE "-ext-max-web-size", cl::Hidden,
cl::desc("Give the maximum size (in number of nodes) of the web of "
"instructions that we will consider for VW expansion"),
cl::init(18));
static cl::opt<bool>
AllowSplatInVW_W(DEBUG_TYPE "-form-vw-w-with-splat", cl::Hidden,
cl::desc("Allow the formation of VW_W operations (e.g., "
"VWADD_W) with splat constants"),
cl::init(false));
static cl::opt<unsigned> NumRepeatedDivisors(
DEBUG_TYPE "-fp-repeated-divisors", cl::Hidden,
cl::desc("Set the minimum number of repetitions of a divisor to allow "
"transformation to multiplications by the reciprocal"),
cl::init(2));
static cl::opt<int>
FPImmCost(DEBUG_TYPE "-fpimm-cost", cl::Hidden,
cl::desc("Give the maximum number of instructions that we will "
"use for creating a floating-point immediate value"),
cl::init(2));
static cl::opt<bool>
RV64LegalI32("riscv-experimental-rv64-legal-i32", cl::ReallyHidden,
cl::desc("Make i32 a legal type for SelectionDAG on RV64."));
RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
const RISCVSubtarget &STI)
: TargetLowering(TM), Subtarget(STI) {
RISCVABI::ABI ABI = Subtarget.getTargetABI();
assert(ABI != RISCVABI::ABI_Unknown && "Improperly initialised target ABI");
if ((ABI == RISCVABI::ABI_ILP32F || ABI == RISCVABI::ABI_LP64F) &&
!Subtarget.hasStdExtF()) {
errs() << "Hard-float 'f' ABI can't be used for a target that "
"doesn't support the F instruction set extension (ignoring "
"target-abi)\n";
ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32;
} else if ((ABI == RISCVABI::ABI_ILP32D || ABI == RISCVABI::ABI_LP64D) &&
!Subtarget.hasStdExtD()) {
errs() << "Hard-float 'd' ABI can't be used for a target that "
"doesn't support the D instruction set extension (ignoring "
"target-abi)\n";
ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32;
}
switch (ABI) {
default:
report_fatal_error("Don't know how to lower this ABI");
case RISCVABI::ABI_ILP32:
case RISCVABI::ABI_ILP32E:
case RISCVABI::ABI_LP64E:
case RISCVABI::ABI_ILP32F:
case RISCVABI::ABI_ILP32D:
case RISCVABI::ABI_LP64:
case RISCVABI::ABI_LP64F:
case RISCVABI::ABI_LP64D:
break;
}
MVT XLenVT = Subtarget.getXLenVT();
// Set up the register classes.
addRegisterClass(XLenVT, &RISCV::GPRRegClass);
if (Subtarget.is64Bit() && RV64LegalI32)
addRegisterClass(MVT::i32, &RISCV::GPRRegClass);
if (Subtarget.hasStdExtZfhmin())
addRegisterClass(MVT::f16, &RISCV::FPR16RegClass);
if (Subtarget.hasStdExtZfbfmin())
addRegisterClass(MVT::bf16, &RISCV::FPR16RegClass);
if (Subtarget.hasStdExtF())
addRegisterClass(MVT::f32, &RISCV::FPR32RegClass);
if (Subtarget.hasStdExtD())
addRegisterClass(MVT::f64, &RISCV::FPR64RegClass);
if (Subtarget.hasStdExtZhinxmin())
addRegisterClass(MVT::f16, &RISCV::GPRF16RegClass);
if (Subtarget.hasStdExtZfinx())
addRegisterClass(MVT::f32, &RISCV::GPRF32RegClass);
if (Subtarget.hasStdExtZdinx()) {
if (Subtarget.is64Bit())
addRegisterClass(MVT::f64, &RISCV::GPRRegClass);
else
addRegisterClass(MVT::f64, &RISCV::GPRPairRegClass);
}
static const MVT::SimpleValueType BoolVecVTs[] = {
MVT::nxv1i1, MVT::nxv2i1, MVT::nxv4i1, MVT::nxv8i1,
MVT::nxv16i1, MVT::nxv32i1, MVT::nxv64i1};
static const MVT::SimpleValueType IntVecVTs[] = {
MVT::nxv1i8, MVT::nxv2i8, MVT::nxv4i8, MVT::nxv8i8, MVT::nxv16i8,
MVT::nxv32i8, MVT::nxv64i8, MVT::nxv1i16, MVT::nxv2i16, MVT::nxv4i16,
MVT::nxv8i16, MVT::nxv16i16, MVT::nxv32i16, MVT::nxv1i32, MVT::nxv2i32,
MVT::nxv4i32, MVT::nxv8i32, MVT::nxv16i32, MVT::nxv1i64, MVT::nxv2i64,
MVT::nxv4i64, MVT::nxv8i64};
static const MVT::SimpleValueType F16VecVTs[] = {
MVT::nxv1f16, MVT::nxv2f16, MVT::nxv4f16,
MVT::nxv8f16, MVT::nxv16f16, MVT::nxv32f16};
static const MVT::SimpleValueType BF16VecVTs[] = {
MVT::nxv1bf16, MVT::nxv2bf16, MVT::nxv4bf16,
MVT::nxv8bf16, MVT::nxv16bf16, MVT::nxv32bf16};
static const MVT::SimpleValueType F32VecVTs[] = {
MVT::nxv1f32, MVT::nxv2f32, MVT::nxv4f32, MVT::nxv8f32, MVT::nxv16f32};
static const MVT::SimpleValueType F64VecVTs[] = {
MVT::nxv1f64, MVT::nxv2f64, MVT::nxv4f64, MVT::nxv8f64};
if (Subtarget.hasVInstructions()) {
auto addRegClassForRVV = [this](MVT VT) {
// Disable the smallest fractional LMUL types if ELEN is less than
// RVVBitsPerBlock.
unsigned MinElts = RISCV::RVVBitsPerBlock / Subtarget.getELen();
if (VT.getVectorMinNumElements() < MinElts)
return;
unsigned Size = VT.getSizeInBits().getKnownMinValue();
const TargetRegisterClass *RC;
if (Size <= RISCV::RVVBitsPerBlock)
RC = &RISCV::VRRegClass;
else if (Size == 2 * RISCV::RVVBitsPerBlock)
RC = &RISCV::VRM2RegClass;
else if (Size == 4 * RISCV::RVVBitsPerBlock)
RC = &RISCV::VRM4RegClass;
else if (Size == 8 * RISCV::RVVBitsPerBlock)
RC = &RISCV::VRM8RegClass;
else
llvm_unreachable("Unexpected size");
addRegisterClass(VT, RC);
};
for (MVT VT : BoolVecVTs)
addRegClassForRVV(VT);
for (MVT VT : IntVecVTs) {
if (VT.getVectorElementType() == MVT::i64 &&
!Subtarget.hasVInstructionsI64())
continue;
addRegClassForRVV(VT);
}
if (Subtarget.hasVInstructionsF16Minimal())
for (MVT VT : F16VecVTs)
addRegClassForRVV(VT);
if (Subtarget.hasVInstructionsBF16())
for (MVT VT : BF16VecVTs)
addRegClassForRVV(VT);
if (Subtarget.hasVInstructionsF32())
for (MVT VT : F32VecVTs)
addRegClassForRVV(VT);
if (Subtarget.hasVInstructionsF64())
for (MVT VT : F64VecVTs)
addRegClassForRVV(VT);
if (Subtarget.useRVVForFixedLengthVectors()) {
auto addRegClassForFixedVectors = [this](MVT VT) {
MVT ContainerVT = getContainerForFixedLengthVector(VT);
unsigned RCID = getRegClassIDForVecVT(ContainerVT);
const RISCVRegisterInfo &TRI = *Subtarget.getRegisterInfo();
addRegisterClass(VT, TRI.getRegClass(RCID));
};
for (MVT VT : MVT::integer_fixedlen_vector_valuetypes())
if (useRVVForFixedLengthVectorVT(VT))
addRegClassForFixedVectors(VT);
for (MVT VT : MVT::fp_fixedlen_vector_valuetypes())
if (useRVVForFixedLengthVectorVT(VT))
addRegClassForFixedVectors(VT);
}
}
// Compute derived properties from the register classes.
computeRegisterProperties(STI.getRegisterInfo());
setStackPointerRegisterToSaveRestore(RISCV::X2);
setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, XLenVT,
MVT::i1, Promote);
// DAGCombiner can call isLoadExtLegal for types that aren't legal.
setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i32,
MVT::i1, Promote);
// TODO: add all necessary setOperationAction calls.
setOperationAction(ISD::DYNAMIC_STACKALLOC, XLenVT, Expand);
setOperationAction(ISD::BR_JT, MVT::Other, Expand);
setOperationAction(ISD::BR_CC, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::BR_CC, MVT::i32, Expand);
setOperationAction(ISD::BRCOND, MVT::Other, Custom);
setOperationAction(ISD::SELECT_CC, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
setCondCodeAction(ISD::SETLE, XLenVT, Expand);
setCondCodeAction(ISD::SETGT, XLenVT, Custom);
setCondCodeAction(ISD::SETGE, XLenVT, Expand);
setCondCodeAction(ISD::SETULE, XLenVT, Expand);
setCondCodeAction(ISD::SETUGT, XLenVT, Custom);
setCondCodeAction(ISD::SETUGE, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::SETCC, MVT::i32, Promote);
setOperationAction({ISD::STACKSAVE, ISD::STACKRESTORE}, MVT::Other, Expand);
setOperationAction(ISD::VASTART, MVT::Other, Custom);
setOperationAction({ISD::VAARG, ISD::VACOPY, ISD::VAEND}, MVT::Other, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::VAARG, MVT::i32, Promote);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
if (!Subtarget.hasStdExtZbb() && !Subtarget.hasVendorXTHeadBb())
setOperationAction(ISD::SIGN_EXTEND_INREG, {MVT::i8, MVT::i16}, Expand);
if (Subtarget.is64Bit()) {
setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom);
if (!RV64LegalI32) {
setOperationAction(ISD::LOAD, MVT::i32, Custom);
setOperationAction({ISD::ADD, ISD::SUB, ISD::SHL, ISD::SRA, ISD::SRL},
MVT::i32, Custom);
setOperationAction({ISD::UADDO, ISD::USUBO, ISD::UADDSAT, ISD::USUBSAT},
MVT::i32, Custom);
if (!Subtarget.hasStdExtZbb())
setOperationAction({ISD::SADDSAT, ISD::SSUBSAT}, MVT::i32, Custom);
} else {
setOperationAction(ISD::SSUBO, MVT::i32, Custom);
if (Subtarget.hasStdExtZbb()) {
setOperationAction({ISD::SADDSAT, ISD::SSUBSAT}, MVT::i32, Custom);
setOperationAction({ISD::UADDSAT, ISD::USUBSAT}, MVT::i32, Custom);
}
}
setOperationAction(ISD::SADDO, MVT::i32, Custom);
} else {
setLibcallName(
{RTLIB::SHL_I128, RTLIB::SRL_I128, RTLIB::SRA_I128, RTLIB::MUL_I128},
nullptr);
setLibcallName(RTLIB::MULO_I64, nullptr);
}
if (!Subtarget.hasStdExtM() && !Subtarget.hasStdExtZmmul()) {
setOperationAction({ISD::MUL, ISD::MULHS, ISD::MULHU}, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::MUL, MVT::i32, Promote);
} else if (Subtarget.is64Bit()) {
setOperationAction(ISD::MUL, MVT::i128, Custom);
if (!RV64LegalI32)
setOperationAction(ISD::MUL, MVT::i32, Custom);
else
setOperationAction(ISD::SMULO, MVT::i32, Custom);
} else {
setOperationAction(ISD::MUL, MVT::i64, Custom);
}
if (!Subtarget.hasStdExtM()) {
setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM},
XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM}, MVT::i32,
Promote);
} else if (Subtarget.is64Bit()) {
if (!RV64LegalI32)
setOperationAction({ISD::SDIV, ISD::UDIV, ISD::UREM},
{MVT::i8, MVT::i16, MVT::i32}, Custom);
}
if (RV64LegalI32 && Subtarget.is64Bit()) {
setOperationAction({ISD::MULHS, ISD::MULHU}, MVT::i32, Expand);
setOperationAction(
{ISD::SDIVREM, ISD::UDIVREM, ISD::SMUL_LOHI, ISD::UMUL_LOHI}, MVT::i32,
Expand);
}
setOperationAction(
{ISD::SDIVREM, ISD::UDIVREM, ISD::SMUL_LOHI, ISD::UMUL_LOHI}, XLenVT,
Expand);
setOperationAction({ISD::SHL_PARTS, ISD::SRL_PARTS, ISD::SRA_PARTS}, XLenVT,
Custom);
if (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()) {
if (!RV64LegalI32 && Subtarget.is64Bit())
setOperationAction({ISD::ROTL, ISD::ROTR}, MVT::i32, Custom);
} else if (Subtarget.hasVendorXTHeadBb()) {
if (Subtarget.is64Bit())
setOperationAction({ISD::ROTL, ISD::ROTR}, MVT::i32, Custom);
setOperationAction({ISD::ROTL, ISD::ROTR}, XLenVT, Custom);
} else if (Subtarget.hasVendorXCVbitmanip()) {
setOperationAction(ISD::ROTL, XLenVT, Expand);
} else {
setOperationAction({ISD::ROTL, ISD::ROTR}, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction({ISD::ROTL, ISD::ROTR}, MVT::i32, Expand);
}
// With Zbb we have an XLen rev8 instruction, but not GREVI. So we'll
// pattern match it directly in isel.
setOperationAction(ISD::BSWAP, XLenVT,
(Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb() ||
Subtarget.hasVendorXTHeadBb())
? Legal
: Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::BSWAP, MVT::i32,
(Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb() ||
Subtarget.hasVendorXTHeadBb())
? Promote
: Expand);
if (Subtarget.hasVendorXCVbitmanip()) {
setOperationAction(ISD::BITREVERSE, XLenVT, Legal);
} else {
// Zbkb can use rev8+brev8 to implement bitreverse.
setOperationAction(ISD::BITREVERSE, XLenVT,
Subtarget.hasStdExtZbkb() ? Custom : Expand);
}
if (Subtarget.hasStdExtZbb()) {
setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, XLenVT,
Legal);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, MVT::i32,
Promote);
if (Subtarget.is64Bit()) {
if (RV64LegalI32)
setOperationAction(ISD::CTTZ, MVT::i32, Legal);
else
setOperationAction({ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF}, MVT::i32, Custom);
}
} else if (!Subtarget.hasVendorXCVbitmanip()) {
setOperationAction({ISD::CTTZ, ISD::CTPOP}, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction({ISD::CTTZ, ISD::CTPOP}, MVT::i32, Expand);
}
if (Subtarget.hasStdExtZbb() || Subtarget.hasVendorXTHeadBb() ||
Subtarget.hasVendorXCVbitmanip()) {
// We need the custom lowering to make sure that the resulting sequence
// for the 32bit case is efficient on 64bit targets.
if (Subtarget.is64Bit()) {
if (RV64LegalI32) {
setOperationAction(ISD::CTLZ, MVT::i32,
Subtarget.hasStdExtZbb() ? Legal : Promote);
if (!Subtarget.hasStdExtZbb())
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Promote);
} else
setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, MVT::i32, Custom);
}
} else {
setOperationAction(ISD::CTLZ, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::CTLZ, MVT::i32, Expand);
}
if (!RV64LegalI32 && Subtarget.is64Bit() &&
!Subtarget.hasShortForwardBranchOpt())
setOperationAction(ISD::ABS, MVT::i32, Custom);
// We can use PseudoCCSUB to implement ABS.
if (Subtarget.hasShortForwardBranchOpt())
setOperationAction(ISD::ABS, XLenVT, Legal);
if (!Subtarget.hasVendorXTHeadCondMov()) {
setOperationAction(ISD::SELECT, XLenVT, Custom);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::SELECT, MVT::i32, Promote);
}
static const unsigned FPLegalNodeTypes[] = {
ISD::FMINNUM, ISD::FMAXNUM, ISD::LRINT,
ISD::LLRINT, ISD::LROUND, ISD::LLROUND,
ISD::STRICT_LRINT, ISD::STRICT_LLRINT, ISD::STRICT_LROUND,
ISD::STRICT_LLROUND, ISD::STRICT_FMA, ISD::STRICT_FADD,
ISD::STRICT_FSUB, ISD::STRICT_FMUL, ISD::STRICT_FDIV,
ISD::STRICT_FSQRT, ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS};
static const ISD::CondCode FPCCToExpand[] = {
ISD::SETOGT, ISD::SETOGE, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUNE, ISD::SETGT,
ISD::SETGE, ISD::SETNE, ISD::SETO, ISD::SETUO};
static const unsigned FPOpToExpand[] = {
ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW,
ISD::FREM};
static const unsigned FPRndMode[] = {
ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FRINT, ISD::FROUND,
ISD::FROUNDEVEN};
if (Subtarget.hasStdExtZfhminOrZhinxmin())
setOperationAction(ISD::BITCAST, MVT::i16, Custom);
static const unsigned ZfhminZfbfminPromoteOps[] = {
ISD::FMINNUM, ISD::FMAXNUM, ISD::FADD,
ISD::FSUB, ISD::FMUL, ISD::FMA,
ISD::FDIV, ISD::FSQRT, ISD::FABS,
ISD::FNEG, ISD::STRICT_FMA, ISD::STRICT_FADD,
ISD::STRICT_FSUB, ISD::STRICT_FMUL, ISD::STRICT_FDIV,
ISD::STRICT_FSQRT, ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS,
ISD::SETCC, ISD::FCEIL, ISD::FFLOOR,
ISD::FTRUNC, ISD::FRINT, ISD::FROUND,
ISD::FROUNDEVEN, ISD::SELECT};
if (Subtarget.hasStdExtZfbfmin()) {
setOperationAction(ISD::BITCAST, MVT::i16, Custom);
setOperationAction(ISD::BITCAST, MVT::bf16, Custom);
setOperationAction(ISD::FP_ROUND, MVT::bf16, Custom);
setOperationAction(ISD::FP_EXTEND, MVT::f32, Custom);
setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
setOperationAction(ISD::ConstantFP, MVT::bf16, Expand);
setOperationAction(ISD::SELECT_CC, MVT::bf16, Expand);
setOperationAction(ISD::BR_CC, MVT::bf16, Expand);
setOperationAction(ZfhminZfbfminPromoteOps, MVT::bf16, Promote);
setOperationAction(ISD::FREM, MVT::bf16, Promote);
// FIXME: Need to promote bf16 FCOPYSIGN to f32, but the
// DAGCombiner::visitFP_ROUND probably needs improvements first.
setOperationAction(ISD::FCOPYSIGN, MVT::bf16, Expand);
}
if (Subtarget.hasStdExtZfhminOrZhinxmin()) {
if (Subtarget.hasStdExtZfhOrZhinx()) {
setOperationAction(FPLegalNodeTypes, MVT::f16, Legal);
setOperationAction(FPRndMode, MVT::f16,
Subtarget.hasStdExtZfa() ? Legal : Custom);
setOperationAction(ISD::SELECT, MVT::f16, Custom);
setOperationAction(ISD::IS_FPCLASS, MVT::f16, Custom);
} else {
setOperationAction(ZfhminZfbfminPromoteOps, MVT::f16, Promote);
setOperationAction({ISD::STRICT_LRINT, ISD::STRICT_LLRINT,
ISD::STRICT_LROUND, ISD::STRICT_LLROUND},
MVT::f16, Legal);
// FIXME: Need to promote f16 FCOPYSIGN to f32, but the
// DAGCombiner::visitFP_ROUND probably needs improvements first.
setOperationAction(ISD::FCOPYSIGN, MVT::f16, Expand);
}
setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Legal);
setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Legal);
setCondCodeAction(FPCCToExpand, MVT::f16, Expand);
setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
setOperationAction(ISD::BR_CC, MVT::f16, Expand);
setOperationAction(ISD::FNEARBYINT, MVT::f16,
Subtarget.hasStdExtZfa() ? Legal : Promote);
setOperationAction({ISD::FREM, ISD::FPOW, ISD::FPOWI,
ISD::FCOS, ISD::FSIN, ISD::FSINCOS, ISD::FEXP,
ISD::FEXP2, ISD::FEXP10, ISD::FLOG, ISD::FLOG2,
ISD::FLOG10},
MVT::f16, Promote);
// FIXME: Need to promote f16 STRICT_* to f32 libcalls, but we don't have
// complete support for all operations in LegalizeDAG.
setOperationAction({ISD::STRICT_FCEIL, ISD::STRICT_FFLOOR,
ISD::STRICT_FNEARBYINT, ISD::STRICT_FRINT,
ISD::STRICT_FROUND, ISD::STRICT_FROUNDEVEN,
ISD::STRICT_FTRUNC},
MVT::f16, Promote);
// We need to custom promote this.
if (Subtarget.is64Bit())
setOperationAction(ISD::FPOWI, MVT::i32, Custom);
if (!Subtarget.hasStdExtZfa())
setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f16, Custom);
}
if (Subtarget.hasStdExtFOrZfinx()) {
setOperationAction(FPLegalNodeTypes, MVT::f32, Legal);
setOperationAction(FPRndMode, MVT::f32,
Subtarget.hasStdExtZfa() ? Legal : Custom);
setCondCodeAction(FPCCToExpand, MVT::f32, Expand);
setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
setOperationAction(ISD::SELECT, MVT::f32, Custom);
setOperationAction(ISD::BR_CC, MVT::f32, Expand);
setOperationAction(FPOpToExpand, MVT::f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
setTruncStoreAction(MVT::f32, MVT::f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::bf16, Expand);
setTruncStoreAction(MVT::f32, MVT::bf16, Expand);
setOperationAction(ISD::IS_FPCLASS, MVT::f32, Custom);
setOperationAction(ISD::BF16_TO_FP, MVT::f32, Custom);
setOperationAction(ISD::FP_TO_BF16, MVT::f32,
Subtarget.isSoftFPABI() ? LibCall : Custom);
setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
setOperationAction(ISD::FP16_TO_FP, MVT::f32, Custom);
if (Subtarget.hasStdExtZfa())
setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
else
setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f32, Custom);
}
if (Subtarget.hasStdExtFOrZfinx() && Subtarget.is64Bit())
setOperationAction(ISD::BITCAST, MVT::i32, Custom);
if (Subtarget.hasStdExtDOrZdinx()) {
setOperationAction(FPLegalNodeTypes, MVT::f64, Legal);
if (!Subtarget.is64Bit())
setOperationAction(ISD::BITCAST, MVT::i64, Custom);
if (Subtarget.hasStdExtZfa()) {
setOperationAction(FPRndMode, MVT::f64, Legal);
setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
} else {
if (Subtarget.is64Bit())
setOperationAction(FPRndMode, MVT::f64, Custom);
setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f64, Custom);
}
setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
setCondCodeAction(FPCCToExpand, MVT::f64, Expand);
setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
setOperationAction(ISD::SELECT, MVT::f64, Custom);
setOperationAction(ISD::BR_CC, MVT::f64, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
setOperationAction(FPOpToExpand, MVT::f64, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
setTruncStoreAction(MVT::f64, MVT::f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::bf16, Expand);
setTruncStoreAction(MVT::f64, MVT::bf16, Expand);
setOperationAction(ISD::IS_FPCLASS, MVT::f64, Custom);
setOperationAction(ISD::BF16_TO_FP, MVT::f64, Custom);
setOperationAction(ISD::FP_TO_BF16, MVT::f64,
Subtarget.isSoftFPABI() ? LibCall : Custom);
setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
}
if (Subtarget.is64Bit()) {
setOperationAction({ISD::FP_TO_UINT, ISD::FP_TO_SINT,
ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT},
MVT::i32, Custom);
setOperationAction(ISD::LROUND, MVT::i32, Custom);
}
if (Subtarget.hasStdExtFOrZfinx()) {
setOperationAction({ISD::FP_TO_UINT_SAT, ISD::FP_TO_SINT_SAT}, XLenVT,
Custom);
setOperationAction({ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT,
ISD::STRICT_UINT_TO_FP, ISD::STRICT_SINT_TO_FP},
XLenVT, Legal);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction({ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT,
ISD::STRICT_UINT_TO_FP, ISD::STRICT_SINT_TO_FP},
MVT::i32, Legal);
setOperationAction(ISD::GET_ROUNDING, XLenVT, Custom);
setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
}
setOperationAction({ISD::GlobalAddress, ISD::BlockAddress, ISD::ConstantPool,
ISD::JumpTable},
XLenVT, Custom);
setOperationAction(ISD::GlobalTLSAddress, XLenVT, Custom);
if (Subtarget.is64Bit())
setOperationAction(ISD::Constant, MVT::i64, Custom);
// TODO: On M-mode only targets, the cycle[h]/time[h] CSR may not be present.
// Unfortunately this can't be determined just from the ISA naming string.
setOperationAction(ISD::READCYCLECOUNTER, MVT::i64,
Subtarget.is64Bit() ? Legal : Custom);
setOperationAction(ISD::READSTEADYCOUNTER, MVT::i64,
Subtarget.is64Bit() ? Legal : Custom);
setOperationAction({ISD::TRAP, ISD::DEBUGTRAP}, MVT::Other, Legal);
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
if (Subtarget.is64Bit())
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i32, Custom);
if (Subtarget.hasStdExtZicbop()) {
setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
}
if (Subtarget.hasStdExtA()) {
setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
if (Subtarget.hasStdExtZabha() && Subtarget.hasStdExtZacas())
setMinCmpXchgSizeInBits(8);
else
setMinCmpXchgSizeInBits(32);
} else if (Subtarget.hasForcedAtomics()) {
setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
} else {
setMaxAtomicSizeInBitsSupported(0);
}
setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
setBooleanContents(ZeroOrOneBooleanContent);
if (Subtarget.hasVInstructions()) {
setBooleanVectorContents(ZeroOrOneBooleanContent);
setOperationAction(ISD::VSCALE, XLenVT, Custom);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::VSCALE, MVT::i32, Custom);
// RVV intrinsics may have illegal operands.
// We also need to custom legalize vmv.x.s.
setOperationAction({ISD::INTRINSIC_WO_CHAIN, ISD::INTRINSIC_W_CHAIN,
ISD::INTRINSIC_VOID},
{MVT::i8, MVT::i16}, Custom);
if (Subtarget.is64Bit())
setOperationAction({ISD::INTRINSIC_W_CHAIN, ISD::INTRINSIC_VOID},
MVT::i32, Custom);
else
setOperationAction({ISD::INTRINSIC_WO_CHAIN, ISD::INTRINSIC_W_CHAIN},
MVT::i64, Custom);
setOperationAction({ISD::INTRINSIC_W_CHAIN, ISD::INTRINSIC_VOID},
MVT::Other, Custom);
static const unsigned IntegerVPOps[] = {
ISD::VP_ADD, ISD::VP_SUB, ISD::VP_MUL,
ISD::VP_SDIV, ISD::VP_UDIV, ISD::VP_SREM,
ISD::VP_UREM, ISD::VP_AND, ISD::VP_OR,
ISD::VP_XOR, ISD::VP_ASHR, ISD::VP_LSHR,
ISD::VP_SHL, ISD::VP_REDUCE_ADD, ISD::VP_REDUCE_AND,
ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR, ISD::VP_REDUCE_SMAX,
ISD::VP_REDUCE_SMIN, ISD::VP_REDUCE_UMAX, ISD::VP_REDUCE_UMIN,
ISD::VP_MERGE, ISD::VP_SELECT, ISD::VP_FP_TO_SINT,
ISD::VP_FP_TO_UINT, ISD::VP_SETCC, ISD::VP_SIGN_EXTEND,
ISD::VP_ZERO_EXTEND, ISD::VP_TRUNCATE, ISD::VP_SMIN,
ISD::VP_SMAX, ISD::VP_UMIN, ISD::VP_UMAX,
ISD::VP_ABS, ISD::EXPERIMENTAL_VP_REVERSE, ISD::EXPERIMENTAL_VP_SPLICE,
ISD::VP_SADDSAT, ISD::VP_UADDSAT, ISD::VP_SSUBSAT,
ISD::VP_USUBSAT};
static const unsigned FloatingPointVPOps[] = {
ISD::VP_FADD, ISD::VP_FSUB, ISD::VP_FMUL,
ISD::VP_FDIV, ISD::VP_FNEG, ISD::VP_FABS,
ISD::VP_FMA, ISD::VP_REDUCE_FADD, ISD::VP_REDUCE_SEQ_FADD,
ISD::VP_REDUCE_FMIN, ISD::VP_REDUCE_FMAX, ISD::VP_MERGE,
ISD::VP_SELECT, ISD::VP_SINT_TO_FP, ISD::VP_UINT_TO_FP,
ISD::VP_SETCC, ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND,
ISD::VP_SQRT, ISD::VP_FMINNUM, ISD::VP_FMAXNUM,
ISD::VP_FCEIL, ISD::VP_FFLOOR, ISD::VP_FROUND,
ISD::VP_FROUNDEVEN, ISD::VP_FCOPYSIGN, ISD::VP_FROUNDTOZERO,
ISD::VP_FRINT, ISD::VP_FNEARBYINT, ISD::VP_IS_FPCLASS,
ISD::VP_FMINIMUM, ISD::VP_FMAXIMUM, ISD::VP_LRINT,
ISD::VP_LLRINT, ISD::EXPERIMENTAL_VP_REVERSE,
ISD::EXPERIMENTAL_VP_SPLICE};
static const unsigned IntegerVecReduceOps[] = {
ISD::VECREDUCE_ADD, ISD::VECREDUCE_AND, ISD::VECREDUCE_OR,
ISD::VECREDUCE_XOR, ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN,
ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN};
static const unsigned FloatingPointVecReduceOps[] = {
ISD::VECREDUCE_FADD, ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_FMIN,
ISD::VECREDUCE_FMAX, ISD::VECREDUCE_FMINIMUM, ISD::VECREDUCE_FMAXIMUM};
if (!Subtarget.is64Bit()) {
// We must custom-lower certain vXi64 operations on RV32 due to the vector
// element type being illegal.
setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT},
MVT::i64, Custom);
setOperationAction(IntegerVecReduceOps, MVT::i64, Custom);
setOperationAction({ISD::VP_REDUCE_ADD, ISD::VP_REDUCE_AND,
ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR,
ISD::VP_REDUCE_SMAX, ISD::VP_REDUCE_SMIN,
ISD::VP_REDUCE_UMAX, ISD::VP_REDUCE_UMIN},
MVT::i64, Custom);
}
for (MVT VT : BoolVecVTs) {
if (!isTypeLegal(VT))
continue;
setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
// Mask VTs are custom-expanded into a series of standard nodes
setOperationAction({ISD::TRUNCATE, ISD::CONCAT_VECTORS,
ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR,
ISD::SCALAR_TO_VECTOR},
VT, Custom);
setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
Custom);
setOperationAction(ISD::SELECT, VT, Custom);
setOperationAction(
{ISD::SELECT_CC, ISD::VSELECT, ISD::VP_MERGE, ISD::VP_SELECT}, VT,
Expand);
setOperationAction({ISD::VP_AND, ISD::VP_OR, ISD::VP_XOR}, VT, Custom);
setOperationAction(
{ISD::VECREDUCE_AND, ISD::VECREDUCE_OR, ISD::VECREDUCE_XOR}, VT,
Custom);
setOperationAction(
{ISD::VP_REDUCE_AND, ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR}, VT,
Custom);
// RVV has native int->float & float->int conversions where the
// element type sizes are within one power-of-two of each other. Any
// wider distances between type sizes have to be lowered as sequences
// which progressively narrow the gap in stages.
setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT,
ISD::FP_TO_UINT, ISD::STRICT_SINT_TO_FP,
ISD::STRICT_UINT_TO_FP, ISD::STRICT_FP_TO_SINT,
ISD::STRICT_FP_TO_UINT},
VT, Custom);
setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
Custom);
// Expand all extending loads to types larger than this, and truncating
// stores from types larger than this.
for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
setTruncStoreAction(VT, OtherVT, Expand);
setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, VT,
OtherVT, Expand);
}
setOperationAction({ISD::VP_FP_TO_SINT, ISD::VP_FP_TO_UINT,
ISD::VP_TRUNCATE, ISD::VP_SETCC},
VT, Custom);
setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
setOperationAction(ISD::VECTOR_REVERSE, VT, Custom);
setOperationAction(ISD::EXPERIMENTAL_VP_SPLICE, VT, Custom);
setOperationAction(ISD::EXPERIMENTAL_VP_REVERSE, VT, Custom);
setOperationPromotedToType(
ISD::VECTOR_SPLICE, VT,
MVT::getVectorVT(MVT::i8, VT.getVectorElementCount()));
}
for (MVT VT : IntVecVTs) {
if (!isTypeLegal(VT))
continue;
setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
// Vectors implement MULHS/MULHU.
setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, VT, Expand);
// nxvXi64 MULHS/MULHU requires the V extension instead of Zve64*.
if (VT.getVectorElementType() == MVT::i64 && !Subtarget.hasStdExtV())
setOperationAction({ISD::MULHU, ISD::MULHS}, VT, Expand);
setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, VT,
Legal);
setOperationAction({ISD::ABDS, ISD::ABDU}, VT, Custom);
// Custom-lower extensions and truncations from/to mask types.
setOperationAction({ISD::ANY_EXTEND, ISD::SIGN_EXTEND, ISD::ZERO_EXTEND},
VT, Custom);
// RVV has native int->float & float->int conversions where the
// element type sizes are within one power-of-two of each other. Any
// wider distances between type sizes have to be lowered as sequences
// which progressively narrow the gap in stages.
setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT,
ISD::FP_TO_UINT, ISD::STRICT_SINT_TO_FP,
ISD::STRICT_UINT_TO_FP, ISD::STRICT_FP_TO_SINT,
ISD::STRICT_FP_TO_UINT},
VT, Custom);
setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
Custom);
setOperationAction({ISD::AVGFLOORU, ISD::AVGCEILU, ISD::SADDSAT,
ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT},
VT, Legal);
// Integer VTs are lowered as a series of "RISCVISD::TRUNCATE_VECTOR_VL"
// nodes which truncate by one power of two at a time.
setOperationAction(ISD::TRUNCATE, VT, Custom);
// Custom-lower insert/extract operations to simplify patterns.
setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
Custom);
// Custom-lower reduction operations to set up the corresponding custom
// nodes' operands.
setOperationAction(IntegerVecReduceOps, VT, Custom);
setOperationAction(IntegerVPOps, VT, Custom);
setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
setOperationAction({ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER},
VT, Custom);
setOperationAction(
{ISD::VP_LOAD, ISD::VP_STORE, ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER, ISD::VP_SCATTER},
VT, Custom);
setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
ISD::EXTRACT_SUBVECTOR, ISD::SCALAR_TO_VECTOR},
VT, Custom);
setOperationAction(ISD::SELECT, VT, Custom);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction({ISD::STEP_VECTOR, ISD::VECTOR_REVERSE}, VT, Custom);
for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
setTruncStoreAction(VT, OtherVT, Expand);
setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, VT,
OtherVT, Expand);
}
setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
// Splice
setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);
if (Subtarget.hasStdExtZvkb()) {
setOperationAction(ISD::BSWAP, VT, Legal);
setOperationAction(ISD::VP_BSWAP, VT, Custom);
} else {
setOperationAction({ISD::BSWAP, ISD::VP_BSWAP}, VT, Expand);
setOperationAction({ISD::ROTL, ISD::ROTR}, VT, Expand);
}
if (Subtarget.hasStdExtZvbb()) {
setOperationAction(ISD::BITREVERSE, VT, Legal);
setOperationAction(ISD::VP_BITREVERSE, VT, Custom);
setOperationAction({ISD::VP_CTLZ, ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_CTTZ,
ISD::VP_CTTZ_ZERO_UNDEF, ISD::VP_CTPOP},
VT, Custom);
} else {
setOperationAction({ISD::BITREVERSE, ISD::VP_BITREVERSE}, VT, Expand);
setOperationAction({ISD::CTLZ, ISD::CTTZ, ISD::CTPOP}, VT, Expand);
setOperationAction({ISD::VP_CTLZ, ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_CTTZ,
ISD::VP_CTTZ_ZERO_UNDEF, ISD::VP_CTPOP},
VT, Expand);
// Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if element of VT in the
// range of f32.
EVT FloatVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
if (isTypeLegal(FloatVT)) {
setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF,
ISD::CTTZ_ZERO_UNDEF, ISD::VP_CTLZ,
ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_CTTZ_ZERO_UNDEF},
VT, Custom);
}
}
}
// Expand various CCs to best match the RVV ISA, which natively supports UNE
// but no other unordered comparisons, and supports all ordered comparisons
// except ONE. Additionally, we expand GT,OGT,GE,OGE for optimization
// purposes; they are expanded to their swapped-operand CCs (LT,OLT,LE,OLE),
// and we pattern-match those back to the "original", swapping operands once
// more. This way we catch both operations and both "vf" and "fv" forms with
// fewer patterns.
static const ISD::CondCode VFPCCToExpand[] = {
ISD::SETO, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUO,
ISD::SETGT, ISD::SETOGT, ISD::SETGE, ISD::SETOGE,
};
// TODO: support more ops.
static const unsigned ZvfhminPromoteOps[] = {
ISD::FMINNUM, ISD::FMAXNUM, ISD::FADD, ISD::FSUB,
ISD::FMUL, ISD::FMA, ISD::FDIV, ISD::FSQRT,
ISD::FABS, ISD::FNEG, ISD::FCOPYSIGN, ISD::FCEIL,
ISD::FFLOOR, ISD::FROUND, ISD::FROUNDEVEN, ISD::FRINT,
ISD::FNEARBYINT, ISD::IS_FPCLASS, ISD::SETCC, ISD::FMAXIMUM,
ISD::FMINIMUM, ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA};
// TODO: support more vp ops.
static const unsigned ZvfhminPromoteVPOps[] = {
ISD::VP_FADD, ISD::VP_FSUB, ISD::VP_FMUL,
ISD::VP_FDIV, ISD::VP_FNEG, ISD::VP_FABS,
ISD::VP_FMA, ISD::VP_REDUCE_FADD, ISD::VP_REDUCE_SEQ_FADD,
ISD::VP_REDUCE_FMIN, ISD::VP_REDUCE_FMAX, ISD::VP_SQRT,
ISD::VP_FMINNUM, ISD::VP_FMAXNUM, ISD::VP_FCEIL,
ISD::VP_FFLOOR, ISD::VP_FROUND, ISD::VP_FROUNDEVEN,
ISD::VP_FCOPYSIGN, ISD::VP_FROUNDTOZERO, ISD::VP_FRINT,
ISD::VP_FNEARBYINT, ISD::VP_SETCC, ISD::VP_FMINIMUM,
ISD::VP_FMAXIMUM};
// Sets common operation actions on RVV floating-point vector types.
const auto SetCommonVFPActions = [&](MVT VT) {
setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
// RVV has native FP_ROUND & FP_EXTEND conversions where the element type
// sizes are within one power-of-two of each other. Therefore conversions
// between vXf16 and vXf64 must be lowered as sequences which convert via
// vXf32.
setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
setOperationAction({ISD::LRINT, ISD::LLRINT}, VT, Custom);
// Custom-lower insert/extract operations to simplify patterns.
setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
Custom);
// Expand various condition codes (explained above).
setCondCodeAction(VFPCCToExpand, VT, Expand);
setOperationAction({ISD::FMINNUM, ISD::FMAXNUM}, VT, Legal);
setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, VT, Custom);
setOperationAction({ISD::FTRUNC, ISD::FCEIL, ISD::FFLOOR, ISD::FROUND,
ISD::FROUNDEVEN, ISD::FRINT, ISD::FNEARBYINT,
ISD::IS_FPCLASS},
VT, Custom);
setOperationAction(FloatingPointVecReduceOps, VT, Custom);
// Expand FP operations that need libcalls.
setOperationAction(ISD::FREM, VT, Expand);
setOperationAction(ISD::FPOW, VT, Expand);
setOperationAction(ISD::FCOS, VT, Expand);
setOperationAction(ISD::FSIN, VT, Expand);
setOperationAction(ISD::FSINCOS, VT, Expand);
setOperationAction(ISD::FEXP, VT, Expand);
setOperationAction(ISD::FEXP2, VT, Expand);
setOperationAction(ISD::FEXP10, VT, Expand);
setOperationAction(ISD::FLOG, VT, Expand);
setOperationAction(ISD::FLOG2, VT, Expand);
setOperationAction(ISD::FLOG10, VT, Expand);
setOperationAction(ISD::FCOPYSIGN, VT, Legal);
setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
setOperationAction({ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER},
VT, Custom);
setOperationAction(
{ISD::VP_LOAD, ISD::VP_STORE, ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER, ISD::VP_SCATTER},
VT, Custom);
setOperationAction(ISD::SELECT, VT, Custom);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
ISD::EXTRACT_SUBVECTOR, ISD::SCALAR_TO_VECTOR},
VT, Custom);
setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
setOperationAction({ISD::VECTOR_REVERSE, ISD::VECTOR_SPLICE}, VT, Custom);
setOperationAction(FloatingPointVPOps, VT, Custom);
setOperationAction({ISD::STRICT_FP_EXTEND, ISD::STRICT_FP_ROUND}, VT,
Custom);
setOperationAction({ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA},
VT, Legal);
setOperationAction({ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS,
ISD::STRICT_FTRUNC, ISD::STRICT_FCEIL,
ISD::STRICT_FFLOOR, ISD::STRICT_FROUND,
ISD::STRICT_FROUNDEVEN, ISD::STRICT_FNEARBYINT},
VT, Custom);
};
// Sets common extload/truncstore actions on RVV floating-point vector
// types.
const auto SetCommonVFPExtLoadTruncStoreActions =
[&](MVT VT, ArrayRef<MVT::SimpleValueType> SmallerVTs) {
for (auto SmallVT : SmallerVTs) {
setTruncStoreAction(VT, SmallVT, Expand);
setLoadExtAction(ISD::EXTLOAD, VT, SmallVT, Expand);
}
};
if (Subtarget.hasVInstructionsF16()) {
for (MVT VT : F16VecVTs) {
if (!isTypeLegal(VT))
continue;
SetCommonVFPActions(VT);
}
} else if (Subtarget.hasVInstructionsF16Minimal()) {
for (MVT VT : F16VecVTs) {
if (!isTypeLegal(VT))
continue;
setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
setOperationAction({ISD::STRICT_FP_ROUND, ISD::STRICT_FP_EXTEND}, VT,
Custom);
setOperationAction({ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND}, VT, Custom);
setOperationAction({ISD::VP_MERGE, ISD::VP_SELECT, ISD::SELECT}, VT,
Custom);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP,
ISD::VP_SINT_TO_FP, ISD::VP_UINT_TO_FP},
VT, Custom);
setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
ISD::EXTRACT_SUBVECTOR, ISD::SCALAR_TO_VECTOR},
VT, Custom);
if (Subtarget.hasStdExtZfhminOrZhinxmin())
setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
// load/store
setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
// Custom split nxv32f16 since nxv32f32 if not legal.
if (VT == MVT::nxv32f16) {
setOperationAction(ZvfhminPromoteOps, VT, Custom);
setOperationAction(ZvfhminPromoteVPOps, VT, Custom);
continue;
}
// Add more promote ops.
MVT F32VecVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
setOperationPromotedToType(ZvfhminPromoteOps, VT, F32VecVT);
setOperationPromotedToType(ZvfhminPromoteVPOps, VT, F32VecVT);
}
}
if (Subtarget.hasVInstructionsF32()) {
for (MVT VT : F32VecVTs) {
if (!isTypeLegal(VT))
continue;
SetCommonVFPActions(VT);
SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
}
}
if (Subtarget.hasVInstructionsF64()) {
for (MVT VT : F64VecVTs) {
if (!isTypeLegal(VT))
continue;
SetCommonVFPActions(VT);
SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
SetCommonVFPExtLoadTruncStoreActions(VT, F32VecVTs);
}
}
if (Subtarget.useRVVForFixedLengthVectors()) {
for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
if (!useRVVForFixedLengthVectorVT(VT))
continue;
// By default everything must be expanded.
for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
setOperationAction(Op, VT, Expand);
for (MVT OtherVT : MVT::integer_fixedlen_vector_valuetypes()) {
setTruncStoreAction(VT, OtherVT, Expand);
setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, VT,
OtherVT, Expand);
}
// Custom lower fixed vector undefs to scalable vector undefs to avoid
// expansion to a build_vector of 0s.
setOperationAction(ISD::UNDEF, VT, Custom);
// We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
setOperationAction({ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR}, VT,
Custom);
setOperationAction({ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS}, VT,
Custom);
setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT},
VT, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
setOperationAction(ISD::SETCC, VT, Custom);
setOperationAction(ISD::SELECT, VT, Custom);
setOperationAction(ISD::TRUNCATE, VT, Custom);
setOperationAction(ISD::BITCAST, VT, Custom);
setOperationAction(
{ISD::VECREDUCE_AND, ISD::VECREDUCE_OR, ISD::VECREDUCE_XOR}, VT,
Custom);
setOperationAction(
{ISD::VP_REDUCE_AND, ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR}, VT,
Custom);
setOperationAction(
{
ISD::SINT_TO_FP,
ISD::UINT_TO_FP,
ISD::FP_TO_SINT,
ISD::FP_TO_UINT,
ISD::STRICT_SINT_TO_FP,
ISD::STRICT_UINT_TO_FP,
ISD::STRICT_FP_TO_SINT,
ISD::STRICT_FP_TO_UINT,
},
VT, Custom);
setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
// Operations below are different for between masks and other vectors.
if (VT.getVectorElementType() == MVT::i1) {
setOperationAction({ISD::VP_AND, ISD::VP_OR, ISD::VP_XOR, ISD::AND,
ISD::OR, ISD::XOR},
VT, Custom);
setOperationAction({ISD::VP_FP_TO_SINT, ISD::VP_FP_TO_UINT,
ISD::VP_SETCC, ISD::VP_TRUNCATE},
VT, Custom);
setOperationAction(ISD::EXPERIMENTAL_VP_SPLICE, VT, Custom);
setOperationAction(ISD::EXPERIMENTAL_VP_REVERSE, VT, Custom);
continue;
}
// Make SPLAT_VECTOR Legal so DAGCombine will convert splat vectors to
// it before type legalization for i64 vectors on RV32. It will then be
// type legalized to SPLAT_VECTOR_PARTS which we need to Custom handle.
// FIXME: Use SPLAT_VECTOR for all types? DAGCombine probably needs
// improvements first.
if (!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64) {
setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
}
setOperationAction(
{ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER}, VT, Custom);
setOperationAction({ISD::VP_LOAD, ISD::VP_STORE,
ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER,
ISD::VP_SCATTER},
VT, Custom);
setOperationAction({ISD::ADD, ISD::MUL, ISD::SUB, ISD::AND, ISD::OR,
ISD::XOR, ISD::SDIV, ISD::SREM, ISD::UDIV,
ISD::UREM, ISD::SHL, ISD::SRA, ISD::SRL},
VT, Custom);
setOperationAction(
{ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX, ISD::ABS}, VT, Custom);
setOperationAction({ISD::ABDS, ISD::ABDU}, VT, Custom);
// vXi64 MULHS/MULHU requires the V extension instead of Zve64*.
if (VT.getVectorElementType() != MVT::i64 || Subtarget.hasStdExtV())
setOperationAction({ISD::MULHS, ISD::MULHU}, VT, Custom);
setOperationAction({ISD::AVGFLOORU, ISD::AVGCEILU, ISD::SADDSAT,
ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT},
VT, Custom);
setOperationAction(ISD::VSELECT, VT, Custom);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction(
{ISD::ANY_EXTEND, ISD::SIGN_EXTEND, ISD::ZERO_EXTEND}, VT, Custom);
// Custom-lower reduction operations to set up the corresponding custom
// nodes' operands.
setOperationAction({ISD::VECREDUCE_ADD, ISD::VECREDUCE_SMAX,
ISD::VECREDUCE_SMIN, ISD::VECREDUCE_UMAX,
ISD::VECREDUCE_UMIN},
VT, Custom);
setOperationAction(IntegerVPOps, VT, Custom);
if (Subtarget.hasStdExtZvkb())
setOperationAction({ISD::BSWAP, ISD::ROTL, ISD::ROTR}, VT, Custom);
if (Subtarget.hasStdExtZvbb()) {
setOperationAction({ISD::BITREVERSE, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF,
ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTPOP},
VT, Custom);
} else {
// Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if element of VT in the
// range of f32.
EVT FloatVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
if (isTypeLegal(FloatVT))
setOperationAction(
{ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF, ISD::CTTZ_ZERO_UNDEF}, VT,
Custom);
}
}
for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) {
// There are no extending loads or truncating stores.
for (MVT InnerVT : MVT::fp_fixedlen_vector_valuetypes()) {
setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
setTruncStoreAction(VT, InnerVT, Expand);
}
if (!useRVVForFixedLengthVectorVT(VT))
continue;
// By default everything must be expanded.
for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
setOperationAction(Op, VT, Expand);
// Custom lower fixed vector undefs to scalable vector undefs to avoid
// expansion to a build_vector of 0s.
setOperationAction(ISD::UNDEF, VT, Custom);
if (VT.getVectorElementType() == MVT::f16 &&
!Subtarget.hasVInstructionsF16()) {
setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
setOperationAction({ISD::STRICT_FP_ROUND, ISD::STRICT_FP_EXTEND}, VT,
Custom);
setOperationAction({ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND}, VT, Custom);
setOperationAction(
{ISD::VP_MERGE, ISD::VP_SELECT, ISD::VSELECT, ISD::SELECT}, VT,
Custom);
setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP,
ISD::VP_SINT_TO_FP, ISD::VP_UINT_TO_FP},
VT, Custom);
setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
ISD::EXTRACT_SUBVECTOR, ISD::SCALAR_TO_VECTOR},
VT, Custom);
setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
MVT F32VecVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
// Don't promote f16 vector operations to f32 if f32 vector type is
// not legal.
// TODO: could split the f16 vector into two vectors and do promotion.
if (!isTypeLegal(F32VecVT))
continue;
setOperationPromotedToType(ZvfhminPromoteOps, VT, F32VecVT);
setOperationPromotedToType(ZvfhminPromoteVPOps, VT, F32VecVT);
continue;
}
// We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
setOperationAction({ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR}, VT,
Custom);
setOperationAction({ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS,
ISD::VECTOR_SHUFFLE, ISD::INSERT_VECTOR_ELT,
ISD::EXTRACT_VECTOR_ELT},
VT, Custom);
setOperationAction({ISD::LOAD, ISD::STORE, ISD::MLOAD, ISD::MSTORE,
ISD::MGATHER, ISD::MSCATTER},
VT, Custom);
setOperationAction({ISD::VP_LOAD, ISD::VP_STORE,
ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER,
ISD::VP_SCATTER},
VT, Custom);
setOperationAction({ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV,
ISD::FNEG, ISD::FABS, ISD::FCOPYSIGN, ISD::FSQRT,
ISD::FMA, ISD::FMINNUM, ISD::FMAXNUM,
ISD::IS_FPCLASS, ISD::FMAXIMUM, ISD::FMINIMUM},
VT, Custom);
setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
setOperationAction({ISD::FTRUNC, ISD::FCEIL, ISD::FFLOOR, ISD::FROUND,
ISD::FROUNDEVEN, ISD::FRINT, ISD::FNEARBYINT},
VT, Custom);
setCondCodeAction(VFPCCToExpand, VT, Expand);
setOperationAction(ISD::SETCC, VT, Custom);
setOperationAction({ISD::VSELECT, ISD::SELECT}, VT, Custom);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction(ISD::BITCAST, VT, Custom);
setOperationAction(FloatingPointVecReduceOps, VT, Custom);
setOperationAction(FloatingPointVPOps, VT, Custom);
setOperationAction({ISD::STRICT_FP_EXTEND, ISD::STRICT_FP_ROUND}, VT,
Custom);
setOperationAction(
{ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA,
ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS, ISD::STRICT_FTRUNC,
ISD::STRICT_FCEIL, ISD::STRICT_FFLOOR, ISD::STRICT_FROUND,
ISD::STRICT_FROUNDEVEN, ISD::STRICT_FNEARBYINT},
VT, Custom);
}
// Custom-legalize bitcasts from fixed-length vectors to scalar types.
setOperationAction(ISD::BITCAST, {MVT::i8, MVT::i16, MVT::i32, MVT::i64},
Custom);
if (Subtarget.hasStdExtZfhminOrZhinxmin())
setOperationAction(ISD::BITCAST, MVT::f16, Custom);
if (Subtarget.hasStdExtFOrZfinx())
setOperationAction(ISD::BITCAST, MVT::f32, Custom);
if (Subtarget.hasStdExtDOrZdinx())
setOperationAction(ISD::BITCAST, MVT::f64, Custom);
}
}
if (Subtarget.hasStdExtA()) {
setOperationAction(ISD::ATOMIC_LOAD_SUB, XLenVT, Expand);
if (RV64LegalI32 && Subtarget.is64Bit())
setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
}
if (Subtarget.hasForcedAtomics()) {
// Force __sync libcalls to be emitted for atomic rmw/cas operations.
setOperationAction(
{ISD::ATOMIC_CMP_SWAP, ISD::ATOMIC_SWAP, ISD::ATOMIC_LOAD_ADD,
ISD::ATOMIC_LOAD_SUB, ISD::ATOMIC_LOAD_AND, ISD::ATOMIC_LOAD_OR,
ISD::ATOMIC_LOAD_XOR, ISD::ATOMIC_LOAD_NAND, ISD::ATOMIC_LOAD_MIN,
ISD::ATOMIC_LOAD_MAX, ISD::ATOMIC_LOAD_UMIN, ISD::ATOMIC_LOAD_UMAX},
XLenVT, LibCall);
}
if (Subtarget.hasVendorXTHeadMemIdx()) {
for (unsigned im : {ISD::PRE_INC, ISD::POST_INC}) {
setIndexedLoadAction(im, MVT::i8, Legal);
setIndexedStoreAction(im, MVT::i8, Legal);
setIndexedLoadAction(im, MVT::i16, Legal);
setIndexedStoreAction(im, MVT::i16, Legal);
setIndexedLoadAction(im, MVT::i32, Legal);
setIndexedStoreAction(im, MVT::i32, Legal);
if (Subtarget.is64Bit()) {
setIndexedLoadAction(im, MVT::i64, Legal);
setIndexedStoreAction(im, MVT::i64, Legal);
}
}
}
// Function alignments.
const Align FunctionAlignment(Subtarget.hasStdExtCOrZca() ? 2 : 4);
setMinFunctionAlignment(FunctionAlignment);
// Set preferred alignments.
setPrefFunctionAlignment(Subtarget.getPrefFunctionAlignment());
setPrefLoopAlignment(Subtarget.getPrefLoopAlignment());
setTargetDAGCombine({ISD::INTRINSIC_VOID, ISD::INTRINSIC_W_CHAIN,
ISD::INTRINSIC_WO_CHAIN, ISD::ADD, ISD::SUB, ISD::MUL,
ISD::AND, ISD::OR, ISD::XOR, ISD::SETCC, ISD::SELECT});
if (Subtarget.is64Bit())
setTargetDAGCombine(ISD::SRA);
if (Subtarget.hasStdExtFOrZfinx())
setTargetDAGCombine({ISD::FADD, ISD::FMAXNUM, ISD::FMINNUM});
if (Subtarget.hasStdExtZbb())
setTargetDAGCombine({ISD::UMAX, ISD::UMIN, ISD::SMAX, ISD::SMIN});
if (Subtarget.hasStdExtZbs() && Subtarget.is64Bit())
setTargetDAGCombine(ISD::TRUNCATE);
if (Subtarget.hasStdExtZbkb())
setTargetDAGCombine(ISD::BITREVERSE);
if (Subtarget.hasStdExtZfhminOrZhinxmin())
setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
if (Subtarget.hasStdExtFOrZfinx())
setTargetDAGCombine({ISD::ZERO_EXTEND, ISD::FP_TO_SINT, ISD::FP_TO_UINT,
ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT});
if (Subtarget.hasVInstructions())
setTargetDAGCombine({ISD::FCOPYSIGN, ISD::MGATHER, ISD::MSCATTER,
ISD::VP_GATHER, ISD::VP_SCATTER, ISD::SRA, ISD::SRL,
ISD::SHL, ISD::STORE, ISD::SPLAT_VECTOR,
ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS,
ISD::EXPERIMENTAL_VP_REVERSE, ISD::MUL,
ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM,
ISD::INSERT_VECTOR_ELT, ISD::ABS});
if (Subtarget.hasVendorXTHeadMemPair())
setTargetDAGCombine({ISD::LOAD, ISD::STORE});
if (Subtarget.useRVVForFixedLengthVectors())
setTargetDAGCombine(ISD::BITCAST);
setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
// Disable strict node mutation.
IsStrictFPEnabled = true;
}
EVT RISCVTargetLowering::getSetCCResultType(const DataLayout &DL,
LLVMContext &Context,
EVT VT) const {
if (!VT.isVector())
return getPointerTy(DL);
if (Subtarget.hasVInstructions() &&
(VT.isScalableVector() || Subtarget.useRVVForFixedLengthVectors()))
return EVT::getVectorVT(Context, MVT::i1, VT.getVectorElementCount());
return VT.changeVectorElementTypeToInteger();
}
MVT RISCVTargetLowering::getVPExplicitVectorLengthTy() const {
return Subtarget.getXLenVT();
}
// Return false if we can lower get_vector_length to a vsetvli intrinsic.
bool RISCVTargetLowering::shouldExpandGetVectorLength(EVT TripCountVT,
unsigned VF,
bool IsScalable) const {
if (!Subtarget.hasVInstructions())
return true;
if (!IsScalable)
return true;
if (TripCountVT != MVT::i32 && TripCountVT != Subtarget.getXLenVT())
return true;
// Don't allow VF=1 if those types are't legal.
if (VF < RISCV::RVVBitsPerBlock / Subtarget.getELen())
return true;
// VLEN=32 support is incomplete.
if (Subtarget.getRealMinVLen() < RISCV::RVVBitsPerBlock)
return true;
// The maximum VF is for the smallest element width with LMUL=8.
// VF must be a power of 2.
unsigned MaxVF = (RISCV::RVVBitsPerBlock / 8) * 8;
return VF > MaxVF || !isPowerOf2_32(VF);
}
bool RISCVTargetLowering::shouldExpandCttzElements(EVT VT) const {
return !Subtarget.hasVInstructions() ||
VT.getVectorElementType() != MVT::i1 || !isTypeLegal(VT);
}
bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
const CallInst &I,
MachineFunction &MF,
unsigned Intrinsic) const {
auto &DL = I.getModule()->getDataLayout();
auto SetRVVLoadStoreInfo = [&](unsigned PtrOp, bool IsStore,
bool IsUnitStrided, bool UsePtrVal = false) {
Info.opc = IsStore ? ISD::INTRINSIC_VOID : ISD::INTRINSIC_W_CHAIN;
// We can't use ptrVal if the intrinsic can access memory before the
// pointer. This means we can't use it for strided or indexed intrinsics.
if (UsePtrVal)
Info.ptrVal = I.getArgOperand(PtrOp);
else
Info.fallbackAddressSpace =
I.getArgOperand(PtrOp)->getType()->getPointerAddressSpace();
Type *MemTy;
if (IsStore) {
// Store value is the first operand.
MemTy = I.getArgOperand(0)->getType();
} else {
// Use return type. If it's segment load, return type is a struct.
MemTy = I.getType();
if (MemTy->isStructTy())
MemTy = MemTy->getStructElementType(0);
}
if (!IsUnitStrided)
MemTy = MemTy->getScalarType();
Info.memVT = getValueType(DL, MemTy);
Info.align = Align(DL.getTypeSizeInBits(MemTy->getScalarType()) / 8);
Info.size = MemoryLocation::UnknownSize;
Info.flags |=
IsStore ? MachineMemOperand::MOStore : MachineMemOperand::MOLoad;
return true;
};
if (I.hasMetadata(LLVMContext::MD_nontemporal))
Info.flags |= MachineMemOperand::MONonTemporal;
Info.flags |= RISCVTargetLowering::getTargetMMOFlags(I);
switch (Intrinsic) {
default:
return false;
case Intrinsic::riscv_masked_atomicrmw_xchg_i32:
case Intrinsic::riscv_masked_atomicrmw_add_i32:
case Intrinsic::riscv_masked_atomicrmw_sub_i32:
case Intrinsic::riscv_masked_atomicrmw_nand_i32:
case Intrinsic::riscv_masked_atomicrmw_max_i32:
case Intrinsic::riscv_masked_atomicrmw_min_i32:
case Intrinsic::riscv_masked_atomicrmw_umax_i32:
case Intrinsic::riscv_masked_atomicrmw_umin_i32:
case Intrinsic::riscv_masked_cmpxchg_i32:
Info.opc = ISD::INTRINSIC_W_CHAIN;
Info.memVT = MVT::i32;
Info.ptrVal = I.getArgOperand(0);
Info.offset = 0;
Info.align = Align(4);
Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
MachineMemOperand::MOVolatile;
return true;
case Intrinsic::riscv_masked_strided_load:
return SetRVVLoadStoreInfo(/*PtrOp*/ 1, /*IsStore*/ false,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_masked_strided_store:
return SetRVVLoadStoreInfo(/*PtrOp*/ 1, /*IsStore*/ true,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_seg2_load:
case Intrinsic::riscv_seg3_load:
case Intrinsic::riscv_seg4_load:
case Intrinsic::riscv_seg5_load:
case Intrinsic::riscv_seg6_load:
case Intrinsic::riscv_seg7_load:
case Intrinsic::riscv_seg8_load:
return SetRVVLoadStoreInfo(/*PtrOp*/ 0, /*IsStore*/ false,
/*IsUnitStrided*/ false, /*UsePtrVal*/ true);
case Intrinsic::riscv_seg2_store:
case Intrinsic::riscv_seg3_store:
case Intrinsic::riscv_seg4_store:
case Intrinsic::riscv_seg5_store:
case Intrinsic::riscv_seg6_store:
case Intrinsic::riscv_seg7_store:
case Intrinsic::riscv_seg8_store:
// Operands are (vec, ..., vec, ptr, vl)
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 2,
/*IsStore*/ true,
/*IsUnitStrided*/ false, /*UsePtrVal*/ true);
case Intrinsic::riscv_vle:
case Intrinsic::riscv_vle_mask:
case Intrinsic::riscv_vleff:
case Intrinsic::riscv_vleff_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
/*IsStore*/ false,
/*IsUnitStrided*/ true,
/*UsePtrVal*/ true);
case Intrinsic::riscv_vse:
case Intrinsic::riscv_vse_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
/*IsStore*/ true,
/*IsUnitStrided*/ true,
/*UsePtrVal*/ true);
case Intrinsic::riscv_vlse:
case Intrinsic::riscv_vlse_mask:
case Intrinsic::riscv_vloxei:
case Intrinsic::riscv_vloxei_mask:
case Intrinsic::riscv_vluxei:
case Intrinsic::riscv_vluxei_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
/*IsStore*/ false,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vsse:
case Intrinsic::riscv_vsse_mask:
case Intrinsic::riscv_vsoxei:
case Intrinsic::riscv_vsoxei_mask:
case Intrinsic::riscv_vsuxei:
case Intrinsic::riscv_vsuxei_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
/*IsStore*/ true,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vlseg2:
case Intrinsic::riscv_vlseg3:
case Intrinsic::riscv_vlseg4:
case Intrinsic::riscv_vlseg5:
case Intrinsic::riscv_vlseg6:
case Intrinsic::riscv_vlseg7:
case Intrinsic::riscv_vlseg8:
case Intrinsic::riscv_vlseg2ff:
case Intrinsic::riscv_vlseg3ff:
case Intrinsic::riscv_vlseg4ff:
case Intrinsic::riscv_vlseg5ff:
case Intrinsic::riscv_vlseg6ff:
case Intrinsic::riscv_vlseg7ff:
case Intrinsic::riscv_vlseg8ff:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 2,
/*IsStore*/ false,
/*IsUnitStrided*/ false, /*UsePtrVal*/ true);
case Intrinsic::riscv_vlseg2_mask:
case Intrinsic::riscv_vlseg3_mask:
case Intrinsic::riscv_vlseg4_mask:
case Intrinsic::riscv_vlseg5_mask:
case Intrinsic::riscv_vlseg6_mask:
case Intrinsic::riscv_vlseg7_mask:
case Intrinsic::riscv_vlseg8_mask:
case Intrinsic::riscv_vlseg2ff_mask:
case Intrinsic::riscv_vlseg3ff_mask:
case Intrinsic::riscv_vlseg4ff_mask:
case Intrinsic::riscv_vlseg5ff_mask:
case Intrinsic::riscv_vlseg6ff_mask:
case Intrinsic::riscv_vlseg7ff_mask:
case Intrinsic::riscv_vlseg8ff_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 4,
/*IsStore*/ false,
/*IsUnitStrided*/ false, /*UsePtrVal*/ true);
case Intrinsic::riscv_vlsseg2:
case Intrinsic::riscv_vlsseg3:
case Intrinsic::riscv_vlsseg4:
case Intrinsic::riscv_vlsseg5:
case Intrinsic::riscv_vlsseg6:
case Intrinsic::riscv_vlsseg7:
case Intrinsic::riscv_vlsseg8:
case Intrinsic::riscv_vloxseg2:
case Intrinsic::riscv_vloxseg3:
case Intrinsic::riscv_vloxseg4:
case Intrinsic::riscv_vloxseg5:
case Intrinsic::riscv_vloxseg6:
case Intrinsic::riscv_vloxseg7:
case Intrinsic::riscv_vloxseg8:
case Intrinsic::riscv_vluxseg2:
case Intrinsic::riscv_vluxseg3:
case Intrinsic::riscv_vluxseg4:
case Intrinsic::riscv_vluxseg5:
case Intrinsic::riscv_vluxseg6:
case Intrinsic::riscv_vluxseg7:
case Intrinsic::riscv_vluxseg8:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
/*IsStore*/ false,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vlsseg2_mask:
case Intrinsic::riscv_vlsseg3_mask:
case Intrinsic::riscv_vlsseg4_mask:
case Intrinsic::riscv_vlsseg5_mask:
case Intrinsic::riscv_vlsseg6_mask:
case Intrinsic::riscv_vlsseg7_mask:
case Intrinsic::riscv_vlsseg8_mask:
case Intrinsic::riscv_vloxseg2_mask:
case Intrinsic::riscv_vloxseg3_mask:
case Intrinsic::riscv_vloxseg4_mask:
case Intrinsic::riscv_vloxseg5_mask:
case Intrinsic::riscv_vloxseg6_mask:
case Intrinsic::riscv_vloxseg7_mask:
case Intrinsic::riscv_vloxseg8_mask:
case Intrinsic::riscv_vluxseg2_mask:
case Intrinsic::riscv_vluxseg3_mask:
case Intrinsic::riscv_vluxseg4_mask:
case Intrinsic::riscv_vluxseg5_mask:
case Intrinsic::riscv_vluxseg6_mask:
case Intrinsic::riscv_vluxseg7_mask:
case Intrinsic::riscv_vluxseg8_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 5,
/*IsStore*/ false,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vsseg2:
case Intrinsic::riscv_vsseg3:
case Intrinsic::riscv_vsseg4:
case Intrinsic::riscv_vsseg5:
case Intrinsic::riscv_vsseg6:
case Intrinsic::riscv_vsseg7:
case Intrinsic::riscv_vsseg8:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 2,
/*IsStore*/ true,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vsseg2_mask:
case Intrinsic::riscv_vsseg3_mask:
case Intrinsic::riscv_vsseg4_mask:
case Intrinsic::riscv_vsseg5_mask:
case Intrinsic::riscv_vsseg6_mask:
case Intrinsic::riscv_vsseg7_mask:
case Intrinsic::riscv_vsseg8_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
/*IsStore*/ true,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vssseg2:
case Intrinsic::riscv_vssseg3:
case Intrinsic::riscv_vssseg4:
case Intrinsic::riscv_vssseg5:
case Intrinsic::riscv_vssseg6:
case Intrinsic::riscv_vssseg7:
case Intrinsic::riscv_vssseg8:
case Intrinsic::riscv_vsoxseg2:
case Intrinsic::riscv_vsoxseg3:
case Intrinsic::riscv_vsoxseg4:
case Intrinsic::riscv_vsoxseg5:
case Intrinsic::riscv_vsoxseg6:
case Intrinsic::riscv_vsoxseg7:
case Intrinsic::riscv_vsoxseg8:
case Intrinsic::riscv_vsuxseg2:
case Intrinsic::riscv_vsuxseg3:
case Intrinsic::riscv_vsuxseg4:
case Intrinsic::riscv_vsuxseg5:
case Intrinsic::riscv_vsuxseg6:
case Intrinsic::riscv_vsuxseg7:
case Intrinsic::riscv_vsuxseg8:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
/*IsStore*/ true,
/*IsUnitStrided*/ false);
case Intrinsic::riscv_vssseg2_mask:
case Intrinsic::riscv_vssseg3_mask:
case Intrinsic::riscv_vssseg4_mask:
case Intrinsic::riscv_vssseg5_mask:
case Intrinsic::riscv_vssseg6_mask:
case Intrinsic::riscv_vssseg7_mask:
case Intrinsic::riscv_vssseg8_mask:
case Intrinsic::riscv_vsoxseg2_mask:
case Intrinsic::riscv_vsoxseg3_mask:
case Intrinsic::riscv_vsoxseg4_mask:
case Intrinsic::riscv_vsoxseg5_mask:
case Intrinsic::riscv_vsoxseg6_mask:
case Intrinsic::riscv_vsoxseg7_mask:
case Intrinsic::riscv_vsoxseg8_mask:
case Intrinsic::riscv_vsuxseg2_mask:
case Intrinsic::riscv_vsuxseg3_mask:
case Intrinsic::riscv_vsuxseg4_mask:
case Intrinsic::riscv_vsuxseg5_mask:
case Intrinsic::riscv_vsuxseg6_mask:
case Intrinsic::riscv_vsuxseg7_mask:
case Intrinsic::riscv_vsuxseg8_mask:
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 4,
/*IsStore*/ true,
/*IsUnitStrided*/ false);
}
}
bool RISCVTargetLowering::isLegalAddressingMode(const DataLayout &DL,
const AddrMode &AM, Type *Ty,
unsigned AS,
Instruction *I) const {
// No global is ever allowed as a base.
if (AM.BaseGV)
return false;
// RVV instructions only support register addressing.
if (Subtarget.hasVInstructions() && isa<VectorType>(Ty))
return AM.HasBaseReg && AM.Scale == 0 && !AM.BaseOffs;
// Require a 12-bit signed offset.
if (!isInt<12>(AM.BaseOffs))
return false;
switch (AM.Scale) {
case 0: // "r+i" or just "i", depending on HasBaseReg.
break;
case 1:
if (!AM.HasBaseReg) // allow "r+i".
break;
return false; // disallow "r+r" or "r+r+i".
default:
return false;
}
return true;
}
bool RISCVTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
return isInt<12>(Imm);
}
bool RISCVTargetLowering::isLegalAddImmediate(int64_t Imm) const {
return isInt<12>(Imm);
}
// On RV32, 64-bit integers are split into their high and low parts and held
// in two different registers, so the trunc is free since the low register can
// just be used.
// FIXME: Should we consider i64->i32 free on RV64 to match the EVT version of
// isTruncateFree?
bool RISCVTargetLowering::isTruncateFree(Type *SrcTy, Type *DstTy) const {
if (Subtarget.is64Bit() || !SrcTy->isIntegerTy() || !DstTy->isIntegerTy())
return false;
unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
unsigned DestBits = DstTy->getPrimitiveSizeInBits();
return (SrcBits == 64 && DestBits == 32);
}
bool RISCVTargetLowering::isTruncateFree(EVT SrcVT, EVT DstVT) const {
// We consider i64->i32 free on RV64 since we have good selection of W
// instructions that make promoting operations back to i64 free in many cases.
if (SrcVT.isVector() || DstVT.isVector() || !SrcVT.isInteger() ||
!DstVT.isInteger())
return false;
unsigned SrcBits = SrcVT.getSizeInBits();
unsigned DestBits = DstVT.getSizeInBits();
return (SrcBits == 64 && DestBits == 32);
}
bool RISCVTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
// Zexts are free if they can be combined with a load.
// Don't advertise i32->i64 zextload as being free for RV64. It interacts
// poorly with type legalization of compares preferring sext.
if (auto *LD = dyn_cast<LoadSDNode>(Val)) {
EVT MemVT = LD->getMemoryVT();
if ((MemVT == MVT::i8 || MemVT == MVT::i16) &&
(LD->getExtensionType() == ISD::NON_EXTLOAD ||
LD->getExtensionType() == ISD::ZEXTLOAD))
return true;
}
return TargetLowering::isZExtFree(Val, VT2);
}
bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const {
return Subtarget.is64Bit() && SrcVT == MVT::i32 && DstVT == MVT::i64;
}
bool RISCVTargetLowering::signExtendConstant(const ConstantInt *CI) const {
return Subtarget.is64Bit() && CI->getType()->isIntegerTy(32);
}
bool RISCVTargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
return Subtarget.hasStdExtZbb() || Subtarget.hasVendorXCVbitmanip();
}
bool RISCVTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
return Subtarget.hasStdExtZbb() || Subtarget.hasVendorXTHeadBb() ||
Subtarget.hasVendorXCVbitmanip();
}
bool RISCVTargetLowering::isMaskAndCmp0FoldingBeneficial(
const Instruction &AndI) const {
// We expect to be able to match a bit extraction instruction if the Zbs
// extension is supported and the mask is a power of two. However, we
// conservatively return false if the mask would fit in an ANDI instruction,
// on the basis that it's possible the sinking+duplication of the AND in
// CodeGenPrepare triggered by this hook wouldn't decrease the instruction
// count and would increase code size (e.g. ANDI+BNEZ => BEXTI+BNEZ).
if (!Subtarget.hasStdExtZbs() && !Subtarget.hasVendorXTHeadBs())
return false;
ConstantInt *Mask = dyn_cast<ConstantInt>(AndI.getOperand(1));
if (!Mask)
return false;
return !Mask->getValue().isSignedIntN(12) && Mask->getValue().isPowerOf2();
}
bool RISCVTargetLowering::hasAndNotCompare(SDValue Y) const {
EVT VT = Y.getValueType();
// FIXME: Support vectors once we have tests.
if (VT.isVector())
return false;
return (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()) &&
!isa<ConstantSDNode>(Y);
}
bool RISCVTargetLowering::hasBitTest(SDValue X, SDValue Y) const {
// Zbs provides BEXT[_I], which can be used with SEQZ/SNEZ as a bit test.
if (Subtarget.hasStdExtZbs())
return X.getValueType().isScalarInteger();
auto *C = dyn_cast<ConstantSDNode>(Y);
// XTheadBs provides th.tst (similar to bexti), if Y is a constant
if (Subtarget.hasVendorXTHeadBs())
return C != nullptr;
// We can use ANDI+SEQZ/SNEZ as a bit test. Y contains the bit position.
return C && C->getAPIntValue().ule(10);
}
bool RISCVTargetLowering::shouldFoldSelectWithIdentityConstant(unsigned Opcode,
EVT VT) const {
// Only enable for rvv.
if (!VT.isVector() || !Subtarget.hasVInstructions())
return false;
if (VT.isFixedLengthVector() && !isTypeLegal(VT))
return false;
return true;
}
bool RISCVTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
Type *Ty) const {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getIntegerBitWidth();
if (BitSize > Subtarget.getXLen())
return false;
// Fast path, assume 32-bit immediates are cheap.
int64_t Val = Imm.getSExtValue();
if (isInt<32>(Val))
return true;
// A constant pool entry may be more aligned thant he load we're trying to
// replace. If we don't support unaligned scalar mem, prefer the constant
// pool.
// TODO: Can the caller pass down the alignment?
if (!Subtarget.enableUnalignedScalarMem())
return true;
// Prefer to keep the load if it would require many instructions.
// This uses the same threshold we use for constant pools but doesn't
// check useConstantPoolForLargeInts.
// TODO: Should we keep the load only when we're definitely going to emit a
// constant pool?
RISCVMatInt::InstSeq Seq = RISCVMatInt::generateInstSeq(Val, Subtarget);
return Seq.size() <= Subtarget.getMaxBuildIntsCost();
}
bool RISCVTargetLowering::
shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
unsigned OldShiftOpcode, unsigned NewShiftOpcode,
SelectionDAG &DAG) const {
// One interesting pattern that we'd want to form is 'bit extract':
// ((1 >> Y) & 1) ==/!= 0
// But we also need to be careful not to try to reverse that fold.
// Is this '((1 >> Y) & 1)'?
if (XC && OldShiftOpcode == ISD::SRL && XC->isOne())
return false; // Keep the 'bit extract' pattern.
// Will this be '((1 >> Y) & 1)' after the transform?
if (NewShiftOpcode == ISD::SRL && CC->isOne())
return true; // Do form the 'bit extract' pattern.
// If 'X' is a constant, and we transform, then we will immediately
// try to undo the fold, thus causing endless combine loop.
// So only do the transform if X is not a constant. This matches the default
// implementation of this function.
return !XC;
}
bool RISCVTargetLowering::canSplatOperand(unsigned Opcode, int Operand) const {
switch (Opcode) {
case Instruction::Add:
case Instruction::Sub:
case Instruction::Mul:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor:
case Instruction::FAdd:
case Instruction::FSub:
case Instruction::FMul:
case Instruction::FDiv:
case Instruction::ICmp:
case Instruction::FCmp:
return true;
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::URem:
case Instruction::SRem:
return Operand == 1;
default:
return false;
}
}
bool RISCVTargetLowering::canSplatOperand(Instruction *I, int Operand) const {
if (!I->getType()->isVectorTy() || !Subtarget.hasVInstructions())
return false;
if (canSplatOperand(I->getOpcode(), Operand))
return true;
auto *II = dyn_cast<IntrinsicInst>(I);
if (!II)
return false;
switch (II->getIntrinsicID()) {
case Intrinsic::fma:
case Intrinsic::vp_fma:
return Operand == 0 || Operand == 1;
case Intrinsic::vp_shl:
case Intrinsic::vp_lshr:
case Intrinsic::vp_ashr:
case Intrinsic::vp_udiv:
case Intrinsic::vp_sdiv:
case Intrinsic::vp_urem:
case Intrinsic::vp_srem:
case Intrinsic::ssub_sat:
case Intrinsic::vp_ssub_sat:
case Intrinsic::usub_sat:
case Intrinsic::vp_usub_sat:
return Operand == 1;
// These intrinsics are commutative.
case Intrinsic::vp_add:
case Intrinsic::vp_mul:
case Intrinsic::vp_and:
case Intrinsic::vp_or:
case Intrinsic::vp_xor:
case Intrinsic::vp_fadd:
case Intrinsic::vp_fmul:
case Intrinsic::vp_icmp:
case Intrinsic::vp_fcmp:
case Intrinsic::smin:
case Intrinsic::vp_smin:
case Intrinsic::umin:
case Intrinsic::vp_umin:
case Intrinsic::smax:
case Intrinsic::vp_smax:
case Intrinsic::umax:
case Intrinsic::vp_umax:
case Intrinsic::sadd_sat:
case Intrinsic::vp_sadd_sat:
case Intrinsic::uadd_sat:
case Intrinsic::vp_uadd_sat:
// These intrinsics have 'vr' versions.
case Intrinsic::vp_sub:
case Intrinsic::vp_fsub:
case Intrinsic::vp_fdiv:
return Operand == 0 || Operand == 1;
default:
return false;
}
}
/// Check if sinking \p I's operands to I's basic block is profitable, because
/// the operands can be folded into a target instruction, e.g.
/// splats of scalars can fold into vector instructions.
bool RISCVTargetLowering::shouldSinkOperands(
Instruction *I, SmallVectorImpl<Use *> &Ops) const {
using namespace llvm::PatternMatch;
if (!I->getType()->isVectorTy() || !Subtarget.hasVInstructions())
return false;
// Don't sink splat operands if the target prefers it. Some targets requires
// S2V transfer buffers and we can run out of them copying the same value
// repeatedly.
// FIXME: It could still be worth doing if it would improve vector register
// pressure and prevent a vector spill.
if (!Subtarget.sinkSplatOperands())
return false;
for (auto OpIdx : enumerate(I->operands())) {
if (!canSplatOperand(I, OpIdx.index()))
continue;
Instruction *Op = dyn_cast<Instruction>(OpIdx.value().get());
// Make sure we are not already sinking this operand
if (!Op || any_of(Ops, [&](Use *U) { return U->get() == Op; }))
continue;
// We are looking for a splat that can be sunk.
if (!match(Op, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
m_Undef(), m_ZeroMask())))
continue;
// Don't sink i1 splats.
if (cast<VectorType>(Op->getType())->getElementType()->isIntegerTy(1))
continue;
// All uses of the shuffle should be sunk to avoid duplicating it across gpr
// and vector registers
for (Use &U : Op->uses()) {
Instruction *Insn = cast<Instruction>(U.getUser());
if (!canSplatOperand(Insn, U.getOperandNo()))
return false;
}
Ops.push_back(&Op->getOperandUse(0));
Ops.push_back(&OpIdx.value());
}
return true;
}
bool RISCVTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
unsigned Opc = VecOp.getOpcode();
// Assume target opcodes can't be scalarized.
// TODO - do we have any exceptions?
if (Opc >= ISD::BUILTIN_OP_END)
return false;
// If the vector op is not supported, try to convert to scalar.
EVT VecVT = VecOp.getValueType();
if (!isOperationLegalOrCustomOrPromote(Opc, VecVT))
return true;
// If the vector op is supported, but the scalar op is not, the transform may
// not be worthwhile.
// Permit a vector binary operation can be converted to scalar binary
// operation which is custom lowered with illegal type.
EVT ScalarVT = VecVT.getScalarType();
return isOperationLegalOrCustomOrPromote(Opc, ScalarVT) ||
isOperationCustom(Opc, ScalarVT);
}
bool RISCVTargetLowering::isOffsetFoldingLegal(
const GlobalAddressSDNode *GA) const {
// In order to maximise the opportunity for common subexpression elimination,
// keep a separate ADD node for the global address offset instead of folding
// it in the global address node. Later peephole optimisations may choose to
// fold it back in when profitable.
return false;
}
// Return one of the followings:
// (1) `{0-31 value, false}` if FLI is available for Imm's type and FP value.
// (2) `{0-31 value, true}` if Imm is negative and FLI is available for its
// positive counterpart, which will be materialized from the first returned
// element. The second returned element indicated that there should be a FNEG
// followed.
// (3) `{-1, _}` if there is no way FLI can be used to materialize Imm.
std::pair<int, bool> RISCVTargetLowering::getLegalZfaFPImm(const APFloat &Imm,
EVT VT) const {
if (!Subtarget.hasStdExtZfa())
return std::make_pair(-1, false);
bool IsSupportedVT = false;
if (VT == MVT::f16) {
IsSupportedVT = Subtarget.hasStdExtZfh() || Subtarget.hasStdExtZvfh();
} else if (VT == MVT::f32) {
IsSupportedVT = true;
} else if (VT == MVT::f64) {
assert(Subtarget.hasStdExtD() && "Expect D extension");
IsSupportedVT = true;
}
if (!IsSupportedVT)
return std::make_pair(-1, false);
int Index = RISCVLoadFPImm::getLoadFPImm(Imm);
if (Index < 0 && Imm.isNegative())
// Try the combination of its positive counterpart + FNEG.
return std::make_pair(RISCVLoadFPImm::getLoadFPImm(-Imm), true);
else
return std::make_pair(Index, false);
}
bool RISCVTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
bool ForCodeSize) const {
bool IsLegalVT = false;
if (VT == MVT::f16)
IsLegalVT = Subtarget.hasStdExtZfhminOrZhinxmin();
else if (VT == MVT::f32)
IsLegalVT = Subtarget.hasStdExtFOrZfinx();
else if (VT == MVT::f64)
IsLegalVT = Subtarget.hasStdExtDOrZdinx();
else if (VT == MVT::bf16)
IsLegalVT = Subtarget.hasStdExtZfbfmin();
if (!IsLegalVT)
return false;
if (getLegalZfaFPImm(Imm, VT).first >= 0)
return true;
// Cannot create a 64 bit floating-point immediate value for rv32.
if (Subtarget.getXLen() < VT.getScalarSizeInBits()) {
// td can handle +0.0 or -0.0 already.
// -0.0 can be created by fmv + fneg.
return Imm.isZero();
}
// Special case: fmv + fneg
if (Imm.isNegZero())
return true;
// Building an integer and then converting requires a fmv at the end of
// the integer sequence.
const int Cost =
1 + RISCVMatInt::getIntMatCost(Imm.bitcastToAPInt(), Subtarget.getXLen(),
Subtarget);
return Cost <= FPImmCost;
}
// TODO: This is very conservative.
bool RISCVTargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
unsigned Index) const {
if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
return false;
// Only support extracting a fixed from a fixed vector for now.
if (ResVT.isScalableVector() || SrcVT.isScalableVector())
return false;
EVT EltVT = ResVT.getVectorElementType();
assert(EltVT == SrcVT.getVectorElementType() && "Should hold for node");
// The smallest type we can slide is i8.
// TODO: We can extract index 0 from a mask vector without a slide.
if (EltVT == MVT::i1)
return false;
unsigned ResElts = ResVT.getVectorNumElements();
unsigned SrcElts = SrcVT.getVectorNumElements();
unsigned MinVLen = Subtarget.getRealMinVLen();
unsigned MinVLMAX = MinVLen / EltVT.getSizeInBits();
// If we're extracting only data from the first VLEN bits of the source
// then we can always do this with an m1 vslidedown.vx. Restricting the
// Index ensures we can use a vslidedown.vi.
// TODO: We can generalize this when the exact VLEN is known.
if (Index + ResElts <= MinVLMAX && Index < 31)
return true;
// Convervatively only handle extracting half of a vector.
// TODO: For sizes which aren't multiples of VLEN sizes, this may not be
// a cheap extract. However, this case is important in practice for
// shuffled extracts of longer vectors. How resolve?
if ((ResElts * 2) != SrcElts)
return false;
// Slide can support arbitrary index, but we only treat vslidedown.vi as
// cheap.
if (Index >= 32)
return false;
// TODO: We can do arbitrary slidedowns, but for now only support extracting
// the upper half of a vector until we have more test coverage.
return Index == 0 || Index == ResElts;
}
MVT RISCVTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
CallingConv::ID CC,
EVT VT) const {
// Use f32 to pass f16 if it is legal and Zfh/Zfhmin is not enabled.
// We might still end up using a GPR but that will be decided based on ABI.
if (VT == MVT::f16 && Subtarget.hasStdExtFOrZfinx() &&
!Subtarget.hasStdExtZfhminOrZhinxmin())
return MVT::f32;
MVT PartVT = TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
if (RV64LegalI32 && Subtarget.is64Bit() && PartVT == MVT::i32)
return MVT::i64;
return PartVT;
}
unsigned RISCVTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
CallingConv::ID CC,
EVT VT) const {
// Use f32 to pass f16 if it is legal and Zfh/Zfhmin is not enabled.
// We might still end up using a GPR but that will be decided based on ABI.
if (VT == MVT::f16 && Subtarget.hasStdExtFOrZfinx() &&
!Subtarget.hasStdExtZfhminOrZhinxmin())
return 1;
return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
}
unsigned RISCVTargetLowering::getVectorTypeBreakdownForCallingConv(
LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
unsigned &NumIntermediates, MVT &RegisterVT) const {
unsigned NumRegs = TargetLowering::getVectorTypeBreakdownForCallingConv(
Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
if (RV64LegalI32 && Subtarget.is64Bit() && IntermediateVT == MVT::i32)
IntermediateVT = MVT::i64;
if (RV64LegalI32 && Subtarget.is64Bit() && RegisterVT == MVT::i32)
RegisterVT = MVT::i64;
return NumRegs;
}
// Changes the condition code and swaps operands if necessary, so the SetCC
// operation matches one of the comparisons supported directly by branches
// in the RISC-V ISA. May adjust compares to favor compare with 0 over compare
// with 1/-1.
static void translateSetCCForBranch(const SDLoc &DL, SDValue &LHS, SDValue &RHS,
ISD::CondCode &CC, SelectionDAG &DAG) {
// If this is a single bit test that can't be handled by ANDI, shift the
// bit to be tested to the MSB and perform a signed compare with 0.
if (isIntEqualitySetCC(CC) && isNullConstant(RHS) &&
LHS.getOpcode() == ISD::AND && LHS.hasOneUse() &&
isa<ConstantSDNode>(LHS.getOperand(1))) {
uint64_t Mask = LHS.getConstantOperandVal(1);
if ((isPowerOf2_64(Mask) || isMask_64(Mask)) && !isInt<12>(Mask)) {
unsigned ShAmt = 0;
if (isPowerOf2_64(Mask)) {
CC = CC == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
ShAmt = LHS.getValueSizeInBits() - 1 - Log2_64(Mask);
} else {
ShAmt = LHS.getValueSizeInBits() - llvm::bit_width(Mask);
}
LHS = LHS.getOperand(0);
if (ShAmt != 0)
LHS = DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS,
DAG.getConstant(ShAmt, DL, LHS.getValueType()));
return;
}
}
if (auto *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
int64_t C = RHSC->getSExtValue();
switch (CC) {
default: break;
case ISD::SETGT:
// Convert X > -1 to X >= 0.
if (C == -1) {
RHS = DAG.getConstant(0, DL, RHS.getValueType());
CC = ISD::SETGE;
return;
}
break;
case ISD::SETLT:
// Convert X < 1 to 0 >= X.
if (C == 1) {
RHS = LHS;
LHS = DAG.getConstant(0, DL, RHS.getValueType());
CC = ISD::SETGE;
return;
}
break;
}
}
switch (CC) {
default:
break;
case ISD::SETGT:
case ISD::SETLE:
case ISD::SETUGT:
case ISD::SETULE:
CC = ISD::getSetCCSwappedOperands(CC);
std::swap(LHS, RHS);
break;
}
}
RISCVII::VLMUL RISCVTargetLowering::getLMUL(MVT VT) {
assert(VT.isScalableVector() && "Expecting a scalable vector type");
unsigned KnownSize = VT.getSizeInBits().getKnownMinValue();
if (VT.getVectorElementType() == MVT::i1)
KnownSize *= 8;
switch (KnownSize) {
default:
llvm_unreachable("Invalid LMUL.");
case 8:
return RISCVII::VLMUL::LMUL_F8;
case 16:
return RISCVII::VLMUL::LMUL_F4;
case 32:
return RISCVII::VLMUL::LMUL_F2;
case 64:
return RISCVII::VLMUL::LMUL_1;
case 128:
return RISCVII::VLMUL::LMUL_2;
case 256:
return RISCVII::VLMUL::LMUL_4;
case 512:
return RISCVII::VLMUL::LMUL_8;
}
}
unsigned RISCVTargetLowering::getRegClassIDForLMUL(RISCVII::VLMUL LMul) {
switch (LMul) {
default:
llvm_unreachable("Invalid LMUL.");
case RISCVII::VLMUL::LMUL_F8:
case RISCVII::VLMUL::LMUL_F4:
case RISCVII::VLMUL::LMUL_F2:
case RISCVII::VLMUL::LMUL_1:
return RISCV::VRRegClassID;
case RISCVII::VLMUL::LMUL_2:
return RISCV::VRM2RegClassID;
case RISCVII::VLMUL::LMUL_4:
return RISCV::VRM4RegClassID;
case RISCVII::VLMUL::LMUL_8:
return RISCV::VRM8RegClassID;
}
}
unsigned RISCVTargetLowering::getSubregIndexByMVT(MVT VT, unsigned Index) {
RISCVII::VLMUL LMUL = getLMUL(VT);
if (LMUL == RISCVII::VLMUL::LMUL_F8 ||
LMUL == RISCVII::VLMUL::LMUL_F4 ||
LMUL == RISCVII::VLMUL::LMUL_F2 ||
LMUL == RISCVII::VLMUL::LMUL_1) {
static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7,
"Unexpected subreg numbering");
return RISCV::sub_vrm1_0 + Index;
}
if (LMUL == RISCVII::VLMUL::LMUL_2) {
static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3,
"Unexpected subreg numbering");
return RISCV::sub_vrm2_0 + Index;
}
if (LMUL == RISCVII::VLMUL::LMUL_4) {
static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1,
"Unexpected subreg numbering");
return RISCV::sub_vrm4_0 + Index;
}
llvm_unreachable("Invalid vector type.");
}
unsigned RISCVTargetLowering::getRegClassIDForVecVT(MVT VT) {
if (VT.getVectorElementType() == MVT::i1)
return RISCV::VRRegClassID;
return getRegClassIDForLMUL(getLMUL(VT));
}
// Attempt to decompose a subvector insert/extract between VecVT and
// SubVecVT via subregister indices. Returns the subregister index that
// can perform the subvector insert/extract with the given element index, as
// well as the index corresponding to any leftover subvectors that must be
// further inserted/extracted within the register class for SubVecVT.
std::pair<unsigned, unsigned>
RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
MVT VecVT, MVT SubVecVT, unsigned InsertExtractIdx,
const RISCVRegisterInfo *TRI) {
static_assert((RISCV::VRM8RegClassID > RISCV::VRM4RegClassID &&
RISCV::VRM4RegClassID > RISCV::VRM2RegClassID &&
RISCV::VRM2RegClassID > RISCV::VRRegClassID),
"Register classes not ordered");
unsigned VecRegClassID = getRegClassIDForVecVT(VecVT);
unsigned SubRegClassID = getRegClassIDForVecVT(SubVecVT);
// Try to compose a subregister index that takes us from the incoming
// LMUL>1 register class down to the outgoing one. At each step we half
// the LMUL:
// nxv16i32@12 -> nxv2i32: sub_vrm4_1_then_sub_vrm2_1_then_sub_vrm1_0
// Note that this is not guaranteed to find a subregister index, such as
// when we are extracting from one VR type to another.
unsigned SubRegIdx = RISCV::NoSubRegister;
for (const unsigned RCID :
{RISCV::VRM4RegClassID, RISCV::VRM2RegClassID, RISCV::VRRegClassID})
if (VecRegClassID > RCID && SubRegClassID <= RCID) {
VecVT = VecVT.getHalfNumVectorElementsVT();
bool IsHi =
InsertExtractIdx >= VecVT.getVectorElementCount().getKnownMinValue();
SubRegIdx = TRI->composeSubRegIndices(SubRegIdx,
getSubregIndexByMVT(VecVT, IsHi));
if (IsHi)
InsertExtractIdx -= VecVT.getVectorElementCount().getKnownMinValue();
}
return {SubRegIdx, InsertExtractIdx};
}
// Permit combining of mask vectors as BUILD_VECTOR never expands to scalar
// stores for those types.
bool RISCVTargetLowering::mergeStoresAfterLegalization(EVT VT) const {
return !Subtarget.useRVVForFixedLengthVectors() ||
(VT.isFixedLengthVector() && VT.getVectorElementType() == MVT::i1);
}
bool RISCVTargetLowering::isLegalElementTypeForRVV(EVT ScalarTy) const {
if (!ScalarTy.isSimple())
return false;
switch (ScalarTy.getSimpleVT().SimpleTy) {
case MVT::iPTR:
return Subtarget.is64Bit() ? Subtarget.hasVInstructionsI64() : true;
case MVT::i8:
case MVT::i16:
case MVT::i32:
return true;
case MVT::i64:
return Subtarget.hasVInstructionsI64();
case MVT::f16:
return Subtarget.hasVInstructionsF16();
case MVT::f32:
return Subtarget.hasVInstructionsF32();
case MVT::f64:
return Subtarget.hasVInstructionsF64();
default:
return false;
}
}
unsigned RISCVTargetLowering::combineRepeatedFPDivisors() const {
return NumRepeatedDivisors;
}
static SDValue getVLOperand(SDValue Op) {
assert((Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN) &&
"Unexpected opcode");
bool HasChain = Op.getOpcode() == ISD::INTRINSIC_W_CHAIN;
unsigned IntNo = Op.getConstantOperandVal(HasChain ? 1 : 0);
const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II =
RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo);
if (!II)
return SDValue();
return Op.getOperand(II->VLOperand + 1 + HasChain);
}
static bool useRVVForFixedLengthVectorVT(MVT VT,
const RISCVSubtarget &Subtarget) {
assert(VT.isFixedLengthVector() && "Expected a fixed length vector type!");
if (!Subtarget.useRVVForFixedLengthVectors())
return false;
// We only support a set of vector types with a consistent maximum fixed size
// across all supported vector element types to avoid legalization issues.
// Therefore -- since the largest is v1024i8/v512i16/etc -- the largest
// fixed-length vector type we support is 1024 bytes.
if (VT.getFixedSizeInBits() > 1024 * 8)
return false;
unsigned MinVLen = Subtarget.getRealMinVLen();
MVT EltVT = VT.getVectorElementType();
// Don't use RVV for vectors we cannot scalarize if required.
switch (EltVT.SimpleTy) {
// i1 is supported but has different rules.
default:
return false;
case MVT::i1:
// Masks can only use a single register.
if (VT.getVectorNumElements() > MinVLen)
return false;
MinVLen /= 8;
break;
case MVT::i8:
case MVT::i16:
case MVT::i32:
break;
case MVT::i64:
if (!Subtarget.hasVInstructionsI64())
return false;
break;
case MVT::f16:
if (!Subtarget.hasVInstructionsF16Minimal())
return false;
break;
case MVT::f32:
if (!Subtarget.hasVInstructionsF32())
return false;
break;
case MVT::f64:
if (!Subtarget.hasVInstructionsF64())
return false;
break;
}
// Reject elements larger than ELEN.
if (EltVT.getSizeInBits() > Subtarget.getELen())
return false;
unsigned LMul = divideCeil(VT.getSizeInBits(), MinVLen);
// Don't use RVV for types that don't fit.
if (LMul > Subtarget.getMaxLMULForFixedLengthVectors())
return false;
// TODO: Perhaps an artificial restriction, but worth having whilst getting
// the base fixed length RVV support in place.
if (!VT.isPow2VectorType())
return false;
return true;
}
bool RISCVTargetLowering::useRVVForFixedLengthVectorVT(MVT VT) const {
return ::useRVVForFixedLengthVectorVT(VT, Subtarget);
}
// Return the largest legal scalable vector type that matches VT's element type.
static MVT getContainerForFixedLengthVector(const TargetLowering &TLI, MVT VT,
const RISCVSubtarget &Subtarget) {
// This may be called before legal types are setup.
assert(((VT.isFixedLengthVector() && TLI.isTypeLegal(VT)) ||
useRVVForFixedLengthVectorVT(VT, Subtarget)) &&
"Expected legal fixed length vector!");
unsigned MinVLen = Subtarget.getRealMinVLen();
unsigned MaxELen = Subtarget.getELen();
MVT EltVT = VT.getVectorElementType();
switch (EltVT.SimpleTy) {
default:
llvm_unreachable("unexpected element type for RVV container");
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
case MVT::f16:
case MVT::f32:
case MVT::f64: {
// We prefer to use LMUL=1 for VLEN sized types. Use fractional lmuls for
// narrower types. The smallest fractional LMUL we support is 8/ELEN. Within
// each fractional LMUL we support SEW between 8 and LMUL*ELEN.
unsigned NumElts =
(VT.getVectorNumElements() * RISCV::RVVBitsPerBlock) / MinVLen;
NumElts = std::max(NumElts, RISCV::RVVBitsPerBlock / MaxELen);
assert(isPowerOf2_32(NumElts) && "Expected power of 2 NumElts");
return MVT::getScalableVectorVT(EltVT, NumElts);
}
}
}
static MVT getContainerForFixedLengthVector(SelectionDAG &DAG, MVT VT,
const RISCVSubtarget &Subtarget) {
return getContainerForFixedLengthVector(DAG.getTargetLoweringInfo(), VT,
Subtarget);
}
MVT RISCVTargetLowering::getContainerForFixedLengthVector(MVT VT) const {
return ::getContainerForFixedLengthVector(*this, VT, getSubtarget());
}
// Grow V to consume an entire RVV register.
static SDValue convertToScalableVector(EVT VT, SDValue V, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(VT.isScalableVector() &&
"Expected to convert into a scalable vector!");
assert(V.getValueType().isFixedLengthVector() &&
"Expected a fixed length vector operand!");
SDLoc DL(V);
SDValue Zero = DAG.getVectorIdxConstant(0, DL);
return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, Zero);
}
// Shrink V so it's just big enough to maintain a VT's worth of data.
static SDValue convertFromScalableVector(EVT VT, SDValue V, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(VT.isFixedLengthVector() &&
"Expected to convert into a fixed length vector!");
assert(V.getValueType().isScalableVector() &&
"Expected a scalable vector operand!");
SDLoc DL(V);
SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, Zero);
}
/// Return the type of the mask type suitable for masking the provided
/// vector type. This is simply an i1 element type vector of the same
/// (possibly scalable) length.
static MVT getMaskTypeFor(MVT VecVT) {
assert(VecVT.isVector());
ElementCount EC = VecVT.getVectorElementCount();
return MVT::getVectorVT(MVT::i1, EC);
}
/// Creates an all ones mask suitable for masking a vector of type VecTy with
/// vector length VL. .
static SDValue getAllOnesMask(MVT VecVT, SDValue VL, const SDLoc &DL,
SelectionDAG &DAG) {
MVT MaskVT = getMaskTypeFor(VecVT);
return DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
}
static SDValue getVLOp(uint64_t NumElts, MVT ContainerVT, const SDLoc &DL,
SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
// If we know the exact VLEN, and our VL is exactly equal to VLMAX,
// canonicalize the representation. InsertVSETVLI will pick the immediate
// encoding later if profitable.
const auto [MinVLMAX, MaxVLMAX] =
RISCVTargetLowering::computeVLMAXBounds(ContainerVT, Subtarget);
if (MinVLMAX == MaxVLMAX && NumElts == MinVLMAX)
return DAG.getRegister(RISCV::X0, Subtarget.getXLenVT());
return DAG.getConstant(NumElts, DL, Subtarget.getXLenVT());
}
static std::pair<SDValue, SDValue>
getDefaultScalableVLOps(MVT VecVT, const SDLoc &DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(VecVT.isScalableVector() && "Expecting a scalable vector");
SDValue VL = DAG.getRegister(RISCV::X0, Subtarget.getXLenVT());
SDValue Mask = getAllOnesMask(VecVT, VL, DL, DAG);
return {Mask, VL};
}
static std::pair<SDValue, SDValue>
getDefaultVLOps(uint64_t NumElts, MVT ContainerVT, const SDLoc &DL,
SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
SDValue VL = getVLOp(NumElts, ContainerVT, DL, DAG, Subtarget);
SDValue Mask = getAllOnesMask(ContainerVT, VL, DL, DAG);
return {Mask, VL};
}
// Gets the two common "VL" operands: an all-ones mask and the vector length.
// VecVT is a vector type, either fixed-length or scalable, and ContainerVT is
// the vector type that the fixed-length vector is contained in. Otherwise if
// VecVT is scalable, then ContainerVT should be the same as VecVT.
static std::pair<SDValue, SDValue>
getDefaultVLOps(MVT VecVT, MVT ContainerVT, const SDLoc &DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (VecVT.isFixedLengthVector())
return getDefaultVLOps(VecVT.getVectorNumElements(), ContainerVT, DL, DAG,
Subtarget);
assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
return getDefaultScalableVLOps(ContainerVT, DL, DAG, Subtarget);
}
SDValue RISCVTargetLowering::computeVLMax(MVT VecVT, const SDLoc &DL,
SelectionDAG &DAG) const {
assert(VecVT.isScalableVector() && "Expected scalable vector");
return DAG.getElementCount(DL, Subtarget.getXLenVT(),
VecVT.getVectorElementCount());
}
std::pair<unsigned, unsigned>
RISCVTargetLowering::computeVLMAXBounds(MVT VecVT,
const RISCVSubtarget &Subtarget) {
assert(VecVT.isScalableVector() && "Expected scalable vector");
unsigned EltSize = VecVT.getScalarSizeInBits();
unsigned MinSize = VecVT.getSizeInBits().getKnownMinValue();
unsigned VectorBitsMax = Subtarget.getRealMaxVLen();
unsigned MaxVLMAX =
RISCVTargetLowering::computeVLMAX(VectorBitsMax, EltSize, MinSize);
unsigned VectorBitsMin = Subtarget.getRealMinVLen();
unsigned MinVLMAX =
RISCVTargetLowering::computeVLMAX(VectorBitsMin, EltSize, MinSize);
return std::make_pair(MinVLMAX, MaxVLMAX);
}
// The state of RVV BUILD_VECTOR and VECTOR_SHUFFLE lowering is that very few
// of either is (currently) supported. This can get us into an infinite loop
// where we try to lower a BUILD_VECTOR as a VECTOR_SHUFFLE as a BUILD_VECTOR
// as a ..., etc.
// Until either (or both) of these can reliably lower any node, reporting that
// we don't want to expand BUILD_VECTORs via VECTOR_SHUFFLEs at least breaks
// the infinite loop. Note that this lowers BUILD_VECTOR through the stack,
// which is not desirable.
bool RISCVTargetLowering::shouldExpandBuildVectorWithShuffles(
EVT VT, unsigned DefinedValues) const {
return false;
}
InstructionCost RISCVTargetLowering::getLMULCost(MVT VT) const {
// TODO: Here assume reciprocal throughput is 1 for LMUL_1, it is
// implementation-defined.
if (!VT.isVector())
return InstructionCost::getInvalid();
unsigned DLenFactor = Subtarget.getDLenFactor();
unsigned Cost;
if (VT.isScalableVector()) {
unsigned LMul;
bool Fractional;
std::tie(LMul, Fractional) =
RISCVVType::decodeVLMUL(RISCVTargetLowering::getLMUL(VT));
if (Fractional)
Cost = LMul <= DLenFactor ? (DLenFactor / LMul) : 1;
else
Cost = (LMul * DLenFactor);
} else {
Cost = divideCeil(VT.getSizeInBits(), Subtarget.getRealMinVLen() / DLenFactor);
}
return Cost;
}
/// Return the cost of a vrgather.vv instruction for the type VT. vrgather.vv
/// is generally quadratic in the number of vreg implied by LMUL. Note that
/// operand (index and possibly mask) are handled separately.
InstructionCost RISCVTargetLowering::getVRGatherVVCost(MVT VT) const {
return getLMULCost(VT) * getLMULCost(VT);
}
/// Return the cost of a vrgather.vi (or vx) instruction for the type VT.
/// vrgather.vi/vx may be linear in the number of vregs implied by LMUL,
/// or may track the vrgather.vv cost. It is implementation-dependent.
InstructionCost RISCVTargetLowering::getVRGatherVICost(MVT VT) const {
return getLMULCost(VT);
}
/// Return the cost of a vslidedown.vx or vslideup.vx instruction
/// for the type VT. (This does not cover the vslide1up or vslide1down
/// variants.) Slides may be linear in the number of vregs implied by LMUL,
/// or may track the vrgather.vv cost. It is implementation-dependent.
InstructionCost RISCVTargetLowering::getVSlideVXCost(MVT VT) const {
return getLMULCost(VT);
}
/// Return the cost of a vslidedown.vi or vslideup.vi instruction
/// for the type VT. (This does not cover the vslide1up or vslide1down
/// variants.) Slides may be linear in the number of vregs implied by LMUL,
/// or may track the vrgather.vv cost. It is implementation-dependent.
InstructionCost RISCVTargetLowering::getVSlideVICost(MVT VT) const {
return getLMULCost(VT);
}
static SDValue lowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
// RISC-V FP-to-int conversions saturate to the destination register size, but
// don't produce 0 for nan. We can use a conversion instruction and fix the
// nan case with a compare and a select.
SDValue Src = Op.getOperand(0);
MVT DstVT = Op.getSimpleValueType();
EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT_SAT;
if (!DstVT.isVector()) {
// For bf16 or for f16 in absense of Zfh, promote to f32, then saturate
// the result.
if ((Src.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfhOrZhinx()) ||
Src.getValueType() == MVT::bf16) {
Src = DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), MVT::f32, Src);
}
unsigned Opc;
if (SatVT == DstVT)
Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
else if (DstVT == MVT::i64 && SatVT == MVT::i32)
Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
else
return SDValue();
// FIXME: Support other SatVTs by clamping before or after the conversion.
SDLoc DL(Op);
SDValue FpToInt = DAG.getNode(
Opc, DL, DstVT, Src,
DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, Subtarget.getXLenVT()));
if (Opc == RISCVISD::FCVT_WU_RV64)
FpToInt = DAG.getZeroExtendInReg(FpToInt, DL, MVT::i32);
SDValue ZeroInt = DAG.getConstant(0, DL, DstVT);
return DAG.getSelectCC(DL, Src, Src, ZeroInt, FpToInt,
ISD::CondCode::SETUO);
}
// Vectors.
MVT DstEltVT = DstVT.getVectorElementType();
MVT SrcVT = Src.getSimpleValueType();
MVT SrcEltVT = SrcVT.getVectorElementType();
unsigned SrcEltSize = SrcEltVT.getSizeInBits();
unsigned DstEltSize = DstEltVT.getSizeInBits();
// Only handle saturating to the destination type.
if (SatVT != DstEltVT)
return SDValue();
// FIXME: Don't support narrowing by more than 1 steps for now.
if (SrcEltSize > (2 * DstEltSize))
return SDValue();
MVT DstContainerVT = DstVT;
MVT SrcContainerVT = SrcVT;
if (DstVT.isFixedLengthVector()) {
DstContainerVT = getContainerForFixedLengthVector(DAG, DstVT, Subtarget);
SrcContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
assert(DstContainerVT.getVectorElementCount() ==
SrcContainerVT.getVectorElementCount() &&
"Expected same element count");
Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
}
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(DstVT, DstContainerVT, DL, DAG, Subtarget);
SDValue IsNan = DAG.getNode(RISCVISD::SETCC_VL, DL, Mask.getValueType(),
{Src, Src, DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(Mask.getValueType()), Mask, VL});
// Need to widen by more than 1 step, promote the FP type, then do a widening
// convert.
if (DstEltSize > (2 * SrcEltSize)) {
assert(SrcContainerVT.getVectorElementType() == MVT::f16 && "Unexpected VT!");
MVT InterVT = SrcContainerVT.changeVectorElementType(MVT::f32);
Src = DAG.getNode(RISCVISD::FP_EXTEND_VL, DL, InterVT, Src, Mask, VL);
}
unsigned RVVOpc =
IsSigned ? RISCVISD::VFCVT_RTZ_X_F_VL : RISCVISD::VFCVT_RTZ_XU_F_VL;
SDValue Res = DAG.getNode(RVVOpc, DL, DstContainerVT, Src, Mask, VL);
SDValue SplatZero = DAG.getNode(
RISCVISD::VMV_V_X_VL, DL, DstContainerVT, DAG.getUNDEF(DstContainerVT),
DAG.getConstant(0, DL, Subtarget.getXLenVT()), VL);
Res = DAG.getNode(RISCVISD::VMERGE_VL, DL, DstContainerVT, IsNan, SplatZero,
Res, DAG.getUNDEF(DstContainerVT), VL);
if (DstVT.isFixedLengthVector())
Res = convertFromScalableVector(DstVT, Res, DAG, Subtarget);
return Res;
}
static RISCVFPRndMode::RoundingMode matchRoundingOp(unsigned Opc) {
switch (Opc) {
case ISD::FROUNDEVEN:
case ISD::STRICT_FROUNDEVEN:
case ISD::VP_FROUNDEVEN:
return RISCVFPRndMode::RNE;
case ISD::FTRUNC:
case ISD::STRICT_FTRUNC:
case ISD::VP_FROUNDTOZERO:
return RISCVFPRndMode::RTZ;
case ISD::FFLOOR:
case ISD::STRICT_FFLOOR:
case ISD::VP_FFLOOR:
return RISCVFPRndMode::RDN;
case ISD::FCEIL:
case ISD::STRICT_FCEIL:
case ISD::VP_FCEIL:
return RISCVFPRndMode::RUP;
case ISD::FROUND:
case ISD::STRICT_FROUND:
case ISD::VP_FROUND:
return RISCVFPRndMode::RMM;
case ISD::FRINT:
return RISCVFPRndMode::DYN;
}
return RISCVFPRndMode::Invalid;
}
// Expand vector FTRUNC, FCEIL, FFLOOR, FROUND, VP_FCEIL, VP_FFLOOR, VP_FROUND
// VP_FROUNDEVEN, VP_FROUNDTOZERO, VP_FRINT and VP_FNEARBYINT by converting to
// the integer domain and back. Taking care to avoid converting values that are
// nan or already correct.
static SDValue
lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
assert(VT.isVector() && "Unexpected type");
SDLoc DL(Op);
SDValue Src = Op.getOperand(0);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
}
SDValue Mask, VL;
if (Op->isVPOpcode()) {
Mask = Op.getOperand(1);
if (VT.isFixedLengthVector())
Mask = convertToScalableVector(getMaskTypeFor(ContainerVT), Mask, DAG,
Subtarget);
VL = Op.getOperand(2);
} else {
std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
}
// Freeze the source since we are increasing the number of uses.
Src = DAG.getFreeze(Src);
// We do the conversion on the absolute value and fix the sign at the end.
SDValue Abs = DAG.getNode(RISCVISD::FABS_VL, DL, ContainerVT, Src, Mask, VL);
// Determine the largest integer that can be represented exactly. This and
// values larger than it don't have any fractional bits so don't need to
// be converted.
const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(ContainerVT);
unsigned Precision = APFloat::semanticsPrecision(FltSem);
APFloat MaxVal = APFloat(FltSem);
MaxVal.convertFromAPInt(APInt::getOneBitSet(Precision, Precision - 1),
/*IsSigned*/ false, APFloat::rmNearestTiesToEven);
SDValue MaxValNode =
DAG.getConstantFP(MaxVal, DL, ContainerVT.getVectorElementType());
SDValue MaxValSplat = DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), MaxValNode, VL);
// If abs(Src) was larger than MaxVal or nan, keep it.
MVT SetccVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
Mask =
DAG.getNode(RISCVISD::SETCC_VL, DL, SetccVT,
{Abs, MaxValSplat, DAG.getCondCode(ISD::SETOLT),
Mask, Mask, VL});
// Truncate to integer and convert back to FP.
MVT IntVT = ContainerVT.changeVectorElementTypeToInteger();
MVT XLenVT = Subtarget.getXLenVT();
SDValue Truncated;
switch (Op.getOpcode()) {
default:
llvm_unreachable("Unexpected opcode");
case ISD::FCEIL:
case ISD::VP_FCEIL:
case ISD::FFLOOR:
case ISD::VP_FFLOOR:
case ISD::FROUND:
case ISD::FROUNDEVEN:
case ISD::VP_FROUND:
case ISD::VP_FROUNDEVEN:
case ISD::VP_FROUNDTOZERO: {
RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
assert(FRM != RISCVFPRndMode::Invalid);
Truncated = DAG.getNode(RISCVISD::VFCVT_RM_X_F_VL, DL, IntVT, Src, Mask,
DAG.getTargetConstant(FRM, DL, XLenVT), VL);
break;
}
case ISD::FTRUNC:
Truncated = DAG.getNode(RISCVISD::VFCVT_RTZ_X_F_VL, DL, IntVT, Src,
Mask, VL);
break;
case ISD::FRINT:
case ISD::VP_FRINT:
Truncated = DAG.getNode(RISCVISD::VFCVT_X_F_VL, DL, IntVT, Src, Mask, VL);
break;
case ISD::FNEARBYINT:
case ISD::VP_FNEARBYINT:
Truncated = DAG.getNode(RISCVISD::VFROUND_NOEXCEPT_VL, DL, ContainerVT, Src,
Mask, VL);
break;
}
// VFROUND_NOEXCEPT_VL includes SINT_TO_FP_VL.
if (Truncated.getOpcode() != RISCVISD::VFROUND_NOEXCEPT_VL)
Truncated = DAG.getNode(RISCVISD::SINT_TO_FP_VL, DL, ContainerVT, Truncated,
Mask, VL);
// Restore the original sign so that -0.0 is preserved.
Truncated = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Truncated,
Src, Src, Mask, VL);
if (!VT.isFixedLengthVector())
return Truncated;
return convertFromScalableVector(VT, Truncated, DAG, Subtarget);
}
// Expand vector STRICT_FTRUNC, STRICT_FCEIL, STRICT_FFLOOR, STRICT_FROUND
// STRICT_FROUNDEVEN and STRICT_FNEARBYINT by converting sNan of the source to
// qNan and coverting the new source to integer and back to FP.
static SDValue
lowerVectorStrictFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue Chain = Op.getOperand(0);
SDValue Src = Op.getOperand(1);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
}
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
// Freeze the source since we are increasing the number of uses.
Src = DAG.getFreeze(Src);
// Covert sNan to qNan by executing x + x for all unordered elemenet x in Src.
MVT MaskVT = Mask.getSimpleValueType();
SDValue Unorder = DAG.getNode(RISCVISD::STRICT_FSETCC_VL, DL,
DAG.getVTList(MaskVT, MVT::Other),
{Chain, Src, Src, DAG.getCondCode(ISD::SETUNE),
DAG.getUNDEF(MaskVT), Mask, VL});
Chain = Unorder.getValue(1);
Src = DAG.getNode(RISCVISD::STRICT_FADD_VL, DL,
DAG.getVTList(ContainerVT, MVT::Other),
{Chain, Src, Src, DAG.getUNDEF(ContainerVT), Unorder, VL});
Chain = Src.getValue(1);
// We do the conversion on the absolute value and fix the sign at the end.
SDValue Abs = DAG.getNode(RISCVISD::FABS_VL, DL, ContainerVT, Src, Mask, VL);
// Determine the largest integer that can be represented exactly. This and
// values larger than it don't have any fractional bits so don't need to
// be converted.
const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(ContainerVT);
unsigned Precision = APFloat::semanticsPrecision(FltSem);
APFloat MaxVal = APFloat(FltSem);
MaxVal.convertFromAPInt(APInt::getOneBitSet(Precision, Precision - 1),
/*IsSigned*/ false, APFloat::rmNearestTiesToEven);
SDValue MaxValNode =
DAG.getConstantFP(MaxVal, DL, ContainerVT.getVectorElementType());
SDValue MaxValSplat = DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), MaxValNode, VL);
// If abs(Src) was larger than MaxVal or nan, keep it.
Mask = DAG.getNode(
RISCVISD::SETCC_VL, DL, MaskVT,
{Abs, MaxValSplat, DAG.getCondCode(ISD::SETOLT), Mask, Mask, VL});
// Truncate to integer and convert back to FP.
MVT IntVT = ContainerVT.changeVectorElementTypeToInteger();
MVT XLenVT = Subtarget.getXLenVT();
SDValue Truncated;
switch (Op.getOpcode()) {
default:
llvm_unreachable("Unexpected opcode");
case ISD::STRICT_FCEIL:
case ISD::STRICT_FFLOOR:
case ISD::STRICT_FROUND:
case ISD::STRICT_FROUNDEVEN: {
RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
assert(FRM != RISCVFPRndMode::Invalid);
Truncated = DAG.getNode(
RISCVISD::STRICT_VFCVT_RM_X_F_VL, DL, DAG.getVTList(IntVT, MVT::Other),
{Chain, Src, Mask, DAG.getTargetConstant(FRM, DL, XLenVT), VL});
break;
}
case ISD::STRICT_FTRUNC:
Truncated =
DAG.getNode(RISCVISD::STRICT_VFCVT_RTZ_X_F_VL, DL,
DAG.getVTList(IntVT, MVT::Other), Chain, Src, Mask, VL);
break;
case ISD::STRICT_FNEARBYINT:
Truncated = DAG.getNode(RISCVISD::STRICT_VFROUND_NOEXCEPT_VL, DL,
DAG.getVTList(ContainerVT, MVT::Other), Chain, Src,
Mask, VL);
break;
}
Chain = Truncated.getValue(1);
// VFROUND_NOEXCEPT_VL includes SINT_TO_FP_VL.
if (Op.getOpcode() != ISD::STRICT_FNEARBYINT) {
Truncated = DAG.getNode(RISCVISD::STRICT_SINT_TO_FP_VL, DL,
DAG.getVTList(ContainerVT, MVT::Other), Chain,
Truncated, Mask, VL);
Chain = Truncated.getValue(1);
}
// Restore the original sign so that -0.0 is preserved.
Truncated = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Truncated,
Src, Src, Mask, VL);
if (VT.isFixedLengthVector())
Truncated = convertFromScalableVector(VT, Truncated, DAG, Subtarget);
return DAG.getMergeValues({Truncated, Chain}, DL);
}
static SDValue
lowerFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
if (VT.isVector())
return lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
if (DAG.shouldOptForSize())
return SDValue();
SDLoc DL(Op);
SDValue Src = Op.getOperand(0);
// Create an integer the size of the mantissa with the MSB set. This and all
// values larger than it don't have any fractional bits so don't need to be
// converted.
const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(VT);
unsigned Precision = APFloat::semanticsPrecision(FltSem);
APFloat MaxVal = APFloat(FltSem);
MaxVal.convertFromAPInt(APInt::getOneBitSet(Precision, Precision - 1),
/*IsSigned*/ false, APFloat::rmNearestTiesToEven);
SDValue MaxValNode = DAG.getConstantFP(MaxVal, DL, VT);
RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
return DAG.getNode(RISCVISD::FROUND, DL, VT, Src, MaxValNode,
DAG.getTargetConstant(FRM, DL, Subtarget.getXLenVT()));
}
// Expand vector LRINT and LLRINT by converting to the integer domain.
static SDValue lowerVectorXRINT(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
assert(VT.isVector() && "Unexpected type");
SDLoc DL(Op);
SDValue Src = Op.getOperand(0);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
}
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
SDValue Truncated =
DAG.getNode(RISCVISD::VFCVT_X_F_VL, DL, ContainerVT, Src, Mask, VL);
if (!VT.isFixedLengthVector())
return Truncated;
return convertFromScalableVector(VT, Truncated, DAG, Subtarget);
}
static SDValue
getVSlidedown(SelectionDAG &DAG, const RISCVSubtarget &Subtarget,
const SDLoc &DL, EVT VT, SDValue Merge, SDValue Op,
SDValue Offset, SDValue Mask, SDValue VL,
unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED) {
if (Merge.isUndef())
Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
SDValue PolicyOp = DAG.getTargetConstant(Policy, DL, Subtarget.getXLenVT());
SDValue Ops[] = {Merge, Op, Offset, Mask, VL, PolicyOp};
return DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, VT, Ops);
}
static SDValue
getVSlideup(SelectionDAG &DAG, const RISCVSubtarget &Subtarget, const SDLoc &DL,
EVT VT, SDValue Merge, SDValue Op, SDValue Offset, SDValue Mask,
SDValue VL,
unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED) {
if (Merge.isUndef())
Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
SDValue PolicyOp = DAG.getTargetConstant(Policy, DL, Subtarget.getXLenVT());
SDValue Ops[] = {Merge, Op, Offset, Mask, VL, PolicyOp};
return DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, VT, Ops);
}
static MVT getLMUL1VT(MVT VT) {
assert(VT.getVectorElementType().getSizeInBits() <= 64 &&
"Unexpected vector MVT");
return MVT::getScalableVectorVT(
VT.getVectorElementType(),
RISCV::RVVBitsPerBlock / VT.getVectorElementType().getSizeInBits());
}
struct VIDSequence {
int64_t StepNumerator;
unsigned StepDenominator;
int64_t Addend;
};
static std::optional<uint64_t> getExactInteger(const APFloat &APF,
uint32_t BitWidth) {
// We will use a SINT_TO_FP to materialize this constant so we should use a
// signed APSInt here.
APSInt ValInt(BitWidth, /*IsUnsigned*/ false);
// We use an arbitrary rounding mode here. If a floating-point is an exact
// integer (e.g., 1.0), the rounding mode does not affect the output value. If
// the rounding mode changes the output value, then it is not an exact
// integer.
RoundingMode ArbitraryRM = RoundingMode::TowardZero;
bool IsExact;
// If it is out of signed integer range, it will return an invalid operation.
// If it is not an exact integer, IsExact is false.
if ((APF.convertToInteger(ValInt, ArbitraryRM, &IsExact) ==
APFloatBase::opInvalidOp) ||
!IsExact)
return std::nullopt;
return ValInt.extractBitsAsZExtValue(BitWidth, 0);
}
// Try to match an arithmetic-sequence BUILD_VECTOR [X,X+S,X+2*S,...,X+(N-1)*S]
// to the (non-zero) step S and start value X. This can be then lowered as the
// RVV sequence (VID * S) + X, for example.
// The step S is represented as an integer numerator divided by a positive
// denominator. Note that the implementation currently only identifies
// sequences in which either the numerator is +/- 1 or the denominator is 1. It
// cannot detect 2/3, for example.
// Note that this method will also match potentially unappealing index
// sequences, like <i32 0, i32 50939494>, however it is left to the caller to
// determine whether this is worth generating code for.
static std::optional<VIDSequence> isSimpleVIDSequence(SDValue Op,
unsigned EltSizeInBits) {
assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unexpected BUILD_VECTOR");
if (!cast<BuildVectorSDNode>(Op)->isConstant())
return std::nullopt;
bool IsInteger = Op.getValueType().isInteger();
std::optional<unsigned> SeqStepDenom;
std::optional<int64_t> SeqStepNum, SeqAddend;
std::optional<std::pair<uint64_t, unsigned>> PrevElt;
assert(EltSizeInBits >= Op.getValueType().getScalarSizeInBits());
// First extract the ops into a list of constant integer values. This may not
// be possible for floats if they're not all representable as integers.
SmallVector<std::optional<uint64_t>> Elts(Op.getNumOperands());
const unsigned OpSize = Op.getScalarValueSizeInBits();
for (auto [Idx, Elt] : enumerate(Op->op_values())) {
if (Elt.isUndef()) {
Elts[Idx] = std::nullopt;
continue;
}
if (IsInteger) {
Elts[Idx] = Elt->getAsZExtVal() & maskTrailingOnes<uint64_t>(OpSize);
} else {
auto ExactInteger =
getExactInteger(cast<ConstantFPSDNode>(Elt)->getValueAPF(), OpSize);
if (!ExactInteger)
return std::nullopt;
Elts[Idx] = *ExactInteger;
}
}
for (auto [Idx, Elt] : enumerate(Elts)) {
// Assume undef elements match the sequence; we just have to be careful
// when interpolating across them.
if (!Elt)
continue;
if (PrevElt) {
// Calculate the step since the last non-undef element, and ensure
// it's consistent across the entire sequence.
unsigned IdxDiff = Idx - PrevElt->second;
int64_t ValDiff = SignExtend64(*Elt - PrevElt->first, EltSizeInBits);
// A zero-value value difference means that we're somewhere in the middle
// of a fractional step, e.g. <0,0,0*,0,1,1,1,1>. Wait until we notice a
// step change before evaluating the sequence.
if (ValDiff == 0)
continue;
int64_t Remainder = ValDiff % IdxDiff;
// Normalize the step if it's greater than 1.
if (Remainder != ValDiff) {
// The difference must cleanly divide the element span.
if (Remainder != 0)
return std::nullopt;
ValDiff /= IdxDiff;
IdxDiff = 1;
}
if (!SeqStepNum)
SeqStepNum = ValDiff;
else if (ValDiff != SeqStepNum)
return std::nullopt;
if (!SeqStepDenom)
SeqStepDenom = IdxDiff;
else if (IdxDiff != *SeqStepDenom)
return std::nullopt;
}
// Record this non-undef element for later.
if (!PrevElt || PrevElt->first != *Elt)
PrevElt = std::make_pair(*Elt, Idx);
}
// We need to have logged a step for this to count as a legal index sequence.
if (!SeqStepNum || !SeqStepDenom)
return std::nullopt;
// Loop back through the sequence and validate elements we might have skipped
// while waiting for a valid step. While doing this, log any sequence addend.
for (auto [Idx, Elt] : enumerate(Elts)) {
if (!Elt)
continue;
uint64_t ExpectedVal =
(int64_t)(Idx * (uint64_t)*SeqStepNum) / *SeqStepDenom;
int64_t Addend = SignExtend64(*Elt - ExpectedVal, EltSizeInBits);
if (!SeqAddend)
SeqAddend = Addend;
else if (Addend != SeqAddend)
return std::nullopt;
}
assert(SeqAddend && "Must have an addend if we have a step");
return VIDSequence{*SeqStepNum, *SeqStepDenom, *SeqAddend};
}
// Match a splatted value (SPLAT_VECTOR/BUILD_VECTOR) of an EXTRACT_VECTOR_ELT
// and lower it as a VRGATHER_VX_VL from the source vector.
static SDValue matchSplatAsGather(SDValue SplatVal, MVT VT, const SDLoc &DL,
SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (SplatVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return SDValue();
SDValue Vec = SplatVal.getOperand(0);
// Only perform this optimization on vectors of the same size for simplicity.
// Don't perform this optimization for i1 vectors.
// FIXME: Support i1 vectors, maybe by promoting to i8?
if (Vec.getValueType() != VT || VT.getVectorElementType() == MVT::i1)
return SDValue();
SDValue Idx = SplatVal.getOperand(1);
// The index must be a legal type.
if (Idx.getValueType() != Subtarget.getXLenVT())
return SDValue();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
SDValue Gather = DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT, Vec,
Idx, DAG.getUNDEF(ContainerVT), Mask, VL);
if (!VT.isFixedLengthVector())
return Gather;
return convertFromScalableVector(VT, Gather, DAG, Subtarget);
}
/// Try and optimize BUILD_VECTORs with "dominant values" - these are values
/// which constitute a large proportion of the elements. In such cases we can
/// splat a vector with the dominant element and make up the shortfall with
/// INSERT_VECTOR_ELTs. Returns SDValue if not profitable.
/// Note that this includes vectors of 2 elements by association. The
/// upper-most element is the "dominant" one, allowing us to use a splat to
/// "insert" the upper element, and an insert of the lower element at position
/// 0, which improves codegen.
static SDValue lowerBuildVectorViaDominantValues(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
assert(VT.isFixedLengthVector() && "Unexpected vector!");
MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
MVT XLenVT = Subtarget.getXLenVT();
unsigned NumElts = Op.getNumOperands();
SDValue DominantValue;
unsigned MostCommonCount = 0;
DenseMap<SDValue, unsigned> ValueCounts;
unsigned NumUndefElts =
count_if(Op->op_values(), [](const SDValue &V) { return V.isUndef(); });
// Track the number of scalar loads we know we'd be inserting, estimated as
// any non-zero floating-point constant. Other kinds of element are either
// already in registers or are materialized on demand. The threshold at which
// a vector load is more desirable than several scalar materializion and
// vector-insertion instructions is not known.
unsigned NumScalarLoads = 0;
for (SDValue V : Op->op_values()) {
if (V.isUndef())
continue;
ValueCounts.insert(std::make_pair(V, 0));
unsigned &Count = ValueCounts[V];
if (0 == Count)
if (auto *CFP = dyn_cast<ConstantFPSDNode>(V))
NumScalarLoads += !CFP->isExactlyValue(+0.0);
// Is this value dominant? In case of a tie, prefer the highest element as
// it's cheaper to insert near the beginning of a vector than it is at the
// end.
if (++Count >= MostCommonCount) {
DominantValue = V;
MostCommonCount = Count;
}
}
assert(DominantValue && "Not expecting an all-undef BUILD_VECTOR");
unsigned NumDefElts = NumElts - NumUndefElts;
unsigned DominantValueCountThreshold = NumDefElts <= 2 ? 0 : NumDefElts - 2;
// Don't perform this optimization when optimizing for size, since
// materializing elements and inserting them tends to cause code bloat.
if (!DAG.shouldOptForSize() && NumScalarLoads < NumElts &&
(NumElts != 2 || ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) &&
((MostCommonCount > DominantValueCountThreshold) ||
(ValueCounts.size() <= Log2_32(NumDefElts)))) {
// Start by splatting the most common element.
SDValue Vec = DAG.getSplatBuildVector(VT, DL, DominantValue);
DenseSet<SDValue> Processed{DominantValue};
// We can handle an insert into the last element (of a splat) via
// v(f)slide1down. This is slightly better than the vslideup insert
// lowering as it avoids the need for a vector group temporary. It
// is also better than using vmerge.vx as it avoids the need to
// materialize the mask in a vector register.
if (SDValue LastOp = Op->getOperand(Op->getNumOperands() - 1);
!LastOp.isUndef() && ValueCounts[LastOp] == 1 &&
LastOp != DominantValue) {
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
auto OpCode =
VT.isFloatingPoint() ? RISCVISD::VFSLIDE1DOWN_VL : RISCVISD::VSLIDE1DOWN_VL;
if (!VT.isFloatingPoint())
LastOp = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, LastOp);
Vec = DAG.getNode(OpCode, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Vec,
LastOp, Mask, VL);
Vec = convertFromScalableVector(VT, Vec, DAG, Subtarget);
Processed.insert(LastOp);
}
MVT SelMaskTy = VT.changeVectorElementType(MVT::i1);
for (const auto &OpIdx : enumerate(Op->ops())) {
const SDValue &V = OpIdx.value();
if (V.isUndef() || !Processed.insert(V).second)
continue;
if (ValueCounts[V] == 1) {
Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V,
DAG.getVectorIdxConstant(OpIdx.index(), DL));
} else {
// Blend in all instances of this value using a VSELECT, using a
// mask where each bit signals whether that element is the one
// we're after.
SmallVector<SDValue> Ops;
transform(Op->op_values(), std::back_inserter(Ops), [&](SDValue V1) {
return DAG.getConstant(V == V1, DL, XLenVT);
});
Vec = DAG.getNode(ISD::VSELECT, DL, VT,
DAG.getBuildVector(SelMaskTy, DL, Ops),
DAG.getSplatBuildVector(VT, DL, V), Vec);
}
}
return Vec;
}
return SDValue();
}
static SDValue lowerBuildVectorOfConstants(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
assert(VT.isFixedLengthVector() && "Unexpected vector!");
MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
MVT XLenVT = Subtarget.getXLenVT();
unsigned NumElts = Op.getNumOperands();
if (VT.getVectorElementType() == MVT::i1) {
if (ISD::isBuildVectorAllZeros(Op.getNode())) {
SDValue VMClr = DAG.getNode(RISCVISD::VMCLR_VL, DL, ContainerVT, VL);
return convertFromScalableVector(VT, VMClr, DAG, Subtarget);
}
if (ISD::isBuildVectorAllOnes(Op.getNode())) {
SDValue VMSet = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
return convertFromScalableVector(VT, VMSet, DAG, Subtarget);
}
// Lower constant mask BUILD_VECTORs via an integer vector type, in
// scalar integer chunks whose bit-width depends on the number of mask
// bits and XLEN.
// First, determine the most appropriate scalar integer type to use. This
// is at most XLenVT, but may be shrunk to a smaller vector element type
// according to the size of the final vector - use i8 chunks rather than
// XLenVT if we're producing a v8i1. This results in more consistent
// codegen across RV32 and RV64.
unsigned NumViaIntegerBits = std::clamp(NumElts, 8u, Subtarget.getXLen());
NumViaIntegerBits = std::min(NumViaIntegerBits, Subtarget.getELen());
// If we have to use more than one INSERT_VECTOR_ELT then this
// optimization is likely to increase code size; avoid peforming it in
// such a case. We can use a load from a constant pool in this case.
if (DAG.shouldOptForSize() && NumElts > NumViaIntegerBits)
return SDValue();
// Now we can create our integer vector type. Note that it may be larger
// than the resulting mask type: v4i1 would use v1i8 as its integer type.
unsigned IntegerViaVecElts = divideCeil(NumElts, NumViaIntegerBits);
MVT IntegerViaVecVT =
MVT::getVectorVT(MVT::getIntegerVT(NumViaIntegerBits),
IntegerViaVecElts);
uint64_t Bits = 0;
unsigned BitPos = 0, IntegerEltIdx = 0;
SmallVector<SDValue, 8> Elts(IntegerViaVecElts);
for (unsigned I = 0; I < NumElts;) {
SDValue V = Op.getOperand(I);
bool BitValue = !V.isUndef() && V->getAsZExtVal();
Bits |= ((uint64_t)BitValue << BitPos);
++BitPos;
++I;
// Once we accumulate enough bits to fill our scalar type or process the
// last element, insert into our vector and clear our accumulated data.
if (I % NumViaIntegerBits == 0 || I == NumElts) {
if (NumViaIntegerBits <= 32)
Bits = SignExtend64<32>(Bits);
SDValue Elt = DAG.getConstant(Bits, DL, XLenVT);
Elts[IntegerEltIdx] = Elt;
Bits = 0;
BitPos = 0;
IntegerEltIdx++;
}
}
SDValue Vec = DAG.getBuildVector(IntegerViaVecVT, DL, Elts);
if (NumElts < NumViaIntegerBits) {
// If we're producing a smaller vector than our minimum legal integer
// type, bitcast to the equivalent (known-legal) mask type, and extract
// our final mask.
assert(IntegerViaVecVT == MVT::v1i8 && "Unexpected mask vector type");
Vec = DAG.getBitcast(MVT::v8i1, Vec);
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Vec,
DAG.getConstant(0, DL, XLenVT));
} else {
// Else we must have produced an integer type with the same size as the
// mask type; bitcast for the final result.
assert(VT.getSizeInBits() == IntegerViaVecVT.getSizeInBits());
Vec = DAG.getBitcast(VT, Vec);
}
return Vec;
}
if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL
: RISCVISD::VMV_V_X_VL;
if (!VT.isFloatingPoint())
Splat = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Splat);
Splat =
DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Splat, VL);
return convertFromScalableVector(VT, Splat, DAG, Subtarget);
}
// Try and match index sequences, which we can lower to the vid instruction
// with optional modifications. An all-undef vector is matched by
// getSplatValue, above.
if (auto SimpleVID = isSimpleVIDSequence(Op, Op.getScalarValueSizeInBits())) {
int64_t StepNumerator = SimpleVID->StepNumerator;
unsigned StepDenominator = SimpleVID->StepDenominator;
int64_t Addend = SimpleVID->Addend;
assert(StepNumerator != 0 && "Invalid step");
bool Negate = false;
int64_t SplatStepVal = StepNumerator;
unsigned StepOpcode = ISD::MUL;
// Exclude INT64_MIN to avoid passing it to std::abs. We won't optimize it
// anyway as the shift of 63 won't fit in uimm5.
if (StepNumerator != 1 && StepNumerator != INT64_MIN &&
isPowerOf2_64(std::abs(StepNumerator))) {
Negate = StepNumerator < 0;
StepOpcode = ISD::SHL;
SplatStepVal = Log2_64(std::abs(StepNumerator));
}
// Only emit VIDs with suitably-small steps/addends. We use imm5 is a
// threshold since it's the immediate value many RVV instructions accept.
// There is no vmul.vi instruction so ensure multiply constant can fit in
// a single addi instruction.
if (((StepOpcode == ISD::MUL && isInt<12>(SplatStepVal)) ||
(StepOpcode == ISD::SHL && isUInt<5>(SplatStepVal))) &&
isPowerOf2_32(StepDenominator) &&
(SplatStepVal >= 0 || StepDenominator == 1) && isInt<5>(Addend)) {
MVT VIDVT =
VT.isFloatingPoint() ? VT.changeVectorElementTypeToInteger() : VT;
MVT VIDContainerVT =
getContainerForFixedLengthVector(DAG, VIDVT, Subtarget);
SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, VIDContainerVT, Mask, VL);
// Convert right out of the scalable type so we can use standard ISD
// nodes for the rest of the computation. If we used scalable types with
// these, we'd lose the fixed-length vector info and generate worse
// vsetvli code.
VID = convertFromScalableVector(VIDVT, VID, DAG, Subtarget);
if ((StepOpcode == ISD::MUL && SplatStepVal != 1) ||
(StepOpcode == ISD::SHL && SplatStepVal != 0)) {
SDValue SplatStep = DAG.getConstant(SplatStepVal, DL, VIDVT);
VID = DAG.getNode(StepOpcode, DL, VIDVT, VID, SplatStep);
}
if (StepDenominator != 1) {
SDValue SplatStep =
DAG.getConstant(Log2_64(StepDenominator), DL, VIDVT);
VID = DAG.getNode(ISD::SRL, DL, VIDVT, VID, SplatStep);
}
if (Addend != 0 || Negate) {
SDValue SplatAddend = DAG.getConstant(Addend, DL, VIDVT);
VID = DAG.getNode(Negate ? ISD::SUB : ISD::ADD, DL, VIDVT, SplatAddend,
VID);
}
if (VT.isFloatingPoint()) {
// TODO: Use vfwcvt to reduce register pressure.
VID = DAG.getNode(ISD::SINT_TO_FP, DL, VT, VID);
}
return VID;
}
}
// For very small build_vectors, use a single scalar insert of a constant.
// TODO: Base this on constant rematerialization cost, not size.
const unsigned EltBitSize = VT.getScalarSizeInBits();
if (VT.getSizeInBits() <= 32 &&
ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
MVT ViaIntVT = MVT::getIntegerVT(VT.getSizeInBits());
assert((ViaIntVT == MVT::i16 || ViaIntVT == MVT::i32) &&
"Unexpected sequence type");
// If we can use the original VL with the modified element type, this
// means we only have a VTYPE toggle, not a VL toggle. TODO: Should this
// be moved into InsertVSETVLI?
unsigned ViaVecLen =
(Subtarget.getRealMinVLen() >= VT.getSizeInBits() * NumElts) ? NumElts : 1;
MVT ViaVecVT = MVT::getVectorVT(ViaIntVT, ViaVecLen);
uint64_t EltMask = maskTrailingOnes<uint64_t>(EltBitSize);
uint64_t SplatValue = 0;
// Construct the amalgamated value at this larger vector type.
for (const auto &OpIdx : enumerate(Op->op_values())) {
const auto &SeqV = OpIdx.value();
if (!SeqV.isUndef())
SplatValue |=
((SeqV->getAsZExtVal() & EltMask) << (OpIdx.index() * EltBitSize));
}
// On RV64, sign-extend from 32 to 64 bits where possible in order to
// achieve better constant materializion.
if (Subtarget.is64Bit() && ViaIntVT == MVT::i32)
SplatValue = SignExtend64<32>(SplatValue);
SDValue Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ViaVecVT,
DAG.getUNDEF(ViaVecVT),
DAG.getConstant(SplatValue, DL, XLenVT),
DAG.getVectorIdxConstant(0, DL));
if (ViaVecLen != 1)
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
MVT::getVectorVT(ViaIntVT, 1), Vec,
DAG.getConstant(0, DL, XLenVT));
return DAG.getBitcast(VT, Vec);
}
// Attempt to detect "hidden" splats, which only reveal themselves as splats
// when re-interpreted as a vector with a larger element type. For example,
// v4i16 = build_vector i16 0, i16 1, i16 0, i16 1
// could be instead splat as
// v2i32 = build_vector i32 0x00010000, i32 0x00010000
// TODO: This optimization could also work on non-constant splats, but it
// would require bit-manipulation instructions to construct the splat value.
SmallVector<SDValue> Sequence;
const auto *BV = cast<BuildVectorSDNode>(Op);
if (VT.isInteger() && EltBitSize < Subtarget.getELen() &&
ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
BV->getRepeatedSequence(Sequence) &&
(Sequence.size() * EltBitSize) <= Subtarget.getELen()) {
unsigned SeqLen = Sequence.size();
MVT ViaIntVT = MVT::getIntegerVT(EltBitSize * SeqLen);
assert((ViaIntVT == MVT::i16 || ViaIntVT == MVT::i32 ||
ViaIntVT == MVT::i64) &&
"Unexpected sequence type");
// If we can use the original VL with the modified element type, this
// means we only have a VTYPE toggle, not a VL toggle. TODO: Should this
// be moved into InsertVSETVLI?
const unsigned RequiredVL = NumElts / SeqLen;
const unsigned ViaVecLen =
(Subtarget.getRealMinVLen() >= ViaIntVT.getSizeInBits() * NumElts) ?
NumElts : RequiredVL;
MVT ViaVecVT = MVT::getVectorVT(ViaIntVT, ViaVecLen);
unsigned EltIdx = 0;
uint64_t EltMask = maskTrailingOnes<uint64_t>(EltBitSize);
uint64_t SplatValue = 0;
// Construct the amalgamated value which can be splatted as this larger
// vector type.
for (const auto &SeqV : Sequence) {
if (!SeqV.isUndef())
SplatValue |=
((SeqV->getAsZExtVal() & EltMask) << (EltIdx * EltBitSize));
EltIdx++;
}
// On RV64, sign-extend from 32 to 64 bits where possible in order to
// achieve better constant materializion.
if (Subtarget.is64Bit() && ViaIntVT == MVT::i32)
SplatValue = SignExtend64<32>(SplatValue);
// Since we can't introduce illegal i64 types at this stage, we can only
// perform an i64 splat on RV32 if it is its own sign-extended value. That
// way we can use RVV instructions to splat.
assert((ViaIntVT.bitsLE(XLenVT) ||
(!Subtarget.is64Bit() && ViaIntVT == MVT::i64)) &&
"Unexpected bitcast sequence");
if (ViaIntVT.bitsLE(XLenVT) || isInt<32>(SplatValue)) {
SDValue ViaVL =
DAG.getConstant(ViaVecVT.getVectorNumElements(), DL, XLenVT);
MVT ViaContainerVT =
getContainerForFixedLengthVector(DAG, ViaVecVT, Subtarget);
SDValue Splat =
DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ViaContainerVT,
DAG.getUNDEF(ViaContainerVT),
DAG.getConstant(SplatValue, DL, XLenVT), ViaVL);
Splat = convertFromScalableVector(ViaVecVT, Splat, DAG, Subtarget);
if (ViaVecLen != RequiredVL)
Splat = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
MVT::getVectorVT(ViaIntVT, RequiredVL), Splat,
DAG.getConstant(0, DL, XLenVT));
return DAG.getBitcast(VT, Splat);
}
}
// If the number of signbits allows, see if we can lower as a <N x i8>.
// Our main goal here is to reduce LMUL (and thus work) required to
// build the constant, but we will also narrow if the resulting
// narrow vector is known to materialize cheaply.
// TODO: We really should be costing the smaller vector. There are
// profitable cases this misses.
if (EltBitSize > 8 && VT.isInteger() &&
(NumElts <= 4 || VT.getSizeInBits() > Subtarget.getRealMinVLen())) {
unsigned SignBits = DAG.ComputeNumSignBits(Op);
if (EltBitSize - SignBits < 8) {
SDValue Source = DAG.getBuildVector(VT.changeVectorElementType(MVT::i8),
DL, Op->ops());
Source = convertToScalableVector(ContainerVT.changeVectorElementType(MVT::i8),
Source, DAG, Subtarget);
SDValue Res = DAG.getNode(RISCVISD::VSEXT_VL, DL, ContainerVT, Source, Mask, VL);
return convertFromScalableVector(VT, Res, DAG, Subtarget);
}
}
if (SDValue Res = lowerBuildVectorViaDominantValues(Op, DAG, Subtarget))
return Res;
// For constant vectors, use generic constant pool lowering. Otherwise,
// we'd have to materialize constants in GPRs just to move them into the
// vector.
return SDValue();
}
static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
assert(VT.isFixedLengthVector() && "Unexpected vector!");
if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()))
return lowerBuildVectorOfConstants(Op, DAG, Subtarget);
MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
MVT XLenVT = Subtarget.getXLenVT();
if (VT.getVectorElementType() == MVT::i1) {
// A BUILD_VECTOR can be lowered as a SETCC. For each fixed-length mask
// vector type, we have a legal equivalently-sized i8 type, so we can use
// that.
MVT WideVecVT = VT.changeVectorElementType(MVT::i8);
SDValue VecZero = DAG.getConstant(0, DL, WideVecVT);
SDValue WideVec;
if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
// For a splat, perform a scalar truncate before creating the wider
// vector.
Splat = DAG.getNode(ISD::AND, DL, Splat.getValueType(), Splat,
DAG.getConstant(1, DL, Splat.getValueType()));
WideVec = DAG.getSplatBuildVector(WideVecVT, DL, Splat);
} else {
SmallVector<SDValue, 8> Ops(Op->op_values());
WideVec = DAG.getBuildVector(WideVecVT, DL, Ops);
SDValue VecOne = DAG.getConstant(1, DL, WideVecVT);
WideVec = DAG.getNode(ISD::AND, DL, WideVecVT, WideVec, VecOne);
}
return DAG.getSetCC(DL, VT, WideVec, VecZero, ISD::SETNE);
}
if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
if (auto Gather = matchSplatAsGather(Splat, VT, DL, DAG, Subtarget))
return Gather;
unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL
: RISCVISD::VMV_V_X_VL;
if (!VT.isFloatingPoint())
Splat = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Splat);
Splat =
DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Splat, VL);
return convertFromScalableVector(VT, Splat, DAG, Subtarget);
}
if (SDValue Res = lowerBuildVectorViaDominantValues(Op, DAG, Subtarget))
return Res;
// If we're compiling for an exact VLEN value, we can split our work per
// register in the register group.
if (const auto VLen = Subtarget.getRealVLen();
VLen && VT.getSizeInBits().getKnownMinValue() > *VLen) {
MVT ElemVT = VT.getVectorElementType();
unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits();
EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
MVT OneRegVT = MVT::getVectorVT(ElemVT, ElemsPerVReg);
MVT M1VT = getContainerForFixedLengthVector(DAG, OneRegVT, Subtarget);
assert(M1VT == getLMUL1VT(M1VT));
// The following semantically builds up a fixed length concat_vector
// of the component build_vectors. We eagerly lower to scalable and
// insert_subvector here to avoid DAG combining it back to a large
// build_vector.
SmallVector<SDValue> BuildVectorOps(Op->op_begin(), Op->op_end());
unsigned NumOpElts = M1VT.getVectorMinNumElements();
SDValue Vec = DAG.getUNDEF(ContainerVT);
for (unsigned i = 0; i < VT.getVectorNumElements(); i += ElemsPerVReg) {
auto OneVRegOfOps = ArrayRef(BuildVectorOps).slice(i, ElemsPerVReg);
SDValue SubBV =
DAG.getNode(ISD::BUILD_VECTOR, DL, OneRegVT, OneVRegOfOps);
SubBV = convertToScalableVector(M1VT, SubBV, DAG, Subtarget);
unsigned InsertIdx = (i / ElemsPerVReg) * NumOpElts;
Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT, Vec, SubBV,
DAG.getVectorIdxConstant(InsertIdx, DL));
}
return convertFromScalableVector(VT, Vec, DAG, Subtarget);
}
// For m1 vectors, if we have non-undef values in both halves of our vector,
// split the vector into low and high halves, build them separately, then
// use a vselect to combine them. For long vectors, this cuts the critical
// path of the vslide1down sequence in half, and gives us an opportunity
// to special case each half independently. Note that we don't change the
// length of the sub-vectors here, so if both fallback to the generic
// vslide1down path, we should be able to fold the vselect into the final
// vslidedown (for the undef tail) for the first half w/ masking.
unsigned NumElts = VT.getVectorNumElements();
unsigned NumUndefElts =
count_if(Op->op_values(), [](const SDValue &V) { return V.isUndef(); });
unsigned NumDefElts = NumElts - NumUndefElts;
if (NumDefElts >= 8 && NumDefElts > NumElts / 2 &&
ContainerVT.bitsLE(getLMUL1VT(ContainerVT))) {
SmallVector<SDValue> SubVecAOps, SubVecBOps;
SmallVector<SDValue> MaskVals;
SDValue UndefElem = DAG.getUNDEF(Op->getOperand(0)->getValueType(0));
SubVecAOps.reserve(NumElts);
SubVecBOps.reserve(NumElts);
for (unsigned i = 0; i < NumElts; i++) {
SDValue Elem = Op->getOperand(i);
if (i < NumElts / 2) {
SubVecAOps.push_back(Elem);
SubVecBOps.push_back(UndefElem);
} else {
SubVecAOps.push_back(UndefElem);
SubVecBOps.push_back(Elem);
}
bool SelectMaskVal = (i < NumElts / 2);
MaskVals.push_back(DAG.getConstant(SelectMaskVal, DL, XLenVT));
}
assert(SubVecAOps.size() == NumElts && SubVecBOps.size() == NumElts &&
MaskVals.size() == NumElts);
SDValue SubVecA = DAG.getBuildVector(VT, DL, SubVecAOps);
SDValue SubVecB = DAG.getBuildVector(VT, DL, SubVecBOps);
MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
SDValue SelectMask = DAG.getBuildVector(MaskVT, DL, MaskVals);
return DAG.getNode(ISD::VSELECT, DL, VT, SelectMask, SubVecA, SubVecB);
}
// Cap the cost at a value linear to the number of elements in the vector.
// The default lowering is to use the stack. The vector store + scalar loads
// is linear in VL. However, at high lmuls vslide1down and vslidedown end up
// being (at least) linear in LMUL. As a result, using the vslidedown
// lowering for every element ends up being VL*LMUL..
// TODO: Should we be directly costing the stack alternative? Doing so might
// give us a more accurate upper bound.
InstructionCost LinearBudget = VT.getVectorNumElements() * 2;
// TODO: unify with TTI getSlideCost.
InstructionCost PerSlideCost = 1;
switch (RISCVTargetLowering::getLMUL(ContainerVT)) {
default: break;
case RISCVII::VLMUL::LMUL_2:
PerSlideCost = 2;
break;
case RISCVII::VLMUL::LMUL_4:
PerSlideCost = 4;
break;
case RISCVII::VLMUL::LMUL_8:
PerSlideCost = 8;
break;
}
// TODO: Should we be using the build instseq then cost + evaluate scheme
// we use for integer constants here?
unsigned UndefCount = 0;
for (const SDValue &V : Op->ops()) {
if (V.isUndef()) {
UndefCount++;
continue;
}
if (UndefCount) {
LinearBudget -= PerSlideCost;
UndefCount = 0;
}
LinearBudget -= PerSlideCost;
}
if (UndefCount) {
LinearBudget -= PerSlideCost;
}
if (LinearBudget < 0)
return SDValue();
assert((!VT.isFloatingPoint() ||
VT.getVectorElementType().getSizeInBits() <= Subtarget.getFLen()) &&
"Illegal type which will result in reserved encoding");
const unsigned Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
SDValue Vec;
UndefCount = 0;
for (SDValue V : Op->ops()) {
if (V.isUndef()) {
UndefCount++;
continue;
}
// Start our sequence with a TA splat in the hopes that hardware is able to
// recognize there's no dependency on the prior value of our temporary
// register.
if (!Vec) {
Vec = DAG.getSplatVector(VT, DL, V);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
UndefCount = 0;
continue;
}
if (UndefCount) {
const SDValue Offset = DAG.getConstant(UndefCount, DL, Subtarget.getXLenVT());
Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
Vec, Offset, Mask, VL, Policy);
UndefCount = 0;
}
auto OpCode =
VT.isFloatingPoint() ? RISCVISD::VFSLIDE1DOWN_VL : RISCVISD::VSLIDE1DOWN_VL;
if (!VT.isFloatingPoint())
V = DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getXLenVT(), V);
Vec = DAG.getNode(OpCode, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Vec,
V, Mask, VL);
}
if (UndefCount) {
const SDValue Offset = DAG.getConstant(UndefCount, DL, Subtarget.getXLenVT());
Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
Vec, Offset, Mask, VL, Policy);
}
return convertFromScalableVector(VT, Vec, DAG, Subtarget);
}
static SDValue splatPartsI64WithVL(const SDLoc &DL, MVT VT, SDValue Passthru,
SDValue Lo, SDValue Hi, SDValue VL,
SelectionDAG &DAG) {
if (!Passthru)
Passthru = DAG.getUNDEF(VT);
if (isa<ConstantSDNode>(Lo) && isa<ConstantSDNode>(Hi)) {
int32_t LoC = cast<ConstantSDNode>(Lo)->getSExtValue();
int32_t HiC = cast<ConstantSDNode>(Hi)->getSExtValue();
// If Hi constant is all the same sign bit as Lo, lower this as a custom
// node in order to try and match RVV vector/scalar instructions.
if ((LoC >> 31) == HiC)
return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Lo, VL);
// If vl is equal to VLMAX or fits in 4 bits and Hi constant is equal to Lo,
// we could use vmv.v.x whose EEW = 32 to lower it. This allows us to use
// vlmax vsetvli or vsetivli to change the VL.
// FIXME: Support larger constants?
// FIXME: Support non-constant VLs by saturating?
if (LoC == HiC) {
SDValue NewVL;
if (isAllOnesConstant(VL) ||
(isa<RegisterSDNode>(VL) &&
cast<RegisterSDNode>(VL)->getReg() == RISCV::X0))
NewVL = DAG.getRegister(RISCV::X0, MVT::i32);
else if (isa<ConstantSDNode>(VL) && isUInt<4>(VL->getAsZExtVal()))
NewVL = DAG.getNode(ISD::ADD, DL, VL.getValueType(), VL, VL);
if (NewVL) {
MVT InterVT =
MVT::getVectorVT(MVT::i32, VT.getVectorElementCount() * 2);
auto InterVec = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, InterVT,
DAG.getUNDEF(InterVT), Lo, NewVL);
return DAG.getNode(ISD::BITCAST, DL, VT, InterVec);
}
}
}
// Detect cases where Hi is (SRA Lo, 31) which means Hi is Lo sign extended.
if (Hi.getOpcode() == ISD::SRA && Hi.getOperand(0) == Lo &&
isa<ConstantSDNode>(Hi.getOperand(1)) &&
Hi.getConstantOperandVal(1) == 31)
return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Lo, VL);
// If the hi bits of the splat are undefined, then it's fine to just splat Lo
// even if it might be sign extended.
if (Hi.isUndef())
return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Lo, VL);
// Fall back to a stack store and stride x0 vector load.
return DAG.getNode(RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL, DL, VT, Passthru, Lo,
Hi, VL);
}
// Called by type legalization to handle splat of i64 on RV32.
// FIXME: We can optimize this when the type has sign or zero bits in one
// of the halves.
static SDValue splatSplitI64WithVL(const SDLoc &DL, MVT VT, SDValue Passthru,
SDValue Scalar, SDValue VL,
SelectionDAG &DAG) {
assert(Scalar.getValueType() == MVT::i64 && "Unexpected VT!");
SDValue Lo, Hi;
std::tie(Lo, Hi) = DAG.SplitScalar(Scalar, DL, MVT::i32, MVT::i32);
return splatPartsI64WithVL(DL, VT, Passthru, Lo, Hi, VL, DAG);
}
// This function lowers a splat of a scalar operand Splat with the vector
// length VL. It ensures the final sequence is type legal, which is useful when
// lowering a splat after type legalization.
static SDValue lowerScalarSplat(SDValue Passthru, SDValue Scalar, SDValue VL,
MVT VT, const SDLoc &DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
bool HasPassthru = Passthru && !Passthru.isUndef();
if (!HasPassthru && !Passthru)
Passthru = DAG.getUNDEF(VT);
if (VT.isFloatingPoint())
return DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, VT, Passthru, Scalar, VL);
MVT XLenVT = Subtarget.getXLenVT();
// Simplest case is that the operand needs to be promoted to XLenVT.
if (Scalar.getValueType().bitsLE(XLenVT)) {
// If the operand is a constant, sign extend to increase our chances
// of being able to use a .vi instruction. ANY_EXTEND would become a
// a zero extend and the simm5 check in isel would fail.
// FIXME: Should we ignore the upper bits in isel instead?
unsigned ExtOpc =
isa<ConstantSDNode>(Scalar) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
Scalar = DAG.getNode(ExtOpc, DL, XLenVT, Scalar);
return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Scalar, VL);
}
assert(XLenVT == MVT::i32 && Scalar.getValueType() == MVT::i64 &&
"Unexpected scalar for splat lowering!");
if (isOneConstant(VL) && isNullConstant(Scalar))
return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT, Passthru,
DAG.getConstant(0, DL, XLenVT), VL);
// Otherwise use the more complicated splatting algorithm.
return splatSplitI64WithVL(DL, VT, Passthru, Scalar, VL, DAG);
}
// This function lowers an insert of a scalar operand Scalar into lane
// 0 of the vector regardless of the value of VL. The contents of the
// remaining lanes of the result vector are unspecified. VL is assumed
// to be non-zero.
static SDValue lowerScalarInsert(SDValue Scalar, SDValue VL, MVT VT,
const SDLoc &DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(VT.isScalableVector() && "Expect VT is scalable vector type.");
const MVT XLenVT = Subtarget.getXLenVT();
SDValue Passthru = DAG.getUNDEF(VT);
if (Scalar.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
isNullConstant(Scalar.getOperand(1))) {
SDValue ExtractedVal = Scalar.getOperand(0);
// The element types must be the same.
if (ExtractedVal.getValueType().getVectorElementType() ==
VT.getVectorElementType()) {
MVT ExtractedVT = ExtractedVal.getSimpleValueType();
MVT ExtractedContainerVT = ExtractedVT;
if (ExtractedContainerVT.isFixedLengthVector()) {
ExtractedContainerVT = getContainerForFixedLengthVector(
DAG, ExtractedContainerVT, Subtarget);
ExtractedVal = convertToScalableVector(ExtractedContainerVT,
ExtractedVal, DAG, Subtarget);
}
if (ExtractedContainerVT.bitsLE(VT))
return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Passthru,
ExtractedVal, DAG.getVectorIdxConstant(0, DL));
return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtractedVal,
DAG.getVectorIdxConstant(0, DL));
}
}
if (VT.isFloatingPoint())
return DAG.getNode(RISCVISD::VFMV_S_F_VL, DL, VT,
DAG.getUNDEF(VT), Scalar, VL);
// Avoid the tricky legalization cases by falling back to using the
// splat code which already handles it gracefully.
if (!Scalar.getValueType().bitsLE(XLenVT))
return lowerScalarSplat(DAG.getUNDEF(VT), Scalar,
DAG.getConstant(1, DL, XLenVT),
VT, DL, DAG, Subtarget);
// If the operand is a constant, sign extend to increase our chances
// of being able to use a .vi instruction. ANY_EXTEND would become a
// a zero extend and the simm5 check in isel would fail.
// FIXME: Should we ignore the upper bits in isel instead?
unsigned ExtOpc =
isa<ConstantSDNode>(Scalar) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
Scalar = DAG.getNode(ExtOpc, DL, XLenVT, Scalar);
return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT,
DAG.getUNDEF(VT), Scalar, VL);
}
// Is this a shuffle extracts either the even or odd elements of a vector?
// That is, specifically, either (a) or (b) below.
// t34: v8i8 = extract_subvector t11, Constant:i64<0>
// t33: v8i8 = extract_subvector t11, Constant:i64<8>
// a) t35: v8i8 = vector_shuffle<0,2,4,6,8,10,12,14> t34, t33
// b) t35: v8i8 = vector_shuffle<1,3,5,7,9,11,13,15> t34, t33
// Returns {Src Vector, Even Elements} om success
static bool isDeinterleaveShuffle(MVT VT, MVT ContainerVT, SDValue V1,
SDValue V2, ArrayRef<int> Mask,
const RISCVSubtarget &Subtarget) {
// Need to be able to widen the vector.
if (VT.getScalarSizeInBits() >= Subtarget.getELen())
return false;
// Both input must be extracts.
if (V1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
V2.getOpcode() != ISD::EXTRACT_SUBVECTOR)
return false;
// Extracting from the same source.
SDValue Src = V1.getOperand(0);
if (Src != V2.getOperand(0))
return false;
// Src needs to have twice the number of elements.
if (Src.getValueType().getVectorNumElements() != (Mask.size() * 2))
return false;
// The extracts must extract the two halves of the source.
if (V1.getConstantOperandVal(1) != 0 ||
V2.getConstantOperandVal(1) != Mask.size())
return false;
// First index must be the first even or odd element from V1.
if (Mask[0] != 0 && Mask[0] != 1)
return false;
// The others must increase by 2 each time.
// TODO: Support undef elements?
for (unsigned i = 1; i != Mask.size(); ++i)
if (Mask[i] != Mask[i - 1] + 2)
return false;
return true;
}
/// Is this shuffle interleaving contiguous elements from one vector into the
/// even elements and contiguous elements from another vector into the odd
/// elements. \p EvenSrc will contain the element that should be in the first
/// even element. \p OddSrc will contain the element that should be in the first
/// odd element. These can be the first element in a source or the element half
/// way through the source.
static bool isInterleaveShuffle(ArrayRef<int> Mask, MVT VT, int &EvenSrc,
int &OddSrc, const RISCVSubtarget &Subtarget) {
// We need to be able to widen elements to the next larger integer type.
if (VT.getScalarSizeInBits() >= Subtarget.getELen())
return false;
int Size = Mask.size();
int NumElts = VT.getVectorNumElements();
assert(Size == (int)NumElts && "Unexpected mask size");
SmallVector<unsigned, 2> StartIndexes;
if (!ShuffleVectorInst::isInterleaveMask(Mask, 2, Size * 2, StartIndexes))
return false;
EvenSrc = StartIndexes[0];
OddSrc = StartIndexes[1];
// One source should be low half of first vector.
if (EvenSrc != 0 && OddSrc != 0)
return false;
// Subvectors will be subtracted from either at the start of the two input
// vectors, or at the start and middle of the first vector if it's an unary
// interleave.
// In both cases, HalfNumElts will be extracted.
// We need to ensure that the extract indices are 0 or HalfNumElts otherwise
// we'll create an illegal extract_subvector.
// FIXME: We could support other values using a slidedown first.
int HalfNumElts = NumElts / 2;
return ((EvenSrc % HalfNumElts) == 0) && ((OddSrc % HalfNumElts) == 0);
}
/// Match shuffles that concatenate two vectors, rotate the concatenation,
/// and then extract the original number of elements from the rotated result.
/// This is equivalent to vector.splice or X86's PALIGNR instruction. The
/// returned rotation amount is for a rotate right, where elements move from
/// higher elements to lower elements. \p LoSrc indicates the first source
/// vector of the rotate or -1 for undef. \p HiSrc indicates the second vector
/// of the rotate or -1 for undef. At least one of \p LoSrc and \p HiSrc will be
/// 0 or 1 if a rotation is found.
///
/// NOTE: We talk about rotate to the right which matches how bit shift and
/// rotate instructions are described where LSBs are on the right, but LLVM IR
/// and the table below write vectors with the lowest elements on the left.
static int isElementRotate(int &LoSrc, int &HiSrc, ArrayRef<int> Mask) {
int Size = Mask.size();
// We need to detect various ways of spelling a rotation:
// [11, 12, 13, 14, 15, 0, 1, 2]
// [-1, 12, 13, 14, -1, -1, 1, -1]
// [-1, -1, -1, -1, -1, -1, 1, 2]
// [ 3, 4, 5, 6, 7, 8, 9, 10]
// [-1, 4, 5, 6, -1, -1, 9, -1]
// [-1, 4, 5, 6, -1, -1, -1, -1]
int Rotation = 0;
LoSrc = -1;
HiSrc = -1;
for (int i = 0; i != Size; ++i) {
int M = Mask[i];
if (M < 0)
continue;
// Determine where a rotate vector would have started.
int StartIdx = i - (M % Size);
// The identity rotation isn't interesting, stop.
if (StartIdx == 0)
return -1;
// If we found the tail of a vector the rotation must be the missing
// front. If we found the head of a vector, it must be how much of the
// head.
int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx;
if (Rotation == 0)
Rotation = CandidateRotation;
else if (Rotation != CandidateRotation)
// The rotations don't match, so we can't match this mask.
return -1;
// Compute which value this mask is pointing at.
int MaskSrc = M < Size ? 0 : 1;
// Compute which of the two target values this index should be assigned to.
// This reflects whether the high elements are remaining or the low elemnts
// are remaining.
int &TargetSrc = StartIdx < 0 ? HiSrc : LoSrc;
// Either set up this value if we've not encountered it before, or check
// that it remains consistent.
if (TargetSrc < 0)
TargetSrc = MaskSrc;
else if (TargetSrc != MaskSrc)
// This may be a rotation, but it pulls from the inputs in some
// unsupported interleaving.
return -1;
}
// Check that we successfully analyzed the mask, and normalize the results.
assert(Rotation != 0 && "Failed to locate a viable rotation!");
assert((LoSrc >= 0 || HiSrc >= 0) &&
"Failed to find a rotated input vector!");
return Rotation;
}
// Lower a deinterleave shuffle to vnsrl.
// [a, p, b, q, c, r, d, s] -> [a, b, c, d] (EvenElts == true)
// -> [p, q, r, s] (EvenElts == false)
// VT is the type of the vector to return, <[vscale x ]n x ty>
// Src is the vector to deinterleave of type <[vscale x ]n*2 x ty>
static SDValue getDeinterleaveViaVNSRL(const SDLoc &DL, MVT VT, SDValue Src,
bool EvenElts,
const RISCVSubtarget &Subtarget,
SelectionDAG &DAG) {
// The result is a vector of type <m x n x ty>
MVT ContainerVT = VT;
// Convert fixed vectors to scalable if needed
if (ContainerVT.isFixedLengthVector()) {
assert(Src.getSimpleValueType().isFixedLengthVector());
ContainerVT = getContainerForFixedLengthVector(DAG, ContainerVT, Subtarget);
// The source is a vector of type <m x n*2 x ty>
MVT SrcContainerVT =
MVT::getVectorVT(ContainerVT.getVectorElementType(),
ContainerVT.getVectorElementCount() * 2);
Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
}
auto [TrueMask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
// Bitcast the source vector from <m x n*2 x ty> -> <m x n x ty*2>
// This also converts FP to int.
unsigned EltBits = ContainerVT.getScalarSizeInBits();
MVT WideSrcContainerVT = MVT::getVectorVT(
MVT::getIntegerVT(EltBits * 2), ContainerVT.getVectorElementCount());
Src = DAG.getBitcast(WideSrcContainerVT, Src);
// The integer version of the container type.
MVT IntContainerVT = ContainerVT.changeVectorElementTypeToInteger();
// If we want even elements, then the shift amount is 0. Otherwise, shift by
// the original element size.
unsigned Shift = EvenElts ? 0 : EltBits;
SDValue SplatShift = DAG.getNode(
RISCVISD::VMV_V_X_VL, DL, IntContainerVT, DAG.getUNDEF(ContainerVT),
DAG.getConstant(Shift, DL, Subtarget.getXLenVT()), VL);
SDValue Res =
DAG.getNode(RISCVISD::VNSRL_VL, DL, IntContainerVT, Src, SplatShift,
DAG.getUNDEF(IntContainerVT), TrueMask, VL);
// Cast back to FP if needed.
Res = DAG.getBitcast(ContainerVT, Res);
if (VT.isFixedLengthVector())
Res = convertFromScalableVector(VT, Res, DAG, Subtarget);
return Res;
}
// Lower the following shuffle to vslidedown.
// a)
// t49: v8i8 = extract_subvector t13, Constant:i64<0>
// t109: v8i8 = extract_subvector t13, Constant:i64<8>
// t108: v8i8 = vector_shuffle<1,2,3,4,5,6,7,8> t49, t106
// b)
// t69: v16i16 = extract_subvector t68, Constant:i64<0>
// t23: v8i16 = extract_subvector t69, Constant:i64<0>
// t29: v4i16 = extract_subvector t23, Constant:i64<4>
// t26: v8i16 = extract_subvector t69, Constant:i64<8>
// t30: v4i16 = extract_subvector t26, Constant:i64<0>
// t54: v4i16 = vector_shuffle<1,2,3,4> t29, t30
static SDValue lowerVECTOR_SHUFFLEAsVSlidedown(const SDLoc &DL, MVT VT,
SDValue V1, SDValue V2,
ArrayRef<int> Mask,
const RISCVSubtarget &Subtarget,
SelectionDAG &DAG) {
auto findNonEXTRACT_SUBVECTORParent =
[](SDValue Parent) -> std::pair<SDValue, uint64_t> {
uint64_t Offset = 0;
while (Parent.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
// EXTRACT_SUBVECTOR can be used to extract a fixed-width vector from
// a scalable vector. But we don't want to match the case.
Parent.getOperand(0).getSimpleValueType().isFixedLengthVector()) {
Offset += Parent.getConstantOperandVal(1);
Parent = Parent.getOperand(0);
}
return std::make_pair(Parent, Offset);
};
auto [V1Src, V1IndexOffset] = findNonEXTRACT_SUBVECTORParent(V1);
auto [V2Src, V2IndexOffset] = findNonEXTRACT_SUBVECTORParent(V2);
// Extracting from the same source.
SDValue Src = V1Src;
if (Src != V2Src)
return SDValue();
// Rebuild mask because Src may be from multiple EXTRACT_SUBVECTORs.
SmallVector<int, 16> NewMask(Mask);
for (size_t i = 0; i != NewMask.size(); ++i) {
if (NewMask[i] == -1)
continue;
if (static_cast<size_t>(NewMask[i]) < NewMask.size()) {
NewMask[i] = NewMask[i] + V1IndexOffset;
} else {
// Minus NewMask.size() is needed. Otherwise, the b case would be
// <5,6,7,12> instead of <5,6,7,8>.
NewMask[i] = NewMask[i] - NewMask.size() + V2IndexOffset;
}
}
// First index must be known and non-zero. It will be used as the slidedown
// amount.
if (NewMask[0] <= 0)
return SDValue();
// NewMask is also continuous.
for (unsigned i = 1; i != NewMask.size(); ++i)
if (NewMask[i - 1] + 1 != NewMask[i])
return SDValue();
MVT XLenVT = Subtarget.getXLenVT();
MVT SrcVT = Src.getSimpleValueType();
MVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
auto [TrueMask, VL] = getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
SDValue Slidedown =
getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
convertToScalableVector(ContainerVT, Src, DAG, Subtarget),
DAG.getConstant(NewMask[0], DL, XLenVT), TrueMask, VL);
return DAG.getNode(
ISD::EXTRACT_SUBVECTOR, DL, VT,
convertFromScalableVector(SrcVT, Slidedown, DAG, Subtarget),
DAG.getConstant(0, DL, XLenVT));
}
// Because vslideup leaves the destination elements at the start intact, we can
// use it to perform shuffles that insert subvectors:
//
// vector_shuffle v8:v8i8, v9:v8i8, <0, 1, 2, 3, 8, 9, 10, 11>
// ->
// vsetvli zero, 8, e8, mf2, ta, ma
// vslideup.vi v8, v9, 4
//
// vector_shuffle v8:v8i8, v9:v8i8 <0, 1, 8, 9, 10, 5, 6, 7>
// ->
// vsetvli zero, 5, e8, mf2, tu, ma
// vslideup.v1 v8, v9, 2
static SDValue lowerVECTOR_SHUFFLEAsVSlideup(const SDLoc &DL, MVT VT,
SDValue V1, SDValue V2,
ArrayRef<int> Mask,
const RISCVSubtarget &Subtarget,
SelectionDAG &DAG) {
unsigned NumElts = VT.getVectorNumElements();
int NumSubElts, Index;
if (!ShuffleVectorInst::isInsertSubvectorMask(Mask, NumElts, NumSubElts,
Index))
return SDValue();
bool OpsSwapped = Mask[Index] < (int)NumElts;
SDValue InPlace = OpsSwapped ? V2 : V1;
SDValue ToInsert = OpsSwapped ? V1 : V2;
MVT XLenVT = Subtarget.getXLenVT();
MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
auto TrueMask = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).first;
// We slide up by the index that the subvector is being inserted at, and set
// VL to the index + the number of elements being inserted.
unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED | RISCVII::MASK_AGNOSTIC;
// If the we're adding a suffix to the in place vector, i.e. inserting right
// up to the very end of it, then we don't actually care about the tail.
if (NumSubElts + Index >= (int)NumElts)
Policy |= RISCVII::TAIL_AGNOSTIC;
InPlace = convertToScalableVector(ContainerVT, InPlace, DAG, Subtarget);
ToInsert = convertToScalableVector(ContainerVT, ToInsert, DAG, Subtarget);
SDValue VL = DAG.getConstant(NumSubElts + Index, DL, XLenVT);
SDValue Res;
// If we're inserting into the lowest elements, use a tail undisturbed
// vmv.v.v.
if (Index == 0)
Res = DAG.getNode(RISCVISD::VMV_V_V_VL, DL, ContainerVT, InPlace, ToInsert,
VL);
else
Res = getVSlideup(DAG, Subtarget, DL, ContainerVT, InPlace, ToInsert,
DAG.getConstant(Index, DL, XLenVT), TrueMask, VL, Policy);
return convertFromScalableVector(VT, Res, DAG, Subtarget);
}
/// Match v(f)slide1up/down idioms. These operations involve sliding
/// N-1 elements to make room for an inserted scalar at one end.
static SDValue lowerVECTOR_SHUFFLEAsVSlide1(const SDLoc &DL, MVT VT,
SDValue V1, SDValue V2,
ArrayRef<int> Mask,
const RISCVSubtarget &Subtarget,
SelectionDAG &DAG) {
bool OpsSwapped = false;
if (!isa<BuildVectorSDNode>(V1)) {
if (!isa<BuildVectorSDNode>(V2))
return SDValue();
std::swap(V1, V2);
OpsSwapped = true;
}
SDValue Splat = cast<BuildVectorSDNode>(V1)->getSplatValue();
if (!Splat)
return SDValue();
// Return true if the mask could describe a slide of Mask.size() - 1
// elements from concat_vector(V1, V2)[Base:] to [Offset:].
auto isSlideMask = [](ArrayRef<int> Mask, unsigned Base, int Offset) {
const unsigned S = (Offset > 0) ? 0 : -Offset;
const unsigned E = Mask.size() - ((Offset > 0) ? Offset : 0);
for (unsigned i = S; i != E; ++i)
if (Mask[i] >= 0 && (unsigned)Mask[i] != Base + i + Offset)
return false;
return true;
};
const unsigned NumElts = VT.getVectorNumElements();
bool IsVSlidedown = isSlideMask(Mask, OpsSwapped ? 0 : NumElts, 1);
if (!IsVSlidedown && !isSlideMask(Mask, OpsSwapped ? 0 : NumElts, -1))
return SDValue();
const int InsertIdx = Mask[IsVSlidedown ? (NumElts - 1) : 0];
// Inserted lane must come from splat, undef scalar is legal but not profitable.
if (InsertIdx < 0 || InsertIdx / NumElts != (unsigned)OpsSwapped)
return SDValue();
MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
auto [TrueMask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
auto OpCode = IsVSlidedown ?
(VT.isFloatingPoint() ? RISCVISD::VFSLIDE1DOWN_VL : RISCVISD::VSLIDE1DOWN_VL) :
(VT.isFloatingPoint() ? RISCVISD::VFSLIDE1UP_VL : RISCVISD::VSLIDE1UP_VL);
if (!VT.isFloatingPoint())
Splat = DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getXLenVT(), Splat);
auto Vec = DAG.getNode(OpCode, DL, ContainerVT,
DAG.getUNDEF(ContainerVT),
convertToScalableVector(ContainerVT, V2, DAG, Subtarget),
Splat, TrueMask, VL);
return convertFromScalableVector(VT, Vec, DAG, Subtarget);
}
// Given two input vectors of <[vscale x ]n x ty>, use vwaddu.vv and vwmaccu.vx
// to create an interleaved vector of <[vscale x] n*2 x ty>.
// This requires that the size of ty is less than the subtarget's maximum ELEN.
static SDValue getWideningInterleave(SDValue EvenV, SDValue OddV,
const SDLoc &DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VecVT = EvenV.getSimpleValueType();
MVT VecContainerVT = VecVT; // <vscale x n x ty>
// Convert fixed vectors to scalable if needed
if (VecContainerVT.isFixedLengthVector()) {
VecContainerVT = getContainerForFixedLengthVector(DAG, VecVT, Subtarget);
EvenV = convertToScalableVector(VecContainerVT, EvenV, DAG, Subtarget);
OddV = convertToScalableVector(VecContainerVT, OddV, DAG, Subtarget);
}
assert(VecVT.getScalarSizeInBits() < Subtarget.getELen());
// We're working with a vector of the same size as the resulting
// interleaved vector, but with half the number of elements and
// twice the SEW (Hence the restriction on not using the maximum
// ELEN)
MVT WideVT =
MVT::getVectorVT(MVT::getIntegerVT(VecVT.getScalarSizeInBits() * 2),
VecVT.getVectorElementCount());
MVT WideContainerVT = WideVT; // <vscale x n x ty*2>
if (WideContainerVT.isFixedLengthVector())
WideContainerVT = getContainerForFixedLengthVector(DAG, WideVT, Subtarget);
// Bitcast the input vectors to integers in case they are FP
VecContainerVT = VecContainerVT.changeTypeToInteger();
EvenV = DAG.getBitcast(VecContainerVT, EvenV);
OddV = DAG.getBitcast(VecContainerVT, OddV);
auto [Mask, VL] = getDefaultVLOps(VecVT, VecContainerVT, DL, DAG, Subtarget);
SDValue Passthru = DAG.getUNDEF(WideContainerVT);
SDValue Interleaved;
if (OddV.isUndef()) {
// If OddV is undef, this is a zero extend.
// FIXME: Not only does this optimize the code, it fixes some correctness
// issues because MIR does not have freeze.
Interleaved =
DAG.getNode(RISCVISD::VZEXT_VL, DL, WideContainerVT, EvenV, Mask, VL);
} else if (Subtarget.hasStdExtZvbb()) {
// Interleaved = (OddV << VecVT.getScalarSizeInBits()) + EvenV.
SDValue OffsetVec =
DAG.getConstant(VecVT.getScalarSizeInBits(), DL, VecContainerVT);
Interleaved = DAG.getNode(RISCVISD::VWSLL_VL, DL, WideContainerVT, OddV,
OffsetVec, Passthru, Mask, VL);
if (!EvenV.isUndef())
Interleaved = DAG.getNode(RISCVISD::VWADDU_W_VL, DL, WideContainerVT,
Interleaved, EvenV, Passthru, Mask, VL);
} else if (EvenV.isUndef()) {
Interleaved =
DAG.getNode(RISCVISD::VZEXT_VL, DL, WideContainerVT, OddV, Mask, VL);
SDValue OffsetVec =
DAG.getConstant(VecVT.getScalarSizeInBits(), DL, WideContainerVT);
Interleaved = DAG.getNode(RISCVISD::SHL_VL, DL, WideContainerVT,
Interleaved, OffsetVec, Passthru, Mask, VL);
} else {
// FIXME: We should freeze the odd vector here. We already handled the case
// of provably undef/poison above.
// Widen EvenV and OddV with 0s and add one copy of OddV to EvenV with
// vwaddu.vv
Interleaved = DAG.getNode(RISCVISD::VWADDU_VL, DL, WideContainerVT, EvenV,
OddV, Passthru, Mask, VL);
// Then get OddV * by 2^(VecVT.getScalarSizeInBits() - 1)
SDValue AllOnesVec = DAG.getSplatVector(
VecContainerVT, DL, DAG.getAllOnesConstant(DL, Subtarget.getXLenVT()));
SDValue OddsMul = DAG.getNode(RISCVISD::VWMULU_VL, DL, WideContainerVT,
OddV, AllOnesVec, Passthru, Mask, VL);
// Add the two together so we get
// (OddV * 0xff...ff) + (OddV + EvenV)
// = (OddV * 0x100...00) + EvenV
// = (OddV << VecVT.getScalarSizeInBits()) + EvenV
// Note the ADD_VL and VLMULU_VL should get selected as vwmaccu.vx
Interleaved = DAG.getNode(RISCVISD::ADD_VL, DL, WideContainerVT,
Interleaved, OddsMul, Passthru, Mask, VL);
}
// Bitcast from <vscale x n * ty*2> to <vscale x 2*n x ty>
MVT ResultContainerVT = MVT::getVectorVT(
VecVT.getVectorElementType(), // Make sure to use original type
VecContainerVT.getVectorElementCount().multiplyCoefficientBy(2));
Interleaved = DAG.getBitcast(ResultContainerVT, Interleaved);
// Convert back to a fixed vector if needed
MVT ResultVT =
MVT::getVectorVT(VecVT.getVectorElementType(),
VecVT.getVectorElementCount().multiplyCoefficientBy(2));
if (ResultVT.isFixedLengthVector())
Interleaved =
convertFromScalableVector(ResultVT, Interleaved, DAG, Subtarget);
return Interleaved;
}
// If we have a vector of bits that we want to reverse, we can use a vbrev on a
// larger element type, e.g. v32i1 can be reversed with a v1i32 bitreverse.
static SDValue lowerBitreverseShuffle(ShuffleVectorSDNode *SVN,
SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDLoc DL(SVN);
MVT VT = SVN->getSimpleValueType(0);
SDValue V = SVN->getOperand(0);
unsigned NumElts = VT.getVectorNumElements();
assert(VT.getVectorElementType() == MVT::i1);
if (!ShuffleVectorInst::isReverseMask(SVN->getMask(),
SVN->getMask().size()) ||
!SVN->getOperand(1).isUndef())
return SDValue();
unsigned ViaEltSize = std::max((uint64_t)8, PowerOf2Ceil(NumElts));
EVT ViaVT = EVT::getVectorVT(
*DAG.getContext(), EVT::getIntegerVT(*DAG.getContext(), ViaEltSize), 1);
EVT ViaBitVT =
EVT::getVectorVT(*DAG.getContext(), MVT::i1, ViaVT.getScalarSizeInBits());
// If we don't have zvbb or the larger element type > ELEN, the operation will
// be illegal.
if (!Subtarget.getTargetLowering()->isOperationLegalOrCustom(ISD::BITREVERSE,
ViaVT) ||
!Subtarget.getTargetLowering()->isTypeLegal(ViaBitVT))
return SDValue();
// If the bit vector doesn't fit exactly into the larger element type, we need
// to insert it into the larger vector and then shift up the reversed bits
// afterwards to get rid of the gap introduced.
if (ViaEltSize > NumElts)
V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ViaBitVT, DAG.getUNDEF(ViaBitVT),
V, DAG.getVectorIdxConstant(0, DL));
SDValue Res =
DAG.getNode(ISD::BITREVERSE, DL, ViaVT, DAG.getBitcast(ViaVT, V));
// Shift up the reversed bits if the vector didn't exactly fit into the larger
// element type.
if (ViaEltSize > NumElts)
Res = DAG.getNode(ISD::SRL, DL, ViaVT, Res,
DAG.getConstant(ViaEltSize - NumElts, DL, ViaVT));
Res = DAG.getBitcast(ViaBitVT, Res);
if (ViaEltSize > NumElts)
Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res,
DAG.getVectorIdxConstant(0, DL));
return Res;
}
static bool isLegalBitRotate(ShuffleVectorSDNode *SVN,
SelectionDAG &DAG,
const RISCVSubtarget &Subtarget,
MVT &RotateVT, unsigned &RotateAmt) {
SDLoc DL(SVN);
EVT VT = SVN->getValueType(0);
unsigned NumElts = VT.getVectorNumElements();
unsigned EltSizeInBits = VT.getScalarSizeInBits();
unsigned NumSubElts;
if (!ShuffleVectorInst::isBitRotateMask(SVN->getMask(), EltSizeInBits, 2,
NumElts, NumSubElts, RotateAmt))
return false;
RotateVT = MVT::getVectorVT(MVT::getIntegerVT(EltSizeInBits * NumSubElts),
NumElts / NumSubElts);
// We might have a RotateVT that isn't legal, e.g. v4i64 on zve32x.
return Subtarget.getTargetLowering()->isTypeLegal(RotateVT);
}
// Given a shuffle mask like <3, 0, 1, 2, 7, 4, 5, 6> for v8i8, we can
// reinterpret it as a v2i32 and rotate it right by 8 instead. We can lower this
// as a vror.vi if we have Zvkb, or otherwise as a vsll, vsrl and vor.
static SDValue lowerVECTOR_SHUFFLEAsRotate(ShuffleVectorSDNode *SVN,
SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDLoc DL(SVN);
EVT VT = SVN->getValueType(0);
unsigned RotateAmt;
MVT RotateVT;
if (!isLegalBitRotate(SVN, DAG, Subtarget, RotateVT, RotateAmt))
return SDValue();
SDValue Op = DAG.getBitcast(RotateVT, SVN->getOperand(0));
SDValue Rotate;
// A rotate of an i16 by 8 bits either direction is equivalent to a byteswap,
// so canonicalize to vrev8.
if (RotateVT.getScalarType() == MVT::i16 && RotateAmt == 8)
Rotate = DAG.getNode(ISD::BSWAP, DL, RotateVT, Op);
else
Rotate = DAG.getNode(ISD::ROTL, DL, RotateVT, Op,
DAG.getConstant(RotateAmt, DL, RotateVT));
return DAG.getBitcast(VT, Rotate);
}
// If compiling with an exactly known VLEN, see if we can split a
// shuffle on m2 or larger into a small number of m1 sized shuffles
// which write each destination registers exactly once.
static SDValue lowerShuffleViaVRegSplitting(ShuffleVectorSDNode *SVN,
SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDLoc DL(SVN);
MVT VT = SVN->getSimpleValueType(0);
SDValue V1 = SVN->getOperand(0);
SDValue V2 = SVN->getOperand(1);
ArrayRef<int> Mask = SVN->getMask();
unsigned NumElts = VT.getVectorNumElements();
// If we don't know exact data layout, not much we can do. If this
// is already m1 or smaller, no point in splitting further.
const auto VLen = Subtarget.getRealVLen();
if (!VLen || VT.getSizeInBits().getFixedValue() <= *VLen)
return SDValue();
// Avoid picking up bitrotate patterns which we have a linear-in-lmul
// expansion for.
unsigned RotateAmt;
MVT RotateVT;
if (isLegalBitRotate(SVN, DAG, Subtarget, RotateVT, RotateAmt))
return SDValue();
MVT ElemVT = VT.getVectorElementType();
unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits();
unsigned VRegsPerSrc = NumElts / ElemsPerVReg;
SmallVector<std::pair<int, SmallVector<int>>>
OutMasks(VRegsPerSrc, {-1, {}});
// Check if our mask can be done as a 1-to-1 mapping from source
// to destination registers in the group without needing to
// write each destination more than once.
for (unsigned DstIdx = 0; DstIdx < Mask.size(); DstIdx++) {
int DstVecIdx = DstIdx / ElemsPerVReg;
int DstSubIdx = DstIdx % ElemsPerVReg;
int SrcIdx = Mask[DstIdx];
if (SrcIdx < 0 || (unsigned)SrcIdx >= 2 * NumElts)
continue;
int SrcVecIdx = SrcIdx / ElemsPerVReg;
int SrcSubIdx = SrcIdx % ElemsPerVReg;
if (OutMasks[DstVecIdx].first == -1)
OutMasks[DstVecIdx].first = SrcVecIdx;
if (OutMasks[DstVecIdx].first != SrcVecIdx)
// Note: This case could easily be handled by keeping track of a chain
// of source values and generating two element shuffles below. This is
// less an implementation question, and more a profitability one.
return SDValue();
OutMasks[DstVecIdx].second.resize(ElemsPerVReg, -1);
OutMasks[DstVecIdx].second[DstSubIdx] = SrcSubIdx;
}
EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
MVT OneRegVT = MVT::getVectorVT(ElemVT, ElemsPerVReg);
MVT M1VT = getContainerForFixedLengthVector(DAG, OneRegVT, Subtarget);
assert(M1VT == getLMUL1VT(M1VT));
unsigned NumOpElts = M1VT.getVectorMinNumElements();
SDValue Vec = DAG.getUNDEF(ContainerVT);
// The following semantically builds up a fixed length concat_vector
// of the component shuffle_vectors. We eagerly lower to scalable here
// to avoid DAG combining it back to a large shuffle_vector again.
V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
V2 = convertToScalableVector(ContainerVT, V2, DAG, Subtarget);
for (unsigned DstVecIdx = 0 ; DstVecIdx < OutMasks.size(); DstVecIdx++) {
auto &[SrcVecIdx, SrcSubMask] = OutMasks[DstVecIdx];
if (SrcVecIdx == -1)
continue;
unsigned ExtractIdx = (SrcVecIdx % VRegsPerSrc) * NumOpElts;
SDValue SrcVec = (unsigned)SrcVecIdx >= VRegsPerSrc ? V2 : V1;
SDValue SubVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, M1VT, SrcVec,
DAG.getVectorIdxConstant(ExtractIdx, DL));
SubVec = convertFromScalableVector(OneRegVT, SubVec, DAG, Subtarget);
SubVec = DAG.getVectorShuffle(OneRegVT, DL, SubVec, SubVec, SrcSubMask);
SubVec = convertToScalableVector(M1VT, SubVec, DAG, Subtarget);
unsigned InsertIdx = DstVecIdx * NumOpElts;
Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT, Vec, SubVec,
DAG.getVectorIdxConstant(InsertIdx, DL));
}
return convertFromScalableVector(VT, Vec, DAG, Subtarget);
}
static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
MVT VT = Op.getSimpleValueType();
unsigned NumElts = VT.getVectorNumElements();
ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
if (VT.getVectorElementType() == MVT::i1) {
// Lower to a vror.vi of a larger element type if possible before we promote
// i1s to i8s.
if (SDValue V = lowerVECTOR_SHUFFLEAsRotate(SVN, DAG, Subtarget))
return V;
if (SDValue V = lowerBitreverseShuffle(SVN, DAG, Subtarget))
return V;
// Promote i1 shuffle to i8 shuffle.
MVT WidenVT = MVT::getVectorVT(MVT::i8, VT.getVectorElementCount());
V1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WidenVT, V1);
V2 = V2.isUndef() ? DAG.getUNDEF(WidenVT)
: DAG.getNode(ISD::ZERO_EXTEND, DL, WidenVT, V2);
SDValue Shuffled = DAG.getVectorShuffle(WidenVT, DL, V1, V2, SVN->getMask());
return DAG.getSetCC(DL, VT, Shuffled, DAG.getConstant(0, DL, WidenVT),
ISD::SETNE);
}
MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
auto [TrueMask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
if (SVN->isSplat()) {
const int Lane = SVN->getSplatIndex();
if (Lane >= 0) {
MVT SVT = VT.getVectorElementType();
// Turn splatted vector load into a strided load with an X0 stride.
SDValue V = V1;
// Peek through CONCAT_VECTORS as VectorCombine can concat a vector
// with undef.
// FIXME: Peek through INSERT_SUBVECTOR, EXTRACT_SUBVECTOR, bitcasts?
int Offset = Lane;
if (V.getOpcode() == ISD::CONCAT_VECTORS) {
int OpElements =
V.getOperand(0).getSimpleValueType().getVectorNumElements();
V = V.getOperand(Offset / OpElements);
Offset %= OpElements;
}
// We need to ensure the load isn't atomic or volatile.
if (ISD::isNormalLoad(V.getNode()) && cast<LoadSDNode>(V)->isSimple()) {
auto *Ld = cast<LoadSDNode>(V);
Offset *= SVT.getStoreSize();
SDValue NewAddr = DAG.getMemBasePlusOffset(
Ld->getBasePtr(), TypeSize::getFixed(Offset), DL);
// If this is SEW=64 on RV32, use a strided load with a stride of x0.
if (SVT.isInteger() && SVT.bitsGT(XLenVT)) {
SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
SDValue IntID =
DAG.getTargetConstant(Intrinsic::riscv_vlse, DL, XLenVT);
SDValue Ops[] = {Ld->getChain(),
IntID,
DAG.getUNDEF(ContainerVT),
NewAddr,
DAG.getRegister(RISCV::X0, XLenVT),
VL};
SDValue NewLoad = DAG.getMemIntrinsicNode(
ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, SVT,
DAG.getMachineFunction().getMachineMemOperand(
Ld->getMemOperand(), Offset, SVT.getStoreSize()));
DAG.makeEquivalentMemoryOrdering(Ld, NewLoad);
return convertFromScalableVector(VT, NewLoad, DAG, Subtarget);
}
// Otherwise use a scalar load and splat. This will give the best
// opportunity to fold a splat into the operation. ISel can turn it into
// the x0 strided load if we aren't able to fold away the select.
if (SVT.isFloatingPoint())
V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
Ld->getPointerInfo().getWithOffset(Offset),
Ld->getOriginalAlign(),
Ld->getMemOperand()->getFlags());
else
V = DAG.getExtLoad(ISD::SEXTLOAD, DL, XLenVT, Ld->getChain(), NewAddr,
Ld->getPointerInfo().getWithOffset(Offset), SVT,
Ld->getOriginalAlign(),
Ld->getMemOperand()->getFlags());
DAG.makeEquivalentMemoryOrdering(Ld, V);
unsigned Opc =
VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL : RISCVISD::VMV_V_X_VL;
SDValue Splat =
DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(ContainerVT), V, VL);
return convertFromScalableVector(VT, Splat, DAG, Subtarget);
}
V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
assert(Lane < (int)NumElts && "Unexpected lane!");
SDValue Gather = DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT,
V1, DAG.getConstant(Lane, DL, XLenVT),
DAG.getUNDEF(ContainerVT), TrueMask, VL);
return convertFromScalableVector(VT, Gather, DAG, Subtarget);
}
}
// For exact VLEN m2 or greater, try to split to m1 operations if we
// can split cleanly.
if (SDValue V = lowerShuffleViaVRegSplitting(SVN, DAG, Subtarget))
return V;
ArrayRef<int> Mask = SVN->getMask();
if (SDValue V =
lowerVECTOR_SHUFFLEAsVSlide1(DL, VT, V1, V2, Mask, Subtarget, DAG))
return V;
if (SDValue V =
lowerVECTOR_SHUFFLEAsVSlidedown(DL, VT, V1, V2, Mask, Subtarget, DAG))
return V;
// A bitrotate will be one instruction on Zvkb, so try to lower to it first if
// available.
if (Subtarget.hasStdExtZvkb())
if (SDValue V = lowerVECTOR_SHUFFLEAsRotate(SVN, DAG, Subtarget))
return V;
// Lower rotations to a SLIDEDOWN and a SLIDEUP. One of the source vectors may
// be undef which can be handled with a single SLIDEDOWN/UP.
int LoSrc, HiSrc;
int Rotation = isElementRotate(LoSrc, HiSrc, Mask);
if (Rotation > 0) {
SDValue LoV, HiV;
if (LoSrc >= 0) {
LoV = LoSrc == 0 ? V1 : V2;
LoV = convertToScalableVector(ContainerVT, LoV, DAG, Subtarget);
}
if (HiSrc >= 0) {
HiV = HiSrc == 0 ? V1 : V2;
HiV = convertToScalableVector(ContainerVT, HiV, DAG, Subtarget);
}
// We found a rotation. We need to slide HiV down by Rotation. Then we need
// to slide LoV up by (NumElts - Rotation).
unsigned InvRotate = NumElts - Rotation;
SDValue Res = DAG.getUNDEF(ContainerVT);
if (HiV) {
// Even though we could use a smaller VL, don't to avoid a vsetivli
// toggle.
Res = getVSlidedown(DAG, Subtarget, DL, ContainerVT, Res, HiV,
DAG.getConstant(Rotation, DL, XLenVT), TrueMask, VL);
}
if (LoV)
Res = getVSlideup(DAG, Subtarget, DL, ContainerVT, Res, LoV,
DAG.getConstant(InvRotate, DL, XLenVT), TrueMask, VL,
RISCVII::TAIL_AGNOSTIC);
return convertFromScalableVector(VT, Res, DAG, Subtarget);
}
// If this is a deinterleave and we can widen the vector, then we can use
// vnsrl to deinterleave.
if (isDeinterleaveShuffle(VT, ContainerVT, V1, V2, Mask, Subtarget)) {
return getDeinterleaveViaVNSRL(DL, VT, V1.getOperand(0), Mask[0] == 0,
Subtarget, DAG);
}
if (SDValue V =
lowerVECTOR_SHUFFLEAsVSlideup(DL, VT, V1, V2, Mask, Subtarget, DAG))
return V;
// Detect an interleave shuffle and lower to
// (vmaccu.vx (vwaddu.vx lohalf(V1), lohalf(V2)), lohalf(V2), (2^eltbits - 1))
int EvenSrc, OddSrc;
if (isInterleaveShuffle(Mask, VT, EvenSrc, OddSrc, Subtarget)) {
// Extract the halves of the vectors.
MVT HalfVT = VT.getHalfNumVectorElementsVT();
int Size = Mask.size();
SDValue EvenV, OddV;
assert(EvenSrc >= 0 && "Undef source?");
EvenV = (EvenSrc / Size) == 0 ? V1 : V2;
EvenV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, EvenV,
DAG.getVectorIdxConstant(EvenSrc % Size, DL));
assert(OddSrc >= 0 && "Undef source?");
OddV = (OddSrc / Size) == 0 ? V1 : V2;
OddV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, OddV,
DAG.getVectorIdxConstant(OddSrc % Size, DL));
return getWideningInterleave(EvenV, OddV, DL, DAG, Subtarget);
}
// Handle any remaining single source shuffles
assert(!V1.isUndef() && "Unexpected shuffle canonicalization");
if (V2.isUndef()) {
// We might be able to express the shuffle as a bitrotate. But even if we
// don't have Zvkb and have to expand, the expanded sequence of approx. 2
// shifts and a vor will have a higher throughput than a vrgather.
if (SDValue V = lowerVECTOR_SHUFFLEAsRotate(SVN, DAG, Subtarget))
return V;
if (VT.getScalarSizeInBits() == 8 &&
any_of(Mask, [&](const auto &Idx) { return Idx > 255; })) {
// On such a vector we're unable to use i8 as the index type.
// FIXME: We could promote the index to i16 and use vrgatherei16, but that
// may involve vector splitting if we're already at LMUL=8, or our
// user-supplied maximum fixed-length LMUL.
return SDValue();
}
// Base case for the two operand recursion below - handle the worst case
// single source shuffle.
unsigned GatherVVOpc = RISCVISD::VRGATHER_VV_VL;
MVT IndexVT = VT.changeTypeToInteger();
// Since we can't introduce illegal index types at this stage, use i16 and
// vrgatherei16 if the corresponding index type for plain vrgather is greater
// than XLenVT.
if (IndexVT.getScalarType().bitsGT(XLenVT)) {
GatherVVOpc = RISCVISD::VRGATHEREI16_VV_VL;
IndexVT = IndexVT.changeVectorElementType(MVT::i16);
}
// If the mask allows, we can do all the index computation in 16 bits. This
// requires less work and less register pressure at high LMUL, and creates
// smaller constants which may be cheaper to materialize.
if (IndexVT.getScalarType().bitsGT(MVT::i16) && isUInt<16>(NumElts - 1) &&
(IndexVT.getSizeInBits() / Subtarget.getRealMinVLen()) > 1) {
GatherVVOpc = RISCVISD::VRGATHEREI16_VV_VL;
IndexVT = IndexVT.changeVectorElementType(MVT::i16);
}
MVT IndexContainerVT =
ContainerVT.changeVectorElementType(IndexVT.getScalarType());
V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
SmallVector<SDValue> GatherIndicesLHS;
for (int MaskIndex : Mask) {
bool IsLHSIndex = MaskIndex < (int)NumElts && MaskIndex >= 0;
GatherIndicesLHS.push_back(IsLHSIndex
? DAG.getConstant(MaskIndex, DL, XLenVT)
: DAG.getUNDEF(XLenVT));
}
SDValue LHSIndices = DAG.getBuildVector(IndexVT, DL, GatherIndicesLHS);
LHSIndices = convertToScalableVector(IndexContainerVT, LHSIndices, DAG,
Subtarget);
SDValue Gather = DAG.getNode(GatherVVOpc, DL, ContainerVT, V1, LHSIndices,
DAG.getUNDEF(ContainerVT), TrueMask, VL);
return convertFromScalableVector(VT, Gather, DAG, Subtarget);
}
// By default we preserve the original operand order, and use a mask to
// select LHS as true and RHS as false. However, since RVV vector selects may
// feature splats but only on the LHS, we may choose to invert our mask and
// instead select between RHS and LHS.
bool SwapOps = DAG.isSplatValue(V2) && !DAG.isSplatValue(V1);
// Detect shuffles which can be re-expressed as vector selects; these are
// shuffles in which each element in the destination is taken from an element
// at the corresponding index in either source vectors.
bool IsSelect = all_of(enumerate(Mask), [&](const auto &MaskIdx) {
int MaskIndex = MaskIdx.value();
return MaskIndex < 0 || MaskIdx.index() == (unsigned)MaskIndex % NumElts;
});
if (IsSelect) {
// Now construct the mask that will be used by the vselect operation.
SmallVector<SDValue> MaskVals;
for (int MaskIndex : Mask) {
bool SelectMaskVal = (MaskIndex < (int)NumElts) ^ SwapOps;
MaskVals.push_back(DAG.getConstant(SelectMaskVal, DL, XLenVT));
}
if (SwapOps)
std::swap(V1, V2);
assert(MaskVals.size() == NumElts && "Unexpected select-like shuffle");
MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
SDValue SelectMask = DAG.getBuildVector(MaskVT, DL, MaskVals);
return DAG.getNode(ISD::VSELECT, DL, VT, SelectMask, V1, V2);
}
// As a backup, shuffles can be lowered via a vrgather instruction, possibly
// merged with a second vrgather.
SmallVector<int> ShuffleMaskLHS, ShuffleMaskRHS;
SmallVector<SDValue> MaskVals;
// Now construct the mask that will be used by the blended vrgather operation.
// Cconstruct the appropriate indices into each vector.
for (int MaskIndex : Mask) {
bool SelectMaskVal = (MaskIndex < (int)NumElts) ^ !SwapOps;
MaskVals.push_back(DAG.getConstant(SelectMaskVal, DL, XLenVT));
bool IsLHSOrUndefIndex = MaskIndex < (int)NumElts;
ShuffleMaskLHS.push_back(IsLHSOrUndefIndex && MaskIndex >= 0
? MaskIndex : -1);
ShuffleMaskRHS.push_back(IsLHSOrUndefIndex ? -1 : (MaskIndex - NumElts));
}
if (SwapOps) {
std::swap(V1, V2);
std::swap(ShuffleMaskLHS, ShuffleMaskRHS);
}
assert(MaskVals.size() == NumElts && "Unexpected select-like shuffle");
MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
SDValue SelectMask = DAG.getBuildVector(MaskVT, DL, MaskVals);
// Recursively invoke lowering for each operand if we had two
// independent single source shuffles, and then combine the result via a
// vselect. Note that the vselect will likely be folded back into the
// second permute (vrgather, or other) by the post-isel combine.
V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), ShuffleMaskLHS);
V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), ShuffleMaskRHS);
return DAG.getNode(ISD::VSELECT, DL, VT, SelectMask, V2, V1);
}
bool RISCVTargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const {
// Support splats for any type. These should type legalize well.
if (ShuffleVectorSDNode::isSplatMask(M.data(), VT))
return true;
// Only support legal VTs for other shuffles for now.
if (!isTypeLegal(VT))
return false;
MVT SVT = VT.getSimpleVT();
// Not for i1 vectors.
if (SVT.getScalarType() == MVT::i1)
return false;
int Dummy1, Dummy2;
return (isElementRotate(Dummy1, Dummy2, M) > 0) ||
isInterleaveShuffle(M, SVT, Dummy1, Dummy2, Subtarget);
}
// Lower CTLZ_ZERO_UNDEF or CTTZ_ZERO_UNDEF by converting to FP and extracting
// the exponent.
SDValue
RISCVTargetLowering::lowerCTLZ_CTTZ_ZERO_UNDEF(SDValue Op,
SelectionDAG &DAG) const {
MVT VT = Op.getSimpleValueType();
unsigned EltSize = VT.getScalarSizeInBits();
SDValue Src = Op.getOperand(0);
SDLoc DL(Op);
MVT ContainerVT = VT;
SDValue Mask, VL;
if (Op->isVPOpcode()) {
Mask = Op.getOperand(1);
if (VT.isFixedLengthVector())
Mask = convertToScalableVector(getMaskTypeFor(ContainerVT), Mask, DAG,
Subtarget);
VL = Op.getOperand(2);
}
// We choose FP type that can represent the value if possible. Otherwise, we
// use rounding to zero conversion for correct exponent of the result.
// TODO: Use f16 for i8 when possible?
MVT FloatEltVT = (EltSize >= 32) ? MVT::f64 : MVT::f32;
if (!isTypeLegal(MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount())))
FloatEltVT = MVT::f32;
MVT FloatVT = MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount());
// Legal types should have been checked in the RISCVTargetLowering
// constructor.
// TODO: Splitting may make sense in some cases.
assert(DAG.getTargetLoweringInfo().isTypeLegal(FloatVT) &&
"Expected legal float type!");
// For CTTZ_ZERO_UNDEF, we need to extract the lowest set bit using X & -X.
// The trailing zero count is equal to log2 of this single bit value.
if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF) {
SDValue Neg = DAG.getNegative(Src, DL, VT);
Src = DAG.getNode(ISD::AND, DL, VT, Src, Neg);
} else if (Op.getOpcode() == ISD::VP_CTTZ_ZERO_UNDEF) {
SDValue Neg = DAG.getNode(ISD::VP_SUB, DL, VT, DAG.getConstant(0, DL, VT),
Src, Mask, VL);
Src = DAG.getNode(ISD::VP_AND, DL, VT, Src, Neg, Mask, VL);
}
// We have a legal FP type, convert to it.
SDValue FloatVal;
if (FloatVT.bitsGT(VT)) {
if (Op->isVPOpcode())
FloatVal = DAG.getNode(ISD::VP_UINT_TO_FP, DL, FloatVT, Src, Mask, VL);
else
FloatVal = DAG.getNode(ISD::UINT_TO_FP, DL, FloatVT, Src);
} else {
// Use RTZ to avoid rounding influencing exponent of FloatVal.
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
}
if (!Op->isVPOpcode())
std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
SDValue RTZRM =
DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, Subtarget.getXLenVT());
MVT ContainerFloatVT =
MVT::getVectorVT(FloatEltVT, ContainerVT.getVectorElementCount());
FloatVal = DAG.getNode(RISCVISD::VFCVT_RM_F_XU_VL, DL, ContainerFloatVT,
Src, Mask, RTZRM, VL);
if (VT.isFixedLengthVector())
FloatVal = convertFromScalableVector(FloatVT, FloatVal, DAG, Subtarget);
}
// Bitcast to integer and shift the exponent to the LSB.
EVT IntVT = FloatVT.changeVectorElementTypeToInteger();
SDValue Bitcast = DAG.getBitcast(IntVT, FloatVal);
unsigned ShiftAmt = FloatEltVT == MVT::f64 ? 52 : 23;
SDValue Exp;
// Restore back to original type. Truncation after SRL is to generate vnsrl.
if (Op->isVPOpcode()) {
Exp = DAG.getNode(ISD::VP_LSHR, DL, IntVT, Bitcast,
DAG.getConstant(ShiftAmt, DL, IntVT), Mask, VL);
Exp = DAG.getVPZExtOrTrunc(DL, VT, Exp, Mask, VL);
} else {
Exp = DAG.getNode(ISD::SRL, DL, IntVT, Bitcast,
DAG.getConstant(ShiftAmt, DL, IntVT));
if (IntVT.bitsLT(VT))
Exp = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Exp);
else if (IntVT.bitsGT(VT))
Exp = DAG.getNode(ISD::TRUNCATE, DL, VT, Exp);
}
// The exponent contains log2 of the value in biased form.
unsigned ExponentBias = FloatEltVT == MVT::f64 ? 1023 : 127;
// For trailing zeros, we just need to subtract the bias.
if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF)
return DAG.getNode(ISD::SUB, DL, VT, Exp,
DAG.getConstant(ExponentBias, DL, VT));
if (Op.getOpcode() == ISD::VP_CTTZ_ZERO_UNDEF)
return DAG.getNode(ISD::VP_SUB, DL, VT, Exp,
DAG.getConstant(ExponentBias, DL, VT), Mask, VL);
// For leading zeros, we need to remove the bias and convert from log2 to
// leading zeros. We can do this by subtracting from (Bias + (EltSize - 1)).
unsigned Adjust = ExponentBias + (EltSize - 1);
SDValue Res;
if (Op->isVPOpcode())
Res = DAG.getNode(ISD::VP_SUB, DL, VT, DAG.getConstant(Adjust, DL, VT), Exp,
Mask, VL);
else
Res = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(Adjust, DL, VT), Exp);
// The above result with zero input equals to Adjust which is greater than
// EltSize. Hence, we can do min(Res, EltSize) for CTLZ.
if (Op.getOpcode() == ISD::CTLZ)
Res = DAG.getNode(ISD::UMIN, DL, VT, Res, DAG.getConstant(EltSize, DL, VT));
else if (Op.getOpcode() == ISD::VP_CTLZ)
Res = DAG.getNode(ISD::VP_UMIN, DL, VT, Res,
DAG.getConstant(EltSize, DL, VT), Mask, VL);
return Res;
}
// While RVV has alignment restrictions, we should always be able to load as a
// legal equivalently-sized byte-typed vector instead. This method is
// responsible for re-expressing a ISD::LOAD via a correctly-aligned type. If
// the load is already correctly-aligned, it returns SDValue().
SDValue RISCVTargetLowering::expandUnalignedRVVLoad(SDValue Op,
SelectionDAG &DAG) const {
auto *Load = cast<LoadSDNode>(Op);
assert(Load && Load->getMemoryVT().isVector() && "Expected vector load");
if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
Load->getMemoryVT(),
*Load->getMemOperand()))
return SDValue();
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
unsigned EltSizeBits = VT.getScalarSizeInBits();
assert((EltSizeBits == 16 || EltSizeBits == 32 || EltSizeBits == 64) &&
"Unexpected unaligned RVV load type");
MVT NewVT =
MVT::getVectorVT(MVT::i8, VT.getVectorElementCount() * (EltSizeBits / 8));
assert(NewVT.isValid() &&
"Expecting equally-sized RVV vector types to be legal");
SDValue L = DAG.getLoad(NewVT, DL, Load->getChain(), Load->getBasePtr(),
Load->getPointerInfo(), Load->getOriginalAlign(),
Load->getMemOperand()->getFlags());
return DAG.getMergeValues({DAG.getBitcast(VT, L), L.getValue(1)}, DL);
}
// While RVV has alignment restrictions, we should always be able to store as a
// legal equivalently-sized byte-typed vector instead. This method is
// responsible for re-expressing a ISD::STORE via a correctly-aligned type. It
// returns SDValue() if the store is already correctly aligned.
SDValue RISCVTargetLowering::expandUnalignedRVVStore(SDValue Op,
SelectionDAG &DAG) const {
auto *Store = cast<StoreSDNode>(Op);
assert(Store && Store->getValue().getValueType().isVector() &&
"Expected vector store");
if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
Store->getMemoryVT(),
*Store->getMemOperand()))
return SDValue();
SDLoc DL(Op);
SDValue StoredVal = Store->getValue();
MVT VT = StoredVal.getSimpleValueType();
unsigned EltSizeBits = VT.getScalarSizeInBits();
assert((EltSizeBits == 16 || EltSizeBits == 32 || EltSizeBits == 64) &&
"Unexpected unaligned RVV store type");
MVT NewVT =
MVT::getVectorVT(MVT::i8, VT.getVectorElementCount() * (EltSizeBits / 8));
assert(NewVT.isValid() &&
"Expecting equally-sized RVV vector types to be legal");
StoredVal = DAG.getBitcast(NewVT, StoredVal);
return DAG.getStore(Store->getChain(), DL, StoredVal, Store->getBasePtr(),
Store->getPointerInfo(), Store->getOriginalAlign(),
Store->getMemOperand()->getFlags());
}
static SDValue lowerConstant(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(Op.getValueType() == MVT::i64 && "Unexpected VT");
int64_t Imm = cast<ConstantSDNode>(Op)->getSExtValue();
// All simm32 constants should be handled by isel.
// NOTE: The getMaxBuildIntsCost call below should return a value >= 2 making
// this check redundant, but small immediates are common so this check
// should have better compile time.
if (isInt<32>(Imm))
return Op;
// We only need to cost the immediate, if constant pool lowering is enabled.
if (!Subtarget.useConstantPoolForLargeInts())
return Op;
RISCVMatInt::InstSeq Seq = RISCVMatInt::generateInstSeq(Imm, Subtarget);
if (Seq.size() <= Subtarget.getMaxBuildIntsCost())
return Op;
// Optimizations below are disabled for opt size. If we're optimizing for
// size, use a constant pool.
if (DAG.shouldOptForSize())
return SDValue();
// Special case. See if we can build the constant as (ADD (SLLI X, C), X) do
// that if it will avoid a constant pool.
// It will require an extra temporary register though.
// If we have Zba we can use (ADD_UW X, (SLLI X, 32)) to handle cases where
// low and high 32 bits are the same and bit 31 and 63 are set.
unsigned ShiftAmt, AddOpc;
RISCVMatInt::InstSeq SeqLo =
RISCVMatInt::generateTwoRegInstSeq(Imm, Subtarget, ShiftAmt, AddOpc);
if (!SeqLo.empty() && (SeqLo.size() + 2) <= Subtarget.getMaxBuildIntsCost())
return Op;
return SDValue();
}
static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDLoc dl(Op);
AtomicOrdering FenceOrdering =
static_cast<AtomicOrdering>(Op.getConstantOperandVal(1));
SyncScope::ID FenceSSID =
static_cast<SyncScope::ID>(Op.getConstantOperandVal(2));
if (Subtarget.hasStdExtZtso()) {
// The only fence that needs an instruction is a sequentially-consistent
// cross-thread fence.
if (FenceOrdering == AtomicOrdering::SequentiallyConsistent &&
FenceSSID == SyncScope::System)
return Op;
// MEMBARRIER is a compiler barrier; it codegens to a no-op.
return DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
}
// singlethread fences only synchronize with signal handlers on the same
// thread and thus only need to preserve instruction order, not actually
// enforce memory ordering.
if (FenceSSID == SyncScope::SingleThread)
// MEMBARRIER is a compiler barrier; it codegens to a no-op.
return DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
return Op;
}
static SDValue lowerSADDSAT_SSUBSAT(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i32 && RV64LegalI32 &&
"Unexpected custom legalisation");
// With Zbb, we can widen to i64 and smin/smax with INT32_MAX/MIN.
bool IsAdd = Op.getOpcode() == ISD::SADDSAT;
SDLoc DL(Op);
SDValue LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(0));
SDValue RHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue Result =
DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, DL, MVT::i64, LHS, RHS);
APInt MinVal = APInt::getSignedMinValue(32).sext(64);
APInt MaxVal = APInt::getSignedMaxValue(32).sext(64);
SDValue SatMin = DAG.getConstant(MinVal, DL, MVT::i64);
SDValue SatMax = DAG.getConstant(MaxVal, DL, MVT::i64);
Result = DAG.getNode(ISD::SMIN, DL, MVT::i64, Result, SatMax);
Result = DAG.getNode(ISD::SMAX, DL, MVT::i64, Result, SatMin);
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Result);
}
static SDValue lowerUADDSAT_USUBSAT(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i32 && RV64LegalI32 &&
"Unexpected custom legalisation");
// With Zbb we can sign extend and let LegalizeDAG use minu/maxu. Using
// sign extend allows overflow of the lower 32 bits to be detected on
// the promoted size.
SDLoc DL(Op);
SDValue LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(0));
SDValue RHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue WideOp = DAG.getNode(Op.getOpcode(), DL, MVT::i64, LHS, RHS);
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, WideOp);
}
// Custom lower i32 SADDO/SSUBO with RV64LegalI32 so we take advantage of addw.
static SDValue lowerSADDO_SSUBO(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i32 && RV64LegalI32 &&
"Unexpected custom legalisation");
if (isa<ConstantSDNode>(Op.getOperand(1)))
return SDValue();
bool IsAdd = Op.getOpcode() == ISD::SADDO;
SDLoc DL(Op);
SDValue LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(0));
SDValue RHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue WideOp =
DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, DL, MVT::i64, LHS, RHS);
SDValue Res = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, WideOp);
SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, WideOp,
DAG.getValueType(MVT::i32));
SDValue Ovf = DAG.getSetCC(DL, Op.getValue(1).getValueType(), WideOp, SExt,
ISD::SETNE);
return DAG.getMergeValues({Res, Ovf}, DL);
}
// Custom lower i32 SMULO with RV64LegalI32 so we take advantage of mulw.
static SDValue lowerSMULO(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i32 && RV64LegalI32 &&
"Unexpected custom legalisation");
SDLoc DL(Op);
SDValue LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(0));
SDValue RHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
SDValue Res = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Mul,
DAG.getValueType(MVT::i32));
SDValue Ovf = DAG.getSetCC(DL, Op.getValue(1).getValueType(), Mul, SExt,
ISD::SETNE);
return DAG.getMergeValues({Res, Ovf}, DL);
}
SDValue RISCVTargetLowering::LowerIS_FPCLASS(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
unsigned Check = Op.getConstantOperandVal(1);
unsigned TDCMask = 0;
if (Check & fcSNan)
TDCMask |= RISCV::FPMASK_Signaling_NaN;
if (Check & fcQNan)
TDCMask |= RISCV::FPMASK_Quiet_NaN;
if (Check & fcPosInf)
TDCMask |= RISCV::FPMASK_Positive_Infinity;
if (Check & fcNegInf)
TDCMask |= RISCV::FPMASK_Negative_Infinity;
if (Check & fcPosNormal)
TDCMask |= RISCV::FPMASK_Positive_Normal;
if (Check & fcNegNormal)
TDCMask |= RISCV::FPMASK_Negative_Normal;
if (Check & fcPosSubnormal)
TDCMask |= RISCV::FPMASK_Positive_Subnormal;
if (Check & fcNegSubnormal)
TDCMask |= RISCV::FPMASK_Negative_Subnormal;
if (Check & fcPosZero)
TDCMask |= RISCV::FPMASK_Positive_Zero;
if (Check & fcNegZero)
TDCMask |= RISCV::FPMASK_Negative_Zero;
bool IsOneBitMask = isPowerOf2_32(TDCMask);
SDValue TDCMaskV = DAG.getConstant(TDCMask, DL, XLenVT);
if (VT.isVector()) {
SDValue Op0 = Op.getOperand(0);
MVT VT0 = Op.getOperand(0).getSimpleValueType();
if (VT.isScalableVector()) {
MVT DstVT = VT0.changeVectorElementTypeToInteger();
auto [Mask, VL] = getDefaultScalableVLOps(VT0, DL, DAG, Subtarget);
if (Op.getOpcode() == ISD::VP_IS_FPCLASS) {
Mask = Op.getOperand(2);
VL = Op.getOperand(3);
}
SDValue FPCLASS = DAG.getNode(RISCVISD::FCLASS_VL, DL, DstVT, Op0, Mask,
VL, Op->getFlags());
if (IsOneBitMask)
return DAG.getSetCC(DL, VT, FPCLASS,
DAG.getConstant(TDCMask, DL, DstVT),
ISD::CondCode::SETEQ);
SDValue AND = DAG.getNode(ISD::AND, DL, DstVT, FPCLASS,
DAG.getConstant(TDCMask, DL, DstVT));
return DAG.getSetCC(DL, VT, AND, DAG.getConstant(0, DL, DstVT),
ISD::SETNE);
}
MVT ContainerVT0 = getContainerForFixedLengthVector(VT0);
MVT ContainerVT = getContainerForFixedLengthVector(VT);
MVT ContainerDstVT = ContainerVT0.changeVectorElementTypeToInteger();
auto [Mask, VL] = getDefaultVLOps(VT0, ContainerVT0, DL, DAG, Subtarget);
if (Op.getOpcode() == ISD::VP_IS_FPCLASS) {
Mask = Op.getOperand(2);
MVT MaskContainerVT =
getContainerForFixedLengthVector(Mask.getSimpleValueType());
Mask = convertToScalableVector(MaskContainerVT, Mask, DAG, Subtarget);
VL = Op.getOperand(3);
}
Op0 = convertToScalableVector(ContainerVT0, Op0, DAG, Subtarget);
SDValue FPCLASS = DAG.getNode(RISCVISD::FCLASS_VL, DL, ContainerDstVT, Op0,
Mask, VL, Op->getFlags());
TDCMaskV = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerDstVT,
DAG.getUNDEF(ContainerDstVT), TDCMaskV, VL);
if (IsOneBitMask) {
SDValue VMSEQ =
DAG.getNode(RISCVISD::SETCC_VL, DL, ContainerVT,
{FPCLASS, TDCMaskV, DAG.getCondCode(ISD::SETEQ),
DAG.getUNDEF(ContainerVT), Mask, VL});
return convertFromScalableVector(VT, VMSEQ, DAG, Subtarget);
}
SDValue AND = DAG.getNode(RISCVISD::AND_VL, DL, ContainerDstVT, FPCLASS,
TDCMaskV, DAG.getUNDEF(ContainerDstVT), Mask, VL);
SDValue SplatZero = DAG.getConstant(0, DL, XLenVT);
SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerDstVT,
DAG.getUNDEF(ContainerDstVT), SplatZero, VL);
SDValue VMSNE = DAG.getNode(RISCVISD::SETCC_VL, DL, ContainerVT,
{AND, SplatZero, DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(ContainerVT), Mask, VL});
return convertFromScalableVector(VT, VMSNE, DAG, Subtarget);
}
SDValue FCLASS = DAG.getNode(RISCVISD::FCLASS, DL, XLenVT, Op.getOperand(0));
SDValue AND = DAG.getNode(ISD::AND, DL, XLenVT, FCLASS, TDCMaskV);
SDValue Res = DAG.getSetCC(DL, XLenVT, AND, DAG.getConstant(0, DL, XLenVT),
ISD::CondCode::SETNE);
return DAG.getNode(ISD::TRUNCATE, DL, VT, Res);
}
// Lower fmaximum and fminimum. Unlike our fmax and fmin instructions, these
// operations propagate nans.
static SDValue lowerFMAXIMUM_FMINIMUM(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue X = Op.getOperand(0);
SDValue Y = Op.getOperand(1);
if (!VT.isVector()) {
MVT XLenVT = Subtarget.getXLenVT();
// If X is a nan, replace Y with X. If Y is a nan, replace X with Y. This
// ensures that when one input is a nan, the other will also be a nan
// allowing the nan to propagate. If both inputs are nan, this will swap the
// inputs which is harmless.
SDValue NewY = Y;
if (!Op->getFlags().hasNoNaNs() && !DAG.isKnownNeverNaN(X)) {
SDValue XIsNonNan = DAG.getSetCC(DL, XLenVT, X, X, ISD::SETOEQ);
NewY = DAG.getSelect(DL, VT, XIsNonNan, Y, X);
}
SDValue NewX = X;
if (!Op->getFlags().hasNoNaNs() && !DAG.isKnownNeverNaN(Y)) {
SDValue YIsNonNan = DAG.getSetCC(DL, XLenVT, Y, Y, ISD::SETOEQ);
NewX = DAG.getSelect(DL, VT, YIsNonNan, X, Y);
}
unsigned Opc =
Op.getOpcode() == ISD::FMAXIMUM ? RISCVISD::FMAX : RISCVISD::FMIN;
return DAG.getNode(Opc, DL, VT, NewX, NewY);
}
// Check no NaNs before converting to fixed vector scalable.
bool XIsNeverNan = Op->getFlags().hasNoNaNs() || DAG.isKnownNeverNaN(X);
bool YIsNeverNan = Op->getFlags().hasNoNaNs() || DAG.isKnownNeverNaN(Y);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
X = convertToScalableVector(ContainerVT, X, DAG, Subtarget);
Y = convertToScalableVector(ContainerVT, Y, DAG, Subtarget);
}
SDValue Mask, VL;
if (Op->isVPOpcode()) {
Mask = Op.getOperand(2);
if (VT.isFixedLengthVector())
Mask = convertToScalableVector(getMaskTypeFor(ContainerVT), Mask, DAG,
Subtarget);
VL = Op.getOperand(3);
} else {
std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
}
SDValue NewY = Y;
if (!XIsNeverNan) {
SDValue XIsNonNan = DAG.getNode(RISCVISD::SETCC_VL, DL, Mask.getValueType(),
{X, X, DAG.getCondCode(ISD::SETOEQ),
DAG.getUNDEF(ContainerVT), Mask, VL});
NewY = DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, XIsNonNan, Y, X,
DAG.getUNDEF(ContainerVT), VL);
}
SDValue NewX = X;
if (!YIsNeverNan) {
SDValue YIsNonNan = DAG.getNode(RISCVISD::SETCC_VL, DL, Mask.getValueType(),
{Y, Y, DAG.getCondCode(ISD::SETOEQ),
DAG.getUNDEF(ContainerVT), Mask, VL});
NewX = DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, YIsNonNan, X, Y,
DAG.getUNDEF(ContainerVT), VL);
}
unsigned Opc =
Op.getOpcode() == ISD::FMAXIMUM || Op->getOpcode() == ISD::VP_FMAXIMUM
? RISCVISD::VFMAX_VL
: RISCVISD::VFMIN_VL;
SDValue Res = DAG.getNode(Opc, DL, ContainerVT, NewX, NewY,
DAG.getUNDEF(ContainerVT), Mask, VL);
if (VT.isFixedLengthVector())
Res = convertFromScalableVector(VT, Res, DAG, Subtarget);
return Res;
}
/// Get a RISC-V target specified VL op for a given SDNode.
static unsigned getRISCVVLOp(SDValue Op) {
#define OP_CASE(NODE) \
case ISD::NODE: \
return RISCVISD::NODE##_VL;
#define VP_CASE(NODE) \
case ISD::VP_##NODE: \
return RISCVISD::NODE##_VL;
// clang-format off
switch (Op.getOpcode()) {
default:
llvm_unreachable("don't have RISC-V specified VL op for this SDNode");
OP_CASE(ADD)
OP_CASE(SUB)
OP_CASE(MUL)
OP_CASE(MULHS)
OP_CASE(MULHU)
OP_CASE(SDIV)
OP_CASE(SREM)
OP_CASE(UDIV)
OP_CASE(UREM)
OP_CASE(SHL)
OP_CASE(SRA)
OP_CASE(SRL)
OP_CASE(ROTL)
OP_CASE(ROTR)
OP_CASE(BSWAP)
OP_CASE(CTTZ)
OP_CASE(CTLZ)
OP_CASE(CTPOP)
OP_CASE(BITREVERSE)
OP_CASE(SADDSAT)
OP_CASE(UADDSAT)
OP_CASE(SSUBSAT)
OP_CASE(USUBSAT)
OP_CASE(AVGFLOORU)
OP_CASE(AVGCEILU)
OP_CASE(FADD)
OP_CASE(FSUB)
OP_CASE(FMUL)
OP_CASE(FDIV)
OP_CASE(FNEG)
OP_CASE(FABS)
OP_CASE(FSQRT)
OP_CASE(SMIN)
OP_CASE(SMAX)
OP_CASE(UMIN)
OP_CASE(UMAX)
OP_CASE(STRICT_FADD)
OP_CASE(STRICT_FSUB)
OP_CASE(STRICT_FMUL)
OP_CASE(STRICT_FDIV)
OP_CASE(STRICT_FSQRT)
VP_CASE(ADD) // VP_ADD
VP_CASE(SUB) // VP_SUB
VP_CASE(MUL) // VP_MUL
VP_CASE(SDIV) // VP_SDIV
VP_CASE(SREM) // VP_SREM
VP_CASE(UDIV) // VP_UDIV
VP_CASE(UREM) // VP_UREM
VP_CASE(SHL) // VP_SHL
VP_CASE(FADD) // VP_FADD
VP_CASE(FSUB) // VP_FSUB
VP_CASE(FMUL) // VP_FMUL
VP_CASE(FDIV) // VP_FDIV
VP_CASE(FNEG) // VP_FNEG
VP_CASE(FABS) // VP_FABS
VP_CASE(SMIN) // VP_SMIN
VP_CASE(SMAX) // VP_SMAX
VP_CASE(UMIN) // VP_UMIN
VP_CASE(UMAX) // VP_UMAX
VP_CASE(FCOPYSIGN) // VP_FCOPYSIGN
VP_CASE(SETCC) // VP_SETCC
VP_CASE(SINT_TO_FP) // VP_SINT_TO_FP
VP_CASE(UINT_TO_FP) // VP_UINT_TO_FP
VP_CASE(BITREVERSE) // VP_BITREVERSE
VP_CASE(SADDSAT) // VP_SADDSAT
VP_CASE(UADDSAT) // VP_UADDSAT
VP_CASE(SSUBSAT) // VP_SSUBSAT
VP_CASE(USUBSAT) // VP_USUBSAT
VP_CASE(BSWAP) // VP_BSWAP
VP_CASE(CTLZ) // VP_CTLZ
VP_CASE(CTTZ) // VP_CTTZ
VP_CASE(CTPOP) // VP_CTPOP
case ISD::CTLZ_ZERO_UNDEF:
case ISD::VP_CTLZ_ZERO_UNDEF:
return RISCVISD::CTLZ_VL;
case ISD::CTTZ_ZERO_UNDEF:
case ISD::VP_CTTZ_ZERO_UNDEF:
return RISCVISD::CTTZ_VL;
case ISD::FMA:
case ISD::VP_FMA:
return RISCVISD::VFMADD_VL;
case ISD::STRICT_FMA:
return RISCVISD::STRICT_VFMADD_VL;
case ISD::AND:
case ISD::VP_AND:
if (Op.getSimpleValueType().getVectorElementType() == MVT::i1)
return RISCVISD::VMAND_VL;
return RISCVISD::AND_VL;
case ISD::OR:
case ISD::VP_OR:
if (Op.getSimpleValueType().getVectorElementType() == MVT::i1)
return RISCVISD::VMOR_VL;
return RISCVISD::OR_VL;
case ISD::XOR:
case ISD::VP_XOR:
if (Op.getSimpleValueType().getVectorElementType() == MVT::i1)
return RISCVISD::VMXOR_VL;
return RISCVISD::XOR_VL;
case ISD::VP_SELECT:
case ISD::VP_MERGE:
return RISCVISD::VMERGE_VL;
case ISD::VP_ASHR:
return RISCVISD::SRA_VL;
case ISD::VP_LSHR:
return RISCVISD::SRL_VL;
case ISD::VP_SQRT:
return RISCVISD::FSQRT_VL;
case ISD::VP_SIGN_EXTEND:
return RISCVISD::VSEXT_VL;
case ISD::VP_ZERO_EXTEND:
return RISCVISD::VZEXT_VL;
case ISD::VP_FP_TO_SINT:
return RISCVISD::VFCVT_RTZ_X_F_VL;
case ISD::VP_FP_TO_UINT:
return RISCVISD::VFCVT_RTZ_XU_F_VL;
case ISD::FMINNUM:
case ISD::VP_FMINNUM:
return RISCVISD::VFMIN_VL;
case ISD::FMAXNUM:
case ISD::VP_FMAXNUM:
return RISCVISD::VFMAX_VL;
case ISD::LRINT:
case ISD::VP_LRINT:
case ISD::LLRINT:
case ISD::VP_LLRINT:
return RISCVISD::VFCVT_X_F_VL;
}
// clang-format on
#undef OP_CASE
#undef VP_CASE
}
/// Return true if a RISC-V target specified op has a merge operand.
static bool hasMergeOp(unsigned Opcode) {
assert(Opcode > RISCVISD::FIRST_NUMBER &&
Opcode <= RISCVISD::LAST_RISCV_STRICTFP_OPCODE &&
"not a RISC-V target specific op");
static_assert(RISCVISD::LAST_VL_VECTOR_OP - RISCVISD::FIRST_VL_VECTOR_OP ==
126 &&
RISCVISD::LAST_RISCV_STRICTFP_OPCODE -
ISD::FIRST_TARGET_STRICTFP_OPCODE ==
21 &&
"adding target specific op should update this function");
if (Opcode >= RISCVISD::ADD_VL && Opcode <= RISCVISD::VFMAX_VL)
return true;
if (Opcode == RISCVISD::FCOPYSIGN_VL)
return true;
if (Opcode >= RISCVISD::VWMUL_VL && Opcode <= RISCVISD::VFWSUB_W_VL)
return true;
if (Opcode == RISCVISD::SETCC_VL)
return true;
if (Opcode >= RISCVISD::STRICT_FADD_VL && Opcode <= RISCVISD::STRICT_FDIV_VL)
return true;
if (Opcode == RISCVISD::VMERGE_VL)
return true;
return false;
}
/// Return true if a RISC-V target specified op has a mask operand.
static bool hasMaskOp(unsigned Opcode) {
assert(Opcode > RISCVISD::FIRST_NUMBER &&
Opcode <= RISCVISD::LAST_RISCV_STRICTFP_OPCODE &&
"not a RISC-V target specific op");
static_assert(RISCVISD::LAST_VL_VECTOR_OP - RISCVISD::FIRST_VL_VECTOR_OP ==
126 &&
RISCVISD::LAST_RISCV_STRICTFP_OPCODE -
ISD::FIRST_TARGET_STRICTFP_OPCODE ==
21 &&
"adding target specific op should update this function");
if (Opcode >= RISCVISD::TRUNCATE_VECTOR_VL && Opcode <= RISCVISD::SETCC_VL)
return true;
if (Opcode >= RISCVISD::VRGATHER_VX_VL && Opcode <= RISCVISD::VFIRST_VL)
return true;
if (Opcode >= RISCVISD::STRICT_FADD_VL &&
Opcode <= RISCVISD::STRICT_VFROUND_NOEXCEPT_VL)
return true;
return false;
}
static SDValue SplitVectorOp(SDValue Op, SelectionDAG &DAG) {
auto [LoVT, HiVT] = DAG.GetSplitDestVTs(Op.getValueType());
SDLoc DL(Op);
SmallVector<SDValue, 4> LoOperands(Op.getNumOperands());
SmallVector<SDValue, 4> HiOperands(Op.getNumOperands());
for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
if (!Op.getOperand(j).getValueType().isVector()) {
LoOperands[j] = Op.getOperand(j);
HiOperands[j] = Op.getOperand(j);
continue;
}
std::tie(LoOperands[j], HiOperands[j]) =
DAG.SplitVector(Op.getOperand(j), DL);
}
SDValue LoRes =
DAG.getNode(Op.getOpcode(), DL, LoVT, LoOperands, Op->getFlags());
SDValue HiRes =
DAG.getNode(Op.getOpcode(), DL, HiVT, HiOperands, Op->getFlags());
return DAG.getNode(ISD::CONCAT_VECTORS, DL, Op.getValueType(), LoRes, HiRes);
}
static SDValue SplitVPOp(SDValue Op, SelectionDAG &DAG) {
assert(ISD::isVPOpcode(Op.getOpcode()) && "Not a VP op");
auto [LoVT, HiVT] = DAG.GetSplitDestVTs(Op.getValueType());
SDLoc DL(Op);
SmallVector<SDValue, 4> LoOperands(Op.getNumOperands());
SmallVector<SDValue, 4> HiOperands(Op.getNumOperands());
for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
if (ISD::getVPExplicitVectorLengthIdx(Op.getOpcode()) == j) {
std::tie(LoOperands[j], HiOperands[j]) =
DAG.SplitEVL(Op.getOperand(j), Op.getValueType(), DL);
continue;
}
if (!Op.getOperand(j).getValueType().isVector()) {
LoOperands[j] = Op.getOperand(j);
HiOperands[j] = Op.getOperand(j);
continue;
}
std::tie(LoOperands[j], HiOperands[j]) =
DAG.SplitVector(Op.getOperand(j), DL);
}
SDValue LoRes =
DAG.getNode(Op.getOpcode(), DL, LoVT, LoOperands, Op->getFlags());
SDValue HiRes =
DAG.getNode(Op.getOpcode(), DL, HiVT, HiOperands, Op->getFlags());
return DAG.getNode(ISD::CONCAT_VECTORS, DL, Op.getValueType(), LoRes, HiRes);
}
static SDValue SplitVectorReductionOp(SDValue Op, SelectionDAG &DAG) {
SDLoc DL(Op);
auto [Lo, Hi] = DAG.SplitVector(Op.getOperand(1), DL);
auto [MaskLo, MaskHi] = DAG.SplitVector(Op.getOperand(2), DL);
auto [EVLLo, EVLHi] =
DAG.SplitEVL(Op.getOperand(3), Op.getOperand(1).getValueType(), DL);
SDValue ResLo =
DAG.getNode(Op.getOpcode(), DL, Op.getValueType(),
{Op.getOperand(0), Lo, MaskLo, EVLLo}, Op->getFlags());
return DAG.getNode(Op.getOpcode(), DL, Op.getValueType(),
{ResLo, Hi, MaskHi, EVLHi}, Op->getFlags());
}
static SDValue SplitStrictFPVectorOp(SDValue Op, SelectionDAG &DAG) {
assert(Op->isStrictFPOpcode());
auto [LoVT, HiVT] = DAG.GetSplitDestVTs(Op->getValueType(0));
SDVTList LoVTs = DAG.getVTList(LoVT, Op->getValueType(1));
SDVTList HiVTs = DAG.getVTList(HiVT, Op->getValueType(1));
SDLoc DL(Op);
SmallVector<SDValue, 4> LoOperands(Op.getNumOperands());
SmallVector<SDValue, 4> HiOperands(Op.getNumOperands());
for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
if (!Op.getOperand(j).getValueType().isVector()) {
LoOperands[j] = Op.getOperand(j);
HiOperands[j] = Op.getOperand(j);
continue;
}
std::tie(LoOperands[j], HiOperands[j]) =
DAG.SplitVector(Op.getOperand(j), DL);
}
SDValue LoRes =
DAG.getNode(Op.getOpcode(), DL, LoVTs, LoOperands, Op->getFlags());
HiOperands[0] = LoRes.getValue(1);
SDValue HiRes =
DAG.getNode(Op.getOpcode(), DL, HiVTs, HiOperands, Op->getFlags());
SDValue V = DAG.getNode(ISD::CONCAT_VECTORS, DL, Op->getValueType(0),
LoRes.getValue(0), HiRes.getValue(0));
return DAG.getMergeValues({V, HiRes.getValue(1)}, DL);
}
SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
SelectionDAG &DAG) const {
switch (Op.getOpcode()) {
default:
report_fatal_error("unimplemented operand");
case ISD::ATOMIC_FENCE:
return LowerATOMIC_FENCE(Op, DAG, Subtarget);
case ISD::GlobalAddress:
return lowerGlobalAddress(Op, DAG);
case ISD::BlockAddress:
return lowerBlockAddress(Op, DAG);
case ISD::ConstantPool:
return lowerConstantPool(Op, DAG);
case ISD::JumpTable:
return lowerJumpTable(Op, DAG);
case ISD::GlobalTLSAddress:
return lowerGlobalTLSAddress(Op, DAG);
case ISD::Constant:
return lowerConstant(Op, DAG, Subtarget);
case ISD::SELECT:
return lowerSELECT(Op, DAG);
case ISD::BRCOND:
return lowerBRCOND(Op, DAG);
case ISD::VASTART:
return lowerVASTART(Op, DAG);
case ISD::FRAMEADDR:
return lowerFRAMEADDR(Op, DAG);
case ISD::RETURNADDR:
return lowerRETURNADDR(Op, DAG);
case ISD::SADDO:
case ISD::SSUBO:
return lowerSADDO_SSUBO(Op, DAG);
case ISD::SMULO:
return lowerSMULO(Op, DAG);
case ISD::SHL_PARTS:
return lowerShiftLeftParts(Op, DAG);
case ISD::SRA_PARTS:
return lowerShiftRightParts(Op, DAG, true);
case ISD::SRL_PARTS:
return lowerShiftRightParts(Op, DAG, false);
case ISD::ROTL:
case ISD::ROTR:
if (Op.getValueType().isFixedLengthVector()) {
assert(Subtarget.hasStdExtZvkb());
return lowerToScalableOp(Op, DAG);
}
assert(Subtarget.hasVendorXTHeadBb() &&
!(Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()) &&
"Unexpected custom legalization");
// XTHeadBb only supports rotate by constant.
if (!isa<ConstantSDNode>(Op.getOperand(1)))
return SDValue();
return Op;
case ISD::BITCAST: {
SDLoc DL(Op);
EVT VT = Op.getValueType();
SDValue Op0 = Op.getOperand(0);
EVT Op0VT = Op0.getValueType();
MVT XLenVT = Subtarget.getXLenVT();
if (VT == MVT::f16 && Op0VT == MVT::i16 &&
Subtarget.hasStdExtZfhminOrZhinxmin()) {
SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Op0);
SDValue FPConv = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::f16, NewOp0);
return FPConv;
}
if (VT == MVT::bf16 && Op0VT == MVT::i16 &&
Subtarget.hasStdExtZfbfmin()) {
SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Op0);
SDValue FPConv = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::bf16, NewOp0);
return FPConv;
}
if (VT == MVT::f32 && Op0VT == MVT::i32 && Subtarget.is64Bit() &&
Subtarget.hasStdExtFOrZfinx()) {
SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
SDValue FPConv =
DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, NewOp0);
return FPConv;
}
if (VT == MVT::f64 && Op0VT == MVT::i64 && XLenVT == MVT::i32) {
SDValue Lo, Hi;
std::tie(Lo, Hi) = DAG.SplitScalar(Op0, DL, MVT::i32, MVT::i32);
SDValue RetReg =
DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
return RetReg;
}
// Consider other scalar<->scalar casts as legal if the types are legal.
// Otherwise expand them.
if (!VT.isVector() && !Op0VT.isVector()) {
if (isTypeLegal(VT) && isTypeLegal(Op0VT))
return Op;
return SDValue();
}
assert(!VT.isScalableVector() && !Op0VT.isScalableVector() &&
"Unexpected types");
if (VT.isFixedLengthVector()) {
// We can handle fixed length vector bitcasts with a simple replacement
// in isel.
if (Op0VT.isFixedLengthVector())
return Op;
// When bitcasting from scalar to fixed-length vector, insert the scalar
// into a one-element vector of the result type, and perform a vector
// bitcast.
if (!Op0VT.isVector()) {
EVT BVT = EVT::getVectorVT(*DAG.getContext(), Op0VT, 1);
if (!isTypeLegal(BVT))
return SDValue();
return DAG.getBitcast(VT, DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, BVT,
DAG.getUNDEF(BVT), Op0,
DAG.getVectorIdxConstant(0, DL)));
}
return SDValue();
}
// Custom-legalize bitcasts from fixed-length vector types to scalar types
// thus: bitcast the vector to a one-element vector type whose element type
// is the same as the result type, and extract the first element.
if (!VT.isVector() && Op0VT.isFixedLengthVector()) {
EVT BVT = EVT::getVectorVT(*DAG.getContext(), VT, 1);
if (!isTypeLegal(BVT))
return SDValue();
SDValue BVec = DAG.getBitcast(BVT, Op0);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, BVec,
DAG.getVectorIdxConstant(0, DL));
}
return SDValue();
}
case ISD::INTRINSIC_WO_CHAIN:
return LowerINTRINSIC_WO_CHAIN(Op, DAG);
case ISD::INTRINSIC_W_CHAIN:
return LowerINTRINSIC_W_CHAIN(Op, DAG);
case ISD::INTRINSIC_VOID:
return LowerINTRINSIC_VOID(Op, DAG);
case ISD::IS_FPCLASS:
return LowerIS_FPCLASS(Op, DAG);
case ISD::BITREVERSE: {
MVT VT = Op.getSimpleValueType();
if (VT.isFixedLengthVector()) {
assert(Subtarget.hasStdExtZvbb());
return lowerToScalableOp(Op, DAG);
}
SDLoc DL(Op);
assert(Subtarget.hasStdExtZbkb() && "Unexpected custom legalization");
assert(Op.getOpcode() == ISD::BITREVERSE && "Unexpected opcode");
// Expand bitreverse to a bswap(rev8) followed by brev8.
SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, Op.getOperand(0));
return DAG.getNode(RISCVISD::BREV8, DL, VT, BSwap);
}
case ISD::TRUNCATE:
// Only custom-lower vector truncates
if (!Op.getSimpleValueType().isVector())
return Op;
return lowerVectorTruncLike(Op, DAG);
case ISD::ANY_EXTEND:
case ISD::ZERO_EXTEND:
if (Op.getOperand(0).getValueType().isVector() &&
Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ 1);
return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VZEXT_VL);
case ISD::SIGN_EXTEND:
if (Op.getOperand(0).getValueType().isVector() &&
Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ -1);
return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VSEXT_VL);
case ISD::SPLAT_VECTOR_PARTS:
return lowerSPLAT_VECTOR_PARTS(Op, DAG);
case ISD::INSERT_VECTOR_ELT:
return lowerINSERT_VECTOR_ELT(Op, DAG);
case ISD::EXTRACT_VECTOR_ELT:
return lowerEXTRACT_VECTOR_ELT(Op, DAG);
case ISD::SCALAR_TO_VECTOR: {
MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
SDValue Scalar = Op.getOperand(0);
if (VT.getVectorElementType() == MVT::i1) {
MVT WideVT = VT.changeVectorElementType(MVT::i8);
SDValue V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, WideVT, Scalar);
return DAG.getNode(ISD::TRUNCATE, DL, VT, V);
}
MVT ContainerVT = VT;
if (VT.isFixedLengthVector())
ContainerVT = getContainerForFixedLengthVector(VT);
SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
Scalar = DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getXLenVT(), Scalar);
SDValue V = DAG.getNode(RISCVISD::VMV_S_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), Scalar, VL);
if (VT.isFixedLengthVector())
V = convertFromScalableVector(VT, V, DAG, Subtarget);
return V;
}
case ISD::VSCALE: {
MVT XLenVT = Subtarget.getXLenVT();
MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
SDValue Res = DAG.getNode(RISCVISD::READ_VLENB, DL, XLenVT);
// We define our scalable vector types for lmul=1 to use a 64 bit known
// minimum size. e.g. <vscale x 2 x i32>. VLENB is in bytes so we calculate
// vscale as VLENB / 8.
static_assert(RISCV::RVVBitsPerBlock == 64, "Unexpected bits per block!");
if (Subtarget.getRealMinVLen() < RISCV::RVVBitsPerBlock)
report_fatal_error("Support for VLEN==32 is incomplete.");
// We assume VLENB is a multiple of 8. We manually choose the best shift
// here because SimplifyDemandedBits isn't always able to simplify it.
uint64_t Val = Op.getConstantOperandVal(0);
if (isPowerOf2_64(Val)) {
uint64_t Log2 = Log2_64(Val);
if (Log2 < 3)
Res = DAG.getNode(ISD::SRL, DL, XLenVT, Res,
DAG.getConstant(3 - Log2, DL, VT));
else if (Log2 > 3)
Res = DAG.getNode(ISD::SHL, DL, XLenVT, Res,
DAG.getConstant(Log2 - 3, DL, XLenVT));
} else if ((Val % 8) == 0) {
// If the multiplier is a multiple of 8, scale it down to avoid needing
// to shift the VLENB value.
Res = DAG.getNode(ISD::MUL, DL, XLenVT, Res,
DAG.getConstant(Val / 8, DL, XLenVT));
} else {
SDValue VScale = DAG.getNode(ISD::SRL, DL, XLenVT, Res,
DAG.getConstant(3, DL, XLenVT));
Res = DAG.getNode(ISD::MUL, DL, XLenVT, VScale,
DAG.getConstant(Val, DL, XLenVT));
}
return DAG.getNode(ISD::TRUNCATE, DL, VT, Res);
}
case ISD::FPOWI: {
// Custom promote f16 powi with illegal i32 integer type on RV64. Once
// promoted this will be legalized into a libcall by LegalizeIntegerTypes.
if (Op.getValueType() == MVT::f16 && Subtarget.is64Bit() &&
Op.getOperand(1).getValueType() == MVT::i32) {
SDLoc DL(Op);
SDValue Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op.getOperand(0));
SDValue Powi =
DAG.getNode(ISD::FPOWI, DL, MVT::f32, Op0, Op.getOperand(1));
return DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, Powi,
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
}
return SDValue();
}
case ISD::FMAXIMUM:
case ISD::FMINIMUM:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVectorOp(Op, DAG);
return lowerFMAXIMUM_FMINIMUM(Op, DAG, Subtarget);
case ISD::FP_EXTEND: {
SDLoc DL(Op);
EVT VT = Op.getValueType();
SDValue Op0 = Op.getOperand(0);
EVT Op0VT = Op0.getValueType();
if (VT == MVT::f32 && Op0VT == MVT::bf16 && Subtarget.hasStdExtZfbfmin())
return DAG.getNode(RISCVISD::FP_EXTEND_BF16, DL, MVT::f32, Op0);
if (VT == MVT::f64 && Op0VT == MVT::bf16 && Subtarget.hasStdExtZfbfmin()) {
SDValue FloatVal =
DAG.getNode(RISCVISD::FP_EXTEND_BF16, DL, MVT::f32, Op0);
return DAG.getNode(ISD::FP_EXTEND, DL, MVT::f64, FloatVal);
}
if (!Op.getValueType().isVector())
return Op;
return lowerVectorFPExtendOrRoundLike(Op, DAG);
}
case ISD::FP_ROUND: {
SDLoc DL(Op);
EVT VT = Op.getValueType();
SDValue Op0 = Op.getOperand(0);
EVT Op0VT = Op0.getValueType();
if (VT == MVT::bf16 && Op0VT == MVT::f32 && Subtarget.hasStdExtZfbfmin())
return DAG.getNode(RISCVISD::FP_ROUND_BF16, DL, MVT::bf16, Op0);
if (VT == MVT::bf16 && Op0VT == MVT::f64 && Subtarget.hasStdExtZfbfmin() &&
Subtarget.hasStdExtDOrZdinx()) {
SDValue FloatVal =
DAG.getNode(ISD::FP_ROUND, DL, MVT::f32, Op0,
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
return DAG.getNode(RISCVISD::FP_ROUND_BF16, DL, MVT::bf16, FloatVal);
}
if (!Op.getValueType().isVector())
return Op;
return lowerVectorFPExtendOrRoundLike(Op, DAG);
}
case ISD::STRICT_FP_ROUND:
case ISD::STRICT_FP_EXTEND:
return lowerStrictFPExtendOrRoundLike(Op, DAG);
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP:
if (Op.getValueType().isVector() &&
Op.getValueType().getScalarType() == MVT::f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16())) {
if (Op.getValueType() == MVT::nxv32f16)
return SplitVectorOp(Op, DAG);
// int -> f32
SDLoc DL(Op);
MVT NVT =
MVT::getVectorVT(MVT::f32, Op.getValueType().getVectorElementCount());
SDValue NC = DAG.getNode(Op.getOpcode(), DL, NVT, Op->ops());
// f32 -> f16
return DAG.getNode(ISD::FP_ROUND, DL, Op.getValueType(), NC,
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
}
[[fallthrough]];
case ISD::FP_TO_SINT:
case ISD::FP_TO_UINT:
if (SDValue Op1 = Op.getOperand(0);
Op1.getValueType().isVector() &&
Op1.getValueType().getScalarType() == MVT::f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16())) {
if (Op1.getValueType() == MVT::nxv32f16)
return SplitVectorOp(Op, DAG);
// f16 -> f32
SDLoc DL(Op);
MVT NVT = MVT::getVectorVT(MVT::f32,
Op1.getValueType().getVectorElementCount());
SDValue WidenVec = DAG.getNode(ISD::FP_EXTEND, DL, NVT, Op1);
// f32 -> int
return DAG.getNode(Op.getOpcode(), DL, Op.getValueType(), WidenVec);
}
[[fallthrough]];
case ISD::STRICT_FP_TO_SINT:
case ISD::STRICT_FP_TO_UINT:
case ISD::STRICT_SINT_TO_FP:
case ISD::STRICT_UINT_TO_FP: {
// RVV can only do fp<->int conversions to types half/double the size as
// the source. We custom-lower any conversions that do two hops into
// sequences.
MVT VT = Op.getSimpleValueType();
if (!VT.isVector())
return Op;
SDLoc DL(Op);
bool IsStrict = Op->isStrictFPOpcode();
SDValue Src = Op.getOperand(0 + IsStrict);
MVT EltVT = VT.getVectorElementType();
MVT SrcVT = Src.getSimpleValueType();
MVT SrcEltVT = SrcVT.getVectorElementType();
unsigned EltSize = EltVT.getSizeInBits();
unsigned SrcEltSize = SrcEltVT.getSizeInBits();
assert(isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) &&
"Unexpected vector element types");
bool IsInt2FP = SrcEltVT.isInteger();
// Widening conversions
if (EltSize > (2 * SrcEltSize)) {
if (IsInt2FP) {
// Do a regular integer sign/zero extension then convert to float.
MVT IVecVT = MVT::getVectorVT(MVT::getIntegerVT(EltSize / 2),
VT.getVectorElementCount());
unsigned ExtOpcode = (Op.getOpcode() == ISD::UINT_TO_FP ||
Op.getOpcode() == ISD::STRICT_UINT_TO_FP)
? ISD::ZERO_EXTEND
: ISD::SIGN_EXTEND;
SDValue Ext = DAG.getNode(ExtOpcode, DL, IVecVT, Src);
if (IsStrict)
return DAG.getNode(Op.getOpcode(), DL, Op->getVTList(),
Op.getOperand(0), Ext);
return DAG.getNode(Op.getOpcode(), DL, VT, Ext);
}
// FP2Int
assert(SrcEltVT == MVT::f16 && "Unexpected FP_TO_[US]INT lowering");
// Do one doubling fp_extend then complete the operation by converting
// to int.
MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
if (IsStrict) {
auto [FExt, Chain] =
DAG.getStrictFPExtendOrRound(Src, Op.getOperand(0), DL, InterimFVT);
return DAG.getNode(Op.getOpcode(), DL, Op->getVTList(), Chain, FExt);
}
SDValue FExt = DAG.getFPExtendOrRound(Src, DL, InterimFVT);
return DAG.getNode(Op.getOpcode(), DL, VT, FExt);
}
// Narrowing conversions
if (SrcEltSize > (2 * EltSize)) {
if (IsInt2FP) {
// One narrowing int_to_fp, then an fp_round.
assert(EltVT == MVT::f16 && "Unexpected [US]_TO_FP lowering");
MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
if (IsStrict) {
SDValue Int2FP = DAG.getNode(Op.getOpcode(), DL,
DAG.getVTList(InterimFVT, MVT::Other),
Op.getOperand(0), Src);
SDValue Chain = Int2FP.getValue(1);
return DAG.getStrictFPExtendOrRound(Int2FP, Chain, DL, VT).first;
}
SDValue Int2FP = DAG.getNode(Op.getOpcode(), DL, InterimFVT, Src);
return DAG.getFPExtendOrRound(Int2FP, DL, VT);
}
// FP2Int
// One narrowing fp_to_int, then truncate the integer. If the float isn't
// representable by the integer, the result is poison.
MVT IVecVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize / 2),
VT.getVectorElementCount());
if (IsStrict) {
SDValue FP2Int =
DAG.getNode(Op.getOpcode(), DL, DAG.getVTList(IVecVT, MVT::Other),
Op.getOperand(0), Src);
SDValue Res = DAG.getNode(ISD::TRUNCATE, DL, VT, FP2Int);
return DAG.getMergeValues({Res, FP2Int.getValue(1)}, DL);
}
SDValue FP2Int = DAG.getNode(Op.getOpcode(), DL, IVecVT, Src);
return DAG.getNode(ISD::TRUNCATE, DL, VT, FP2Int);
}
// Scalable vectors can exit here. Patterns will handle equally-sized
// conversions halving/doubling ones.
if (!VT.isFixedLengthVector())
return Op;
// For fixed-length vectors we lower to a custom "VL" node.
unsigned RVVOpc = 0;
switch (Op.getOpcode()) {
default:
llvm_unreachable("Impossible opcode");
case ISD::FP_TO_SINT:
RVVOpc = RISCVISD::VFCVT_RTZ_X_F_VL;
break;
case ISD::FP_TO_UINT:
RVVOpc = RISCVISD::VFCVT_RTZ_XU_F_VL;
break;
case ISD::SINT_TO_FP:
RVVOpc = RISCVISD::SINT_TO_FP_VL;
break;
case ISD::UINT_TO_FP:
RVVOpc = RISCVISD::UINT_TO_FP_VL;
break;
case ISD::STRICT_FP_TO_SINT:
RVVOpc = RISCVISD::STRICT_VFCVT_RTZ_X_F_VL;
break;
case ISD::STRICT_FP_TO_UINT:
RVVOpc = RISCVISD::STRICT_VFCVT_RTZ_XU_F_VL;
break;
case ISD::STRICT_SINT_TO_FP:
RVVOpc = RISCVISD::STRICT_SINT_TO_FP_VL;
break;
case ISD::STRICT_UINT_TO_FP:
RVVOpc = RISCVISD::STRICT_UINT_TO_FP_VL;
break;
}
MVT ContainerVT = getContainerForFixedLengthVector(VT);
MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
assert(ContainerVT.getVectorElementCount() == SrcContainerVT.getVectorElementCount() &&
"Expected same element count");
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
if (IsStrict) {
Src = DAG.getNode(RVVOpc, DL, DAG.getVTList(ContainerVT, MVT::Other),
Op.getOperand(0), Src, Mask, VL);
SDValue SubVec = convertFromScalableVector(VT, Src, DAG, Subtarget);
return DAG.getMergeValues({SubVec, Src.getValue(1)}, DL);
}
Src = DAG.getNode(RVVOpc, DL, ContainerVT, Src, Mask, VL);
return convertFromScalableVector(VT, Src, DAG, Subtarget);
}
case ISD::FP_TO_SINT_SAT:
case ISD::FP_TO_UINT_SAT:
return lowerFP_TO_INT_SAT(Op, DAG, Subtarget);
case ISD::FP_TO_BF16: {
// Custom lower to ensure the libcall return is passed in an FPR on hard
// float ABIs.
assert(!Subtarget.isSoftFPABI() && "Unexpected custom legalization");
SDLoc DL(Op);
MakeLibCallOptions CallOptions;
RTLIB::Libcall LC =
RTLIB::getFPROUND(Op.getOperand(0).getValueType(), MVT::bf16);
SDValue Res =
makeLibCall(DAG, LC, MVT::f32, Op.getOperand(0), CallOptions, DL).first;
if (Subtarget.is64Bit() && !RV64LegalI32)
return DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Res);
return DAG.getBitcast(MVT::i32, Res);
}
case ISD::BF16_TO_FP: {
assert(Subtarget.hasStdExtFOrZfinx() && "Unexpected custom legalization");
MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
Op = DAG.getNode(
ISD::SHL, DL, Op.getOperand(0).getValueType(), Op.getOperand(0),
DAG.getShiftAmountConstant(16, Op.getOperand(0).getValueType(), DL));
SDValue Res = Subtarget.is64Bit()
? DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, Op)
: DAG.getBitcast(MVT::f32, Op);
// fp_extend if the target VT is bigger than f32.
if (VT != MVT::f32)
return DAG.getNode(ISD::FP_EXTEND, DL, VT, Res);
return Res;
}
case ISD::FP_TO_FP16: {
// Custom lower to ensure the libcall return is passed in an FPR on hard
// float ABIs.
assert(Subtarget.hasStdExtFOrZfinx() && "Unexpected custom legalisation");
SDLoc DL(Op);
MakeLibCallOptions CallOptions;
RTLIB::Libcall LC =
RTLIB::getFPROUND(Op.getOperand(0).getValueType(), MVT::f16);
SDValue Res =
makeLibCall(DAG, LC, MVT::f32, Op.getOperand(0), CallOptions, DL).first;
if (Subtarget.is64Bit() && !RV64LegalI32)
return DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Res);
return DAG.getBitcast(MVT::i32, Res);
}
case ISD::FP16_TO_FP: {
// Custom lower to ensure the libcall argument is passed in an FPR on hard
// float ABIs.
assert(Subtarget.hasStdExtFOrZfinx() && "Unexpected custom legalisation");
SDLoc DL(Op);
MakeLibCallOptions CallOptions;
SDValue Arg = Subtarget.is64Bit()
? DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32,
Op.getOperand(0))
: DAG.getBitcast(MVT::f32, Op.getOperand(0));
SDValue Res =
makeLibCall(DAG, RTLIB::FPEXT_F16_F32, MVT::f32, Arg, CallOptions, DL)
.first;
return Res;
}
case ISD::FTRUNC:
case ISD::FCEIL:
case ISD::FFLOOR:
case ISD::FNEARBYINT:
case ISD::FRINT:
case ISD::FROUND:
case ISD::FROUNDEVEN:
return lowerFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
case ISD::LRINT:
case ISD::LLRINT:
return lowerVectorXRINT(Op, DAG, Subtarget);
case ISD::VECREDUCE_ADD:
case ISD::VECREDUCE_UMAX:
case ISD::VECREDUCE_SMAX:
case ISD::VECREDUCE_UMIN:
case ISD::VECREDUCE_SMIN:
return lowerVECREDUCE(Op, DAG);
case ISD::VECREDUCE_AND:
case ISD::VECREDUCE_OR:
case ISD::VECREDUCE_XOR:
if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
return lowerVectorMaskVecReduction(Op, DAG, /*IsVP*/ false);
return lowerVECREDUCE(Op, DAG);
case ISD::VECREDUCE_FADD:
case ISD::VECREDUCE_SEQ_FADD:
case ISD::VECREDUCE_FMIN:
case ISD::VECREDUCE_FMAX:
case ISD::VECREDUCE_FMAXIMUM:
case ISD::VECREDUCE_FMINIMUM:
return lowerFPVECREDUCE(Op, DAG);
case ISD::VP_REDUCE_ADD:
case ISD::VP_REDUCE_UMAX:
case ISD::VP_REDUCE_SMAX:
case ISD::VP_REDUCE_UMIN:
case ISD::VP_REDUCE_SMIN:
case ISD::VP_REDUCE_FADD:
case ISD::VP_REDUCE_SEQ_FADD:
case ISD::VP_REDUCE_FMIN:
case ISD::VP_REDUCE_FMAX:
if (Op.getOperand(1).getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVectorReductionOp(Op, DAG);
return lowerVPREDUCE(Op, DAG);
case ISD::VP_REDUCE_AND:
case ISD::VP_REDUCE_OR:
case ISD::VP_REDUCE_XOR:
if (Op.getOperand(1).getValueType().getVectorElementType() == MVT::i1)
return lowerVectorMaskVecReduction(Op, DAG, /*IsVP*/ true);
return lowerVPREDUCE(Op, DAG);
case ISD::UNDEF: {
MVT ContainerVT = getContainerForFixedLengthVector(Op.getSimpleValueType());
return convertFromScalableVector(Op.getSimpleValueType(),
DAG.getUNDEF(ContainerVT), DAG, Subtarget);
}
case ISD::INSERT_SUBVECTOR:
return lowerINSERT_SUBVECTOR(Op, DAG);
case ISD::EXTRACT_SUBVECTOR:
return lowerEXTRACT_SUBVECTOR(Op, DAG);
case ISD::VECTOR_DEINTERLEAVE:
return lowerVECTOR_DEINTERLEAVE(Op, DAG);
case ISD::VECTOR_INTERLEAVE:
return lowerVECTOR_INTERLEAVE(Op, DAG);
case ISD::STEP_VECTOR:
return lowerSTEP_VECTOR(Op, DAG);
case ISD::VECTOR_REVERSE:
return lowerVECTOR_REVERSE(Op, DAG);
case ISD::VECTOR_SPLICE:
return lowerVECTOR_SPLICE(Op, DAG);
case ISD::BUILD_VECTOR:
return lowerBUILD_VECTOR(Op, DAG, Subtarget);
case ISD::SPLAT_VECTOR:
if (Op.getValueType().getScalarType() == MVT::f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16())) {
if (Op.getValueType() == MVT::nxv32f16)
return SplitVectorOp(Op, DAG);
SDLoc DL(Op);
SDValue NewScalar =
DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op.getOperand(0));
SDValue NewSplat = DAG.getNode(
ISD::SPLAT_VECTOR, DL,
MVT::getVectorVT(MVT::f32, Op.getValueType().getVectorElementCount()),
NewScalar);
return DAG.getNode(ISD::FP_ROUND, DL, Op.getValueType(), NewSplat,
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
}
if (Op.getValueType().getVectorElementType() == MVT::i1)
return lowerVectorMaskSplat(Op, DAG);
return SDValue();
case ISD::VECTOR_SHUFFLE:
return lowerVECTOR_SHUFFLE(Op, DAG, Subtarget);
case ISD::CONCAT_VECTORS: {
// Split CONCAT_VECTORS into a series of INSERT_SUBVECTOR nodes. This is
// better than going through the stack, as the default expansion does.
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector())
ContainerVT = ::getContainerForFixedLengthVector(DAG, VT, Subtarget);
// Recursively split concat_vectors with more than 2 operands:
//
// concat_vector op1, op2, op3, op4
// ->
// concat_vector (concat_vector op1, op2), (concat_vector op3, op4)
//
// This reduces the length of the chain of vslideups and allows us to
// perform the vslideups at a smaller LMUL, limited to MF2.
if (Op.getNumOperands() > 2 &&
ContainerVT.bitsGE(getLMUL1VT(ContainerVT))) {
MVT HalfVT = VT.getHalfNumVectorElementsVT();
assert(isPowerOf2_32(Op.getNumOperands()));
size_t HalfNumOps = Op.getNumOperands() / 2;
SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, DL, HalfVT,
Op->ops().take_front(HalfNumOps));
SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, DL, HalfVT,
Op->ops().drop_front(HalfNumOps));
return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
}
unsigned NumOpElts =
Op.getOperand(0).getSimpleValueType().getVectorMinNumElements();
SDValue Vec = DAG.getUNDEF(VT);
for (const auto &OpIdx : enumerate(Op->ops())) {
SDValue SubVec = OpIdx.value();
// Don't insert undef subvectors.
if (SubVec.isUndef())
continue;
Vec =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Vec, SubVec,
DAG.getVectorIdxConstant(OpIdx.index() * NumOpElts, DL));
}
return Vec;
}
case ISD::LOAD:
if (auto V = expandUnalignedRVVLoad(Op, DAG))
return V;
if (Op.getValueType().isFixedLengthVector())
return lowerFixedLengthVectorLoadToRVV(Op, DAG);
return Op;
case ISD::STORE:
if (auto V = expandUnalignedRVVStore(Op, DAG))
return V;
if (Op.getOperand(1).getValueType().isFixedLengthVector())
return lowerFixedLengthVectorStoreToRVV(Op, DAG);
return Op;
case ISD::MLOAD:
case ISD::VP_LOAD:
return lowerMaskedLoad(Op, DAG);
case ISD::MSTORE:
case ISD::VP_STORE:
return lowerMaskedStore(Op, DAG);
case ISD::SELECT_CC: {
// This occurs because we custom legalize SETGT and SETUGT for setcc. That
// causes LegalizeDAG to think we need to custom legalize select_cc. Expand
// into separate SETCC+SELECT just like LegalizeDAG.
SDValue Tmp1 = Op.getOperand(0);
SDValue Tmp2 = Op.getOperand(1);
SDValue True = Op.getOperand(2);
SDValue False = Op.getOperand(3);
EVT VT = Op.getValueType();
SDValue CC = Op.getOperand(4);
EVT CmpVT = Tmp1.getValueType();
EVT CCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT);
SDLoc DL(Op);
SDValue Cond =
DAG.getNode(ISD::SETCC, DL, CCVT, Tmp1, Tmp2, CC, Op->getFlags());
return DAG.getSelect(DL, VT, Cond, True, False);
}
case ISD::SETCC: {
MVT OpVT = Op.getOperand(0).getSimpleValueType();
if (OpVT.isScalarInteger()) {
MVT VT = Op.getSimpleValueType();
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
ISD::CondCode CCVal = cast<CondCodeSDNode>(Op.getOperand(2))->get();
assert((CCVal == ISD::SETGT || CCVal == ISD::SETUGT) &&
"Unexpected CondCode");
SDLoc DL(Op);
// If the RHS is a constant in the range [-2049, 0) or (0, 2046], we can
// convert this to the equivalent of (set(u)ge X, C+1) by using
// (xori (slti(u) X, C+1), 1). This avoids materializing a small constant
// in a register.
if (isa<ConstantSDNode>(RHS)) {
int64_t Imm = cast<ConstantSDNode>(RHS)->getSExtValue();
if (Imm != 0 && isInt<12>((uint64_t)Imm + 1)) {
// If this is an unsigned compare and the constant is -1, incrementing
// the constant would change behavior. The result should be false.
if (CCVal == ISD::SETUGT && Imm == -1)
return DAG.getConstant(0, DL, VT);
// Using getSetCCSwappedOperands will convert SET(U)GT->SET(U)LT.
CCVal = ISD::getSetCCSwappedOperands(CCVal);
SDValue SetCC = DAG.getSetCC(
DL, VT, LHS, DAG.getConstant(Imm + 1, DL, OpVT), CCVal);
return DAG.getLogicalNOT(DL, SetCC, VT);
}
}
// Not a constant we could handle, swap the operands and condition code to
// SETLT/SETULT.
CCVal = ISD::getSetCCSwappedOperands(CCVal);
return DAG.getSetCC(DL, VT, RHS, LHS, CCVal);
}
if (Op.getOperand(0).getSimpleValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVectorOp(Op, DAG);
return lowerFixedLengthVectorSetccToRVV(Op, DAG);
}
case ISD::ADD:
case ISD::SUB:
case ISD::MUL:
case ISD::MULHS:
case ISD::MULHU:
case ISD::AND:
case ISD::OR:
case ISD::XOR:
case ISD::SDIV:
case ISD::SREM:
case ISD::UDIV:
case ISD::UREM:
case ISD::BSWAP:
case ISD::CTPOP:
return lowerToScalableOp(Op, DAG);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL:
if (Op.getSimpleValueType().isFixedLengthVector())
return lowerToScalableOp(Op, DAG);
// This can be called for an i32 shift amount that needs to be promoted.
assert(Op.getOperand(1).getValueType() == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
return SDValue();
case ISD::FADD:
case ISD::FSUB:
case ISD::FMUL:
case ISD::FDIV:
case ISD::FNEG:
case ISD::FABS:
case ISD::FSQRT:
case ISD::FMA:
case ISD::FMINNUM:
case ISD::FMAXNUM:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVectorOp(Op, DAG);
[[fallthrough]];
case ISD::AVGFLOORU:
case ISD::AVGCEILU:
case ISD::SMIN:
case ISD::SMAX:
case ISD::UMIN:
case ISD::UMAX:
return lowerToScalableOp(Op, DAG);
case ISD::UADDSAT:
case ISD::USUBSAT:
if (!Op.getValueType().isVector())
return lowerUADDSAT_USUBSAT(Op, DAG);
return lowerToScalableOp(Op, DAG);
case ISD::SADDSAT:
case ISD::SSUBSAT:
if (!Op.getValueType().isVector())
return lowerSADDSAT_SSUBSAT(Op, DAG);
return lowerToScalableOp(Op, DAG);
case ISD::ABDS:
case ISD::ABDU: {
SDLoc dl(Op);
EVT VT = Op->getValueType(0);
SDValue LHS = DAG.getFreeze(Op->getOperand(0));
SDValue RHS = DAG.getFreeze(Op->getOperand(1));
bool IsSigned = Op->getOpcode() == ISD::ABDS;
// abds(lhs, rhs) -> sub(smax(lhs,rhs), smin(lhs,rhs))
// abdu(lhs, rhs) -> sub(umax(lhs,rhs), umin(lhs,rhs))
unsigned MaxOpc = IsSigned ? ISD::SMAX : ISD::UMAX;
unsigned MinOpc = IsSigned ? ISD::SMIN : ISD::UMIN;
SDValue Max = DAG.getNode(MaxOpc, dl, VT, LHS, RHS);
SDValue Min = DAG.getNode(MinOpc, dl, VT, LHS, RHS);
return DAG.getNode(ISD::SUB, dl, VT, Max, Min);
}
case ISD::ABS:
case ISD::VP_ABS:
return lowerABS(Op, DAG);
case ISD::CTLZ:
case ISD::CTLZ_ZERO_UNDEF:
case ISD::CTTZ:
case ISD::CTTZ_ZERO_UNDEF:
if (Subtarget.hasStdExtZvbb())
return lowerToScalableOp(Op, DAG);
assert(Op.getOpcode() != ISD::CTTZ);
return lowerCTLZ_CTTZ_ZERO_UNDEF(Op, DAG);
case ISD::VSELECT:
return lowerFixedLengthVectorSelectToRVV(Op, DAG);
case ISD::FCOPYSIGN:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVectorOp(Op, DAG);
return lowerFixedLengthVectorFCOPYSIGNToRVV(Op, DAG);
case ISD::STRICT_FADD:
case ISD::STRICT_FSUB:
case ISD::STRICT_FMUL:
case ISD::STRICT_FDIV:
case ISD::STRICT_FSQRT:
case ISD::STRICT_FMA:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitStrictFPVectorOp(Op, DAG);
return lowerToScalableOp(Op, DAG);
case ISD::STRICT_FSETCC:
case ISD::STRICT_FSETCCS:
return lowerVectorStrictFSetcc(Op, DAG);
case ISD::STRICT_FCEIL:
case ISD::STRICT_FRINT:
case ISD::STRICT_FFLOOR:
case ISD::STRICT_FTRUNC:
case ISD::STRICT_FNEARBYINT:
case ISD::STRICT_FROUND:
case ISD::STRICT_FROUNDEVEN:
return lowerVectorStrictFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
case ISD::MGATHER:
case ISD::VP_GATHER:
return lowerMaskedGather(Op, DAG);
case ISD::MSCATTER:
case ISD::VP_SCATTER:
return lowerMaskedScatter(Op, DAG);
case ISD::GET_ROUNDING:
return lowerGET_ROUNDING(Op, DAG);
case ISD::SET_ROUNDING:
return lowerSET_ROUNDING(Op, DAG);
case ISD::EH_DWARF_CFA:
return lowerEH_DWARF_CFA(Op, DAG);
case ISD::VP_SELECT:
case ISD::VP_MERGE:
case ISD::VP_ADD:
case ISD::VP_SUB:
case ISD::VP_MUL:
case ISD::VP_SDIV:
case ISD::VP_UDIV:
case ISD::VP_SREM:
case ISD::VP_UREM:
case ISD::VP_UADDSAT:
case ISD::VP_USUBSAT:
case ISD::VP_SADDSAT:
case ISD::VP_SSUBSAT:
case ISD::VP_LRINT:
case ISD::VP_LLRINT:
return lowerVPOp(Op, DAG);
case ISD::VP_AND:
case ISD::VP_OR:
case ISD::VP_XOR:
return lowerLogicVPOp(Op, DAG);
case ISD::VP_FADD:
case ISD::VP_FSUB:
case ISD::VP_FMUL:
case ISD::VP_FDIV:
case ISD::VP_FNEG:
case ISD::VP_FABS:
case ISD::VP_SQRT:
case ISD::VP_FMA:
case ISD::VP_FMINNUM:
case ISD::VP_FMAXNUM:
case ISD::VP_FCOPYSIGN:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVPOp(Op, DAG);
[[fallthrough]];
case ISD::VP_ASHR:
case ISD::VP_LSHR:
case ISD::VP_SHL:
return lowerVPOp(Op, DAG);
case ISD::VP_IS_FPCLASS:
return LowerIS_FPCLASS(Op, DAG);
case ISD::VP_SIGN_EXTEND:
case ISD::VP_ZERO_EXTEND:
if (Op.getOperand(0).getSimpleValueType().getVectorElementType() == MVT::i1)
return lowerVPExtMaskOp(Op, DAG);
return lowerVPOp(Op, DAG);
case ISD::VP_TRUNCATE:
return lowerVectorTruncLike(Op, DAG);
case ISD::VP_FP_EXTEND:
case ISD::VP_FP_ROUND:
return lowerVectorFPExtendOrRoundLike(Op, DAG);
case ISD::VP_SINT_TO_FP:
case ISD::VP_UINT_TO_FP:
if (Op.getValueType().isVector() &&
Op.getValueType().getScalarType() == MVT::f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16())) {
if (Op.getValueType() == MVT::nxv32f16)
return SplitVPOp(Op, DAG);
// int -> f32
SDLoc DL(Op);
MVT NVT =
MVT::getVectorVT(MVT::f32, Op.getValueType().getVectorElementCount());
auto NC = DAG.getNode(Op.getOpcode(), DL, NVT, Op->ops());
// f32 -> f16
return DAG.getNode(ISD::FP_ROUND, DL, Op.getValueType(), NC,
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
}
[[fallthrough]];
case ISD::VP_FP_TO_SINT:
case ISD::VP_FP_TO_UINT:
if (SDValue Op1 = Op.getOperand(0);
Op1.getValueType().isVector() &&
Op1.getValueType().getScalarType() == MVT::f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16())) {
if (Op1.getValueType() == MVT::nxv32f16)
return SplitVPOp(Op, DAG);
// f16 -> f32
SDLoc DL(Op);
MVT NVT = MVT::getVectorVT(MVT::f32,
Op1.getValueType().getVectorElementCount());
SDValue WidenVec = DAG.getNode(ISD::FP_EXTEND, DL, NVT, Op1);
// f32 -> int
return DAG.getNode(Op.getOpcode(), DL, Op.getValueType(),
{WidenVec, Op.getOperand(1), Op.getOperand(2)});
}
return lowerVPFPIntConvOp(Op, DAG);
case ISD::VP_SETCC:
if (Op.getOperand(0).getSimpleValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVPOp(Op, DAG);
if (Op.getOperand(0).getSimpleValueType().getVectorElementType() == MVT::i1)
return lowerVPSetCCMaskOp(Op, DAG);
[[fallthrough]];
case ISD::VP_SMIN:
case ISD::VP_SMAX:
case ISD::VP_UMIN:
case ISD::VP_UMAX:
case ISD::VP_BITREVERSE:
case ISD::VP_BSWAP:
return lowerVPOp(Op, DAG);
case ISD::VP_CTLZ:
case ISD::VP_CTLZ_ZERO_UNDEF:
if (Subtarget.hasStdExtZvbb())
return lowerVPOp(Op, DAG);
return lowerCTLZ_CTTZ_ZERO_UNDEF(Op, DAG);
case ISD::VP_CTTZ:
case ISD::VP_CTTZ_ZERO_UNDEF:
if (Subtarget.hasStdExtZvbb())
return lowerVPOp(Op, DAG);
return lowerCTLZ_CTTZ_ZERO_UNDEF(Op, DAG);
case ISD::VP_CTPOP:
return lowerVPOp(Op, DAG);
case ISD::EXPERIMENTAL_VP_STRIDED_LOAD:
return lowerVPStridedLoad(Op, DAG);
case ISD::EXPERIMENTAL_VP_STRIDED_STORE:
return lowerVPStridedStore(Op, DAG);
case ISD::VP_FCEIL:
case ISD::VP_FFLOOR:
case ISD::VP_FRINT:
case ISD::VP_FNEARBYINT:
case ISD::VP_FROUND:
case ISD::VP_FROUNDEVEN:
case ISD::VP_FROUNDTOZERO:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVPOp(Op, DAG);
return lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
case ISD::VP_FMAXIMUM:
case ISD::VP_FMINIMUM:
if (Op.getValueType() == MVT::nxv32f16 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SplitVPOp(Op, DAG);
return lowerFMAXIMUM_FMINIMUM(Op, DAG, Subtarget);
case ISD::EXPERIMENTAL_VP_SPLICE:
return lowerVPSpliceExperimental(Op, DAG);
case ISD::EXPERIMENTAL_VP_REVERSE:
return lowerVPReverseExperimental(Op, DAG);
}
}
static SDValue getTargetNode(GlobalAddressSDNode *N, const SDLoc &DL, EVT Ty,
SelectionDAG &DAG, unsigned Flags) {
return DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, Flags);
}
static SDValue getTargetNode(BlockAddressSDNode *N, const SDLoc &DL, EVT Ty,
SelectionDAG &DAG, unsigned Flags) {
return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, N->getOffset(),
Flags);
}
static SDValue getTargetNode(ConstantPoolSDNode *N, const SDLoc &DL, EVT Ty,
SelectionDAG &DAG, unsigned Flags) {
return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(),
N->getOffset(), Flags);
}
static SDValue getTargetNode(JumpTableSDNode *N, const SDLoc &DL, EVT Ty,
SelectionDAG &DAG, unsigned Flags) {
return DAG.getTargetJumpTable(N->getIndex(), Ty, Flags);
}
template <class NodeTy>
SDValue RISCVTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
bool IsLocal, bool IsExternWeak) const {
SDLoc DL(N);
EVT Ty = getPointerTy(DAG.getDataLayout());
// When HWASAN is used and tagging of global variables is enabled
// they should be accessed via the GOT, since the tagged address of a global
// is incompatible with existing code models. This also applies to non-pic
// mode.
if (isPositionIndependent() || Subtarget.allowTaggedGlobals()) {
SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
if (IsLocal && !Subtarget.allowTaggedGlobals())
// Use PC-relative addressing to access the symbol. This generates the
// pattern (PseudoLLA sym), which expands to (addi (auipc %pcrel_hi(sym))
// %pcrel_lo(auipc)).
return DAG.getNode(RISCVISD::LLA, DL, Ty, Addr);
// Use PC-relative addressing to access the GOT for this symbol, then load
// the address from the GOT. This generates the pattern (PseudoLGA sym),
// which expands to (ld (addi (auipc %got_pcrel_hi(sym)) %pcrel_lo(auipc))).
SDValue Load =
SDValue(DAG.getMachineNode(RISCV::PseudoLGA, DL, Ty, Addr), 0);
MachineFunction &MF = DAG.getMachineFunction();
MachineMemOperand *MemOp = MF.getMachineMemOperand(
MachinePointerInfo::getGOT(MF),
MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
MachineMemOperand::MOInvariant,
LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
DAG.setNodeMemRefs(cast<MachineSDNode>(Load.getNode()), {MemOp});
return Load;
}
switch (getTargetMachine().getCodeModel()) {
default:
report_fatal_error("Unsupported code model for lowering");
case CodeModel::Small: {
// Generate a sequence for accessing addresses within the first 2 GiB of
// address space. This generates the pattern (addi (lui %hi(sym)) %lo(sym)).
SDValue AddrHi = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_HI);
SDValue AddrLo = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_LO);
SDValue MNHi = DAG.getNode(RISCVISD::HI, DL, Ty, AddrHi);
return DAG.getNode(RISCVISD::ADD_LO, DL, Ty, MNHi, AddrLo);
}
case CodeModel::Medium: {
SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
if (IsExternWeak) {
// An extern weak symbol may be undefined, i.e. have value 0, which may
// not be within 2GiB of PC, so use GOT-indirect addressing to access the
// symbol. This generates the pattern (PseudoLGA sym), which expands to
// (ld (addi (auipc %got_pcrel_hi(sym)) %pcrel_lo(auipc))).
SDValue Load =
SDValue(DAG.getMachineNode(RISCV::PseudoLGA, DL, Ty, Addr), 0);
MachineFunction &MF = DAG.getMachineFunction();
MachineMemOperand *MemOp = MF.getMachineMemOperand(
MachinePointerInfo::getGOT(MF),
MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
MachineMemOperand::MOInvariant,
LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
DAG.setNodeMemRefs(cast<MachineSDNode>(Load.getNode()), {MemOp});
return Load;
}
// Generate a sequence for accessing addresses within any 2GiB range within
// the address space. This generates the pattern (PseudoLLA sym), which
// expands to (addi (auipc %pcrel_hi(sym)) %pcrel_lo(auipc)).
return DAG.getNode(RISCVISD::LLA, DL, Ty, Addr);
}
}
}
SDValue RISCVTargetLowering::lowerGlobalAddress(SDValue Op,
SelectionDAG &DAG) const {
GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
assert(N->getOffset() == 0 && "unexpected offset in global node");
const GlobalValue *GV = N->getGlobal();
return getAddr(N, DAG, GV->isDSOLocal(), GV->hasExternalWeakLinkage());
}
SDValue RISCVTargetLowering::lowerBlockAddress(SDValue Op,
SelectionDAG &DAG) const {
BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
return getAddr(N, DAG);
}
SDValue RISCVTargetLowering::lowerConstantPool(SDValue Op,
SelectionDAG &DAG) const {
ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
return getAddr(N, DAG);
}
SDValue RISCVTargetLowering::lowerJumpTable(SDValue Op,
SelectionDAG &DAG) const {
JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
return getAddr(N, DAG);
}
SDValue RISCVTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N,
SelectionDAG &DAG,
bool UseGOT) const {
SDLoc DL(N);
EVT Ty = getPointerTy(DAG.getDataLayout());
const GlobalValue *GV = N->getGlobal();
MVT XLenVT = Subtarget.getXLenVT();
if (UseGOT) {
// Use PC-relative addressing to access the GOT for this TLS symbol, then
// load the address from the GOT and add the thread pointer. This generates
// the pattern (PseudoLA_TLS_IE sym), which expands to
// (ld (auipc %tls_ie_pcrel_hi(sym)) %pcrel_lo(auipc)).
SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
SDValue Load =
SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_IE, DL, Ty, Addr), 0);
MachineFunction &MF = DAG.getMachineFunction();
MachineMemOperand *MemOp = MF.getMachineMemOperand(
MachinePointerInfo::getGOT(MF),
MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
MachineMemOperand::MOInvariant,
LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
DAG.setNodeMemRefs(cast<MachineSDNode>(Load.getNode()), {MemOp});
// Add the thread pointer.
SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
return DAG.getNode(ISD::ADD, DL, Ty, Load, TPReg);
}
// Generate a sequence for accessing the address relative to the thread
// pointer, with the appropriate adjustment for the thread pointer offset.
// This generates the pattern
// (add (add_tprel (lui %tprel_hi(sym)) tp %tprel_add(sym)) %tprel_lo(sym))
SDValue AddrHi =
DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_HI);
SDValue AddrAdd =
DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_ADD);
SDValue AddrLo =
DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_LO);
SDValue MNHi = DAG.getNode(RISCVISD::HI, DL, Ty, AddrHi);
SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
SDValue MNAdd =
DAG.getNode(RISCVISD::ADD_TPREL, DL, Ty, MNHi, TPReg, AddrAdd);
return DAG.getNode(RISCVISD::ADD_LO, DL, Ty, MNAdd, AddrLo);
}
SDValue RISCVTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N,
SelectionDAG &DAG) const {
SDLoc DL(N);
EVT Ty = getPointerTy(DAG.getDataLayout());
IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits());
const GlobalValue *GV = N->getGlobal();
// Use a PC-relative addressing mode to access the global dynamic GOT address.
// This generates the pattern (PseudoLA_TLS_GD sym), which expands to
// (addi (auipc %tls_gd_pcrel_hi(sym)) %pcrel_lo(auipc)).
SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
SDValue Load =
SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_GD, DL, Ty, Addr), 0);
// Prepare argument list to generate call.
ArgListTy Args;
ArgListEntry Entry;
Entry.Node = Load;
Entry.Ty = CallTy;
Args.push_back(Entry);
// Setup call to __tls_get_addr.
TargetLowering::CallLoweringInfo CLI(DAG);
CLI.setDebugLoc(DL)
.setChain(DAG.getEntryNode())
.setLibCallee(CallingConv::C, CallTy,
DAG.getExternalSymbol("__tls_get_addr", Ty),
std::move(Args));
return LowerCallTo(CLI).first;
}
SDValue RISCVTargetLowering::getTLSDescAddr(GlobalAddressSDNode *N,
SelectionDAG &DAG) const {
SDLoc DL(N);
EVT Ty = getPointerTy(DAG.getDataLayout());
const GlobalValue *GV = N->getGlobal();
// Use a PC-relative addressing mode to access the global dynamic GOT address.
// This generates the pattern (PseudoLA_TLSDESC sym), which expands to
//
// auipc tX, %tlsdesc_hi(symbol) // R_RISCV_TLSDESC_HI20(symbol)
// lw tY, tX, %tlsdesc_load_lo(label) // R_RISCV_TLSDESC_LOAD_LO12(label)
// addi a0, tX, %tlsdesc_add_lo(label) // R_RISCV_TLSDESC_ADD_LO12(label)
// jalr t0, tY // R_RISCV_TLSDESC_CALL(label)
SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
return SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLSDESC, DL, Ty, Addr), 0);
}
SDValue RISCVTargetLowering::lowerGlobalTLSAddress(SDValue Op,
SelectionDAG &DAG) const {
GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
assert(N->getOffset() == 0 && "unexpected offset in global node");
if (DAG.getTarget().useEmulatedTLS())
return LowerToTLSEmulatedModel(N, DAG);
TLSModel::Model Model = getTargetMachine().getTLSModel(N->getGlobal());
if (DAG.getMachineFunction().getFunction().getCallingConv() ==
CallingConv::GHC)
report_fatal_error("In GHC calling convention TLS is not supported");
SDValue Addr;
switch (Model) {
case TLSModel::LocalExec:
Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/false);
break;
case TLSModel::InitialExec:
Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/true);
break;
case TLSModel::LocalDynamic:
case TLSModel::GeneralDynamic:
Addr = DAG.getTarget().useTLSDESC() ? getTLSDescAddr(N, DAG)
: getDynamicTLSAddr(N, DAG);
break;
}
return Addr;
}
// Return true if Val is equal to (setcc LHS, RHS, CC).
// Return false if Val is the inverse of (setcc LHS, RHS, CC).
// Otherwise, return std::nullopt.
static std::optional<bool> matchSetCC(SDValue LHS, SDValue RHS,
ISD::CondCode CC, SDValue Val) {
assert(Val->getOpcode() == ISD::SETCC);
SDValue LHS2 = Val.getOperand(0);
SDValue RHS2 = Val.getOperand(1);
ISD::CondCode CC2 = cast<CondCodeSDNode>(Val.getOperand(2))->get();
if (LHS == LHS2 && RHS == RHS2) {
if (CC == CC2)
return true;
if (CC == ISD::getSetCCInverse(CC2, LHS2.getValueType()))
return false;
} else if (LHS == RHS2 && RHS == LHS2) {
CC2 = ISD::getSetCCSwappedOperands(CC2);
if (CC == CC2)
return true;
if (CC == ISD::getSetCCInverse(CC2, LHS2.getValueType()))
return false;
}
return std::nullopt;
}
static SDValue combineSelectToBinOp(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue CondV = N->getOperand(0);
SDValue TrueV = N->getOperand(1);
SDValue FalseV = N->getOperand(2);
MVT VT = N->getSimpleValueType(0);
SDLoc DL(N);
if (!Subtarget.hasConditionalMoveFusion()) {
// (select c, -1, y) -> -c | y
if (isAllOnesConstant(TrueV)) {
SDValue Neg = DAG.getNegative(CondV, DL, VT);
return DAG.getNode(ISD::OR, DL, VT, Neg, DAG.getFreeze(FalseV));
}
// (select c, y, -1) -> (c-1) | y
if (isAllOnesConstant(FalseV)) {
SDValue Neg = DAG.getNode(ISD::ADD, DL, VT, CondV,
DAG.getAllOnesConstant(DL, VT));
return DAG.getNode(ISD::OR, DL, VT, Neg, DAG.getFreeze(TrueV));
}
// (select c, 0, y) -> (c-1) & y
if (isNullConstant(TrueV)) {
SDValue Neg = DAG.getNode(ISD::ADD, DL, VT, CondV,
DAG.getAllOnesConstant(DL, VT));
return DAG.getNode(ISD::AND, DL, VT, Neg, DAG.getFreeze(FalseV));
}
// (select c, y, 0) -> -c & y
if (isNullConstant(FalseV)) {
SDValue Neg = DAG.getNegative(CondV, DL, VT);
return DAG.getNode(ISD::AND, DL, VT, Neg, DAG.getFreeze(TrueV));
}
}
// select c, ~x, x --> xor -c, x
if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV)) {
const APInt &TrueVal = TrueV->getAsAPIntVal();
const APInt &FalseVal = FalseV->getAsAPIntVal();
if (~TrueVal == FalseVal) {
SDValue Neg = DAG.getNegative(CondV, DL, VT);
return DAG.getNode(ISD::XOR, DL, VT, Neg, FalseV);
}
}
// Try to fold (select (setcc lhs, rhs, cc), truev, falsev) into bitwise ops
// when both truev and falsev are also setcc.
if (CondV.getOpcode() == ISD::SETCC && TrueV.getOpcode() == ISD::SETCC &&
FalseV.getOpcode() == ISD::SETCC) {
SDValue LHS = CondV.getOperand(0);
SDValue RHS = CondV.getOperand(1);
ISD::CondCode CC = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
// (select x, x, y) -> x | y
// (select !x, x, y) -> x & y
if (std::optional<bool> MatchResult = matchSetCC(LHS, RHS, CC, TrueV)) {
return DAG.getNode(*MatchResult ? ISD::OR : ISD::AND, DL, VT, TrueV,
DAG.getFreeze(FalseV));
}
// (select x, y, x) -> x & y
// (select !x, y, x) -> x | y
if (std::optional<bool> MatchResult = matchSetCC(LHS, RHS, CC, FalseV)) {
return DAG.getNode(*MatchResult ? ISD::AND : ISD::OR, DL, VT,
DAG.getFreeze(TrueV), FalseV);
}
}
return SDValue();
}
// Transform `binOp (select cond, x, c0), c1` where `c0` and `c1` are constants
// into `select cond, binOp(x, c1), binOp(c0, c1)` if profitable.
// For now we only consider transformation profitable if `binOp(c0, c1)` ends up
// being `0` or `-1`. In such cases we can replace `select` with `and`.
// TODO: Should we also do this if `binOp(c0, c1)` is cheaper to materialize
// than `c0`?
static SDValue
foldBinOpIntoSelectIfProfitable(SDNode *BO, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (Subtarget.hasShortForwardBranchOpt())
return SDValue();
unsigned SelOpNo = 0;
SDValue Sel = BO->getOperand(0);
if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) {
SelOpNo = 1;
Sel = BO->getOperand(1);
}
if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse())
return SDValue();
unsigned ConstSelOpNo = 1;
unsigned OtherSelOpNo = 2;
if (!dyn_cast<ConstantSDNode>(Sel->getOperand(ConstSelOpNo))) {
ConstSelOpNo = 2;
OtherSelOpNo = 1;
}
SDValue ConstSelOp = Sel->getOperand(ConstSelOpNo);
ConstantSDNode *ConstSelOpNode = dyn_cast<ConstantSDNode>(ConstSelOp);
if (!ConstSelOpNode || ConstSelOpNode->isOpaque())
return SDValue();
SDValue ConstBinOp = BO->getOperand(SelOpNo ^ 1);
ConstantSDNode *ConstBinOpNode = dyn_cast<ConstantSDNode>(ConstBinOp);
if (!ConstBinOpNode || ConstBinOpNode->isOpaque())
return SDValue();
SDLoc DL(Sel);
EVT VT = BO->getValueType(0);
SDValue NewConstOps[2] = {ConstSelOp, ConstBinOp};
if (SelOpNo == 1)
std::swap(NewConstOps[0], NewConstOps[1]);
SDValue NewConstOp =
DAG.FoldConstantArithmetic(BO->getOpcode(), DL, VT, NewConstOps);
if (!NewConstOp)
return SDValue();
const APInt &NewConstAPInt = NewConstOp->getAsAPIntVal();
if (!NewConstAPInt.isZero() && !NewConstAPInt.isAllOnes())
return SDValue();
SDValue OtherSelOp = Sel->getOperand(OtherSelOpNo);
SDValue NewNonConstOps[2] = {OtherSelOp, ConstBinOp};
if (SelOpNo == 1)
std::swap(NewNonConstOps[0], NewNonConstOps[1]);
SDValue NewNonConstOp = DAG.getNode(BO->getOpcode(), DL, VT, NewNonConstOps);
SDValue NewT = (ConstSelOpNo == 1) ? NewConstOp : NewNonConstOp;
SDValue NewF = (ConstSelOpNo == 1) ? NewNonConstOp : NewConstOp;
return DAG.getSelect(DL, VT, Sel.getOperand(0), NewT, NewF);
}
SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const {
SDValue CondV = Op.getOperand(0);
SDValue TrueV = Op.getOperand(1);
SDValue FalseV = Op.getOperand(2);
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
// Lower vector SELECTs to VSELECTs by splatting the condition.
if (VT.isVector()) {
MVT SplatCondVT = VT.changeVectorElementType(MVT::i1);
SDValue CondSplat = DAG.getSplat(SplatCondVT, DL, CondV);
return DAG.getNode(ISD::VSELECT, DL, VT, CondSplat, TrueV, FalseV);
}
// When Zicond or XVentanaCondOps is present, emit CZERO_EQZ and CZERO_NEZ
// nodes to implement the SELECT. Performing the lowering here allows for
// greater control over when CZERO_{EQZ/NEZ} are used vs another branchless
// sequence or RISCVISD::SELECT_CC node (branch-based select).
if ((Subtarget.hasStdExtZicond() || Subtarget.hasVendorXVentanaCondOps()) &&
VT.isScalarInteger()) {
// (select c, t, 0) -> (czero_eqz t, c)
if (isNullConstant(FalseV))
return DAG.getNode(RISCVISD::CZERO_EQZ, DL, VT, TrueV, CondV);
// (select c, 0, f) -> (czero_nez f, c)
if (isNullConstant(TrueV))
return DAG.getNode(RISCVISD::CZERO_NEZ, DL, VT, FalseV, CondV);
// (select c, (and f, x), f) -> (or (and f, x), (czero_nez f, c))
if (TrueV.getOpcode() == ISD::AND &&
(TrueV.getOperand(0) == FalseV || TrueV.getOperand(1) == FalseV))
return DAG.getNode(
ISD::OR, DL, VT, TrueV,
DAG.getNode(RISCVISD::CZERO_NEZ, DL, VT, FalseV, CondV));
// (select c, t, (and t, x)) -> (or (czero_eqz t, c), (and t, x))
if (FalseV.getOpcode() == ISD::AND &&
(FalseV.getOperand(0) == TrueV || FalseV.getOperand(1) == TrueV))
return DAG.getNode(
ISD::OR, DL, VT, FalseV,
DAG.getNode(RISCVISD::CZERO_EQZ, DL, VT, TrueV, CondV));
// Try some other optimizations before falling back to generic lowering.
if (SDValue V = combineSelectToBinOp(Op.getNode(), DAG, Subtarget))
return V;
// (select c, c1, c2) -> (add (czero_nez c2 - c1, c), c1)
// (select c, c1, c2) -> (add (czero_eqz c1 - c2, c), c2)
if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV)) {
const APInt &TrueVal = TrueV->getAsAPIntVal();
const APInt &FalseVal = FalseV->getAsAPIntVal();
const int TrueValCost = RISCVMatInt::getIntMatCost(
TrueVal, Subtarget.getXLen(), Subtarget, /*CompressionCost=*/true);
const int FalseValCost = RISCVMatInt::getIntMatCost(
FalseVal, Subtarget.getXLen(), Subtarget, /*CompressionCost=*/true);
bool IsCZERO_NEZ = TrueValCost <= FalseValCost;
SDValue LHSVal = DAG.getConstant(
IsCZERO_NEZ ? FalseVal - TrueVal : TrueVal - FalseVal, DL, VT);
SDValue RHSVal =
DAG.getConstant(IsCZERO_NEZ ? TrueVal : FalseVal, DL, VT);
SDValue CMOV =
DAG.getNode(IsCZERO_NEZ ? RISCVISD::CZERO_NEZ : RISCVISD::CZERO_EQZ,
DL, VT, LHSVal, CondV);
return DAG.getNode(ISD::ADD, DL, VT, CMOV, RHSVal);
}
// (select c, t, f) -> (or (czero_eqz t, c), (czero_nez f, c))
// Unless we have the short forward branch optimization.
if (!Subtarget.hasConditionalMoveFusion())
return DAG.getNode(
ISD::OR, DL, VT,
DAG.getNode(RISCVISD::CZERO_EQZ, DL, VT, TrueV, CondV),
DAG.getNode(RISCVISD::CZERO_NEZ, DL, VT, FalseV, CondV));
}
if (SDValue V = combineSelectToBinOp(Op.getNode(), DAG, Subtarget))
return V;
if (Op.hasOneUse()) {
unsigned UseOpc = Op->use_begin()->getOpcode();
if (isBinOp(UseOpc) && DAG.isSafeToSpeculativelyExecute(UseOpc)) {
SDNode *BinOp = *Op->use_begin();
if (SDValue NewSel = foldBinOpIntoSelectIfProfitable(*Op->use_begin(),
DAG, Subtarget)) {
DAG.ReplaceAllUsesWith(BinOp, &NewSel);
return lowerSELECT(NewSel, DAG);
}
}
}
// (select cc, 1.0, 0.0) -> (sint_to_fp (zext cc))
// (select cc, 0.0, 1.0) -> (sint_to_fp (zext (xor cc, 1)))
const ConstantFPSDNode *FPTV = dyn_cast<ConstantFPSDNode>(TrueV);
const ConstantFPSDNode *FPFV = dyn_cast<ConstantFPSDNode>(FalseV);
if (FPTV && FPFV) {
if (FPTV->isExactlyValue(1.0) && FPFV->isExactlyValue(0.0))
return DAG.getNode(ISD::SINT_TO_FP, DL, VT, CondV);
if (FPTV->isExactlyValue(0.0) && FPFV->isExactlyValue(1.0)) {
SDValue XOR = DAG.getNode(ISD::XOR, DL, XLenVT, CondV,
DAG.getConstant(1, DL, XLenVT));
return DAG.getNode(ISD::SINT_TO_FP, DL, VT, XOR);
}
}
// If the condition is not an integer SETCC which operates on XLenVT, we need
// to emit a RISCVISD::SELECT_CC comparing the condition to zero. i.e.:
// (select condv, truev, falsev)
// -> (riscvisd::select_cc condv, zero, setne, truev, falsev)
if (CondV.getOpcode() != ISD::SETCC ||
CondV.getOperand(0).getSimpleValueType() != XLenVT) {
SDValue Zero = DAG.getConstant(0, DL, XLenVT);
SDValue SetNE = DAG.getCondCode(ISD::SETNE);
SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV};
return DAG.getNode(RISCVISD::SELECT_CC, DL, VT, Ops);
}
// If the CondV is the output of a SETCC node which operates on XLenVT inputs,
// then merge the SETCC node into the lowered RISCVISD::SELECT_CC to take
// advantage of the integer compare+branch instructions. i.e.:
// (select (setcc lhs, rhs, cc), truev, falsev)
// -> (riscvisd::select_cc lhs, rhs, cc, truev, falsev)
SDValue LHS = CondV.getOperand(0);
SDValue RHS = CondV.getOperand(1);
ISD::CondCode CCVal = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
// Special case for a select of 2 constants that have a diffence of 1.
// Normally this is done by DAGCombine, but if the select is introduced by
// type legalization or op legalization, we miss it. Restricting to SETLT
// case for now because that is what signed saturating add/sub need.
// FIXME: We don't need the condition to be SETLT or even a SETCC,
// but we would probably want to swap the true/false values if the condition
// is SETGE/SETLE to avoid an XORI.
if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV) &&
CCVal == ISD::SETLT) {
const APInt &TrueVal = TrueV->getAsAPIntVal();
const APInt &FalseVal = FalseV->getAsAPIntVal();
if (TrueVal - 1 == FalseVal)
return DAG.getNode(ISD::ADD, DL, VT, CondV, FalseV);
if (TrueVal + 1 == FalseVal)
return DAG.getNode(ISD::SUB, DL, VT, FalseV, CondV);
}
translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
// 1 < x ? x : 1 -> 0 < x ? x : 1
if (isOneConstant(LHS) && (CCVal == ISD::SETLT || CCVal == ISD::SETULT) &&
RHS == TrueV && LHS == FalseV) {
LHS = DAG.getConstant(0, DL, VT);
// 0 <u x is the same as x != 0.
if (CCVal == ISD::SETULT) {
std::swap(LHS, RHS);
CCVal = ISD::SETNE;
}
}
// x <s -1 ? x : -1 -> x <s 0 ? x : -1
if (isAllOnesConstant(RHS) && CCVal == ISD::SETLT && LHS == TrueV &&
RHS == FalseV) {
RHS = DAG.getConstant(0, DL, VT);
}
SDValue TargetCC = DAG.getCondCode(CCVal);
if (isa<ConstantSDNode>(TrueV) && !isa<ConstantSDNode>(FalseV)) {
// (select (setcc lhs, rhs, CC), constant, falsev)
// -> (select (setcc lhs, rhs, InverseCC), falsev, constant)
std::swap(TrueV, FalseV);
TargetCC = DAG.getCondCode(ISD::getSetCCInverse(CCVal, LHS.getValueType()));
}
SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
return DAG.getNode(RISCVISD::SELECT_CC, DL, VT, Ops);
}
SDValue RISCVTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
SDValue CondV = Op.getOperand(1);
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
if (CondV.getOpcode() == ISD::SETCC &&
CondV.getOperand(0).getValueType() == XLenVT) {
SDValue LHS = CondV.getOperand(0);
SDValue RHS = CondV.getOperand(1);
ISD::CondCode CCVal = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
SDValue TargetCC = DAG.getCondCode(CCVal);
return DAG.getNode(RISCVISD::BR_CC, DL, Op.getValueType(), Op.getOperand(0),
LHS, RHS, TargetCC, Op.getOperand(2));
}
return DAG.getNode(RISCVISD::BR_CC, DL, Op.getValueType(), Op.getOperand(0),
CondV, DAG.getConstant(0, DL, XLenVT),
DAG.getCondCode(ISD::SETNE), Op.getOperand(2));
}
SDValue RISCVTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>();
SDLoc DL(Op);
SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
getPointerTy(MF.getDataLayout()));
// vastart just stores the address of the VarArgsFrameIndex slot into the
// memory location argument.
const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
MachinePointerInfo(SV));
}
SDValue RISCVTargetLowering::lowerFRAMEADDR(SDValue Op,
SelectionDAG &DAG) const {
const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo &MFI = MF.getFrameInfo();
MFI.setFrameAddressIsTaken(true);
Register FrameReg = RI.getFrameRegister(MF);
int XLenInBytes = Subtarget.getXLen() / 8;
EVT VT = Op.getValueType();
SDLoc DL(Op);
SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT);
unsigned Depth = Op.getConstantOperandVal(0);
while (Depth--) {
int Offset = -(XLenInBytes * 2);
SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
DAG.getIntPtrConstant(Offset, DL));
FrameAddr =
DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
}
return FrameAddr;
}
SDValue RISCVTargetLowering::lowerRETURNADDR(SDValue Op,
SelectionDAG &DAG) const {
const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo &MFI = MF.getFrameInfo();
MFI.setReturnAddressIsTaken(true);
MVT XLenVT = Subtarget.getXLenVT();
int XLenInBytes = Subtarget.getXLen() / 8;
if (verifyReturnAddressArgumentIsConstant(Op, DAG))
return SDValue();
EVT VT = Op.getValueType();
SDLoc DL(Op);
unsigned Depth = Op.getConstantOperandVal(0);
if (Depth) {
int Off = -XLenInBytes;
SDValue FrameAddr = lowerFRAMEADDR(Op, DAG);
SDValue Offset = DAG.getConstant(Off, DL, VT);
return DAG.getLoad(VT, DL, DAG.getEntryNode(),
DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
MachinePointerInfo());
}
// Return the value of the return address register, marking it an implicit
// live-in.
Register Reg = MF.addLiveIn(RI.getRARegister(), getRegClassFor(XLenVT));
return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, XLenVT);
}
SDValue RISCVTargetLowering::lowerShiftLeftParts(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Lo = Op.getOperand(0);
SDValue Hi = Op.getOperand(1);
SDValue Shamt = Op.getOperand(2);
EVT VT = Lo.getValueType();
// if Shamt-XLEN < 0: // Shamt < XLEN
// Lo = Lo << Shamt
// Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (XLEN-1 - Shamt))
// else:
// Lo = 0
// Hi = Lo << (Shamt-XLEN)
SDValue Zero = DAG.getConstant(0, DL, VT);
SDValue One = DAG.getConstant(1, DL, VT);
SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One);
SDValue ShiftRightLo =
DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, XLenMinus1Shamt);
SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusXLen);
SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero);
Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
SDValue Parts[2] = {Lo, Hi};
return DAG.getMergeValues(Parts, DL);
}
SDValue RISCVTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
bool IsSRA) const {
SDLoc DL(Op);
SDValue Lo = Op.getOperand(0);
SDValue Hi = Op.getOperand(1);
SDValue Shamt = Op.getOperand(2);
EVT VT = Lo.getValueType();
// SRA expansion:
// if Shamt-XLEN < 0: // Shamt < XLEN
// Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - ShAmt))
// Hi = Hi >>s Shamt
// else:
// Lo = Hi >>s (Shamt-XLEN);
// Hi = Hi >>s (XLEN-1)
//
// SRL expansion:
// if Shamt-XLEN < 0: // Shamt < XLEN
// Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - ShAmt))
// Hi = Hi >>u Shamt
// else:
// Lo = Hi >>u (Shamt-XLEN);
// Hi = 0;
unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL;
SDValue Zero = DAG.getConstant(0, DL, VT);
SDValue One = DAG.getConstant(1, DL, VT);
SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One);
SDValue ShiftLeftHi =
DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, XLenMinus1Shamt);
SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi);
SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt);
SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusXLen);
SDValue HiFalse =
IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, XLenMinus1) : Zero;
SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse);
Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
SDValue Parts[2] = {Lo, Hi};
return DAG.getMergeValues(Parts, DL);
}
// Lower splats of i1 types to SETCC. For each mask vector type, we have a
// legal equivalently-sized i8 type, so we can use that as a go-between.
SDValue RISCVTargetLowering::lowerVectorMaskSplat(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue SplatVal = Op.getOperand(0);
// All-zeros or all-ones splats are handled specially.
if (ISD::isConstantSplatVectorAllOnes(Op.getNode())) {
SDValue VL = getDefaultScalableVLOps(VT, DL, DAG, Subtarget).second;
return DAG.getNode(RISCVISD::VMSET_VL, DL, VT, VL);
}
if (ISD::isConstantSplatVectorAllZeros(Op.getNode())) {
SDValue VL = getDefaultScalableVLOps(VT, DL, DAG, Subtarget).second;
return DAG.getNode(RISCVISD::VMCLR_VL, DL, VT, VL);
}
MVT InterVT = VT.changeVectorElementType(MVT::i8);
SplatVal = DAG.getNode(ISD::AND, DL, SplatVal.getValueType(), SplatVal,
DAG.getConstant(1, DL, SplatVal.getValueType()));
SDValue LHS = DAG.getSplatVector(InterVT, DL, SplatVal);
SDValue Zero = DAG.getConstant(0, DL, InterVT);
return DAG.getSetCC(DL, VT, LHS, Zero, ISD::SETNE);
}
// Custom-lower a SPLAT_VECTOR_PARTS where XLEN<SEW, as the SEW element type is
// illegal (currently only vXi64 RV32).
// FIXME: We could also catch non-constant sign-extended i32 values and lower
// them to VMV_V_X_VL.
SDValue RISCVTargetLowering::lowerSPLAT_VECTOR_PARTS(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VecVT = Op.getSimpleValueType();
assert(!Subtarget.is64Bit() && VecVT.getVectorElementType() == MVT::i64 &&
"Unexpected SPLAT_VECTOR_PARTS lowering");
assert(Op.getNumOperands() == 2 && "Unexpected number of operands!");
SDValue Lo = Op.getOperand(0);
SDValue Hi = Op.getOperand(1);
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector())
ContainerVT = getContainerForFixedLengthVector(VecVT);
auto VL = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).second;
SDValue Res =
splatPartsI64WithVL(DL, ContainerVT, SDValue(), Lo, Hi, VL, DAG);
if (VecVT.isFixedLengthVector())
Res = convertFromScalableVector(VecVT, Res, DAG, Subtarget);
return Res;
}
// Custom-lower extensions from mask vectors by using a vselect either with 1
// for zero/any-extension or -1 for sign-extension:
// (vXiN = (s|z)ext vXi1:vmask) -> (vXiN = vselect vmask, (-1 or 1), 0)
// Note that any-extension is lowered identically to zero-extension.
SDValue RISCVTargetLowering::lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG,
int64_t ExtTrueVal) const {
SDLoc DL(Op);
MVT VecVT = Op.getSimpleValueType();
SDValue Src = Op.getOperand(0);
// Only custom-lower extensions from mask types
assert(Src.getValueType().isVector() &&
Src.getValueType().getVectorElementType() == MVT::i1);
if (VecVT.isScalableVector()) {
SDValue SplatZero = DAG.getConstant(0, DL, VecVT);
SDValue SplatTrueVal = DAG.getConstant(ExtTrueVal, DL, VecVT);
return DAG.getNode(ISD::VSELECT, DL, VecVT, Src, SplatTrueVal, SplatZero);
}
MVT ContainerVT = getContainerForFixedLengthVector(VecVT);
MVT I1ContainerVT =
MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
SDValue CC = convertToScalableVector(I1ContainerVT, Src, DAG, Subtarget);
SDValue VL = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).second;
MVT XLenVT = Subtarget.getXLenVT();
SDValue SplatZero = DAG.getConstant(0, DL, XLenVT);
SDValue SplatTrueVal = DAG.getConstant(ExtTrueVal, DL, XLenVT);
SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SplatZero, VL);
SplatTrueVal = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SplatTrueVal, VL);
SDValue Select =
DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, CC, SplatTrueVal,
SplatZero, DAG.getUNDEF(ContainerVT), VL);
return convertFromScalableVector(VecVT, Select, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerFixedLengthVectorExtendToRVV(
SDValue Op, SelectionDAG &DAG, unsigned ExtendOpc) const {
MVT ExtVT = Op.getSimpleValueType();
// Only custom-lower extensions from fixed-length vector types.
if (!ExtVT.isFixedLengthVector())
return Op;
MVT VT = Op.getOperand(0).getSimpleValueType();
// Grab the canonical container type for the extended type. Infer the smaller
// type from that to ensure the same number of vector elements, as we know
// the LMUL will be sufficient to hold the smaller type.
MVT ContainerExtVT = getContainerForFixedLengthVector(ExtVT);
// Get the extended container type manually to ensure the same number of
// vector elements between source and dest.
MVT ContainerVT = MVT::getVectorVT(VT.getVectorElementType(),
ContainerExtVT.getVectorElementCount());
SDValue Op1 =
convertToScalableVector(ContainerVT, Op.getOperand(0), DAG, Subtarget);
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
SDValue Ext = DAG.getNode(ExtendOpc, DL, ContainerExtVT, Op1, Mask, VL);
return convertFromScalableVector(ExtVT, Ext, DAG, Subtarget);
}
// Custom-lower truncations from vectors to mask vectors by using a mask and a
// setcc operation:
// (vXi1 = trunc vXiN vec) -> (vXi1 = setcc (and vec, 1), 0, ne)
SDValue RISCVTargetLowering::lowerVectorMaskTruncLike(SDValue Op,
SelectionDAG &DAG) const {
bool IsVPTrunc = Op.getOpcode() == ISD::VP_TRUNCATE;
SDLoc DL(Op);
EVT MaskVT = Op.getValueType();
// Only expect to custom-lower truncations to mask types
assert(MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 &&
"Unexpected type for vector mask lowering");
SDValue Src = Op.getOperand(0);
MVT VecVT = Src.getSimpleValueType();
SDValue Mask, VL;
if (IsVPTrunc) {
Mask = Op.getOperand(1);
VL = Op.getOperand(2);
}
// If this is a fixed vector, we need to convert it to a scalable vector.
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
if (IsVPTrunc) {
MVT MaskContainerVT =
getContainerForFixedLengthVector(Mask.getSimpleValueType());
Mask = convertToScalableVector(MaskContainerVT, Mask, DAG, Subtarget);
}
}
if (!IsVPTrunc) {
std::tie(Mask, VL) =
getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
}
SDValue SplatOne = DAG.getConstant(1, DL, Subtarget.getXLenVT());
SDValue SplatZero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
SplatOne = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SplatOne, VL);
SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SplatZero, VL);
MVT MaskContainerVT = ContainerVT.changeVectorElementType(MVT::i1);
SDValue Trunc = DAG.getNode(RISCVISD::AND_VL, DL, ContainerVT, Src, SplatOne,
DAG.getUNDEF(ContainerVT), Mask, VL);
Trunc = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskContainerVT,
{Trunc, SplatZero, DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(MaskContainerVT), Mask, VL});
if (MaskVT.isFixedLengthVector())
Trunc = convertFromScalableVector(MaskVT, Trunc, DAG, Subtarget);
return Trunc;
}
SDValue RISCVTargetLowering::lowerVectorTruncLike(SDValue Op,
SelectionDAG &DAG) const {
bool IsVPTrunc = Op.getOpcode() == ISD::VP_TRUNCATE;
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
// Only custom-lower vector truncates
assert(VT.isVector() && "Unexpected type for vector truncate lowering");
// Truncates to mask types are handled differently
if (VT.getVectorElementType() == MVT::i1)
return lowerVectorMaskTruncLike(Op, DAG);
// RVV only has truncates which operate from SEW*2->SEW, so lower arbitrary
// truncates as a series of "RISCVISD::TRUNCATE_VECTOR_VL" nodes which
// truncate by one power of two at a time.
MVT DstEltVT = VT.getVectorElementType();
SDValue Src = Op.getOperand(0);
MVT SrcVT = Src.getSimpleValueType();
MVT SrcEltVT = SrcVT.getVectorElementType();
assert(DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT.getSizeInBits()) &&
isPowerOf2_64(SrcEltVT.getSizeInBits()) &&
"Unexpected vector truncate lowering");
MVT ContainerVT = SrcVT;
SDValue Mask, VL;
if (IsVPTrunc) {
Mask = Op.getOperand(1);
VL = Op.getOperand(2);
}
if (SrcVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(SrcVT);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
if (IsVPTrunc) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
}
SDValue Result = Src;
if (!IsVPTrunc) {
std::tie(Mask, VL) =
getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
}
LLVMContext &Context = *DAG.getContext();
const ElementCount Count = ContainerVT.getVectorElementCount();
do {
SrcEltVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2);
EVT ResultVT = EVT::getVectorVT(Context, SrcEltVT, Count);
Result = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, ResultVT, Result,
Mask, VL);
} while (SrcEltVT != DstEltVT);
if (SrcVT.isFixedLengthVector())
Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
return Result;
}
SDValue
RISCVTargetLowering::lowerStrictFPExtendOrRoundLike(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Chain = Op.getOperand(0);
SDValue Src = Op.getOperand(1);
MVT VT = Op.getSimpleValueType();
MVT SrcVT = Src.getSimpleValueType();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
ContainerVT =
SrcContainerVT.changeVectorElementType(VT.getVectorElementType());
Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
}
auto [Mask, VL] = getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
// RVV can only widen/truncate fp to types double/half the size as the source.
if ((VT.getVectorElementType() == MVT::f64 &&
SrcVT.getVectorElementType() == MVT::f16) ||
(VT.getVectorElementType() == MVT::f16 &&
SrcVT.getVectorElementType() == MVT::f64)) {
// For double rounding, the intermediate rounding should be round-to-odd.
unsigned InterConvOpc = Op.getOpcode() == ISD::STRICT_FP_EXTEND
? RISCVISD::STRICT_FP_EXTEND_VL
: RISCVISD::STRICT_VFNCVT_ROD_VL;
MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32);
Src = DAG.getNode(InterConvOpc, DL, DAG.getVTList(InterVT, MVT::Other),
Chain, Src, Mask, VL);
Chain = Src.getValue(1);
}
unsigned ConvOpc = Op.getOpcode() == ISD::STRICT_FP_EXTEND
? RISCVISD::STRICT_FP_EXTEND_VL
: RISCVISD::STRICT_FP_ROUND_VL;
SDValue Res = DAG.getNode(ConvOpc, DL, DAG.getVTList(ContainerVT, MVT::Other),
Chain, Src, Mask, VL);
if (VT.isFixedLengthVector()) {
// StrictFP operations have two result values. Their lowered result should
// have same result count.
SDValue SubVec = convertFromScalableVector(VT, Res, DAG, Subtarget);
Res = DAG.getMergeValues({SubVec, Res.getValue(1)}, DL);
}
return Res;
}
SDValue
RISCVTargetLowering::lowerVectorFPExtendOrRoundLike(SDValue Op,
SelectionDAG &DAG) const {
bool IsVP =
Op.getOpcode() == ISD::VP_FP_ROUND || Op.getOpcode() == ISD::VP_FP_EXTEND;
bool IsExtend =
Op.getOpcode() == ISD::VP_FP_EXTEND || Op.getOpcode() == ISD::FP_EXTEND;
// RVV can only do truncate fp to types half the size as the source. We
// custom-lower f64->f16 rounds via RVV's round-to-odd float
// conversion instruction.
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
assert(VT.isVector() && "Unexpected type for vector truncate lowering");
SDValue Src = Op.getOperand(0);
MVT SrcVT = Src.getSimpleValueType();
bool IsDirectExtend = IsExtend && (VT.getVectorElementType() != MVT::f64 ||
SrcVT.getVectorElementType() != MVT::f16);
bool IsDirectTrunc = !IsExtend && (VT.getVectorElementType() != MVT::f16 ||
SrcVT.getVectorElementType() != MVT::f64);
bool IsDirectConv = IsDirectExtend || IsDirectTrunc;
// Prepare any fixed-length vector operands.
MVT ContainerVT = VT;
SDValue Mask, VL;
if (IsVP) {
Mask = Op.getOperand(1);
VL = Op.getOperand(2);
}
if (VT.isFixedLengthVector()) {
MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
ContainerVT =
SrcContainerVT.changeVectorElementType(VT.getVectorElementType());
Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
if (IsVP) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
}
if (!IsVP)
std::tie(Mask, VL) =
getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
unsigned ConvOpc = IsExtend ? RISCVISD::FP_EXTEND_VL : RISCVISD::FP_ROUND_VL;
if (IsDirectConv) {
Src = DAG.getNode(ConvOpc, DL, ContainerVT, Src, Mask, VL);
if (VT.isFixedLengthVector())
Src = convertFromScalableVector(VT, Src, DAG, Subtarget);
return Src;
}
unsigned InterConvOpc =
IsExtend ? RISCVISD::FP_EXTEND_VL : RISCVISD::VFNCVT_ROD_VL;
MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32);
SDValue IntermediateConv =
DAG.getNode(InterConvOpc, DL, InterVT, Src, Mask, VL);
SDValue Result =
DAG.getNode(ConvOpc, DL, ContainerVT, IntermediateConv, Mask, VL);
if (VT.isFixedLengthVector())
return convertFromScalableVector(VT, Result, DAG, Subtarget);
return Result;
}
// Given a scalable vector type and an index into it, returns the type for the
// smallest subvector that the index fits in. This can be used to reduce LMUL
// for operations like vslidedown.
//
// E.g. With Zvl128b, index 3 in a nxv4i32 fits within the first nxv2i32.
static std::optional<MVT>
getSmallestVTForIndex(MVT VecVT, unsigned MaxIdx, SDLoc DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(VecVT.isScalableVector());
const unsigned EltSize = VecVT.getScalarSizeInBits();
const unsigned VectorBitsMin = Subtarget.getRealMinVLen();
const unsigned MinVLMAX = VectorBitsMin / EltSize;
MVT SmallerVT;
if (MaxIdx < MinVLMAX)
SmallerVT = getLMUL1VT(VecVT);
else if (MaxIdx < MinVLMAX * 2)
SmallerVT = getLMUL1VT(VecVT).getDoubleNumVectorElementsVT();
else if (MaxIdx < MinVLMAX * 4)
SmallerVT = getLMUL1VT(VecVT)
.getDoubleNumVectorElementsVT()
.getDoubleNumVectorElementsVT();
if (!SmallerVT.isValid() || !VecVT.bitsGT(SmallerVT))
return std::nullopt;
return SmallerVT;
}
// Custom-legalize INSERT_VECTOR_ELT so that the value is inserted into the
// first position of a vector, and that vector is slid up to the insert index.
// By limiting the active vector length to index+1 and merging with the
// original vector (with an undisturbed tail policy for elements >= VL), we
// achieve the desired result of leaving all elements untouched except the one
// at VL-1, which is replaced with the desired value.
SDValue RISCVTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VecVT = Op.getSimpleValueType();
SDValue Vec = Op.getOperand(0);
SDValue Val = Op.getOperand(1);
SDValue Idx = Op.getOperand(2);
if (VecVT.getVectorElementType() == MVT::i1) {
// FIXME: For now we just promote to an i8 vector and insert into that,
// but this is probably not optimal.
MVT WideVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorElementCount());
Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT, Vec);
Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideVT, Vec, Val, Idx);
return DAG.getNode(ISD::TRUNCATE, DL, VecVT, Vec);
}
MVT ContainerVT = VecVT;
// If the operand is a fixed-length vector, convert to a scalable one.
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
// If we know the index we're going to insert at, we can shrink Vec so that
// we're performing the scalar inserts and slideup on a smaller LMUL.
MVT OrigContainerVT = ContainerVT;
SDValue OrigVec = Vec;
SDValue AlignedIdx;
if (auto *IdxC = dyn_cast<ConstantSDNode>(Idx)) {
const unsigned OrigIdx = IdxC->getZExtValue();
// Do we know an upper bound on LMUL?
if (auto ShrunkVT = getSmallestVTForIndex(ContainerVT, OrigIdx,
DL, DAG, Subtarget)) {
ContainerVT = *ShrunkVT;
AlignedIdx = DAG.getVectorIdxConstant(0, DL);
}
// If we're compiling for an exact VLEN value, we can always perform
// the insert in m1 as we can determine the register corresponding to
// the index in the register group.
const MVT M1VT = getLMUL1VT(ContainerVT);
if (auto VLEN = Subtarget.getRealVLen();
VLEN && ContainerVT.bitsGT(M1VT)) {
EVT ElemVT = VecVT.getVectorElementType();
unsigned ElemsPerVReg = *VLEN / ElemVT.getFixedSizeInBits();
unsigned RemIdx = OrigIdx % ElemsPerVReg;
unsigned SubRegIdx = OrigIdx / ElemsPerVReg;
unsigned ExtractIdx =
SubRegIdx * M1VT.getVectorElementCount().getKnownMinValue();
AlignedIdx = DAG.getVectorIdxConstant(ExtractIdx, DL);
Idx = DAG.getVectorIdxConstant(RemIdx, DL);
ContainerVT = M1VT;
}
if (AlignedIdx)
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ContainerVT, Vec,
AlignedIdx);
}
MVT XLenVT = Subtarget.getXLenVT();
bool IsLegalInsert = Subtarget.is64Bit() || Val.getValueType() != MVT::i64;
// Even i64-element vectors on RV32 can be lowered without scalar
// legalization if the most-significant 32 bits of the value are not affected
// by the sign-extension of the lower 32 bits.
// TODO: We could also catch sign extensions of a 32-bit value.
if (!IsLegalInsert && isa<ConstantSDNode>(Val)) {
const auto *CVal = cast<ConstantSDNode>(Val);
if (isInt<32>(CVal->getSExtValue())) {
IsLegalInsert = true;
Val = DAG.getConstant(CVal->getSExtValue(), DL, MVT::i32);
}
}
auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
SDValue ValInVec;
if (IsLegalInsert) {
unsigned Opc =
VecVT.isFloatingPoint() ? RISCVISD::VFMV_S_F_VL : RISCVISD::VMV_S_X_VL;
if (isNullConstant(Idx)) {
if (!VecVT.isFloatingPoint())
Val = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Val);
Vec = DAG.getNode(Opc, DL, ContainerVT, Vec, Val, VL);
if (AlignedIdx)
Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, OrigContainerVT, OrigVec,
Vec, AlignedIdx);
if (!VecVT.isFixedLengthVector())
return Vec;
return convertFromScalableVector(VecVT, Vec, DAG, Subtarget);
}
ValInVec = lowerScalarInsert(Val, VL, ContainerVT, DL, DAG, Subtarget);
} else {
// On RV32, i64-element vectors must be specially handled to place the
// value at element 0, by using two vslide1down instructions in sequence on
// the i32 split lo/hi value. Use an equivalently-sized i32 vector for
// this.
SDValue ValLo, ValHi;
std::tie(ValLo, ValHi) = DAG.SplitScalar(Val, DL, MVT::i32, MVT::i32);
MVT I32ContainerVT =
MVT::getVectorVT(MVT::i32, ContainerVT.getVectorElementCount() * 2);
SDValue I32Mask =
getDefaultScalableVLOps(I32ContainerVT, DL, DAG, Subtarget).first;
// Limit the active VL to two.
SDValue InsertI64VL = DAG.getConstant(2, DL, XLenVT);
// If the Idx is 0 we can insert directly into the vector.
if (isNullConstant(Idx)) {
// First slide in the lo value, then the hi in above it. We use slide1down
// to avoid the register group overlap constraint of vslide1up.
ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
Vec, Vec, ValLo, I32Mask, InsertI64VL);
// If the source vector is undef don't pass along the tail elements from
// the previous slide1down.
SDValue Tail = Vec.isUndef() ? Vec : ValInVec;
ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
Tail, ValInVec, ValHi, I32Mask, InsertI64VL);
// Bitcast back to the right container type.
ValInVec = DAG.getBitcast(ContainerVT, ValInVec);
if (AlignedIdx)
ValInVec =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, OrigContainerVT, OrigVec,
ValInVec, AlignedIdx);
if (!VecVT.isFixedLengthVector())
return ValInVec;
return convertFromScalableVector(VecVT, ValInVec, DAG, Subtarget);
}
// First slide in the lo value, then the hi in above it. We use slide1down
// to avoid the register group overlap constraint of vslide1up.
ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
DAG.getUNDEF(I32ContainerVT),
DAG.getUNDEF(I32ContainerVT), ValLo,
I32Mask, InsertI64VL);
ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
DAG.getUNDEF(I32ContainerVT), ValInVec, ValHi,
I32Mask, InsertI64VL);
// Bitcast back to the right container type.
ValInVec = DAG.getBitcast(ContainerVT, ValInVec);
}
// Now that the value is in a vector, slide it into position.
SDValue InsertVL =
DAG.getNode(ISD::ADD, DL, XLenVT, Idx, DAG.getConstant(1, DL, XLenVT));
// Use tail agnostic policy if Idx is the last index of Vec.
unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
if (VecVT.isFixedLengthVector() && isa<ConstantSDNode>(Idx) &&
Idx->getAsZExtVal() + 1 == VecVT.getVectorNumElements())
Policy = RISCVII::TAIL_AGNOSTIC;
SDValue Slideup = getVSlideup(DAG, Subtarget, DL, ContainerVT, Vec, ValInVec,
Idx, Mask, InsertVL, Policy);
if (AlignedIdx)
Slideup = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, OrigContainerVT, OrigVec,
Slideup, AlignedIdx);
if (!VecVT.isFixedLengthVector())
return Slideup;
return convertFromScalableVector(VecVT, Slideup, DAG, Subtarget);
}
// Custom-lower EXTRACT_VECTOR_ELT operations to slide the vector down, then
// extract the first element: (extractelt (slidedown vec, idx), 0). For integer
// types this is done using VMV_X_S to allow us to glean information about the
// sign bits of the result.
SDValue RISCVTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Idx = Op.getOperand(1);
SDValue Vec = Op.getOperand(0);
EVT EltVT = Op.getValueType();
MVT VecVT = Vec.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
if (VecVT.getVectorElementType() == MVT::i1) {
// Use vfirst.m to extract the first bit.
if (isNullConstant(Idx)) {
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
SDValue Vfirst =
DAG.getNode(RISCVISD::VFIRST_VL, DL, XLenVT, Vec, Mask, VL);
SDValue Res = DAG.getSetCC(DL, XLenVT, Vfirst,
DAG.getConstant(0, DL, XLenVT), ISD::SETEQ);
return DAG.getNode(ISD::TRUNCATE, DL, EltVT, Res);
}
if (VecVT.isFixedLengthVector()) {
unsigned NumElts = VecVT.getVectorNumElements();
if (NumElts >= 8) {
MVT WideEltVT;
unsigned WidenVecLen;
SDValue ExtractElementIdx;
SDValue ExtractBitIdx;
unsigned MaxEEW = Subtarget.getELen();
MVT LargestEltVT = MVT::getIntegerVT(
std::min(MaxEEW, unsigned(XLenVT.getSizeInBits())));
if (NumElts <= LargestEltVT.getSizeInBits()) {
assert(isPowerOf2_32(NumElts) &&
"the number of elements should be power of 2");
WideEltVT = MVT::getIntegerVT(NumElts);
WidenVecLen = 1;
ExtractElementIdx = DAG.getConstant(0, DL, XLenVT);
ExtractBitIdx = Idx;
} else {
WideEltVT = LargestEltVT;
WidenVecLen = NumElts / WideEltVT.getSizeInBits();
// extract element index = index / element width
ExtractElementIdx = DAG.getNode(
ISD::SRL, DL, XLenVT, Idx,
DAG.getConstant(Log2_64(WideEltVT.getSizeInBits()), DL, XLenVT));
// mask bit index = index % element width
ExtractBitIdx = DAG.getNode(
ISD::AND, DL, XLenVT, Idx,
DAG.getConstant(WideEltVT.getSizeInBits() - 1, DL, XLenVT));
}
MVT WideVT = MVT::getVectorVT(WideEltVT, WidenVecLen);
Vec = DAG.getNode(ISD::BITCAST, DL, WideVT, Vec);
SDValue ExtractElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, XLenVT,
Vec, ExtractElementIdx);
// Extract the bit from GPR.
SDValue ShiftRight =
DAG.getNode(ISD::SRL, DL, XLenVT, ExtractElt, ExtractBitIdx);
SDValue Res = DAG.getNode(ISD::AND, DL, XLenVT, ShiftRight,
DAG.getConstant(1, DL, XLenVT));
return DAG.getNode(ISD::TRUNCATE, DL, EltVT, Res);
}
}
// Otherwise, promote to an i8 vector and extract from that.
MVT WideVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorElementCount());
Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT, Vec);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vec, Idx);
}
// If this is a fixed vector, we need to convert it to a scalable vector.
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
// If we're compiling for an exact VLEN value and we have a known
// constant index, we can always perform the extract in m1 (or
// smaller) as we can determine the register corresponding to
// the index in the register group.
const auto VLen = Subtarget.getRealVLen();
if (auto *IdxC = dyn_cast<ConstantSDNode>(Idx);
IdxC && VLen && VecVT.getSizeInBits().getKnownMinValue() > *VLen) {
MVT M1VT = getLMUL1VT(ContainerVT);
unsigned OrigIdx = IdxC->getZExtValue();
EVT ElemVT = VecVT.getVectorElementType();
unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits();
unsigned RemIdx = OrigIdx % ElemsPerVReg;
unsigned SubRegIdx = OrigIdx / ElemsPerVReg;
unsigned ExtractIdx =
SubRegIdx * M1VT.getVectorElementCount().getKnownMinValue();
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, M1VT, Vec,
DAG.getVectorIdxConstant(ExtractIdx, DL));
Idx = DAG.getVectorIdxConstant(RemIdx, DL);
ContainerVT = M1VT;
}
// Reduce the LMUL of our slidedown and vmv.x.s to the smallest LMUL which
// contains our index.
std::optional<uint64_t> MaxIdx;
if (VecVT.isFixedLengthVector())
MaxIdx = VecVT.getVectorNumElements() - 1;
if (auto *IdxC = dyn_cast<ConstantSDNode>(Idx))
MaxIdx = IdxC->getZExtValue();
if (MaxIdx) {
if (auto SmallerVT =
getSmallestVTForIndex(ContainerVT, *MaxIdx, DL, DAG, Subtarget)) {
ContainerVT = *SmallerVT;
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ContainerVT, Vec,
DAG.getConstant(0, DL, XLenVT));
}
}
// If after narrowing, the required slide is still greater than LMUL2,
// fallback to generic expansion and go through the stack. This is done
// for a subtle reason: extracting *all* elements out of a vector is
// widely expected to be linear in vector size, but because vslidedown
// is linear in LMUL, performing N extracts using vslidedown becomes
// O(n^2) / (VLEN/ETYPE) work. On the surface, going through the stack
// seems to have the same problem (the store is linear in LMUL), but the
// generic expansion *memoizes* the store, and thus for many extracts of
// the same vector we end up with one store and a bunch of loads.
// TODO: We don't have the same code for insert_vector_elt because we
// have BUILD_VECTOR and handle the degenerate case there. Should we
// consider adding an inverse BUILD_VECTOR node?
MVT LMUL2VT = getLMUL1VT(ContainerVT).getDoubleNumVectorElementsVT();
if (ContainerVT.bitsGT(LMUL2VT) && VecVT.isFixedLengthVector())
return SDValue();
// If the index is 0, the vector is already in the right position.
if (!isNullConstant(Idx)) {
// Use a VL of 1 to avoid processing more elements than we need.
auto [Mask, VL] = getDefaultVLOps(1, ContainerVT, DL, DAG, Subtarget);
Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL);
}
if (!EltVT.isInteger()) {
// Floating-point extracts are handled in TableGen.
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vec,
DAG.getVectorIdxConstant(0, DL));
}
SDValue Elt0 = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec);
return DAG.getNode(ISD::TRUNCATE, DL, EltVT, Elt0);
}
// Some RVV intrinsics may claim that they want an integer operand to be
// promoted or expanded.
static SDValue lowerVectorIntrinsicScalars(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert((Op.getOpcode() == ISD::INTRINSIC_VOID ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN) &&
"Unexpected opcode");
if (!Subtarget.hasVInstructions())
return SDValue();
bool HasChain = Op.getOpcode() == ISD::INTRINSIC_VOID ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN;
unsigned IntNo = Op.getConstantOperandVal(HasChain ? 1 : 0);
SDLoc DL(Op);
const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II =
RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo);
if (!II || !II->hasScalarOperand())
return SDValue();
unsigned SplatOp = II->ScalarOperand + 1 + HasChain;
assert(SplatOp < Op.getNumOperands());
SmallVector<SDValue, 8> Operands(Op->op_begin(), Op->op_end());
SDValue &ScalarOp = Operands[SplatOp];
MVT OpVT = ScalarOp.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
// If this isn't a scalar, or its type is XLenVT we're done.
if (!OpVT.isScalarInteger() || OpVT == XLenVT)
return SDValue();
// Simplest case is that the operand needs to be promoted to XLenVT.
if (OpVT.bitsLT(XLenVT)) {
// If the operand is a constant, sign extend to increase our chances
// of being able to use a .vi instruction. ANY_EXTEND would become a
// a zero extend and the simm5 check in isel would fail.
// FIXME: Should we ignore the upper bits in isel instead?
unsigned ExtOpc =
isa<ConstantSDNode>(ScalarOp) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
ScalarOp = DAG.getNode(ExtOpc, DL, XLenVT, ScalarOp);
return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
}
// Use the previous operand to get the vXi64 VT. The result might be a mask
// VT for compares. Using the previous operand assumes that the previous
// operand will never have a smaller element size than a scalar operand and
// that a widening operation never uses SEW=64.
// NOTE: If this fails the below assert, we can probably just find the
// element count from any operand or result and use it to construct the VT.
assert(II->ScalarOperand > 0 && "Unexpected splat operand!");
MVT VT = Op.getOperand(SplatOp - 1).getSimpleValueType();
// The more complex case is when the scalar is larger than XLenVT.
assert(XLenVT == MVT::i32 && OpVT == MVT::i64 &&
VT.getVectorElementType() == MVT::i64 && "Unexpected VTs!");
// If this is a sign-extended 32-bit value, we can truncate it and rely on the
// instruction to sign-extend since SEW>XLEN.
if (DAG.ComputeNumSignBits(ScalarOp) > 32) {
ScalarOp = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, ScalarOp);
return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
}
switch (IntNo) {
case Intrinsic::riscv_vslide1up:
case Intrinsic::riscv_vslide1down:
case Intrinsic::riscv_vslide1up_mask:
case Intrinsic::riscv_vslide1down_mask: {
// We need to special case these when the scalar is larger than XLen.
unsigned NumOps = Op.getNumOperands();
bool IsMasked = NumOps == 7;
// Convert the vector source to the equivalent nxvXi32 vector.
MVT I32VT = MVT::getVectorVT(MVT::i32, VT.getVectorElementCount() * 2);
SDValue Vec = DAG.getBitcast(I32VT, Operands[2]);
SDValue ScalarLo, ScalarHi;
std::tie(ScalarLo, ScalarHi) =
DAG.SplitScalar(ScalarOp, DL, MVT::i32, MVT::i32);
// Double the VL since we halved SEW.
SDValue AVL = getVLOperand(Op);
SDValue I32VL;
// Optimize for constant AVL
if (isa<ConstantSDNode>(AVL)) {
const auto [MinVLMAX, MaxVLMAX] =
RISCVTargetLowering::computeVLMAXBounds(VT, Subtarget);
uint64_t AVLInt = AVL->getAsZExtVal();
if (AVLInt <= MinVLMAX) {
I32VL = DAG.getConstant(2 * AVLInt, DL, XLenVT);
} else if (AVLInt >= 2 * MaxVLMAX) {
// Just set vl to VLMAX in this situation
RISCVII::VLMUL Lmul = RISCVTargetLowering::getLMUL(I32VT);
SDValue LMUL = DAG.getConstant(Lmul, DL, XLenVT);
unsigned Sew = RISCVVType::encodeSEW(I32VT.getScalarSizeInBits());
SDValue SEW = DAG.getConstant(Sew, DL, XLenVT);
SDValue SETVLMAX = DAG.getTargetConstant(
Intrinsic::riscv_vsetvlimax, DL, MVT::i32);
I32VL = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, XLenVT, SETVLMAX, SEW,
LMUL);
} else {
// For AVL between (MinVLMAX, 2 * MaxVLMAX), the actual working vl
// is related to the hardware implementation.
// So let the following code handle
}
}
if (!I32VL) {
RISCVII::VLMUL Lmul = RISCVTargetLowering::getLMUL(VT);
SDValue LMUL = DAG.getConstant(Lmul, DL, XLenVT);
unsigned Sew = RISCVVType::encodeSEW(VT.getScalarSizeInBits());
SDValue SEW = DAG.getConstant(Sew, DL, XLenVT);
SDValue SETVL =
DAG.getTargetConstant(Intrinsic::riscv_vsetvli, DL, MVT::i32);
// Using vsetvli instruction to get actually used length which related to
// the hardware implementation
SDValue VL = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, XLenVT, SETVL, AVL,
SEW, LMUL);
I32VL =
DAG.getNode(ISD::SHL, DL, XLenVT, VL, DAG.getConstant(1, DL, XLenVT));
}
SDValue I32Mask = getAllOnesMask(I32VT, I32VL, DL, DAG);
// Shift the two scalar parts in using SEW=32 slide1up/slide1down
// instructions.
SDValue Passthru;
if (IsMasked)
Passthru = DAG.getUNDEF(I32VT);
else
Passthru = DAG.getBitcast(I32VT, Operands[1]);
if (IntNo == Intrinsic::riscv_vslide1up ||
IntNo == Intrinsic::riscv_vslide1up_mask) {
Vec = DAG.getNode(RISCVISD::VSLIDE1UP_VL, DL, I32VT, Passthru, Vec,
ScalarHi, I32Mask, I32VL);
Vec = DAG.getNode(RISCVISD::VSLIDE1UP_VL, DL, I32VT, Passthru, Vec,
ScalarLo, I32Mask, I32VL);
} else {
Vec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32VT, Passthru, Vec,
ScalarLo, I32Mask, I32VL);
Vec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32VT, Passthru, Vec,
ScalarHi, I32Mask, I32VL);
}
// Convert back to nxvXi64.
Vec = DAG.getBitcast(VT, Vec);
if (!IsMasked)
return Vec;
// Apply mask after the operation.
SDValue Mask = Operands[NumOps - 3];
SDValue MaskedOff = Operands[1];
// Assume Policy operand is the last operand.
uint64_t Policy = Operands[NumOps - 1]->getAsZExtVal();
// We don't need to select maskedoff if it's undef.
if (MaskedOff.isUndef())
return Vec;
// TAMU
if (Policy == RISCVII::TAIL_AGNOSTIC)
return DAG.getNode(RISCVISD::VMERGE_VL, DL, VT, Mask, Vec, MaskedOff,
DAG.getUNDEF(VT), AVL);
// TUMA or TUMU: Currently we always emit tumu policy regardless of tuma.
// It's fine because vmerge does not care mask policy.
return DAG.getNode(RISCVISD::VMERGE_VL, DL, VT, Mask, Vec, MaskedOff,
MaskedOff, AVL);
}
}
// We need to convert the scalar to a splat vector.
SDValue VL = getVLOperand(Op);
assert(VL.getValueType() == XLenVT);
ScalarOp = splatSplitI64WithVL(DL, VT, SDValue(), ScalarOp, VL, DAG);
return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
}
// Lower the llvm.get.vector.length intrinsic to vsetvli. We only support
// scalable vector llvm.get.vector.length for now.
//
// We need to convert from a scalable VF to a vsetvli with VLMax equal to
// (vscale * VF). The vscale and VF are independent of element width. We use
// SEW=8 for the vsetvli because it is the only element width that supports all
// fractional LMULs. The LMUL is choosen so that with SEW=8 the VLMax is
// (vscale * VF). Where vscale is defined as VLEN/RVVBitsPerBlock. The
// InsertVSETVLI pass can fix up the vtype of the vsetvli if a different
// SEW and LMUL are better for the surrounding vector instructions.
static SDValue lowerGetVectorLength(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT XLenVT = Subtarget.getXLenVT();
// The smallest LMUL is only valid for the smallest element width.
const unsigned ElementWidth = 8;
// Determine the VF that corresponds to LMUL 1 for ElementWidth.
unsigned LMul1VF = RISCV::RVVBitsPerBlock / ElementWidth;
// We don't support VF==1 with ELEN==32.
[[maybe_unused]] unsigned MinVF =
RISCV::RVVBitsPerBlock / Subtarget.getELen();
[[maybe_unused]] unsigned VF = N->getConstantOperandVal(2);
assert(VF >= MinVF && VF <= (LMul1VF * 8) && isPowerOf2_32(VF) &&
"Unexpected VF");
bool Fractional = VF < LMul1VF;
unsigned LMulVal = Fractional ? LMul1VF / VF : VF / LMul1VF;
unsigned VLMUL = (unsigned)RISCVVType::encodeLMUL(LMulVal, Fractional);
unsigned VSEW = RISCVVType::encodeSEW(ElementWidth);
SDLoc DL(N);
SDValue LMul = DAG.getTargetConstant(VLMUL, DL, XLenVT);
SDValue Sew = DAG.getTargetConstant(VSEW, DL, XLenVT);
SDValue AVL = DAG.getNode(ISD::ZERO_EXTEND, DL, XLenVT, N->getOperand(1));
SDValue ID = DAG.getTargetConstant(Intrinsic::riscv_vsetvli, DL, XLenVT);
SDValue Res =
DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, XLenVT, ID, AVL, Sew, LMul);
return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), Res);
}
static SDValue lowerCttzElts(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue Op0 = N->getOperand(1);
MVT OpVT = Op0.getSimpleValueType();
MVT ContainerVT = OpVT;
if (OpVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(DAG, OpVT, Subtarget);
Op0 = convertToScalableVector(ContainerVT, Op0, DAG, Subtarget);
}
MVT XLenVT = Subtarget.getXLenVT();
SDLoc DL(N);
auto [Mask, VL] = getDefaultVLOps(OpVT, ContainerVT, DL, DAG, Subtarget);
SDValue Res = DAG.getNode(RISCVISD::VFIRST_VL, DL, XLenVT, Op0, Mask, VL);
if (isOneConstant(N->getOperand(2)))
return Res;
// Convert -1 to VL.
SDValue Setcc =
DAG.getSetCC(DL, XLenVT, Res, DAG.getConstant(0, DL, XLenVT), ISD::SETLT);
VL = DAG.getElementCount(DL, XLenVT, OpVT.getVectorElementCount());
return DAG.getSelect(DL, XLenVT, Setcc, VL, Res);
}
static inline void promoteVCIXScalar(const SDValue &Op,
SmallVectorImpl<SDValue> &Operands,
SelectionDAG &DAG) {
const RISCVSubtarget &Subtarget =
DAG.getMachineFunction().getSubtarget<RISCVSubtarget>();
bool HasChain = Op.getOpcode() == ISD::INTRINSIC_VOID ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN;
unsigned IntNo = Op.getConstantOperandVal(HasChain ? 1 : 0);
SDLoc DL(Op);
const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II =
RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo);
if (!II || !II->hasScalarOperand())
return;
unsigned SplatOp = II->ScalarOperand + 1;
assert(SplatOp < Op.getNumOperands());
SDValue &ScalarOp = Operands[SplatOp];
MVT OpVT = ScalarOp.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
// The code below is partially copied from lowerVectorIntrinsicScalars.
// If this isn't a scalar, or its type is XLenVT we're done.
if (!OpVT.isScalarInteger() || OpVT == XLenVT)
return;
// Manually emit promote operation for scalar operation.
if (OpVT.bitsLT(XLenVT)) {
unsigned ExtOpc =
isa<ConstantSDNode>(ScalarOp) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
ScalarOp = DAG.getNode(ExtOpc, DL, XLenVT, ScalarOp);
}
return;
}
static void processVCIXOperands(SDValue &OrigOp,
SmallVectorImpl<SDValue> &Operands,
SelectionDAG &DAG) {
promoteVCIXScalar(OrigOp, Operands, DAG);
const RISCVSubtarget &Subtarget =
DAG.getMachineFunction().getSubtarget<RISCVSubtarget>();
for (SDValue &V : Operands) {
EVT ValType = V.getValueType();
if (ValType.isVector() && ValType.isFloatingPoint()) {
MVT InterimIVT =
MVT::getVectorVT(MVT::getIntegerVT(ValType.getScalarSizeInBits()),
ValType.getVectorElementCount());
V = DAG.getBitcast(InterimIVT, V);
}
if (ValType.isFixedLengthVector()) {
MVT OpContainerVT = getContainerForFixedLengthVector(
DAG, V.getSimpleValueType(), Subtarget);
V = convertToScalableVector(OpContainerVT, V, DAG, Subtarget);
}
}
}
// LMUL * VLEN should be greater than or equal to EGS * SEW
static inline bool isValidEGW(int EGS, EVT VT,
const RISCVSubtarget &Subtarget) {
return (Subtarget.getRealMinVLen() *
VT.getSizeInBits().getKnownMinValue()) / RISCV::RVVBitsPerBlock >=
EGS * VT.getScalarSizeInBits();
}
SDValue RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
SelectionDAG &DAG) const {
unsigned IntNo = Op.getConstantOperandVal(0);
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
switch (IntNo) {
default:
break; // Don't custom lower most intrinsics.
case Intrinsic::thread_pointer: {
EVT PtrVT = getPointerTy(DAG.getDataLayout());
return DAG.getRegister(RISCV::X4, PtrVT);
}
case Intrinsic::riscv_orc_b:
case Intrinsic::riscv_brev8:
case Intrinsic::riscv_sha256sig0:
case Intrinsic::riscv_sha256sig1:
case Intrinsic::riscv_sha256sum0:
case Intrinsic::riscv_sha256sum1:
case Intrinsic::riscv_sm3p0:
case Intrinsic::riscv_sm3p1: {
unsigned Opc;
switch (IntNo) {
case Intrinsic::riscv_orc_b: Opc = RISCVISD::ORC_B; break;
case Intrinsic::riscv_brev8: Opc = RISCVISD::BREV8; break;
case Intrinsic::riscv_sha256sig0: Opc = RISCVISD::SHA256SIG0; break;
case Intrinsic::riscv_sha256sig1: Opc = RISCVISD::SHA256SIG1; break;
case Intrinsic::riscv_sha256sum0: Opc = RISCVISD::SHA256SUM0; break;
case Intrinsic::riscv_sha256sum1: Opc = RISCVISD::SHA256SUM1; break;
case Intrinsic::riscv_sm3p0: Opc = RISCVISD::SM3P0; break;
case Intrinsic::riscv_sm3p1: Opc = RISCVISD::SM3P1; break;
}
if (RV64LegalI32 && Subtarget.is64Bit() && Op.getValueType() == MVT::i32) {
SDValue NewOp =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp);
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res);
}
return DAG.getNode(Opc, DL, XLenVT, Op.getOperand(1));
}
case Intrinsic::riscv_sm4ks:
case Intrinsic::riscv_sm4ed: {
unsigned Opc =
IntNo == Intrinsic::riscv_sm4ks ? RISCVISD::SM4KS : RISCVISD::SM4ED;
if (RV64LegalI32 && Subtarget.is64Bit() && Op.getValueType() == MVT::i32) {
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(2));
SDValue Res =
DAG.getNode(Opc, DL, MVT::i64, NewOp0, NewOp1, Op.getOperand(3));
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res);
}
return DAG.getNode(Opc, DL, XLenVT, Op.getOperand(1), Op.getOperand(2),
Op.getOperand(3));
}
case Intrinsic::riscv_zip:
case Intrinsic::riscv_unzip: {
unsigned Opc =
IntNo == Intrinsic::riscv_zip ? RISCVISD::ZIP : RISCVISD::UNZIP;
return DAG.getNode(Opc, DL, XLenVT, Op.getOperand(1));
}
case Intrinsic::riscv_mopr: {
if (RV64LegalI32 && Subtarget.is64Bit() && Op.getValueType() == MVT::i32) {
SDValue NewOp =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue Res = DAG.getNode(
RISCVISD::MOPR, DL, MVT::i64, NewOp,
DAG.getTargetConstant(Op.getConstantOperandVal(2), DL, MVT::i64));
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res);
}
return DAG.getNode(RISCVISD::MOPR, DL, XLenVT, Op.getOperand(1),
Op.getOperand(2));
}
case Intrinsic::riscv_moprr: {
if (RV64LegalI32 && Subtarget.is64Bit() && Op.getValueType() == MVT::i32) {
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(2));
SDValue Res = DAG.getNode(
RISCVISD::MOPRR, DL, MVT::i64, NewOp0, NewOp1,
DAG.getTargetConstant(Op.getConstantOperandVal(3), DL, MVT::i64));
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res);
}
return DAG.getNode(RISCVISD::MOPRR, DL, XLenVT, Op.getOperand(1),
Op.getOperand(2), Op.getOperand(3));
}
case Intrinsic::riscv_clmul:
if (RV64LegalI32 && Subtarget.is64Bit() && Op.getValueType() == MVT::i32) {
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(2));
SDValue Res = DAG.getNode(RISCVISD::CLMUL, DL, MVT::i64, NewOp0, NewOp1);
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res);
}
return DAG.getNode(RISCVISD::CLMUL, DL, XLenVT, Op.getOperand(1),
Op.getOperand(2));
case Intrinsic::riscv_clmulh:
case Intrinsic::riscv_clmulr: {
unsigned Opc =
IntNo == Intrinsic::riscv_clmulh ? RISCVISD::CLMULH : RISCVISD::CLMULR;
if (RV64LegalI32 && Subtarget.is64Bit() && Op.getValueType() == MVT::i32) {
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(2));
NewOp0 = DAG.getNode(ISD::SHL, DL, MVT::i64, NewOp0,
DAG.getConstant(32, DL, MVT::i64));
NewOp1 = DAG.getNode(ISD::SHL, DL, MVT::i64, NewOp1,
DAG.getConstant(32, DL, MVT::i64));
SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp0, NewOp1);
Res = DAG.getNode(ISD::SRL, DL, MVT::i64, Res,
DAG.getConstant(32, DL, MVT::i64));
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res);
}
return DAG.getNode(Opc, DL, XLenVT, Op.getOperand(1), Op.getOperand(2));
}
case Intrinsic::experimental_get_vector_length:
return lowerGetVectorLength(Op.getNode(), DAG, Subtarget);
case Intrinsic::experimental_cttz_elts:
return lowerCttzElts(Op.getNode(), DAG, Subtarget);
case Intrinsic::riscv_vmv_x_s: {
SDValue Res = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Op.getOperand(1));
return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Res);
}
case Intrinsic::riscv_vfmv_f_s:
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op.getValueType(),
Op.getOperand(1), DAG.getVectorIdxConstant(0, DL));
case Intrinsic::riscv_vmv_v_x:
return lowerScalarSplat(Op.getOperand(1), Op.getOperand(2),
Op.getOperand(3), Op.getSimpleValueType(), DL, DAG,
Subtarget);
case Intrinsic::riscv_vfmv_v_f:
return DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, Op.getValueType(),
Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
case Intrinsic::riscv_vmv_s_x: {
SDValue Scalar = Op.getOperand(2);
if (Scalar.getValueType().bitsLE(XLenVT)) {
Scalar = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Scalar);
return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, Op.getValueType(),
Op.getOperand(1), Scalar, Op.getOperand(3));
}
assert(Scalar.getValueType() == MVT::i64 && "Unexpected scalar VT!");
// This is an i64 value that lives in two scalar registers. We have to
// insert this in a convoluted way. First we build vXi64 splat containing
// the two values that we assemble using some bit math. Next we'll use
// vid.v and vmseq to build a mask with bit 0 set. Then we'll use that mask
// to merge element 0 from our splat into the source vector.
// FIXME: This is probably not the best way to do this, but it is
// consistent with INSERT_VECTOR_ELT lowering so it is a good starting
// point.
// sw lo, (a0)
// sw hi, 4(a0)
// vlse vX, (a0)
//
// vid.v vVid
// vmseq.vx mMask, vVid, 0
// vmerge.vvm vDest, vSrc, vVal, mMask
MVT VT = Op.getSimpleValueType();
SDValue Vec = Op.getOperand(1);
SDValue VL = getVLOperand(Op);
SDValue SplattedVal = splatSplitI64WithVL(DL, VT, SDValue(), Scalar, VL, DAG);
if (Op.getOperand(1).isUndef())
return SplattedVal;
SDValue SplattedIdx =
DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
DAG.getConstant(0, DL, MVT::i32), VL);
MVT MaskVT = getMaskTypeFor(VT);
SDValue Mask = getAllOnesMask(VT, VL, DL, DAG);
SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, VT, Mask, VL);
SDValue SelectCond =
DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT,
{VID, SplattedIdx, DAG.getCondCode(ISD::SETEQ),
DAG.getUNDEF(MaskVT), Mask, VL});
return DAG.getNode(RISCVISD::VMERGE_VL, DL, VT, SelectCond, SplattedVal,
Vec, DAG.getUNDEF(VT), VL);
}
case Intrinsic::riscv_vfmv_s_f:
return DAG.getNode(RISCVISD::VFMV_S_F_VL, DL, Op.getSimpleValueType(),
Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
// EGS * EEW >= 128 bits
case Intrinsic::riscv_vaesdf_vv:
case Intrinsic::riscv_vaesdf_vs:
case Intrinsic::riscv_vaesdm_vv:
case Intrinsic::riscv_vaesdm_vs:
case Intrinsic::riscv_vaesef_vv:
case Intrinsic::riscv_vaesef_vs:
case Intrinsic::riscv_vaesem_vv:
case Intrinsic::riscv_vaesem_vs:
case Intrinsic::riscv_vaeskf1:
case Intrinsic::riscv_vaeskf2:
case Intrinsic::riscv_vaesz_vs:
case Intrinsic::riscv_vsm4k:
case Intrinsic::riscv_vsm4r_vv:
case Intrinsic::riscv_vsm4r_vs: {
if (!isValidEGW(4, Op.getSimpleValueType(), Subtarget) ||
!isValidEGW(4, Op->getOperand(1).getSimpleValueType(), Subtarget) ||
!isValidEGW(4, Op->getOperand(2).getSimpleValueType(), Subtarget))
report_fatal_error("EGW should be greater than or equal to 4 * SEW.");
return Op;
}
// EGS * EEW >= 256 bits
case Intrinsic::riscv_vsm3c:
case Intrinsic::riscv_vsm3me: {
if (!isValidEGW(8, Op.getSimpleValueType(), Subtarget) ||
!isValidEGW(8, Op->getOperand(1).getSimpleValueType(), Subtarget))
report_fatal_error("EGW should be greater than or equal to 8 * SEW.");
return Op;
}
// zvknha(SEW=32)/zvknhb(SEW=[32|64])
case Intrinsic::riscv_vsha2ch:
case Intrinsic::riscv_vsha2cl:
case Intrinsic::riscv_vsha2ms: {
if (Op->getSimpleValueType(0).getScalarSizeInBits() == 64 &&
!Subtarget.hasStdExtZvknhb())
report_fatal_error("SEW=64 needs Zvknhb to be enabled.");
if (!isValidEGW(4, Op.getSimpleValueType(), Subtarget) ||
!isValidEGW(4, Op->getOperand(1).getSimpleValueType(), Subtarget) ||
!isValidEGW(4, Op->getOperand(2).getSimpleValueType(), Subtarget))
report_fatal_error("EGW should be greater than or equal to 4 * SEW.");
return Op;
}
case Intrinsic::riscv_sf_vc_v_x:
case Intrinsic::riscv_sf_vc_v_i:
case Intrinsic::riscv_sf_vc_v_xv:
case Intrinsic::riscv_sf_vc_v_iv:
case Intrinsic::riscv_sf_vc_v_vv:
case Intrinsic::riscv_sf_vc_v_fv:
case Intrinsic::riscv_sf_vc_v_xvv:
case Intrinsic::riscv_sf_vc_v_ivv:
case Intrinsic::riscv_sf_vc_v_vvv:
case Intrinsic::riscv_sf_vc_v_fvv:
case Intrinsic::riscv_sf_vc_v_xvw:
case Intrinsic::riscv_sf_vc_v_ivw:
case Intrinsic::riscv_sf_vc_v_vvw:
case Intrinsic::riscv_sf_vc_v_fvw: {
MVT VT = Op.getSimpleValueType();
SmallVector<SDValue> Operands{Op->op_values()};
processVCIXOperands(Op, Operands, DAG);
MVT RetVT = VT;
if (VT.isFixedLengthVector())
RetVT = getContainerForFixedLengthVector(VT);
else if (VT.isFloatingPoint())
RetVT = MVT::getVectorVT(MVT::getIntegerVT(VT.getScalarSizeInBits()),
VT.getVectorElementCount());
SDValue NewNode = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, RetVT, Operands);
if (VT.isFixedLengthVector())
NewNode = convertFromScalableVector(VT, NewNode, DAG, Subtarget);
else if (VT.isFloatingPoint())
NewNode = DAG.getBitcast(VT, NewNode);
if (Op == NewNode)
break;
return NewNode;
}
}
return lowerVectorIntrinsicScalars(Op, DAG, Subtarget);
}
static inline SDValue getVCIXISDNodeWCHAIN(SDValue &Op, SelectionDAG &DAG,
unsigned Type) {
SDLoc DL(Op);
SmallVector<SDValue> Operands{Op->op_values()};
Operands.erase(Operands.begin() + 1);
const RISCVSubtarget &Subtarget =
DAG.getMachineFunction().getSubtarget<RISCVSubtarget>();
MVT VT = Op.getSimpleValueType();
MVT RetVT = VT;
MVT FloatVT = VT;
if (VT.isFloatingPoint()) {
RetVT = MVT::getVectorVT(MVT::getIntegerVT(VT.getScalarSizeInBits()),
VT.getVectorElementCount());
FloatVT = RetVT;
}
if (VT.isFixedLengthVector())
RetVT = getContainerForFixedLengthVector(DAG.getTargetLoweringInfo(), RetVT,
Subtarget);
processVCIXOperands(Op, Operands, DAG);
SDVTList VTs = DAG.getVTList({RetVT, MVT::Other});
SDValue NewNode = DAG.getNode(Type, DL, VTs, Operands);
SDValue Chain = NewNode.getValue(1);
if (VT.isFixedLengthVector())
NewNode = convertFromScalableVector(FloatVT, NewNode, DAG, Subtarget);
if (VT.isFloatingPoint())
NewNode = DAG.getBitcast(VT, NewNode);
NewNode = DAG.getMergeValues({NewNode, Chain}, DL);
return NewNode;
}
static inline SDValue getVCIXISDNodeVOID(SDValue &Op, SelectionDAG &DAG,
unsigned Type) {
SmallVector<SDValue> Operands{Op->op_values()};
Operands.erase(Operands.begin() + 1);
processVCIXOperands(Op, Operands, DAG);
return DAG.getNode(Type, SDLoc(Op), Op.getValueType(), Operands);
}
SDValue RISCVTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
SelectionDAG &DAG) const {
unsigned IntNo = Op.getConstantOperandVal(1);
switch (IntNo) {
default:
break;
case Intrinsic::riscv_masked_strided_load: {
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
// If the mask is known to be all ones, optimize to an unmasked intrinsic;
// the selection of the masked intrinsics doesn't do this for us.
SDValue Mask = Op.getOperand(5);
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
MVT VT = Op->getSimpleValueType(0);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector())
ContainerVT = getContainerForFixedLengthVector(VT);
SDValue PassThru = Op.getOperand(2);
if (!IsUnmasked) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
if (VT.isFixedLengthVector()) {
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
PassThru = convertToScalableVector(ContainerVT, PassThru, DAG, Subtarget);
}
}
auto *Load = cast<MemIntrinsicSDNode>(Op);
SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
SDValue Ptr = Op.getOperand(3);
SDValue Stride = Op.getOperand(4);
SDValue Result, Chain;
// TODO: We restrict this to unmasked loads currently in consideration of
// the complexity of handling all falses masks.
MVT ScalarVT = ContainerVT.getVectorElementType();
if (IsUnmasked && isNullConstant(Stride) && ContainerVT.isInteger()) {
SDValue ScalarLoad =
DAG.getExtLoad(ISD::ZEXTLOAD, DL, XLenVT, Load->getChain(), Ptr,
ScalarVT, Load->getMemOperand());
Chain = ScalarLoad.getValue(1);
Result = lowerScalarSplat(SDValue(), ScalarLoad, VL, ContainerVT, DL, DAG,
Subtarget);
} else if (IsUnmasked && isNullConstant(Stride) && isTypeLegal(ScalarVT)) {
SDValue ScalarLoad = DAG.getLoad(ScalarVT, DL, Load->getChain(), Ptr,
Load->getMemOperand());
Chain = ScalarLoad.getValue(1);
Result = DAG.getSplat(ContainerVT, DL, ScalarLoad);
} else {
SDValue IntID = DAG.getTargetConstant(
IsUnmasked ? Intrinsic::riscv_vlse : Intrinsic::riscv_vlse_mask, DL,
XLenVT);
SmallVector<SDValue, 8> Ops{Load->getChain(), IntID};
if (IsUnmasked)
Ops.push_back(DAG.getUNDEF(ContainerVT));
else
Ops.push_back(PassThru);
Ops.push_back(Ptr);
Ops.push_back(Stride);
if (!IsUnmasked)
Ops.push_back(Mask);
Ops.push_back(VL);
if (!IsUnmasked) {
SDValue Policy =
DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT);
Ops.push_back(Policy);
}
SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
Result =
DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
Load->getMemoryVT(), Load->getMemOperand());
Chain = Result.getValue(1);
}
if (VT.isFixedLengthVector())
Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
return DAG.getMergeValues({Result, Chain}, DL);
}
case Intrinsic::riscv_seg2_load:
case Intrinsic::riscv_seg3_load:
case Intrinsic::riscv_seg4_load:
case Intrinsic::riscv_seg5_load:
case Intrinsic::riscv_seg6_load:
case Intrinsic::riscv_seg7_load:
case Intrinsic::riscv_seg8_load: {
SDLoc DL(Op);
static const Intrinsic::ID VlsegInts[7] = {
Intrinsic::riscv_vlseg2, Intrinsic::riscv_vlseg3,
Intrinsic::riscv_vlseg4, Intrinsic::riscv_vlseg5,
Intrinsic::riscv_vlseg6, Intrinsic::riscv_vlseg7,
Intrinsic::riscv_vlseg8};
unsigned NF = Op->getNumValues() - 1;
assert(NF >= 2 && NF <= 8 && "Unexpected seg number");
MVT XLenVT = Subtarget.getXLenVT();
MVT VT = Op->getSimpleValueType(0);
MVT ContainerVT = getContainerForFixedLengthVector(VT);
SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG,
Subtarget);
SDValue IntID = DAG.getTargetConstant(VlsegInts[NF - 2], DL, XLenVT);
auto *Load = cast<MemIntrinsicSDNode>(Op);
SmallVector<EVT, 9> ContainerVTs(NF, ContainerVT);
ContainerVTs.push_back(MVT::Other);
SDVTList VTs = DAG.getVTList(ContainerVTs);
SmallVector<SDValue, 12> Ops = {Load->getChain(), IntID};
Ops.insert(Ops.end(), NF, DAG.getUNDEF(ContainerVT));
Ops.push_back(Op.getOperand(2));
Ops.push_back(VL);
SDValue Result =
DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
Load->getMemoryVT(), Load->getMemOperand());
SmallVector<SDValue, 9> Results;
for (unsigned int RetIdx = 0; RetIdx < NF; RetIdx++)
Results.push_back(convertFromScalableVector(VT, Result.getValue(RetIdx),
DAG, Subtarget));
Results.push_back(Result.getValue(NF));
return DAG.getMergeValues(Results, DL);
}
case Intrinsic::riscv_sf_vc_v_x_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_X_SE);
case Intrinsic::riscv_sf_vc_v_i_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_I_SE);
case Intrinsic::riscv_sf_vc_v_xv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_XV_SE);
case Intrinsic::riscv_sf_vc_v_iv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_IV_SE);
case Intrinsic::riscv_sf_vc_v_vv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_VV_SE);
case Intrinsic::riscv_sf_vc_v_fv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_FV_SE);
case Intrinsic::riscv_sf_vc_v_xvv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_XVV_SE);
case Intrinsic::riscv_sf_vc_v_ivv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_IVV_SE);
case Intrinsic::riscv_sf_vc_v_vvv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_VVV_SE);
case Intrinsic::riscv_sf_vc_v_fvv_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_FVV_SE);
case Intrinsic::riscv_sf_vc_v_xvw_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_XVW_SE);
case Intrinsic::riscv_sf_vc_v_ivw_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_IVW_SE);
case Intrinsic::riscv_sf_vc_v_vvw_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_VVW_SE);
case Intrinsic::riscv_sf_vc_v_fvw_se:
return getVCIXISDNodeWCHAIN(Op, DAG, RISCVISD::SF_VC_V_FVW_SE);
}
return lowerVectorIntrinsicScalars(Op, DAG, Subtarget);
}
SDValue RISCVTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
SelectionDAG &DAG) const {
unsigned IntNo = Op.getConstantOperandVal(1);
switch (IntNo) {
default:
break;
case Intrinsic::riscv_masked_strided_store: {
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
// If the mask is known to be all ones, optimize to an unmasked intrinsic;
// the selection of the masked intrinsics doesn't do this for us.
SDValue Mask = Op.getOperand(5);
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
SDValue Val = Op.getOperand(2);
MVT VT = Val.getSimpleValueType();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
Val = convertToScalableVector(ContainerVT, Val, DAG, Subtarget);
}
if (!IsUnmasked) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
if (VT.isFixedLengthVector())
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
SDValue IntID = DAG.getTargetConstant(
IsUnmasked ? Intrinsic::riscv_vsse : Intrinsic::riscv_vsse_mask, DL,
XLenVT);
auto *Store = cast<MemIntrinsicSDNode>(Op);
SmallVector<SDValue, 8> Ops{Store->getChain(), IntID};
Ops.push_back(Val);
Ops.push_back(Op.getOperand(3)); // Ptr
Ops.push_back(Op.getOperand(4)); // Stride
if (!IsUnmasked)
Ops.push_back(Mask);
Ops.push_back(VL);
return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL, Store->getVTList(),
Ops, Store->getMemoryVT(),
Store->getMemOperand());
}
case Intrinsic::riscv_seg2_store:
case Intrinsic::riscv_seg3_store:
case Intrinsic::riscv_seg4_store:
case Intrinsic::riscv_seg5_store:
case Intrinsic::riscv_seg6_store:
case Intrinsic::riscv_seg7_store:
case Intrinsic::riscv_seg8_store: {
SDLoc DL(Op);
static const Intrinsic::ID VssegInts[] = {
Intrinsic::riscv_vsseg2, Intrinsic::riscv_vsseg3,
Intrinsic::riscv_vsseg4, Intrinsic::riscv_vsseg5,
Intrinsic::riscv_vsseg6, Intrinsic::riscv_vsseg7,
Intrinsic::riscv_vsseg8};
// Operands are (chain, int_id, vec*, ptr, vl)
unsigned NF = Op->getNumOperands() - 4;
assert(NF >= 2 && NF <= 8 && "Unexpected seg number");
MVT XLenVT = Subtarget.getXLenVT();
MVT VT = Op->getOperand(2).getSimpleValueType();
MVT ContainerVT = getContainerForFixedLengthVector(VT);
SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG,
Subtarget);
SDValue IntID = DAG.getTargetConstant(VssegInts[NF - 2], DL, XLenVT);
SDValue Ptr = Op->getOperand(NF + 2);
auto *FixedIntrinsic = cast<MemIntrinsicSDNode>(Op);
SmallVector<SDValue, 12> Ops = {FixedIntrinsic->getChain(), IntID};
for (unsigned i = 0; i < NF; i++)
Ops.push_back(convertToScalableVector(
ContainerVT, FixedIntrinsic->getOperand(2 + i), DAG, Subtarget));
Ops.append({Ptr, VL});
return DAG.getMemIntrinsicNode(
ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other), Ops,
FixedIntrinsic->getMemoryVT(), FixedIntrinsic->getMemOperand());
}
case Intrinsic::riscv_sf_vc_xv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_XV_SE);
case Intrinsic::riscv_sf_vc_iv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_IV_SE);
case Intrinsic::riscv_sf_vc_vv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_VV_SE);
case Intrinsic::riscv_sf_vc_fv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_FV_SE);
case Intrinsic::riscv_sf_vc_xvv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_XVV_SE);
case Intrinsic::riscv_sf_vc_ivv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_IVV_SE);
case Intrinsic::riscv_sf_vc_vvv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_VVV_SE);
case Intrinsic::riscv_sf_vc_fvv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_FVV_SE);
case Intrinsic::riscv_sf_vc_xvw_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_XVW_SE);
case Intrinsic::riscv_sf_vc_ivw_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_IVW_SE);
case Intrinsic::riscv_sf_vc_vvw_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_VVW_SE);
case Intrinsic::riscv_sf_vc_fvw_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_FVW_SE);
}
return lowerVectorIntrinsicScalars(Op, DAG, Subtarget);
}
static unsigned getRVVReductionOp(unsigned ISDOpcode) {
switch (ISDOpcode) {
default:
llvm_unreachable("Unhandled reduction");
case ISD::VP_REDUCE_ADD:
case ISD::VECREDUCE_ADD:
return RISCVISD::VECREDUCE_ADD_VL;
case ISD::VP_REDUCE_UMAX:
case ISD::VECREDUCE_UMAX:
return RISCVISD::VECREDUCE_UMAX_VL;
case ISD::VP_REDUCE_SMAX:
case ISD::VECREDUCE_SMAX:
return RISCVISD::VECREDUCE_SMAX_VL;
case ISD::VP_REDUCE_UMIN:
case ISD::VECREDUCE_UMIN:
return RISCVISD::VECREDUCE_UMIN_VL;
case ISD::VP_REDUCE_SMIN:
case ISD::VECREDUCE_SMIN:
return RISCVISD::VECREDUCE_SMIN_VL;
case ISD::VP_REDUCE_AND:
case ISD::VECREDUCE_AND:
return RISCVISD::VECREDUCE_AND_VL;
case ISD::VP_REDUCE_OR:
case ISD::VECREDUCE_OR:
return RISCVISD::VECREDUCE_OR_VL;
case ISD::VP_REDUCE_XOR:
case ISD::VECREDUCE_XOR:
return RISCVISD::VECREDUCE_XOR_VL;
case ISD::VP_REDUCE_FADD:
return RISCVISD::VECREDUCE_FADD_VL;
case ISD::VP_REDUCE_SEQ_FADD:
return RISCVISD::VECREDUCE_SEQ_FADD_VL;
case ISD::VP_REDUCE_FMAX:
return RISCVISD::VECREDUCE_FMAX_VL;
case ISD::VP_REDUCE_FMIN:
return RISCVISD::VECREDUCE_FMIN_VL;
}
}
SDValue RISCVTargetLowering::lowerVectorMaskVecReduction(SDValue Op,
SelectionDAG &DAG,
bool IsVP) const {
SDLoc DL(Op);
SDValue Vec = Op.getOperand(IsVP ? 1 : 0);
MVT VecVT = Vec.getSimpleValueType();
assert((Op.getOpcode() == ISD::VECREDUCE_AND ||
Op.getOpcode() == ISD::VECREDUCE_OR ||
Op.getOpcode() == ISD::VECREDUCE_XOR ||
Op.getOpcode() == ISD::VP_REDUCE_AND ||
Op.getOpcode() == ISD::VP_REDUCE_OR ||
Op.getOpcode() == ISD::VP_REDUCE_XOR) &&
"Unexpected reduction lowering");
MVT XLenVT = Subtarget.getXLenVT();
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
SDValue Mask, VL;
if (IsVP) {
Mask = Op.getOperand(2);
VL = Op.getOperand(3);
} else {
std::tie(Mask, VL) =
getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
}
unsigned BaseOpc;
ISD::CondCode CC;
SDValue Zero = DAG.getConstant(0, DL, XLenVT);
switch (Op.getOpcode()) {
default:
llvm_unreachable("Unhandled reduction");
case ISD::VECREDUCE_AND:
case ISD::VP_REDUCE_AND: {
// vcpop ~x == 0
SDValue TrueMask = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
Vec = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Vec, TrueMask, VL);
Vec = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Vec, Mask, VL);
CC = ISD::SETEQ;
BaseOpc = ISD::AND;
break;
}
case ISD::VECREDUCE_OR:
case ISD::VP_REDUCE_OR:
// vcpop x != 0
Vec = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Vec, Mask, VL);
CC = ISD::SETNE;
BaseOpc = ISD::OR;
break;
case ISD::VECREDUCE_XOR:
case ISD::VP_REDUCE_XOR: {
// ((vcpop x) & 1) != 0
SDValue One = DAG.getConstant(1, DL, XLenVT);
Vec = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Vec, Mask, VL);
Vec = DAG.getNode(ISD::AND, DL, XLenVT, Vec, One);
CC = ISD::SETNE;
BaseOpc = ISD::XOR;
break;
}
}
SDValue SetCC = DAG.getSetCC(DL, XLenVT, Vec, Zero, CC);
SetCC = DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), SetCC);
if (!IsVP)
return SetCC;
// Now include the start value in the operation.
// Note that we must return the start value when no elements are operated
// upon. The vcpop instructions we've emitted in each case above will return
// 0 for an inactive vector, and so we've already received the neutral value:
// AND gives us (0 == 0) -> 1 and OR/XOR give us (0 != 0) -> 0. Therefore we
// can simply include the start value.
return DAG.getNode(BaseOpc, DL, Op.getValueType(), SetCC, Op.getOperand(0));
}
static bool isNonZeroAVL(SDValue AVL) {
auto *RegisterAVL = dyn_cast<RegisterSDNode>(AVL);
auto *ImmAVL = dyn_cast<ConstantSDNode>(AVL);
return (RegisterAVL && RegisterAVL->getReg() == RISCV::X0) ||
(ImmAVL && ImmAVL->getZExtValue() >= 1);
}
/// Helper to lower a reduction sequence of the form:
/// scalar = reduce_op vec, scalar_start
static SDValue lowerReductionSeq(unsigned RVVOpcode, MVT ResVT,
SDValue StartValue, SDValue Vec, SDValue Mask,
SDValue VL, const SDLoc &DL, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
const MVT VecVT = Vec.getSimpleValueType();
const MVT M1VT = getLMUL1VT(VecVT);
const MVT XLenVT = Subtarget.getXLenVT();
const bool NonZeroAVL = isNonZeroAVL(VL);
// The reduction needs an LMUL1 input; do the splat at either LMUL1
// or the original VT if fractional.
auto InnerVT = VecVT.bitsLE(M1VT) ? VecVT : M1VT;
// We reuse the VL of the reduction to reduce vsetvli toggles if we can
// prove it is non-zero. For the AVL=0 case, we need the scalar to
// be the result of the reduction operation.
auto InnerVL = NonZeroAVL ? VL : DAG.getConstant(1, DL, XLenVT);
SDValue InitialValue = lowerScalarInsert(StartValue, InnerVL, InnerVT, DL,
DAG, Subtarget);
if (M1VT != InnerVT)
InitialValue =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, M1VT, DAG.getUNDEF(M1VT),
InitialValue, DAG.getVectorIdxConstant(0, DL));
SDValue PassThru = NonZeroAVL ? DAG.getUNDEF(M1VT) : InitialValue;
SDValue Policy = DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT);
SDValue Ops[] = {PassThru, Vec, InitialValue, Mask, VL, Policy};
SDValue Reduction = DAG.getNode(RVVOpcode, DL, M1VT, Ops);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Reduction,
DAG.getVectorIdxConstant(0, DL));
}
SDValue RISCVTargetLowering::lowerVECREDUCE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Vec = Op.getOperand(0);
EVT VecEVT = Vec.getValueType();
unsigned BaseOpc = ISD::getVecReduceBaseOpcode(Op.getOpcode());
// Due to ordering in legalize types we may have a vector type that needs to
// be split. Do that manually so we can get down to a legal type.
while (getTypeAction(*DAG.getContext(), VecEVT) ==
TargetLowering::TypeSplitVector) {
auto [Lo, Hi] = DAG.SplitVector(Vec, DL);
VecEVT = Lo.getValueType();
Vec = DAG.getNode(BaseOpc, DL, VecEVT, Lo, Hi);
}
// TODO: The type may need to be widened rather than split. Or widened before
// it can be split.
if (!isTypeLegal(VecEVT))
return SDValue();
MVT VecVT = VecEVT.getSimpleVT();
MVT VecEltVT = VecVT.getVectorElementType();
unsigned RVVOpcode = getRVVReductionOp(Op.getOpcode());
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
SDValue StartV = DAG.getNeutralElement(BaseOpc, DL, VecEltVT, SDNodeFlags());
switch (BaseOpc) {
case ISD::AND:
case ISD::OR:
case ISD::UMAX:
case ISD::UMIN:
case ISD::SMAX:
case ISD::SMIN:
StartV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VecEltVT, Vec,
DAG.getVectorIdxConstant(0, DL));
}
return lowerReductionSeq(RVVOpcode, Op.getSimpleValueType(), StartV, Vec,
Mask, VL, DL, DAG, Subtarget);
}
// Given a reduction op, this function returns the matching reduction opcode,
// the vector SDValue and the scalar SDValue required to lower this to a
// RISCVISD node.
static std::tuple<unsigned, SDValue, SDValue>
getRVVFPReductionOpAndOperands(SDValue Op, SelectionDAG &DAG, EVT EltVT,
const RISCVSubtarget &Subtarget) {
SDLoc DL(Op);
auto Flags = Op->getFlags();
unsigned Opcode = Op.getOpcode();
switch (Opcode) {
default:
llvm_unreachable("Unhandled reduction");
case ISD::VECREDUCE_FADD: {
// Use positive zero if we can. It is cheaper to materialize.
SDValue Zero =
DAG.getConstantFP(Flags.hasNoSignedZeros() ? 0.0 : -0.0, DL, EltVT);
return std::make_tuple(RISCVISD::VECREDUCE_FADD_VL, Op.getOperand(0), Zero);
}
case ISD::VECREDUCE_SEQ_FADD:
return std::make_tuple(RISCVISD::VECREDUCE_SEQ_FADD_VL, Op.getOperand(1),
Op.getOperand(0));
case ISD::VECREDUCE_FMINIMUM:
case ISD::VECREDUCE_FMAXIMUM:
case ISD::VECREDUCE_FMIN:
case ISD::VECREDUCE_FMAX: {
SDValue Front =
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Op.getOperand(0),
DAG.getVectorIdxConstant(0, DL));
unsigned RVVOpc =
(Opcode == ISD::VECREDUCE_FMIN || Opcode == ISD::VECREDUCE_FMINIMUM)
? RISCVISD::VECREDUCE_FMIN_VL
: RISCVISD::VECREDUCE_FMAX_VL;
return std::make_tuple(RVVOpc, Op.getOperand(0), Front);
}
}
}
SDValue RISCVTargetLowering::lowerFPVECREDUCE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VecEltVT = Op.getSimpleValueType();
unsigned RVVOpcode;
SDValue VectorVal, ScalarVal;
std::tie(RVVOpcode, VectorVal, ScalarVal) =
getRVVFPReductionOpAndOperands(Op, DAG, VecEltVT, Subtarget);
MVT VecVT = VectorVal.getSimpleValueType();
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
VectorVal = convertToScalableVector(ContainerVT, VectorVal, DAG, Subtarget);
}
MVT ResVT = Op.getSimpleValueType();
auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
SDValue Res = lowerReductionSeq(RVVOpcode, ResVT, ScalarVal, VectorVal, Mask,
VL, DL, DAG, Subtarget);
if (Op.getOpcode() != ISD::VECREDUCE_FMINIMUM &&
Op.getOpcode() != ISD::VECREDUCE_FMAXIMUM)
return Res;
if (Op->getFlags().hasNoNaNs())
return Res;
// Force output to NaN if any element is Nan.
SDValue IsNan =
DAG.getNode(RISCVISD::SETCC_VL, DL, Mask.getValueType(),
{VectorVal, VectorVal, DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(Mask.getValueType()), Mask, VL});
MVT XLenVT = Subtarget.getXLenVT();
SDValue CPop = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, IsNan, Mask, VL);
SDValue NoNaNs = DAG.getSetCC(DL, XLenVT, CPop,
DAG.getConstant(0, DL, XLenVT), ISD::SETEQ);
return DAG.getSelect(
DL, ResVT, NoNaNs, Res,
DAG.getConstantFP(APFloat::getNaN(DAG.EVTToAPFloatSemantics(ResVT)), DL,
ResVT));
}
SDValue RISCVTargetLowering::lowerVPREDUCE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Vec = Op.getOperand(1);
EVT VecEVT = Vec.getValueType();
// TODO: The type may need to be widened rather than split. Or widened before
// it can be split.
if (!isTypeLegal(VecEVT))
return SDValue();
MVT VecVT = VecEVT.getSimpleVT();
unsigned RVVOpcode = getRVVReductionOp(Op.getOpcode());
if (VecVT.isFixedLengthVector()) {
auto ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
SDValue VL = Op.getOperand(3);
SDValue Mask = Op.getOperand(2);
return lowerReductionSeq(RVVOpcode, Op.getSimpleValueType(), Op.getOperand(0),
Vec, Mask, VL, DL, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op,
SelectionDAG &DAG) const {
SDValue Vec = Op.getOperand(0);
SDValue SubVec = Op.getOperand(1);
MVT VecVT = Vec.getSimpleValueType();
MVT SubVecVT = SubVec.getSimpleValueType();
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
unsigned OrigIdx = Op.getConstantOperandVal(2);
const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
// We don't have the ability to slide mask vectors up indexed by their i1
// elements; the smallest we can do is i8. Often we are able to bitcast to
// equivalent i8 vectors. Note that when inserting a fixed-length vector
// into a scalable one, we might not necessarily have enough scalable
// elements to safely divide by 8: nxv1i1 = insert nxv1i1, v4i1 is valid.
if (SubVecVT.getVectorElementType() == MVT::i1 &&
(OrigIdx != 0 || !Vec.isUndef())) {
if (VecVT.getVectorMinNumElements() >= 8 &&
SubVecVT.getVectorMinNumElements() >= 8) {
assert(OrigIdx % 8 == 0 && "Invalid index");
assert(VecVT.getVectorMinNumElements() % 8 == 0 &&
SubVecVT.getVectorMinNumElements() % 8 == 0 &&
"Unexpected mask vector lowering");
OrigIdx /= 8;
SubVecVT =
MVT::getVectorVT(MVT::i8, SubVecVT.getVectorMinNumElements() / 8,
SubVecVT.isScalableVector());
VecVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorMinNumElements() / 8,
VecVT.isScalableVector());
Vec = DAG.getBitcast(VecVT, Vec);
SubVec = DAG.getBitcast(SubVecVT, SubVec);
} else {
// We can't slide this mask vector up indexed by its i1 elements.
// This poses a problem when we wish to insert a scalable vector which
// can't be re-expressed as a larger type. Just choose the slow path and
// extend to a larger type, then truncate back down.
MVT ExtVecVT = VecVT.changeVectorElementType(MVT::i8);
MVT ExtSubVecVT = SubVecVT.changeVectorElementType(MVT::i8);
Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVecVT, Vec);
SubVec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtSubVecVT, SubVec);
Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ExtVecVT, Vec, SubVec,
Op.getOperand(2));
SDValue SplatZero = DAG.getConstant(0, DL, ExtVecVT);
return DAG.getSetCC(DL, VecVT, Vec, SplatZero, ISD::SETNE);
}
}
// If the subvector vector is a fixed-length type, we cannot use subregister
// manipulation to simplify the codegen; we don't know which register of a
// LMUL group contains the specific subvector as we only know the minimum
// register size. Therefore we must slide the vector group up the full
// amount.
if (SubVecVT.isFixedLengthVector()) {
if (OrigIdx == 0 && Vec.isUndef() && !VecVT.isFixedLengthVector())
return Op;
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
if (OrigIdx == 0 && Vec.isUndef() && VecVT.isFixedLengthVector()) {
SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SubVec,
DAG.getVectorIdxConstant(0, DL));
SubVec = convertFromScalableVector(VecVT, SubVec, DAG, Subtarget);
return DAG.getBitcast(Op.getValueType(), SubVec);
}
SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SubVec,
DAG.getVectorIdxConstant(0, DL));
SDValue Mask =
getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first;
// Set the vector length to only the number of elements we care about. Note
// that for slideup this includes the offset.
unsigned EndIndex = OrigIdx + SubVecVT.getVectorNumElements();
SDValue VL = getVLOp(EndIndex, ContainerVT, DL, DAG, Subtarget);
// Use tail agnostic policy if we're inserting over Vec's tail.
unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
if (VecVT.isFixedLengthVector() && EndIndex == VecVT.getVectorNumElements())
Policy = RISCVII::TAIL_AGNOSTIC;
// If we're inserting into the lowest elements, use a tail undisturbed
// vmv.v.v.
if (OrigIdx == 0) {
SubVec =
DAG.getNode(RISCVISD::VMV_V_V_VL, DL, ContainerVT, Vec, SubVec, VL);
} else {
SDValue SlideupAmt = DAG.getConstant(OrigIdx, DL, XLenVT);
SubVec = getVSlideup(DAG, Subtarget, DL, ContainerVT, Vec, SubVec,
SlideupAmt, Mask, VL, Policy);
}
if (VecVT.isFixedLengthVector())
SubVec = convertFromScalableVector(VecVT, SubVec, DAG, Subtarget);
return DAG.getBitcast(Op.getValueType(), SubVec);
}
unsigned SubRegIdx, RemIdx;
std::tie(SubRegIdx, RemIdx) =
RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
VecVT, SubVecVT, OrigIdx, TRI);
RISCVII::VLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecVT);
bool IsSubVecPartReg = SubVecLMUL == RISCVII::VLMUL::LMUL_F2 ||
SubVecLMUL == RISCVII::VLMUL::LMUL_F4 ||
SubVecLMUL == RISCVII::VLMUL::LMUL_F8;
// 1. If the Idx has been completely eliminated and this subvector's size is
// a vector register or a multiple thereof, or the surrounding elements are
// undef, then this is a subvector insert which naturally aligns to a vector
// register. These can easily be handled using subregister manipulation.
// 2. If the subvector is smaller than a vector register, then the insertion
// must preserve the undisturbed elements of the register. We do this by
// lowering to an EXTRACT_SUBVECTOR grabbing the nearest LMUL=1 vector type
// (which resolves to a subregister copy), performing a VSLIDEUP to place the
// subvector within the vector register, and an INSERT_SUBVECTOR of that
// LMUL=1 type back into the larger vector (resolving to another subregister
// operation). See below for how our VSLIDEUP works. We go via a LMUL=1 type
// to avoid allocating a large register group to hold our subvector.
if (RemIdx == 0 && (!IsSubVecPartReg || Vec.isUndef()))
return Op;
// VSLIDEUP works by leaving elements 0<i<OFFSET undisturbed, elements
// OFFSET<=i<VL set to the "subvector" and vl<=i<VLMAX set to the tail policy
// (in our case undisturbed). This means we can set up a subvector insertion
// where OFFSET is the insertion offset, and the VL is the OFFSET plus the
// size of the subvector.
MVT InterSubVT = VecVT;
SDValue AlignedExtract = Vec;
unsigned AlignedIdx = OrigIdx - RemIdx;
if (VecVT.bitsGT(getLMUL1VT(VecVT))) {
InterSubVT = getLMUL1VT(VecVT);
// Extract a subvector equal to the nearest full vector register type. This
// should resolve to a EXTRACT_SUBREG instruction.
AlignedExtract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InterSubVT, Vec,
DAG.getVectorIdxConstant(AlignedIdx, DL));
}
SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InterSubVT,
DAG.getUNDEF(InterSubVT), SubVec,
DAG.getVectorIdxConstant(0, DL));
auto [Mask, VL] = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget);
ElementCount EndIndex =
ElementCount::getScalable(RemIdx) + SubVecVT.getVectorElementCount();
VL = computeVLMax(SubVecVT, DL, DAG);
// Use tail agnostic policy if we're inserting over InterSubVT's tail.
unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
if (EndIndex == InterSubVT.getVectorElementCount())
Policy = RISCVII::TAIL_AGNOSTIC;
// If we're inserting into the lowest elements, use a tail undisturbed
// vmv.v.v.
if (RemIdx == 0) {
SubVec = DAG.getNode(RISCVISD::VMV_V_V_VL, DL, InterSubVT, AlignedExtract,
SubVec, VL);
} else {
SDValue SlideupAmt =
DAG.getVScale(DL, XLenVT, APInt(XLenVT.getSizeInBits(), RemIdx));
// Construct the vector length corresponding to RemIdx + length(SubVecVT).
VL = DAG.getNode(ISD::ADD, DL, XLenVT, SlideupAmt, VL);
SubVec = getVSlideup(DAG, Subtarget, DL, InterSubVT, AlignedExtract, SubVec,
SlideupAmt, Mask, VL, Policy);
}
// If required, insert this subvector back into the correct vector register.
// This should resolve to an INSERT_SUBREG instruction.
if (VecVT.bitsGT(InterSubVT))
SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, Vec, SubVec,
DAG.getVectorIdxConstant(AlignedIdx, DL));
// We might have bitcast from a mask type: cast back to the original type if
// required.
return DAG.getBitcast(Op.getSimpleValueType(), SubVec);
}
SDValue RISCVTargetLowering::lowerEXTRACT_SUBVECTOR(SDValue Op,
SelectionDAG &DAG) const {
SDValue Vec = Op.getOperand(0);
MVT SubVecVT = Op.getSimpleValueType();
MVT VecVT = Vec.getSimpleValueType();
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
unsigned OrigIdx = Op.getConstantOperandVal(1);
const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
// We don't have the ability to slide mask vectors down indexed by their i1
// elements; the smallest we can do is i8. Often we are able to bitcast to
// equivalent i8 vectors. Note that when extracting a fixed-length vector
// from a scalable one, we might not necessarily have enough scalable
// elements to safely divide by 8: v8i1 = extract nxv1i1 is valid.
if (SubVecVT.getVectorElementType() == MVT::i1 && OrigIdx != 0) {
if (VecVT.getVectorMinNumElements() >= 8 &&
SubVecVT.getVectorMinNumElements() >= 8) {
assert(OrigIdx % 8 == 0 && "Invalid index");
assert(VecVT.getVectorMinNumElements() % 8 == 0 &&
SubVecVT.getVectorMinNumElements() % 8 == 0 &&
"Unexpected mask vector lowering");
OrigIdx /= 8;
SubVecVT =
MVT::getVectorVT(MVT::i8, SubVecVT.getVectorMinNumElements() / 8,
SubVecVT.isScalableVector());
VecVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorMinNumElements() / 8,
VecVT.isScalableVector());
Vec = DAG.getBitcast(VecVT, Vec);
} else {
// We can't slide this mask vector down, indexed by its i1 elements.
// This poses a problem when we wish to extract a scalable vector which
// can't be re-expressed as a larger type. Just choose the slow path and
// extend to a larger type, then truncate back down.
// TODO: We could probably improve this when extracting certain fixed
// from fixed, where we can extract as i8 and shift the correct element
// right to reach the desired subvector?
MVT ExtVecVT = VecVT.changeVectorElementType(MVT::i8);
MVT ExtSubVecVT = SubVecVT.changeVectorElementType(MVT::i8);
Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVecVT, Vec);
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtSubVecVT, Vec,
Op.getOperand(1));
SDValue SplatZero = DAG.getConstant(0, DL, ExtSubVecVT);
return DAG.getSetCC(DL, SubVecVT, Vec, SplatZero, ISD::SETNE);
}
}
// With an index of 0 this is a cast-like subvector, which can be performed
// with subregister operations.
if (OrigIdx == 0)
return Op;
const auto VLen = Subtarget.getRealVLen();
// If the subvector vector is a fixed-length type and we don't know VLEN
// exactly, we cannot use subregister manipulation to simplify the codegen; we
// don't know which register of a LMUL group contains the specific subvector
// as we only know the minimum register size. Therefore we must slide the
// vector group down the full amount.
if (SubVecVT.isFixedLengthVector() && !VLen) {
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
// Shrink down Vec so we're performing the slidedown on a smaller LMUL.
unsigned LastIdx = OrigIdx + SubVecVT.getVectorNumElements() - 1;
if (auto ShrunkVT =
getSmallestVTForIndex(ContainerVT, LastIdx, DL, DAG, Subtarget)) {
ContainerVT = *ShrunkVT;
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ContainerVT, Vec,
DAG.getVectorIdxConstant(0, DL));
}
SDValue Mask =
getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first;
// Set the vector length to only the number of elements we care about. This
// avoids sliding down elements we're going to discard straight away.
SDValue VL = getVLOp(SubVecVT.getVectorNumElements(), ContainerVT, DL, DAG,
Subtarget);
SDValue SlidedownAmt = DAG.getConstant(OrigIdx, DL, XLenVT);
SDValue Slidedown =
getVSlidedown(DAG, Subtarget, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), Vec, SlidedownAmt, Mask, VL);
// Now we can use a cast-like subvector extract to get the result.
Slidedown = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, Slidedown,
DAG.getVectorIdxConstant(0, DL));
return DAG.getBitcast(Op.getValueType(), Slidedown);
}
if (VecVT.isFixedLengthVector()) {
VecVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(VecVT, Vec, DAG, Subtarget);
}
MVT ContainerSubVecVT = SubVecVT;
if (SubVecVT.isFixedLengthVector())
ContainerSubVecVT = getContainerForFixedLengthVector(SubVecVT);
unsigned SubRegIdx;
ElementCount RemIdx;
// extract_subvector scales the index by vscale if the subvector is scalable,
// and decomposeSubvectorInsertExtractToSubRegs takes this into account. So if
// we have a fixed length subvector, we need to adjust the index by 1/vscale.
if (SubVecVT.isFixedLengthVector()) {
assert(VLen);
unsigned Vscale = *VLen / RISCV::RVVBitsPerBlock;
auto Decompose =
RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
VecVT, ContainerSubVecVT, OrigIdx / Vscale, TRI);
SubRegIdx = Decompose.first;
RemIdx = ElementCount::getFixed((Decompose.second * Vscale) +
(OrigIdx % Vscale));
} else {
auto Decompose =
RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
VecVT, ContainerSubVecVT, OrigIdx, TRI);
SubRegIdx = Decompose.first;
RemIdx = ElementCount::getScalable(Decompose.second);
}
// If the Idx has been completely eliminated then this is a subvector extract
// which naturally aligns to a vector register. These can easily be handled
// using subregister manipulation.
if (RemIdx.isZero()) {
if (SubVecVT.isFixedLengthVector()) {
Vec = DAG.getTargetExtractSubreg(SubRegIdx, DL, ContainerSubVecVT, Vec);
return convertFromScalableVector(SubVecVT, Vec, DAG, Subtarget);
}
return Op;
}
// Else SubVecVT is M1 or smaller and may need to be slid down: if SubVecVT
// was > M1 then the index would need to be a multiple of VLMAX, and so would
// divide exactly.
assert(RISCVVType::decodeVLMUL(getLMUL(ContainerSubVecVT)).second ||
getLMUL(ContainerSubVecVT) == RISCVII::VLMUL::LMUL_1);
// If the vector type is an LMUL-group type, extract a subvector equal to the
// nearest full vector register type.
MVT InterSubVT = VecVT;
if (VecVT.bitsGT(getLMUL1VT(VecVT))) {
// If VecVT has an LMUL > 1, then SubVecVT should have a smaller LMUL, and
// we should have successfully decomposed the extract into a subregister.
assert(SubRegIdx != RISCV::NoSubRegister);
InterSubVT = getLMUL1VT(VecVT);
Vec = DAG.getTargetExtractSubreg(SubRegIdx, DL, InterSubVT, Vec);
}
// Slide this vector register down by the desired number of elements in order
// to place the desired subvector starting at element 0.
SDValue SlidedownAmt = DAG.getElementCount(DL, XLenVT, RemIdx);
auto [Mask, VL] = getDefaultScalableVLOps(InterSubVT, DL, DAG, Subtarget);
if (SubVecVT.isFixedLengthVector())
VL = getVLOp(SubVecVT.getVectorNumElements(), InterSubVT, DL, DAG,
Subtarget);
SDValue Slidedown =
getVSlidedown(DAG, Subtarget, DL, InterSubVT, DAG.getUNDEF(InterSubVT),
Vec, SlidedownAmt, Mask, VL);
// Now the vector is in the right position, extract our final subvector. This
// should resolve to a COPY.
Slidedown = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, Slidedown,
DAG.getVectorIdxConstant(0, DL));
// We might have bitcast from a mask type: cast back to the original type if
// required.
return DAG.getBitcast(Op.getSimpleValueType(), Slidedown);
}
// Widen a vector's operands to i8, then truncate its results back to the
// original type, typically i1. All operand and result types must be the same.
static SDValue widenVectorOpsToi8(SDValue N, const SDLoc &DL,
SelectionDAG &DAG) {
MVT VT = N.getSimpleValueType();
MVT WideVT = VT.changeVectorElementType(MVT::i8);
SmallVector<SDValue, 4> WideOps;
for (SDValue Op : N->ops()) {
assert(Op.getSimpleValueType() == VT &&
"Operands and result must be same type");
WideOps.push_back(DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT, Op));
}
unsigned NumVals = N->getNumValues();
SDVTList VTs = DAG.getVTList(SmallVector<EVT, 4>(
NumVals, N.getValueType().changeVectorElementType(MVT::i8)));
SDValue WideN = DAG.getNode(N.getOpcode(), DL, VTs, WideOps);
SmallVector<SDValue, 4> TruncVals;
for (unsigned I = 0; I < NumVals; I++) {
TruncVals.push_back(
DAG.getSetCC(DL, N->getSimpleValueType(I), WideN.getValue(I),
DAG.getConstant(0, DL, WideVT), ISD::SETNE));
}
if (TruncVals.size() > 1)
return DAG.getMergeValues(TruncVals, DL);
return TruncVals.front();
}
SDValue RISCVTargetLowering::lowerVECTOR_DEINTERLEAVE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VecVT = Op.getSimpleValueType();
assert(VecVT.isScalableVector() &&
"vector_interleave on non-scalable vector!");
// 1 bit element vectors need to be widened to e8
if (VecVT.getVectorElementType() == MVT::i1)
return widenVectorOpsToi8(Op, DL, DAG);
// If the VT is LMUL=8, we need to split and reassemble.
if (VecVT.getSizeInBits().getKnownMinValue() ==
(8 * RISCV::RVVBitsPerBlock)) {
auto [Op0Lo, Op0Hi] = DAG.SplitVectorOperand(Op.getNode(), 0);
auto [Op1Lo, Op1Hi] = DAG.SplitVectorOperand(Op.getNode(), 1);
EVT SplitVT = Op0Lo.getValueType();
SDValue ResLo = DAG.getNode(ISD::VECTOR_DEINTERLEAVE, DL,
DAG.getVTList(SplitVT, SplitVT), Op0Lo, Op0Hi);
SDValue ResHi = DAG.getNode(ISD::VECTOR_DEINTERLEAVE, DL,
DAG.getVTList(SplitVT, SplitVT), Op1Lo, Op1Hi);
SDValue Even = DAG.getNode(ISD::CONCAT_VECTORS, DL, VecVT,
ResLo.getValue(0), ResHi.getValue(0));
SDValue Odd = DAG.getNode(ISD::CONCAT_VECTORS, DL, VecVT, ResLo.getValue(1),
ResHi.getValue(1));
return DAG.getMergeValues({Even, Odd}, DL);
}
// Concatenate the two vectors as one vector to deinterleave
MVT ConcatVT =
MVT::getVectorVT(VecVT.getVectorElementType(),
VecVT.getVectorElementCount().multiplyCoefficientBy(2));
SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatVT,
Op.getOperand(0), Op.getOperand(1));
// We want to operate on all lanes, so get the mask and VL and mask for it
auto [Mask, VL] = getDefaultScalableVLOps(ConcatVT, DL, DAG, Subtarget);
SDValue Passthru = DAG.getUNDEF(ConcatVT);
// We can deinterleave through vnsrl.wi if the element type is smaller than
// ELEN
if (VecVT.getScalarSizeInBits() < Subtarget.getELen()) {
SDValue Even =
getDeinterleaveViaVNSRL(DL, VecVT, Concat, true, Subtarget, DAG);
SDValue Odd =
getDeinterleaveViaVNSRL(DL, VecVT, Concat, false, Subtarget, DAG);
return DAG.getMergeValues({Even, Odd}, DL);
}
// For the indices, use the same SEW to avoid an extra vsetvli
MVT IdxVT = ConcatVT.changeVectorElementTypeToInteger();
// Create a vector of even indices {0, 2, 4, ...}
SDValue EvenIdx =
DAG.getStepVector(DL, IdxVT, APInt(IdxVT.getScalarSizeInBits(), 2));
// Create a vector of odd indices {1, 3, 5, ... }
SDValue OddIdx =
DAG.getNode(ISD::ADD, DL, IdxVT, EvenIdx, DAG.getConstant(1, DL, IdxVT));
// Gather the even and odd elements into two separate vectors
SDValue EvenWide = DAG.getNode(RISCVISD::VRGATHER_VV_VL, DL, ConcatVT,
Concat, EvenIdx, Passthru, Mask, VL);
SDValue OddWide = DAG.getNode(RISCVISD::VRGATHER_VV_VL, DL, ConcatVT,
Concat, OddIdx, Passthru, Mask, VL);
// Extract the result half of the gather for even and odd
SDValue Even = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VecVT, EvenWide,
DAG.getVectorIdxConstant(0, DL));
SDValue Odd = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VecVT, OddWide,
DAG.getVectorIdxConstant(0, DL));
return DAG.getMergeValues({Even, Odd}, DL);
}
SDValue RISCVTargetLowering::lowerVECTOR_INTERLEAVE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VecVT = Op.getSimpleValueType();
assert(VecVT.isScalableVector() &&
"vector_interleave on non-scalable vector!");
// i1 vectors need to be widened to i8
if (VecVT.getVectorElementType() == MVT::i1)
return widenVectorOpsToi8(Op, DL, DAG);
MVT XLenVT = Subtarget.getXLenVT();
SDValue VL = DAG.getRegister(RISCV::X0, XLenVT);
// If the VT is LMUL=8, we need to split and reassemble.
if (VecVT.getSizeInBits().getKnownMinValue() == (8 * RISCV::RVVBitsPerBlock)) {
auto [Op0Lo, Op0Hi] = DAG.SplitVectorOperand(Op.getNode(), 0);
auto [Op1Lo, Op1Hi] = DAG.SplitVectorOperand(Op.getNode(), 1);
EVT SplitVT = Op0Lo.getValueType();
SDValue ResLo = DAG.getNode(ISD::VECTOR_INTERLEAVE, DL,
DAG.getVTList(SplitVT, SplitVT), Op0Lo, Op1Lo);
SDValue ResHi = DAG.getNode(ISD::VECTOR_INTERLEAVE, DL,
DAG.getVTList(SplitVT, SplitVT), Op0Hi, Op1Hi);
SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, DL, VecVT,
ResLo.getValue(0), ResLo.getValue(1));
SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, DL, VecVT,
ResHi.getValue(0), ResHi.getValue(1));
return DAG.getMergeValues({Lo, Hi}, DL);
}
SDValue Interleaved;
// If the element type is smaller than ELEN, then we can interleave with
// vwaddu.vv and vwmaccu.vx
if (VecVT.getScalarSizeInBits() < Subtarget.getELen()) {
Interleaved = getWideningInterleave(Op.getOperand(0), Op.getOperand(1), DL,
DAG, Subtarget);
} else {
// Otherwise, fallback to using vrgathere16.vv
MVT ConcatVT =
MVT::getVectorVT(VecVT.getVectorElementType(),
VecVT.getVectorElementCount().multiplyCoefficientBy(2));
SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatVT,
Op.getOperand(0), Op.getOperand(1));
MVT IdxVT = ConcatVT.changeVectorElementType(MVT::i16);
// 0 1 2 3 4 5 6 7 ...
SDValue StepVec = DAG.getStepVector(DL, IdxVT);
// 1 1 1 1 1 1 1 1 ...
SDValue Ones = DAG.getSplatVector(IdxVT, DL, DAG.getConstant(1, DL, XLenVT));
// 1 0 1 0 1 0 1 0 ...
SDValue OddMask = DAG.getNode(ISD::AND, DL, IdxVT, StepVec, Ones);
OddMask = DAG.getSetCC(
DL, IdxVT.changeVectorElementType(MVT::i1), OddMask,
DAG.getSplatVector(IdxVT, DL, DAG.getConstant(0, DL, XLenVT)),
ISD::CondCode::SETNE);
SDValue VLMax = DAG.getSplatVector(IdxVT, DL, computeVLMax(VecVT, DL, DAG));
// Build up the index vector for interleaving the concatenated vector
// 0 0 1 1 2 2 3 3 ...
SDValue Idx = DAG.getNode(ISD::SRL, DL, IdxVT, StepVec, Ones);
// 0 n 1 n+1 2 n+2 3 n+3 ...
Idx =
DAG.getNode(RISCVISD::ADD_VL, DL, IdxVT, Idx, VLMax, Idx, OddMask, VL);
// Then perform the interleave
// v[0] v[n] v[1] v[n+1] v[2] v[n+2] v[3] v[n+3] ...
SDValue TrueMask = getAllOnesMask(IdxVT, VL, DL, DAG);
Interleaved = DAG.getNode(RISCVISD::VRGATHEREI16_VV_VL, DL, ConcatVT,
Concat, Idx, DAG.getUNDEF(ConcatVT), TrueMask, VL);
}
// Extract the two halves from the interleaved result
SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VecVT, Interleaved,
DAG.getVectorIdxConstant(0, DL));
SDValue Hi = DAG.getNode(
ISD::EXTRACT_SUBVECTOR, DL, VecVT, Interleaved,
DAG.getVectorIdxConstant(VecVT.getVectorMinNumElements(), DL));
return DAG.getMergeValues({Lo, Hi}, DL);
}
// Lower step_vector to the vid instruction. Any non-identity step value must
// be accounted for my manual expansion.
SDValue RISCVTargetLowering::lowerSTEP_VECTOR(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
assert(VT.isScalableVector() && "Expected scalable vector");
MVT XLenVT = Subtarget.getXLenVT();
auto [Mask, VL] = getDefaultScalableVLOps(VT, DL, DAG, Subtarget);
SDValue StepVec = DAG.getNode(RISCVISD::VID_VL, DL, VT, Mask, VL);
uint64_t StepValImm = Op.getConstantOperandVal(0);
if (StepValImm != 1) {
if (isPowerOf2_64(StepValImm)) {
SDValue StepVal =
DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
DAG.getConstant(Log2_64(StepValImm), DL, XLenVT), VL);
StepVec = DAG.getNode(ISD::SHL, DL, VT, StepVec, StepVal);
} else {
SDValue StepVal = lowerScalarSplat(
SDValue(), DAG.getConstant(StepValImm, DL, VT.getVectorElementType()),
VL, VT, DL, DAG, Subtarget);
StepVec = DAG.getNode(ISD::MUL, DL, VT, StepVec, StepVal);
}
}
return StepVec;
}
// Implement vector_reverse using vrgather.vv with indices determined by
// subtracting the id of each element from (VLMAX-1). This will convert
// the indices like so:
// (0, 1,..., VLMAX-2, VLMAX-1) -> (VLMAX-1, VLMAX-2,..., 1, 0).
// TODO: This code assumes VLMAX <= 65536 for LMUL=8 SEW=16.
SDValue RISCVTargetLowering::lowerVECTOR_REVERSE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VecVT = Op.getSimpleValueType();
if (VecVT.getVectorElementType() == MVT::i1) {
MVT WidenVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorElementCount());
SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WidenVT, Op.getOperand(0));
SDValue Op2 = DAG.getNode(ISD::VECTOR_REVERSE, DL, WidenVT, Op1);
return DAG.getNode(ISD::TRUNCATE, DL, VecVT, Op2);
}
unsigned EltSize = VecVT.getScalarSizeInBits();
unsigned MinSize = VecVT.getSizeInBits().getKnownMinValue();
unsigned VectorBitsMax = Subtarget.getRealMaxVLen();
unsigned MaxVLMAX =
RISCVTargetLowering::computeVLMAX(VectorBitsMax, EltSize, MinSize);
unsigned GatherOpc = RISCVISD::VRGATHER_VV_VL;
MVT IntVT = VecVT.changeVectorElementTypeToInteger();
// If this is SEW=8 and VLMAX is potentially more than 256, we need
// to use vrgatherei16.vv.
// TODO: It's also possible to use vrgatherei16.vv for other types to
// decrease register width for the index calculation.
if (MaxVLMAX > 256 && EltSize == 8) {
// If this is LMUL=8, we have to split before can use vrgatherei16.vv.
// Reverse each half, then reassemble them in reverse order.
// NOTE: It's also possible that after splitting that VLMAX no longer
// requires vrgatherei16.vv.
if (MinSize == (8 * RISCV::RVVBitsPerBlock)) {
auto [Lo, Hi] = DAG.SplitVectorOperand(Op.getNode(), 0);
auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VecVT);
Lo = DAG.getNode(ISD::VECTOR_REVERSE, DL, LoVT, Lo);
Hi = DAG.getNode(ISD::VECTOR_REVERSE, DL, HiVT, Hi);
// Reassemble the low and high pieces reversed.
// FIXME: This is a CONCAT_VECTORS.
SDValue Res =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, DAG.getUNDEF(VecVT), Hi,
DAG.getVectorIdxConstant(0, DL));
return DAG.getNode(
ISD::INSERT_SUBVECTOR, DL, VecVT, Res, Lo,
DAG.getVectorIdxConstant(LoVT.getVectorMinNumElements(), DL));
}
// Just promote the int type to i16 which will double the LMUL.
IntVT = MVT::getVectorVT(MVT::i16, VecVT.getVectorElementCount());
GatherOpc = RISCVISD::VRGATHEREI16_VV_VL;
}
MVT XLenVT = Subtarget.getXLenVT();
auto [Mask, VL] = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget);
// Calculate VLMAX-1 for the desired SEW.
SDValue VLMinus1 = DAG.getNode(ISD::SUB, DL, XLenVT,
computeVLMax(VecVT, DL, DAG),
DAG.getConstant(1, DL, XLenVT));
// Splat VLMAX-1 taking care to handle SEW==64 on RV32.
bool IsRV32E64 =
!Subtarget.is64Bit() && IntVT.getVectorElementType() == MVT::i64;
SDValue SplatVL;
if (!IsRV32E64)
SplatVL = DAG.getSplatVector(IntVT, DL, VLMinus1);
else
SplatVL = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntVT, DAG.getUNDEF(IntVT),
VLMinus1, DAG.getRegister(RISCV::X0, XLenVT));
SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, IntVT, Mask, VL);
SDValue Indices = DAG.getNode(RISCVISD::SUB_VL, DL, IntVT, SplatVL, VID,
DAG.getUNDEF(IntVT), Mask, VL);
return DAG.getNode(GatherOpc, DL, VecVT, Op.getOperand(0), Indices,
DAG.getUNDEF(VecVT), Mask, VL);
}
SDValue RISCVTargetLowering::lowerVECTOR_SPLICE(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
MVT XLenVT = Subtarget.getXLenVT();
MVT VecVT = Op.getSimpleValueType();
SDValue VLMax = computeVLMax(VecVT, DL, DAG);
int64_t ImmValue = cast<ConstantSDNode>(Op.getOperand(2))->getSExtValue();
SDValue DownOffset, UpOffset;
if (ImmValue >= 0) {
// The operand is a TargetConstant, we need to rebuild it as a regular
// constant.
DownOffset = DAG.getConstant(ImmValue, DL, XLenVT);
UpOffset = DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, DownOffset);
} else {
// The operand is a TargetConstant, we need to rebuild it as a regular
// constant rather than negating the original operand.
UpOffset = DAG.getConstant(-ImmValue, DL, XLenVT);
DownOffset = DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, UpOffset);
}
SDValue TrueMask = getAllOnesMask(VecVT, VLMax, DL, DAG);
SDValue SlideDown =
getVSlidedown(DAG, Subtarget, DL, VecVT, DAG.getUNDEF(VecVT), V1,
DownOffset, TrueMask, UpOffset);
return getVSlideup(DAG, Subtarget, DL, VecVT, SlideDown, V2, UpOffset,
TrueMask, DAG.getRegister(RISCV::X0, XLenVT),
RISCVII::TAIL_AGNOSTIC);
}
SDValue
RISCVTargetLowering::lowerFixedLengthVectorLoadToRVV(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
auto *Load = cast<LoadSDNode>(Op);
assert(allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
Load->getMemoryVT(),
*Load->getMemOperand()) &&
"Expecting a correctly-aligned load");
MVT VT = Op.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
MVT ContainerVT = getContainerForFixedLengthVector(VT);
// If we know the exact VLEN and our fixed length vector completely fills
// the container, use a whole register load instead.
const auto [MinVLMAX, MaxVLMAX] =
RISCVTargetLowering::computeVLMAXBounds(ContainerVT, Subtarget);
if (MinVLMAX == MaxVLMAX && MinVLMAX == VT.getVectorNumElements() &&
getLMUL1VT(ContainerVT).bitsLE(ContainerVT)) {
MachineMemOperand *MMO = Load->getMemOperand();
SDValue NewLoad =
DAG.getLoad(ContainerVT, DL, Load->getChain(), Load->getBasePtr(),
MMO->getPointerInfo(), MMO->getBaseAlign(), MMO->getFlags(),
MMO->getAAInfo(), MMO->getRanges());
SDValue Result = convertFromScalableVector(VT, NewLoad, DAG, Subtarget);
return DAG.getMergeValues({Result, NewLoad.getValue(1)}, DL);
}
SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG, Subtarget);
bool IsMaskOp = VT.getVectorElementType() == MVT::i1;
SDValue IntID = DAG.getTargetConstant(
IsMaskOp ? Intrinsic::riscv_vlm : Intrinsic::riscv_vle, DL, XLenVT);
SmallVector<SDValue, 4> Ops{Load->getChain(), IntID};
if (!IsMaskOp)
Ops.push_back(DAG.getUNDEF(ContainerVT));
Ops.push_back(Load->getBasePtr());
Ops.push_back(VL);
SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
SDValue NewLoad =
DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
Load->getMemoryVT(), Load->getMemOperand());
SDValue Result = convertFromScalableVector(VT, NewLoad, DAG, Subtarget);
return DAG.getMergeValues({Result, NewLoad.getValue(1)}, DL);
}
SDValue
RISCVTargetLowering::lowerFixedLengthVectorStoreToRVV(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
auto *Store = cast<StoreSDNode>(Op);
assert(allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
Store->getMemoryVT(),
*Store->getMemOperand()) &&
"Expecting a correctly-aligned store");
SDValue StoreVal = Store->getValue();
MVT VT = StoreVal.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
// If the size less than a byte, we need to pad with zeros to make a byte.
if (VT.getVectorElementType() == MVT::i1 && VT.getVectorNumElements() < 8) {
VT = MVT::v8i1;
StoreVal =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getConstant(0, DL, VT),
StoreVal, DAG.getVectorIdxConstant(0, DL));
}
MVT ContainerVT = getContainerForFixedLengthVector(VT);
SDValue NewValue =
convertToScalableVector(ContainerVT, StoreVal, DAG, Subtarget);
// If we know the exact VLEN and our fixed length vector completely fills
// the container, use a whole register store instead.
const auto [MinVLMAX, MaxVLMAX] =
RISCVTargetLowering::computeVLMAXBounds(ContainerVT, Subtarget);
if (MinVLMAX == MaxVLMAX && MinVLMAX == VT.getVectorNumElements() &&
getLMUL1VT(ContainerVT).bitsLE(ContainerVT)) {
MachineMemOperand *MMO = Store->getMemOperand();
return DAG.getStore(Store->getChain(), DL, NewValue, Store->getBasePtr(),
MMO->getPointerInfo(), MMO->getBaseAlign(),
MMO->getFlags(), MMO->getAAInfo());
}
SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG,
Subtarget);
bool IsMaskOp = VT.getVectorElementType() == MVT::i1;
SDValue IntID = DAG.getTargetConstant(
IsMaskOp ? Intrinsic::riscv_vsm : Intrinsic::riscv_vse, DL, XLenVT);
return DAG.getMemIntrinsicNode(
ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other),
{Store->getChain(), IntID, NewValue, Store->getBasePtr(), VL},
Store->getMemoryVT(), Store->getMemOperand());
}
SDValue RISCVTargetLowering::lowerMaskedLoad(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
const auto *MemSD = cast<MemSDNode>(Op);
EVT MemVT = MemSD->getMemoryVT();
MachineMemOperand *MMO = MemSD->getMemOperand();
SDValue Chain = MemSD->getChain();
SDValue BasePtr = MemSD->getBasePtr();
SDValue Mask, PassThru, VL;
if (const auto *VPLoad = dyn_cast<VPLoadSDNode>(Op)) {
Mask = VPLoad->getMask();
PassThru = DAG.getUNDEF(VT);
VL = VPLoad->getVectorLength();
} else {
const auto *MLoad = cast<MaskedLoadSDNode>(Op);
Mask = MLoad->getMask();
PassThru = MLoad->getPassThru();
}
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
MVT XLenVT = Subtarget.getXLenVT();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
PassThru = convertToScalableVector(ContainerVT, PassThru, DAG, Subtarget);
if (!IsUnmasked) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
}
if (!VL)
VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
unsigned IntID =
IsUnmasked ? Intrinsic::riscv_vle : Intrinsic::riscv_vle_mask;
SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
if (IsUnmasked)
Ops.push_back(DAG.getUNDEF(ContainerVT));
else
Ops.push_back(PassThru);
Ops.push_back(BasePtr);
if (!IsUnmasked)
Ops.push_back(Mask);
Ops.push_back(VL);
if (!IsUnmasked)
Ops.push_back(DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT));
SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
SDValue Result =
DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, MemVT, MMO);
Chain = Result.getValue(1);
if (VT.isFixedLengthVector())
Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
return DAG.getMergeValues({Result, Chain}, DL);
}
SDValue RISCVTargetLowering::lowerMaskedStore(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
const auto *MemSD = cast<MemSDNode>(Op);
EVT MemVT = MemSD->getMemoryVT();
MachineMemOperand *MMO = MemSD->getMemOperand();
SDValue Chain = MemSD->getChain();
SDValue BasePtr = MemSD->getBasePtr();
SDValue Val, Mask, VL;
bool IsCompressingStore = false;
if (const auto *VPStore = dyn_cast<VPStoreSDNode>(Op)) {
Val = VPStore->getValue();
Mask = VPStore->getMask();
VL = VPStore->getVectorLength();
} else {
const auto *MStore = cast<MaskedStoreSDNode>(Op);
Val = MStore->getValue();
Mask = MStore->getMask();
IsCompressingStore = MStore->isCompressingStore();
}
bool IsUnmasked =
ISD::isConstantSplatVectorAllOnes(Mask.getNode()) || IsCompressingStore;
MVT VT = Val.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
Val = convertToScalableVector(ContainerVT, Val, DAG, Subtarget);
if (!IsUnmasked || IsCompressingStore) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
}
if (!VL)
VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
if (IsCompressingStore) {
Val = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ContainerVT,
DAG.getConstant(Intrinsic::riscv_vcompress, DL, XLenVT),
DAG.getUNDEF(ContainerVT), Val, Mask, VL);
VL =
DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Mask,
getAllOnesMask(Mask.getSimpleValueType(), VL, DL, DAG), VL);
}
unsigned IntID =
IsUnmasked ? Intrinsic::riscv_vse : Intrinsic::riscv_vse_mask;
SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
Ops.push_back(Val);
Ops.push_back(BasePtr);
if (!IsUnmasked)
Ops.push_back(Mask);
Ops.push_back(VL);
return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL,
DAG.getVTList(MVT::Other), Ops, MemVT, MMO);
}
SDValue
RISCVTargetLowering::lowerFixedLengthVectorSetccToRVV(SDValue Op,
SelectionDAG &DAG) const {
MVT InVT = Op.getOperand(0).getSimpleValueType();
MVT ContainerVT = getContainerForFixedLengthVector(InVT);
MVT VT = Op.getSimpleValueType();
SDValue Op1 =
convertToScalableVector(ContainerVT, Op.getOperand(0), DAG, Subtarget);
SDValue Op2 =
convertToScalableVector(ContainerVT, Op.getOperand(1), DAG, Subtarget);
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(VT.getVectorNumElements(), ContainerVT, DL,
DAG, Subtarget);
MVT MaskVT = getMaskTypeFor(ContainerVT);
SDValue Cmp =
DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT,
{Op1, Op2, Op.getOperand(2), DAG.getUNDEF(MaskVT), Mask, VL});
return convertFromScalableVector(VT, Cmp, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerVectorStrictFSetcc(SDValue Op,
SelectionDAG &DAG) const {
unsigned Opc = Op.getOpcode();
SDLoc DL(Op);
SDValue Chain = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
SDValue Op2 = Op.getOperand(2);
SDValue CC = Op.getOperand(3);
ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
MVT VT = Op.getSimpleValueType();
MVT InVT = Op1.getSimpleValueType();
// RVV VMFEQ/VMFNE ignores qNan, so we expand strict_fsetccs with OEQ/UNE
// condition code.
if (Opc == ISD::STRICT_FSETCCS) {
// Expand strict_fsetccs(x, oeq) to
// (and strict_fsetccs(x, y, oge), strict_fsetccs(x, y, ole))
SDVTList VTList = Op->getVTList();
if (CCVal == ISD::SETEQ || CCVal == ISD::SETOEQ) {
SDValue OLECCVal = DAG.getCondCode(ISD::SETOLE);
SDValue Tmp1 = DAG.getNode(ISD::STRICT_FSETCCS, DL, VTList, Chain, Op1,
Op2, OLECCVal);
SDValue Tmp2 = DAG.getNode(ISD::STRICT_FSETCCS, DL, VTList, Chain, Op2,
Op1, OLECCVal);
SDValue OutChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
Tmp1.getValue(1), Tmp2.getValue(1));
// Tmp1 and Tmp2 might be the same node.
if (Tmp1 != Tmp2)
Tmp1 = DAG.getNode(ISD::AND, DL, VT, Tmp1, Tmp2);
return DAG.getMergeValues({Tmp1, OutChain}, DL);
}
// Expand (strict_fsetccs x, y, une) to (not (strict_fsetccs x, y, oeq))
if (CCVal == ISD::SETNE || CCVal == ISD::SETUNE) {
SDValue OEQCCVal = DAG.getCondCode(ISD::SETOEQ);
SDValue OEQ = DAG.getNode(ISD::STRICT_FSETCCS, DL, VTList, Chain, Op1,
Op2, OEQCCVal);
SDValue Res = DAG.getNOT(DL, OEQ, VT);
return DAG.getMergeValues({Res, OEQ.getValue(1)}, DL);
}
}
MVT ContainerInVT = InVT;
if (InVT.isFixedLengthVector()) {
ContainerInVT = getContainerForFixedLengthVector(InVT);
Op1 = convertToScalableVector(ContainerInVT, Op1, DAG, Subtarget);
Op2 = convertToScalableVector(ContainerInVT, Op2, DAG, Subtarget);
}
MVT MaskVT = getMaskTypeFor(ContainerInVT);
auto [Mask, VL] = getDefaultVLOps(InVT, ContainerInVT, DL, DAG, Subtarget);
SDValue Res;
if (Opc == ISD::STRICT_FSETCC &&
(CCVal == ISD::SETLT || CCVal == ISD::SETOLT || CCVal == ISD::SETLE ||
CCVal == ISD::SETOLE)) {
// VMFLT/VMFLE/VMFGT/VMFGE raise exception for qNan. Generate a mask to only
// active when both input elements are ordered.
SDValue True = getAllOnesMask(ContainerInVT, VL, DL, DAG);
SDValue OrderMask1 = DAG.getNode(
RISCVISD::STRICT_FSETCC_VL, DL, DAG.getVTList(MaskVT, MVT::Other),
{Chain, Op1, Op1, DAG.getCondCode(ISD::SETOEQ), DAG.getUNDEF(MaskVT),
True, VL});
SDValue OrderMask2 = DAG.getNode(
RISCVISD::STRICT_FSETCC_VL, DL, DAG.getVTList(MaskVT, MVT::Other),
{Chain, Op2, Op2, DAG.getCondCode(ISD::SETOEQ), DAG.getUNDEF(MaskVT),
True, VL});
Mask =
DAG.getNode(RISCVISD::VMAND_VL, DL, MaskVT, OrderMask1, OrderMask2, VL);
// Use Mask as the merge operand to let the result be 0 if either of the
// inputs is unordered.
Res = DAG.getNode(RISCVISD::STRICT_FSETCCS_VL, DL,
DAG.getVTList(MaskVT, MVT::Other),
{Chain, Op1, Op2, CC, Mask, Mask, VL});
} else {
unsigned RVVOpc = Opc == ISD::STRICT_FSETCC ? RISCVISD::STRICT_FSETCC_VL
: RISCVISD::STRICT_FSETCCS_VL;
Res = DAG.getNode(RVVOpc, DL, DAG.getVTList(MaskVT, MVT::Other),
{Chain, Op1, Op2, CC, DAG.getUNDEF(MaskVT), Mask, VL});
}
if (VT.isFixedLengthVector()) {
SDValue SubVec = convertFromScalableVector(VT, Res, DAG, Subtarget);
return DAG.getMergeValues({SubVec, Res.getValue(1)}, DL);
}
return Res;
}
// Lower vector ABS to smax(X, sub(0, X)).
SDValue RISCVTargetLowering::lowerABS(SDValue Op, SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue X = Op.getOperand(0);
assert((Op.getOpcode() == ISD::VP_ABS || VT.isFixedLengthVector()) &&
"Unexpected type for ISD::ABS");
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
X = convertToScalableVector(ContainerVT, X, DAG, Subtarget);
}
SDValue Mask, VL;
if (Op->getOpcode() == ISD::VP_ABS) {
Mask = Op->getOperand(1);
if (VT.isFixedLengthVector())
Mask = convertToScalableVector(getMaskTypeFor(ContainerVT), Mask, DAG,
Subtarget);
VL = Op->getOperand(2);
} else
std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
SDValue SplatZero = DAG.getNode(
RISCVISD::VMV_V_X_VL, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
DAG.getConstant(0, DL, Subtarget.getXLenVT()), VL);
SDValue NegX = DAG.getNode(RISCVISD::SUB_VL, DL, ContainerVT, SplatZero, X,
DAG.getUNDEF(ContainerVT), Mask, VL);
SDValue Max = DAG.getNode(RISCVISD::SMAX_VL, DL, ContainerVT, X, NegX,
DAG.getUNDEF(ContainerVT), Mask, VL);
if (VT.isFixedLengthVector())
Max = convertFromScalableVector(VT, Max, DAG, Subtarget);
return Max;
}
SDValue RISCVTargetLowering::lowerFixedLengthVectorFCOPYSIGNToRVV(
SDValue Op, SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue Mag = Op.getOperand(0);
SDValue Sign = Op.getOperand(1);
assert(Mag.getValueType() == Sign.getValueType() &&
"Can only handle COPYSIGN with matching types.");
MVT ContainerVT = getContainerForFixedLengthVector(VT);
Mag = convertToScalableVector(ContainerVT, Mag, DAG, Subtarget);
Sign = convertToScalableVector(ContainerVT, Sign, DAG, Subtarget);
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
SDValue CopySign = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Mag,
Sign, DAG.getUNDEF(ContainerVT), Mask, VL);
return convertFromScalableVector(VT, CopySign, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerFixedLengthVectorSelectToRVV(
SDValue Op, SelectionDAG &DAG) const {
MVT VT = Op.getSimpleValueType();
MVT ContainerVT = getContainerForFixedLengthVector(VT);
MVT I1ContainerVT =
MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
SDValue CC =
convertToScalableVector(I1ContainerVT, Op.getOperand(0), DAG, Subtarget);
SDValue Op1 =
convertToScalableVector(ContainerVT, Op.getOperand(1), DAG, Subtarget);
SDValue Op2 =
convertToScalableVector(ContainerVT, Op.getOperand(2), DAG, Subtarget);
SDLoc DL(Op);
SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
SDValue Select = DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, CC, Op1,
Op2, DAG.getUNDEF(ContainerVT), VL);
return convertFromScalableVector(VT, Select, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerToScalableOp(SDValue Op,
SelectionDAG &DAG) const {
unsigned NewOpc = getRISCVVLOp(Op);
bool HasMergeOp = hasMergeOp(NewOpc);
bool HasMask = hasMaskOp(NewOpc);
MVT VT = Op.getSimpleValueType();
MVT ContainerVT = getContainerForFixedLengthVector(VT);
// Create list of operands by converting existing ones to scalable types.
SmallVector<SDValue, 6> Ops;
for (const SDValue &V : Op->op_values()) {
assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!");
// Pass through non-vector operands.
if (!V.getValueType().isVector()) {
Ops.push_back(V);
continue;
}
// "cast" fixed length vector to a scalable vector.
assert(useRVVForFixedLengthVectorVT(V.getSimpleValueType()) &&
"Only fixed length vectors are supported!");
Ops.push_back(convertToScalableVector(ContainerVT, V, DAG, Subtarget));
}
SDLoc DL(Op);
auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
if (HasMergeOp)
Ops.push_back(DAG.getUNDEF(ContainerVT));
if (HasMask)
Ops.push_back(Mask);
Ops.push_back(VL);
// StrictFP operations have two result values. Their lowered result should
// have same result count.
if (Op->isStrictFPOpcode()) {
SDValue ScalableRes =
DAG.getNode(NewOpc, DL, DAG.getVTList(ContainerVT, MVT::Other), Ops,
Op->getFlags());
SDValue SubVec = convertFromScalableVector(VT, ScalableRes, DAG, Subtarget);
return DAG.getMergeValues({SubVec, ScalableRes.getValue(1)}, DL);
}
SDValue ScalableRes =
DAG.getNode(NewOpc, DL, ContainerVT, Ops, Op->getFlags());
return convertFromScalableVector(VT, ScalableRes, DAG, Subtarget);
}
// Lower a VP_* ISD node to the corresponding RISCVISD::*_VL node:
// * Operands of each node are assumed to be in the same order.
// * The EVL operand is promoted from i32 to i64 on RV64.
// * Fixed-length vectors are converted to their scalable-vector container
// types.
SDValue RISCVTargetLowering::lowerVPOp(SDValue Op, SelectionDAG &DAG) const {
unsigned RISCVISDOpc = getRISCVVLOp(Op);
bool HasMergeOp = hasMergeOp(RISCVISDOpc);
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SmallVector<SDValue, 4> Ops;
MVT ContainerVT = VT;
if (VT.isFixedLengthVector())
ContainerVT = getContainerForFixedLengthVector(VT);
for (const auto &OpIdx : enumerate(Op->ops())) {
SDValue V = OpIdx.value();
assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!");
// Add dummy merge value before the mask. Or if there isn't a mask, before
// EVL.
if (HasMergeOp) {
auto MaskIdx = ISD::getVPMaskIdx(Op.getOpcode());
if (MaskIdx) {
if (*MaskIdx == OpIdx.index())
Ops.push_back(DAG.getUNDEF(ContainerVT));
} else if (ISD::getVPExplicitVectorLengthIdx(Op.getOpcode()) ==
OpIdx.index()) {
if (Op.getOpcode() == ISD::VP_MERGE) {
// For VP_MERGE, copy the false operand instead of an undef value.
Ops.push_back(Ops.back());
} else {
assert(Op.getOpcode() == ISD::VP_SELECT);
// For VP_SELECT, add an undef value.
Ops.push_back(DAG.getUNDEF(ContainerVT));
}
}
}
// Pass through operands which aren't fixed-length vectors.
if (!V.getValueType().isFixedLengthVector()) {
Ops.push_back(V);
continue;
}
// "cast" fixed length vector to a scalable vector.
MVT OpVT = V.getSimpleValueType();
MVT ContainerVT = getContainerForFixedLengthVector(OpVT);
assert(useRVVForFixedLengthVectorVT(OpVT) &&
"Only fixed length vectors are supported!");
Ops.push_back(convertToScalableVector(ContainerVT, V, DAG, Subtarget));
}
if (!VT.isFixedLengthVector())
return DAG.getNode(RISCVISDOpc, DL, VT, Ops, Op->getFlags());
SDValue VPOp = DAG.getNode(RISCVISDOpc, DL, ContainerVT, Ops, Op->getFlags());
return convertFromScalableVector(VT, VPOp, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerVPExtMaskOp(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue Src = Op.getOperand(0);
// NOTE: Mask is dropped.
SDValue VL = Op.getOperand(2);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
MVT SrcVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
Src = convertToScalableVector(SrcVT, Src, DAG, Subtarget);
}
MVT XLenVT = Subtarget.getXLenVT();
SDValue Zero = DAG.getConstant(0, DL, XLenVT);
SDValue ZeroSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), Zero, VL);
SDValue SplatValue = DAG.getConstant(
Op.getOpcode() == ISD::VP_ZERO_EXTEND ? 1 : -1, DL, XLenVT);
SDValue Splat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), SplatValue, VL);
SDValue Result = DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, Src, Splat,
ZeroSplat, DAG.getUNDEF(ContainerVT), VL);
if (!VT.isFixedLengthVector())
return Result;
return convertFromScalableVector(VT, Result, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerVPSetCCMaskOp(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
SDValue Op1 = Op.getOperand(0);
SDValue Op2 = Op.getOperand(1);
ISD::CondCode Condition = cast<CondCodeSDNode>(Op.getOperand(2))->get();
// NOTE: Mask is dropped.
SDValue VL = Op.getOperand(4);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
Op2 = convertToScalableVector(ContainerVT, Op2, DAG, Subtarget);
}
SDValue Result;
SDValue AllOneMask = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
switch (Condition) {
default:
break;
// X != Y --> (X^Y)
case ISD::SETNE:
Result = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, Op2, VL);
break;
// X == Y --> ~(X^Y)
case ISD::SETEQ: {
SDValue Temp =
DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, Op2, VL);
Result =
DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Temp, AllOneMask, VL);
break;
}
// X >s Y --> X == 0 & Y == 1 --> ~X & Y
// X <u Y --> X == 0 & Y == 1 --> ~X & Y
case ISD::SETGT:
case ISD::SETULT: {
SDValue Temp =
DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, AllOneMask, VL);
Result = DAG.getNode(RISCVISD::VMAND_VL, DL, ContainerVT, Temp, Op2, VL);
break;
}
// X <s Y --> X == 1 & Y == 0 --> ~Y & X
// X >u Y --> X == 1 & Y == 0 --> ~Y & X
case ISD::SETLT:
case ISD::SETUGT: {
SDValue Temp =
DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op2, AllOneMask, VL);
Result = DAG.getNode(RISCVISD::VMAND_VL, DL, ContainerVT, Op1, Temp, VL);
break;
}
// X >=s Y --> X == 0 | Y == 1 --> ~X | Y
// X <=u Y --> X == 0 | Y == 1 --> ~X | Y
case ISD::SETGE:
case ISD::SETULE: {
SDValue Temp =
DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, AllOneMask, VL);
Result = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Temp, Op2, VL);
break;
}
// X <=s Y --> X == 1 | Y == 0 --> ~Y | X
// X >=u Y --> X == 1 | Y == 0 --> ~Y | X
case ISD::SETLE:
case ISD::SETUGE: {
SDValue Temp =
DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op2, AllOneMask, VL);
Result = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Temp, Op1, VL);
break;
}
}
if (!VT.isFixedLengthVector())
return Result;
return convertFromScalableVector(VT, Result, DAG, Subtarget);
}
// Lower Floating-Point/Integer Type-Convert VP SDNodes
SDValue RISCVTargetLowering::lowerVPFPIntConvOp(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Src = Op.getOperand(0);
SDValue Mask = Op.getOperand(1);
SDValue VL = Op.getOperand(2);
unsigned RISCVISDOpc = getRISCVVLOp(Op);
MVT DstVT = Op.getSimpleValueType();
MVT SrcVT = Src.getSimpleValueType();
if (DstVT.isFixedLengthVector()) {
DstVT = getContainerForFixedLengthVector(DstVT);
SrcVT = getContainerForFixedLengthVector(SrcVT);
Src = convertToScalableVector(SrcVT, Src, DAG, Subtarget);
MVT MaskVT = getMaskTypeFor(DstVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
unsigned DstEltSize = DstVT.getScalarSizeInBits();
unsigned SrcEltSize = SrcVT.getScalarSizeInBits();
SDValue Result;
if (DstEltSize >= SrcEltSize) { // Single-width and widening conversion.
if (SrcVT.isInteger()) {
assert(DstVT.isFloatingPoint() && "Wrong input/output vector types");
unsigned RISCVISDExtOpc = RISCVISDOpc == RISCVISD::SINT_TO_FP_VL
? RISCVISD::VSEXT_VL
: RISCVISD::VZEXT_VL;
// Do we need to do any pre-widening before converting?
if (SrcEltSize == 1) {
MVT IntVT = DstVT.changeVectorElementTypeToInteger();
MVT XLenVT = Subtarget.getXLenVT();
SDValue Zero = DAG.getConstant(0, DL, XLenVT);
SDValue ZeroSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntVT,
DAG.getUNDEF(IntVT), Zero, VL);
SDValue One = DAG.getConstant(
RISCVISDExtOpc == RISCVISD::VZEXT_VL ? 1 : -1, DL, XLenVT);
SDValue OneSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntVT,
DAG.getUNDEF(IntVT), One, VL);
Src = DAG.getNode(RISCVISD::VMERGE_VL, DL, IntVT, Src, OneSplat,
ZeroSplat, DAG.getUNDEF(IntVT), VL);
} else if (DstEltSize > (2 * SrcEltSize)) {
// Widen before converting.
MVT IntVT = MVT::getVectorVT(MVT::getIntegerVT(DstEltSize / 2),
DstVT.getVectorElementCount());
Src = DAG.getNode(RISCVISDExtOpc, DL, IntVT, Src, Mask, VL);
}
Result = DAG.getNode(RISCVISDOpc, DL, DstVT, Src, Mask, VL);
} else {
assert(SrcVT.isFloatingPoint() && DstVT.isInteger() &&
"Wrong input/output vector types");
// Convert f16 to f32 then convert f32 to i64.
if (DstEltSize > (2 * SrcEltSize)) {
assert(SrcVT.getVectorElementType() == MVT::f16 && "Unexpected type!");
MVT InterimFVT =
MVT::getVectorVT(MVT::f32, DstVT.getVectorElementCount());
Src =
DAG.getNode(RISCVISD::FP_EXTEND_VL, DL, InterimFVT, Src, Mask, VL);
}
Result = DAG.getNode(RISCVISDOpc, DL, DstVT, Src, Mask, VL);
}
} else { // Narrowing + Conversion
if (SrcVT.isInteger()) {
assert(DstVT.isFloatingPoint() && "Wrong input/output vector types");
// First do a narrowing convert to an FP type half the size, then round
// the FP type to a small FP type if needed.
MVT InterimFVT = DstVT;
if (SrcEltSize > (2 * DstEltSize)) {
assert(SrcEltSize == (4 * DstEltSize) && "Unexpected types!");
assert(DstVT.getVectorElementType() == MVT::f16 && "Unexpected type!");
InterimFVT = MVT::getVectorVT(MVT::f32, DstVT.getVectorElementCount());
}
Result = DAG.getNode(RISCVISDOpc, DL, InterimFVT, Src, Mask, VL);
if (InterimFVT != DstVT) {
Src = Result;
Result = DAG.getNode(RISCVISD::FP_ROUND_VL, DL, DstVT, Src, Mask, VL);
}
} else {
assert(SrcVT.isFloatingPoint() && DstVT.isInteger() &&
"Wrong input/output vector types");
// First do a narrowing conversion to an integer half the size, then
// truncate if needed.
if (DstEltSize == 1) {
// First convert to the same size integer, then convert to mask using
// setcc.
assert(SrcEltSize >= 16 && "Unexpected FP type!");
MVT InterimIVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize),
DstVT.getVectorElementCount());
Result = DAG.getNode(RISCVISDOpc, DL, InterimIVT, Src, Mask, VL);
// Compare the integer result to 0. The integer should be 0 or 1/-1,
// otherwise the conversion was undefined.
MVT XLenVT = Subtarget.getXLenVT();
SDValue SplatZero = DAG.getConstant(0, DL, XLenVT);
SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, InterimIVT,
DAG.getUNDEF(InterimIVT), SplatZero, VL);
Result = DAG.getNode(RISCVISD::SETCC_VL, DL, DstVT,
{Result, SplatZero, DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(DstVT), Mask, VL});
} else {
MVT InterimIVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize / 2),
DstVT.getVectorElementCount());
Result = DAG.getNode(RISCVISDOpc, DL, InterimIVT, Src, Mask, VL);
while (InterimIVT != DstVT) {
SrcEltSize /= 2;
Src = Result;
InterimIVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize / 2),
DstVT.getVectorElementCount());
Result = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, InterimIVT,
Src, Mask, VL);
}
}
}
}
MVT VT = Op.getSimpleValueType();
if (!VT.isFixedLengthVector())
return Result;
return convertFromScalableVector(VT, Result, DAG, Subtarget);
}
SDValue
RISCVTargetLowering::lowerVPSpliceExperimental(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Op1 = Op.getOperand(0);
SDValue Op2 = Op.getOperand(1);
SDValue Offset = Op.getOperand(2);
SDValue Mask = Op.getOperand(3);
SDValue EVL1 = Op.getOperand(4);
SDValue EVL2 = Op.getOperand(5);
const MVT XLenVT = Subtarget.getXLenVT();
MVT VT = Op.getSimpleValueType();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
Op2 = convertToScalableVector(ContainerVT, Op2, DAG, Subtarget);
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
// EVL1 may need to be extended to XLenVT with RV64LegalI32.
EVL1 = DAG.getNode(ISD::ZERO_EXTEND, DL, XLenVT, EVL1);
bool IsMaskVector = VT.getVectorElementType() == MVT::i1;
if (IsMaskVector) {
ContainerVT = ContainerVT.changeVectorElementType(MVT::i8);
// Expand input operands
SDValue SplatOneOp1 = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT),
DAG.getConstant(1, DL, XLenVT), EVL1);
SDValue SplatZeroOp1 = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT),
DAG.getConstant(0, DL, XLenVT), EVL1);
Op1 = DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, Op1, SplatOneOp1,
SplatZeroOp1, DAG.getUNDEF(ContainerVT), EVL1);
SDValue SplatOneOp2 = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT),
DAG.getConstant(1, DL, XLenVT), EVL2);
SDValue SplatZeroOp2 = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT),
DAG.getConstant(0, DL, XLenVT), EVL2);
Op2 = DAG.getNode(RISCVISD::VMERGE_VL, DL, ContainerVT, Op2, SplatOneOp2,
SplatZeroOp2, DAG.getUNDEF(ContainerVT), EVL2);
}
int64_t ImmValue = cast<ConstantSDNode>(Offset)->getSExtValue();
SDValue DownOffset, UpOffset;
if (ImmValue >= 0) {
// The operand is a TargetConstant, we need to rebuild it as a regular
// constant.
DownOffset = DAG.getConstant(ImmValue, DL, XLenVT);
UpOffset = DAG.getNode(ISD::SUB, DL, XLenVT, EVL1, DownOffset);
} else {
// The operand is a TargetConstant, we need to rebuild it as a regular
// constant rather than negating the original operand.
UpOffset = DAG.getConstant(-ImmValue, DL, XLenVT);
DownOffset = DAG.getNode(ISD::SUB, DL, XLenVT, EVL1, UpOffset);
}
SDValue SlideDown =
getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
Op1, DownOffset, Mask, UpOffset);
SDValue Result = getVSlideup(DAG, Subtarget, DL, ContainerVT, SlideDown, Op2,
UpOffset, Mask, EVL2, RISCVII::TAIL_AGNOSTIC);
if (IsMaskVector) {
// Truncate Result back to a mask vector (Result has same EVL as Op2)
Result = DAG.getNode(
RISCVISD::SETCC_VL, DL, ContainerVT.changeVectorElementType(MVT::i1),
{Result, DAG.getConstant(0, DL, ContainerVT),
DAG.getCondCode(ISD::SETNE), DAG.getUNDEF(getMaskTypeFor(ContainerVT)),
Mask, EVL2});
}
if (!VT.isFixedLengthVector())
return Result;
return convertFromScalableVector(VT, Result, DAG, Subtarget);
}
SDValue
RISCVTargetLowering::lowerVPReverseExperimental(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
SDValue Op1 = Op.getOperand(0);
SDValue Mask = Op.getOperand(1);
SDValue EVL = Op.getOperand(2);
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
MVT GatherVT = ContainerVT;
MVT IndicesVT = ContainerVT.changeVectorElementTypeToInteger();
// Check if we are working with mask vectors
bool IsMaskVector = ContainerVT.getVectorElementType() == MVT::i1;
if (IsMaskVector) {
GatherVT = IndicesVT = ContainerVT.changeVectorElementType(MVT::i8);
// Expand input operand
SDValue SplatOne = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IndicesVT,
DAG.getUNDEF(IndicesVT),
DAG.getConstant(1, DL, XLenVT), EVL);
SDValue SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IndicesVT,
DAG.getUNDEF(IndicesVT),
DAG.getConstant(0, DL, XLenVT), EVL);
Op1 = DAG.getNode(RISCVISD::VMERGE_VL, DL, IndicesVT, Op1, SplatOne,
SplatZero, DAG.getUNDEF(IndicesVT), EVL);
}
unsigned EltSize = GatherVT.getScalarSizeInBits();
unsigned MinSize = GatherVT.getSizeInBits().getKnownMinValue();
unsigned VectorBitsMax = Subtarget.getRealMaxVLen();
unsigned MaxVLMAX =
RISCVTargetLowering::computeVLMAX(VectorBitsMax, EltSize, MinSize);
unsigned GatherOpc = RISCVISD::VRGATHER_VV_VL;
// If this is SEW=8 and VLMAX is unknown or more than 256, we need
// to use vrgatherei16.vv.
// TODO: It's also possible to use vrgatherei16.vv for other types to
// decrease register width for the index calculation.
// NOTE: This code assumes VLMAX <= 65536 for LMUL=8 SEW=16.
if (MaxVLMAX > 256 && EltSize == 8) {
// If this is LMUL=8, we have to split before using vrgatherei16.vv.
// Split the vector in half and reverse each half using a full register
// reverse.
// Swap the halves and concatenate them.
// Slide the concatenated result by (VLMax - VL).
if (MinSize == (8 * RISCV::RVVBitsPerBlock)) {
auto [LoVT, HiVT] = DAG.GetSplitDestVTs(GatherVT);
auto [Lo, Hi] = DAG.SplitVector(Op1, DL);
SDValue LoRev = DAG.getNode(ISD::VECTOR_REVERSE, DL, LoVT, Lo);
SDValue HiRev = DAG.getNode(ISD::VECTOR_REVERSE, DL, HiVT, Hi);
// Reassemble the low and high pieces reversed.
// NOTE: this Result is unmasked (because we do not need masks for
// shuffles). If in the future this has to change, we can use a SELECT_VL
// between Result and UNDEF using the mask originally passed to VP_REVERSE
SDValue Result =
DAG.getNode(ISD::CONCAT_VECTORS, DL, GatherVT, HiRev, LoRev);
// Slide off any elements from past EVL that were reversed into the low
// elements.
unsigned MinElts = GatherVT.getVectorMinNumElements();
SDValue VLMax =
DAG.getVScale(DL, XLenVT, APInt(XLenVT.getSizeInBits(), MinElts));
SDValue Diff = DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, EVL);
Result = getVSlidedown(DAG, Subtarget, DL, GatherVT,
DAG.getUNDEF(GatherVT), Result, Diff, Mask, EVL);
if (IsMaskVector) {
// Truncate Result back to a mask vector
Result =
DAG.getNode(RISCVISD::SETCC_VL, DL, ContainerVT,
{Result, DAG.getConstant(0, DL, GatherVT),
DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(getMaskTypeFor(ContainerVT)), Mask, EVL});
}
if (!VT.isFixedLengthVector())
return Result;
return convertFromScalableVector(VT, Result, DAG, Subtarget);
}
// Just promote the int type to i16 which will double the LMUL.
IndicesVT = MVT::getVectorVT(MVT::i16, IndicesVT.getVectorElementCount());
GatherOpc = RISCVISD::VRGATHEREI16_VV_VL;
}
SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, IndicesVT, Mask, EVL);
SDValue VecLen =
DAG.getNode(ISD::SUB, DL, XLenVT, EVL, DAG.getConstant(1, DL, XLenVT));
SDValue VecLenSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IndicesVT,
DAG.getUNDEF(IndicesVT), VecLen, EVL);
SDValue VRSUB = DAG.getNode(RISCVISD::SUB_VL, DL, IndicesVT, VecLenSplat, VID,
DAG.getUNDEF(IndicesVT), Mask, EVL);
SDValue Result = DAG.getNode(GatherOpc, DL, GatherVT, Op1, VRSUB,
DAG.getUNDEF(GatherVT), Mask, EVL);
if (IsMaskVector) {
// Truncate Result back to a mask vector
Result = DAG.getNode(
RISCVISD::SETCC_VL, DL, ContainerVT,
{Result, DAG.getConstant(0, DL, GatherVT), DAG.getCondCode(ISD::SETNE),
DAG.getUNDEF(getMaskTypeFor(ContainerVT)), Mask, EVL});
}
if (!VT.isFixedLengthVector())
return Result;
return convertFromScalableVector(VT, Result, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerLogicVPOp(SDValue Op,
SelectionDAG &DAG) const {
MVT VT = Op.getSimpleValueType();
if (VT.getVectorElementType() != MVT::i1)
return lowerVPOp(Op, DAG);
// It is safe to drop mask parameter as masked-off elements are undef.
SDValue Op1 = Op->getOperand(0);
SDValue Op2 = Op->getOperand(1);
SDValue VL = Op->getOperand(3);
MVT ContainerVT = VT;
const bool IsFixed = VT.isFixedLengthVector();
if (IsFixed) {
ContainerVT = getContainerForFixedLengthVector(VT);
Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
Op2 = convertToScalableVector(ContainerVT, Op2, DAG, Subtarget);
}
SDLoc DL(Op);
SDValue Val = DAG.getNode(getRISCVVLOp(Op), DL, ContainerVT, Op1, Op2, VL);
if (!IsFixed)
return Val;
return convertFromScalableVector(VT, Val, DAG, Subtarget);
}
SDValue RISCVTargetLowering::lowerVPStridedLoad(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
MVT VT = Op.getSimpleValueType();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector())
ContainerVT = getContainerForFixedLengthVector(VT);
SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
auto *VPNode = cast<VPStridedLoadSDNode>(Op);
// Check if the mask is known to be all ones
SDValue Mask = VPNode->getMask();
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
SDValue IntID = DAG.getTargetConstant(IsUnmasked ? Intrinsic::riscv_vlse
: Intrinsic::riscv_vlse_mask,
DL, XLenVT);
SmallVector<SDValue, 8> Ops{VPNode->getChain(), IntID,
DAG.getUNDEF(ContainerVT), VPNode->getBasePtr(),
VPNode->getStride()};
if (!IsUnmasked) {
if (VT.isFixedLengthVector()) {
MVT MaskVT = ContainerVT.changeVectorElementType(MVT::i1);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
Ops.push_back(Mask);
}
Ops.push_back(VPNode->getVectorLength());
if (!IsUnmasked) {
SDValue Policy = DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT);
Ops.push_back(Policy);
}
SDValue Result =
DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
VPNode->getMemoryVT(), VPNode->getMemOperand());
SDValue Chain = Result.getValue(1);
if (VT.isFixedLengthVector())
Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
return DAG.getMergeValues({Result, Chain}, DL);
}
SDValue RISCVTargetLowering::lowerVPStridedStore(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT XLenVT = Subtarget.getXLenVT();
auto *VPNode = cast<VPStridedStoreSDNode>(Op);
SDValue StoreVal = VPNode->getValue();
MVT VT = StoreVal.getSimpleValueType();
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
StoreVal = convertToScalableVector(ContainerVT, StoreVal, DAG, Subtarget);
}
// Check if the mask is known to be all ones
SDValue Mask = VPNode->getMask();
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
SDValue IntID = DAG.getTargetConstant(IsUnmasked ? Intrinsic::riscv_vsse
: Intrinsic::riscv_vsse_mask,
DL, XLenVT);
SmallVector<SDValue, 8> Ops{VPNode->getChain(), IntID, StoreVal,
VPNode->getBasePtr(), VPNode->getStride()};
if (!IsUnmasked) {
if (VT.isFixedLengthVector()) {
MVT MaskVT = ContainerVT.changeVectorElementType(MVT::i1);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
Ops.push_back(Mask);
}
Ops.push_back(VPNode->getVectorLength());
return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL, VPNode->getVTList(),
Ops, VPNode->getMemoryVT(),
VPNode->getMemOperand());
}
// Custom lower MGATHER/VP_GATHER to a legalized form for RVV. It will then be
// matched to a RVV indexed load. The RVV indexed load instructions only
// support the "unsigned unscaled" addressing mode; indices are implicitly
// zero-extended or truncated to XLEN and are treated as byte offsets. Any
// signed or scaled indexing is extended to the XLEN value type and scaled
// accordingly.
SDValue RISCVTargetLowering::lowerMaskedGather(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
MVT VT = Op.getSimpleValueType();
const auto *MemSD = cast<MemSDNode>(Op.getNode());
EVT MemVT = MemSD->getMemoryVT();
MachineMemOperand *MMO = MemSD->getMemOperand();
SDValue Chain = MemSD->getChain();
SDValue BasePtr = MemSD->getBasePtr();
[[maybe_unused]] ISD::LoadExtType LoadExtType;
SDValue Index, Mask, PassThru, VL;
if (auto *VPGN = dyn_cast<VPGatherSDNode>(Op.getNode())) {
Index = VPGN->getIndex();
Mask = VPGN->getMask();
PassThru = DAG.getUNDEF(VT);
VL = VPGN->getVectorLength();
// VP doesn't support extending loads.
LoadExtType = ISD::NON_EXTLOAD;
} else {
// Else it must be a MGATHER.
auto *MGN = cast<MaskedGatherSDNode>(Op.getNode());
Index = MGN->getIndex();
Mask = MGN->getMask();
PassThru = MGN->getPassThru();
LoadExtType = MGN->getExtensionType();
}
MVT IndexVT = Index.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
"Unexpected VTs!");
assert(BasePtr.getSimpleValueType() == XLenVT && "Unexpected pointer type");
// Targets have to explicitly opt-in for extending vector loads.
assert(LoadExtType == ISD::NON_EXTLOAD &&
"Unexpected extending MGATHER/VP_GATHER");
// If the mask is known to be all ones, optimize to an unmasked intrinsic;
// the selection of the masked intrinsics doesn't do this for us.
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
IndexVT = MVT::getVectorVT(IndexVT.getVectorElementType(),
ContainerVT.getVectorElementCount());
Index = convertToScalableVector(IndexVT, Index, DAG, Subtarget);
if (!IsUnmasked) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
PassThru = convertToScalableVector(ContainerVT, PassThru, DAG, Subtarget);
}
}
if (!VL)
VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
if (XLenVT == MVT::i32 && IndexVT.getVectorElementType().bitsGT(XLenVT)) {
IndexVT = IndexVT.changeVectorElementType(XLenVT);
Index = DAG.getNode(ISD::TRUNCATE, DL, IndexVT, Index);
}
unsigned IntID =
IsUnmasked ? Intrinsic::riscv_vluxei : Intrinsic::riscv_vluxei_mask;
SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
if (IsUnmasked)
Ops.push_back(DAG.getUNDEF(ContainerVT));
else
Ops.push_back(PassThru);
Ops.push_back(BasePtr);
Ops.push_back(Index);
if (!IsUnmasked)
Ops.push_back(Mask);
Ops.push_back(VL);
if (!IsUnmasked)
Ops.push_back(DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT));
SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
SDValue Result =
DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, MemVT, MMO);
Chain = Result.getValue(1);
if (VT.isFixedLengthVector())
Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
return DAG.getMergeValues({Result, Chain}, DL);
}
// Custom lower MSCATTER/VP_SCATTER to a legalized form for RVV. It will then be
// matched to a RVV indexed store. The RVV indexed store instructions only
// support the "unsigned unscaled" addressing mode; indices are implicitly
// zero-extended or truncated to XLEN and are treated as byte offsets. Any
// signed or scaled indexing is extended to the XLEN value type and scaled
// accordingly.
SDValue RISCVTargetLowering::lowerMaskedScatter(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
const auto *MemSD = cast<MemSDNode>(Op.getNode());
EVT MemVT = MemSD->getMemoryVT();
MachineMemOperand *MMO = MemSD->getMemOperand();
SDValue Chain = MemSD->getChain();
SDValue BasePtr = MemSD->getBasePtr();
[[maybe_unused]] bool IsTruncatingStore = false;
SDValue Index, Mask, Val, VL;
if (auto *VPSN = dyn_cast<VPScatterSDNode>(Op.getNode())) {
Index = VPSN->getIndex();
Mask = VPSN->getMask();
Val = VPSN->getValue();
VL = VPSN->getVectorLength();
// VP doesn't support truncating stores.
IsTruncatingStore = false;
} else {
// Else it must be a MSCATTER.
auto *MSN = cast<MaskedScatterSDNode>(Op.getNode());
Index = MSN->getIndex();
Mask = MSN->getMask();
Val = MSN->getValue();
IsTruncatingStore = MSN->isTruncatingStore();
}
MVT VT = Val.getSimpleValueType();
MVT IndexVT = Index.getSimpleValueType();
MVT XLenVT = Subtarget.getXLenVT();
assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
"Unexpected VTs!");
assert(BasePtr.getSimpleValueType() == XLenVT && "Unexpected pointer type");
// Targets have to explicitly opt-in for extending vector loads and
// truncating vector stores.
assert(!IsTruncatingStore && "Unexpected truncating MSCATTER/VP_SCATTER");
// If the mask is known to be all ones, optimize to an unmasked intrinsic;
// the selection of the masked intrinsics doesn't do this for us.
bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
MVT ContainerVT = VT;
if (VT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VT);
IndexVT = MVT::getVectorVT(IndexVT.getVectorElementType(),
ContainerVT.getVectorElementCount());
Index = convertToScalableVector(IndexVT, Index, DAG, Subtarget);
Val = convertToScalableVector(ContainerVT, Val, DAG, Subtarget);
if (!IsUnmasked) {
MVT MaskVT = getMaskTypeFor(ContainerVT);
Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
}
}
if (!VL)
VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
if (XLenVT == MVT::i32 && IndexVT.getVectorElementType().bitsGT(XLenVT)) {
IndexVT = IndexVT.changeVectorElementType(XLenVT);
Index = DAG.getNode(ISD::TRUNCATE, DL, IndexVT, Index);
}
unsigned IntID =
IsUnmasked ? Intrinsic::riscv_vsoxei : Intrinsic::riscv_vsoxei_mask;
SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
Ops.push_back(Val);
Ops.push_back(BasePtr);
Ops.push_back(Index);
if (!IsUnmasked)
Ops.push_back(Mask);
Ops.push_back(VL);
return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL,
DAG.getVTList(MVT::Other), Ops, MemVT, MMO);
}
SDValue RISCVTargetLowering::lowerGET_ROUNDING(SDValue Op,
SelectionDAG &DAG) const {
const MVT XLenVT = Subtarget.getXLenVT();
SDLoc DL(Op);
SDValue Chain = Op->getOperand(0);
SDValue SysRegNo = DAG.getTargetConstant(
RISCVSysReg::lookupSysRegByName("FRM")->Encoding, DL, XLenVT);
SDVTList VTs = DAG.getVTList(XLenVT, MVT::Other);
SDValue RM = DAG.getNode(RISCVISD::READ_CSR, DL, VTs, Chain, SysRegNo);
// Encoding used for rounding mode in RISC-V differs from that used in
// FLT_ROUNDS. To convert it the RISC-V rounding mode is used as an index in a
// table, which consists of a sequence of 4-bit fields, each representing
// corresponding FLT_ROUNDS mode.
static const int Table =
(int(RoundingMode::NearestTiesToEven) << 4 * RISCVFPRndMode::RNE) |
(int(RoundingMode::TowardZero) << 4 * RISCVFPRndMode::RTZ) |
(int(RoundingMode::TowardNegative) << 4 * RISCVFPRndMode::RDN) |
(int(RoundingMode::TowardPositive) << 4 * RISCVFPRndMode::RUP) |
(int(RoundingMode::NearestTiesToAway) << 4 * RISCVFPRndMode::RMM);
SDValue Shift =
DAG.getNode(ISD::SHL, DL, XLenVT, RM, DAG.getConstant(2, DL, XLenVT));
SDValue Shifted = DAG.getNode(ISD::SRL, DL, XLenVT,
DAG.getConstant(Table, DL, XLenVT), Shift);
SDValue Masked = DAG.getNode(ISD::AND, DL, XLenVT, Shifted,
DAG.getConstant(7, DL, XLenVT));
return DAG.getMergeValues({Masked, Chain}, DL);
}
SDValue RISCVTargetLowering::lowerSET_ROUNDING(SDValue Op,
SelectionDAG &DAG) const {
const MVT XLenVT = Subtarget.getXLenVT();
SDLoc DL(Op);
SDValue Chain = Op->getOperand(0);
SDValue RMValue = Op->getOperand(1);
SDValue SysRegNo = DAG.getTargetConstant(
RISCVSysReg::lookupSysRegByName("FRM")->Encoding, DL, XLenVT);
// Encoding used for rounding mode in RISC-V differs from that used in
// FLT_ROUNDS. To convert it the C rounding mode is used as an index in
// a table, which consists of a sequence of 4-bit fields, each representing
// corresponding RISC-V mode.
static const unsigned Table =
(RISCVFPRndMode::RNE << 4 * int(RoundingMode::NearestTiesToEven)) |
(RISCVFPRndMode::RTZ << 4 * int(RoundingMode::TowardZero)) |
(RISCVFPRndMode::RDN << 4 * int(RoundingMode::TowardNegative)) |
(RISCVFPRndMode::RUP << 4 * int(RoundingMode::TowardPositive)) |
(RISCVFPRndMode::RMM << 4 * int(RoundingMode::NearestTiesToAway));
RMValue = DAG.getNode(ISD::ZERO_EXTEND, DL, XLenVT, RMValue);
SDValue Shift = DAG.getNode(ISD::SHL, DL, XLenVT, RMValue,
DAG.getConstant(2, DL, XLenVT));
SDValue Shifted = DAG.getNode(ISD::SRL, DL, XLenVT,
DAG.getConstant(Table, DL, XLenVT), Shift);
RMValue = DAG.getNode(ISD::AND, DL, XLenVT, Shifted,
DAG.getConstant(0x7, DL, XLenVT));
return DAG.getNode(RISCVISD::WRITE_CSR, DL, MVT::Other, Chain, SysRegNo,
RMValue);
}
SDValue RISCVTargetLowering::lowerEH_DWARF_CFA(SDValue Op,
SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
bool isRISCV64 = Subtarget.is64Bit();
EVT PtrVT = getPointerTy(DAG.getDataLayout());
int FI = MF.getFrameInfo().CreateFixedObject(isRISCV64 ? 8 : 4, 0, false);
return DAG.getFrameIndex(FI, PtrVT);
}
// Returns the opcode of the target-specific SDNode that implements the 32-bit
// form of the given Opcode.
static RISCVISD::NodeType getRISCVWOpcode(unsigned Opcode) {
switch (Opcode) {
default:
llvm_unreachable("Unexpected opcode");
case ISD::SHL:
return RISCVISD::SLLW;
case ISD::SRA:
return RISCVISD::SRAW;
case ISD::SRL:
return RISCVISD::SRLW;
case ISD::SDIV:
return RISCVISD::DIVW;
case ISD::UDIV:
return RISCVISD::DIVUW;
case ISD::UREM:
return RISCVISD::REMUW;
case ISD::ROTL:
return RISCVISD::ROLW;
case ISD::ROTR:
return RISCVISD::RORW;
}
}
// Converts the given i8/i16/i32 operation to a target-specific SelectionDAG
// node. Because i8/i16/i32 isn't a legal type for RV64, these operations would
// otherwise be promoted to i64, making it difficult to select the
// SLLW/DIVUW/.../*W later one because the fact the operation was originally of
// type i8/i16/i32 is lost.
static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG,
unsigned ExtOpc = ISD::ANY_EXTEND) {
SDLoc DL(N);
RISCVISD::NodeType WOpcode = getRISCVWOpcode(N->getOpcode());
SDValue NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0));
SDValue NewOp1 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(1));
SDValue NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, NewOp1);
// ReplaceNodeResults requires we maintain the same type for the return value.
return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewRes);
}
// Converts the given 32-bit operation to a i64 operation with signed extension
// semantic to reduce the signed extension instructions.
static SDValue customLegalizeToWOpWithSExt(SDNode *N, SelectionDAG &DAG) {
SDLoc DL(N);
SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
SDValue NewOp1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue NewWOp = DAG.getNode(N->getOpcode(), DL, MVT::i64, NewOp0, NewOp1);
SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp,
DAG.getValueType(MVT::i32));
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes);
}
void RISCVTargetLowering::ReplaceNodeResults(SDNode *N,
SmallVectorImpl<SDValue> &Results,
SelectionDAG &DAG) const {
SDLoc DL(N);
switch (N->getOpcode()) {
default:
llvm_unreachable("Don't know how to custom type legalize this operation!");
case ISD::STRICT_FP_TO_SINT:
case ISD::STRICT_FP_TO_UINT:
case ISD::FP_TO_SINT:
case ISD::FP_TO_UINT: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
bool IsStrict = N->isStrictFPOpcode();
bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT ||
N->getOpcode() == ISD::STRICT_FP_TO_SINT;
SDValue Op0 = IsStrict ? N->getOperand(1) : N->getOperand(0);
if (getTypeAction(*DAG.getContext(), Op0.getValueType()) !=
TargetLowering::TypeSoftenFloat) {
if (!isTypeLegal(Op0.getValueType()))
return;
if (IsStrict) {
SDValue Chain = N->getOperand(0);
// In absense of Zfh, promote f16 to f32, then convert.
if (Op0.getValueType() == MVT::f16 &&
!Subtarget.hasStdExtZfhOrZhinx()) {
Op0 = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {MVT::f32, MVT::Other},
{Chain, Op0});
Chain = Op0.getValue(1);
}
unsigned Opc = IsSigned ? RISCVISD::STRICT_FCVT_W_RV64
: RISCVISD::STRICT_FCVT_WU_RV64;
SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other);
SDValue Res = DAG.getNode(
Opc, DL, VTs, Chain, Op0,
DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, MVT::i64));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
Results.push_back(Res.getValue(1));
return;
}
// For bf16, or f16 in absense of Zfh, promote [b]f16 to f32 and then
// convert.
if ((Op0.getValueType() == MVT::f16 &&
!Subtarget.hasStdExtZfhOrZhinx()) ||
Op0.getValueType() == MVT::bf16)
Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op0);
unsigned Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
SDValue Res =
DAG.getNode(Opc, DL, MVT::i64, Op0,
DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, MVT::i64));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
// If the FP type needs to be softened, emit a library call using the 'si'
// version. If we left it to default legalization we'd end up with 'di'. If
// the FP type doesn't need to be softened just let generic type
// legalization promote the result type.
RTLIB::Libcall LC;
if (IsSigned)
LC = RTLIB::getFPTOSINT(Op0.getValueType(), N->getValueType(0));
else
LC = RTLIB::getFPTOUINT(Op0.getValueType(), N->getValueType(0));
MakeLibCallOptions CallOptions;
EVT OpVT = Op0.getValueType();
CallOptions.setTypeListBeforeSoften(OpVT, N->getValueType(0), true);
SDValue Chain = IsStrict ? N->getOperand(0) : SDValue();
SDValue Result;
std::tie(Result, Chain) =
makeLibCall(DAG, LC, N->getValueType(0), Op0, CallOptions, DL, Chain);
Results.push_back(Result);
if (IsStrict)
Results.push_back(Chain);
break;
}
case ISD::LROUND: {
SDValue Op0 = N->getOperand(0);
EVT Op0VT = Op0.getValueType();
if (getTypeAction(*DAG.getContext(), Op0.getValueType()) !=
TargetLowering::TypeSoftenFloat) {
if (!isTypeLegal(Op0VT))
return;
// In absense of Zfh, promote f16 to f32, then convert.
if (Op0.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfhOrZhinx())
Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op0);
SDValue Res =
DAG.getNode(RISCVISD::FCVT_W_RV64, DL, MVT::i64, Op0,
DAG.getTargetConstant(RISCVFPRndMode::RMM, DL, MVT::i64));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
// If the FP type needs to be softened, emit a library call to lround. We'll
// need to truncate the result. We assume any value that doesn't fit in i32
// is allowed to return an unspecified value.
RTLIB::Libcall LC =
Op0.getValueType() == MVT::f64 ? RTLIB::LROUND_F64 : RTLIB::LROUND_F32;
MakeLibCallOptions CallOptions;
EVT OpVT = Op0.getValueType();
CallOptions.setTypeListBeforeSoften(OpVT, MVT::i64, true);
SDValue Result = makeLibCall(DAG, LC, MVT::i64, Op0, CallOptions, DL).first;
Result = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Result);
Results.push_back(Result);
break;
}
case ISD::READCYCLECOUNTER:
case ISD::READSTEADYCOUNTER: {
assert(!Subtarget.is64Bit() && "READCYCLECOUNTER/READSTEADYCOUNTER only "
"has custom type legalization on riscv32");
SDValue LoCounter, HiCounter;
MVT XLenVT = Subtarget.getXLenVT();
if (N->getOpcode() == ISD::READCYCLECOUNTER) {
LoCounter = DAG.getTargetConstant(
RISCVSysReg::lookupSysRegByName("CYCLE")->Encoding, DL, XLenVT);
HiCounter = DAG.getTargetConstant(
RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding, DL, XLenVT);
} else {
LoCounter = DAG.getTargetConstant(
RISCVSysReg::lookupSysRegByName("TIME")->Encoding, DL, XLenVT);
HiCounter = DAG.getTargetConstant(
RISCVSysReg::lookupSysRegByName("TIMEH")->Encoding, DL, XLenVT);
}
SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
SDValue RCW = DAG.getNode(RISCVISD::READ_COUNTER_WIDE, DL, VTs,
N->getOperand(0), LoCounter, HiCounter);
Results.push_back(
DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, RCW, RCW.getValue(1)));
Results.push_back(RCW.getValue(2));
break;
}
case ISD::LOAD: {
if (!ISD::isNON_EXTLoad(N))
return;
// Use a SEXTLOAD instead of the default EXTLOAD. Similar to the
// sext_inreg we emit for ADD/SUB/MUL/SLLI.
LoadSDNode *Ld = cast<LoadSDNode>(N);
SDLoc dl(N);
SDValue Res = DAG.getExtLoad(ISD::SEXTLOAD, dl, MVT::i64, Ld->getChain(),
Ld->getBasePtr(), Ld->getMemoryVT(),
Ld->getMemOperand());
Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Res));
Results.push_back(Res.getValue(1));
return;
}
case ISD::MUL: {
unsigned Size = N->getSimpleValueType(0).getSizeInBits();
unsigned XLen = Subtarget.getXLen();
// This multiply needs to be expanded, try to use MULHSU+MUL if possible.
if (Size > XLen) {
assert(Size == (XLen * 2) && "Unexpected custom legalisation");
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
APInt HighMask = APInt::getHighBitsSet(Size, XLen);
bool LHSIsU = DAG.MaskedValueIsZero(LHS, HighMask);
bool RHSIsU = DAG.MaskedValueIsZero(RHS, HighMask);
// We need exactly one side to be unsigned.
if (LHSIsU == RHSIsU)
return;
auto MakeMULPair = [&](SDValue S, SDValue U) {
MVT XLenVT = Subtarget.getXLenVT();
S = DAG.getNode(ISD::TRUNCATE, DL, XLenVT, S);
U = DAG.getNode(ISD::TRUNCATE, DL, XLenVT, U);
SDValue Lo = DAG.getNode(ISD::MUL, DL, XLenVT, S, U);
SDValue Hi = DAG.getNode(RISCVISD::MULHSU, DL, XLenVT, S, U);
return DAG.getNode(ISD::BUILD_PAIR, DL, N->getValueType(0), Lo, Hi);
};
bool LHSIsS = DAG.ComputeNumSignBits(LHS) > XLen;
bool RHSIsS = DAG.ComputeNumSignBits(RHS) > XLen;
// The other operand should be signed, but still prefer MULH when
// possible.
if (RHSIsU && LHSIsS && !RHSIsS)
Results.push_back(MakeMULPair(LHS, RHS));
else if (LHSIsU && RHSIsS && !LHSIsS)
Results.push_back(MakeMULPair(RHS, LHS));
return;
}
[[fallthrough]];
}
case ISD::ADD:
case ISD::SUB:
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
Results.push_back(customLegalizeToWOpWithSExt(N, DAG));
break;
case ISD::SHL:
case ISD::SRA:
case ISD::SRL:
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
if (N->getOperand(1).getOpcode() != ISD::Constant) {
// If we can use a BSET instruction, allow default promotion to apply.
if (N->getOpcode() == ISD::SHL && Subtarget.hasStdExtZbs() &&
isOneConstant(N->getOperand(0)))
break;
Results.push_back(customLegalizeToWOp(N, DAG));
break;
}
// Custom legalize ISD::SHL by placing a SIGN_EXTEND_INREG after. This is
// similar to customLegalizeToWOpWithSExt, but we must zero_extend the
// shift amount.
if (N->getOpcode() == ISD::SHL) {
SDLoc DL(N);
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
SDValue NewOp1 =
DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue NewWOp = DAG.getNode(ISD::SHL, DL, MVT::i64, NewOp0, NewOp1);
SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp,
DAG.getValueType(MVT::i32));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes));
}
break;
case ISD::ROTL:
case ISD::ROTR:
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
assert((Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb() ||
Subtarget.hasVendorXTHeadBb()) &&
"Unexpected custom legalization");
if (!isa<ConstantSDNode>(N->getOperand(1)) &&
!(Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()))
return;
Results.push_back(customLegalizeToWOp(N, DAG));
break;
case ISD::CTTZ:
case ISD::CTTZ_ZERO_UNDEF:
case ISD::CTLZ:
case ISD::CTLZ_ZERO_UNDEF: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
bool IsCTZ =
N->getOpcode() == ISD::CTTZ || N->getOpcode() == ISD::CTTZ_ZERO_UNDEF;
unsigned Opc = IsCTZ ? RISCVISD::CTZW : RISCVISD::CLZW;
SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp0);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case ISD::SDIV:
case ISD::UDIV:
case ISD::UREM: {
MVT VT = N->getSimpleValueType(0);
assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
Subtarget.is64Bit() && Subtarget.hasStdExtM() &&
"Unexpected custom legalisation");
// Don't promote division/remainder by constant since we should expand those
// to multiply by magic constant.
AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
if (N->getOperand(1).getOpcode() == ISD::Constant &&
!isIntDivCheap(N->getValueType(0), Attr))
return;
// If the input is i32, use ANY_EXTEND since the W instructions don't read
// the upper 32 bits. For other types we need to sign or zero extend
// based on the opcode.
unsigned ExtOpc = ISD::ANY_EXTEND;
if (VT != MVT::i32)
ExtOpc = N->getOpcode() == ISD::SDIV ? ISD::SIGN_EXTEND
: ISD::ZERO_EXTEND;
Results.push_back(customLegalizeToWOp(N, DAG, ExtOpc));
break;
}
case ISD::SADDO: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
// If the RHS is a constant, we can simplify ConditionRHS below. Otherwise
// use the default legalization.
if (!isa<ConstantSDNode>(N->getOperand(1)))
return;
SDValue LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(0));
SDValue RHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue Res = DAG.getNode(ISD::ADD, DL, MVT::i64, LHS, RHS);
Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Res,
DAG.getValueType(MVT::i32));
SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
// For an addition, the result should be less than one of the operands (LHS)
// if and only if the other operand (RHS) is negative, otherwise there will
// be overflow.
// For a subtraction, the result should be less than one of the operands
// (LHS) if and only if the other operand (RHS) is (non-zero) positive,
// otherwise there will be overflow.
EVT OType = N->getValueType(1);
SDValue ResultLowerThanLHS = DAG.getSetCC(DL, OType, Res, LHS, ISD::SETLT);
SDValue ConditionRHS = DAG.getSetCC(DL, OType, RHS, Zero, ISD::SETLT);
SDValue Overflow =
DAG.getNode(ISD::XOR, DL, OType, ConditionRHS, ResultLowerThanLHS);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
Results.push_back(Overflow);
return;
}
case ISD::UADDO:
case ISD::USUBO: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
bool IsAdd = N->getOpcode() == ISD::UADDO;
// Create an ADDW or SUBW.
SDValue LHS = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
SDValue RHS = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue Res =
DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, DL, MVT::i64, LHS, RHS);
Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Res,
DAG.getValueType(MVT::i32));
SDValue Overflow;
if (IsAdd && isOneConstant(RHS)) {
// Special case uaddo X, 1 overflowed if the addition result is 0.
// The general case (X + C) < C is not necessarily beneficial. Although we
// reduce the live range of X, we may introduce the materialization of
// constant C, especially when the setcc result is used by branch. We have
// no compare with constant and branch instructions.
Overflow = DAG.getSetCC(DL, N->getValueType(1), Res,
DAG.getConstant(0, DL, MVT::i64), ISD::SETEQ);
} else if (IsAdd && isAllOnesConstant(RHS)) {
// Special case uaddo X, -1 overflowed if X != 0.
Overflow = DAG.getSetCC(DL, N->getValueType(1), N->getOperand(0),
DAG.getConstant(0, DL, MVT::i32), ISD::SETNE);
} else {
// Sign extend the LHS and perform an unsigned compare with the ADDW
// result. Since the inputs are sign extended from i32, this is equivalent
// to comparing the lower 32 bits.
LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(0));
Overflow = DAG.getSetCC(DL, N->getValueType(1), Res, LHS,
IsAdd ? ISD::SETULT : ISD::SETUGT);
}
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
Results.push_back(Overflow);
return;
}
case ISD::UADDSAT:
case ISD::USUBSAT: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
if (Subtarget.hasStdExtZbb()) {
// With Zbb we can sign extend and let LegalizeDAG use minu/maxu. Using
// sign extend allows overflow of the lower 32 bits to be detected on
// the promoted size.
SDValue LHS =
DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(0));
SDValue RHS =
DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue Res = DAG.getNode(N->getOpcode(), DL, MVT::i64, LHS, RHS);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
// Without Zbb, expand to UADDO/USUBO+select which will trigger our custom
// promotion for UADDO/USUBO.
Results.push_back(expandAddSubSat(N, DAG));
return;
}
case ISD::SADDSAT:
case ISD::SSUBSAT: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
Results.push_back(expandAddSubSat(N, DAG));
return;
}
case ISD::ABS: {
assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
"Unexpected custom legalisation");
if (Subtarget.hasStdExtZbb()) {
// Emit a special ABSW node that will be expanded to NEGW+MAX at isel.
// This allows us to remember that the result is sign extended. Expanding
// to NEGW+MAX here requires a Freeze which breaks ComputeNumSignBits.
SDValue Src = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64,
N->getOperand(0));
SDValue Abs = DAG.getNode(RISCVISD::ABSW, DL, MVT::i64, Src);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Abs));
return;
}
// Expand abs to Y = (sraiw X, 31); subw(xor(X, Y), Y)
SDValue Src = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
// Freeze the source so we can increase it's use count.
Src = DAG.getFreeze(Src);
// Copy sign bit to all bits using the sraiw pattern.
SDValue SignFill = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Src,
DAG.getValueType(MVT::i32));
SignFill = DAG.getNode(ISD::SRA, DL, MVT::i64, SignFill,
DAG.getConstant(31, DL, MVT::i64));
SDValue NewRes = DAG.getNode(ISD::XOR, DL, MVT::i64, Src, SignFill);
NewRes = DAG.getNode(ISD::SUB, DL, MVT::i64, NewRes, SignFill);
// NOTE: The result is only required to be anyextended, but sext is
// consistent with type legalization of sub.
NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewRes,
DAG.getValueType(MVT::i32));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes));
return;
}
case ISD::BITCAST: {
EVT VT = N->getValueType(0);
assert(VT.isInteger() && !VT.isVector() && "Unexpected VT!");
SDValue Op0 = N->getOperand(0);
EVT Op0VT = Op0.getValueType();
MVT XLenVT = Subtarget.getXLenVT();
if (VT == MVT::i16 && Op0VT == MVT::f16 &&
Subtarget.hasStdExtZfhminOrZhinxmin()) {
SDValue FPConv = DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, XLenVT, Op0);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FPConv));
} else if (VT == MVT::i16 && Op0VT == MVT::bf16 &&
Subtarget.hasStdExtZfbfmin()) {
SDValue FPConv = DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, XLenVT, Op0);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FPConv));
} else if (VT == MVT::i32 && Op0VT == MVT::f32 && Subtarget.is64Bit() &&
Subtarget.hasStdExtFOrZfinx()) {
SDValue FPConv =
DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Op0);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, FPConv));
} else if (VT == MVT::i64 && Op0VT == MVT::f64 && XLenVT == MVT::i32) {
SDValue NewReg = DAG.getNode(RISCVISD::SplitF64, DL,
DAG.getVTList(MVT::i32, MVT::i32), Op0);
SDValue RetReg = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
NewReg.getValue(0), NewReg.getValue(1));
Results.push_back(RetReg);
} else if (!VT.isVector() && Op0VT.isFixedLengthVector() &&
isTypeLegal(Op0VT)) {
// Custom-legalize bitcasts from fixed-length vector types to illegal
// scalar types in order to improve codegen. Bitcast the vector to a
// one-element vector type whose element type is the same as the result
// type, and extract the first element.
EVT BVT = EVT::getVectorVT(*DAG.getContext(), VT, 1);
if (isTypeLegal(BVT)) {
SDValue BVec = DAG.getBitcast(BVT, Op0);
Results.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, BVec,
DAG.getVectorIdxConstant(0, DL)));
}
}
break;
}
case RISCVISD::BREV8: {
MVT VT = N->getSimpleValueType(0);
MVT XLenVT = Subtarget.getXLenVT();
assert((VT == MVT::i16 || (VT == MVT::i32 && Subtarget.is64Bit())) &&
"Unexpected custom legalisation");
assert(Subtarget.hasStdExtZbkb() && "Unexpected extension");
SDValue NewOp = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, N->getOperand(0));
SDValue NewRes = DAG.getNode(N->getOpcode(), DL, XLenVT, NewOp);
// ReplaceNodeResults requires we maintain the same type for the return
// value.
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, NewRes));
break;
}
case ISD::EXTRACT_VECTOR_ELT: {
// Custom-legalize an EXTRACT_VECTOR_ELT where XLEN<SEW, as the SEW element
// type is illegal (currently only vXi64 RV32).
// With vmv.x.s, when SEW > XLEN, only the least-significant XLEN bits are
// transferred to the destination register. We issue two of these from the
// upper- and lower- halves of the SEW-bit vector element, slid down to the
// first element.
SDValue Vec = N->getOperand(0);
SDValue Idx = N->getOperand(1);
// The vector type hasn't been legalized yet so we can't issue target
// specific nodes if it needs legalization.
// FIXME: We would manually legalize if it's important.
if (!isTypeLegal(Vec.getValueType()))
return;
MVT VecVT = Vec.getSimpleValueType();
assert(!Subtarget.is64Bit() && N->getValueType(0) == MVT::i64 &&
VecVT.getVectorElementType() == MVT::i64 &&
"Unexpected EXTRACT_VECTOR_ELT legalization");
// If this is a fixed vector, we need to convert it to a scalable vector.
MVT ContainerVT = VecVT;
if (VecVT.isFixedLengthVector()) {
ContainerVT = getContainerForFixedLengthVector(VecVT);
Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
}
MVT XLenVT = Subtarget.getXLenVT();
// Use a VL of 1 to avoid processing more elements than we need.
auto [Mask, VL] = getDefaultVLOps(1, ContainerVT, DL, DAG, Subtarget);
// Unless the index is known to be 0, we must slide the vector down to get
// the desired element into index 0.
if (!isNullConstant(Idx)) {
Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT,
DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL);
}
// Extract the lower XLEN bits of the correct vector element.
SDValue EltLo = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec);
// To extract the upper XLEN bits of the vector element, shift the first
// element right by 32 bits and re-extract the lower XLEN bits.
SDValue ThirtyTwoV = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
DAG.getUNDEF(ContainerVT),
DAG.getConstant(32, DL, XLenVT), VL);
SDValue LShr32 =
DAG.getNode(RISCVISD::SRL_VL, DL, ContainerVT, Vec, ThirtyTwoV,
DAG.getUNDEF(ContainerVT), Mask, VL);
SDValue EltHi = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, LShr32);
Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, EltLo, EltHi));
break;
}
case ISD::INTRINSIC_WO_CHAIN: {
unsigned IntNo = N->getConstantOperandVal(0);
switch (IntNo) {
default:
llvm_unreachable(
"Don't know how to custom type legalize this intrinsic!");
case Intrinsic::experimental_get_vector_length: {
SDValue Res = lowerGetVectorLength(N, DAG, Subtarget);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::experimental_cttz_elts: {
SDValue Res = lowerCttzElts(N, DAG, Subtarget);
Results.push_back(
DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), Res));
return;
}
case Intrinsic::riscv_orc_b:
case Intrinsic::riscv_brev8:
case Intrinsic::riscv_sha256sig0:
case Intrinsic::riscv_sha256sig1:
case Intrinsic::riscv_sha256sum0:
case Intrinsic::riscv_sha256sum1:
case Intrinsic::riscv_sm3p0:
case Intrinsic::riscv_sm3p1: {
if (!Subtarget.is64Bit() || N->getValueType(0) != MVT::i32)
return;
unsigned Opc;
switch (IntNo) {
case Intrinsic::riscv_orc_b: Opc = RISCVISD::ORC_B; break;
case Intrinsic::riscv_brev8: Opc = RISCVISD::BREV8; break;
case Intrinsic::riscv_sha256sig0: Opc = RISCVISD::SHA256SIG0; break;
case Intrinsic::riscv_sha256sig1: Opc = RISCVISD::SHA256SIG1; break;
case Intrinsic::riscv_sha256sum0: Opc = RISCVISD::SHA256SUM0; break;
case Intrinsic::riscv_sha256sum1: Opc = RISCVISD::SHA256SUM1; break;
case Intrinsic::riscv_sm3p0: Opc = RISCVISD::SM3P0; break;
case Intrinsic::riscv_sm3p1: Opc = RISCVISD::SM3P1; break;
}
SDValue NewOp =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::riscv_sm4ks:
case Intrinsic::riscv_sm4ed: {
unsigned Opc =
IntNo == Intrinsic::riscv_sm4ks ? RISCVISD::SM4KS : RISCVISD::SM4ED;
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(2));
SDValue Res =
DAG.getNode(Opc, DL, MVT::i64, NewOp0, NewOp1, N->getOperand(3));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::riscv_mopr: {
if (!Subtarget.is64Bit() || N->getValueType(0) != MVT::i32)
return;
SDValue NewOp =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue Res = DAG.getNode(
RISCVISD::MOPR, DL, MVT::i64, NewOp,
DAG.getTargetConstant(N->getConstantOperandVal(2), DL, MVT::i64));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::riscv_moprr: {
if (!Subtarget.is64Bit() || N->getValueType(0) != MVT::i32)
return;
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(2));
SDValue Res = DAG.getNode(
RISCVISD::MOPRR, DL, MVT::i64, NewOp0, NewOp1,
DAG.getTargetConstant(N->getConstantOperandVal(3), DL, MVT::i64));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::riscv_clmul: {
if (!Subtarget.is64Bit() || N->getValueType(0) != MVT::i32)
return;
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(2));
SDValue Res = DAG.getNode(RISCVISD::CLMUL, DL, MVT::i64, NewOp0, NewOp1);
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::riscv_clmulh:
case Intrinsic::riscv_clmulr: {
if (!Subtarget.is64Bit() || N->getValueType(0) != MVT::i32)
return;
// Extend inputs to XLen, and shift by 32. This will add 64 trailing zeros
// to the full 128-bit clmul result of multiplying two xlen values.
// Perform clmulr or clmulh on the shifted values. Finally, extract the
// upper 32 bits.
//
// The alternative is to mask the inputs to 32 bits and use clmul, but
// that requires two shifts to mask each input without zext.w.
// FIXME: If the inputs are known zero extended or could be freely
// zero extended, the mask form would be better.
SDValue NewOp0 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
SDValue NewOp1 =
DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(2));
NewOp0 = DAG.getNode(ISD::SHL, DL, MVT::i64, NewOp0,
DAG.getConstant(32, DL, MVT::i64));
NewOp1 = DAG.getNode(ISD::SHL, DL, MVT::i64, NewOp1,
DAG.getConstant(32, DL, MVT::i64));
unsigned Opc = IntNo == Intrinsic::riscv_clmulh ? RISCVISD::CLMULH
: RISCVISD::CLMULR;
SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp0, NewOp1);
Res = DAG.getNode(ISD::SRL, DL, MVT::i64, Res,
DAG.getConstant(32, DL, MVT::i64));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
return;
}
case Intrinsic::riscv_vmv_x_s: {
EVT VT = N->getValueType(0);
MVT XLenVT = Subtarget.getXLenVT();
if (VT.bitsLT(XLenVT)) {
// Simple case just extract using vmv.x.s and truncate.
SDValue Extract = DAG.getNode(RISCVISD::VMV_X_S, DL,
Subtarget.getXLenVT(), N->getOperand(1));
Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Extract));
return;
}
assert(VT == MVT::i64 && !Subtarget.is64Bit() &&
"Unexpected custom legalization");
// We need to do the move in two steps.
SDValue Vec = N->getOperand(1);
MVT VecVT = Vec.getSimpleValueType();
// First extract the lower XLEN bits of the element.
SDValue EltLo = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec);
// To extract the upper XLEN bits of the vector element, shift the first
// element right by 32 bits and re-extract the lower XLEN bits.
auto [Mask, VL] = getDefaultVLOps(1, VecVT, DL, DAG, Subtarget);
SDValue ThirtyTwoV =
DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VecVT, DAG.getUNDEF(VecVT),
DAG.getConstant(32, DL, XLenVT), VL);
SDValue LShr32 = DAG.getNode(RISCVISD::SRL_VL, DL, VecVT, Vec, ThirtyTwoV,
DAG.getUNDEF(VecVT), Mask, VL);
SDValue EltHi = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, LShr32);
Results.push_back(
DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, EltLo, EltHi));
break;
}
}
break;
}
case ISD::VECREDUCE_ADD:
case ISD::VECREDUCE_AND:
case ISD::VECREDUCE_OR:
case ISD::VECREDUCE_XOR:
case ISD::VECREDUCE_SMAX:
case ISD::VECREDUCE_UMAX:
case ISD::VECREDUCE_SMIN:
case ISD::VECREDUCE_UMIN:
if (SDValue V = lowerVECREDUCE(SDValue(N, 0), DAG))
Results.push_back(V);
break;
case ISD::VP_REDUCE_ADD:
case ISD::VP_REDUCE_AND:
case ISD::VP_REDUCE_OR:
case ISD::VP_REDUCE_XOR:
case ISD::VP_REDUCE_SMAX:
case ISD::VP_REDUCE_UMAX:
case ISD::VP_REDUCE_SMIN:
case ISD::VP_REDUCE_UMIN:
if (SDValue V = lowerVPREDUCE(SDValue(N, 0), DAG))
Results.push_back(V);
break;
case ISD::GET_ROUNDING: {
SDVTList VTs = DAG.getVTList(Subtarget.getXLenVT(), MVT::Other);
SDValue Res = DAG.getNode(ISD::GET_ROUNDING, DL, VTs, N->getOperand(0));
Results.push_back(Res.getValue(0));
Results.push_back(Res.getValue(1));
break;
}
}
}
/// Given a binary operator, return the *associative* generic ISD::VECREDUCE_OP
/// which corresponds to it.
static unsigned getVecReduceOpcode(unsigned Opc) {
switch (Opc) {
default:
llvm_unreachable("Unhandled binary to transfrom reduction");
case ISD::ADD:
return ISD::VECREDUCE_ADD;
case ISD::UMAX:
return ISD::VECREDUCE_UMAX;
case ISD::SMAX:
return ISD::VECREDUCE_SMAX;
case ISD::UMIN:
return ISD::VECREDUCE_UMIN;
case ISD::SMIN:
return ISD::VECREDUCE_SMIN;
case ISD::AND:
return ISD::VECREDUCE_AND;
case ISD::OR:
return ISD::VECREDUCE_OR;
case ISD::XOR:
return ISD::VECREDUCE_XOR;
case ISD::FADD:
// Note: This is the associative form of the generic reduction opcode.
return ISD::VECREDUCE_FADD;
}
}
/// Perform two related transforms whose purpose is to incrementally recognize
/// an explode_vector followed by scalar reduction as a vector reduction node.
/// This exists to recover from a deficiency in SLP which can't handle
/// forests with multiple roots sharing common nodes. In some cases, one
/// of the trees will be vectorized, and the other will remain (unprofitably)
/// scalarized.
static SDValue
combineBinOpOfExtractToReduceTree(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
// This transforms need to run before all integer types have been legalized
// to i64 (so that the vector element type matches the add type), and while
// it's safe to introduce odd sized vector types.
if (DAG.NewNodesMustHaveLegalTypes)
return SDValue();
// Without V, this transform isn't useful. We could form the (illegal)
// operations and let them be scalarized again, but there's really no point.
if (!Subtarget.hasVInstructions())
return SDValue();
const SDLoc DL(N);
const EVT VT = N->getValueType(0);
const unsigned Opc = N->getOpcode();
// For FADD, we only handle the case with reassociation allowed. We
// could handle strict reduction order, but at the moment, there's no
// known reason to, and the complexity isn't worth it.
// TODO: Handle fminnum and fmaxnum here
if (!VT.isInteger() &&
(Opc != ISD::FADD || !N->getFlags().hasAllowReassociation()))
return SDValue();
const unsigned ReduceOpc = getVecReduceOpcode(Opc);
assert(Opc == ISD::getVecReduceBaseOpcode(ReduceOpc) &&
"Inconsistent mappings");
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
if (!LHS.hasOneUse() || !RHS.hasOneUse())
return SDValue();
if (RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
std::swap(LHS, RHS);
if (RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
!isa<ConstantSDNode>(RHS.getOperand(1)))
return SDValue();
uint64_t RHSIdx = cast<ConstantSDNode>(RHS.getOperand(1))->getLimitedValue();
SDValue SrcVec = RHS.getOperand(0);
EVT SrcVecVT = SrcVec.getValueType();
assert(SrcVecVT.getVectorElementType() == VT);
if (SrcVecVT.isScalableVector())
return SDValue();
if (SrcVecVT.getScalarSizeInBits() > Subtarget.getELen())
return SDValue();
// match binop (extract_vector_elt V, 0), (extract_vector_elt V, 1) to
// reduce_op (extract_subvector [2 x VT] from V). This will form the
// root of our reduction tree. TODO: We could extend this to any two
// adjacent aligned constant indices if desired.
if (LHS.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
LHS.getOperand(0) == SrcVec && isa<ConstantSDNode>(LHS.getOperand(1))) {
uint64_t LHSIdx =
cast<ConstantSDNode>(LHS.getOperand(1))->getLimitedValue();
if (0 == std::min(LHSIdx, RHSIdx) && 1 == std::max(LHSIdx, RHSIdx)) {
EVT ReduceVT = EVT::getVectorVT(*DAG.getContext(), VT, 2);
SDValue Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ReduceVT, SrcVec,
DAG.getVectorIdxConstant(0, DL));
return DAG.getNode(ReduceOpc, DL, VT, Vec, N->getFlags());
}
}
// Match (binop (reduce (extract_subvector V, 0),
// (extract_vector_elt V, sizeof(SubVec))))
// into a reduction of one more element from the original vector V.
if (LHS.getOpcode() != ReduceOpc)
return SDValue();
SDValue ReduceVec = LHS.getOperand(0);
if (ReduceVec.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
ReduceVec.hasOneUse() && ReduceVec.getOperand(0) == RHS.getOperand(0) &&
isNullConstant(ReduceVec.getOperand(1)) &&
ReduceVec.getValueType().getVectorNumElements() == RHSIdx) {
// For illegal types (e.g. 3xi32), most will be combined again into a
// wider (hopefully legal) type. If this is a terminal state, we are
// relying on type legalization here to produce something reasonable
// and this lowering quality could probably be improved. (TODO)
EVT ReduceVT = EVT::getVectorVT(*DAG.getContext(), VT, RHSIdx + 1);
SDValue Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ReduceVT, SrcVec,
DAG.getVectorIdxConstant(0, DL));
auto Flags = ReduceVec->getFlags();
Flags.intersectWith(N->getFlags());
return DAG.getNode(ReduceOpc, DL, VT, Vec, Flags);
}
return SDValue();
}
// Try to fold (<bop> x, (reduction.<bop> vec, start))
static SDValue combineBinOpToReduce(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
auto BinOpToRVVReduce = [](unsigned Opc) {
switch (Opc) {
default:
llvm_unreachable("Unhandled binary to transfrom reduction");
case ISD::ADD:
return RISCVISD::VECREDUCE_ADD_VL;
case ISD::UMAX:
return RISCVISD::VECREDUCE_UMAX_VL;
case ISD::SMAX:
return RISCVISD::VECREDUCE_SMAX_VL;
case ISD::UMIN:
return RISCVISD::VECREDUCE_UMIN_VL;
case ISD::SMIN:
return RISCVISD::VECREDUCE_SMIN_VL;
case ISD::AND:
return RISCVISD::VECREDUCE_AND_VL;
case ISD::OR:
return RISCVISD::VECREDUCE_OR_VL;
case ISD::XOR:
return RISCVISD::VECREDUCE_XOR_VL;
case ISD::FADD:
return RISCVISD::VECREDUCE_FADD_VL;
case ISD::FMAXNUM:
return RISCVISD::VECREDUCE_FMAX_VL;
case ISD::FMINNUM:
return RISCVISD::VECREDUCE_FMIN_VL;
}
};
auto IsReduction = [&BinOpToRVVReduce](SDValue V, unsigned Opc) {
return V.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
isNullConstant(V.getOperand(1)) &&
V.getOperand(0).getOpcode() == BinOpToRVVReduce(Opc);
};
unsigned Opc = N->getOpcode();
unsigned ReduceIdx;
if (IsReduction(N->getOperand(0), Opc))
ReduceIdx = 0;
else if (IsReduction(N->getOperand(1), Opc))
ReduceIdx = 1;
else
return SDValue();
// Skip if FADD disallows reassociation but the combiner needs.
if (Opc == ISD::FADD && !N->getFlags().hasAllowReassociation())
return SDValue();
SDValue Extract = N->getOperand(ReduceIdx);
SDValue Reduce = Extract.getOperand(0);
if (!Extract.hasOneUse() || !Reduce.hasOneUse())
return SDValue();
SDValue ScalarV = Reduce.getOperand(2);
EVT ScalarVT = ScalarV.getValueType();
if (ScalarV.getOpcode() == ISD::INSERT_SUBVECTOR &&
ScalarV.getOperand(0)->isUndef() &&
isNullConstant(ScalarV.getOperand(2)))
ScalarV = ScalarV.getOperand(1);
// Make sure that ScalarV is a splat with VL=1.
if (ScalarV.getOpcode() != RISCVISD::VFMV_S_F_VL &&
ScalarV.getOpcode() != RISCVISD::VMV_S_X_VL &&
ScalarV.getOpcode() != RISCVISD::VMV_V_X_VL)
return SDValue();
if (!isNonZeroAVL(ScalarV.getOperand(2)))
return SDValue();
// Check the scalar of ScalarV is neutral element
// TODO: Deal with value other than neutral element.
if (!isNeutralConstant(N->getOpcode(), N->getFlags(), ScalarV.getOperand(1),
0))
return SDValue();
// If the AVL is zero, operand 0 will be returned. So it's not safe to fold.
// FIXME: We might be able to improve this if operand 0 is undef.
if (!isNonZeroAVL(Reduce.getOperand(5)))
return SDValue();
SDValue NewStart = N->getOperand(1 - ReduceIdx);
SDLoc DL(N);
SDValue NewScalarV =
lowerScalarInsert(NewStart, ScalarV.getOperand(2),
ScalarV.getSimpleValueType(), DL, DAG, Subtarget);
// If we looked through an INSERT_SUBVECTOR we need to restore it.
if (ScalarVT != ScalarV.getValueType())
NewScalarV =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ScalarVT, DAG.getUNDEF(ScalarVT),
NewScalarV, DAG.getVectorIdxConstant(0, DL));
SDValue Ops[] = {Reduce.getOperand(0), Reduce.getOperand(1),
NewScalarV, Reduce.getOperand(3),
Reduce.getOperand(4), Reduce.getOperand(5)};
SDValue NewReduce =
DAG.getNode(Reduce.getOpcode(), DL, Reduce.getValueType(), Ops);
return DAG.getNode(Extract.getOpcode(), DL, Extract.getValueType(), NewReduce,
Extract.getOperand(1));
}
// Optimize (add (shl x, c0), (shl y, c1)) ->
// (SLLI (SH*ADD x, y), c0), if c1-c0 equals to [1|2|3].
static SDValue transformAddShlImm(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
// Perform this optimization only in the zba extension.
if (!Subtarget.hasStdExtZba())
return SDValue();
// Skip for vector types and larger types.
EVT VT = N->getValueType(0);
if (VT.isVector() || VT.getSizeInBits() > Subtarget.getXLen())
return SDValue();
// The two operand nodes must be SHL and have no other use.
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
if (N0->getOpcode() != ISD::SHL || N1->getOpcode() != ISD::SHL ||
!N0->hasOneUse() || !N1->hasOneUse())
return SDValue();
// Check c0 and c1.
auto *N0C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
auto *N1C = dyn_cast<ConstantSDNode>(N1->getOperand(1));
if (!N0C || !N1C)
return SDValue();
int64_t C0 = N0C->getSExtValue();
int64_t C1 = N1C->getSExtValue();
if (C0 <= 0 || C1 <= 0)
return SDValue();
// Skip if SH1ADD/SH2ADD/SH3ADD are not applicable.
int64_t Bits = std::min(C0, C1);
int64_t Diff = std::abs(C0 - C1);
if (Diff != 1 && Diff != 2 && Diff != 3)
return SDValue();
// Build nodes.
SDLoc DL(N);
SDValue NS = (C0 < C1) ? N0->getOperand(0) : N1->getOperand(0);
SDValue NL = (C0 > C1) ? N0->getOperand(0) : N1->getOperand(0);
SDValue NA0 =
DAG.getNode(ISD::SHL, DL, VT, NL, DAG.getConstant(Diff, DL, VT));
SDValue NA1 = DAG.getNode(ISD::ADD, DL, VT, NA0, NS);
return DAG.getNode(ISD::SHL, DL, VT, NA1, DAG.getConstant(Bits, DL, VT));
}
// Combine a constant select operand into its use:
//
// (and (select cond, -1, c), x)
// -> (select cond, x, (and x, c)) [AllOnes=1]
// (or (select cond, 0, c), x)
// -> (select cond, x, (or x, c)) [AllOnes=0]
// (xor (select cond, 0, c), x)
// -> (select cond, x, (xor x, c)) [AllOnes=0]
// (add (select cond, 0, c), x)
// -> (select cond, x, (add x, c)) [AllOnes=0]
// (sub x, (select cond, 0, c))
// -> (select cond, x, (sub x, c)) [AllOnes=0]
static SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
SelectionDAG &DAG, bool AllOnes,
const RISCVSubtarget &Subtarget) {
EVT VT = N->getValueType(0);
// Skip vectors.
if (VT.isVector())
return SDValue();
if (!Subtarget.hasConditionalMoveFusion()) {
// (select cond, x, (and x, c)) has custom lowering with Zicond.
if ((!Subtarget.hasStdExtZicond() &&
!Subtarget.hasVendorXVentanaCondOps()) ||
N->getOpcode() != ISD::AND)
return SDValue();
// Maybe harmful when condition code has multiple use.
if (Slct.getOpcode() == ISD::SELECT && !Slct.getOperand(0).hasOneUse())
return SDValue();
// Maybe harmful when VT is wider than XLen.
if (VT.getSizeInBits() > Subtarget.getXLen())
return SDValue();
}
if ((Slct.getOpcode() != ISD::SELECT &&
Slct.getOpcode() != RISCVISD::SELECT_CC) ||
!Slct.hasOneUse())
return SDValue();
auto isZeroOrAllOnes = [](SDValue N, bool AllOnes) {
return AllOnes ? isAllOnesConstant(N) : isNullConstant(N);
};
bool SwapSelectOps;
unsigned OpOffset = Slct.getOpcode() == RISCVISD::SELECT_CC ? 2 : 0;
SDValue TrueVal = Slct.getOperand(1 + OpOffset);
SDValue FalseVal = Slct.getOperand(2 + OpOffset);
SDValue NonConstantVal;
if (isZeroOrAllOnes(TrueVal, AllOnes)) {
SwapSelectOps = false;
NonConstantVal = FalseVal;
} else if (isZeroOrAllOnes(FalseVal, AllOnes)) {
SwapSelectOps = true;
NonConstantVal = TrueVal;
} else
return SDValue();
// Slct is now know to be the desired identity constant when CC is true.
TrueVal = OtherOp;
FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT, OtherOp, NonConstantVal);
// Unless SwapSelectOps says the condition should be false.
if (SwapSelectOps)
std::swap(TrueVal, FalseVal);
if (Slct.getOpcode() == RISCVISD::SELECT_CC)
return DAG.getNode(RISCVISD::SELECT_CC, SDLoc(N), VT,
{Slct.getOperand(0), Slct.getOperand(1),
Slct.getOperand(2), TrueVal, FalseVal});
return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
{Slct.getOperand(0), TrueVal, FalseVal});
}
// Attempt combineSelectAndUse on each operand of a commutative operator N.
static SDValue combineSelectAndUseCommutative(SDNode *N, SelectionDAG &DAG,
bool AllOnes,
const RISCVSubtarget &Subtarget) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
if (SDValue Result = combineSelectAndUse(N, N0, N1, DAG, AllOnes, Subtarget))
return Result;
if (SDValue Result = combineSelectAndUse(N, N1, N0, DAG, AllOnes, Subtarget))
return Result;
return SDValue();
}
// Transform (add (mul x, c0), c1) ->
// (add (mul (add x, c1/c0), c0), c1%c0).
// if c1/c0 and c1%c0 are simm12, while c1 is not. A special corner case
// that should be excluded is when c0*(c1/c0) is simm12, which will lead
// to an infinite loop in DAGCombine if transformed.
// Or transform (add (mul x, c0), c1) ->
// (add (mul (add x, c1/c0+1), c0), c1%c0-c0),
// if c1/c0+1 and c1%c0-c0 are simm12, while c1 is not. A special corner
// case that should be excluded is when c0*(c1/c0+1) is simm12, which will
// lead to an infinite loop in DAGCombine if transformed.
// Or transform (add (mul x, c0), c1) ->
// (add (mul (add x, c1/c0-1), c0), c1%c0+c0),
// if c1/c0-1 and c1%c0+c0 are simm12, while c1 is not. A special corner
// case that should be excluded is when c0*(c1/c0-1) is simm12, which will
// lead to an infinite loop in DAGCombine if transformed.
// Or transform (add (mul x, c0), c1) ->
// (mul (add x, c1/c0), c0).
// if c1%c0 is zero, and c1/c0 is simm12 while c1 is not.
static SDValue transformAddImmMulImm(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
// Skip for vector types and larger types.
EVT VT = N->getValueType(0);
if (VT.isVector() || VT.getSizeInBits() > Subtarget.getXLen())
return SDValue();
// The first operand node must be a MUL and has no other use.
SDValue N0 = N->getOperand(0);
if (!N0->hasOneUse() || N0->getOpcode() != ISD::MUL)
return SDValue();
// Check if c0 and c1 match above conditions.
auto *N0C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
auto *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (!N0C || !N1C)
return SDValue();
// If N0C has multiple uses it's possible one of the cases in
// DAGCombiner::isMulAddWithConstProfitable will be true, which would result
// in an infinite loop.
if (!N0C->hasOneUse())
return SDValue();
int64_t C0 = N0C->getSExtValue();
int64_t C1 = N1C->getSExtValue();
int64_t CA, CB;
if (C0 == -1 || C0 == 0 || C0 == 1 || isInt<12>(C1))
return SDValue();
// Search for proper CA (non-zero) and CB that both are simm12.
if ((C1 / C0) != 0 && isInt<12>(C1 / C0) && isInt<12>(C1 % C0) &&
!isInt<12>(C0 * (C1 / C0))) {
CA = C1 / C0;
CB = C1 % C0;
} else if ((C1 / C0 + 1) != 0 && isInt<12>(C1 / C0 + 1) &&
isInt<12>(C1 % C0 - C0) && !isInt<12>(C0 * (C1 / C0 + 1))) {
CA = C1 / C0 + 1;
CB = C1 % C0 - C0;
} else if ((C1 / C0 - 1) != 0 && isInt<12>(C1 / C0 - 1) &&
isInt<12>(C1 % C0 + C0) && !isInt<12>(C0 * (C1 / C0 - 1))) {
CA = C1 / C0 - 1;
CB = C1 % C0 + C0;
} else
return SDValue();
// Build new nodes (add (mul (add x, c1/c0), c0), c1%c0).
SDLoc DL(N);
SDValue New0 = DAG.getNode(ISD::ADD, DL, VT, N0->getOperand(0),
DAG.getConstant(CA, DL, VT));
SDValue New1 =
DAG.getNode(ISD::MUL, DL, VT, New0, DAG.getConstant(C0, DL, VT));
return DAG.getNode(ISD::ADD, DL, VT, New1, DAG.getConstant(CB, DL, VT));
}
// add (zext, zext) -> zext (add (zext, zext))
// sub (zext, zext) -> sext (sub (zext, zext))
// mul (zext, zext) -> zext (mul (zext, zext))
// sdiv (zext, zext) -> zext (sdiv (zext, zext))
// udiv (zext, zext) -> zext (udiv (zext, zext))
// srem (zext, zext) -> zext (srem (zext, zext))
// urem (zext, zext) -> zext (urem (zext, zext))
//
// where the sum of the extend widths match, and the the range of the bin op
// fits inside the width of the narrower bin op. (For profitability on rvv, we
// use a power of two for both inner and outer extend.)
static SDValue combineBinOpOfZExt(SDNode *N, SelectionDAG &DAG) {
EVT VT = N->getValueType(0);
if (!VT.isVector() || !DAG.getTargetLoweringInfo().isTypeLegal(VT))
return SDValue();
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
if (N0.getOpcode() != ISD::ZERO_EXTEND || N1.getOpcode() != ISD::ZERO_EXTEND)
return SDValue();
if (!N0.hasOneUse() || !N1.hasOneUse())
return SDValue();
SDValue Src0 = N0.getOperand(0);
SDValue Src1 = N1.getOperand(0);
EVT SrcVT = Src0.getValueType();
if (!DAG.getTargetLoweringInfo().isTypeLegal(SrcVT) ||
SrcVT != Src1.getValueType() || SrcVT.getScalarSizeInBits() < 8 ||
SrcVT.getScalarSizeInBits() >= VT.getScalarSizeInBits() / 2)
return SDValue();
LLVMContext &C = *DAG.getContext();
EVT ElemVT = VT.getVectorElementType().getHalfSizedIntegerVT(C);
EVT NarrowVT = EVT::getVectorVT(C, ElemVT, VT.getVectorElementCount());
Src0 = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(Src0), NarrowVT, Src0);
Src1 = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(Src1), NarrowVT, Src1);
// Src0 and Src1 are zero extended, so they're always positive if signed.
//
// sub can produce a negative from two positive operands, so it needs sign
// extended. Other nodes produce a positive from two positive operands, so
// zero extend instead.
unsigned OuterExtend =
N->getOpcode() == ISD::SUB ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
return DAG.getNode(
OuterExtend, SDLoc(N), VT,
DAG.getNode(N->getOpcode(), SDLoc(N), NarrowVT, Src0, Src1));
}
// Try to turn (add (xor bool, 1) -1) into (neg bool).
static SDValue combineAddOfBooleanXor(SDNode *N, SelectionDAG &DAG) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
EVT VT = N->getValueType(0);
SDLoc DL(N);
// RHS should be -1.
if (!isAllOnesConstant(N1))
return SDValue();
// Look for (xor X, 1).
if (N0.getOpcode() != ISD::XOR || !isOneConstant(N0.getOperand(1)))
return SDValue();
// First xor input should be 0 or 1.
APInt Mask = APInt::getBitsSetFrom(VT.getSizeInBits(), 1);
if (!DAG.MaskedValueIsZero(N0.getOperand(0), Mask))
return SDValue();
// Emit a negate of the setcc.
return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
N0.getOperand(0));
}
static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (SDValue V = combineAddOfBooleanXor(N, DAG))
return V;
if (SDValue V = transformAddImmMulImm(N, DAG, Subtarget))
return V;
if (SDValue V = transformAddShlImm(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpOfExtractToReduceTree(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpOfZExt(N, DAG))
return V;
// fold (add (select lhs, rhs, cc, 0, y), x) ->
// (select lhs, rhs, cc, x, (add x, y))
return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
}
// Try to turn a sub boolean RHS and constant LHS into an addi.
static SDValue combineSubOfBoolean(SDNode *N, SelectionDAG &DAG) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
EVT VT = N->getValueType(0);
SDLoc DL(N);
// Require a constant LHS.
auto *N0C = dyn_cast<ConstantSDNode>(N0);
if (!N0C)
return SDValue();
// All our optimizations involve subtracting 1 from the immediate and forming
// an ADDI. Make sure the new immediate is valid for an ADDI.
APInt ImmValMinus1 = N0C->getAPIntValue() - 1;
if (!ImmValMinus1.isSignedIntN(12))
return SDValue();
SDValue NewLHS;
if (N1.getOpcode() == ISD::SETCC && N1.hasOneUse()) {
// (sub constant, (setcc x, y, eq/neq)) ->
// (add (setcc x, y, neq/eq), constant - 1)
ISD::CondCode CCVal = cast<CondCodeSDNode>(N1.getOperand(2))->get();
EVT SetCCOpVT = N1.getOperand(0).getValueType();
if (!isIntEqualitySetCC(CCVal) || !SetCCOpVT.isInteger())
return SDValue();
CCVal = ISD::getSetCCInverse(CCVal, SetCCOpVT);
NewLHS =
DAG.getSetCC(SDLoc(N1), VT, N1.getOperand(0), N1.getOperand(1), CCVal);
} else if (N1.getOpcode() == ISD::XOR && isOneConstant(N1.getOperand(1)) &&
N1.getOperand(0).getOpcode() == ISD::SETCC) {
// (sub C, (xor (setcc), 1)) -> (add (setcc), C-1).
// Since setcc returns a bool the xor is equivalent to 1-setcc.
NewLHS = N1.getOperand(0);
} else
return SDValue();
SDValue NewRHS = DAG.getConstant(ImmValMinus1, DL, VT);
return DAG.getNode(ISD::ADD, DL, VT, NewLHS, NewRHS);
}
static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (SDValue V = combineSubOfBoolean(N, DAG))
return V;
EVT VT = N->getValueType(0);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
// fold (sub 0, (setcc x, 0, setlt)) -> (sra x, xlen - 1)
if (isNullConstant(N0) && N1.getOpcode() == ISD::SETCC && N1.hasOneUse() &&
isNullConstant(N1.getOperand(1))) {
ISD::CondCode CCVal = cast<CondCodeSDNode>(N1.getOperand(2))->get();
if (CCVal == ISD::SETLT) {
SDLoc DL(N);
unsigned ShAmt = N0.getValueSizeInBits() - 1;
return DAG.getNode(ISD::SRA, DL, VT, N1.getOperand(0),
DAG.getConstant(ShAmt, DL, VT));
}
}
if (SDValue V = combineBinOpOfZExt(N, DAG))
return V;
// fold (sub x, (select lhs, rhs, cc, 0, y)) ->
// (select lhs, rhs, cc, x, (sub x, y))
return combineSelectAndUse(N, N1, N0, DAG, /*AllOnes*/ false, Subtarget);
}
// Apply DeMorgan's law to (and/or (xor X, 1), (xor Y, 1)) if X and Y are 0/1.
// Legalizing setcc can introduce xors like this. Doing this transform reduces
// the number of xors and may allow the xor to fold into a branch condition.
static SDValue combineDeMorganOfBoolean(SDNode *N, SelectionDAG &DAG) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
bool IsAnd = N->getOpcode() == ISD::AND;
if (N0.getOpcode() != ISD::XOR || N1.getOpcode() != ISD::XOR)
return SDValue();
if (!N0.hasOneUse() || !N1.hasOneUse())
return SDValue();
SDValue N01 = N0.getOperand(1);
SDValue N11 = N1.getOperand(1);
// For AND, SimplifyDemandedBits may have turned one of the (xor X, 1) into
// (xor X, -1) based on the upper bits of the other operand being 0. If the
// operation is And, allow one of the Xors to use -1.
if (isOneConstant(N01)) {
if (!isOneConstant(N11) && !(IsAnd && isAllOnesConstant(N11)))
return SDValue();
} else if (isOneConstant(N11)) {
// N01 and N11 being 1 was already handled. Handle N11==1 and N01==-1.
if (!(IsAnd && isAllOnesConstant(N01)))
return SDValue();
} else
return SDValue();
EVT VT = N->getValueType(0);
SDValue N00 = N0.getOperand(0);
SDValue N10 = N1.getOperand(0);
// The LHS of the xors needs to be 0/1.
APInt Mask = APInt::getBitsSetFrom(VT.getSizeInBits(), 1);
if (!DAG.MaskedValueIsZero(N00, Mask) || !DAG.MaskedValueIsZero(N10, Mask))
return SDValue();
// Invert the opcode and insert a new xor.
SDLoc DL(N);
unsigned Opc = IsAnd ? ISD::OR : ISD::AND;
SDValue Logic = DAG.getNode(Opc, DL, VT, N00, N10);
return DAG.getNode(ISD::XOR, DL, VT, Logic, DAG.getConstant(1, DL, VT));
}
static SDValue performTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
// Pre-promote (i1 (truncate (srl X, Y))) on RV64 with Zbs without zero
// extending X. This is safe since we only need the LSB after the shift and
// shift amounts larger than 31 would produce poison. If we wait until
// type legalization, we'll create RISCVISD::SRLW and we can't recover it
// to use a BEXT instruction.
if (!RV64LegalI32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbs() && VT == MVT::i1 &&
N0.getValueType() == MVT::i32 && N0.getOpcode() == ISD::SRL &&
!isa<ConstantSDNode>(N0.getOperand(1)) && N0.hasOneUse()) {
SDLoc DL(N0);
SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
SDValue Srl = DAG.getNode(ISD::SRL, DL, MVT::i64, Op0, Op1);
return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Srl);
}
return SDValue();
}
// Combines two comparison operation and logic operation to one selection
// operation(min, max) and logic operation. Returns new constructed Node if
// conditions for optimization are satisfied.
static SDValue performANDCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
SelectionDAG &DAG = DCI.DAG;
SDValue N0 = N->getOperand(0);
// Pre-promote (i32 (and (srl X, Y), 1)) on RV64 with Zbs without zero
// extending X. This is safe since we only need the LSB after the shift and
// shift amounts larger than 31 would produce poison. If we wait until
// type legalization, we'll create RISCVISD::SRLW and we can't recover it
// to use a BEXT instruction.
if (!RV64LegalI32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbs() &&
N->getValueType(0) == MVT::i32 && isOneConstant(N->getOperand(1)) &&
N0.getOpcode() == ISD::SRL && !isa<ConstantSDNode>(N0.getOperand(1)) &&
N0.hasOneUse()) {
SDLoc DL(N);
SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
SDValue Srl = DAG.getNode(ISD::SRL, DL, MVT::i64, Op0, Op1);
SDValue And = DAG.getNode(ISD::AND, DL, MVT::i64, Srl,
DAG.getConstant(1, DL, MVT::i64));
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, And);
}
if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpOfExtractToReduceTree(N, DAG, Subtarget))
return V;
if (DCI.isAfterLegalizeDAG())
if (SDValue V = combineDeMorganOfBoolean(N, DAG))
return V;
// fold (and (select lhs, rhs, cc, -1, y), x) ->
// (select lhs, rhs, cc, x, (and x, y))
return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ true, Subtarget);
}
// Try to pull an xor with 1 through a select idiom that uses czero_eqz/nez.
// FIXME: Generalize to other binary operators with same operand.
static SDValue combineOrOfCZERO(SDNode *N, SDValue N0, SDValue N1,
SelectionDAG &DAG) {
assert(N->getOpcode() == ISD::OR && "Unexpected opcode");
if (N0.getOpcode() != RISCVISD::CZERO_EQZ ||
N1.getOpcode() != RISCVISD::CZERO_NEZ ||
!N0.hasOneUse() || !N1.hasOneUse())
return SDValue();
// Should have the same condition.
SDValue Cond = N0.getOperand(1);
if (Cond != N1.getOperand(1))
return SDValue();
SDValue TrueV = N0.getOperand(0);
SDValue FalseV = N1.getOperand(0);
if (TrueV.getOpcode() != ISD::XOR || FalseV.getOpcode() != ISD::XOR ||
TrueV.getOperand(1) != FalseV.getOperand(1) ||
!isOneConstant(TrueV.getOperand(1)) ||
!TrueV.hasOneUse() || !FalseV.hasOneUse())
return SDValue();
EVT VT = N->getValueType(0);
SDLoc DL(N);
SDValue NewN0 = DAG.getNode(RISCVISD::CZERO_EQZ, DL, VT, TrueV.getOperand(0),
Cond);
SDValue NewN1 = DAG.getNode(RISCVISD::CZERO_NEZ, DL, VT, FalseV.getOperand(0),
Cond);
SDValue NewOr = DAG.getNode(ISD::OR, DL, VT, NewN0, NewN1);
return DAG.getNode(ISD::XOR, DL, VT, NewOr, TrueV.getOperand(1));
}
static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
SelectionDAG &DAG = DCI.DAG;
if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpOfExtractToReduceTree(N, DAG, Subtarget))
return V;
if (DCI.isAfterLegalizeDAG())
if (SDValue V = combineDeMorganOfBoolean(N, DAG))
return V;
// Look for Or of CZERO_EQZ/NEZ with same condition which is the select idiom.
// We may be able to pull a common operation out of the true and false value.
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
if (SDValue V = combineOrOfCZERO(N, N0, N1, DAG))
return V;
if (SDValue V = combineOrOfCZERO(N, N1, N0, DAG))
return V;
// fold (or (select cond, 0, y), x) ->
// (select cond, x, (or x, y))
return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
}
static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
// Pre-promote (i32 (xor (shl -1, X), ~0)) on RV64 with Zbs so we can use
// (ADDI (BSET X0, X), -1). If we wait until/ type legalization, we'll create
// RISCVISD:::SLLW and we can't recover it to use a BSET instruction.
if (!RV64LegalI32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbs() &&
N->getValueType(0) == MVT::i32 && isAllOnesConstant(N1) &&
N0.getOpcode() == ISD::SHL && isAllOnesConstant(N0.getOperand(0)) &&
!isa<ConstantSDNode>(N0.getOperand(1)) && N0.hasOneUse()) {
SDLoc DL(N);
SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
SDValue Shl = DAG.getNode(ISD::SHL, DL, MVT::i64, Op0, Op1);
SDValue And = DAG.getNOT(DL, Shl, MVT::i64);
return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, And);
}
// fold (xor (sllw 1, x), -1) -> (rolw ~1, x)
// NOTE: Assumes ROL being legal means ROLW is legal.
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
if (N0.getOpcode() == RISCVISD::SLLW &&
isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0)) &&
TLI.isOperationLegal(ISD::ROTL, MVT::i64)) {
SDLoc DL(N);
return DAG.getNode(RISCVISD::ROLW, DL, MVT::i64,
DAG.getConstant(~1, DL, MVT::i64), N0.getOperand(1));
}
// Fold (xor (setcc constant, y, setlt), 1) -> (setcc y, constant + 1, setlt)
if (N0.getOpcode() == ISD::SETCC && isOneConstant(N1) && N0.hasOneUse()) {
auto *ConstN00 = dyn_cast<ConstantSDNode>(N0.getOperand(0));
ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
if (ConstN00 && CC == ISD::SETLT) {
EVT VT = N0.getValueType();
SDLoc DL(N0);
const APInt &Imm = ConstN00->getAPIntValue();
if ((Imm + 1).isSignedIntN(12))
return DAG.getSetCC(DL, VT, N0.getOperand(1),
DAG.getConstant(Imm + 1, DL, VT), CC);
}
}
// Combine (xor (trunc (X cc Y)) 1) -> (trunc (X !cc Y)). This is needed with
// RV64LegalI32 when the setcc is created after type legalization. An i1 xor
// would have been promoted to i32, but the setcc would have i64 result.
if (N->getValueType(0) == MVT::i32 && N0.getOpcode() == ISD::TRUNCATE &&
isOneConstant(N1) && N0.getOperand(0).getOpcode() == ISD::SETCC) {
SDValue N00 = N0.getOperand(0);
SDLoc DL(N);
SDValue LHS = N00.getOperand(0);
SDValue RHS = N00.getOperand(1);
SDValue CC = N00.getOperand(2);
ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
LHS.getValueType());
SDValue Setcc = DAG.getSetCC(SDLoc(N00), N0.getOperand(0).getValueType(),
LHS, RHS, NotCC);
return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N->getValueType(0), Setcc);
}
if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpOfExtractToReduceTree(N, DAG, Subtarget))
return V;
// fold (xor (select cond, 0, y), x) ->
// (select cond, x, (xor x, y))
return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
}
// Try to expand a scalar multiply to a faster sequence.
static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
EVT VT = N->getValueType(0);
// LI + MUL is usually smaller than the alternative sequence.
if (DAG.getMachineFunction().getFunction().hasMinSize())
return SDValue();
if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
return SDValue();
if (VT != Subtarget.getXLenVT())
return SDValue();
if (!Subtarget.hasStdExtZba() && !Subtarget.hasVendorXTHeadBa())
return SDValue();
ConstantSDNode *CNode = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (!CNode)
return SDValue();
uint64_t MulAmt = CNode->getZExtValue();
// 3/5/9 * 2^N -> shXadd (sll X, C), (sll X, C)
// Matched in tablegen, avoid perturbing patterns.
for (uint64_t Divisor : {3, 5, 9})
if (MulAmt % Divisor == 0 && isPowerOf2_64(MulAmt / Divisor))
return SDValue();
// If this is a power 2 + 2/4/8, we can use a shift followed by a single
// shXadd. First check if this a sum of two power of 2s because that's
// easy. Then count how many zeros are up to the first bit.
if (isPowerOf2_64(MulAmt & (MulAmt - 1))) {
unsigned ScaleShift = llvm::countr_zero(MulAmt);
if (ScaleShift >= 1 && ScaleShift < 4) {
unsigned ShiftAmt = Log2_64((MulAmt & (MulAmt - 1)));
SDLoc DL(N);
SDValue Shift1 = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
DAG.getConstant(ShiftAmt, DL, VT));
SDValue Shift2 = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
DAG.getConstant(ScaleShift, DL, VT));
return DAG.getNode(ISD::ADD, DL, VT, Shift1, Shift2);
}
}
// 2^(1,2,3) * 3,5,9 + 1 -> (shXadd (shYadd x, x), x)
// Matched in tablegen, avoid perturbing patterns.
switch (MulAmt) {
case 11:
case 13:
case 19:
case 21:
case 25:
case 27:
case 29:
case 37:
case 41:
case 45:
case 73:
case 91:
return SDValue();
default:
break;
}
// 2^n + 2/4/8 + 1 -> (add (shl X, C1), (shXadd X, X))
if (MulAmt > 2 && isPowerOf2_64((MulAmt - 1) & (MulAmt - 2))) {
unsigned ScaleShift = llvm::countr_zero(MulAmt - 1);
if (ScaleShift >= 1 && ScaleShift < 4) {
unsigned ShiftAmt = Log2_64(((MulAmt - 1) & (MulAmt - 2)));
SDLoc DL(N);
SDValue Shift1 = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
DAG.getConstant(ShiftAmt, DL, VT));
SDValue Shift2 = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
DAG.getConstant(ScaleShift, DL, VT));
return DAG.getNode(
ISD::ADD, DL, VT, Shift1,
DAG.getNode(ISD::ADD, DL, VT, Shift2, N->getOperand(0)));
}
}
return SDValue();
}
static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
EVT VT = N->getValueType(0);
if (!VT.isVector())
return expandMul(N, DAG, DCI, Subtarget);
SDLoc DL(N);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
SDValue MulOper;
unsigned AddSubOpc;
// vmadd: (mul (add x, 1), y) -> (add (mul x, y), y)
// (mul x, add (y, 1)) -> (add x, (mul x, y))
// vnmsub: (mul (sub 1, x), y) -> (sub y, (mul x, y))
// (mul x, (sub 1, y)) -> (sub x, (mul x, y))
auto IsAddSubWith1 = [&](SDValue V) -> bool {
AddSubOpc = V->getOpcode();
if ((AddSubOpc == ISD::ADD || AddSubOpc == ISD::SUB) && V->hasOneUse()) {
SDValue Opnd = V->getOperand(1);
MulOper = V->getOperand(0);
if (AddSubOpc == ISD::SUB)
std::swap(Opnd, MulOper);
if (isOneOrOneSplat(Opnd))
return true;
}
return false;
};
if (IsAddSubWith1(N0)) {
SDValue MulVal = DAG.getNode(ISD::MUL, DL, VT, N1, MulOper);
return DAG.getNode(AddSubOpc, DL, VT, N1, MulVal);
}
if (IsAddSubWith1(N1)) {
SDValue MulVal = DAG.getNode(ISD::MUL, DL, VT, N0, MulOper);
return DAG.getNode(AddSubOpc, DL, VT, N0, MulVal);
}
if (SDValue V = combineBinOpOfZExt(N, DAG))
return V;
return SDValue();
}
/// According to the property that indexed load/store instructions zero-extend
/// their indices, try to narrow the type of index operand.
static bool narrowIndex(SDValue &N, ISD::MemIndexType IndexType, SelectionDAG &DAG) {
if (isIndexTypeSigned(IndexType))
return false;
if (!N->hasOneUse())
return false;
EVT VT = N.getValueType();
SDLoc DL(N);
// In general, what we're doing here is seeing if we can sink a truncate to
// a smaller element type into the expression tree building our index.
// TODO: We can generalize this and handle a bunch more cases if useful.
// Narrow a buildvector to the narrowest element type. This requires less
// work and less register pressure at high LMUL, and creates smaller constants
// which may be cheaper to materialize.
if (ISD::isBuildVectorOfConstantSDNodes(N.getNode())) {
KnownBits Known = DAG.computeKnownBits(N);
unsigned ActiveBits = std::max(8u, Known.countMaxActiveBits());
LLVMContext &C = *DAG.getContext();
EVT ResultVT = EVT::getIntegerVT(C, ActiveBits).getRoundIntegerType(C);
if (ResultVT.bitsLT(VT.getVectorElementType())) {
N = DAG.getNode(ISD::TRUNCATE, DL,
VT.changeVectorElementType(ResultVT), N);
return true;
}
}
// Handle the pattern (shl (zext x to ty), C) and bits(x) + C < bits(ty).
if (N.getOpcode() != ISD::SHL)
return false;
SDValue N0 = N.getOperand(0);
if (N0.getOpcode() != ISD::ZERO_EXTEND &&
N0.getOpcode() != RISCVISD::VZEXT_VL)
return false;
if (!N0->hasOneUse())
return false;
APInt ShAmt;
SDValue N1 = N.getOperand(1);
if (!ISD::isConstantSplatVector(N1.getNode(), ShAmt))
return false;
SDValue Src = N0.getOperand(0);
EVT SrcVT = Src.getValueType();
unsigned SrcElen = SrcVT.getScalarSizeInBits();
unsigned ShAmtV = ShAmt.getZExtValue();
unsigned NewElen = PowerOf2Ceil(SrcElen + ShAmtV);
NewElen = std::max(NewElen, 8U);
// Skip if NewElen is not narrower than the original extended type.
if (NewElen >= N0.getValueType().getScalarSizeInBits())
return false;
EVT NewEltVT = EVT::getIntegerVT(*DAG.getContext(), NewElen);
EVT NewVT = SrcVT.changeVectorElementType(NewEltVT);
SDValue NewExt = DAG.getNode(N0->getOpcode(), DL, NewVT, N0->ops());
SDValue NewShAmtVec = DAG.getConstant(ShAmtV, DL, NewVT);
N = DAG.getNode(ISD::SHL, DL, NewVT, NewExt, NewShAmtVec);
return true;
}
// Replace (seteq (i64 (and X, 0xffffffff)), C1) with
// (seteq (i64 (sext_inreg (X, i32)), C1')) where C1' is C1 sign extended from
// bit 31. Same for setne. C1' may be cheaper to materialize and the sext_inreg
// can become a sext.w instead of a shift pair.
static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
EVT VT = N->getValueType(0);
EVT OpVT = N0.getValueType();
if (OpVT != MVT::i64 || !Subtarget.is64Bit())
return SDValue();
// RHS needs to be a constant.
auto *N1C = dyn_cast<ConstantSDNode>(N1);
if (!N1C)
return SDValue();
// LHS needs to be (and X, 0xffffffff).
if (N0.getOpcode() != ISD::AND || !N0.hasOneUse() ||
!isa<ConstantSDNode>(N0.getOperand(1)) ||
N0.getConstantOperandVal(1) != UINT64_C(0xffffffff))
return SDValue();
// Looking for an equality compare.
ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(2))->get();
if (!isIntEqualitySetCC(Cond))
return SDValue();
// Don't do this if the sign bit is provably zero, it will be turned back into
// an AND.
APInt SignMask = APInt::getOneBitSet(64, 31);
if (DAG.MaskedValueIsZero(N0.getOperand(0), SignMask))
return SDValue();
const APInt &C1 = N1C->getAPIntValue();
SDLoc dl(N);
// If the constant is larger than 2^32 - 1 it is impossible for both sides
// to be equal.
if (C1.getActiveBits() > 32)
return DAG.getBoolConstant(Cond == ISD::SETNE, dl, VT, OpVT);
SDValue SExtOp = DAG.getNode(ISD::SIGN_EXTEND_INREG, N, OpVT,
N0.getOperand(0), DAG.getValueType(MVT::i32));
return DAG.getSetCC(dl, VT, SExtOp, DAG.getConstant(C1.trunc(32).sext(64),
dl, OpVT), Cond);
}
static SDValue
performSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue Src = N->getOperand(0);
EVT VT = N->getValueType(0);
// Fold (sext_inreg (fmv_x_anyexth X), i16) -> (fmv_x_signexth X)
if (Src.getOpcode() == RISCVISD::FMV_X_ANYEXTH &&
cast<VTSDNode>(N->getOperand(1))->getVT().bitsGE(MVT::i16))
return DAG.getNode(RISCVISD::FMV_X_SIGNEXTH, SDLoc(N), VT,
Src.getOperand(0));
return SDValue();
}
namespace {
// Forward declaration of the structure holding the necessary information to
// apply a combine.
struct CombineResult;
enum ExtKind : uint8_t { ZExt = 1 << 0, SExt = 1 << 1, FPExt = 1 << 2 };
/// Helper class for folding sign/zero extensions.
/// In particular, this class is used for the following combines:
/// add | add_vl | or disjoint -> vwadd(u) | vwadd(u)_w
/// sub | sub_vl -> vwsub(u) | vwsub(u)_w
/// mul | mul_vl -> vwmul(u) | vwmul_su
/// shl | shl_vl -> vwsll
/// fadd -> vfwadd | vfwadd_w
/// fsub -> vfwsub | vfwsub_w
/// fmul -> vfwmul
/// An object of this class represents an operand of the operation we want to
/// combine.
/// E.g., when trying to combine `mul_vl a, b`, we will have one instance of
/// NodeExtensionHelper for `a` and one for `b`.
///
/// This class abstracts away how the extension is materialized and
/// how its number of users affect the combines.
///
/// In particular:
/// - VWADD_W is conceptually == add(op0, sext(op1))
/// - VWADDU_W == add(op0, zext(op1))
/// - VWSUB_W == sub(op0, sext(op1))
/// - VWSUBU_W == sub(op0, zext(op1))
/// - VFWADD_W == fadd(op0, fpext(op1))
/// - VFWSUB_W == fsub(op0, fpext(op1))
/// And VMV_V_X_VL, depending on the value, is conceptually equivalent to
/// zext|sext(smaller_value).
struct NodeExtensionHelper {
/// Records if this operand is like being zero extended.
bool SupportsZExt;
/// Records if this operand is like being sign extended.
/// Note: SupportsZExt and SupportsSExt are not mutually exclusive. For
/// instance, a splat constant (e.g., 3), would support being both sign and
/// zero extended.
bool SupportsSExt;
/// Records if this operand is like being floating-Point extended.
bool SupportsFPExt;
/// This boolean captures whether we care if this operand would still be
/// around after the folding happens.
bool EnforceOneUse;
/// Original value that this NodeExtensionHelper represents.
SDValue OrigOperand;
/// Get the value feeding the extension or the value itself.
/// E.g., for zext(a), this would return a.
SDValue getSource() const {
switch (OrigOperand.getOpcode()) {
case ISD::ZERO_EXTEND:
case ISD::SIGN_EXTEND:
case RISCVISD::VSEXT_VL:
case RISCVISD::VZEXT_VL:
case RISCVISD::FP_EXTEND_VL:
return OrigOperand.getOperand(0);
default:
return OrigOperand;
}
}
/// Check if this instance represents a splat.
bool isSplat() const {
return OrigOperand.getOpcode() == RISCVISD::VMV_V_X_VL ||
OrigOperand.getOpcode() == ISD::SPLAT_VECTOR;
}
/// Get the extended opcode.
unsigned getExtOpc(ExtKind SupportsExt) const {
switch (SupportsExt) {
case ExtKind::SExt:
return RISCVISD::VSEXT_VL;
case ExtKind::ZExt:
return RISCVISD::VZEXT_VL;
case ExtKind::FPExt:
return RISCVISD::FP_EXTEND_VL;
}
llvm_unreachable("Unknown ExtKind enum");
}
/// Get or create a value that can feed \p Root with the given extension \p
/// SupportsExt. If \p SExt is std::nullopt, this returns the source of this
/// operand. \see ::getSource().
SDValue getOrCreateExtendedOp(SDNode *Root, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget,
std::optional<ExtKind> SupportsExt) const {
if (!SupportsExt.has_value())
return OrigOperand;
MVT NarrowVT = getNarrowType(Root, *SupportsExt);
SDValue Source = getSource();
assert(Subtarget.getTargetLowering()->isTypeLegal(Source.getValueType()));
if (Source.getValueType() == NarrowVT)
return Source;
unsigned ExtOpc = getExtOpc(*SupportsExt);
// If we need an extension, we should be changing the type.
SDLoc DL(OrigOperand);
auto [Mask, VL] = getMaskAndVL(Root, DAG, Subtarget);
switch (OrigOperand.getOpcode()) {
case ISD::ZERO_EXTEND:
case ISD::SIGN_EXTEND:
case RISCVISD::VSEXT_VL:
case RISCVISD::VZEXT_VL:
case RISCVISD::FP_EXTEND_VL:
return DAG.getNode(ExtOpc, DL, NarrowVT, Source, Mask, VL);
case ISD::SPLAT_VECTOR:
return DAG.getSplat(NarrowVT, DL, Source.getOperand(0));
case RISCVISD::VMV_V_X_VL:
return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, NarrowVT,
DAG.getUNDEF(NarrowVT), Source.getOperand(1), VL);
default:
// Other opcodes can only come from the original LHS of VW(ADD|SUB)_W_VL
// and that operand should already have the right NarrowVT so no
// extension should be required at this point.
llvm_unreachable("Unsupported opcode");
}
}
/// Helper function to get the narrow type for \p Root.
/// The narrow type is the type of \p Root where we divided the size of each
/// element by 2. E.g., if Root's type <2xi16> -> narrow type <2xi8>.
/// \pre Both the narrow type and the original type should be legal.
static MVT getNarrowType(const SDNode *Root, ExtKind SupportsExt) {
MVT VT = Root->getSimpleValueType(0);
// Determine the narrow size.
unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
MVT EltVT = SupportsExt == ExtKind::FPExt
? MVT::getFloatingPointVT(NarrowSize)
: MVT::getIntegerVT(NarrowSize);
assert((int)NarrowSize >= (SupportsExt == ExtKind::FPExt ? 16 : 8) &&
"Trying to extend something we can't represent");
MVT NarrowVT = MVT::getVectorVT(EltVT, VT.getVectorElementCount());
return NarrowVT;
}
/// Get the opcode to materialize:
/// Opcode(sext(a), sext(b)) -> newOpcode(a, b)
static unsigned getSExtOpcode(unsigned Opcode) {
switch (Opcode) {
case ISD::ADD:
case RISCVISD::ADD_VL:
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWADDU_W_VL:
case ISD::OR:
return RISCVISD::VWADD_VL;
case ISD::SUB:
case RISCVISD::SUB_VL:
case RISCVISD::VWSUB_W_VL:
case RISCVISD::VWSUBU_W_VL:
return RISCVISD::VWSUB_VL;
case ISD::MUL:
case RISCVISD::MUL_VL:
return RISCVISD::VWMUL_VL;
default:
llvm_unreachable("Unexpected opcode");
}
}
/// Get the opcode to materialize:
/// Opcode(zext(a), zext(b)) -> newOpcode(a, b)
static unsigned getZExtOpcode(unsigned Opcode) {
switch (Opcode) {
case ISD::ADD:
case RISCVISD::ADD_VL:
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWADDU_W_VL:
case ISD::OR:
return RISCVISD::VWADDU_VL;
case ISD::SUB:
case RISCVISD::SUB_VL:
case RISCVISD::VWSUB_W_VL:
case RISCVISD::VWSUBU_W_VL:
return RISCVISD::VWSUBU_VL;
case ISD::MUL:
case RISCVISD::MUL_VL:
return RISCVISD::VWMULU_VL;
case ISD::SHL:
case RISCVISD::SHL_VL:
return RISCVISD::VWSLL_VL;
default:
llvm_unreachable("Unexpected opcode");
}
}
/// Get the opcode to materialize:
/// Opcode(fpext(a), fpext(b)) -> newOpcode(a, b)
static unsigned getFPExtOpcode(unsigned Opcode) {
switch (Opcode) {
case RISCVISD::FADD_VL:
case RISCVISD::VFWADD_W_VL:
return RISCVISD::VFWADD_VL;
case RISCVISD::FSUB_VL:
case RISCVISD::VFWSUB_W_VL:
return RISCVISD::VFWSUB_VL;
case RISCVISD::FMUL_VL:
return RISCVISD::VFWMUL_VL;
default:
llvm_unreachable("Unexpected opcode");
}
}
/// Get the opcode to materialize \p Opcode(sext(a), zext(b)) ->
/// newOpcode(a, b).
static unsigned getSUOpcode(unsigned Opcode) {
assert((Opcode == RISCVISD::MUL_VL || Opcode == ISD::MUL) &&
"SU is only supported for MUL");
return RISCVISD::VWMULSU_VL;
}
/// Get the opcode to materialize
/// \p Opcode(a, s|z|fpext(b)) -> newOpcode(a, b).
static unsigned getWOpcode(unsigned Opcode, ExtKind SupportsExt) {
switch (Opcode) {
case ISD::ADD:
case RISCVISD::ADD_VL:
case ISD::OR:
return SupportsExt == ExtKind::SExt ? RISCVISD::VWADD_W_VL
: RISCVISD::VWADDU_W_VL;
case ISD::SUB:
case RISCVISD::SUB_VL:
return SupportsExt == ExtKind::SExt ? RISCVISD::VWSUB_W_VL
: RISCVISD::VWSUBU_W_VL;
case RISCVISD::FADD_VL:
return RISCVISD::VFWADD_W_VL;
case RISCVISD::FSUB_VL:
return RISCVISD::VFWSUB_W_VL;
default:
llvm_unreachable("Unexpected opcode");
}
}
using CombineToTry = std::function<std::optional<CombineResult>(
SDNode * /*Root*/, const NodeExtensionHelper & /*LHS*/,
const NodeExtensionHelper & /*RHS*/, SelectionDAG &,
const RISCVSubtarget &)>;
/// Check if this node needs to be fully folded or extended for all users.
bool needToPromoteOtherUsers() const { return EnforceOneUse; }
void fillUpExtensionSupportForSplat(SDNode *Root, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
unsigned Opc = OrigOperand.getOpcode();
MVT VT = OrigOperand.getSimpleValueType();
assert((Opc == ISD::SPLAT_VECTOR || Opc == RISCVISD::VMV_V_X_VL) &&
"Unexpected Opcode");
// The pasthru must be undef for tail agnostic.
if (Opc == RISCVISD::VMV_V_X_VL && !OrigOperand.getOperand(0).isUndef())
return;
// Get the scalar value.
SDValue Op = Opc == ISD::SPLAT_VECTOR ? OrigOperand.getOperand(0)
: OrigOperand.getOperand(1);
// See if we have enough sign bits or zero bits in the scalar to use a
// widening opcode by splatting to smaller element size.
unsigned EltBits = VT.getScalarSizeInBits();
unsigned ScalarBits = Op.getValueSizeInBits();
// Make sure we're getting all element bits from the scalar register.
// FIXME: Support implicit sign extension of vmv.v.x?
if (ScalarBits < EltBits)
return;
unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
// If the narrow type cannot be expressed with a legal VMV,
// this is not a valid candidate.
if (NarrowSize < 8)
return;
if (DAG.ComputeMaxSignificantBits(Op) <= NarrowSize)
SupportsSExt = true;
if (DAG.MaskedValueIsZero(Op,
APInt::getBitsSetFrom(ScalarBits, NarrowSize)))
SupportsZExt = true;
EnforceOneUse = false;
}
/// Helper method to set the various fields of this struct based on the
/// type of \p Root.
void fillUpExtensionSupport(SDNode *Root, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SupportsZExt = false;
SupportsSExt = false;
SupportsFPExt = false;
EnforceOneUse = true;
unsigned Opc = OrigOperand.getOpcode();
// For the nodes we handle below, we end up using their inputs directly: see
// getSource(). However since they either don't have a passthru or we check
// that their passthru is undef, we can safely ignore their mask and VL.
switch (Opc) {
case ISD::ZERO_EXTEND:
case ISD::SIGN_EXTEND: {
MVT VT = OrigOperand.getSimpleValueType();
if (!VT.isVector())
break;
SDValue NarrowElt = OrigOperand.getOperand(0);
MVT NarrowVT = NarrowElt.getSimpleValueType();
// i1 types are legal but we can't select V{S,Z}EXT_VLs with them.
if (NarrowVT.getVectorElementType() == MVT::i1)
break;
SupportsZExt = Opc == ISD::ZERO_EXTEND;
SupportsSExt = Opc == ISD::SIGN_EXTEND;
break;
}
case RISCVISD::VZEXT_VL:
SupportsZExt = true;
break;
case RISCVISD::VSEXT_VL:
SupportsSExt = true;
break;
case RISCVISD::FP_EXTEND_VL:
SupportsFPExt = true;
break;
case ISD::SPLAT_VECTOR:
case RISCVISD::VMV_V_X_VL:
fillUpExtensionSupportForSplat(Root, DAG, Subtarget);
break;
default:
break;
}
}
/// Check if \p Root supports any extension folding combines.
static bool isSupportedRoot(const SDNode *Root,
const RISCVSubtarget &Subtarget) {
switch (Root->getOpcode()) {
case ISD::ADD:
case ISD::SUB:
case ISD::MUL: {
return Root->getValueType(0).isScalableVector();
}
case ISD::OR: {
return Root->getValueType(0).isScalableVector() &&
Root->getFlags().hasDisjoint();
}
// Vector Widening Integer Add/Sub/Mul Instructions
case RISCVISD::ADD_VL:
case RISCVISD::MUL_VL:
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWADDU_W_VL:
case RISCVISD::SUB_VL:
case RISCVISD::VWSUB_W_VL:
case RISCVISD::VWSUBU_W_VL:
// Vector Widening Floating-Point Add/Sub/Mul Instructions
case RISCVISD::FADD_VL:
case RISCVISD::FSUB_VL:
case RISCVISD::FMUL_VL:
case RISCVISD::VFWADD_W_VL:
case RISCVISD::VFWSUB_W_VL:
return true;
case ISD::SHL:
return Root->getValueType(0).isScalableVector() &&
Subtarget.hasStdExtZvbb();
case RISCVISD::SHL_VL:
return Subtarget.hasStdExtZvbb();
default:
return false;
}
}
/// Build a NodeExtensionHelper for \p Root.getOperand(\p OperandIdx).
NodeExtensionHelper(SDNode *Root, unsigned OperandIdx, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(isSupportedRoot(Root, Subtarget) &&
"Trying to build an helper with an "
"unsupported root");
assert(OperandIdx < 2 && "Requesting something else than LHS or RHS");
assert(DAG.getTargetLoweringInfo().isTypeLegal(Root->getValueType(0)));
OrigOperand = Root->getOperand(OperandIdx);
unsigned Opc = Root->getOpcode();
switch (Opc) {
// We consider
// VW<ADD|SUB>_W(LHS, RHS) -> <ADD|SUB>(LHS, SEXT(RHS))
// VW<ADD|SUB>U_W(LHS, RHS) -> <ADD|SUB>(LHS, ZEXT(RHS))
// VFW<ADD|SUB>_W(LHS, RHS) -> F<ADD|SUB>(LHS, FPEXT(RHS))
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWADDU_W_VL:
case RISCVISD::VWSUB_W_VL:
case RISCVISD::VWSUBU_W_VL:
case RISCVISD::VFWADD_W_VL:
case RISCVISD::VFWSUB_W_VL:
if (OperandIdx == 1) {
SupportsZExt =
Opc == RISCVISD::VWADDU_W_VL || Opc == RISCVISD::VWSUBU_W_VL;
SupportsSExt =
Opc == RISCVISD::VWADD_W_VL || Opc == RISCVISD::VWSUB_W_VL;
SupportsFPExt =
Opc == RISCVISD::VFWADD_W_VL || Opc == RISCVISD::VFWSUB_W_VL;
// There's no existing extension here, so we don't have to worry about
// making sure it gets removed.
EnforceOneUse = false;
break;
}
[[fallthrough]];
default:
fillUpExtensionSupport(Root, DAG, Subtarget);
break;
}
}
/// Helper function to get the Mask and VL from \p Root.
static std::pair<SDValue, SDValue>
getMaskAndVL(const SDNode *Root, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(isSupportedRoot(Root, Subtarget) && "Unexpected root");
switch (Root->getOpcode()) {
case ISD::ADD:
case ISD::SUB:
case ISD::MUL:
case ISD::OR:
case ISD::SHL: {
SDLoc DL(Root);
MVT VT = Root->getSimpleValueType(0);
return getDefaultScalableVLOps(VT, DL, DAG, Subtarget);
}
default:
return std::make_pair(Root->getOperand(3), Root->getOperand(4));
}
}
/// Helper function to check if \p N is commutative with respect to the
/// foldings that are supported by this class.
static bool isCommutative(const SDNode *N) {
switch (N->getOpcode()) {
case ISD::ADD:
case ISD::MUL:
case ISD::OR:
case RISCVISD::ADD_VL:
case RISCVISD::MUL_VL:
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWADDU_W_VL:
case RISCVISD::FADD_VL:
case RISCVISD::FMUL_VL:
case RISCVISD::VFWADD_W_VL:
return true;
case ISD::SUB:
case RISCVISD::SUB_VL:
case RISCVISD::VWSUB_W_VL:
case RISCVISD::VWSUBU_W_VL:
case RISCVISD::FSUB_VL:
case RISCVISD::VFWSUB_W_VL:
case ISD::SHL:
case RISCVISD::SHL_VL:
return false;
default:
llvm_unreachable("Unexpected opcode");
}
}
/// Get a list of combine to try for folding extensions in \p Root.
/// Note that each returned CombineToTry function doesn't actually modify
/// anything. Instead they produce an optional CombineResult that if not None,
/// need to be materialized for the combine to be applied.
/// \see CombineResult::materialize.
/// If the related CombineToTry function returns std::nullopt, that means the
/// combine didn't match.
static SmallVector<CombineToTry> getSupportedFoldings(const SDNode *Root);
};
/// Helper structure that holds all the necessary information to materialize a
/// combine that does some extension folding.
struct CombineResult {
/// Opcode to be generated when materializing the combine.
unsigned TargetOpcode;
// No value means no extension is needed.
std::optional<ExtKind> LHSExt;
std::optional<ExtKind> RHSExt;
/// Root of the combine.
SDNode *Root;
/// LHS of the TargetOpcode.
NodeExtensionHelper LHS;
/// RHS of the TargetOpcode.
NodeExtensionHelper RHS;
CombineResult(unsigned TargetOpcode, SDNode *Root,
const NodeExtensionHelper &LHS, std::optional<ExtKind> LHSExt,
const NodeExtensionHelper &RHS, std::optional<ExtKind> RHSExt)
: TargetOpcode(TargetOpcode), LHSExt(LHSExt), RHSExt(RHSExt), Root(Root),
LHS(LHS), RHS(RHS) {}
/// Return a value that uses TargetOpcode and that can be used to replace
/// Root.
/// The actual replacement is *not* done in that method.
SDValue materialize(SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) const {
SDValue Mask, VL, Merge;
std::tie(Mask, VL) =
NodeExtensionHelper::getMaskAndVL(Root, DAG, Subtarget);
switch (Root->getOpcode()) {
default:
Merge = Root->getOperand(2);
break;
case ISD::ADD:
case ISD::SUB:
case ISD::MUL:
case ISD::OR:
case ISD::SHL:
Merge = DAG.getUNDEF(Root->getValueType(0));
break;
}
return DAG.getNode(TargetOpcode, SDLoc(Root), Root->getValueType(0),
LHS.getOrCreateExtendedOp(Root, DAG, Subtarget, LHSExt),
RHS.getOrCreateExtendedOp(Root, DAG, Subtarget, RHSExt),
Merge, Mask, VL);
}
};
/// Check if \p Root follows a pattern Root(ext(LHS), ext(RHS))
/// where `ext` is the same for both LHS and RHS (i.e., both are sext or both
/// are zext) and LHS and RHS can be folded into Root.
/// AllowExtMask define which form `ext` can take in this pattern.
///
/// \note If the pattern can match with both zext and sext, the returned
/// CombineResult will feature the zext result.
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVWWithSameExtensionImpl(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS,
uint8_t AllowExtMask, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if ((AllowExtMask & ExtKind::ZExt) && LHS.SupportsZExt && RHS.SupportsZExt)
return CombineResult(NodeExtensionHelper::getZExtOpcode(Root->getOpcode()),
Root, LHS, /*LHSExt=*/{ExtKind::ZExt}, RHS,
/*RHSExt=*/{ExtKind::ZExt});
if ((AllowExtMask & ExtKind::SExt) && LHS.SupportsSExt && RHS.SupportsSExt)
return CombineResult(NodeExtensionHelper::getSExtOpcode(Root->getOpcode()),
Root, LHS, /*LHSExt=*/{ExtKind::SExt}, RHS,
/*RHSExt=*/{ExtKind::SExt});
if ((AllowExtMask & ExtKind::FPExt) && LHS.SupportsFPExt && RHS.SupportsFPExt)
return CombineResult(NodeExtensionHelper::getFPExtOpcode(Root->getOpcode()),
Root, LHS, /*LHSExt=*/{ExtKind::FPExt}, RHS,
/*RHSExt=*/{ExtKind::FPExt});
return std::nullopt;
}
/// Check if \p Root follows a pattern Root(ext(LHS), ext(RHS))
/// where `ext` is the same for both LHS and RHS (i.e., both are sext or both
/// are zext) and LHS and RHS can be folded into Root.
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVWWithSameExtension(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
return canFoldToVWWithSameExtensionImpl(
Root, LHS, RHS, ExtKind::ZExt | ExtKind::SExt | ExtKind::FPExt, DAG,
Subtarget);
}
/// Check if \p Root follows a pattern Root(LHS, ext(RHS))
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVW_W(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (RHS.SupportsFPExt)
return CombineResult(
NodeExtensionHelper::getWOpcode(Root->getOpcode(), ExtKind::FPExt),
Root, LHS, /*LHSExt=*/std::nullopt, RHS, /*RHSExt=*/{ExtKind::FPExt});
// FIXME: Is it useful to form a vwadd.wx or vwsub.wx if it removes a scalar
// sext/zext?
// Control this behavior behind an option (AllowSplatInVW_W) for testing
// purposes.
if (RHS.SupportsZExt && (!RHS.isSplat() || AllowSplatInVW_W))
return CombineResult(
NodeExtensionHelper::getWOpcode(Root->getOpcode(), ExtKind::ZExt), Root,
LHS, /*LHSExt=*/std::nullopt, RHS, /*RHSExt=*/{ExtKind::ZExt});
if (RHS.SupportsSExt && (!RHS.isSplat() || AllowSplatInVW_W))
return CombineResult(
NodeExtensionHelper::getWOpcode(Root->getOpcode(), ExtKind::SExt), Root,
LHS, /*LHSExt=*/std::nullopt, RHS, /*RHSExt=*/{ExtKind::SExt});
return std::nullopt;
}
/// Check if \p Root follows a pattern Root(sext(LHS), sext(RHS))
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVWWithSEXT(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
return canFoldToVWWithSameExtensionImpl(Root, LHS, RHS, ExtKind::SExt, DAG,
Subtarget);
}
/// Check if \p Root follows a pattern Root(zext(LHS), zext(RHS))
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVWWithZEXT(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
return canFoldToVWWithSameExtensionImpl(Root, LHS, RHS, ExtKind::ZExt, DAG,
Subtarget);
}
/// Check if \p Root follows a pattern Root(fpext(LHS), fpext(RHS))
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVWWithFPEXT(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
return canFoldToVWWithSameExtensionImpl(Root, LHS, RHS, ExtKind::FPExt, DAG,
Subtarget);
}
/// Check if \p Root follows a pattern Root(sext(LHS), zext(RHS))
///
/// \returns std::nullopt if the pattern doesn't match or a CombineResult that
/// can be used to apply the pattern.
static std::optional<CombineResult>
canFoldToVW_SU(SDNode *Root, const NodeExtensionHelper &LHS,
const NodeExtensionHelper &RHS, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (!LHS.SupportsSExt || !RHS.SupportsZExt)
return std::nullopt;
return CombineResult(NodeExtensionHelper::getSUOpcode(Root->getOpcode()),
Root, LHS, /*LHSExt=*/{ExtKind::SExt}, RHS,
/*RHSExt=*/{ExtKind::ZExt});
}
SmallVector<NodeExtensionHelper::CombineToTry>
NodeExtensionHelper::getSupportedFoldings(const SDNode *Root) {
SmallVector<CombineToTry> Strategies;
switch (Root->getOpcode()) {
case ISD::ADD:
case ISD::SUB:
case ISD::OR:
case RISCVISD::ADD_VL:
case RISCVISD::SUB_VL:
case RISCVISD::FADD_VL:
case RISCVISD::FSUB_VL:
// add|sub|fadd|fsub-> vwadd(u)|vwsub(u)|vfwadd|vfwsub
Strategies.push_back(canFoldToVWWithSameExtension);
// add|sub|fadd|fsub -> vwadd(u)_w|vwsub(u)_w}|vfwadd_w|vfwsub_w
Strategies.push_back(canFoldToVW_W);
break;
case RISCVISD::FMUL_VL:
Strategies.push_back(canFoldToVWWithSameExtension);
break;
case ISD::MUL:
case RISCVISD::MUL_VL:
// mul -> vwmul(u)
Strategies.push_back(canFoldToVWWithSameExtension);
// mul -> vwmulsu
Strategies.push_back(canFoldToVW_SU);
break;
case ISD::SHL:
case RISCVISD::SHL_VL:
// shl -> vwsll
Strategies.push_back(canFoldToVWWithZEXT);
break;
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWSUB_W_VL:
// vwadd_w|vwsub_w -> vwadd|vwsub
Strategies.push_back(canFoldToVWWithSEXT);
break;
case RISCVISD::VWADDU_W_VL:
case RISCVISD::VWSUBU_W_VL:
// vwaddu_w|vwsubu_w -> vwaddu|vwsubu
Strategies.push_back(canFoldToVWWithZEXT);
break;
case RISCVISD::VFWADD_W_VL:
case RISCVISD::VFWSUB_W_VL:
// vfwadd_w|vfwsub_w -> vfwadd|vfwsub
Strategies.push_back(canFoldToVWWithFPEXT);
break;
default:
llvm_unreachable("Unexpected opcode");
}
return Strategies;
}
} // End anonymous namespace.
/// Combine a binary operation to its equivalent VW or VW_W form.
/// The supported combines are:
/// add | add_vl | or disjoint -> vwadd(u) | vwadd(u)_w
/// sub | sub_vl -> vwsub(u) | vwsub(u)_w
/// mul | mul_vl -> vwmul(u) | vwmul_su
/// shl | shl_vl -> vwsll
/// fadd_vl -> vfwadd | vfwadd_w
/// fsub_vl -> vfwsub | vfwsub_w
/// fmul_vl -> vfwmul
/// vwadd_w(u) -> vwadd(u)
/// vwsub_w(u) -> vwsub(u)
/// vfwadd_w -> vfwadd
/// vfwsub_w -> vfwsub
static SDValue combineBinOp_VLToVWBinOp_VL(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
SelectionDAG &DAG = DCI.DAG;
if (DCI.isBeforeLegalize())
return SDValue();
if (!NodeExtensionHelper::isSupportedRoot(N, Subtarget))
return SDValue();
SmallVector<SDNode *> Worklist;
SmallSet<SDNode *, 8> Inserted;
Worklist.push_back(N);
Inserted.insert(N);
SmallVector<CombineResult> CombinesToApply;
while (!Worklist.empty()) {
SDNode *Root = Worklist.pop_back_val();
if (!NodeExtensionHelper::isSupportedRoot(Root, Subtarget))
return SDValue();
NodeExtensionHelper LHS(N, 0, DAG, Subtarget);
NodeExtensionHelper RHS(N, 1, DAG, Subtarget);
auto AppendUsersIfNeeded = [&Worklist,
&Inserted](const NodeExtensionHelper &Op) {
if (Op.needToPromoteOtherUsers()) {
for (SDNode *TheUse : Op.OrigOperand->uses()) {
if (Inserted.insert(TheUse).second)
Worklist.push_back(TheUse);
}
}
};
// Control the compile time by limiting the number of node we look at in
// total.
if (Inserted.size() > ExtensionMaxWebSize)
return SDValue();
SmallVector<NodeExtensionHelper::CombineToTry> FoldingStrategies =
NodeExtensionHelper::getSupportedFoldings(N);
assert(!FoldingStrategies.empty() && "Nothing to be folded");
bool Matched = false;
for (int Attempt = 0;
(Attempt != 1 + NodeExtensionHelper::isCommutative(N)) && !Matched;
++Attempt) {
for (NodeExtensionHelper::CombineToTry FoldingStrategy :
FoldingStrategies) {
std::optional<CombineResult> Res =
FoldingStrategy(N, LHS, RHS, DAG, Subtarget);
if (Res) {
Matched = true;
CombinesToApply.push_back(*Res);
// All the inputs that are extended need to be folded, otherwise
// we would be leaving the old input (since it is may still be used),
// and the new one.
if (Res->LHSExt.has_value())
AppendUsersIfNeeded(LHS);
if (Res->RHSExt.has_value())
AppendUsersIfNeeded(RHS);
break;
}
}
std::swap(LHS, RHS);
}
// Right now we do an all or nothing approach.
if (!Matched)
return SDValue();
}
// Store the value for the replacement of the input node separately.
SDValue InputRootReplacement;
// We do the RAUW after we materialize all the combines, because some replaced
// nodes may be feeding some of the yet-to-be-replaced nodes. Put differently,
// some of these nodes may appear in the NodeExtensionHelpers of some of the
// yet-to-be-visited CombinesToApply roots.
SmallVector<std::pair<SDValue, SDValue>> ValuesToReplace;
ValuesToReplace.reserve(CombinesToApply.size());
for (CombineResult Res : CombinesToApply) {
SDValue NewValue = Res.materialize(DAG, Subtarget);
if (!InputRootReplacement) {
assert(Res.Root == N &&
"First element is expected to be the current node");
InputRootReplacement = NewValue;
} else {
ValuesToReplace.emplace_back(SDValue(Res.Root, 0), NewValue);
}
}
for (std::pair<SDValue, SDValue> OldNewValues : ValuesToReplace) {
DAG.ReplaceAllUsesOfValueWith(OldNewValues.first, OldNewValues.second);
DCI.AddToWorklist(OldNewValues.second.getNode());
}
return InputRootReplacement;
}
// Fold (vwadd(u).wv y, (vmerge cond, x, 0)) -> vwadd(u).wv y, x, y, cond
// (vwsub(u).wv y, (vmerge cond, x, 0)) -> vwsub(u).wv y, x, y, cond
// y will be the Passthru and cond will be the Mask.
static SDValue combineVWADDSUBWSelect(SDNode *N, SelectionDAG &DAG) {
unsigned Opc = N->getOpcode();
assert(Opc == RISCVISD::VWADD_W_VL || Opc == RISCVISD::VWADDU_W_VL ||
Opc == RISCVISD::VWSUB_W_VL || Opc == RISCVISD::VWSUBU_W_VL);
SDValue Y = N->getOperand(0);
SDValue MergeOp = N->getOperand(1);
unsigned MergeOpc = MergeOp.getOpcode();
if (MergeOpc != RISCVISD::VMERGE_VL && MergeOpc != ISD::VSELECT)
return SDValue();
SDValue X = MergeOp->getOperand(1);
if (!MergeOp.hasOneUse())
return SDValue();
// Passthru should be undef
SDValue Passthru = N->getOperand(2);
if (!Passthru.isUndef())
return SDValue();
// Mask should be all ones
SDValue Mask = N->getOperand(3);
if (Mask.getOpcode() != RISCVISD::VMSET_VL)
return SDValue();
// False value of MergeOp should be all zeros
SDValue Z = MergeOp->getOperand(2);
if (Z.getOpcode() == ISD::INSERT_SUBVECTOR &&
(isNullOrNullSplat(Z.getOperand(0)) || Z.getOperand(0).isUndef()))
Z = Z.getOperand(1);
if (!ISD::isConstantSplatVectorAllZeros(Z.getNode()))
return SDValue();
return DAG.getNode(Opc, SDLoc(N), N->getValueType(0),
{Y, X, Y, MergeOp->getOperand(0), N->getOperand(4)},
N->getFlags());
}
static SDValue performVWADDSUBW_VLCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
[[maybe_unused]] unsigned Opc = N->getOpcode();
assert(Opc == RISCVISD::VWADD_W_VL || Opc == RISCVISD::VWADDU_W_VL ||
Opc == RISCVISD::VWSUB_W_VL || Opc == RISCVISD::VWSUBU_W_VL);
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
return combineVWADDSUBWSelect(N, DCI.DAG);
}
// Helper function for performMemPairCombine.
// Try to combine the memory loads/stores LSNode1 and LSNode2
// into a single memory pair operation.
static SDValue tryMemPairCombine(SelectionDAG &DAG, LSBaseSDNode *LSNode1,
LSBaseSDNode *LSNode2, SDValue BasePtr,
uint64_t Imm) {
SmallPtrSet<const SDNode *, 32> Visited;
SmallVector<const SDNode *, 8> Worklist = {LSNode1, LSNode2};
if (SDNode::hasPredecessorHelper(LSNode1, Visited, Worklist) ||
SDNode::hasPredecessorHelper(LSNode2, Visited, Worklist))
return SDValue();
MachineFunction &MF = DAG.getMachineFunction();
const RISCVSubtarget &Subtarget = MF.getSubtarget<RISCVSubtarget>();
// The new operation has twice the width.
MVT XLenVT = Subtarget.getXLenVT();
EVT MemVT = LSNode1->getMemoryVT();
EVT NewMemVT = (MemVT == MVT::i32) ? MVT::i64 : MVT::i128;
MachineMemOperand *MMO = LSNode1->getMemOperand();
MachineMemOperand *NewMMO = MF.getMachineMemOperand(
MMO, MMO->getPointerInfo(), MemVT == MVT::i32 ? 8 : 16);
if (LSNode1->getOpcode() == ISD::LOAD) {
auto Ext = cast<LoadSDNode>(LSNode1)->getExtensionType();
unsigned Opcode;
if (MemVT == MVT::i32)
Opcode = (Ext == ISD::ZEXTLOAD) ? RISCVISD::TH_LWUD : RISCVISD::TH_LWD;
else
Opcode = RISCVISD::TH_LDD;
SDValue Res = DAG.getMemIntrinsicNode(
Opcode, SDLoc(LSNode1), DAG.getVTList({XLenVT, XLenVT, MVT::Other}),
{LSNode1->getChain(), BasePtr,
DAG.getConstant(Imm, SDLoc(LSNode1), XLenVT)},
NewMemVT, NewMMO);
SDValue Node1 =
DAG.getMergeValues({Res.getValue(0), Res.getValue(2)}, SDLoc(LSNode1));
SDValue Node2 =
DAG.getMergeValues({Res.getValue(1), Res.getValue(2)}, SDLoc(LSNode2));
DAG.ReplaceAllUsesWith(LSNode2, Node2.getNode());
return Node1;
} else {
unsigned Opcode = (MemVT == MVT::i32) ? RISCVISD::TH_SWD : RISCVISD::TH_SDD;
SDValue Res = DAG.getMemIntrinsicNode(
Opcode, SDLoc(LSNode1), DAG.getVTList(MVT::Other),
{LSNode1->getChain(), LSNode1->getOperand(1), LSNode2->getOperand(1),
BasePtr, DAG.getConstant(Imm, SDLoc(LSNode1), XLenVT)},
NewMemVT, NewMMO);
DAG.ReplaceAllUsesWith(LSNode2, Res.getNode());
return Res;
}
}
// Try to combine two adjacent loads/stores to a single pair instruction from
// the XTHeadMemPair vendor extension.
static SDValue performMemPairCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI) {
SelectionDAG &DAG = DCI.DAG;
MachineFunction &MF = DAG.getMachineFunction();
const RISCVSubtarget &Subtarget = MF.getSubtarget<RISCVSubtarget>();
// Target does not support load/store pair.
if (!Subtarget.hasVendorXTHeadMemPair())
return SDValue();
LSBaseSDNode *LSNode1 = cast<LSBaseSDNode>(N);
EVT MemVT = LSNode1->getMemoryVT();
unsigned OpNum = LSNode1->getOpcode() == ISD::LOAD ? 1 : 2;
// No volatile, indexed or atomic loads/stores.
if (!LSNode1->isSimple() || LSNode1->isIndexed())
return SDValue();
// Function to get a base + constant representation from a memory value.
auto ExtractBaseAndOffset = [](SDValue Ptr) -> std::pair<SDValue, uint64_t> {
if (Ptr->getOpcode() == ISD::ADD)
if (auto *C1 = dyn_cast<ConstantSDNode>(Ptr->getOperand(1)))
return {Ptr->getOperand(0), C1->getZExtValue()};
return {Ptr, 0};
};
auto [Base1, Offset1] = ExtractBaseAndOffset(LSNode1->getOperand(OpNum));
SDValue Chain = N->getOperand(0);
for (SDNode::use_iterator UI = Chain->use_begin(), UE = Chain->use_end();
UI != UE; ++UI) {
SDUse &Use = UI.getUse();
if (Use.getUser() != N && Use.getResNo() == 0 &&
Use.getUser()->getOpcode() == N->getOpcode()) {
LSBaseSDNode *LSNode2 = cast<LSBaseSDNode>(Use.getUser());
// No volatile, indexed or atomic loads/stores.
if (!LSNode2->isSimple() || LSNode2->isIndexed())
continue;
// Check if LSNode1 and LSNode2 have the same type and extension.
if (LSNode1->getOpcode() == ISD::LOAD)
if (cast<LoadSDNode>(LSNode2)->getExtensionType() !=
cast<LoadSDNode>(LSNode1)->getExtensionType())
continue;
if (LSNode1->getMemoryVT() != LSNode2->getMemoryVT())
continue;
auto [Base2, Offset2] = ExtractBaseAndOffset(LSNode2->getOperand(OpNum));
// Check if the base pointer is the same for both instruction.
if (Base1 != Base2)
continue;
// Check if the offsets match the XTHeadMemPair encoding contraints.
bool Valid = false;
if (MemVT == MVT::i32) {
// Check for adjacent i32 values and a 2-bit index.
if ((Offset1 + 4 == Offset2) && isShiftedUInt<2, 3>(Offset1))
Valid = true;
} else if (MemVT == MVT::i64) {
// Check for adjacent i64 values and a 2-bit index.
if ((Offset1 + 8 == Offset2) && isShiftedUInt<2, 4>(Offset1))
Valid = true;
}
if (!Valid)
continue;
// Try to combine.
if (SDValue Res =
tryMemPairCombine(DAG, LSNode1, LSNode2, Base1, Offset1))
return Res;
}
}
return SDValue();
}
// Fold
// (fp_to_int (froundeven X)) -> fcvt X, rne
// (fp_to_int (ftrunc X)) -> fcvt X, rtz
// (fp_to_int (ffloor X)) -> fcvt X, rdn
// (fp_to_int (fceil X)) -> fcvt X, rup
// (fp_to_int (fround X)) -> fcvt X, rmm
// (fp_to_int (frint X)) -> fcvt X
static SDValue performFP_TO_INTCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
SelectionDAG &DAG = DCI.DAG;
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
MVT XLenVT = Subtarget.getXLenVT();
SDValue Src = N->getOperand(0);
// Don't do this for strict-fp Src.
if (Src->isStrictFPOpcode() || Src->isTargetStrictFPOpcode())
return SDValue();
// Ensure the FP type is legal.
if (!TLI.isTypeLegal(Src.getValueType()))
return SDValue();
// Don't do this for f16 with Zfhmin and not Zfh.
if (Src.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfh())
return SDValue();
RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Src.getOpcode());
// If the result is invalid, we didn't find a foldable instruction.
if (FRM == RISCVFPRndMode::Invalid)
return SDValue();
SDLoc DL(N);
bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
EVT VT = N->getValueType(0);
if (VT.isVector() && TLI.isTypeLegal(VT)) {
MVT SrcVT = Src.getSimpleValueType();
MVT SrcContainerVT = SrcVT;
MVT ContainerVT = VT.getSimpleVT();
SDValue XVal = Src.getOperand(0);
// For widening and narrowing conversions we just combine it into a
// VFCVT_..._VL node, as there are no specific VFWCVT/VFNCVT VL nodes. They
// end up getting lowered to their appropriate pseudo instructions based on
// their operand types
if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits() * 2 ||
VT.getScalarSizeInBits() * 2 < SrcVT.getScalarSizeInBits())
return SDValue();
// Make fixed-length vectors scalable first
if (SrcVT.isFixedLengthVector()) {
SrcContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
XVal = convertToScalableVector(SrcContainerVT, XVal, DAG, Subtarget);
ContainerVT =
getContainerForFixedLengthVector(DAG, ContainerVT, Subtarget);
}
auto [Mask, VL] =
getDefaultVLOps(SrcVT, SrcContainerVT, DL, DAG, Subtarget);
SDValue FpToInt;
if (FRM == RISCVFPRndMode::RTZ) {
// Use the dedicated trunc static rounding mode if we're truncating so we
// don't need to generate calls to fsrmi/fsrm
unsigned Opc =
IsSigned ? RISCVISD::VFCVT_RTZ_X_F_VL : RISCVISD::VFCVT_RTZ_XU_F_VL;
FpToInt = DAG.getNode(Opc, DL, ContainerVT, XVal, Mask, VL);
} else if (FRM == RISCVFPRndMode::DYN) {
unsigned Opc =
IsSigned ? RISCVISD::VFCVT_X_F_VL : RISCVISD::VFCVT_XU_F_VL;
FpToInt = DAG.getNode(Opc, DL, ContainerVT, XVal, Mask, VL);
} else {
unsigned Opc =
IsSigned ? RISCVISD::VFCVT_RM_X_F_VL : RISCVISD::VFCVT_RM_XU_F_VL;
FpToInt = DAG.getNode(Opc, DL, ContainerVT, XVal, Mask,
DAG.getTargetConstant(FRM, DL, XLenVT), VL);
}
// If converted from fixed-length to scalable, convert back
if (VT.isFixedLengthVector())
FpToInt = convertFromScalableVector(VT, FpToInt, DAG, Subtarget);
return FpToInt;
}
// Only handle XLen or i32 types. Other types narrower than XLen will
// eventually be legalized to XLenVT.
if (VT != MVT::i32 && VT != XLenVT)
return SDValue();
unsigned Opc;
if (VT == XLenVT)
Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
else
Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
SDValue FpToInt = DAG.getNode(Opc, DL, XLenVT, Src.getOperand(0),
DAG.getTargetConstant(FRM, DL, XLenVT));
return DAG.getNode(ISD::TRUNCATE, DL, VT, FpToInt);
}
// Fold
// (fp_to_int_sat (froundeven X)) -> (select X == nan, 0, (fcvt X, rne))
// (fp_to_int_sat (ftrunc X)) -> (select X == nan, 0, (fcvt X, rtz))
// (fp_to_int_sat (ffloor X)) -> (select X == nan, 0, (fcvt X, rdn))
// (fp_to_int_sat (fceil X)) -> (select X == nan, 0, (fcvt X, rup))
// (fp_to_int_sat (fround X)) -> (select X == nan, 0, (fcvt X, rmm))
// (fp_to_int_sat (frint X)) -> (select X == nan, 0, (fcvt X, dyn))
static SDValue performFP_TO_INT_SATCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const RISCVSubtarget &Subtarget) {
SelectionDAG &DAG = DCI.DAG;
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
MVT XLenVT = Subtarget.getXLenVT();
// Only handle XLen types. Other types narrower than XLen will eventually be
// legalized to XLenVT.
EVT DstVT = N->getValueType(0);
if (DstVT != XLenVT)
return SDValue();
SDValue Src = N->getOperand(0);
// Don't do this for strict-fp Src.
if (Src->isStrictFPOpcode() || Src->isTargetStrictFPOpcode())
return SDValue();
// Ensure the FP type is also legal.
if (!TLI.isTypeLegal(Src.getValueType()))
return SDValue();
// Don't do this for f16 with Zfhmin and not Zfh.
if (Src.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfh())
return SDValue();
EVT SatVT = cast<VTSDNode>(N->getOperand(1))->getVT();
RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Src.getOpcode());
if (FRM == RISCVFPRndMode::Invalid)
return SDValue();
bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT_SAT;
unsigned Opc;
if (SatVT == DstVT)
Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
else if (DstVT == MVT::i64 && SatVT == MVT::i32)
Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
else
return SDValue();
// FIXME: Support other SatVTs by clamping before or after the conversion.
Src = Src.getOperand(0);
SDLoc DL(N);
SDValue FpToInt = DAG.getNode(Opc, DL, XLenVT, Src,
DAG.getTargetConstant(FRM, DL, XLenVT));
// fcvt.wu.* sign extends bit 31 on RV64. FP_TO_UINT_SAT expects to zero
// extend.
if (Opc == RISCVISD::FCVT_WU_RV64)
FpToInt = DAG.getZeroExtendInReg(FpToInt, DL, MVT::i32);
// RISC-V FP-to-int conversions saturate to the destination register size, but
// don't produce 0 for nan.
SDValue ZeroInt = DAG.getConstant(0, DL, DstVT);
return DAG.getSelectCC(DL, Src, Src, ZeroInt, FpToInt, ISD::CondCode::SETUO);
}
// Combine (bitreverse (bswap X)) to the BREV8 GREVI encoding if the type is
// smaller than XLenVT.
static SDValue performBITREVERSECombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(Subtarget.hasStdExtZbkb() && "Unexpected extension");
SDValue Src = N->getOperand(0);
if (Src.getOpcode() != ISD::BSWAP)
return SDValue();
EVT VT = N->getValueType(0);
if (!VT.isScalarInteger() || VT.getSizeInBits() >= Subtarget.getXLen() ||
!llvm::has_single_bit<uint32_t>(VT.getSizeInBits()))
return SDValue();
SDLoc DL(N);
return DAG.getNode(RISCVISD::BREV8, DL, VT, Src.getOperand(0));
}
// Convert from one FMA opcode to another based on whether we are negating the
// multiply result and/or the accumulator.
// NOTE: Only supports RVV operations with VL.
static unsigned negateFMAOpcode(unsigned Opcode, bool NegMul, bool NegAcc) {
// Negating the multiply result changes ADD<->SUB and toggles 'N'.
if (NegMul) {
// clang-format off
switch (Opcode) {
default: llvm_unreachable("Unexpected opcode");
case RISCVISD::VFMADD_VL: Opcode = RISCVISD::VFNMSUB_VL; break;
case RISCVISD::VFNMSUB_VL: Opcode = RISCVISD::VFMADD_VL; break;
case RISCVISD::VFNMADD_VL: Opcode = RISCVISD::VFMSUB_VL; break;
case RISCVISD::VFMSUB_VL: Opcode = RISCVISD::VFNMADD_VL; break;
case RISCVISD::STRICT_VFMADD_VL: Opcode = RISCVISD::STRICT_VFNMSUB_VL; break;
case RISCVISD::STRICT_VFNMSUB_VL: Opcode = RISCVISD::STRICT_VFMADD_VL; break;
case RISCVISD::STRICT_VFNMADD_VL: Opcode = RISCVISD::STRICT_VFMSUB_VL; break;
case RISCVISD::STRICT_VFMSUB_VL: Opcode = RISCVISD::STRICT_VFNMADD_VL; break;
}
// clang-format on
}
// Negating the accumulator changes ADD<->SUB.
if (NegAcc) {
// clang-format off
switch (Opcode) {
default: llvm_unreachable("Unexpected opcode");
case RISCVISD::VFMADD_VL: Opcode = RISCVISD::VFMSUB_VL; break;
case RISCVISD::VFMSUB_VL: Opcode = RISCVISD::VFMADD_VL; break;
case RISCVISD::VFNMADD_VL: Opcode = RISCVISD::VFNMSUB_VL; break;
case RISCVISD::VFNMSUB_VL: Opcode = RISCVISD::VFNMADD_VL; break;
case RISCVISD::STRICT_VFMADD_VL: Opcode = RISCVISD::STRICT_VFMSUB_VL; break;
case RISCVISD::STRICT_VFMSUB_VL: Opcode = RISCVISD::STRICT_VFMADD_VL; break;
case RISCVISD::STRICT_VFNMADD_VL: Opcode = RISCVISD::STRICT_VFNMSUB_VL; break;
case RISCVISD::STRICT_VFNMSUB_VL: Opcode = RISCVISD::STRICT_VFNMADD_VL; break;
}
// clang-format on
}
return Opcode;
}
static SDValue combineVFMADD_VLWithVFNEG_VL(SDNode *N, SelectionDAG &DAG) {
// Fold FNEG_VL into FMA opcodes.
// The first operand of strict-fp is chain.
unsigned Offset = N->isTargetStrictFPOpcode();
SDValue A = N->getOperand(0 + Offset);
SDValue B = N->getOperand(1 + Offset);
SDValue C = N->getOperand(2 + Offset);
SDValue Mask = N->getOperand(3 + Offset);
SDValue VL = N->getOperand(4 + Offset);
auto invertIfNegative = [&Mask, &VL](SDValue &V) {
if (V.getOpcode() == RISCVISD::FNEG_VL && V.getOperand(1) == Mask &&
V.getOperand(2) == VL) {
// Return the negated input.
V = V.getOperand(0);
return true;
}
return false;
};
bool NegA = invertIfNegative(A);
bool NegB = invertIfNegative(B);
bool NegC = invertIfNegative(C);
// If no operands are negated, we're done.
if (!NegA && !NegB && !NegC)
return SDValue();
unsigned NewOpcode = negateFMAOpcode(N->getOpcode(), NegA != NegB, NegC);
if (N->isTargetStrictFPOpcode())
return DAG.getNode(NewOpcode, SDLoc(N), N->getVTList(),
{N->getOperand(0), A, B, C, Mask, VL});
return DAG.getNode(NewOpcode, SDLoc(N), N->getValueType(0), A, B, C, Mask,
VL);
}
static SDValue performVFMADD_VLCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (SDValue V = combineVFMADD_VLWithVFNEG_VL(N, DAG))
return V;
if (N->getValueType(0).isScalableVector() &&
N->getValueType(0).getVectorElementType() == MVT::f32 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16())) {
return SDValue();
}
// FIXME: Ignore strict opcodes for now.
if (N->isTargetStrictFPOpcode())
return SDValue();
// Try to form widening FMA.
SDValue Op0 = N->getOperand(0);
SDValue Op1 = N->getOperand(1);
SDValue Mask = N->getOperand(3);
SDValue VL = N->getOperand(4);
if (Op0.getOpcode() != RISCVISD::FP_EXTEND_VL ||
Op1.getOpcode() != RISCVISD::FP_EXTEND_VL)
return SDValue();
// TODO: Refactor to handle more complex cases similar to
// combineBinOp_VLToVWBinOp_VL.
if ((!Op0.hasOneUse() || !Op1.hasOneUse()) &&
(Op0 != Op1 || !Op0->hasNUsesOfValue(2, 0)))
return SDValue();
// Check the mask and VL are the same.
if (Op0.getOperand(1) != Mask || Op0.getOperand(2) != VL ||
Op1.getOperand(1) != Mask || Op1.getOperand(2) != VL)
return SDValue();
unsigned NewOpc;
switch (N->getOpcode()) {
default:
llvm_unreachable("Unexpected opcode");
case RISCVISD::VFMADD_VL:
NewOpc = RISCVISD::VFWMADD_VL;
break;
case RISCVISD::VFNMSUB_VL:
NewOpc = RISCVISD::VFWNMSUB_VL;
break;
case RISCVISD::VFNMADD_VL:
NewOpc = RISCVISD::VFWNMADD_VL;
break;
case RISCVISD::VFMSUB_VL:
NewOpc = RISCVISD::VFWMSUB_VL;
break;
}
Op0 = Op0.getOperand(0);
Op1 = Op1.getOperand(0);
return DAG.getNode(NewOpc, SDLoc(N), N->getValueType(0), Op0, Op1,
N->getOperand(2), Mask, VL);
}
static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(N->getOpcode() == ISD::SRA && "Unexpected opcode");
if (N->getValueType(0) != MVT::i64 || !Subtarget.is64Bit())
return SDValue();
if (!isa<ConstantSDNode>(N->getOperand(1)))
return SDValue();
uint64_t ShAmt = N->getConstantOperandVal(1);
if (ShAmt > 32)
return SDValue();
SDValue N0 = N->getOperand(0);
// Combine (sra (sext_inreg (shl X, C1), i32), C2) ->
// (sra (shl X, C1+32), C2+32) so it gets selected as SLLI+SRAI instead of
// SLLIW+SRAIW. SLLI+SRAI have compressed forms.
if (ShAmt < 32 &&
N0.getOpcode() == ISD::SIGN_EXTEND_INREG && N0.hasOneUse() &&
cast<VTSDNode>(N0.getOperand(1))->getVT() == MVT::i32 &&
N0.getOperand(0).getOpcode() == ISD::SHL && N0.getOperand(0).hasOneUse() &&
isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
uint64_t LShAmt = N0.getOperand(0).getConstantOperandVal(1);
if (LShAmt < 32) {
SDLoc ShlDL(N0.getOperand(0));
SDValue Shl = DAG.getNode(ISD::SHL, ShlDL, MVT::i64,
N0.getOperand(0).getOperand(0),
DAG.getConstant(LShAmt + 32, ShlDL, MVT::i64));
SDLoc DL(N);
return DAG.getNode(ISD::SRA, DL, MVT::i64, Shl,
DAG.getConstant(ShAmt + 32, DL, MVT::i64));
}
}
// Combine (sra (shl X, 32), 32 - C) -> (shl (sext_inreg X, i32), C)
// FIXME: Should this be a generic combine? There's a similar combine on X86.
//
// Also try these folds where an add or sub is in the middle.
// (sra (add (shl X, 32), C1), 32 - C) -> (shl (sext_inreg (add X, C1), C)
// (sra (sub C1, (shl X, 32)), 32 - C) -> (shl (sext_inreg (sub C1, X), C)
SDValue Shl;
ConstantSDNode *AddC = nullptr;
// We might have an ADD or SUB between the SRA and SHL.
bool IsAdd = N0.getOpcode() == ISD::ADD;
if ((IsAdd || N0.getOpcode() == ISD::SUB)) {
// Other operand needs to be a constant we can modify.
AddC = dyn_cast<ConstantSDNode>(N0.getOperand(IsAdd ? 1 : 0));
if (!AddC)
return SDValue();
// AddC needs to have at least 32 trailing zeros.
if (AddC->getAPIntValue().countr_zero() < 32)
return SDValue();
// All users should be a shift by constant less than or equal to 32. This
// ensures we'll do this optimization for each of them to produce an
// add/sub+sext_inreg they can all share.
for (SDNode *U : N0->uses()) {
if (U->getOpcode() != ISD::SRA ||
!isa<ConstantSDNode>(U->getOperand(1)) ||
U->getConstantOperandVal(1) > 32)
return SDValue();
}
Shl = N0.getOperand(IsAdd ? 0 : 1);
} else {
// Not an ADD or SUB.
Shl = N0;
}
// Look for a shift left by 32.
if (Shl.getOpcode() != ISD::SHL || !isa<ConstantSDNode>(Shl.getOperand(1)) ||
Shl.getConstantOperandVal(1) != 32)
return SDValue();
// We if we didn't look through an add/sub, then the shl should have one use.
// If we did look through an add/sub, the sext_inreg we create is free so
// we're only creating 2 new instructions. It's enough to only remove the
// original sra+add/sub.
if (!AddC && !Shl.hasOneUse())
return SDValue();
SDLoc DL(N);
SDValue In = Shl.getOperand(0);
// If we looked through an ADD or SUB, we need to rebuild it with the shifted
// constant.
if (AddC) {
SDValue ShiftedAddC =
DAG.getConstant(AddC->getAPIntValue().lshr(32), DL, MVT::i64);
if (IsAdd)
In = DAG.getNode(ISD::ADD, DL, MVT::i64, In, ShiftedAddC);
else
In = DAG.getNode(ISD::SUB, DL, MVT::i64, ShiftedAddC, In);
}
SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, In,
DAG.getValueType(MVT::i32));
if (ShAmt == 32)
return SExt;
return DAG.getNode(
ISD::SHL, DL, MVT::i64, SExt,
DAG.getConstant(32 - ShAmt, DL, MVT::i64));
}
// Invert (and/or (set cc X, Y), (xor Z, 1)) to (or/and (set !cc X, Y)), Z) if
// the result is used as the conditon of a br_cc or select_cc we can invert,
// inverting the setcc is free, and Z is 0/1. Caller will invert the
// br_cc/select_cc.
static SDValue tryDemorganOfBooleanCondition(SDValue Cond, SelectionDAG &DAG) {
bool IsAnd = Cond.getOpcode() == ISD::AND;
if (!IsAnd && Cond.getOpcode() != ISD::OR)
return SDValue();
if (!Cond.hasOneUse())
return SDValue();
SDValue Setcc = Cond.getOperand(0);
SDValue Xor = Cond.getOperand(1);
// Canonicalize setcc to LHS.
if (Setcc.getOpcode() != ISD::SETCC)
std::swap(Setcc, Xor);
// LHS should be a setcc and RHS should be an xor.
if (Setcc.getOpcode() != ISD::SETCC || !Setcc.hasOneUse() ||
Xor.getOpcode() != ISD::XOR || !Xor.hasOneUse())
return SDValue();
// If the condition is an And, SimplifyDemandedBits may have changed
// (xor Z, 1) to (not Z).
SDValue Xor1 = Xor.getOperand(1);
if (!isOneConstant(Xor1) && !(IsAnd && isAllOnesConstant(Xor1)))
return SDValue();
EVT VT = Cond.getValueType();
SDValue Xor0 = Xor.getOperand(0);
// The LHS of the xor needs to be 0/1.
APInt Mask = APInt::getBitsSetFrom(VT.getSizeInBits(), 1);
if (!DAG.MaskedValueIsZero(Xor0, Mask))
return SDValue();
// We can only invert integer setccs.
EVT SetCCOpVT = Setcc.getOperand(0).getValueType();
if (!SetCCOpVT.isScalarInteger())
return SDValue();
ISD::CondCode CCVal = cast<CondCodeSDNode>(Setcc.getOperand(2))->get();
if (ISD::isIntEqualitySetCC(CCVal)) {
CCVal = ISD::getSetCCInverse(CCVal, SetCCOpVT);
Setcc = DAG.getSetCC(SDLoc(Setcc), VT, Setcc.getOperand(0),
Setcc.getOperand(1), CCVal);
} else if (CCVal == ISD::SETLT && isNullConstant(Setcc.getOperand(0))) {
// Invert (setlt 0, X) by converting to (setlt X, 1).
Setcc = DAG.getSetCC(SDLoc(Setcc), VT, Setcc.getOperand(1),
DAG.getConstant(1, SDLoc(Setcc), VT), CCVal);
} else if (CCVal == ISD::SETLT && isOneConstant(Setcc.getOperand(1))) {
// (setlt X, 1) by converting to (setlt 0, X).
Setcc = DAG.getSetCC(SDLoc(Setcc), VT,
DAG.getConstant(0, SDLoc(Setcc), VT),
Setcc.getOperand(0), CCVal);
} else
return SDValue();
unsigned Opc = IsAnd ? ISD::OR : ISD::AND;
return DAG.getNode(Opc, SDLoc(Cond), VT, Setcc, Xor.getOperand(0));
}
// Perform common combines for BR_CC and SELECT_CC condtions.
static bool combine_CC(SDValue &LHS, SDValue &RHS, SDValue &CC, const SDLoc &DL,
SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
// As far as arithmetic right shift always saves the sign,
// shift can be omitted.
// Fold setlt (sra X, N), 0 -> setlt X, 0 and
// setge (sra X, N), 0 -> setge X, 0
if (isNullConstant(RHS) && (CCVal == ISD::SETGE || CCVal == ISD::SETLT) &&
LHS.getOpcode() == ISD::SRA) {
LHS = LHS.getOperand(0);
return true;
}
if (!ISD::isIntEqualitySetCC(CCVal))
return false;
// Fold ((setlt X, Y), 0, ne) -> (X, Y, lt)
// Sometimes the setcc is introduced after br_cc/select_cc has been formed.
if (LHS.getOpcode() == ISD::SETCC && isNullConstant(RHS) &&
LHS.getOperand(0).getValueType() == Subtarget.getXLenVT()) {
// If we're looking for eq 0 instead of ne 0, we need to invert the
// condition.
bool Invert = CCVal == ISD::SETEQ;
CCVal = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
if (Invert)
CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
RHS = LHS.getOperand(1);
LHS = LHS.getOperand(0);
translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
CC = DAG.getCondCode(CCVal);
return true;
}
// Fold ((xor X, Y), 0, eq/ne) -> (X, Y, eq/ne)
if (LHS.getOpcode() == ISD::XOR && isNullConstant(RHS)) {
RHS = LHS.getOperand(1);
LHS = LHS.getOperand(0);
return true;
}
// Fold ((srl (and X, 1<<C), C), 0, eq/ne) -> ((shl X, XLen-1-C), 0, ge/lt)
if (isNullConstant(RHS) && LHS.getOpcode() == ISD::SRL && LHS.hasOneUse() &&
LHS.getOperand(1).getOpcode() == ISD::Constant) {
SDValue LHS0 = LHS.getOperand(0);
if (LHS0.getOpcode() == ISD::AND &&
LHS0.getOperand(1).getOpcode() == ISD::Constant) {
uint64_t Mask = LHS0.getConstantOperandVal(1);
uint64_t ShAmt = LHS.getConstantOperandVal(1);
if (isPowerOf2_64(Mask) && Log2_64(Mask) == ShAmt) {
CCVal = CCVal == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
CC = DAG.getCondCode(CCVal);
ShAmt = LHS.getValueSizeInBits() - 1 - ShAmt;
LHS = LHS0.getOperand(0);
if (ShAmt != 0)
LHS =
DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS0.getOperand(0),
DAG.getConstant(ShAmt, DL, LHS.getValueType()));
return true;
}
}
}
// (X, 1, setne) -> // (X, 0, seteq) if we can prove X is 0/1.
// This can occur when legalizing some floating point comparisons.
APInt Mask = APInt::getBitsSetFrom(LHS.getValueSizeInBits(), 1);
if (isOneConstant(RHS) && DAG.MaskedValueIsZero(LHS, Mask)) {
CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
CC = DAG.getCondCode(CCVal);
RHS = DAG.getConstant(0, DL, LHS.getValueType());
return true;
}
if (isNullConstant(RHS)) {
if (SDValue NewCond = tryDemorganOfBooleanCondition(LHS, DAG)) {
CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
CC = DAG.getCondCode(CCVal);
LHS = NewCond;
return true;
}
}
return false;
}
// Fold
// (select C, (add Y, X), Y) -> (add Y, (select C, X, 0)).
// (select C, (sub Y, X), Y) -> (sub Y, (select C, X, 0)).
// (select C, (or Y, X), Y) -> (or Y, (select C, X, 0)).
// (select C, (xor Y, X), Y) -> (xor Y, (select C, X, 0)).
static SDValue tryFoldSelectIntoOp(SDNode *N, SelectionDAG &DAG,
SDValue TrueVal, SDValue FalseVal,
bool Swapped) {
bool Commutative = true;
unsigned Opc = TrueVal.getOpcode();
switch (Opc) {
default:
return SDValue();
case ISD::SHL:
case ISD::SRA:
case ISD::SRL:
case ISD::SUB:
Commutative = false;
break;
case ISD::ADD:
case ISD::OR:
case ISD::XOR:
break;
}
if (!TrueVal.hasOneUse() || isa<ConstantSDNode>(FalseVal))
return SDValue();
unsigned OpToFold;
if (FalseVal == TrueVal.getOperand(0))
OpToFold = 0;
else if (Commutative && FalseVal == TrueVal.getOperand(1))
OpToFold = 1;
else
return SDValue();
EVT VT = N->getValueType(0);
SDLoc DL(N);
SDValue OtherOp = TrueVal.getOperand(1 - OpToFold);
EVT OtherOpVT = OtherOp->getValueType(0);
SDValue IdentityOperand =
DAG.getNeutralElement(Opc, DL, OtherOpVT, N->getFlags());
if (!Commutative)
IdentityOperand = DAG.getConstant(0, DL, OtherOpVT);
assert(IdentityOperand && "No identity operand!");
if (Swapped)
std::swap(OtherOp, IdentityOperand);
SDValue NewSel =
DAG.getSelect(DL, OtherOpVT, N->getOperand(0), OtherOp, IdentityOperand);
return DAG.getNode(TrueVal.getOpcode(), DL, VT, FalseVal, NewSel);
}
// This tries to get rid of `select` and `icmp` that are being used to handle
// `Targets` that do not support `cttz(0)`/`ctlz(0)`.
static SDValue foldSelectOfCTTZOrCTLZ(SDNode *N, SelectionDAG &DAG) {
SDValue Cond = N->getOperand(0);
// This represents either CTTZ or CTLZ instruction.
SDValue CountZeroes;
SDValue ValOnZero;
if (Cond.getOpcode() != ISD::SETCC)
return SDValue();
if (!isNullConstant(Cond->getOperand(1)))
return SDValue();
ISD::CondCode CCVal = cast<CondCodeSDNode>(Cond->getOperand(2))->get();
if (CCVal == ISD::CondCode::SETEQ) {
CountZeroes = N->getOperand(2);
ValOnZero = N->getOperand(1);
} else if (CCVal == ISD::CondCode::SETNE) {
CountZeroes = N->getOperand(1);
ValOnZero = N->getOperand(2);
} else {
return SDValue();
}
if (CountZeroes.getOpcode() == ISD::TRUNCATE ||
CountZeroes.getOpcode() == ISD::ZERO_EXTEND)
CountZeroes = CountZeroes.getOperand(0);
if (CountZeroes.getOpcode() != ISD::CTTZ &&
CountZeroes.getOpcode() != ISD::CTTZ_ZERO_UNDEF &&
CountZeroes.getOpcode() != ISD::CTLZ &&
CountZeroes.getOpcode() != ISD::CTLZ_ZERO_UNDEF)
return SDValue();
if (!isNullConstant(ValOnZero))
return SDValue();
SDValue CountZeroesArgument = CountZeroes->getOperand(0);
if (Cond->getOperand(0) != CountZeroesArgument)
return SDValue();
if (CountZeroes.getOpcode() == ISD::CTTZ_ZERO_UNDEF) {
CountZeroes = DAG.getNode(ISD::CTTZ, SDLoc(CountZeroes),
CountZeroes.getValueType(), CountZeroesArgument);
} else if (CountZeroes.getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
CountZeroes = DAG.getNode(ISD::CTLZ, SDLoc(CountZeroes),
CountZeroes.getValueType(), CountZeroesArgument);
}
unsigned BitWidth = CountZeroes.getValueSizeInBits();
SDValue BitWidthMinusOne =
DAG.getConstant(BitWidth - 1, SDLoc(N), CountZeroes.getValueType());
auto AndNode = DAG.getNode(ISD::AND, SDLoc(N), CountZeroes.getValueType(),
CountZeroes, BitWidthMinusOne);
return DAG.getZExtOrTrunc(AndNode, SDLoc(N), N->getValueType(0));
}
static SDValue useInversedSetcc(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
SDValue Cond = N->getOperand(0);
SDValue True = N->getOperand(1);
SDValue False = N->getOperand(2);
SDLoc DL(N);
EVT VT = N->getValueType(0);
EVT CondVT = Cond.getValueType();
if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse())
return SDValue();
// Replace (setcc eq (and x, C)) with (setcc ne (and x, C))) to generate
// BEXTI, where C is power of 2.
if (Subtarget.hasStdExtZbs() && VT.isScalarInteger() &&
(Subtarget.hasStdExtZicond() || Subtarget.hasVendorXVentanaCondOps())) {
SDValue LHS = Cond.getOperand(0);
SDValue RHS = Cond.getOperand(1);
ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
if (CC == ISD::SETEQ && LHS.getOpcode() == ISD::AND &&
isa<ConstantSDNode>(LHS.getOperand(1)) && isNullConstant(RHS)) {
const APInt &MaskVal = LHS.getConstantOperandAPInt(1);
if (MaskVal.isPowerOf2() && !MaskVal.isSignedIntN(12))
return DAG.getSelect(DL, VT,
DAG.getSetCC(DL, CondVT, LHS, RHS, ISD::SETNE),
False, True);
}
}
return SDValue();
}
static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (SDValue Folded = foldSelectOfCTTZOrCTLZ(N, DAG))
return Folded;
if (SDValue V = useInversedSetcc(N, DAG, Subtarget))
return V;
if (Subtarget.hasConditionalMoveFusion())
return SDValue();
SDValue TrueVal = N->getOperand(1);
SDValue FalseVal = N->getOperand(2);
if (SDValue V = tryFoldSelectIntoOp(N, DAG, TrueVal, FalseVal, /*Swapped*/false))
return V;
return tryFoldSelectIntoOp(N, DAG, FalseVal, TrueVal, /*Swapped*/true);
}
/// If we have a build_vector where each lane is binop X, C, where C
/// is a constant (but not necessarily the same constant on all lanes),
/// form binop (build_vector x1, x2, ...), (build_vector c1, c2, c3, ..).
/// We assume that materializing a constant build vector will be no more
/// expensive that performing O(n) binops.
static SDValue performBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget,
const RISCVTargetLowering &TLI) {
SDLoc DL(N);
EVT VT = N->getValueType(0);
assert(!VT.isScalableVector() && "unexpected build vector");
if (VT.getVectorNumElements() == 1)
return SDValue();
const unsigned Opcode = N->op_begin()->getNode()->getOpcode();
if (!TLI.isBinOp(Opcode))
return SDValue();
if (!TLI.isOperationLegalOrCustom(Opcode, VT) || !TLI.isTypeLegal(VT))
return SDValue();
// This BUILD_VECTOR involves an implicit truncation, and sinking
// truncates through binops is non-trivial.
if (N->op_begin()->getValueType() != VT.getVectorElementType())
return SDValue();
SmallVector<SDValue> LHSOps;
SmallVector<SDValue> RHSOps;
for (SDValue Op : N->ops()) {
if (Op.isUndef()) {
// We can't form a divide or remainder from undef.
if (!DAG.isSafeToSpeculativelyExecute(Opcode))
return SDValue();
LHSOps.push_back(Op);
RHSOps.push_back(Op);
continue;
}
// TODO: We can handle operations which have an neutral rhs value
// (e.g. x + 0, a * 1 or a << 0), but we then have to keep track
// of profit in a more explicit manner.
if (Op.getOpcode() != Opcode || !Op.hasOneUse())
return SDValue();
LHSOps.push_back(Op.getOperand(0));
if (!isa<ConstantSDNode>(Op.getOperand(1)) &&
!isa<ConstantFPSDNode>(Op.getOperand(1)))
return SDValue();
// FIXME: Return failure if the RHS type doesn't match the LHS. Shifts may
// have different LHS and RHS types.
if (Op.getOperand(0).getValueType() != Op.getOperand(1).getValueType())
return SDValue();
RHSOps.push_back(Op.getOperand(1));
}
return DAG.getNode(Opcode, DL, VT, DAG.getBuildVector(VT, DL, LHSOps),
DAG.getBuildVector(VT, DL, RHSOps));
}
static SDValue performINSERT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget,
const RISCVTargetLowering &TLI) {
SDValue InVec = N->getOperand(0);
SDValue InVal = N->getOperand(1);
SDValue EltNo = N->getOperand(2);
SDLoc DL(N);
EVT VT = InVec.getValueType();
if (VT.isScalableVector())
return SDValue();
if (!InVec.hasOneUse())
return SDValue();
// Given insert_vector_elt (binop a, VecC), (same_binop b, C2), Elt
// move the insert_vector_elts into the arms of the binop. Note that
// the new RHS must be a constant.
const unsigned InVecOpcode = InVec->getOpcode();
if (InVecOpcode == InVal->getOpcode() && TLI.isBinOp(InVecOpcode) &&
InVal.hasOneUse()) {
SDValue InVecLHS = InVec->getOperand(0);
SDValue InVecRHS = InVec->getOperand(1);
SDValue InValLHS = InVal->getOperand(0);
SDValue InValRHS = InVal->getOperand(1);
if (!ISD::isBuildVectorOfConstantSDNodes(InVecRHS.getNode()))
return SDValue();
if (!isa<ConstantSDNode>(InValRHS) && !isa<ConstantFPSDNode>(InValRHS))
return SDValue();
// FIXME: Return failure if the RHS type doesn't match the LHS. Shifts may
// have different LHS and RHS types.
if (InVec.getOperand(0).getValueType() != InVec.getOperand(1).getValueType())
return SDValue();
SDValue LHS = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT,
InVecLHS, InValLHS, EltNo);
SDValue RHS = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT,
InVecRHS, InValRHS, EltNo);
return DAG.getNode(InVecOpcode, DL, VT, LHS, RHS);
}
// Given insert_vector_elt (concat_vectors ...), InVal, Elt
// move the insert_vector_elt to the source operand of the concat_vector.
if (InVec.getOpcode() != ISD::CONCAT_VECTORS)
return SDValue();
auto *IndexC = dyn_cast<ConstantSDNode>(EltNo);
if (!IndexC)
return SDValue();
unsigned Elt = IndexC->getZExtValue();
EVT ConcatVT = InVec.getOperand(0).getValueType();
if (ConcatVT.getVectorElementType() != InVal.getValueType())
return SDValue();
unsigned ConcatNumElts = ConcatVT.getVectorNumElements();
SDValue NewIdx = DAG.getVectorIdxConstant(Elt % ConcatNumElts, DL);
unsigned ConcatOpIdx = Elt / ConcatNumElts;
SDValue ConcatOp = InVec.getOperand(ConcatOpIdx);
ConcatOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ConcatVT,
ConcatOp, InVal, NewIdx);
SmallVector<SDValue> ConcatOps;
ConcatOps.append(InVec->op_begin(), InVec->op_end());
ConcatOps[ConcatOpIdx] = ConcatOp;
return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps);
}
// If we're concatenating a series of vector loads like
// concat_vectors (load v4i8, p+0), (load v4i8, p+n), (load v4i8, p+n*2) ...
// Then we can turn this into a strided load by widening the vector elements
// vlse32 p, stride=n
static SDValue performCONCAT_VECTORSCombine(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget,
const RISCVTargetLowering &TLI) {
SDLoc DL(N);
EVT VT = N->getValueType(0);
// Only perform this combine on legal MVTs.
if (!TLI.isTypeLegal(VT))
return SDValue();
// TODO: Potentially extend this to scalable vectors
if (VT.isScalableVector())
return SDValue();
auto *BaseLd = dyn_cast<LoadSDNode>(N->getOperand(0));
if (!BaseLd || !BaseLd->isSimple() || !ISD::isNormalLoad(BaseLd) ||
!SDValue(BaseLd, 0).hasOneUse())
return SDValue();
EVT BaseLdVT = BaseLd->getValueType(0);
// Go through the loads and check that they're strided
SmallVector<LoadSDNode *> Lds;
Lds.push_back(BaseLd);
Align Align = BaseLd->getAlign();
for (SDValue Op : N->ops().drop_front()) {
auto *Ld = dyn_cast<LoadSDNode>(Op);
if (!Ld || !Ld->isSimple() || !Op.hasOneUse() ||
Ld->getChain() != BaseLd->getChain() || !ISD::isNormalLoad(Ld) ||
Ld->getValueType(0) != BaseLdVT)
return SDValue();
Lds.push_back(Ld);
// The common alignment is the most restrictive (smallest) of all the loads
Align = std::min(Align, Ld->getAlign());
}
using PtrDiff = std::pair<std::variant<int64_t, SDValue>, bool>;
auto GetPtrDiff = [&DAG](LoadSDNode *Ld1,
LoadSDNode *Ld2) -> std::optional<PtrDiff> {
// If the load ptrs can be decomposed into a common (Base + Index) with a
// common constant stride, then return the constant stride.
BaseIndexOffset BIO1 = BaseIndexOffset::match(Ld1, DAG);
BaseIndexOffset BIO2 = BaseIndexOffset::match(Ld2, DAG);
if (BIO1.equalBaseIndex(BIO2, DAG))
return {{BIO2.getOffset() - BIO1.getOffset(), false}};
// Otherwise try to match (add LastPtr, Stride) or (add NextPtr, Stride)
SDValue P1 = Ld1->getBasePtr();
SDValue P2 = Ld2->getBasePtr();
if (P2.getOpcode() == ISD::ADD && P2.getOperand(0) == P1)
return {{P2.getOperand(1), false}};
if (P1.getOpcode() == ISD::ADD && P1.getOperand(0) == P2)
return {{P1.getOperand(1), true}};
return std::nullopt;
};
// Get the distance between the first and second loads
auto BaseDiff = GetPtrDiff(Lds[0], Lds[1]);
if (!BaseDiff)
return SDValue();
// Check all the loads are the same distance apart
for (auto *It = Lds.begin() + 1; It != Lds.end() - 1; It++)
if (GetPtrDiff(*It, *std::next(It)) != BaseDiff)
return SDValue();
// TODO: At this point, we've successfully matched a generalized gather
// load. Maybe we should emit that, and then move the specialized
// matchers above and below into a DAG combine?
// Get the widened scalar type, e.g. v4i8 -> i64
unsigned WideScalarBitWidth =
BaseLdVT.getScalarSizeInBits() * BaseLdVT.getVectorNumElements();
MVT WideScalarVT = MVT::getIntegerVT(WideScalarBitWidth);
// Get the vector type for the strided load, e.g. 4 x v4i8 -> v4i64
MVT WideVecVT = MVT::getVectorVT(WideScalarVT, N->getNumOperands());
if (!TLI.isTypeLegal(WideVecVT))
return SDValue();
// Check that the operation is legal
if (!TLI.isLegalStridedLoadStore(WideVecVT, Align))
return SDValue();
auto [StrideVariant, MustNegateStride] = *BaseDiff;
SDValue Stride = std::holds_alternative<SDValue>(StrideVariant)
? std::get<SDValue>(StrideVariant)
: DAG.getConstant(std::get<int64_t>(StrideVariant), DL,
Lds[0]->getOffset().getValueType());
if (MustNegateStride)
Stride = DAG.getNegative(Stride, DL, Stride.getValueType());
SDVTList VTs = DAG.getVTList({WideVecVT, MVT::Other});
SDValue IntID =
DAG.getTargetConstant(Intrinsic::riscv_masked_strided_load, DL,
Subtarget.getXLenVT());
SDValue AllOneMask =
DAG.getSplat(WideVecVT.changeVectorElementType(MVT::i1), DL,
DAG.getConstant(1, DL, MVT::i1));
SDValue Ops[] = {BaseLd->getChain(), IntID, DAG.getUNDEF(WideVecVT),
BaseLd->getBasePtr(), Stride, AllOneMask};
uint64_t MemSize;
if (auto *ConstStride = dyn_cast<ConstantSDNode>(Stride);
ConstStride && ConstStride->getSExtValue() >= 0)
// total size = (elsize * n) + (stride - elsize) * (n-1)
// = elsize + stride * (n-1)
MemSize = WideScalarVT.getSizeInBits() +
ConstStride->getSExtValue() * (N->getNumOperands() - 1);
else
// If Stride isn't constant, then we can't know how much it will load
MemSize = MemoryLocation::UnknownSize;
MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
BaseLd->getPointerInfo(), BaseLd->getMemOperand()->getFlags(), MemSize,
Align);
SDValue StridedLoad = DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs,
Ops, WideVecVT, MMO);
for (SDValue Ld : N->ops())
DAG.makeEquivalentMemoryOrdering(cast<LoadSDNode>(Ld), StridedLoad);
return DAG.getBitcast(VT.getSimpleVT(), StridedLoad);
}
static SDValue combineToVWMACC(SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
assert(N->getOpcode() == RISCVISD::ADD_VL || N->getOpcode() == ISD::ADD);
if (N->getValueType(0).isFixedLengthVector())
return SDValue();
SDValue Addend = N->getOperand(0);
SDValue MulOp = N->getOperand(1);
if (N->getOpcode() == RISCVISD::ADD_VL) {
SDValue AddMergeOp = N->getOperand(2);
if (!AddMergeOp.isUndef())
return SDValue();
}
auto IsVWMulOpc = [](unsigned Opc) {
switch (Opc) {
case RISCVISD::VWMUL_VL:
case RISCVISD::VWMULU_VL:
case RISCVISD::VWMULSU_VL:
return true;
default:
return false;
}
};
if (!IsVWMulOpc(MulOp.getOpcode()))
std::swap(Addend, MulOp);
if (!IsVWMulOpc(MulOp.getOpcode()))
return SDValue();
SDValue MulMergeOp = MulOp.getOperand(2);
if (!MulMergeOp.isUndef())
return SDValue();
auto [AddMask, AddVL] = [](SDNode *N, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
if (N->getOpcode() == ISD::ADD) {
SDLoc DL(N);
return getDefaultScalableVLOps(N->getSimpleValueType(0), DL, DAG,
Subtarget);
}
return std::make_pair(N->getOperand(3), N->getOperand(4));
}(N, DAG, Subtarget);
SDValue MulMask = MulOp.getOperand(3);
SDValue MulVL = MulOp.getOperand(4);
if (AddMask != MulMask || AddVL != MulVL)
return SDValue();
unsigned Opc = RISCVISD::VWMACC_VL + MulOp.getOpcode() - RISCVISD::VWMUL_VL;
static_assert(RISCVISD::VWMACC_VL + 1 == RISCVISD::VWMACCU_VL,
"Unexpected opcode after VWMACC_VL");
static_assert(RISCVISD::VWMACC_VL + 2 == RISCVISD::VWMACCSU_VL,
"Unexpected opcode after VWMACC_VL!");
static_assert(RISCVISD::VWMUL_VL + 1 == RISCVISD::VWMULU_VL,
"Unexpected opcode after VWMUL_VL!");
static_assert(RISCVISD::VWMUL_VL + 2 == RISCVISD::VWMULSU_VL,
"Unexpected opcode after VWMUL_VL!");
SDLoc DL(N);
EVT VT = N->getValueType(0);
SDValue Ops[] = {MulOp.getOperand(0), MulOp.getOperand(1), Addend, AddMask,
AddVL};
return DAG.getNode(Opc, DL, VT, Ops);
}
static bool legalizeScatterGatherIndexType(SDLoc DL, SDValue &Index,
ISD::MemIndexType &IndexType,
RISCVTargetLowering::DAGCombinerInfo &DCI) {
if (!DCI.isBeforeLegalize())
return false;
SelectionDAG &DAG = DCI.DAG;
const MVT XLenVT =
DAG.getMachineFunction().getSubtarget<RISCVSubtarget>().getXLenVT();
const EVT IndexVT = Index.getValueType();
// RISC-V indexed loads only support the "unsigned unscaled" addressing
// mode, so anything else must be manually legalized.
if (!isIndexTypeSigned(IndexType))
return false;
if (IndexVT.getVectorElementType().bitsLT(XLenVT)) {
// Any index legalization should first promote to XLenVT, so we don't lose
// bits when scaling. This may create an illegal index type so we let
// LLVM's legalization take care of the splitting.
// FIXME: LLVM can't split VP_GATHER or VP_SCATTER yet.
Index = DAG.getNode(ISD::SIGN_EXTEND, DL,
IndexVT.changeVectorElementType(XLenVT), Index);
}
IndexType = ISD::UNSIGNED_SCALED;
return true;
}
/// Match the index vector of a scatter or gather node as the shuffle mask
/// which performs the rearrangement if possible. Will only match if
/// all lanes are touched, and thus replacing the scatter or gather with
/// a unit strided access and shuffle is legal.
static bool matchIndexAsShuffle(EVT VT, SDValue Index, SDValue Mask,
SmallVector<int> &ShuffleMask) {
if (!ISD::isConstantSplatVectorAllOnes(Mask.getNode()))
return false;
if (!ISD::isBuildVectorOfConstantSDNodes(Index.getNode()))
return false;
const unsigned ElementSize = VT.getScalarStoreSize();
const unsigned NumElems = VT.getVectorNumElements();
// Create the shuffle mask and check all bits active
assert(ShuffleMask.empty());
BitVector ActiveLanes(NumElems);
for (unsigned i = 0; i < Index->getNumOperands(); i++) {
// TODO: We've found an active bit of UB, and could be
// more aggressive here if desired.
if (Index->getOperand(i)->isUndef())
return false;
uint64_t C = Index->getConstantOperandVal(i);
if (C % ElementSize != 0)
return false;
C = C / ElementSize;
if (C >= NumElems)
return false;
ShuffleMask.push_back(C);
ActiveLanes.set(C);
}
return ActiveLanes.all();
}
/// Match the index of a gather or scatter operation as an operation
/// with twice the element width and half the number of elements. This is
/// generally profitable (if legal) because these operations are linear
/// in VL, so even if we cause some extract VTYPE/VL toggles, we still
/// come out ahead.
static bool matchIndexAsWiderOp(EVT VT, SDValue Index, SDValue Mask,
Align BaseAlign, const RISCVSubtarget &ST) {
if (!ISD::isConstantSplatVectorAllOnes(Mask.getNode()))
return false;
if (!ISD::isBuildVectorOfConstantSDNodes(Index.getNode()))
return false;
// Attempt a doubling. If we can use a element type 4x or 8x in
// size, this will happen via multiply iterations of the transform.
const unsigned NumElems = VT.getVectorNumElements();
if (NumElems % 2 != 0)
return false;
const unsigned ElementSize = VT.getScalarStoreSize();
const unsigned WiderElementSize = ElementSize * 2;
if (WiderElementSize > ST.getELen()/8)
return false;
if (!ST.enableUnalignedVectorMem() && BaseAlign < WiderElementSize)
return false;
for (unsigned i = 0; i < Index->getNumOperands(); i++) {
// TODO: We've found an active bit of UB, and could be
// more aggressive here if desired.
if (Index->getOperand(i)->isUndef())
return false;
// TODO: This offset check is too strict if we support fully
// misaligned memory operations.
uint64_t C = Index->getConstantOperandVal(i);
if (i % 2 == 0) {
if (C % WiderElementSize != 0)
return false;
continue;
}
uint64_t Last = Index->getConstantOperandVal(i-1);
if (C != Last + ElementSize)
return false;
}
return true;
}
SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
const MVT XLenVT = Subtarget.getXLenVT();
SDLoc DL(N);
// Helper to call SimplifyDemandedBits on an operand of N where only some low
// bits are demanded. N will be added to the Worklist if it was not deleted.
// Caller should return SDValue(N, 0) if this returns true.
auto SimplifyDemandedLowBitsHelper = [&](unsigned OpNo, unsigned LowBits) {
SDValue Op = N->getOperand(OpNo);
APInt Mask = APInt::getLowBitsSet(Op.getValueSizeInBits(), LowBits);
if (!SimplifyDemandedBits(Op, Mask, DCI))
return false;
if (N->getOpcode() != ISD::DELETED_NODE)
DCI.AddToWorklist(N);
return true;
};
switch (N->getOpcode()) {
default:
break;
case RISCVISD::SplitF64: {
SDValue Op0 = N->getOperand(0);
// If the input to SplitF64 is just BuildPairF64 then the operation is
// redundant. Instead, use BuildPairF64's operands directly.
if (Op0->getOpcode() == RISCVISD::BuildPairF64)
return DCI.CombineTo(N, Op0.getOperand(0), Op0.getOperand(1));
if (Op0->isUndef()) {
SDValue Lo = DAG.getUNDEF(MVT::i32);
SDValue Hi = DAG.getUNDEF(MVT::i32);
return DCI.CombineTo(N, Lo, Hi);
}
// It's cheaper to materialise two 32-bit integers than to load a double
// from the constant pool and transfer it to integer registers through the
// stack.
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op0)) {
APInt V = C->getValueAPF().bitcastToAPInt();
SDValue Lo = DAG.getConstant(V.trunc(32), DL, MVT::i32);
SDValue Hi = DAG.getConstant(V.lshr(32).trunc(32), DL, MVT::i32);
return DCI.CombineTo(N, Lo, Hi);
}
// This is a target-specific version of a DAGCombine performed in
// DAGCombiner::visitBITCAST. It performs the equivalent of:
// fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
// fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) ||
!Op0.getNode()->hasOneUse())
break;
SDValue NewSplitF64 =
DAG.getNode(RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32),
Op0.getOperand(0));
SDValue Lo = NewSplitF64.getValue(0);
SDValue Hi = NewSplitF64.getValue(1);
APInt SignBit = APInt::getSignMask(32);
if (Op0.getOpcode() == ISD::FNEG) {
SDValue NewHi = DAG.getNode(ISD::XOR, DL, MVT::i32, Hi,
DAG.getConstant(SignBit, DL, MVT::i32));
return DCI.CombineTo(N, Lo, NewHi);
}
assert(Op0.getOpcode() == ISD::FABS);
SDValue NewHi = DAG.getNode(ISD::AND, DL, MVT::i32, Hi,
DAG.getConstant(~SignBit, DL, MVT::i32));
return DCI.CombineTo(N, Lo, NewHi);
}
case RISCVISD::SLLW:
case RISCVISD::SRAW:
case RISCVISD::SRLW:
case RISCVISD::RORW:
case RISCVISD::ROLW: {
// Only the lower 32 bits of LHS and lower 5 bits of RHS are read.
if (SimplifyDemandedLowBitsHelper(0, 32) ||
SimplifyDemandedLowBitsHelper(1, 5))
return SDValue(N, 0);
break;
}
case RISCVISD::CLZW:
case RISCVISD::CTZW: {
// Only the lower 32 bits of the first operand are read
if (SimplifyDemandedLowBitsHelper(0, 32))
return SDValue(N, 0);
break;
}
case RISCVISD::FMV_W_X_RV64: {
// If the input to FMV_W_X_RV64 is just FMV_X_ANYEXTW_RV64 the the
// conversion is unnecessary and can be replaced with the
// FMV_X_ANYEXTW_RV64 operand.
SDValue Op0 = N->getOperand(0);
if (Op0.getOpcode() == RISCVISD::FMV_X_ANYEXTW_RV64)
return Op0.getOperand(0);
break;
}
case RISCVISD::FMV_X_ANYEXTH:
case RISCVISD::FMV_X_ANYEXTW_RV64: {
SDLoc DL(N);
SDValue Op0 = N->getOperand(0);
MVT VT = N->getSimpleValueType(0);
// If the input to FMV_X_ANYEXTW_RV64 is just FMV_W_X_RV64 then the
// conversion is unnecessary and can be replaced with the FMV_W_X_RV64
// operand. Similar for FMV_X_ANYEXTH and FMV_H_X.
if ((N->getOpcode() == RISCVISD::FMV_X_ANYEXTW_RV64 &&
Op0->getOpcode() == RISCVISD::FMV_W_X_RV64) ||
(N->getOpcode() == RISCVISD::FMV_X_ANYEXTH &&
Op0->getOpcode() == RISCVISD::FMV_H_X)) {
assert(Op0.getOperand(0).getValueType() == VT &&
"Unexpected value type!");
return Op0.getOperand(0);
}
// This is a target-specific version of a DAGCombine performed in
// DAGCombiner::visitBITCAST. It performs the equivalent of:
// fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
// fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) ||
!Op0.getNode()->hasOneUse())
break;
SDValue NewFMV = DAG.getNode(N->getOpcode(), DL, VT, Op0.getOperand(0));
unsigned FPBits = N->getOpcode() == RISCVISD::FMV_X_ANYEXTW_RV64 ? 32 : 16;
APInt SignBit = APInt::getSignMask(FPBits).sext(VT.getSizeInBits());
if (Op0.getOpcode() == ISD::FNEG)
return DAG.getNode(ISD::XOR, DL, VT, NewFMV,
DAG.getConstant(SignBit, DL, VT));
assert(Op0.getOpcode() == ISD::FABS);
return DAG.getNode(ISD::AND, DL, VT, NewFMV,
DAG.getConstant(~SignBit, DL, VT));
}
case ISD::ABS: {
EVT VT = N->getValueType(0);
SDValue N0 = N->getOperand(0);
// abs (sext) -> zext (abs)
// abs (zext) -> zext (handled elsewhere)
if (VT.isVector() && N0.hasOneUse() && N0.getOpcode() == ISD::SIGN_EXTEND) {
SDValue Src = N0.getOperand(0);
SDLoc DL(N);
return DAG.getNode(ISD::ZERO_EXTEND, DL, VT,
DAG.getNode(ISD::ABS, DL, Src.getValueType(), Src));
}
break;
}
case ISD::ADD: {
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
if (SDValue V = combineToVWMACC(N, DAG, Subtarget))
return V;
return performADDCombine(N, DAG, Subtarget);
}
case ISD::SUB: {
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
return performSUBCombine(N, DAG, Subtarget);
}
case ISD::AND:
return performANDCombine(N, DCI, Subtarget);
case ISD::OR: {
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
return performORCombine(N, DCI, Subtarget);
}
case ISD::XOR:
return performXORCombine(N, DAG, Subtarget);
case ISD::MUL:
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
return performMULCombine(N, DAG, DCI, Subtarget);
case ISD::SDIV:
case ISD::UDIV:
case ISD::SREM:
case ISD::UREM:
if (SDValue V = combineBinOpOfZExt(N, DAG))
return V;
break;
case ISD::FADD:
case ISD::UMAX:
case ISD::UMIN:
case ISD::SMAX:
case ISD::SMIN:
case ISD::FMAXNUM:
case ISD::FMINNUM: {
if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
return V;
if (SDValue V = combineBinOpOfExtractToReduceTree(N, DAG, Subtarget))
return V;
return SDValue();
}
case ISD::SETCC:
return performSETCCCombine(N, DAG, Subtarget);
case ISD::SIGN_EXTEND_INREG:
return performSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
case ISD::ZERO_EXTEND:
// Fold (zero_extend (fp_to_uint X)) to prevent forming fcvt+zexti32 during
// type legalization. This is safe because fp_to_uint produces poison if
// it overflows.
if (N->getValueType(0) == MVT::i64 && Subtarget.is64Bit()) {
SDValue Src = N->getOperand(0);
if (Src.getOpcode() == ISD::FP_TO_UINT &&
isTypeLegal(Src.getOperand(0).getValueType()))
return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), MVT::i64,
Src.getOperand(0));
if (Src.getOpcode() == ISD::STRICT_FP_TO_UINT && Src.hasOneUse() &&
isTypeLegal(Src.getOperand(1).getValueType())) {
SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other);
SDValue Res = DAG.getNode(ISD::STRICT_FP_TO_UINT, SDLoc(N), VTs,
Src.getOperand(0), Src.getOperand(1));
DCI.CombineTo(N, Res);
DAG.ReplaceAllUsesOfValueWith(Src.getValue(1), Res.getValue(1));
DCI.recursivelyDeleteUnusedNodes(Src.getNode());
return SDValue(N, 0); // Return N so it doesn't get rechecked.
}
}
return SDValue();
case RISCVISD::TRUNCATE_VECTOR_VL: {
// trunc (sra sext (X), zext (Y)) -> sra (X, smin (Y, scalarsize(Y) - 1))
// This would be benefit for the cases where X and Y are both the same value
// type of low precision vectors. Since the truncate would be lowered into
// n-levels TRUNCATE_VECTOR_VL to satisfy RVV's SEW*2->SEW truncate
// restriction, such pattern would be expanded into a series of "vsetvli"
// and "vnsrl" instructions later to reach this point.
auto IsTruncNode = [](SDValue V) {
if (V.getOpcode() != RISCVISD::TRUNCATE_VECTOR_VL)
return false;
SDValue VL = V.getOperand(2);
auto *C = dyn_cast<ConstantSDNode>(VL);
// Assume all TRUNCATE_VECTOR_VL nodes use VLMAX for VMSET_VL operand
bool IsVLMAXForVMSET = (C && C->isAllOnes()) ||
(isa<RegisterSDNode>(VL) &&
cast<RegisterSDNode>(VL)->getReg() == RISCV::X0);
return V.getOperand(1).getOpcode() == RISCVISD::VMSET_VL &&
IsVLMAXForVMSET;
};
SDValue Op = N->getOperand(0);
// We need to first find the inner level of TRUNCATE_VECTOR_VL node
// to distinguish such pattern.
while (IsTruncNode(Op)) {
if (!Op.hasOneUse())
return SDValue();
Op = Op.getOperand(0);
}
if (Op.getOpcode() == ISD::SRA && Op.hasOneUse()) {
SDValue N0 = Op.getOperand(0);
SDValue N1 = Op.getOperand(1);
if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.hasOneUse() &&
N1.getOpcode() == ISD::ZERO_EXTEND && N1.hasOneUse()) {
SDValue N00 = N0.getOperand(0);
SDValue N10 = N1.getOperand(0);
if (N00.getValueType().isVector() &&
N00.getValueType() == N10.getValueType() &&
N->getValueType(0) == N10.getValueType()) {
unsigned MaxShAmt = N10.getValueType().getScalarSizeInBits() - 1;
SDValue SMin = DAG.getNode(
ISD::SMIN, SDLoc(N1), N->getValueType(0), N10,
DAG.getConstant(MaxShAmt, SDLoc(N1), N->getValueType(0)));
return DAG.getNode(ISD::SRA, SDLoc(N), N->getValueType(0), N00, SMin);
}
}
}
break;
}
case ISD::TRUNCATE:
return performTRUNCATECombine(N, DAG, Subtarget);
case ISD::SELECT:
return performSELECTCombine(N, DAG, Subtarget);
case RISCVISD::CZERO_EQZ:
case RISCVISD::CZERO_NEZ:
// czero_eq X, (xor Y, 1) -> czero_ne X, Y if Y is 0 or 1.
// czero_ne X, (xor Y, 1) -> czero_eq X, Y if Y is 0 or 1.
if (N->getOperand(1).getOpcode() == ISD::XOR &&
isOneConstant(N->getOperand(1).getOperand(1))) {
SDValue Cond = N->getOperand(1).getOperand(0);
APInt Mask = APInt::getBitsSetFrom(Cond.getValueSizeInBits(), 1);
if (DAG.MaskedValueIsZero(Cond, Mask)) {
unsigned NewOpc = N->getOpcode() == RISCVISD::CZERO_EQZ
? RISCVISD::CZERO_NEZ
: RISCVISD::CZERO_EQZ;
return DAG.getNode(NewOpc, SDLoc(N), N->getValueType(0),
N->getOperand(0), Cond);
}
}
return SDValue();
case RISCVISD::SELECT_CC: {
// Transform
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
SDValue CC = N->getOperand(2);
ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
SDValue TrueV = N->getOperand(3);
SDValue FalseV = N->getOperand(4);
SDLoc DL(N);
EVT VT = N->getValueType(0);
// If the True and False values are the same, we don't need a select_cc.
if (TrueV == FalseV)
return TrueV;
// (select (x < 0), y, z) -> x >> (XLEN - 1) & (y - z) + z
// (select (x >= 0), y, z) -> x >> (XLEN - 1) & (z - y) + y
if (!Subtarget.hasShortForwardBranchOpt() && isa<ConstantSDNode>(TrueV) &&
isa<ConstantSDNode>(FalseV) && isNullConstant(RHS) &&
(CCVal == ISD::CondCode::SETLT || CCVal == ISD::CondCode::SETGE)) {
if (CCVal == ISD::CondCode::SETGE)
std::swap(TrueV, FalseV);
int64_t TrueSImm = cast<ConstantSDNode>(TrueV)->getSExtValue();
int64_t FalseSImm = cast<ConstantSDNode>(FalseV)->getSExtValue();
// Only handle simm12, if it is not in this range, it can be considered as
// register.
if (isInt<12>(TrueSImm) && isInt<12>(FalseSImm) &&
isInt<12>(TrueSImm - FalseSImm)) {
SDValue SRA =
DAG.getNode(ISD::SRA, DL, VT, LHS,
DAG.getConstant(Subtarget.getXLen() - 1, DL, VT));
SDValue AND =
DAG.getNode(ISD::AND, DL, VT, SRA,
DAG.getConstant(TrueSImm - FalseSImm, DL, VT));
return DAG.getNode(ISD::ADD, DL, VT, AND, FalseV);
}
if (CCVal == ISD::CondCode::SETGE)
std::swap(TrueV, FalseV);
}
if (combine_CC(LHS, RHS, CC, DL, DAG, Subtarget))
return DAG.getNode(RISCVISD::SELECT_CC, DL, N->getValueType(0),
{LHS, RHS, CC, TrueV, FalseV});
if (!Subtarget.hasConditionalMoveFusion()) {
// (select c, -1, y) -> -c | y
if (isAllOnesConstant(TrueV)) {
SDValue C = DAG.getSetCC(DL, VT, LHS, RHS, CCVal);
SDValue Neg = DAG.getNegative(C, DL, VT);
return DAG.getNode(ISD::OR, DL, VT, Neg, FalseV);
}
// (select c, y, -1) -> -!c | y
if (isAllOnesConstant(FalseV)) {
SDValue C =
DAG.getSetCC(DL, VT, LHS, RHS, ISD::getSetCCInverse(CCVal, VT));
SDValue Neg = DAG.getNegative(C, DL, VT);
return DAG.getNode(ISD::OR, DL, VT, Neg, TrueV);
}
// (select c, 0, y) -> -!c & y
if (isNullConstant(TrueV)) {
SDValue C =
DAG.getSetCC(DL, VT, LHS, RHS, ISD::getSetCCInverse(CCVal, VT));
SDValue Neg = DAG.getNegative(C, DL, VT);
return DAG.getNode(ISD::AND, DL, VT, Neg, FalseV);
}
// (select c, y, 0) -> -c & y
if (isNullConstant(FalseV)) {
SDValue C = DAG.getSetCC(DL, VT, LHS, RHS, CCVal);
SDValue Neg = DAG.getNegative(C, DL, VT);
return DAG.getNode(ISD::AND, DL, VT, Neg, TrueV);
}
// (riscvisd::select_cc x, 0, ne, x, 1) -> (add x, (setcc x, 0, eq))
// (riscvisd::select_cc x, 0, eq, 1, x) -> (add x, (setcc x, 0, eq))
if (((isOneConstant(FalseV) && LHS == TrueV &&
CCVal == ISD::CondCode::SETNE) ||
(isOneConstant(TrueV) && LHS == FalseV &&
CCVal == ISD::CondCode::SETEQ)) &&
isNullConstant(RHS)) {
// freeze it to be safe.
LHS = DAG.getFreeze(LHS);
SDValue C = DAG.getSetCC(DL, VT, LHS, RHS, ISD::CondCode::SETEQ);
return DAG.getNode(ISD::ADD, DL, VT, LHS, C);
}
}
// If both true/false are an xor with 1, pull through the select.
// This can occur after op legalization if both operands are setccs that
// require an xor to invert.
// FIXME: Generalize to other binary ops with identical operand?
if (TrueV.getOpcode() == ISD::XOR && FalseV.getOpcode() == ISD::XOR &&
TrueV.getOperand(1) == FalseV.getOperand(1) &&
isOneConstant(TrueV.getOperand(1)) &&
TrueV.hasOneUse() && FalseV.hasOneUse()) {
SDValue NewSel = DAG.getNode(RISCVISD::SELECT_CC, DL, VT, LHS, RHS, CC,
TrueV.getOperand(0), FalseV.getOperand(0));
return DAG.getNode(ISD::XOR, DL, VT, NewSel, TrueV.getOperand(1));
}
return SDValue();
}
case RISCVISD::BR_CC: {
SDValue LHS = N->getOperand(1);
SDValue RHS = N->getOperand(2);
SDValue CC = N->getOperand(3);
SDLoc DL(N);
if (combine_CC(LHS, RHS, CC, DL, DAG, Subtarget))
return DAG.getNode(RISCVISD::BR_CC, DL, N->getValueType(0),
N->getOperand(0), LHS, RHS, CC, N->getOperand(4));
return SDValue();
}
case ISD::BITREVERSE:
return performBITREVERSECombine(N, DAG, Subtarget);
case ISD::FP_TO_SINT:
case ISD::FP_TO_UINT:
return performFP_TO_INTCombine(N, DCI, Subtarget);
case ISD::FP_TO_SINT_SAT:
case ISD::FP_TO_UINT_SAT:
return performFP_TO_INT_SATCombine(N, DCI, Subtarget);
case ISD::FCOPYSIGN: {
EVT VT = N->getValueType(0);
if (!VT.isVector())
break;
// There is a form of VFSGNJ which injects the negated sign of its second
// operand. Try and bubble any FNEG up after the extend/round to produce
// this optimized pattern. Avoid modifying cases where FP_ROUND and
// TRUNC=1.
SDValue In2 = N->getOperand(1);
// Avoid cases where the extend/round has multiple uses, as duplicating
// those is typically more expensive than removing a fneg.
if (!In2.hasOneUse())
break;
if (In2.getOpcode() != ISD::FP_EXTEND &&
(In2.getOpcode() != ISD::FP_ROUND || In2.getConstantOperandVal(1) != 0))
break;
In2 = In2.getOperand(0);
if (In2.getOpcode() != ISD::FNEG)
break;
SDLoc DL(N);
SDValue NewFPExtRound = DAG.getFPExtendOrRound(In2.getOperand(0), DL, VT);
return DAG.getNode(ISD::FCOPYSIGN, DL, VT, N->getOperand(0),
DAG.getNode(ISD::FNEG, DL, VT, NewFPExtRound));
}
case ISD::MGATHER: {
const auto *MGN = dyn_cast<MaskedGatherSDNode>(N);
const EVT VT = N->getValueType(0);
SDValue Index = MGN->getIndex();
SDValue ScaleOp = MGN->getScale();
ISD::MemIndexType IndexType = MGN->getIndexType();
assert(!MGN->isIndexScaled() &&
"Scaled gather/scatter should not be formed");
SDLoc DL(N);
if (legalizeScatterGatherIndexType(DL, Index, IndexType, DCI))
return DAG.getMaskedGather(
N->getVTList(), MGN->getMemoryVT(), DL,
{MGN->getChain(), MGN->getPassThru(), MGN->getMask(),
MGN->getBasePtr(), Index, ScaleOp},
MGN->getMemOperand(), IndexType, MGN->getExtensionType());
if (narrowIndex(Index, IndexType, DAG))
return DAG.getMaskedGather(
N->getVTList(), MGN->getMemoryVT(), DL,
{MGN->getChain(), MGN->getPassThru(), MGN->getMask(),
MGN->getBasePtr(), Index, ScaleOp},
MGN->getMemOperand(), IndexType, MGN->getExtensionType());
if (Index.getOpcode() == ISD::BUILD_VECTOR &&
MGN->getExtensionType() == ISD::NON_EXTLOAD && isTypeLegal(VT)) {
// The sequence will be XLenVT, not the type of Index. Tell
// isSimpleVIDSequence this so we avoid overflow.
if (std::optional<VIDSequence> SimpleVID =
isSimpleVIDSequence(Index, Subtarget.getXLen());
SimpleVID && SimpleVID->StepDenominator == 1) {
const int64_t StepNumerator = SimpleVID->StepNumerator;
const int64_t Addend = SimpleVID->Addend;
// Note: We don't need to check alignment here since (by assumption
// from the existance of the gather), our offsets must be sufficiently
// aligned.
const EVT PtrVT = getPointerTy(DAG.getDataLayout());
assert(MGN->getBasePtr()->getValueType(0) == PtrVT);
assert(IndexType == ISD::UNSIGNED_SCALED);
SDValue BasePtr = DAG.getNode(ISD::ADD, DL, PtrVT, MGN->getBasePtr(),
DAG.getConstant(Addend, DL, PtrVT));
SDVTList VTs = DAG.getVTList({VT, MVT::Other});
SDValue IntID =
DAG.getTargetConstant(Intrinsic::riscv_masked_strided_load, DL,
XLenVT);
SDValue Ops[] =
{MGN->getChain(), IntID, MGN->getPassThru(), BasePtr,
DAG.getConstant(StepNumerator, DL, XLenVT), MGN->getMask()};
return DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs,
Ops, VT, MGN->getMemOperand());
}
}
SmallVector<int> ShuffleMask;
if (MGN->getExtensionType() == ISD::NON_EXTLOAD &&
matchIndexAsShuffle(VT, Index, MGN->getMask(), ShuffleMask)) {
SDValue Load = DAG.getMaskedLoad(VT, DL, MGN->getChain(),
MGN->getBasePtr(), DAG.getUNDEF(XLenVT),
MGN->getMask(), DAG.getUNDEF(VT),
MGN->getMemoryVT(), MGN->getMemOperand(),
ISD::UNINDEXED, ISD::NON_EXTLOAD);
SDValue Shuffle =
DAG.getVectorShuffle(VT, DL, Load, DAG.getUNDEF(VT), ShuffleMask);
return DAG.getMergeValues({Shuffle, Load.getValue(1)}, DL);
}
if (MGN->getExtensionType() == ISD::NON_EXTLOAD &&
matchIndexAsWiderOp(VT, Index, MGN->getMask(),
MGN->getMemOperand()->getBaseAlign(), Subtarget)) {
SmallVector<SDValue> NewIndices;
for (unsigned i = 0; i < Index->getNumOperands(); i += 2)
NewIndices.push_back(Index.getOperand(i));
EVT IndexVT = Index.getValueType()
.getHalfNumVectorElementsVT(*DAG.getContext());
Index = DAG.getBuildVector(IndexVT, DL, NewIndices);
unsigned ElementSize = VT.getScalarStoreSize();
EVT WideScalarVT = MVT::getIntegerVT(ElementSize * 8 * 2);
auto EltCnt = VT.getVectorElementCount();
assert(EltCnt.isKnownEven() && "Splitting vector, but not in half!");
EVT WideVT = EVT::getVectorVT(*DAG.getContext(), WideScalarVT,
EltCnt.divideCoefficientBy(2));
SDValue Passthru = DAG.getBitcast(WideVT, MGN->getPassThru());
EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
EltCnt.divideCoefficientBy(2));
SDValue Mask = DAG.getSplat(MaskVT, DL, DAG.getConstant(1, DL, MVT::i1));
SDValue Gather =
DAG.getMaskedGather(DAG.getVTList(WideVT, MVT::Other), WideVT, DL,
{MGN->getChain(), Passthru, Mask, MGN->getBasePtr(),
Index, ScaleOp},
MGN->getMemOperand(), IndexType, ISD::NON_EXTLOAD);
SDValue Result = DAG.getBitcast(VT, Gather.getValue(0));
return DAG.getMergeValues({Result, Gather.getValue(1)}, DL);
}
break;
}
case ISD::MSCATTER:{
const auto *MSN = dyn_cast<MaskedScatterSDNode>(N);
SDValue Index = MSN->getIndex();
SDValue ScaleOp = MSN->getScale();
ISD::MemIndexType IndexType = MSN->getIndexType();
assert(!MSN->isIndexScaled() &&
"Scaled gather/scatter should not be formed");
SDLoc DL(N);
if (legalizeScatterGatherIndexType(DL, Index, IndexType, DCI))
return DAG.getMaskedScatter(
N->getVTList(), MSN->getMemoryVT(), DL,
{MSN->getChain(), MSN->getValue(), MSN->getMask(), MSN->getBasePtr(),
Index, ScaleOp},
MSN->getMemOperand(), IndexType, MSN->isTruncatingStore());
if (narrowIndex(Index, IndexType, DAG))
return DAG.getMaskedScatter(
N->getVTList(), MSN->getMemoryVT(), DL,
{MSN->getChain(), MSN->getValue(), MSN->getMask(), MSN->getBasePtr(),
Index, ScaleOp},
MSN->getMemOperand(), IndexType, MSN->isTruncatingStore());
EVT VT = MSN->getValue()->getValueType(0);
SmallVector<int> ShuffleMask;
if (!MSN->isTruncatingStore() &&
matchIndexAsShuffle(VT, Index, MSN->getMask(), ShuffleMask)) {
SDValue Shuffle = DAG.getVectorShuffle(VT, DL, MSN->getValue(),
DAG.getUNDEF(VT), ShuffleMask);
return DAG.getMaskedStore(MSN->getChain(), DL, Shuffle, MSN->getBasePtr(),
DAG.getUNDEF(XLenVT), MSN->getMask(),
MSN->getMemoryVT(), MSN->getMemOperand(),
ISD::UNINDEXED, false);
}
break;
}
case ISD::VP_GATHER: {
const auto *VPGN = dyn_cast<VPGatherSDNode>(N);
SDValue Index = VPGN->getIndex();
SDValue ScaleOp = VPGN->getScale();
ISD::MemIndexType IndexType = VPGN->getIndexType();
assert(!VPGN->isIndexScaled() &&
"Scaled gather/scatter should not be formed");
SDLoc DL(N);
if (legalizeScatterGatherIndexType(DL, Index, IndexType, DCI))
return DAG.getGatherVP(N->getVTList(), VPGN->getMemoryVT(), DL,
{VPGN->getChain(), VPGN->getBasePtr(), Index,
ScaleOp, VPGN->getMask(),
VPGN->getVectorLength()},
VPGN->getMemOperand(), IndexType);
if (narrowIndex(Index, IndexType, DAG))
return DAG.getGatherVP(N->getVTList(), VPGN->getMemoryVT(), DL,
{VPGN->getChain(), VPGN->getBasePtr(), Index,
ScaleOp, VPGN->getMask(),
VPGN->getVectorLength()},
VPGN->getMemOperand(), IndexType);
break;
}
case ISD::VP_SCATTER: {
const auto *VPSN = dyn_cast<VPScatterSDNode>(N);
SDValue Index = VPSN->getIndex();
SDValue ScaleOp = VPSN->getScale();
ISD::MemIndexType IndexType = VPSN->getIndexType();
assert(!VPSN->isIndexScaled() &&
"Scaled gather/scatter should not be formed");
SDLoc DL(N);
if (legalizeScatterGatherIndexType(DL, Index, IndexType, DCI))
return DAG.getScatterVP(N->getVTList(), VPSN->getMemoryVT(), DL,
{VPSN->getChain(), VPSN->getValue(),
VPSN->getBasePtr(), Index, ScaleOp,
VPSN->getMask(), VPSN->getVectorLength()},
VPSN->getMemOperand(), IndexType);
if (narrowIndex(Index, IndexType, DAG))
return DAG.getScatterVP(N->getVTList(), VPSN->getMemoryVT(), DL,
{VPSN->getChain(), VPSN->getValue(),
VPSN->getBasePtr(), Index, ScaleOp,
VPSN->getMask(), VPSN->getVectorLength()},
VPSN->getMemOperand(), IndexType);
break;
}
case RISCVISD::SHL_VL:
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
[[fallthrough]];
case RISCVISD::SRA_VL:
case RISCVISD::SRL_VL: {
SDValue ShAmt = N->getOperand(1);
if (ShAmt.getOpcode() == RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL) {
// We don't need the upper 32 bits of a 64-bit element for a shift amount.
SDLoc DL(N);
SDValue VL = N->getOperand(4);
EVT VT = N->getValueType(0);
ShAmt = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
ShAmt.getOperand(1), VL);
return DAG.getNode(N->getOpcode(), DL, VT, N->getOperand(0), ShAmt,
N->getOperand(2), N->getOperand(3), N->getOperand(4));
}
break;
}
case ISD::SRA:
if (SDValue V = performSRACombine(N, DAG, Subtarget))
return V;
[[fallthrough]];
case ISD::SRL:
case ISD::SHL: {
if (N->getOpcode() == ISD::SHL) {
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
}
SDValue ShAmt = N->getOperand(1);
if (ShAmt.getOpcode() == RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL) {
// We don't need the upper 32 bits of a 64-bit element for a shift amount.
SDLoc DL(N);
EVT VT = N->getValueType(0);
ShAmt = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
ShAmt.getOperand(1),
DAG.getRegister(RISCV::X0, Subtarget.getXLenVT()));
return DAG.getNode(N->getOpcode(), DL, VT, N->getOperand(0), ShAmt);
}
break;
}
case RISCVISD::ADD_VL:
if (SDValue V = combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget))
return V;
return combineToVWMACC(N, DAG, Subtarget);
case RISCVISD::VWADD_W_VL:
case RISCVISD::VWADDU_W_VL:
case RISCVISD::VWSUB_W_VL:
case RISCVISD::VWSUBU_W_VL:
return performVWADDSUBW_VLCombine(N, DCI, Subtarget);
case RISCVISD::SUB_VL:
case RISCVISD::MUL_VL:
return combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget);
case RISCVISD::VFMADD_VL:
case RISCVISD::VFNMADD_VL:
case RISCVISD::VFMSUB_VL:
case RISCVISD::VFNMSUB_VL:
case RISCVISD::STRICT_VFMADD_VL:
case RISCVISD::STRICT_VFNMADD_VL:
case RISCVISD::STRICT_VFMSUB_VL:
case RISCVISD::STRICT_VFNMSUB_VL:
return performVFMADD_VLCombine(N, DAG, Subtarget);
case RISCVISD::FADD_VL:
case RISCVISD::FSUB_VL:
case RISCVISD::FMUL_VL:
case RISCVISD::VFWADD_W_VL:
case RISCVISD::VFWSUB_W_VL: {
if (N->getValueType(0).isScalableVector() &&
N->getValueType(0).getVectorElementType() == MVT::f32 &&
(Subtarget.hasVInstructionsF16Minimal() &&
!Subtarget.hasVInstructionsF16()))
return SDValue();
return combineBinOp_VLToVWBinOp_VL(N, DCI, Subtarget);
}
case ISD::LOAD:
case ISD::STORE: {
if (DCI.isAfterLegalizeDAG())
if (SDValue V = performMemPairCombine(N, DCI))
return V;
if (N->getOpcode() != ISD::STORE)
break;
auto *Store = cast<StoreSDNode>(N);
SDValue Chain = Store->getChain();
EVT MemVT = Store->getMemoryVT();
SDValue Val = Store->getValue();
SDLoc DL(N);
bool IsScalarizable =
MemVT.isFixedLengthVector() && ISD::isNormalStore(Store) &&
Store->isSimple() &&
MemVT.getVectorElementType().bitsLE(Subtarget.getXLenVT()) &&
isPowerOf2_64(MemVT.getSizeInBits()) &&
MemVT.getSizeInBits() <= Subtarget.getXLen();
// If sufficiently aligned we can scalarize stores of constant vectors of
// any power-of-two size up to XLen bits, provided that they aren't too
// expensive to materialize.
// vsetivli zero, 2, e8, m1, ta, ma
// vmv.v.i v8, 4
// vse64.v v8, (a0)
// ->
// li a1, 1028
// sh a1, 0(a0)
if (DCI.isBeforeLegalize() && IsScalarizable &&
ISD::isBuildVectorOfConstantSDNodes(Val.getNode())) {
// Get the constant vector bits
APInt NewC(Val.getValueSizeInBits(), 0);
uint64_t EltSize = Val.getScalarValueSizeInBits();
for (unsigned i = 0; i < Val.getNumOperands(); i++) {
if (Val.getOperand(i).isUndef())
continue;
NewC.insertBits(Val.getConstantOperandAPInt(i).trunc(EltSize),
i * EltSize);
}
MVT NewVT = MVT::getIntegerVT(MemVT.getSizeInBits());
if (RISCVMatInt::getIntMatCost(NewC, Subtarget.getXLen(), Subtarget,
true) <= 2 &&
allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
NewVT, *Store->getMemOperand())) {
SDValue NewV = DAG.getConstant(NewC, DL, NewVT);
return DAG.getStore(Chain, DL, NewV, Store->getBasePtr(),
Store->getPointerInfo(), Store->getOriginalAlign(),
Store->getMemOperand()->getFlags());
}
}
// Similarly, if sufficiently aligned we can scalarize vector copies, e.g.
// vsetivli zero, 2, e16, m1, ta, ma
// vle16.v v8, (a0)
// vse16.v v8, (a1)
if (auto *L = dyn_cast<LoadSDNode>(Val);
L && DCI.isBeforeLegalize() && IsScalarizable && L->isSimple() &&
L->hasNUsesOfValue(1, 0) && L->hasNUsesOfValue(1, 1) &&
Store->getChain() == SDValue(L, 1) && ISD::isNormalLoad(L) &&
L->getMemoryVT() == MemVT) {
MVT NewVT = MVT::getIntegerVT(MemVT.getSizeInBits());
if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
NewVT, *Store->getMemOperand()) &&
allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
NewVT, *L->getMemOperand())) {
SDValue NewL = DAG.getLoad(NewVT, DL, L->getChain(), L->getBasePtr(),
L->getPointerInfo(), L->getOriginalAlign(),
L->getMemOperand()->getFlags());
return DAG.getStore(Chain, DL, NewL, Store->getBasePtr(),
Store->getPointerInfo(), Store->getOriginalAlign(),
Store->getMemOperand()->getFlags());
}
}
// Combine store of vmv.x.s/vfmv.f.s to vse with VL of 1.
// vfmv.f.s is represented as extract element from 0. Match it late to avoid
// any illegal types.
if (Val.getOpcode() == RISCVISD::VMV_X_S ||
(DCI.isAfterLegalizeDAG() &&
Val.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
isNullConstant(Val.getOperand(1)))) {
SDValue Src = Val.getOperand(0);
MVT VecVT = Src.getSimpleValueType();
// VecVT should be scalable and memory VT should match the element type.
if (!Store->isIndexed() && VecVT.isScalableVector() &&
MemVT == VecVT.getVectorElementType()) {
SDLoc DL(N);
MVT MaskVT = getMaskTypeFor(VecVT);
return DAG.getStoreVP(
Store->getChain(), DL, Src, Store->getBasePtr(), Store->getOffset(),
DAG.getConstant(1, DL, MaskVT),
DAG.getConstant(1, DL, Subtarget.getXLenVT()), MemVT,
Store->getMemOperand(), Store->getAddressingMode(),
Store->isTruncatingStore(), /*IsCompress*/ false);
}
}
break;
}
case ISD::SPLAT_VECTOR: {
EVT VT = N->getValueType(0);
// Only perform this combine on legal MVT types.
if (!isTypeLegal(VT))
break;
if (auto Gather = matchSplatAsGather(N->getOperand(0), VT.getSimpleVT(), N,
DAG, Subtarget))
return Gather;
break;
}
case ISD::BUILD_VECTOR:
if (SDValue V = performBUILD_VECTORCombine(N, DAG, Subtarget, *this))
return V;
break;
case ISD::CONCAT_VECTORS:
if (SDValue V = performCONCAT_VECTORSCombine(N, DAG, Subtarget, *this))
return V;
break;
case ISD::INSERT_VECTOR_ELT:
if (SDValue V = performINSERT_VECTOR_ELTCombine(N, DAG, Subtarget, *this))
return V;
break;
case RISCVISD::VFMV_V_F_VL: {
const MVT VT = N->getSimpleValueType(0);
SDValue Passthru = N->getOperand(0);
SDValue Scalar = N->getOperand(1);
SDValue VL = N->getOperand(2);
// If VL is 1, we can use vfmv.s.f.
if (isOneConstant(VL))
return DAG.getNode(RISCVISD::VFMV_S_F_VL, DL, VT, Passthru, Scalar, VL);
break;
}
case RISCVISD::VMV_V_X_VL: {
const MVT VT = N->getSimpleValueType(0);
SDValue Passthru = N->getOperand(0);
SDValue Scalar = N->getOperand(1);
SDValue VL = N->getOperand(2);
// Tail agnostic VMV.V.X only demands the vector element bitwidth from the
// scalar input.
unsigned ScalarSize = Scalar.getValueSizeInBits();
unsigned EltWidth = VT.getScalarSizeInBits();
if (ScalarSize > EltWidth && Passthru.isUndef())
if (SimplifyDemandedLowBitsHelper(1, EltWidth))
return SDValue(N, 0);
// If VL is 1 and the scalar value won't benefit from immediate, we can
// use vmv.s.x.
ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Scalar);
if (isOneConstant(VL) &&
(!Const || Const->isZero() ||
!Const->getAPIntValue().sextOrTrunc(EltWidth).isSignedIntN(5)))
return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT, Passthru, Scalar, VL);
break;
}
case RISCVISD::VFMV_S_F_VL: {
SDValue Src = N->getOperand(1);
// Try to remove vector->scalar->vector if the scalar->vector is inserting
// into an undef vector.
// TODO: Could use a vslide or vmv.v.v for non-undef.
if (N->getOperand(0).isUndef() &&
Src.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
isNullConstant(Src.getOperand(1)) &&
Src.getOperand(0).getValueType().isScalableVector()) {
EVT VT = N->getValueType(0);
EVT SrcVT = Src.getOperand(0).getValueType();
assert(SrcVT.getVectorElementType() == VT.getVectorElementType());
// Widths match, just return the original vector.
if (SrcVT == VT)
return Src.getOperand(0);
// TODO: Use insert_subvector/extract_subvector to change widen/narrow?
}
[[fallthrough]];
}
case RISCVISD::VMV_S_X_VL: {
const MVT VT = N->getSimpleValueType(0);
SDValue Passthru = N->getOperand(0);
SDValue Scalar = N->getOperand(1);
SDValue VL = N->getOperand(2);
// Use M1 or smaller to avoid over constraining register allocation
const MVT M1VT = getLMUL1VT(VT);
if (M1VT.bitsLT(VT)) {
SDValue M1Passthru =
DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, M1VT, Passthru,
DAG.getVectorIdxConstant(0, DL));
SDValue Result =
DAG.getNode(N->getOpcode(), DL, M1VT, M1Passthru, Scalar, VL);
Result = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Passthru, Result,
DAG.getVectorIdxConstant(0, DL));
return Result;
}
// We use a vmv.v.i if possible. We limit this to LMUL1. LMUL2 or
// higher would involve overly constraining the register allocator for
// no purpose.
if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Scalar);
Const && !Const->isZero() && isInt<5>(Const->getSExtValue()) &&
VT.bitsLE(getLMUL1VT(VT)) && Passthru.isUndef())
return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Scalar, VL);
break;
}
case RISCVISD::VMV_X_S: {
SDValue Vec = N->getOperand(0);
MVT VecVT = N->getOperand(0).getSimpleValueType();
const MVT M1VT = getLMUL1VT(VecVT);
if (M1VT.bitsLT(VecVT)) {
Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, M1VT, Vec,
DAG.getVectorIdxConstant(0, DL));
return DAG.getNode(RISCVISD::VMV_X_S, DL, N->getSimpleValueType(0), Vec);
}
break;
}
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN:
case ISD::INTRINSIC_WO_CHAIN: {
unsigned IntOpNo = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1;
unsigned IntNo = N->getConstantOperandVal(IntOpNo);
switch (IntNo) {
// By default we do not combine any intrinsic.
default:
return SDValue();
case Intrinsic::riscv_masked_strided_load: {
MVT VT = N->getSimpleValueType(0);
auto *Load = cast<MemIntrinsicSDNode>(N);
SDValue PassThru = N->getOperand(2);
SDValue Base = N->getOperand(3);
SDValue Stride = N->getOperand(4);
SDValue Mask = N->getOperand(5);
// If the stride is equal to the element size in bytes, we can use
// a masked.load.
const unsigned ElementSize = VT.getScalarStoreSize();
if (auto *StrideC = dyn_cast<ConstantSDNode>(Stride);
StrideC && StrideC->getZExtValue() == ElementSize)
return DAG.getMaskedLoad(VT, DL, Load->getChain(), Base,
DAG.getUNDEF(XLenVT), Mask, PassThru,
Load->getMemoryVT(), Load->getMemOperand(),
ISD::UNINDEXED, ISD::NON_EXTLOAD);
return SDValue();
}
case Intrinsic::riscv_masked_strided_store: {
auto *Store = cast<MemIntrinsicSDNode>(N);
SDValue Value = N->getOperand(2);
SDValue Base = N->getOperand(3);
SDValue Stride = N->getOperand(4);
SDValue Mask = N->getOperand(5);
// If the stride is equal to the element size in bytes, we can use
// a masked.store.
const unsigned ElementSize = Value.getValueType().getScalarStoreSize();
if (auto *StrideC = dyn_cast<ConstantSDNode>(Stride);
StrideC && StrideC->getZExtValue() == ElementSize)
return DAG.getMaskedStore(Store->getChain(), DL, Value, Base,
DAG.getUNDEF(XLenVT), Mask,
Store->getMemoryVT(), Store->getMemOperand(),
ISD::UNINDEXED, false);
return SDValue();
}
case Intrinsic::riscv_vcpop:
case Intrinsic::riscv_vcpop_mask:
case Intrinsic::riscv_vfirst:
case Intrinsic::riscv_vfirst_mask: {
SDValue VL = N->getOperand(2);
if (IntNo == Intrinsic::riscv_vcpop_mask ||
IntNo == Intrinsic::riscv_vfirst_mask)
VL = N->getOperand(3);
if (!isNullConstant(VL))
return SDValue();
// If VL is 0, vcpop -> li 0, vfirst -> li -1.
SDLoc DL(N);
EVT VT = N->getValueType(0);
if (IntNo == Intrinsic::riscv_vfirst ||
IntNo == Intrinsic::riscv_vfirst_mask)
return DAG.getConstant(-1, DL, VT);
return DAG.getConstant(0, DL, VT);
}
}
}
case ISD::BITCAST: {
assert(Subtarget.useRVVForFixedLengthVectors());
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
EVT SrcVT = N0.getValueType();
// If this is a bitcast between a MVT::v4i1/v2i1/v1i1 and an illegal integer
// type, widen both sides to avoid a trip through memory.
if ((SrcVT == MVT::v1i1 || SrcVT == MVT::v2i1 || SrcVT == MVT::v4i1) &&
VT.isScalarInteger()) {
unsigned NumConcats = 8 / SrcVT.getVectorNumElements();
SmallVector<SDValue, 4> Ops(NumConcats, DAG.getUNDEF(SrcVT));
Ops[0] = N0;
SDLoc DL(N);
N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i1, Ops);
N0 = DAG.getBitcast(MVT::i8, N0);
return DAG.getNode(ISD::TRUNCATE, DL, VT, N0);
}
return SDValue();
}
}
return SDValue();
}
bool RISCVTargetLowering::shouldTransformSignedTruncationCheck(
EVT XVT, unsigned KeptBits) const {
// For vectors, we don't have a preference..
if (XVT.isVector())
return false;
if (XVT != MVT::i32 && XVT != MVT::i64)
return false;
// We can use sext.w for RV64 or an srai 31 on RV32.
if (KeptBits == 32 || KeptBits == 64)
return true;
// With Zbb we can use sext.h/sext.b.
return Subtarget.hasStdExtZbb() &&
((KeptBits == 8 && XVT == MVT::i64 && !Subtarget.is64Bit()) ||
KeptBits == 16);
}
bool RISCVTargetLowering::isDesirableToCommuteWithShift(
const SDNode *N, CombineLevel Level) const {
assert((N->getOpcode() == ISD::SHL || N->getOpcode() == ISD::SRA ||
N->getOpcode() == ISD::SRL) &&
"Expected shift op");
// The following folds are only desirable if `(OP _, c1 << c2)` can be
// materialised in fewer instructions than `(OP _, c1)`:
//
// (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
// (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2)
SDValue N0 = N->getOperand(0);
EVT Ty = N0.getValueType();
if (Ty.isScalarInteger() &&
(N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR)) {
auto *C1 = dyn_cast<ConstantSDNode>(N0->getOperand(1));
auto *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (C1 && C2) {
const APInt &C1Int = C1->getAPIntValue();
APInt ShiftedC1Int = C1Int << C2->getAPIntValue();
// We can materialise `c1 << c2` into an add immediate, so it's "free",
// and the combine should happen, to potentially allow further combines
// later.
if (ShiftedC1Int.getSignificantBits() <= 64 &&
isLegalAddImmediate(ShiftedC1Int.getSExtValue()))
return true;
// We can materialise `c1` in an add immediate, so it's "free", and the
// combine should be prevented.
if (C1Int.getSignificantBits() <= 64 &&
isLegalAddImmediate(C1Int.getSExtValue()))
return false;
// Neither constant will fit into an immediate, so find materialisation
// costs.
int C1Cost =
RISCVMatInt::getIntMatCost(C1Int, Ty.getSizeInBits(), Subtarget,
/*CompressionCost*/ true);
int ShiftedC1Cost = RISCVMatInt::getIntMatCost(
ShiftedC1Int, Ty.getSizeInBits(), Subtarget,
/*CompressionCost*/ true);
// Materialising `c1` is cheaper than materialising `c1 << c2`, so the
// combine should be prevented.
if (C1Cost < ShiftedC1Cost)
return false;
}
}
return true;
}
bool RISCVTargetLowering::targetShrinkDemandedConstant(
SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
TargetLoweringOpt &TLO) const {
// Delay this optimization as late as possible.
if (!TLO.LegalOps)
return false;
EVT VT = Op.getValueType();
if (VT.isVector())
return false;
unsigned Opcode = Op.getOpcode();
if (Opcode != ISD::AND && Opcode != ISD::OR && Opcode != ISD::XOR)
return false;
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
if (!C)
return false;
const APInt &Mask = C->getAPIntValue();
// Clear all non-demanded bits initially.
APInt ShrunkMask = Mask & DemandedBits;
// Try to make a smaller immediate by setting undemanded bits.
APInt ExpandedMask = Mask | ~DemandedBits;
auto IsLegalMask = [ShrunkMask, ExpandedMask](const APInt &Mask) -> bool {
return ShrunkMask.isSubsetOf(Mask) && Mask.isSubsetOf(ExpandedMask);
};
auto UseMask = [Mask, Op, &TLO](const APInt &NewMask) -> bool {
if (NewMask == Mask)
return true;
SDLoc DL(Op);
SDValue NewC = TLO.DAG.getConstant(NewMask, DL, Op.getValueType());
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), DL, Op.getValueType(),
Op.getOperand(0), NewC);
return TLO.CombineTo(Op, NewOp);
};
// If the shrunk mask fits in sign extended 12 bits, let the target
// independent code apply it.
if (ShrunkMask.isSignedIntN(12))
return false;
// And has a few special cases for zext.
if (Opcode == ISD::AND) {
// Preserve (and X, 0xffff), if zext.h exists use zext.h,
// otherwise use SLLI + SRLI.
APInt NewMask = APInt(Mask.getBitWidth(), 0xffff);
if (IsLegalMask(NewMask))
return UseMask(NewMask);
// Try to preserve (and X, 0xffffffff), the (zext_inreg X, i32) pattern.
if (VT == MVT::i64) {
APInt NewMask = APInt(64, 0xffffffff);
if (IsLegalMask(NewMask))
return UseMask(NewMask);
}
}
// For the remaining optimizations, we need to be able to make a negative
// number through a combination of mask and undemanded bits.
if (!ExpandedMask.isNegative())
return false;
// What is the fewest number of bits we need to represent the negative number.
unsigned MinSignedBits = ExpandedMask.getSignificantBits();
// Try to make a 12 bit negative immediate. If that fails try to make a 32
// bit negative immediate unless the shrunk immediate already fits in 32 bits.
// If we can't create a simm12, we shouldn't change opaque constants.
APInt NewMask = ShrunkMask;
if (MinSignedBits <= 12)
NewMask.setBitsFrom(11);
else if (!C->isOpaque() && MinSignedBits <= 32 && !ShrunkMask.isSignedIntN(32))
NewMask.setBitsFrom(31);
else
return false;
// Check that our new mask is a subset of the demanded mask.
assert(IsLegalMask(NewMask));
return UseMask(NewMask);
}
static uint64_t computeGREVOrGORC(uint64_t x, unsigned ShAmt, bool IsGORC) {
static const uint64_t GREVMasks[] = {
0x5555555555555555ULL, 0x3333333333333333ULL, 0x0F0F0F0F0F0F0F0FULL,
0x00FF00FF00FF00FFULL, 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL};
for (unsigned Stage = 0; Stage != 6; ++Stage) {
unsigned Shift = 1 << Stage;
if (ShAmt & Shift) {
uint64_t Mask = GREVMasks[Stage];
uint64_t Res = ((x & Mask) << Shift) | ((x >> Shift) & Mask);
if (IsGORC)
Res |= x;
x = Res;
}
}
return x;
}
void RISCVTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
KnownBits &Known,
const APInt &DemandedElts,
const SelectionDAG &DAG,
unsigned Depth) const {
unsigned BitWidth = Known.getBitWidth();
unsigned Opc = Op.getOpcode();
assert((Opc >= ISD::BUILTIN_OP_END ||
Opc == ISD::INTRINSIC_WO_CHAIN ||
Opc == ISD::INTRINSIC_W_CHAIN ||
Opc == ISD::INTRINSIC_VOID) &&
"Should use MaskedValueIsZero if you don't know whether Op"
" is a target node!");
Known.resetAll();
switch (Opc) {
default: break;
case RISCVISD::SELECT_CC: {
Known = DAG.computeKnownBits(Op.getOperand(4), Depth + 1);
// If we don't know any bits, early out.
if (Known.isUnknown())
break;
KnownBits Known2 = DAG.computeKnownBits(Op.getOperand(3), Depth + 1);
// Only known if known in both the LHS and RHS.
Known = Known.intersectWith(Known2);
break;
}
case RISCVISD::CZERO_EQZ:
case RISCVISD::CZERO_NEZ:
Known = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
// Result is either all zero or operand 0. We can propagate zeros, but not
// ones.
Known.One.clearAllBits();
break;
case RISCVISD::REMUW: {
KnownBits Known2;
Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
// We only care about the lower 32 bits.
Known = KnownBits::urem(Known.trunc(32), Known2.trunc(32));
// Restore the original width by sign extending.
Known = Known.sext(BitWidth);
break;
}
case RISCVISD::DIVUW: {
KnownBits Known2;
Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
// We only care about the lower 32 bits.
Known = KnownBits::udiv(Known.trunc(32), Known2.trunc(32));
// Restore the original width by sign extending.
Known = Known.sext(BitWidth);
break;
}
case RISCVISD::SLLW: {
KnownBits Known2;
Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
Known = KnownBits::shl(Known.trunc(32), Known2.trunc(5).zext(32));
// Restore the original width by sign extending.
Known = Known.sext(BitWidth);
break;
}
case RISCVISD::CTZW: {
KnownBits Known2 = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
unsigned PossibleTZ = Known2.trunc(32).countMaxTrailingZeros();
unsigned LowBits = llvm::bit_width(PossibleTZ);
Known.Zero.setBitsFrom(LowBits);
break;
}
case RISCVISD::CLZW: {
KnownBits Known2 = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
unsigned PossibleLZ = Known2.trunc(32).countMaxLeadingZeros();
unsigned LowBits = llvm::bit_width(PossibleLZ);
Known.Zero.setBitsFrom(LowBits);
break;
}
case RISCVISD::BREV8:
case RISCVISD::ORC_B: {
// FIXME: This is based on the non-ratified Zbp GREV and GORC where a
// control value of 7 is equivalent to brev8 and orc.b.
Known = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
bool IsGORC = Op.getOpcode() == RISCVISD::ORC_B;
// To compute zeros, we need to invert the value and invert it back after.
Known.Zero =
~computeGREVOrGORC(~Known.Zero.getZExtValue(), 7, IsGORC);
Known.One = computeGREVOrGORC(Known.One.getZExtValue(), 7, IsGORC);
break;
}
case RISCVISD::READ_VLENB: {
// We can use the minimum and maximum VLEN values to bound VLENB. We
// know VLEN must be a power of two.
const unsigned MinVLenB = Subtarget.getRealMinVLen() / 8;
const unsigned MaxVLenB = Subtarget.getRealMaxVLen() / 8;
assert(MinVLenB > 0 && "READ_VLENB without vector extension enabled?");
Known.Zero.setLowBits(Log2_32(MinVLenB));
Known.Zero.setBitsFrom(Log2_32(MaxVLenB)+1);
if (MaxVLenB == MinVLenB)
Known.One.setBit(Log2_32(MinVLenB));
break;
}
case RISCVISD::FCLASS: {
// fclass will only set one of the low 10 bits.
Known.Zero.setBitsFrom(10);
break;
}
case ISD::INTRINSIC_W_CHAIN:
case ISD::INTRINSIC_WO_CHAIN: {
unsigned IntNo =
Op.getConstantOperandVal(Opc == ISD::INTRINSIC_WO_CHAIN ? 0 : 1);
switch (IntNo) {
default:
// We can't do anything for most intrinsics.
break;
case Intrinsic::riscv_vsetvli:
case Intrinsic::riscv_vsetvlimax: {
bool HasAVL = IntNo == Intrinsic::riscv_vsetvli;
unsigned VSEW = Op.getConstantOperandVal(HasAVL + 1);
RISCVII::VLMUL VLMUL =
static_cast<RISCVII::VLMUL>(Op.getConstantOperandVal(HasAVL + 2));
unsigned SEW = RISCVVType::decodeVSEW(VSEW);
auto [LMul, Fractional] = RISCVVType::decodeVLMUL(VLMUL);
uint64_t MaxVL = Subtarget.getRealMaxVLen() / SEW;
MaxVL = (Fractional) ? MaxVL / LMul : MaxVL * LMul;
// Result of vsetvli must be not larger than AVL.
if (HasAVL && isa<ConstantSDNode>(Op.getOperand(1)))
MaxVL = std::min(MaxVL, Op.getConstantOperandVal(1));
unsigned KnownZeroFirstBit = Log2_32(MaxVL) + 1;
if (BitWidth > KnownZeroFirstBit)
Known.Zero.setBitsFrom(KnownZeroFirstBit);
break;
}
}
break;
}
}
}
unsigned RISCVTargetLowering::ComputeNumSignBitsForTargetNode(
SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
unsigned Depth) const {
switch (Op.getOpcode()) {
default:
break;
case RISCVISD::SELECT_CC: {
unsigned Tmp =
DAG.ComputeNumSignBits(Op.getOperand(3), DemandedElts, Depth + 1);
if (Tmp == 1) return 1; // Early out.
unsigned Tmp2 =
DAG.ComputeNumSignBits(Op.getOperand(4), DemandedElts, Depth + 1);
return std::min(Tmp, Tmp2);
}
case RISCVISD::CZERO_EQZ:
case RISCVISD::CZERO_NEZ:
// Output is either all zero or operand 0. We can propagate sign bit count
// from operand 0.
return DAG.ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
case RISCVISD::ABSW: {
// We expand this at isel to negw+max. The result will have 33 sign bits
// if the input has at least 33 sign bits.
unsigned Tmp =
DAG.ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
if (Tmp < 33) return 1;
return 33;
}
case RISCVISD::SLLW:
case RISCVISD::SRAW:
case RISCVISD::SRLW:
case RISCVISD::DIVW:
case RISCVISD::DIVUW:
case RISCVISD::REMUW:
case RISCVISD::ROLW:
case RISCVISD::RORW:
case RISCVISD::FCVT_W_RV64:
case RISCVISD::FCVT_WU_RV64:
case RISCVISD::STRICT_FCVT_W_RV64:
case RISCVISD::STRICT_FCVT_WU_RV64:
// TODO: As the result is sign-extended, this is conservatively correct. A
// more precise answer could be calculated for SRAW depending on known
// bits in the shift amount.
return 33;
case RISCVISD::VMV_X_S: {
// The number of sign bits of the scalar result is computed by obtaining the
// element type of the input vector operand, subtracting its width from the
// XLEN, and then adding one (sign bit within the element type). If the
// element type is wider than XLen, the least-significant XLEN bits are
// taken.
unsigned XLen = Subtarget.getXLen();
unsigned EltBits = Op.getOperand(0).getScalarValueSizeInBits();
if (EltBits <= XLen)
return XLen - EltBits + 1;
break;
}
case ISD::INTRINSIC_W_CHAIN: {
unsigned IntNo = Op.getConstantOperandVal(1);
switch (IntNo) {
default:
break;
case Intrinsic::riscv_masked_atomicrmw_xchg_i64:
case Intrinsic::riscv_masked_atomicrmw_add_i64:
case Intrinsic::riscv_masked_atomicrmw_sub_i64:
case Intrinsic::riscv_masked_atomicrmw_nand_i64:
case Intrinsic::riscv_masked_atomicrmw_max_i64:
case Intrinsic::riscv_masked_atomicrmw_min_i64:
case Intrinsic::riscv_masked_atomicrmw_umax_i64:
case Intrinsic::riscv_masked_atomicrmw_umin_i64:
case Intrinsic::riscv_masked_cmpxchg_i64:
// riscv_masked_{atomicrmw_*,cmpxchg} intrinsics represent an emulated
// narrow atomic operation. These are implemented using atomic
// operations at the minimum supported atomicrmw/cmpxchg width whose
// result is then sign extended to XLEN. With +A, the minimum width is
// 32 for both 64 and 32.
assert(Subtarget.getXLen() == 64);
assert(getMinCmpXchgSizeInBits() == 32);
assert(Subtarget.hasStdExtA());
return 33;
}
break;
}
}
return 1;
}
bool RISCVTargetLowering::canCreateUndefOrPoisonForTargetNode(
SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
bool PoisonOnly, bool ConsiderFlags, unsigned Depth) const {
// TODO: Add more target nodes.
switch (Op.getOpcode()) {
case RISCVISD::SELECT_CC:
// Integer select_cc cannot create poison.
// TODO: What are the FP poison semantics?
// TODO: This instruction blocks poison from the unselected operand, can
// we do anything with that?
return !Op.getValueType().isInteger();
}
return TargetLowering::canCreateUndefOrPoisonForTargetNode(
Op, DemandedElts, DAG, PoisonOnly, ConsiderFlags, Depth);
}
const Constant *
RISCVTargetLowering::getTargetConstantFromLoad(LoadSDNode *Ld) const {
assert(Ld && "Unexpected null LoadSDNode");
if (!ISD::isNormalLoad(Ld))
return nullptr;
SDValue Ptr = Ld->getBasePtr();
// Only constant pools with no offset are supported.
auto GetSupportedConstantPool = [](SDValue Ptr) -> ConstantPoolSDNode * {
auto *CNode = dyn_cast<ConstantPoolSDNode>(Ptr);
if (!CNode || CNode->isMachineConstantPoolEntry() ||
CNode->getOffset() != 0)
return nullptr;
return CNode;
};
// Simple case, LLA.
if (Ptr.getOpcode() == RISCVISD::LLA) {
auto *CNode = GetSupportedConstantPool(Ptr);
if (!CNode || CNode->getTargetFlags() != 0)
return nullptr;
return CNode->getConstVal();
}
// Look for a HI and ADD_LO pair.
if (Ptr.getOpcode() != RISCVISD::ADD_LO ||
Ptr.getOperand(0).getOpcode() != RISCVISD::HI)
return nullptr;
auto *CNodeLo = GetSupportedConstantPool(Ptr.getOperand(1));
auto *CNodeHi = GetSupportedConstantPool(Ptr.getOperand(0).getOperand(0));
if (!CNodeLo || CNodeLo->getTargetFlags() != RISCVII::MO_LO ||
!CNodeHi || CNodeHi->getTargetFlags() != RISCVII::MO_HI)
return nullptr;
if (CNodeLo->getConstVal() != CNodeHi->getConstVal())
return nullptr;
return CNodeLo->getConstVal();
}
static MachineBasicBlock *emitReadCounterWidePseudo(MachineInstr &MI,
MachineBasicBlock *BB) {
assert(MI.getOpcode() == RISCV::ReadCounterWide && "Unexpected instruction");
// To read a 64-bit counter CSR on a 32-bit target, we read the two halves.
// Should the count have wrapped while it was being read, we need to try
// again.
// For example:
// ```
// read:
// csrrs x3, counterh # load high word of counter
// csrrs x2, counter # load low word of counter
// csrrs x4, counterh # load high word of counter
// bne x3, x4, read # check if high word reads match, otherwise try again
// ```
MachineFunction &MF = *BB->getParent();
const BasicBlock *LLVMBB = BB->getBasicBlock();
MachineFunction::iterator It = ++BB->getIterator();
MachineBasicBlock *LoopMBB = MF.CreateMachineBasicBlock(LLVMBB);
MF.insert(It, LoopMBB);
MachineBasicBlock *DoneMBB = MF.CreateMachineBasicBlock(LLVMBB);
MF.insert(It, DoneMBB);
// Transfer the remainder of BB and its successor edges to DoneMBB.
DoneMBB->splice(DoneMBB->begin(), BB,
std::next(MachineBasicBlock::iterator(MI)), BB->end());
DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
BB->addSuccessor(LoopMBB);
MachineRegisterInfo &RegInfo = MF.getRegInfo();
Register ReadAgainReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
Register LoReg = MI.getOperand(0).getReg();
Register HiReg = MI.getOperand(1).getReg();
int64_t LoCounter = MI.getOperand(2).getImm();
int64_t HiCounter = MI.getOperand(3).getImm();
DebugLoc DL = MI.getDebugLoc();
const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), HiReg)
.addImm(HiCounter)
.addReg(RISCV::X0);
BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), LoReg)
.addImm(LoCounter)
.addReg(RISCV::X0);
BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), ReadAgainReg)
.addImm(HiCounter)
.addReg(RISCV::X0);
BuildMI(LoopMBB, DL, TII->get(RISCV::BNE))
.addReg(HiReg)
.addReg(ReadAgainReg)
.addMBB(LoopMBB);
LoopMBB->addSuccessor(LoopMBB);
LoopMBB->addSuccessor(DoneMBB);
MI.eraseFromParent();
return DoneMBB;
}
static MachineBasicBlock *emitSplitF64Pseudo(MachineInstr &MI,
MachineBasicBlock *BB,
const RISCVSubtarget &Subtarget) {
assert(MI.getOpcode() == RISCV::SplitF64Pseudo && "Unexpected instruction");
MachineFunction &MF = *BB->getParent();
DebugLoc DL = MI.getDebugLoc();
const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
Register LoReg = MI.getOperand(0).getReg();
Register HiReg = MI.getOperand(1).getReg();
Register SrcReg = MI.getOperand(2).getReg();
const TargetRegisterClass *SrcRC = &RISCV::FPR64RegClass;
int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex(MF);
TII.storeRegToStackSlot(*BB, MI, SrcReg, MI.getOperand(2).isKill(), FI, SrcRC,
RI, Register());
MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
MachineMemOperand *MMOLo =
MF.getMachineMemOperand(MPI, MachineMemOperand::MOLoad, 4, Align(8));
MachineMemOperand *MMOHi = MF.getMachineMemOperand(
MPI.getWithOffset(4), MachineMemOperand::MOLoad, 4, Align(8));
BuildMI(*BB, MI, DL, TII.get(RISCV::LW), LoReg)
.addFrameIndex(FI)
.addImm(0)
.addMemOperand(MMOLo);
BuildMI(*BB, MI, DL, TII.get(RISCV::LW), HiReg)
.addFrameIndex(FI)
.addImm(4)
.addMemOperand(MMOHi);
MI.eraseFromParent(); // The pseudo instruction is gone now.
return BB;
}
static MachineBasicBlock *emitBuildPairF64Pseudo(MachineInstr &MI,
MachineBasicBlock *BB,
const RISCVSubtarget &Subtarget) {
assert(MI.getOpcode() == RISCV::BuildPairF64Pseudo &&
"Unexpected instruction");
MachineFunction &MF = *BB->getParent();
DebugLoc DL = MI.getDebugLoc();
const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
Register DstReg = MI.getOperand(0).getReg();
Register LoReg = MI.getOperand(1).getReg();
Register HiReg = MI.getOperand(2).getReg();
const TargetRegisterClass *DstRC = &RISCV::FPR64RegClass;
int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex(MF);
MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
MachineMemOperand *MMOLo =
MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, Align(8));
MachineMemOperand *MMOHi = MF.getMachineMemOperand(
MPI.getWithOffset(4), MachineMemOperand::MOStore, 4, Align(8));
BuildMI(*BB, MI, DL, TII.get(RISCV::SW))
.addReg(LoReg, getKillRegState(MI.getOperand(1).isKill()))
.addFrameIndex(FI)
.addImm(0)
.addMemOperand(MMOLo);
BuildMI(*BB, MI, DL, TII.get(RISCV::SW))
.addReg(HiReg, getKillRegState(MI.getOperand(2).isKill()))
.addFrameIndex(FI)
.addImm(4)
.addMemOperand(MMOHi);
TII.loadRegFromStackSlot(*BB, MI, DstReg, FI, DstRC, RI, Register());
MI.eraseFromParent(); // The pseudo instruction is gone now.
return BB;
}
static bool isSelectPseudo(MachineInstr &MI) {
switch (MI.getOpcode()) {
default:
return false;
case RISCV::Select_GPR_Using_CC_GPR:
case RISCV::Select_FPR16_Using_CC_GPR:
case RISCV::Select_FPR16INX_Using_CC_GPR:
case RISCV::Select_FPR32_Using_CC_GPR:
case RISCV::Select_FPR32INX_Using_CC_GPR:
case RISCV::Select_FPR64_Using_CC_GPR:
case RISCV::Select_FPR64INX_Using_CC_GPR:
case RISCV::Select_FPR64IN32X_Using_CC_GPR:
return true;
}
}
static MachineBasicBlock *emitQuietFCMP(MachineInstr &MI, MachineBasicBlock *BB,
unsigned RelOpcode, unsigned EqOpcode,
const RISCVSubtarget &Subtarget) {
DebugLoc DL = MI.getDebugLoc();
Register DstReg = MI.getOperand(0).getReg();
Register Src1Reg = MI.getOperand(1).getReg();
Register Src2Reg = MI.getOperand(2).getReg();
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
Register SavedFFlags = MRI.createVirtualRegister(&RISCV::GPRRegClass);
const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
// Save the current FFLAGS.
BuildMI(*BB, MI, DL, TII.get(RISCV::ReadFFLAGS), SavedFFlags);
auto MIB = BuildMI(*BB, MI, DL, TII.get(RelOpcode), DstReg)
.addReg(Src1Reg)
.addReg(Src2Reg);
if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
// Restore the FFLAGS.
BuildMI(*BB, MI, DL, TII.get(RISCV::WriteFFLAGS))
.addReg(SavedFFlags, RegState::Kill);
// Issue a dummy FEQ opcode to raise exception for signaling NaNs.
auto MIB2 = BuildMI(*BB, MI, DL, TII.get(EqOpcode), RISCV::X0)
.addReg(Src1Reg, getKillRegState(MI.getOperand(1).isKill()))
.addReg(Src2Reg, getKillRegState(MI.getOperand(2).isKill()));
if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
MIB2->setFlag(MachineInstr::MIFlag::NoFPExcept);
// Erase the pseudoinstruction.
MI.eraseFromParent();
return BB;
}
static MachineBasicBlock *
EmitLoweredCascadedSelect(MachineInstr &First, MachineInstr &Second,
MachineBasicBlock *ThisMBB,
const RISCVSubtarget &Subtarget) {
// Select_FPRX_ (rs1, rs2, imm, rs4, (Select_FPRX_ rs1, rs2, imm, rs4, rs5)
// Without this, custom-inserter would have generated:
//
// A
// | \
// | B
// | /
// C
// | \
// | D
// | /
// E
//
// A: X = ...; Y = ...
// B: empty
// C: Z = PHI [X, A], [Y, B]
// D: empty
// E: PHI [X, C], [Z, D]
//
// If we lower both Select_FPRX_ in a single step, we can instead generate:
//
// A
// | \
// | C
// | /|
// |/ |
// | |
// | D
// | /
// E
//
// A: X = ...; Y = ...
// D: empty
// E: PHI [X, A], [X, C], [Y, D]
const RISCVInstrInfo &TII = *Subtarget.getInstrInfo();
const DebugLoc &DL = First.getDebugLoc();
const BasicBlock *LLVM_BB = ThisMBB->getBasicBlock();
MachineFunction *F = ThisMBB->getParent();
MachineBasicBlock *FirstMBB = F->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *SecondMBB = F->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *SinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
MachineFunction::iterator It = ++ThisMBB->getIterator();
F->insert(It, FirstMBB);
F->insert(It, SecondMBB);
F->insert(It, SinkMBB);
// Transfer the remainder of ThisMBB and its successor edges to SinkMBB.
SinkMBB->splice(SinkMBB->begin(), ThisMBB,
std::next(MachineBasicBlock::iterator(First)),
ThisMBB->end());
SinkMBB->transferSuccessorsAndUpdatePHIs(ThisMBB);
// Fallthrough block for ThisMBB.
ThisMBB->addSuccessor(FirstMBB);
// Fallthrough block for FirstMBB.
FirstMBB->addSuccessor(SecondMBB);
ThisMBB->addSuccessor(SinkMBB);
FirstMBB->addSuccessor(SinkMBB);
// This is fallthrough.
SecondMBB->addSuccessor(SinkMBB);
auto FirstCC = static_cast<RISCVCC::CondCode>(First.getOperand(3).getImm());
Register FLHS = First.getOperand(1).getReg();
Register FRHS = First.getOperand(2).getReg();
// Insert appropriate branch.
BuildMI(FirstMBB, DL, TII.getBrCond(FirstCC))
.addReg(FLHS)
.addReg(FRHS)
.addMBB(SinkMBB);
Register SLHS = Second.getOperand(1).getReg();
Register SRHS = Second.getOperand(2).getReg();
Register Op1Reg4 = First.getOperand(4).getReg();
Register Op1Reg5 = First.getOperand(5).getReg();
auto SecondCC = static_cast<RISCVCC::CondCode>(Second.getOperand(3).getImm());
// Insert appropriate branch.
BuildMI(ThisMBB, DL, TII.getBrCond(SecondCC))
.addReg(SLHS)
.addReg(SRHS)
.addMBB(SinkMBB);
Register DestReg = Second.getOperand(0).getReg();
Register Op2Reg4 = Second.getOperand(4).getReg();
BuildMI(*SinkMBB, SinkMBB->begin(), DL, TII.get(RISCV::PHI), DestReg)
.addReg(Op2Reg4)
.addMBB(ThisMBB)
.addReg(Op1Reg4)
.addMBB(FirstMBB)
.addReg(Op1Reg5)
.addMBB(SecondMBB);
// Now remove the Select_FPRX_s.
First.eraseFromParent();
Second.eraseFromParent();
return SinkMBB;
}
static MachineBasicBlock *emitSelectPseudo(MachineInstr &MI,
MachineBasicBlock *BB,
const RISCVSubtarget &Subtarget) {
// To "insert" Select_* instructions, we actually have to insert the triangle
// control-flow pattern. The incoming instructions know the destination vreg
// to set, the condition code register to branch on, the true/false values to
// select between, and the condcode to use to select the appropriate branch.
//
// We produce the following control flow:
// HeadMBB
// | \
// | IfFalseMBB
// | /
// TailMBB
//
// When we find a sequence of selects we attempt to optimize their emission
// by sharing the control flow. Currently we only handle cases where we have
// multiple selects with the exact same condition (same LHS, RHS and CC).
// The selects may be interleaved with other instructions if the other
// instructions meet some requirements we deem safe:
// - They are not pseudo instructions.
// - They are debug instructions. Otherwise,
// - They do not have side-effects, do not access memory and their inputs do
// not depend on the results of the select pseudo-instructions.
// The TrueV/FalseV operands of the selects cannot depend on the result of
// previous selects in the sequence.
// These conditions could be further relaxed. See the X86 target for a
// related approach and more information.
//
// Select_FPRX_ (rs1, rs2, imm, rs4, (Select_FPRX_ rs1, rs2, imm, rs4, rs5))
// is checked here and handled by a separate function -
// EmitLoweredCascadedSelect.
Register LHS = MI.getOperand(1).getReg();
Register RHS = MI.getOperand(2).getReg();
auto CC = static_cast<RISCVCC::CondCode>(MI.getOperand(3).getImm());
SmallVector<MachineInstr *, 4> SelectDebugValues;
SmallSet<Register, 4> SelectDests;
SelectDests.insert(MI.getOperand(0).getReg());
MachineInstr *LastSelectPseudo = &MI;
auto Next = next_nodbg(MI.getIterator(), BB->instr_end());
if (MI.getOpcode() != RISCV::Select_GPR_Using_CC_GPR && Next != BB->end() &&
Next->getOpcode() == MI.getOpcode() &&
Next->getOperand(5).getReg() == MI.getOperand(0).getReg() &&
Next->getOperand(5).isKill()) {
return EmitLoweredCascadedSelect(MI, *Next, BB, Subtarget);
}
for (auto E = BB->end(), SequenceMBBI = MachineBasicBlock::iterator(MI);
SequenceMBBI != E; ++SequenceMBBI) {
if (SequenceMBBI->isDebugInstr())
continue;
if (isSelectPseudo(*SequenceMBBI)) {
if (SequenceMBBI->getOperand(1).getReg() != LHS ||
SequenceMBBI->getOperand(2).getReg() != RHS ||
SequenceMBBI->getOperand(3).getImm() != CC ||
SelectDests.count(SequenceMBBI->getOperand(4).getReg()) ||
SelectDests.count(SequenceMBBI->getOperand(5).getReg()))
break;
LastSelectPseudo = &*SequenceMBBI;
SequenceMBBI->collectDebugValues(SelectDebugValues);
SelectDests.insert(SequenceMBBI->getOperand(0).getReg());
continue;
}
if (SequenceMBBI->hasUnmodeledSideEffects() ||
SequenceMBBI->mayLoadOrStore() ||
SequenceMBBI->usesCustomInsertionHook())
break;
if (llvm::any_of(SequenceMBBI->operands(), [&](MachineOperand &MO) {
return MO.isReg() && MO.isUse() && SelectDests.count(MO.getReg());
}))
break;
}
const RISCVInstrInfo &TII = *Subtarget.getInstrInfo();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
DebugLoc DL = MI.getDebugLoc();
MachineFunction::iterator I = ++BB->getIterator();
MachineBasicBlock *HeadMBB = BB;
MachineFunction *F = BB->getParent();
MachineBasicBlock *TailMBB = F->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *IfFalseMBB = F->CreateMachineBasicBlock(LLVM_BB);
F->insert(I, IfFalseMBB);
F->insert(I, TailMBB);
// Transfer debug instructions associated with the selects to TailMBB.
for (MachineInstr *DebugInstr : SelectDebugValues) {
TailMBB->push_back(DebugInstr->removeFromParent());
}
// Move all instructions after the sequence to TailMBB.
TailMBB->splice(TailMBB->end(), HeadMBB,
std::next(LastSelectPseudo->getIterator()), HeadMBB->end());
// Update machine-CFG edges by transferring all successors of the current
// block to the new block which will contain the Phi nodes for the selects.
TailMBB->transferSuccessorsAndUpdatePHIs(HeadMBB);
// Set the successors for HeadMBB.
HeadMBB->addSuccessor(IfFalseMBB);
HeadMBB->addSuccessor(TailMBB);
// Insert appropriate branch.
BuildMI(HeadMBB, DL, TII.getBrCond(CC))
.addReg(LHS)
.addReg(RHS)
.addMBB(TailMBB);
// IfFalseMBB just falls through to TailMBB.
IfFalseMBB->addSuccessor(TailMBB);
// Create PHIs for all of the select pseudo-instructions.
auto SelectMBBI = MI.getIterator();
auto SelectEnd = std::next(LastSelectPseudo->getIterator());
auto InsertionPoint = TailMBB->begin();
while (SelectMBBI != SelectEnd) {
auto Next = std::next(SelectMBBI);
if (isSelectPseudo(*SelectMBBI)) {
// %Result = phi [ %TrueValue, HeadMBB ], [ %FalseValue, IfFalseMBB ]
BuildMI(*TailMBB, InsertionPoint, SelectMBBI->getDebugLoc(),
TII.get(RISCV::PHI), SelectMBBI->getOperand(0).getReg())
.addReg(SelectMBBI->getOperand(4).getReg())
.addMBB(HeadMBB)
.addReg(SelectMBBI->getOperand(5).getReg())
.addMBB(IfFalseMBB);
SelectMBBI->eraseFromParent();
}
SelectMBBI = Next;
}
F->getProperties().reset(MachineFunctionProperties::Property::NoPHIs);
return TailMBB;
}
static MachineBasicBlock *emitVFROUND_NOEXCEPT_MASK(MachineInstr &MI,
MachineBasicBlock *BB,
unsigned CVTXOpc) {
DebugLoc DL = MI.getDebugLoc();
const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
Register SavedFFLAGS = MRI.createVirtualRegister(&RISCV::GPRRegClass);
// Save the old value of FFLAGS.
BuildMI(*BB, MI, DL, TII.get(RISCV::ReadFFLAGS), SavedFFLAGS);
assert(MI.getNumOperands() == 7);
// Emit a VFCVT_X_F
const TargetRegisterInfo *TRI =
BB->getParent()->getSubtarget().getRegisterInfo();
const TargetRegisterClass *RC = MI.getRegClassConstraint(0, &TII, TRI);
Register Tmp = MRI.createVirtualRegister(RC);
BuildMI(*BB, MI, DL, TII.get(CVTXOpc), Tmp)
.add(MI.getOperand(1))
.add(MI.getOperand(2))
.add(MI.getOperand(3))
.add(MachineOperand::CreateImm(7)) // frm = DYN
.add(MI.getOperand(4))
.add(MI.getOperand(5))
.add(MI.getOperand(6))
.add(MachineOperand::CreateReg(RISCV::FRM,
/*IsDef*/ false,
/*IsImp*/ true));
// Emit a VFCVT_F_X
RISCVII::VLMUL LMul = RISCVII::getLMul(MI.getDesc().TSFlags);
unsigned Log2SEW = MI.getOperand(RISCVII::getSEWOpNum(MI.getDesc())).getImm();
// There is no E8 variant for VFCVT_F_X.
assert(Log2SEW >= 4);
// Since MI (VFROUND) isn't SEW specific, we cannot use a macro to make
// handling of different (LMUL, SEW) pairs easier because we need to pull the
// SEW immediate from MI, and that information is not avaliable during macro
// expansion.
unsigned CVTFOpc;
if (Log2SEW == 4) {
switch (LMul) {
case RISCVII::LMUL_1:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M1_E16_MASK;
break;
case RISCVII::LMUL_2:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M2_E16_MASK;
break;
case RISCVII::LMUL_4:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M4_E16_MASK;
break;
case RISCVII::LMUL_8:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M8_E16_MASK;
break;
case RISCVII::LMUL_F2:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_MF2_E16_MASK;
break;
case RISCVII::LMUL_F4:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_MF4_E16_MASK;
break;
case RISCVII::LMUL_F8:
case RISCVII::LMUL_RESERVED:
llvm_unreachable("Unexpected LMUL and SEW combination value for MI.");
}
} else if (Log2SEW == 5) {
switch (LMul) {
case RISCVII::LMUL_1:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M1_E32_MASK;
break;
case RISCVII::LMUL_2:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M2_E32_MASK;
break;
case RISCVII::LMUL_4:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M4_E32_MASK;
break;
case RISCVII::LMUL_8:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M8_E32_MASK;
break;
case RISCVII::LMUL_F2:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_MF2_E32_MASK;
break;
case RISCVII::LMUL_F4:
case RISCVII::LMUL_F8:
case RISCVII::LMUL_RESERVED:
llvm_unreachable("Unexpected LMUL and SEW combination value for MI.");
}
} else if (Log2SEW == 6) {
switch (LMul) {
case RISCVII::LMUL_1:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M1_E64_MASK;
break;
case RISCVII::LMUL_2:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M2_E64_MASK;
break;
case RISCVII::LMUL_4:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M4_E64_MASK;
break;
case RISCVII::LMUL_8:
CVTFOpc = RISCV::PseudoVFCVT_F_X_V_M8_E64_MASK;
break;
case RISCVII::LMUL_F2:
case RISCVII::LMUL_F4:
case RISCVII::LMUL_F8:
case RISCVII::LMUL_RESERVED:
llvm_unreachable("Unexpected LMUL and SEW combination value for MI.");
}
} else {
llvm_unreachable("Unexpected LMUL and SEW combination value for MI.");
}
BuildMI(*BB, MI, DL, TII.get(CVTFOpc))
.add(MI.getOperand(0))
.add(MI.getOperand(1))
.addReg(Tmp)
.add(MI.getOperand(3))
.add(MachineOperand::CreateImm(7)) // frm = DYN
.add(MI.getOperand(4))
.add(MI.getOperand(5))
.add(MI.getOperand(6))
.add(MachineOperand::CreateReg(RISCV::FRM,
/*IsDef*/ false,
/*IsImp*/ true));
// Restore FFLAGS.
BuildMI(*BB, MI, DL, TII.get(RISCV::WriteFFLAGS))
.addReg(SavedFFLAGS, RegState::Kill);
// Erase the pseudoinstruction.
MI.eraseFromParent();
return BB;
}
static MachineBasicBlock *emitFROUND(MachineInstr &MI, MachineBasicBlock *MBB,
const RISCVSubtarget &Subtarget) {
unsigned CmpOpc, F2IOpc, I2FOpc, FSGNJOpc, FSGNJXOpc;
const TargetRegisterClass *RC;
switch (MI.getOpcode()) {
default:
llvm_unreachable("Unexpected opcode");
case RISCV::PseudoFROUND_H:
CmpOpc = RISCV::FLT_H;
F2IOpc = RISCV::FCVT_W_H;
I2FOpc = RISCV::FCVT_H_W;
FSGNJOpc = RISCV::FSGNJ_H;
FSGNJXOpc = RISCV::FSGNJX_H;
RC = &RISCV::FPR16RegClass;
break;
case RISCV::PseudoFROUND_H_INX:
CmpOpc = RISCV::FLT_H_INX;
F2IOpc = RISCV::FCVT_W_H_INX;
I2FOpc = RISCV::FCVT_H_W_INX;
FSGNJOpc = RISCV::FSGNJ_H_INX;
FSGNJXOpc = RISCV::FSGNJX_H_INX;
RC = &RISCV::GPRF16RegClass;
break;
case RISCV::PseudoFROUND_S:
CmpOpc = RISCV::FLT_S;
F2IOpc = RISCV::FCVT_W_S;
I2FOpc = RISCV::FCVT_S_W;
FSGNJOpc = RISCV::FSGNJ_S;
FSGNJXOpc = RISCV::FSGNJX_S;
RC = &RISCV::FPR32RegClass;
break;
case RISCV::PseudoFROUND_S_INX:
CmpOpc = RISCV::FLT_S_INX;
F2IOpc = RISCV::FCVT_W_S_INX;
I2FOpc = RISCV::FCVT_S_W_INX;
FSGNJOpc = RISCV::FSGNJ_S_INX;
FSGNJXOpc = RISCV::FSGNJX_S_INX;
RC = &RISCV::GPRF32RegClass;
break;
case RISCV::PseudoFROUND_D:
assert(Subtarget.is64Bit() && "Expected 64-bit GPR.");
CmpOpc = RISCV::FLT_D;
F2IOpc = RISCV::FCVT_L_D;
I2FOpc = RISCV::FCVT_D_L;
FSGNJOpc = RISCV::FSGNJ_D;
FSGNJXOpc = RISCV::FSGNJX_D;
RC = &RISCV::FPR64RegClass;
break;
case RISCV::PseudoFROUND_D_INX:
assert(Subtarget.is64Bit() && "Expected 64-bit GPR.");
CmpOpc = RISCV::FLT_D_INX;
F2IOpc = RISCV::FCVT_L_D_INX;
I2FOpc = RISCV::FCVT_D_L_INX;
FSGNJOpc = RISCV::FSGNJ_D_INX;
FSGNJXOpc = RISCV::FSGNJX_D_INX;
RC = &RISCV::GPRRegClass;
break;
}
const BasicBlock *BB = MBB->getBasicBlock();
DebugLoc DL = MI.getDebugLoc();
MachineFunction::iterator I = ++MBB->getIterator();
MachineFunction *F = MBB->getParent();
MachineBasicBlock *CvtMBB = F->CreateMachineBasicBlock(BB);
MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB);
F->insert(I, CvtMBB);
F->insert(I, DoneMBB);
// Move all instructions after the sequence to DoneMBB.
DoneMBB->splice(DoneMBB->end(), MBB, MachineBasicBlock::iterator(MI),
MBB->end());
// Update machine-CFG edges by transferring all successors of the current
// block to the new block which will contain the Phi nodes for the selects.
DoneMBB->transferSuccessorsAndUpdatePHIs(MBB);
// Set the successors for MBB.
MBB->addSuccessor(CvtMBB);
MBB->addSuccessor(DoneMBB);
Register DstReg = MI.getOperand(0).getReg();
Register SrcReg = MI.getOperand(1).getReg();
Register MaxReg = MI.getOperand(2).getReg();
int64_t FRM = MI.getOperand(3).getImm();
const RISCVInstrInfo &TII = *Subtarget.getInstrInfo();
MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
Register FabsReg = MRI.createVirtualRegister(RC);
BuildMI(MBB, DL, TII.get(FSGNJXOpc), FabsReg).addReg(SrcReg).addReg(SrcReg);
// Compare the FP value to the max value.
Register CmpReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
auto MIB =
BuildMI(MBB, DL, TII.get(CmpOpc), CmpReg).addReg(FabsReg).addReg(MaxReg);
if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
// Insert branch.
BuildMI(MBB, DL, TII.get(RISCV::BEQ))
.addReg(CmpReg)
.addReg(RISCV::X0)
.addMBB(DoneMBB);
CvtMBB->addSuccessor(DoneMBB);
// Convert to integer.
Register F2IReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
MIB = BuildMI(CvtMBB, DL, TII.get(F2IOpc), F2IReg).addReg(SrcReg).addImm(FRM);
if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
// Convert back to FP.
Register I2FReg = MRI.createVirtualRegister(RC);
MIB = BuildMI(CvtMBB, DL, TII.get(I2FOpc), I2FReg).addReg(F2IReg).addImm(FRM);
if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
// Restore the sign bit.
Register CvtReg = MRI.createVirtualRegister(RC);
BuildMI(CvtMBB, DL, TII.get(FSGNJOpc), CvtReg).addReg(I2FReg).addReg(SrcReg);
// Merge the results.
BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(RISCV::PHI), DstReg)
.addReg(SrcReg)
.addMBB(MBB)
.addReg(CvtReg)
.addMBB(CvtMBB);
MI.eraseFromParent();
return DoneMBB;
}
MachineBasicBlock *
RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
MachineBasicBlock *BB) const {
switch (MI.getOpcode()) {
default:
llvm_unreachable("Unexpected instr type to insert");
case RISCV::ReadCounterWide:
assert(!Subtarget.is64Bit() &&
"ReadCounterWide is only to be used on riscv32");
return emitReadCounterWidePseudo(MI, BB);
case RISCV::Select_GPR_Using_CC_GPR:
case RISCV::Select_FPR16_Using_CC_GPR:
case RISCV::Select_FPR16INX_Using_CC_GPR:
case RISCV::Select_FPR32_Using_CC_GPR:
case RISCV::Select_FPR32INX_Using_CC_GPR:
case RISCV::Select_FPR64_Using_CC_GPR:
case RISCV::Select_FPR64INX_Using_CC_GPR:
case RISCV::Select_FPR64IN32X_Using_CC_GPR:
return emitSelectPseudo(MI, BB, Subtarget);
case RISCV::BuildPairF64Pseudo:
return emitBuildPairF64Pseudo(MI, BB, Subtarget);
case RISCV::SplitF64Pseudo:
return emitSplitF64Pseudo(MI, BB, Subtarget);
case RISCV::PseudoQuietFLE_H:
return emitQuietFCMP(MI, BB, RISCV::FLE_H, RISCV::FEQ_H, Subtarget);
case RISCV::PseudoQuietFLE_H_INX:
return emitQuietFCMP(MI, BB, RISCV::FLE_H_INX, RISCV::FEQ_H_INX, Subtarget);
case RISCV::PseudoQuietFLT_H:
return emitQuietFCMP(MI, BB, RISCV::FLT_H, RISCV::FEQ_H, Subtarget);
case RISCV::PseudoQuietFLT_H_INX:
return emitQuietFCMP(MI, BB, RISCV::FLT_H_INX, RISCV::FEQ_H_INX, Subtarget);
case RISCV::PseudoQuietFLE_S:
return emitQuietFCMP(MI, BB, RISCV::FLE_S, RISCV::FEQ_S, Subtarget);
case RISCV::PseudoQuietFLE_S_INX:
return emitQuietFCMP(MI, BB, RISCV::FLE_S_INX, RISCV::FEQ_S_INX, Subtarget);
case RISCV::PseudoQuietFLT_S:
return emitQuietFCMP(MI, BB, RISCV::FLT_S, RISCV::FEQ_S, Subtarget);
case RISCV::PseudoQuietFLT_S_INX:
return emitQuietFCMP(MI, BB, RISCV::FLT_S_INX, RISCV::FEQ_S_INX, Subtarget);
case RISCV::PseudoQuietFLE_D:
return emitQuietFCMP(MI, BB, RISCV::FLE_D, RISCV::FEQ_D, Subtarget);
case RISCV::PseudoQuietFLE_D_INX:
return emitQuietFCMP(MI, BB, RISCV::FLE_D_INX, RISCV::FEQ_D_INX, Subtarget);
case RISCV::PseudoQuietFLE_D_IN32X:
return emitQuietFCMP(MI, BB, RISCV::FLE_D_IN32X, RISCV::FEQ_D_IN32X,
Subtarget);
case RISCV::PseudoQuietFLT_D:
return emitQuietFCMP(MI, BB, RISCV::FLT_D, RISCV::FEQ_D, Subtarget);
case RISCV::PseudoQuietFLT_D_INX:
return emitQuietFCMP(MI, BB, RISCV::FLT_D_INX, RISCV::FEQ_D_INX, Subtarget);
case RISCV::PseudoQuietFLT_D_IN32X:
return emitQuietFCMP(MI, BB, RISCV::FLT_D_IN32X, RISCV::FEQ_D_IN32X,
Subtarget);
case RISCV::PseudoVFROUND_NOEXCEPT_V_M1_MASK:
return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M1_MASK);
case RISCV::PseudoVFROUND_NOEXCEPT_V_M2_MASK:
return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M2_MASK);
case RISCV::PseudoVFROUND_NOEXCEPT_V_M4_MASK:
return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M4_MASK);
case RISCV::PseudoVFROUND_NOEXCEPT_V_M8_MASK:
return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M8_MASK);
case RISCV::PseudoVFROUND_NOEXCEPT_V_MF2_MASK:
return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_MF2_MASK);
case RISCV::PseudoVFROUND_NOEXCEPT_V_MF4_MASK:
return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_MF4_MASK);
case RISCV::PseudoFROUND_H:
case RISCV::PseudoFROUND_H_INX:
case RISCV::PseudoFROUND_S:
case RISCV::PseudoFROUND_S_INX:
case RISCV::PseudoFROUND_D:
case RISCV::PseudoFROUND_D_INX:
case RISCV::PseudoFROUND_D_IN32X:
return emitFROUND(MI, BB, Subtarget);
case TargetOpcode::STATEPOINT:
// STATEPOINT is a pseudo instruction which has no implicit defs/uses
// while jal call instruction (where statepoint will be lowered at the end)
// has implicit def. This def is early-clobber as it will be set at
// the moment of the call and earlier than any use is read.
// Add this implicit dead def here as a workaround.
MI.addOperand(*MI.getMF(),
MachineOperand::CreateReg(
RISCV::X1, /*isDef*/ true,
/*isImp*/ true, /*isKill*/ false, /*isDead*/ true,
/*isUndef*/ false, /*isEarlyClobber*/ true));
[[fallthrough]];
case TargetOpcode::STACKMAP:
case TargetOpcode::PATCHPOINT:
if (!Subtarget.is64Bit())
report_fatal_error("STACKMAP, PATCHPOINT and STATEPOINT are only "
"supported on 64-bit targets");
return emitPatchPoint(MI, BB);
}
}
void RISCVTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
SDNode *Node) const {
// Add FRM dependency to any instructions with dynamic rounding mode.
int Idx = RISCV::getNamedOperandIdx(MI.getOpcode(), RISCV::OpName::frm);
if (Idx < 0) {
// Vector pseudos have FRM index indicated by TSFlags.
Idx = RISCVII::getFRMOpNum(MI.getDesc());
if (Idx < 0)
return;
}
if (MI.getOperand(Idx).getImm() != RISCVFPRndMode::DYN)
return;
// If the instruction already reads FRM, don't add another read.
if (MI.readsRegister(RISCV::FRM))
return;
MI.addOperand(
MachineOperand::CreateReg(RISCV::FRM, /*isDef*/ false, /*isImp*/ true));
}
// Calling Convention Implementation.
// The expectations for frontend ABI lowering vary from target to target.
// Ideally, an LLVM frontend would be able to avoid worrying about many ABI
// details, but this is a longer term goal. For now, we simply try to keep the
// role of the frontend as simple and well-defined as possible. The rules can
// be summarised as:
// * Never split up large scalar arguments. We handle them here.
// * If a hardfloat calling convention is being used, and the struct may be
// passed in a pair of registers (fp+fp, int+fp), and both registers are
// available, then pass as two separate arguments. If either the GPRs or FPRs
// are exhausted, then pass according to the rule below.
// * If a struct could never be passed in registers or directly in a stack
// slot (as it is larger than 2*XLEN and the floating point rules don't
// apply), then pass it using a pointer with the byval attribute.
// * If a struct is less than 2*XLEN, then coerce to either a two-element
// word-sized array or a 2*XLEN scalar (depending on alignment).
// * The frontend can determine whether a struct is returned by reference or
// not based on its size and fields. If it will be returned by reference, the
// frontend must modify the prototype so a pointer with the sret annotation is
// passed as the first argument. This is not necessary for large scalar
// returns.
// * Struct return values and varargs should be coerced to structs containing
// register-size fields in the same situations they would be for fixed
// arguments.
static const MCPhysReg ArgFPR16s[] = {
RISCV::F10_H, RISCV::F11_H, RISCV::F12_H, RISCV::F13_H,
RISCV::F14_H, RISCV::F15_H, RISCV::F16_H, RISCV::F17_H
};
static const MCPhysReg ArgFPR32s[] = {
RISCV::F10_F, RISCV::F11_F, RISCV::F12_F, RISCV::F13_F,
RISCV::F14_F, RISCV::F15_F, RISCV::F16_F, RISCV::F17_F
};
static const MCPhysReg ArgFPR64s[] = {
RISCV::F10_D, RISCV::F11_D, RISCV::F12_D, RISCV::F13_D,
RISCV::F14_D, RISCV::F15_D, RISCV::F16_D, RISCV::F17_D
};
// This is an interim calling convention and it may be changed in the future.
static const MCPhysReg ArgVRs[] = {
RISCV::V8, RISCV::V9, RISCV::V10, RISCV::V11, RISCV::V12, RISCV::V13,
RISCV::V14, RISCV::V15, RISCV::V16, RISCV::V17, RISCV::V18, RISCV::V19,
RISCV::V20, RISCV::V21, RISCV::V22, RISCV::V23};
static const MCPhysReg ArgVRM2s[] = {RISCV::V8M2, RISCV::V10M2, RISCV::V12M2,
RISCV::V14M2, RISCV::V16M2, RISCV::V18M2,
RISCV::V20M2, RISCV::V22M2};
static const MCPhysReg ArgVRM4s[] = {RISCV::V8M4, RISCV::V12M4, RISCV::V16M4,
RISCV::V20M4};
static const MCPhysReg ArgVRM8s[] = {RISCV::V8M8, RISCV::V16M8};
ArrayRef<MCPhysReg> RISCV::getArgGPRs(const RISCVABI::ABI ABI) {
// The GPRs used for passing arguments in the ILP32* and LP64* ABIs, except
// the ILP32E ABI.
static const MCPhysReg ArgIGPRs[] = {RISCV::X10, RISCV::X11, RISCV::X12,
RISCV::X13, RISCV::X14, RISCV::X15,
RISCV::X16, RISCV::X17};
// The GPRs used for passing arguments in the ILP32E/ILP64E ABI.
static const MCPhysReg ArgEGPRs[] = {RISCV::X10, RISCV::X11, RISCV::X12,
RISCV::X13, RISCV::X14, RISCV::X15};
if (ABI == RISCVABI::ABI_ILP32E || ABI == RISCVABI::ABI_LP64E)
return ArrayRef(ArgEGPRs);
return ArrayRef(ArgIGPRs);
}
static ArrayRef<MCPhysReg> getFastCCArgGPRs(const RISCVABI::ABI ABI) {
// The GPRs used for passing arguments in the FastCC, X5 and X6 might be used
// for save-restore libcall, so we don't use them.
static const MCPhysReg FastCCIGPRs[] = {
RISCV::X10, RISCV::X11, RISCV::X12, RISCV::X13, RISCV::X14,
RISCV::X15, RISCV::X16, RISCV::X17, RISCV::X7, RISCV::X28,
RISCV::X29, RISCV::X30, RISCV::X31};
// The GPRs used for passing arguments in the FastCC when using ILP32E/ILP64E.
static const MCPhysReg FastCCEGPRs[] = {RISCV::X10, RISCV::X11, RISCV::X12,
RISCV::X13, RISCV::X14, RISCV::X15,
RISCV::X7};
if (ABI == RISCVABI::ABI_ILP32E || ABI == RISCVABI::ABI_LP64E)
return ArrayRef(FastCCEGPRs);
return ArrayRef(FastCCIGPRs);
}
// Pass a 2*XLEN argument that has been split into two XLEN values through
// registers or the stack as necessary.
static bool CC_RISCVAssign2XLen(unsigned XLen, CCState &State, CCValAssign VA1,
ISD::ArgFlagsTy ArgFlags1, unsigned ValNo2,
MVT ValVT2, MVT LocVT2,
ISD::ArgFlagsTy ArgFlags2, bool EABI) {
unsigned XLenInBytes = XLen / 8;
const RISCVSubtarget &STI =
State.getMachineFunction().getSubtarget<RISCVSubtarget>();
ArrayRef<MCPhysReg> ArgGPRs = RISCV::getArgGPRs(STI.getTargetABI());
if (Register Reg = State.AllocateReg(ArgGPRs)) {
// At least one half can be passed via register.
State.addLoc(CCValAssign::getReg(VA1.getValNo(), VA1.getValVT(), Reg,
VA1.getLocVT(), CCValAssign::Full));
} else {
// Both halves must be passed on the stack, with proper alignment.
// TODO: To be compatible with GCC's behaviors, we force them to have 4-byte
// alignment. This behavior may be changed when RV32E/ILP32E is ratified.
Align StackAlign(XLenInBytes);
if (!EABI || XLen != 32)
StackAlign = std::max(StackAlign, ArgFlags1.getNonZeroOrigAlign());
State.addLoc(
CCValAssign::getMem(VA1.getValNo(), VA1.getValVT(),
State.AllocateStack(XLenInBytes, StackAlign),
VA1.getLocVT(), CCValAssign::Full));
State.addLoc(CCValAssign::getMem(
ValNo2, ValVT2, State.AllocateStack(XLenInBytes, Align(XLenInBytes)),
LocVT2, CCValAssign::Full));
return false;
}
if (Register Reg = State.AllocateReg(ArgGPRs)) {
// The second half can also be passed via register.
State.addLoc(
CCValAssign::getReg(ValNo2, ValVT2, Reg, LocVT2, CCValAssign::Full));
} else {
// The second half is passed via the stack, without additional alignment.
State.addLoc(CCValAssign::getMem(
ValNo2, ValVT2, State.AllocateStack(XLenInBytes, Align(XLenInBytes)),
LocVT2, CCValAssign::Full));
}
return false;
}
// Implements the RISC-V calling convention. Returns true upon failure.
bool RISCV::CC_RISCV(const DataLayout &DL, RISCVABI::ABI ABI, unsigned ValNo,
MVT ValVT, MVT LocVT, CCValAssign::LocInfo LocInfo,
ISD::ArgFlagsTy ArgFlags, CCState &State, bool IsFixed,
bool IsRet, Type *OrigTy, const RISCVTargetLowering &TLI,
RVVArgDispatcher &RVVDispatcher) {
unsigned XLen = DL.getLargestLegalIntTypeSizeInBits();
assert(XLen == 32 || XLen == 64);
MVT XLenVT = XLen == 32 ? MVT::i32 : MVT::i64;
// Static chain parameter must not be passed in normal argument registers,
// so we assign t2 for it as done in GCC's __builtin_call_with_static_chain
if (ArgFlags.isNest()) {
if (unsigned Reg = State.AllocateReg(RISCV::X7)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
// Any return value split in to more than two values can't be returned
// directly. Vectors are returned via the available vector registers.
if (!LocVT.isVector() && IsRet && ValNo > 1)
return true;
// UseGPRForF16_F32 if targeting one of the soft-float ABIs, if passing a
// variadic argument, or if no F16/F32 argument registers are available.
bool UseGPRForF16_F32 = true;
// UseGPRForF64 if targeting soft-float ABIs or an FLEN=32 ABI, if passing a
// variadic argument, or if no F64 argument registers are available.
bool UseGPRForF64 = true;
switch (ABI) {
default:
llvm_unreachable("Unexpected ABI");
case RISCVABI::ABI_ILP32:
case RISCVABI::ABI_ILP32E:
case RISCVABI::ABI_LP64:
case RISCVABI::ABI_LP64E:
break;
case RISCVABI::ABI_ILP32F:
case RISCVABI::ABI_LP64F:
UseGPRForF16_F32 = !IsFixed;
break;
case RISCVABI::ABI_ILP32D:
case RISCVABI::ABI_LP64D:
UseGPRForF16_F32 = !IsFixed;
UseGPRForF64 = !IsFixed;
break;
}
// FPR16, FPR32, and FPR64 alias each other.
if (State.getFirstUnallocated(ArgFPR32s) == std::size(ArgFPR32s)) {
UseGPRForF16_F32 = true;
UseGPRForF64 = true;
}
// From this point on, rely on UseGPRForF16_F32, UseGPRForF64 and
// similar local variables rather than directly checking against the target
// ABI.
if (UseGPRForF16_F32 &&
(ValVT == MVT::f16 || ValVT == MVT::bf16 || ValVT == MVT::f32)) {
LocVT = XLenVT;
LocInfo = CCValAssign::BCvt;
} else if (UseGPRForF64 && XLen == 64 && ValVT == MVT::f64) {
LocVT = MVT::i64;
LocInfo = CCValAssign::BCvt;
}
ArrayRef<MCPhysReg> ArgGPRs = RISCV::getArgGPRs(ABI);
// If this is a variadic argument, the RISC-V calling convention requires
// that it is assigned an 'even' or 'aligned' register if it has 8-byte
// alignment (RV32) or 16-byte alignment (RV64). An aligned register should
// be used regardless of whether the original argument was split during
// legalisation or not. The argument will not be passed by registers if the
// original type is larger than 2*XLEN, so the register alignment rule does
// not apply.
// TODO: To be compatible with GCC's behaviors, we don't align registers
// currently if we are using ILP32E calling convention. This behavior may be
// changed when RV32E/ILP32E is ratified.
unsigned TwoXLenInBytes = (2 * XLen) / 8;
if (!IsFixed && ArgFlags.getNonZeroOrigAlign() == TwoXLenInBytes &&
DL.getTypeAllocSize(OrigTy) == TwoXLenInBytes &&
ABI != RISCVABI::ABI_ILP32E) {
unsigned RegIdx = State.getFirstUnallocated(ArgGPRs);
// Skip 'odd' register if necessary.
if (RegIdx != std::size(ArgGPRs) && RegIdx % 2 == 1)
State.AllocateReg(ArgGPRs);
}
SmallVectorImpl<CCValAssign> &PendingLocs = State.getPendingLocs();
SmallVectorImpl<ISD::ArgFlagsTy> &PendingArgFlags =
State.getPendingArgFlags();
assert(PendingLocs.size() == PendingArgFlags.size() &&
"PendingLocs and PendingArgFlags out of sync");
// Handle passing f64 on RV32D with a soft float ABI or when floating point
// registers are exhausted.
if (UseGPRForF64 && XLen == 32 && ValVT == MVT::f64) {
assert(PendingLocs.empty() && "Can't lower f64 if it is split");
// Depending on available argument GPRS, f64 may be passed in a pair of
// GPRs, split between a GPR and the stack, or passed completely on the
// stack. LowerCall/LowerFormalArguments/LowerReturn must recognise these
// cases.
Register Reg = State.AllocateReg(ArgGPRs);
if (!Reg) {
unsigned StackOffset = State.AllocateStack(8, Align(8));
State.addLoc(
CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
return false;
}
LocVT = MVT::i32;
State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
Register HiReg = State.AllocateReg(ArgGPRs);
if (HiReg) {
State.addLoc(
CCValAssign::getCustomReg(ValNo, ValVT, HiReg, LocVT, LocInfo));
} else {
unsigned StackOffset = State.AllocateStack(4, Align(4));
State.addLoc(
CCValAssign::getCustomMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
}
return false;
}
// Fixed-length vectors are located in the corresponding scalable-vector
// container types.
if (ValVT.isFixedLengthVector())
LocVT = TLI.getContainerForFixedLengthVector(LocVT);
// Split arguments might be passed indirectly, so keep track of the pending
// values. Split vectors are passed via a mix of registers and indirectly, so
// treat them as we would any other argument.
if (ValVT.isScalarInteger() && (ArgFlags.isSplit() || !PendingLocs.empty())) {
LocVT = XLenVT;
LocInfo = CCValAssign::Indirect;
PendingLocs.push_back(
CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));
PendingArgFlags.push_back(ArgFlags);
if (!ArgFlags.isSplitEnd()) {
return false;
}
}
// If the split argument only had two elements, it should be passed directly
// in registers or on the stack.
if (ValVT.isScalarInteger() && ArgFlags.isSplitEnd() &&
PendingLocs.size() <= 2) {
assert(PendingLocs.size() == 2 && "Unexpected PendingLocs.size()");
// Apply the normal calling convention rules to the first half of the
// split argument.
CCValAssign VA = PendingLocs[0];
ISD::ArgFlagsTy AF = PendingArgFlags[0];
PendingLocs.clear();
PendingArgFlags.clear();
return CC_RISCVAssign2XLen(
XLen, State, VA, AF, ValNo, ValVT, LocVT, ArgFlags,
ABI == RISCVABI::ABI_ILP32E || ABI == RISCVABI::ABI_LP64E);
}
// Allocate to a register if possible, or else a stack slot.
Register Reg;
unsigned StoreSizeBytes = XLen / 8;
Align StackAlign = Align(XLen / 8);
if ((ValVT == MVT::f16 || ValVT == MVT::bf16) && !UseGPRForF16_F32)
Reg = State.AllocateReg(ArgFPR16s);
else if (ValVT == MVT::f32 && !UseGPRForF16_F32)
Reg = State.AllocateReg(ArgFPR32s);
else if (ValVT == MVT::f64 && !UseGPRForF64)
Reg = State.AllocateReg(ArgFPR64s);
else if (ValVT.isVector()) {
Reg = RVVDispatcher.getNextPhysReg();
if (!Reg) {
// For return values, the vector must be passed fully via registers or
// via the stack.
// FIXME: The proposed vector ABI only mandates v8-v15 for return values,
// but we're using all of them.
if (IsRet)
return true;
// Try using a GPR to pass the address
if ((Reg = State.AllocateReg(ArgGPRs))) {
LocVT = XLenVT;
LocInfo = CCValAssign::Indirect;
} else if (ValVT.isScalableVector()) {
LocVT = XLenVT;
LocInfo = CCValAssign::Indirect;
} else {
// Pass fixed-length vectors on the stack.
LocVT = ValVT;
StoreSizeBytes = ValVT.getStoreSize();
// Align vectors to their element sizes, being careful for vXi1
// vectors.
StackAlign = MaybeAlign(ValVT.getScalarSizeInBits() / 8).valueOrOne();
}
}
} else {
Reg = State.AllocateReg(ArgGPRs);
}
unsigned StackOffset =
Reg ? 0 : State.AllocateStack(StoreSizeBytes, StackAlign);
// If we reach this point and PendingLocs is non-empty, we must be at the
// end of a split argument that must be passed indirectly.
if (!PendingLocs.empty()) {
assert(ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()");
assert(PendingLocs.size() > 2 && "Unexpected PendingLocs.size()");
for (auto &It : PendingLocs) {
if (Reg)
It.convertToReg(Reg);
else
It.convertToMem(StackOffset);
State.addLoc(It);
}
PendingLocs.clear();
PendingArgFlags.clear();
return false;
}
assert((!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT ||
(TLI.getSubtarget().hasVInstructions() && ValVT.isVector())) &&
"Expected an XLenVT or vector types at this stage");
if (Reg) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
// When a scalar floating-point value is passed on the stack, no
// bit-conversion is needed.
if (ValVT.isFloatingPoint() && LocInfo != CCValAssign::Indirect) {
assert(!ValVT.isVector());
LocVT = ValVT;
LocInfo = CCValAssign::Full;
}
State.addLoc(CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
return false;
}
template <typename ArgTy>
static std::optional<unsigned> preAssignMask(const ArgTy &Args) {
for (const auto &ArgIdx : enumerate(Args)) {
MVT ArgVT = ArgIdx.value().VT;
if (ArgVT.isVector() && ArgVT.getVectorElementType() == MVT::i1)
return ArgIdx.index();
}
return std::nullopt;
}
void RISCVTargetLowering::analyzeInputArgs(
MachineFunction &MF, CCState &CCInfo,
const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet,
RISCVCCAssignFn Fn) const {
unsigned NumArgs = Ins.size();
FunctionType *FType = MF.getFunction().getFunctionType();
RVVArgDispatcher Dispatcher;
if (IsRet) {
Dispatcher = RVVArgDispatcher{&MF, this, ArrayRef(Ins)};
} else {
SmallVector<Type *, 4> TypeList;
for (const Argument &Arg : MF.getFunction().args())
TypeList.push_back(Arg.getType());
Dispatcher = RVVArgDispatcher{&MF, this, ArrayRef(TypeList)};
}
for (unsigned i = 0; i != NumArgs; ++i) {
MVT ArgVT = Ins[i].VT;
ISD::ArgFlagsTy ArgFlags = Ins[i].Flags;
Type *ArgTy = nullptr;
if (IsRet)
ArgTy = FType->getReturnType();
else if (Ins[i].isOrigArg())
ArgTy = FType->getParamType(Ins[i].getOrigArgIndex());
RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
if (Fn(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full,
ArgFlags, CCInfo, /*IsFixed=*/true, IsRet, ArgTy, *this,
Dispatcher)) {
LLVM_DEBUG(dbgs() << "InputArg #" << i << " has unhandled type "
<< ArgVT << '\n');
llvm_unreachable(nullptr);
}
}
}
void RISCVTargetLowering::analyzeOutputArgs(
MachineFunction &MF, CCState &CCInfo,
const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet,
CallLoweringInfo *CLI, RISCVCCAssignFn Fn) const {
unsigned NumArgs = Outs.size();
SmallVector<Type *, 4> TypeList;
if (IsRet)
TypeList.push_back(MF.getFunction().getReturnType());
else if (CLI)
for (const TargetLowering::ArgListEntry &Arg : CLI->getArgs())
TypeList.push_back(Arg.Ty);
RVVArgDispatcher Dispatcher{&MF, this, ArrayRef(TypeList)};
for (unsigned i = 0; i != NumArgs; i++) {
MVT ArgVT = Outs[i].VT;
ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
Type *OrigTy = CLI ? CLI->getArgs()[Outs[i].OrigArgIndex].Ty : nullptr;
RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
if (Fn(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full,
ArgFlags, CCInfo, Outs[i].IsFixed, IsRet, OrigTy, *this,
Dispatcher)) {
LLVM_DEBUG(dbgs() << "OutputArg #" << i << " has unhandled type "
<< ArgVT << "\n");
llvm_unreachable(nullptr);
}
}
}
// Convert Val to a ValVT. Should not be called for CCValAssign::Indirect
// values.
static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDValue Val,
const CCValAssign &VA, const SDLoc &DL,
const RISCVSubtarget &Subtarget) {
switch (VA.getLocInfo()) {
default:
llvm_unreachable("Unexpected CCValAssign::LocInfo");
case CCValAssign::Full:
if (VA.getValVT().isFixedLengthVector() && VA.getLocVT().isScalableVector())
Val = convertFromScalableVector(VA.getValVT(), Val, DAG, Subtarget);
break;
case CCValAssign::BCvt:
if (VA.getLocVT().isInteger() &&
(VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16)) {
Val = DAG.getNode(RISCVISD::FMV_H_X, DL, VA.getValVT(), Val);
} else if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) {
if (RV64LegalI32) {
Val = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Val);
Val = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
} else {
Val = DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, Val);
}
} else {
Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
}
break;
}
return Val;
}
// The caller is responsible for loading the full value if the argument is
// passed with CCValAssign::Indirect.
static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain,
const CCValAssign &VA, const SDLoc &DL,
const ISD::InputArg &In,
const RISCVTargetLowering &TLI) {
MachineFunction &MF = DAG.getMachineFunction();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
EVT LocVT = VA.getLocVT();
SDValue Val;
const TargetRegisterClass *RC = TLI.getRegClassFor(LocVT.getSimpleVT());
Register VReg = RegInfo.createVirtualRegister(RC);
RegInfo.addLiveIn(VA.getLocReg(), VReg);
Val = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
// If input is sign extended from 32 bits, note it for the SExtWRemoval pass.
if (In.isOrigArg()) {
Argument *OrigArg = MF.getFunction().getArg(In.getOrigArgIndex());
if (OrigArg->getType()->isIntegerTy()) {
unsigned BitWidth = OrigArg->getType()->getIntegerBitWidth();
// An input zero extended from i31 can also be considered sign extended.
if ((BitWidth <= 32 && In.Flags.isSExt()) ||
(BitWidth < 32 && In.Flags.isZExt())) {
RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
RVFI->addSExt32Register(VReg);
}
}
}
if (VA.getLocInfo() == CCValAssign::Indirect)
return Val;
return convertLocVTToValVT(DAG, Val, VA, DL, TLI.getSubtarget());
}
static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDValue Val,
const CCValAssign &VA, const SDLoc &DL,
const RISCVSubtarget &Subtarget) {
EVT LocVT = VA.getLocVT();
switch (VA.getLocInfo()) {
default:
llvm_unreachable("Unexpected CCValAssign::LocInfo");
case CCValAssign::Full:
if (VA.getValVT().isFixedLengthVector() && LocVT.isScalableVector())
Val = convertToScalableVector(LocVT, Val, DAG, Subtarget);
break;
case CCValAssign::BCvt:
if (LocVT.isInteger() &&
(VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16)) {
Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, LocVT, Val);
} else if (LocVT == MVT::i64 && VA.getValVT() == MVT::f32) {
if (RV64LegalI32) {
Val = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Val);
Val = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Val);
} else {
Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Val);
}
} else {
Val = DAG.getNode(ISD::BITCAST, DL, LocVT, Val);
}
break;
}
return Val;
}
// The caller is responsible for loading the full value if the argument is
// passed with CCValAssign::Indirect.
static SDValue unpackFromMemLoc(SelectionDAG &DAG, SDValue Chain,
const CCValAssign &VA, const SDLoc &DL) {
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo &MFI = MF.getFrameInfo();
EVT LocVT = VA.getLocVT();
EVT ValVT = VA.getValVT();
EVT PtrVT = MVT::getIntegerVT(DAG.getDataLayout().getPointerSizeInBits(0));
if (ValVT.isScalableVector()) {
// When the value is a scalable vector, we save the pointer which points to
// the scalable vector value in the stack. The ValVT will be the pointer
// type, instead of the scalable vector type.
ValVT = LocVT;
}
int FI = MFI.CreateFixedObject(ValVT.getStoreSize(), VA.getLocMemOffset(),
/*IsImmutable=*/true);
SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
SDValue Val;
ISD::LoadExtType ExtType;
switch (VA.getLocInfo()) {
default:
llvm_unreachable("Unexpected CCValAssign::LocInfo");
case CCValAssign::Full:
case CCValAssign::Indirect:
case CCValAssign::BCvt:
ExtType = ISD::NON_EXTLOAD;
break;
}
Val = DAG.getExtLoad(
ExtType, DL, LocVT, Chain, FIN,
MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), ValVT);
return Val;
}
static SDValue unpackF64OnRV32DSoftABI(SelectionDAG &DAG, SDValue Chain,
const CCValAssign &VA,
const CCValAssign &HiVA,
const SDLoc &DL) {
assert(VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64 &&
"Unexpected VA");
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo &MFI = MF.getFrameInfo();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
assert(VA.isRegLoc() && "Expected register VA assignment");
Register LoVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
RegInfo.addLiveIn(VA.getLocReg(), LoVReg);
SDValue Lo = DAG.getCopyFromReg(Chain, DL, LoVReg, MVT::i32);
SDValue Hi;
if (HiVA.isMemLoc()) {
// Second half of f64 is passed on the stack.
int FI = MFI.CreateFixedObject(4, HiVA.getLocMemOffset(),
/*IsImmutable=*/true);
SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
Hi = DAG.getLoad(MVT::i32, DL, Chain, FIN,
MachinePointerInfo::getFixedStack(MF, FI));
} else {
// Second half of f64 is passed in another GPR.
Register HiVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
RegInfo.addLiveIn(HiVA.getLocReg(), HiVReg);
Hi = DAG.getCopyFromReg(Chain, DL, HiVReg, MVT::i32);
}
return DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
}
// FastCC has less than 1% performance improvement for some particular
// benchmark. But theoretically, it may has benenfit for some cases.
bool RISCV::CC_RISCV_FastCC(const DataLayout &DL, RISCVABI::ABI ABI,
unsigned ValNo, MVT ValVT, MVT LocVT,
CCValAssign::LocInfo LocInfo,
ISD::ArgFlagsTy ArgFlags, CCState &State,
bool IsFixed, bool IsRet, Type *OrigTy,
const RISCVTargetLowering &TLI,
RVVArgDispatcher &RVVDispatcher) {
if (LocVT == MVT::i32 || LocVT == MVT::i64) {
if (unsigned Reg = State.AllocateReg(getFastCCArgGPRs(ABI))) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
const RISCVSubtarget &Subtarget = TLI.getSubtarget();
if (LocVT == MVT::f16 &&
(Subtarget.hasStdExtZfh() || Subtarget.hasStdExtZfhmin())) {
static const MCPhysReg FPR16List[] = {
RISCV::F10_H, RISCV::F11_H, RISCV::F12_H, RISCV::F13_H, RISCV::F14_H,
RISCV::F15_H, RISCV::F16_H, RISCV::F17_H, RISCV::F0_H, RISCV::F1_H,
RISCV::F2_H, RISCV::F3_H, RISCV::F4_H, RISCV::F5_H, RISCV::F6_H,
RISCV::F7_H, RISCV::F28_H, RISCV::F29_H, RISCV::F30_H, RISCV::F31_H};
if (unsigned Reg = State.AllocateReg(FPR16List)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
if (LocVT == MVT::f32 && Subtarget.hasStdExtF()) {
static const MCPhysReg FPR32List[] = {
RISCV::F10_F, RISCV::F11_F, RISCV::F12_F, RISCV::F13_F, RISCV::F14_F,
RISCV::F15_F, RISCV::F16_F, RISCV::F17_F, RISCV::F0_F, RISCV::F1_F,
RISCV::F2_F, RISCV::F3_F, RISCV::F4_F, RISCV::F5_F, RISCV::F6_F,
RISCV::F7_F, RISCV::F28_F, RISCV::F29_F, RISCV::F30_F, RISCV::F31_F};
if (unsigned Reg = State.AllocateReg(FPR32List)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
if (LocVT == MVT::f64 && Subtarget.hasStdExtD()) {
static const MCPhysReg FPR64List[] = {
RISCV::F10_D, RISCV::F11_D, RISCV::F12_D, RISCV::F13_D, RISCV::F14_D,
RISCV::F15_D, RISCV::F16_D, RISCV::F17_D, RISCV::F0_D, RISCV::F1_D,
RISCV::F2_D, RISCV::F3_D, RISCV::F4_D, RISCV::F5_D, RISCV::F6_D,
RISCV::F7_D, RISCV::F28_D, RISCV::F29_D, RISCV::F30_D, RISCV::F31_D};
if (unsigned Reg = State.AllocateReg(FPR64List)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
// Check if there is an available GPR before hitting the stack.
if ((LocVT == MVT::f16 &&
(Subtarget.hasStdExtZhinx() || Subtarget.hasStdExtZhinxmin())) ||
(LocVT == MVT::f32 && Subtarget.hasStdExtZfinx()) ||
(LocVT == MVT::f64 && Subtarget.is64Bit() &&
Subtarget.hasStdExtZdinx())) {
if (unsigned Reg = State.AllocateReg(getFastCCArgGPRs(ABI))) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
if (LocVT == MVT::f16) {
unsigned Offset2 = State.AllocateStack(2, Align(2));
State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset2, LocVT, LocInfo));
return false;
}
if (LocVT == MVT::i32 || LocVT == MVT::f32) {
unsigned Offset4 = State.AllocateStack(4, Align(4));
State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset4, LocVT, LocInfo));
return false;
}
if (LocVT == MVT::i64 || LocVT == MVT::f64) {
unsigned Offset5 = State.AllocateStack(8, Align(8));
State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset5, LocVT, LocInfo));
return false;
}
if (LocVT.isVector()) {
MCPhysReg AllocatedVReg = RVVDispatcher.getNextPhysReg();
if (AllocatedVReg) {
// Fixed-length vectors are located in the corresponding scalable-vector
// container types.
if (ValVT.isFixedLengthVector())
LocVT = TLI.getContainerForFixedLengthVector(LocVT);
State.addLoc(
CCValAssign::getReg(ValNo, ValVT, AllocatedVReg, LocVT, LocInfo));
} else {
// Try and pass the address via a "fast" GPR.
if (unsigned GPRReg = State.AllocateReg(getFastCCArgGPRs(ABI))) {
LocInfo = CCValAssign::Indirect;
LocVT = TLI.getSubtarget().getXLenVT();
State.addLoc(CCValAssign::getReg(ValNo, ValVT, GPRReg, LocVT, LocInfo));
} else if (ValVT.isFixedLengthVector()) {
auto StackAlign =
MaybeAlign(ValVT.getScalarSizeInBits() / 8).valueOrOne();
unsigned StackOffset =
State.AllocateStack(ValVT.getStoreSize(), StackAlign);
State.addLoc(
CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
} else {
// Can't pass scalable vectors on the stack.
return true;
}
}
return false;
}
return true; // CC didn't match.
}
bool RISCV::CC_RISCV_GHC(unsigned ValNo, MVT ValVT, MVT LocVT,
CCValAssign::LocInfo LocInfo,
ISD::ArgFlagsTy ArgFlags, CCState &State) {
if (ArgFlags.isNest()) {
report_fatal_error(
"Attribute 'nest' is not supported in GHC calling convention");
}
static const MCPhysReg GPRList[] = {
RISCV::X9, RISCV::X18, RISCV::X19, RISCV::X20, RISCV::X21, RISCV::X22,
RISCV::X23, RISCV::X24, RISCV::X25, RISCV::X26, RISCV::X27};
if (LocVT == MVT::i32 || LocVT == MVT::i64) {
// Pass in STG registers: Base, Sp, Hp, R1, R2, R3, R4, R5, R6, R7, SpLim
// s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11
if (unsigned Reg = State.AllocateReg(GPRList)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
const RISCVSubtarget &Subtarget =
State.getMachineFunction().getSubtarget<RISCVSubtarget>();
if (LocVT == MVT::f32 && Subtarget.hasStdExtF()) {
// Pass in STG registers: F1, ..., F6
// fs0 ... fs5
static const MCPhysReg FPR32List[] = {RISCV::F8_F, RISCV::F9_F,
RISCV::F18_F, RISCV::F19_F,
RISCV::F20_F, RISCV::F21_F};
if (unsigned Reg = State.AllocateReg(FPR32List)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
if (LocVT == MVT::f64 && Subtarget.hasStdExtD()) {
// Pass in STG registers: D1, ..., D6
// fs6 ... fs11
static const MCPhysReg FPR64List[] = {RISCV::F22_D, RISCV::F23_D,
RISCV::F24_D, RISCV::F25_D,
RISCV::F26_D, RISCV::F27_D};
if (unsigned Reg = State.AllocateReg(FPR64List)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
if ((LocVT == MVT::f32 && Subtarget.hasStdExtZfinx()) ||
(LocVT == MVT::f64 && Subtarget.hasStdExtZdinx() &&
Subtarget.is64Bit())) {
if (unsigned Reg = State.AllocateReg(GPRList)) {
State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
return false;
}
}
report_fatal_error("No registers left in GHC calling convention");
return true;
}
// Transform physical registers into virtual registers.
SDValue RISCVTargetLowering::LowerFormalArguments(
SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
switch (CallConv) {
default:
report_fatal_error("Unsupported calling convention");
case CallingConv::C:
case CallingConv::Fast:
case CallingConv::SPIR_KERNEL:
case CallingConv::GRAAL:
case CallingConv::RISCV_VectorCall:
break;
case CallingConv::GHC:
if (Subtarget.hasStdExtE())
report_fatal_error("GHC calling convention is not supported on RVE!");
if (!Subtarget.hasStdExtFOrZfinx() || !Subtarget.hasStdExtDOrZdinx())
report_fatal_error("GHC calling convention requires the (Zfinx/F) and "
"(Zdinx/D) instruction set extensions");
}
const Function &Func = MF.getFunction();
if (Func.hasFnAttribute("interrupt")) {
if (!Func.arg_empty())
report_fatal_error(
"Functions with the interrupt attribute cannot have arguments!");
StringRef Kind =
MF.getFunction().getFnAttribute("interrupt").getValueAsString();
if (!(Kind == "user" || Kind == "supervisor" || Kind == "machine"))
report_fatal_error(
"Function interrupt attribute argument not supported!");
}
EVT PtrVT = getPointerTy(DAG.getDataLayout());
MVT XLenVT = Subtarget.getXLenVT();
unsigned XLenInBytes = Subtarget.getXLen() / 8;
// Used with vargs to acumulate store chains.
std::vector<SDValue> OutChains;
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
if (CallConv == CallingConv::GHC)
CCInfo.AnalyzeFormalArguments(Ins, RISCV::CC_RISCV_GHC);
else
analyzeInputArgs(MF, CCInfo, Ins, /*IsRet=*/false,
CallConv == CallingConv::Fast ? RISCV::CC_RISCV_FastCC
: RISCV::CC_RISCV);
for (unsigned i = 0, e = ArgLocs.size(), InsIdx = 0; i != e; ++i, ++InsIdx) {
CCValAssign &VA = ArgLocs[i];
SDValue ArgValue;
// Passing f64 on RV32D with a soft float ABI must be handled as a special
// case.
if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
assert(VA.needsCustom());
ArgValue = unpackF64OnRV32DSoftABI(DAG, Chain, VA, ArgLocs[++i], DL);
} else if (VA.isRegLoc())
ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL, Ins[InsIdx], *this);
else
ArgValue = unpackFromMemLoc(DAG, Chain, VA, DL);
if (VA.getLocInfo() == CCValAssign::Indirect) {
// If the original argument was split and passed by reference (e.g. i128
// on RV32), we need to load all parts of it here (using the same
// address). Vectors may be partly split to registers and partly to the
// stack, in which case the base address is partly offset and subsequent
// stores are relative to that.
InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, ArgValue,
MachinePointerInfo()));
unsigned ArgIndex = Ins[InsIdx].OrigArgIndex;
unsigned ArgPartOffset = Ins[InsIdx].PartOffset;
assert(VA.getValVT().isVector() || ArgPartOffset == 0);
while (i + 1 != e && Ins[InsIdx + 1].OrigArgIndex == ArgIndex) {
CCValAssign &PartVA = ArgLocs[i + 1];
unsigned PartOffset = Ins[InsIdx + 1].PartOffset - ArgPartOffset;
SDValue Offset = DAG.getIntPtrConstant(PartOffset, DL);
if (PartVA.getValVT().isScalableVector())
Offset = DAG.getNode(ISD::VSCALE, DL, XLenVT, Offset);
SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue, Offset);
InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address,
MachinePointerInfo()));
++i;
++InsIdx;
}
continue;
}
InVals.push_back(ArgValue);
}
if (any_of(ArgLocs,
[](CCValAssign &VA) { return VA.getLocVT().isScalableVector(); }))
MF.getInfo<RISCVMachineFunctionInfo>()->setIsVectorCall();
if (IsVarArg) {
ArrayRef<MCPhysReg> ArgRegs = RISCV::getArgGPRs(Subtarget.getTargetABI());
unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs);
const TargetRegisterClass *RC = &RISCV::GPRRegClass;
MachineFrameInfo &MFI = MF.getFrameInfo();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
// Size of the vararg save area. For now, the varargs save area is either
// zero or large enough to hold a0-a7.
int VarArgsSaveSize = XLenInBytes * (ArgRegs.size() - Idx);
int FI;
// If all registers are allocated, then all varargs must be passed on the
// stack and we don't need to save any argregs.
if (VarArgsSaveSize == 0) {
int VaArgOffset = CCInfo.getStackSize();
FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
} else {
int VaArgOffset = -VarArgsSaveSize;
FI = MFI.CreateFixedObject(VarArgsSaveSize, VaArgOffset, true);
// If saving an odd number of registers then create an extra stack slot to
// ensure that the frame pointer is 2*XLEN-aligned, which in turn ensures
// offsets to even-numbered registered remain 2*XLEN-aligned.
if (Idx % 2) {
MFI.CreateFixedObject(
XLenInBytes, VaArgOffset - static_cast<int>(XLenInBytes), true);
VarArgsSaveSize += XLenInBytes;
}
SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
// Copy the integer registers that may have been used for passing varargs
// to the vararg save area.
for (unsigned I = Idx; I < ArgRegs.size(); ++I) {
const Register Reg = RegInfo.createVirtualRegister(RC);
RegInfo.addLiveIn(ArgRegs[I], Reg);
SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, XLenVT);
SDValue Store = DAG.getStore(
Chain, DL, ArgValue, FIN,
MachinePointerInfo::getFixedStack(MF, FI, (I - Idx) * XLenInBytes));
OutChains.push_back(Store);
FIN =
DAG.getMemBasePlusOffset(FIN, TypeSize::getFixed(XLenInBytes), DL);
}
}
// Record the frame index of the first variable argument
// which is a value necessary to VASTART.
RVFI->setVarArgsFrameIndex(FI);
RVFI->setVarArgsSaveSize(VarArgsSaveSize);
}
// All stores are grouped in one node to allow the matching between
// the size of Ins and InVals. This only happens for vararg functions.
if (!OutChains.empty()) {
OutChains.push_back(Chain);
Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
}
return Chain;
}
/// isEligibleForTailCallOptimization - Check whether the call is eligible
/// for tail call optimization.
/// Note: This is modelled after ARM's IsEligibleForTailCallOptimization.
bool RISCVTargetLowering::isEligibleForTailCallOptimization(
CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
const SmallVector<CCValAssign, 16> &ArgLocs) const {
auto CalleeCC = CLI.CallConv;
auto &Outs = CLI.Outs;
auto &Caller = MF.getFunction();
auto CallerCC = Caller.getCallingConv();
// Exception-handling functions need a special set of instructions to
// indicate a return to the hardware. Tail-calling another function would
// probably break this.
// TODO: The "interrupt" attribute isn't currently defined by RISC-V. This
// should be expanded as new function attributes are introduced.
if (Caller.hasFnAttribute("interrupt"))
return false;
// Do not tail call opt if the stack is used to pass parameters.
if (CCInfo.getStackSize() != 0)
return false;
// Do not tail call opt if any parameters need to be passed indirectly.
// Since long doubles (fp128) and i128 are larger than 2*XLEN, they are
// passed indirectly. So the address of the value will be passed in a
// register, or if not available, then the address is put on the stack. In
// order to pass indirectly, space on the stack often needs to be allocated
// in order to store the value. In this case the CCInfo.getNextStackOffset()
// != 0 check is not enough and we need to check if any CCValAssign ArgsLocs
// are passed CCValAssign::Indirect.
for (auto &VA : ArgLocs)
if (VA.getLocInfo() == CCValAssign::Indirect)
return false;
// Do not tail call opt if either caller or callee uses struct return
// semantics.
auto IsCallerStructRet = Caller.hasStructRetAttr();
auto IsCalleeStructRet = Outs.empty() ? false : Outs[0].Flags.isSRet();
if (IsCallerStructRet || IsCalleeStructRet)
return false;
// The callee has to preserve all registers the caller needs to preserve.
const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
if (CalleeCC != CallerCC) {
const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
return false;
}
// Byval parameters hand the function a pointer directly into the stack area
// we want to reuse during a tail call. Working around this *is* possible
// but less efficient and uglier in LowerCall.
for (auto &Arg : Outs)
if (Arg.Flags.isByVal())
return false;
return true;
}
static Align getPrefTypeAlign(EVT VT, SelectionDAG &DAG) {
return DAG.getDataLayout().getPrefTypeAlign(
VT.getTypeForEVT(*DAG.getContext()));
}
// Lower a call to a callseq_start + CALL + callseq_end chain, and add input
// and output parameter nodes.
SDValue RISCVTargetLowering::LowerCall(CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
SelectionDAG &DAG = CLI.DAG;
SDLoc &DL = CLI.DL;
SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
SDValue Chain = CLI.Chain;
SDValue Callee = CLI.Callee;
bool &IsTailCall = CLI.IsTailCall;
CallingConv::ID CallConv = CLI.CallConv;
bool IsVarArg = CLI.IsVarArg;
EVT PtrVT = getPointerTy(DAG.getDataLayout());
MVT XLenVT = Subtarget.getXLenVT();
MachineFunction &MF = DAG.getMachineFunction();
// Analyze the operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
if (CallConv == CallingConv::GHC) {
if (Subtarget.hasStdExtE())
report_fatal_error("GHC calling convention is not supported on RVE!");
ArgCCInfo.AnalyzeCallOperands(Outs, RISCV::CC_RISCV_GHC);
} else
analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false, &CLI,
CallConv == CallingConv::Fast ? RISCV::CC_RISCV_FastCC
: RISCV::CC_RISCV);
// Check if it's really possible to do a tail call.
if (IsTailCall)
IsTailCall = isEligibleForTailCallOptimization(ArgCCInfo, CLI, MF, ArgLocs);
if (IsTailCall)
++NumTailCalls;
else if (CLI.CB && CLI.CB->isMustTailCall())
report_fatal_error("failed to perform tail call elimination on a call "
"site marked musttail");
// Get a count of how many bytes are to be pushed on the stack.
unsigned NumBytes = ArgCCInfo.getStackSize();
// Create local copies for byval args
SmallVector<SDValue, 8> ByValArgs;
for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
ISD::ArgFlagsTy Flags = Outs[i].Flags;
if (!Flags.isByVal())
continue;
SDValue Arg = OutVals[i];
unsigned Size = Flags.getByValSize();
Align Alignment = Flags.getNonZeroByValAlign();
int FI =
MF.getFrameInfo().CreateStackObject(Size, Alignment, /*isSS=*/false);
SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
SDValue SizeNode = DAG.getConstant(Size, DL, XLenVT);
Chain = DAG.getMemcpy(Chain, DL, FIPtr, Arg, SizeNode, Alignment,
/*IsVolatile=*/false,
/*AlwaysInline=*/false, IsTailCall,
MachinePointerInfo(), MachinePointerInfo());
ByValArgs.push_back(FIPtr);
}
if (!IsTailCall)
Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, CLI.DL);
// Copy argument values to their designated locations.
SmallVector<std::pair<Register, SDValue>, 8> RegsToPass;
SmallVector<SDValue, 8> MemOpChains;
SDValue StackPtr;
for (unsigned i = 0, j = 0, e = ArgLocs.size(), OutIdx = 0; i != e;
++i, ++OutIdx) {
CCValAssign &VA = ArgLocs[i];
SDValue ArgValue = OutVals[OutIdx];
ISD::ArgFlagsTy Flags = Outs[OutIdx].Flags;
// Handle passing f64 on RV32D with a soft float ABI as a special case.
if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
assert(VA.isRegLoc() && "Expected register VA assignment");
assert(VA.needsCustom());
SDValue SplitF64 = DAG.getNode(
RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32), ArgValue);
SDValue Lo = SplitF64.getValue(0);
SDValue Hi = SplitF64.getValue(1);
Register RegLo = VA.getLocReg();
RegsToPass.push_back(std::make_pair(RegLo, Lo));
// Get the CCValAssign for the Hi part.
CCValAssign &HiVA = ArgLocs[++i];
if (HiVA.isMemLoc()) {
// Second half of f64 is passed on the stack.
if (!StackPtr.getNode())
StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT);
SDValue Address =
DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
DAG.getIntPtrConstant(HiVA.getLocMemOffset(), DL));
// Emit the store.
MemOpChains.push_back(
DAG.getStore(Chain, DL, Hi, Address, MachinePointerInfo()));
} else {
// Second half of f64 is passed in another GPR.
Register RegHigh = HiVA.getLocReg();
RegsToPass.push_back(std::make_pair(RegHigh, Hi));
}
continue;
}
// Promote the value if needed.
// For now, only handle fully promoted and indirect arguments.
if (VA.getLocInfo() == CCValAssign::Indirect) {
// Store the argument in a stack slot and pass its address.
Align StackAlign =
std::max(getPrefTypeAlign(Outs[OutIdx].ArgVT, DAG),
getPrefTypeAlign(ArgValue.getValueType(), DAG));
TypeSize StoredSize = ArgValue.getValueType().getStoreSize();
// If the original argument was split (e.g. i128), we need
// to store the required parts of it here (and pass just one address).
// Vectors may be partly split to registers and partly to the stack, in
// which case the base address is partly offset and subsequent stores are
// relative to that.
unsigned ArgIndex = Outs[OutIdx].OrigArgIndex;
unsigned ArgPartOffset = Outs[OutIdx].PartOffset;
assert(VA.getValVT().isVector() || ArgPartOffset == 0);
// Calculate the total size to store. We don't have access to what we're
// actually storing other than performing the loop and collecting the
// info.
SmallVector<std::pair<SDValue, SDValue>> Parts;
while (i + 1 != e && Outs[OutIdx + 1].OrigArgIndex == ArgIndex) {
SDValue PartValue = OutVals[OutIdx + 1];
unsigned PartOffset = Outs[OutIdx + 1].PartOffset - ArgPartOffset;
SDValue Offset = DAG.getIntPtrConstant(PartOffset, DL);
EVT PartVT = PartValue.getValueType();
if (PartVT.isScalableVector())
Offset = DAG.getNode(ISD::VSCALE, DL, XLenVT, Offset);
StoredSize += PartVT.getStoreSize();
StackAlign = std::max(StackAlign, getPrefTypeAlign(PartVT, DAG));
Parts.push_back(std::make_pair(PartValue, Offset));
++i;
++OutIdx;
}
SDValue SpillSlot = DAG.CreateStackTemporary(StoredSize, StackAlign);
int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
MemOpChains.push_back(
DAG.getStore(Chain, DL, ArgValue, SpillSlot,
MachinePointerInfo::getFixedStack(MF, FI)));
for (const auto &Part : Parts) {
SDValue PartValue = Part.first;
SDValue PartOffset = Part.second;
SDValue Address =
DAG.getNode(ISD::ADD, DL, PtrVT, SpillSlot, PartOffset);
MemOpChains.push_back(
DAG.getStore(Chain, DL, PartValue, Address,
MachinePointerInfo::getFixedStack(MF, FI)));
}
ArgValue = SpillSlot;
} else {
ArgValue = convertValVTToLocVT(DAG, ArgValue, VA, DL, Subtarget);
}
// Use local copy if it is a byval arg.
if (Flags.isByVal())
ArgValue = ByValArgs[j++];
if (VA.isRegLoc()) {
// Queue up the argument copies and emit them at the end.
RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
} else {
assert(VA.isMemLoc() && "Argument not register or memory");
assert(!IsTailCall && "Tail call not allowed if stack is used "
"for passing parameters");
// Work out the address of the stack slot.
if (!StackPtr.getNode())
StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT);
SDValue Address =
DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
// Emit the store.
MemOpChains.push_back(
DAG.getStore(Chain, DL, ArgValue, Address, MachinePointerInfo()));
}
}
// Join the stores, which are independent of one another.
if (!MemOpChains.empty())
Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
SDValue Glue;
// Build a sequence of copy-to-reg nodes, chained and glued together.
for (auto &Reg : RegsToPass) {
Chain = DAG.getCopyToReg(Chain, DL, Reg.first, Reg.second, Glue);
Glue = Chain.getValue(1);
}
// Validate that none of the argument registers have been marked as
// reserved, if so report an error. Do the same for the return address if this
// is not a tailcall.
validateCCReservedRegs(RegsToPass, MF);
if (!IsTailCall &&
MF.getSubtarget<RISCVSubtarget>().isRegisterReservedByUser(RISCV::X1))
MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
MF.getFunction(),
"Return address register required, but has been reserved."});
// If the callee is a GlobalAddress/ExternalSymbol node, turn it into a
// TargetGlobalAddress/TargetExternalSymbol node so that legalize won't
// split it and then direct call can be matched by PseudoCALL.
if (GlobalAddressSDNode *S = dyn_cast<GlobalAddressSDNode>(Callee)) {
const GlobalValue *GV = S->getGlobal();
Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, RISCVII::MO_CALL);
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT, RISCVII::MO_CALL);
}
// The first call operand is the chain and the second is the target address.
SmallVector<SDValue, 8> Ops;
Ops.push_back(Chain);
Ops.push_back(Callee);
// Add argument registers to the end of the list so that they are
// known live into the call.
for (auto &Reg : RegsToPass)
Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType()));
if (!IsTailCall) {
// Add a register mask operand representing the call-preserved registers.
const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
assert(Mask && "Missing call preserved mask for calling convention");
Ops.push_back(DAG.getRegisterMask(Mask));
}
// Glue the call to the argument copies, if any.
if (Glue.getNode())
Ops.push_back(Glue);
assert((!CLI.CFIType || CLI.CB->isIndirectCall()) &&
"Unexpected CFI type for a direct call");
// Emit the call.
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
if (IsTailCall) {
MF.getFrameInfo().setHasTailCall();
SDValue Ret = DAG.getNode(RISCVISD::TAIL, DL, NodeTys, Ops);
if (CLI.CFIType)
Ret.getNode()->setCFIType(CLI.CFIType->getZExtValue());
DAG.addNoMergeSiteInfo(Ret.getNode(), CLI.NoMerge);
return Ret;
}
Chain = DAG.getNode(RISCVISD::CALL, DL, NodeTys, Ops);
if (CLI.CFIType)
Chain.getNode()->setCFIType(CLI.CFIType->getZExtValue());
DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge);
Glue = Chain.getValue(1);
// Mark the end of the call, which is glued to the call itself.
Chain = DAG.getCALLSEQ_END(Chain, NumBytes, 0, Glue, DL);
Glue = Chain.getValue(1);
// Assign locations to each value returned by this call.
SmallVector<CCValAssign, 16> RVLocs;
CCState RetCCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
analyzeInputArgs(MF, RetCCInfo, Ins, /*IsRet=*/true, RISCV::CC_RISCV);
// Copy all of the result registers out of their specified physreg.
for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
auto &VA = RVLocs[i];
// Copy the value out
SDValue RetValue =
DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), Glue);
// Glue the RetValue to the end of the call sequence
Chain = RetValue.getValue(1);
Glue = RetValue.getValue(2);
if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
assert(VA.needsCustom());
SDValue RetValue2 = DAG.getCopyFromReg(Chain, DL, RVLocs[++i].getLocReg(),
MVT::i32, Glue);
Chain = RetValue2.getValue(1);
Glue = RetValue2.getValue(2);
RetValue = DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, RetValue,
RetValue2);
}
RetValue = convertLocVTToValVT(DAG, RetValue, VA, DL, Subtarget);
InVals.push_back(RetValue);
}
return Chain;
}
bool RISCVTargetLowering::CanLowerReturn(
CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
SmallVector<CCValAssign, 16> RVLocs;
CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
RVVArgDispatcher Dispatcher{&MF, this, ArrayRef(Outs)};
for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
MVT VT = Outs[i].VT;
ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
if (RISCV::CC_RISCV(MF.getDataLayout(), ABI, i, VT, VT, CCValAssign::Full,
ArgFlags, CCInfo, /*IsFixed=*/true, /*IsRet=*/true,
nullptr, *this, Dispatcher))
return false;
}
return true;
}
SDValue
RISCVTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
bool IsVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
const SDLoc &DL, SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>();
// Stores the assignment of the return value to a location.
SmallVector<CCValAssign, 16> RVLocs;
// Info about the registers and stack slot.
CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
*DAG.getContext());
analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true,
nullptr, RISCV::CC_RISCV);
if (CallConv == CallingConv::GHC && !RVLocs.empty())
report_fatal_error("GHC functions return void only");
SDValue Glue;
SmallVector<SDValue, 4> RetOps(1, Chain);
// Copy the result values into the output registers.
for (unsigned i = 0, e = RVLocs.size(), OutIdx = 0; i < e; ++i, ++OutIdx) {
SDValue Val = OutVals[OutIdx];
CCValAssign &VA = RVLocs[i];
assert(VA.isRegLoc() && "Can only return in registers!");
if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
// Handle returning f64 on RV32D with a soft float ABI.
assert(VA.isRegLoc() && "Expected return via registers");
assert(VA.needsCustom());
SDValue SplitF64 = DAG.getNode(RISCVISD::SplitF64, DL,
DAG.getVTList(MVT::i32, MVT::i32), Val);
SDValue Lo = SplitF64.getValue(0);
SDValue Hi = SplitF64.getValue(1);
Register RegLo = VA.getLocReg();
Register RegHi = RVLocs[++i].getLocReg();
if (STI.isRegisterReservedByUser(RegLo) ||
STI.isRegisterReservedByUser(RegHi))
MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
MF.getFunction(),
"Return value register required, but has been reserved."});
Chain = DAG.getCopyToReg(Chain, DL, RegLo, Lo, Glue);
Glue = Chain.getValue(1);
RetOps.push_back(DAG.getRegister(RegLo, MVT::i32));
Chain = DAG.getCopyToReg(Chain, DL, RegHi, Hi, Glue);
Glue = Chain.getValue(1);
RetOps.push_back(DAG.getRegister(RegHi, MVT::i32));
} else {
// Handle a 'normal' return.
Val = convertValVTToLocVT(DAG, Val, VA, DL, Subtarget);
Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Glue);
if (STI.isRegisterReservedByUser(VA.getLocReg()))
MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
MF.getFunction(),
"Return value register required, but has been reserved."});
// Guarantee that all emitted copies are stuck together.
Glue = Chain.getValue(1);
RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
}
}
RetOps[0] = Chain; // Update chain.
// Add the glue node if we have it.
if (Glue.getNode()) {
RetOps.push_back(Glue);
}
if (any_of(RVLocs,
[](CCValAssign &VA) { return VA.getLocVT().isScalableVector(); }))
MF.getInfo<RISCVMachineFunctionInfo>()->setIsVectorCall();
unsigned RetOpc = RISCVISD::RET_GLUE;
// Interrupt service routines use different return instructions.
const Function &Func = DAG.getMachineFunction().getFunction();
if (Func.hasFnAttribute("interrupt")) {
if (!Func.getReturnType()->isVoidTy())
report_fatal_error(
"Functions with the interrupt attribute must have void return type!");
MachineFunction &MF = DAG.getMachineFunction();
StringRef Kind =
MF.getFunction().getFnAttribute("interrupt").getValueAsString();
if (Kind == "supervisor")
RetOpc = RISCVISD::SRET_GLUE;
else
RetOpc = RISCVISD::MRET_GLUE;
}
return DAG.getNode(RetOpc, DL, MVT::Other, RetOps);
}
void RISCVTargetLowering::validateCCReservedRegs(
const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs,
MachineFunction &MF) const {
const Function &F = MF.getFunction();
const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>();
if (llvm::any_of(Regs, [&STI](auto Reg) {
return STI.isRegisterReservedByUser(Reg.first);
}))
F.getContext().diagnose(DiagnosticInfoUnsupported{
F, "Argument register required, but has been reserved."});
}
// Check if the result of the node is only used as a return value, as
// otherwise we can't perform a tail-call.
bool RISCVTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
if (N->getNumValues() != 1)
return false;
if (!N->hasNUsesOfValue(1, 0))
return false;
SDNode *Copy = *N->use_begin();
if (Copy->getOpcode() == ISD::BITCAST) {
return isUsedByReturnOnly(Copy, Chain);
}
// TODO: Handle additional opcodes in order to support tail-calling libcalls
// with soft float ABIs.
if (Copy->getOpcode() != ISD::CopyToReg) {
return false;
}
// If the ISD::CopyToReg has a glue operand, we conservatively assume it
// isn't safe to perform a tail call.
if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() == MVT::Glue)
return false;
// The copy must be used by a RISCVISD::RET_GLUE, and nothing else.
bool HasRet = false;
for (SDNode *Node : Copy->uses()) {
if (Node->getOpcode() != RISCVISD::RET_GLUE)
return false;
HasRet = true;
}
if (!HasRet)
return false;
Chain = Copy->getOperand(0);
return true;
}
bool RISCVTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
return CI->isTailCall();
}
const char *RISCVTargetLowering::getTargetNodeName(unsigned Opcode) const {
#define NODE_NAME_CASE(NODE) \
case RISCVISD::NODE: \
return "RISCVISD::" #NODE;
// clang-format off
switch ((RISCVISD::NodeType)Opcode) {
case RISCVISD::FIRST_NUMBER:
break;
NODE_NAME_CASE(RET_GLUE)
NODE_NAME_CASE(SRET_GLUE)
NODE_NAME_CASE(MRET_GLUE)
NODE_NAME_CASE(CALL)
NODE_NAME_CASE(SELECT_CC)
NODE_NAME_CASE(BR_CC)
NODE_NAME_CASE(BuildPairF64)
NODE_NAME_CASE(SplitF64)
NODE_NAME_CASE(TAIL)
NODE_NAME_CASE(ADD_LO)
NODE_NAME_CASE(HI)
NODE_NAME_CASE(LLA)
NODE_NAME_CASE(ADD_TPREL)
NODE_NAME_CASE(MULHSU)
NODE_NAME_CASE(SLLW)
NODE_NAME_CASE(SRAW)
NODE_NAME_CASE(SRLW)
NODE_NAME_CASE(DIVW)
NODE_NAME_CASE(DIVUW)
NODE_NAME_CASE(REMUW)
NODE_NAME_CASE(ROLW)
NODE_NAME_CASE(RORW)
NODE_NAME_CASE(CLZW)
NODE_NAME_CASE(CTZW)
NODE_NAME_CASE(ABSW)
NODE_NAME_CASE(FMV_H_X)
NODE_NAME_CASE(FMV_X_ANYEXTH)
NODE_NAME_CASE(FMV_X_SIGNEXTH)
NODE_NAME_CASE(FMV_W_X_RV64)
NODE_NAME_CASE(FMV_X_ANYEXTW_RV64)
NODE_NAME_CASE(FCVT_X)
NODE_NAME_CASE(FCVT_XU)
NODE_NAME_CASE(FCVT_W_RV64)
NODE_NAME_CASE(FCVT_WU_RV64)
NODE_NAME_CASE(STRICT_FCVT_W_RV64)
NODE_NAME_CASE(STRICT_FCVT_WU_RV64)
NODE_NAME_CASE(FP_ROUND_BF16)
NODE_NAME_CASE(FP_EXTEND_BF16)
NODE_NAME_CASE(FROUND)
NODE_NAME_CASE(FCLASS)
NODE_NAME_CASE(FMAX)
NODE_NAME_CASE(FMIN)
NODE_NAME_CASE(READ_COUNTER_WIDE)
NODE_NAME_CASE(BREV8)
NODE_NAME_CASE(ORC_B)
NODE_NAME_CASE(ZIP)
NODE_NAME_CASE(UNZIP)
NODE_NAME_CASE(CLMUL)
NODE_NAME_CASE(CLMULH)
NODE_NAME_CASE(CLMULR)
NODE_NAME_CASE(MOPR)
NODE_NAME_CASE(MOPRR)
NODE_NAME_CASE(SHA256SIG0)
NODE_NAME_CASE(SHA256SIG1)
NODE_NAME_CASE(SHA256SUM0)
NODE_NAME_CASE(SHA256SUM1)
NODE_NAME_CASE(SM4KS)
NODE_NAME_CASE(SM4ED)
NODE_NAME_CASE(SM3P0)
NODE_NAME_CASE(SM3P1)
NODE_NAME_CASE(TH_LWD)
NODE_NAME_CASE(TH_LWUD)
NODE_NAME_CASE(TH_LDD)
NODE_NAME_CASE(TH_SWD)
NODE_NAME_CASE(TH_SDD)
NODE_NAME_CASE(VMV_V_V_VL)
NODE_NAME_CASE(VMV_V_X_VL)
NODE_NAME_CASE(VFMV_V_F_VL)
NODE_NAME_CASE(VMV_X_S)
NODE_NAME_CASE(VMV_S_X_VL)
NODE_NAME_CASE(VFMV_S_F_VL)
NODE_NAME_CASE(SPLAT_VECTOR_SPLIT_I64_VL)
NODE_NAME_CASE(READ_VLENB)
NODE_NAME_CASE(TRUNCATE_VECTOR_VL)
NODE_NAME_CASE(VSLIDEUP_VL)
NODE_NAME_CASE(VSLIDE1UP_VL)
NODE_NAME_CASE(VSLIDEDOWN_VL)
NODE_NAME_CASE(VSLIDE1DOWN_VL)
NODE_NAME_CASE(VFSLIDE1UP_VL)
NODE_NAME_CASE(VFSLIDE1DOWN_VL)
NODE_NAME_CASE(VID_VL)
NODE_NAME_CASE(VFNCVT_ROD_VL)
NODE_NAME_CASE(VECREDUCE_ADD_VL)
NODE_NAME_CASE(VECREDUCE_UMAX_VL)
NODE_NAME_CASE(VECREDUCE_SMAX_VL)
NODE_NAME_CASE(VECREDUCE_UMIN_VL)
NODE_NAME_CASE(VECREDUCE_SMIN_VL)
NODE_NAME_CASE(VECREDUCE_AND_VL)
NODE_NAME_CASE(VECREDUCE_OR_VL)
NODE_NAME_CASE(VECREDUCE_XOR_VL)
NODE_NAME_CASE(VECREDUCE_FADD_VL)
NODE_NAME_CASE(VECREDUCE_SEQ_FADD_VL)
NODE_NAME_CASE(VECREDUCE_FMIN_VL)
NODE_NAME_CASE(VECREDUCE_FMAX_VL)
NODE_NAME_CASE(ADD_VL)
NODE_NAME_CASE(AND_VL)
NODE_NAME_CASE(MUL_VL)
NODE_NAME_CASE(OR_VL)
NODE_NAME_CASE(SDIV_VL)
NODE_NAME_CASE(SHL_VL)
NODE_NAME_CASE(SREM_VL)
NODE_NAME_CASE(SRA_VL)
NODE_NAME_CASE(SRL_VL)
NODE_NAME_CASE(ROTL_VL)
NODE_NAME_CASE(ROTR_VL)
NODE_NAME_CASE(SUB_VL)
NODE_NAME_CASE(UDIV_VL)
NODE_NAME_CASE(UREM_VL)
NODE_NAME_CASE(XOR_VL)
NODE_NAME_CASE(AVGFLOORU_VL)
NODE_NAME_CASE(AVGCEILU_VL)
NODE_NAME_CASE(SADDSAT_VL)
NODE_NAME_CASE(UADDSAT_VL)
NODE_NAME_CASE(SSUBSAT_VL)
NODE_NAME_CASE(USUBSAT_VL)
NODE_NAME_CASE(FADD_VL)
NODE_NAME_CASE(FSUB_VL)
NODE_NAME_CASE(FMUL_VL)
NODE_NAME_CASE(FDIV_VL)
NODE_NAME_CASE(FNEG_VL)
NODE_NAME_CASE(FABS_VL)
NODE_NAME_CASE(FSQRT_VL)
NODE_NAME_CASE(FCLASS_VL)
NODE_NAME_CASE(VFMADD_VL)
NODE_NAME_CASE(VFNMADD_VL)
NODE_NAME_CASE(VFMSUB_VL)
NODE_NAME_CASE(VFNMSUB_VL)
NODE_NAME_CASE(VFWMADD_VL)
NODE_NAME_CASE(VFWNMADD_VL)
NODE_NAME_CASE(VFWMSUB_VL)
NODE_NAME_CASE(VFWNMSUB_VL)
NODE_NAME_CASE(FCOPYSIGN_VL)
NODE_NAME_CASE(SMIN_VL)
NODE_NAME_CASE(SMAX_VL)
NODE_NAME_CASE(UMIN_VL)
NODE_NAME_CASE(UMAX_VL)
NODE_NAME_CASE(BITREVERSE_VL)
NODE_NAME_CASE(BSWAP_VL)
NODE_NAME_CASE(CTLZ_VL)
NODE_NAME_CASE(CTTZ_VL)
NODE_NAME_CASE(CTPOP_VL)
NODE_NAME_CASE(VFMIN_VL)
NODE_NAME_CASE(VFMAX_VL)
NODE_NAME_CASE(MULHS_VL)
NODE_NAME_CASE(MULHU_VL)
NODE_NAME_CASE(VFCVT_RTZ_X_F_VL)
NODE_NAME_CASE(VFCVT_RTZ_XU_F_VL)
NODE_NAME_CASE(VFCVT_RM_X_F_VL)
NODE_NAME_CASE(VFCVT_RM_XU_F_VL)
NODE_NAME_CASE(VFCVT_X_F_VL)
NODE_NAME_CASE(VFCVT_XU_F_VL)
NODE_NAME_CASE(VFROUND_NOEXCEPT_VL)
NODE_NAME_CASE(SINT_TO_FP_VL)
NODE_NAME_CASE(UINT_TO_FP_VL)
NODE_NAME_CASE(VFCVT_RM_F_XU_VL)
NODE_NAME_CASE(VFCVT_RM_F_X_VL)
NODE_NAME_CASE(FP_EXTEND_VL)
NODE_NAME_CASE(FP_ROUND_VL)
NODE_NAME_CASE(STRICT_FADD_VL)
NODE_NAME_CASE(STRICT_FSUB_VL)
NODE_NAME_CASE(STRICT_FMUL_VL)
NODE_NAME_CASE(STRICT_FDIV_VL)
NODE_NAME_CASE(STRICT_FSQRT_VL)
NODE_NAME_CASE(STRICT_VFMADD_VL)
NODE_NAME_CASE(STRICT_VFNMADD_VL)
NODE_NAME_CASE(STRICT_VFMSUB_VL)
NODE_NAME_CASE(STRICT_VFNMSUB_VL)
NODE_NAME_CASE(STRICT_FP_ROUND_VL)
NODE_NAME_CASE(STRICT_FP_EXTEND_VL)
NODE_NAME_CASE(STRICT_VFNCVT_ROD_VL)
NODE_NAME_CASE(STRICT_SINT_TO_FP_VL)
NODE_NAME_CASE(STRICT_UINT_TO_FP_VL)
NODE_NAME_CASE(STRICT_VFCVT_RM_X_F_VL)
NODE_NAME_CASE(STRICT_VFCVT_RTZ_X_F_VL)
NODE_NAME_CASE(STRICT_VFCVT_RTZ_XU_F_VL)
NODE_NAME_CASE(STRICT_FSETCC_VL)
NODE_NAME_CASE(STRICT_FSETCCS_VL)
NODE_NAME_CASE(STRICT_VFROUND_NOEXCEPT_VL)
NODE_NAME_CASE(VWMUL_VL)
NODE_NAME_CASE(VWMULU_VL)
NODE_NAME_CASE(VWMULSU_VL)
NODE_NAME_CASE(VWADD_VL)
NODE_NAME_CASE(VWADDU_VL)
NODE_NAME_CASE(VWSUB_VL)
NODE_NAME_CASE(VWSUBU_VL)
NODE_NAME_CASE(VWADD_W_VL)
NODE_NAME_CASE(VWADDU_W_VL)
NODE_NAME_CASE(VWSUB_W_VL)
NODE_NAME_CASE(VWSUBU_W_VL)
NODE_NAME_CASE(VWSLL_VL)
NODE_NAME_CASE(VFWMUL_VL)
NODE_NAME_CASE(VFWADD_VL)
NODE_NAME_CASE(VFWSUB_VL)
NODE_NAME_CASE(VFWADD_W_VL)
NODE_NAME_CASE(VFWSUB_W_VL)
NODE_NAME_CASE(VWMACC_VL)
NODE_NAME_CASE(VWMACCU_VL)
NODE_NAME_CASE(VWMACCSU_VL)
NODE_NAME_CASE(VNSRL_VL)
NODE_NAME_CASE(SETCC_VL)
NODE_NAME_CASE(VMERGE_VL)
NODE_NAME_CASE(VMAND_VL)
NODE_NAME_CASE(VMOR_VL)
NODE_NAME_CASE(VMXOR_VL)
NODE_NAME_CASE(VMCLR_VL)
NODE_NAME_CASE(VMSET_VL)
NODE_NAME_CASE(VRGATHER_VX_VL)
NODE_NAME_CASE(VRGATHER_VV_VL)
NODE_NAME_CASE(VRGATHEREI16_VV_VL)
NODE_NAME_CASE(VSEXT_VL)
NODE_NAME_CASE(VZEXT_VL)
NODE_NAME_CASE(VCPOP_VL)
NODE_NAME_CASE(VFIRST_VL)
NODE_NAME_CASE(READ_CSR)
NODE_NAME_CASE(WRITE_CSR)
NODE_NAME_CASE(SWAP_CSR)
NODE_NAME_CASE(CZERO_EQZ)
NODE_NAME_CASE(CZERO_NEZ)
NODE_NAME_CASE(SF_VC_XV_SE)
NODE_NAME_CASE(SF_VC_IV_SE)
NODE_NAME_CASE(SF_VC_VV_SE)
NODE_NAME_CASE(SF_VC_FV_SE)
NODE_NAME_CASE(SF_VC_XVV_SE)
NODE_NAME_CASE(SF_VC_IVV_SE)
NODE_NAME_CASE(SF_VC_VVV_SE)
NODE_NAME_CASE(SF_VC_FVV_SE)
NODE_NAME_CASE(SF_VC_XVW_SE)
NODE_NAME_CASE(SF_VC_IVW_SE)
NODE_NAME_CASE(SF_VC_VVW_SE)
NODE_NAME_CASE(SF_VC_FVW_SE)
NODE_NAME_CASE(SF_VC_V_X_SE)
NODE_NAME_CASE(SF_VC_V_I_SE)
NODE_NAME_CASE(SF_VC_V_XV_SE)
NODE_NAME_CASE(SF_VC_V_IV_SE)
NODE_NAME_CASE(SF_VC_V_VV_SE)
NODE_NAME_CASE(SF_VC_V_FV_SE)
NODE_NAME_CASE(SF_VC_V_XVV_SE)
NODE_NAME_CASE(SF_VC_V_IVV_SE)
NODE_NAME_CASE(SF_VC_V_VVV_SE)
NODE_NAME_CASE(SF_VC_V_FVV_SE)
NODE_NAME_CASE(SF_VC_V_XVW_SE)
NODE_NAME_CASE(SF_VC_V_IVW_SE)
NODE_NAME_CASE(SF_VC_V_VVW_SE)
NODE_NAME_CASE(SF_VC_V_FVW_SE)
}
// clang-format on
return nullptr;
#undef NODE_NAME_CASE
}
/// getConstraintType - Given a constraint letter, return the type of
/// constraint it is for this target.
RISCVTargetLowering::ConstraintType
RISCVTargetLowering::getConstraintType(StringRef Constraint) const {
if (Constraint.size() == 1) {
switch (Constraint[0]) {
default:
break;
case 'f':
return C_RegisterClass;
case 'I':
case 'J':
case 'K':
return C_Immediate;
case 'A':
return C_Memory;
case 's':
case 'S': // A symbolic address
return C_Other;
}
} else {
if (Constraint == "vr" || Constraint == "vm")
return C_RegisterClass;
}
return TargetLowering::getConstraintType(Constraint);
}
std::pair<unsigned, const TargetRegisterClass *>
RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
StringRef Constraint,
MVT VT) const {
// First, see if this is a constraint that directly corresponds to a RISC-V
// register class.
if (Constraint.size() == 1) {
switch (Constraint[0]) {
case 'r':
// TODO: Support fixed vectors up to XLen for P extension?
if (VT.isVector())
break;
if (VT == MVT::f16 && Subtarget.hasStdExtZhinxmin())
return std::make_pair(0U, &RISCV::GPRF16RegClass);
if (VT == MVT::f32 && Subtarget.hasStdExtZfinx())
return std::make_pair(0U, &RISCV::GPRF32RegClass);
if (VT == MVT::f64 && Subtarget.hasStdExtZdinx() && !Subtarget.is64Bit())
return std::make_pair(0U, &RISCV::GPRPairRegClass);
return std::make_pair(0U, &RISCV::GPRNoX0RegClass);
case 'f':
if (Subtarget.hasStdExtZfhmin() && VT == MVT::f16)
return std::make_pair(0U, &RISCV::FPR16RegClass);
if (Subtarget.hasStdExtF() && VT == MVT::f32)
return std::make_pair(0U, &RISCV::FPR32RegClass);
if (Subtarget.hasStdExtD() && VT == MVT::f64)
return std::make_pair(0U, &RISCV::FPR64RegClass);
break;
default:
break;
}
} else if (Constraint == "vr") {
for (const auto *RC : {&RISCV::VRRegClass, &RISCV::VRM2RegClass,
&RISCV::VRM4RegClass, &RISCV::VRM8RegClass}) {
if (TRI->isTypeLegalForClass(*RC, VT.SimpleTy))
return std::make_pair(0U, RC);
}
} else if (Constraint == "vm") {
if (TRI->isTypeLegalForClass(RISCV::VMV0RegClass, VT.SimpleTy))
return std::make_pair(0U, &RISCV::VMV0RegClass);
}
// Clang will correctly decode the usage of register name aliases into their
// official names. However, other frontends like `rustc` do not. This allows
// users of these frontends to use the ABI names for registers in LLVM-style
// register constraints.
unsigned XRegFromAlias = StringSwitch<unsigned>(Constraint.lower())
.Case("{zero}", RISCV::X0)
.Case("{ra}", RISCV::X1)
.Case("{sp}", RISCV::X2)
.Case("{gp}", RISCV::X3)
.Case("{tp}", RISCV::X4)
.Case("{t0}", RISCV::X5)
.Case("{t1}", RISCV::X6)
.Case("{t2}", RISCV::X7)
.Cases("{s0}", "{fp}", RISCV::X8)
.Case("{s1}", RISCV::X9)
.Case("{a0}", RISCV::X10)
.Case("{a1}", RISCV::X11)
.Case("{a2}", RISCV::X12)
.Case("{a3}", RISCV::X13)
.Case("{a4}", RISCV::X14)
.Case("{a5}", RISCV::X15)
.Case("{a6}", RISCV::X16)
.Case("{a7}", RISCV::X17)
.Case("{s2}", RISCV::X18)
.Case("{s3}", RISCV::X19)
.Case("{s4}", RISCV::X20)
.Case("{s5}", RISCV::X21)
.Case("{s6}", RISCV::X22)
.Case("{s7}", RISCV::X23)
.Case("{s8}", RISCV::X24)
.Case("{s9}", RISCV::X25)
.Case("{s10}", RISCV::X26)
.Case("{s11}", RISCV::X27)
.Case("{t3}", RISCV::X28)
.Case("{t4}", RISCV::X29)
.Case("{t5}", RISCV::X30)
.Case("{t6}", RISCV::X31)
.Default(RISCV::NoRegister);
if (XRegFromAlias != RISCV::NoRegister)
return std::make_pair(XRegFromAlias, &RISCV::GPRRegClass);
// Since TargetLowering::getRegForInlineAsmConstraint uses the name of the
// TableGen record rather than the AsmName to choose registers for InlineAsm
// constraints, plus we want to match those names to the widest floating point
// register type available, manually select floating point registers here.
//
// The second case is the ABI name of the register, so that frontends can also
// use the ABI names in register constraint lists.
if (Subtarget.hasStdExtF()) {
unsigned FReg = StringSwitch<unsigned>(Constraint.lower())
.Cases("{f0}", "{ft0}", RISCV::F0_F)
.Cases("{f1}", "{ft1}", RISCV::F1_F)
.Cases("{f2}", "{ft2}", RISCV::F2_F)
.Cases("{f3}", "{ft3}", RISCV::F3_F)
.Cases("{f4}", "{ft4}", RISCV::F4_F)
.Cases("{f5}", "{ft5}", RISCV::F5_F)
.Cases("{f6}", "{ft6}", RISCV::F6_F)
.Cases("{f7}", "{ft7}", RISCV::F7_F)
.Cases("{f8}", "{fs0}", RISCV::F8_F)
.Cases("{f9}", "{fs1}", RISCV::F9_F)
.Cases("{f10}", "{fa0}", RISCV::F10_F)
.Cases("{f11}", "{fa1}", RISCV::F11_F)
.Cases("{f12}", "{fa2}", RISCV::F12_F)
.Cases("{f13}", "{fa3}", RISCV::F13_F)
.Cases("{f14}", "{fa4}", RISCV::F14_F)
.Cases("{f15}", "{fa5}", RISCV::F15_F)
.Cases("{f16}", "{fa6}", RISCV::F16_F)
.Cases("{f17}", "{fa7}", RISCV::F17_F)
.Cases("{f18}", "{fs2}", RISCV::F18_F)
.Cases("{f19}", "{fs3}", RISCV::F19_F)
.Cases("{f20}", "{fs4}", RISCV::F20_F)
.Cases("{f21}", "{fs5}", RISCV::F21_F)
.Cases("{f22}", "{fs6}", RISCV::F22_F)
.Cases("{f23}", "{fs7}", RISCV::F23_F)
.Cases("{f24}", "{fs8}", RISCV::F24_F)
.Cases("{f25}", "{fs9}", RISCV::F25_F)
.Cases("{f26}", "{fs10}", RISCV::F26_F)
.Cases("{f27}", "{fs11}", RISCV::F27_F)
.Cases("{f28}", "{ft8}", RISCV::F28_F)
.Cases("{f29}", "{ft9}", RISCV::F29_F)
.Cases("{f30}", "{ft10}", RISCV::F30_F)
.Cases("{f31}", "{ft11}", RISCV::F31_F)
.Default(RISCV::NoRegister);
if (FReg != RISCV::NoRegister) {
assert(RISCV::F0_F <= FReg && FReg <= RISCV::F31_F && "Unknown fp-reg");
if (Subtarget.hasStdExtD() && (VT == MVT::f64 || VT == MVT::Other)) {
unsigned RegNo = FReg - RISCV::F0_F;
unsigned DReg = RISCV::F0_D + RegNo;
return std::make_pair(DReg, &RISCV::FPR64RegClass);
}
if (VT == MVT::f32 || VT == MVT::Other)
return std::make_pair(FReg, &RISCV::FPR32RegClass);
if (Subtarget.hasStdExtZfhmin() && VT == MVT::f16) {
unsigned RegNo = FReg - RISCV::F0_F;
unsigned HReg = RISCV::F0_H + RegNo;
return std::make_pair(HReg, &RISCV::FPR16RegClass);
}
}
}
if (Subtarget.hasVInstructions()) {
Register VReg = StringSwitch<Register>(Constraint.lower())
.Case("{v0}", RISCV::V0)
.Case("{v1}", RISCV::V1)
.Case("{v2}", RISCV::V2)
.Case("{v3}", RISCV::V3)
.Case("{v4}", RISCV::V4)
.Case("{v5}", RISCV::V5)
.Case("{v6}", RISCV::V6)
.Case("{v7}", RISCV::V7)
.Case("{v8}", RISCV::V8)
.Case("{v9}", RISCV::V9)
.Case("{v10}", RISCV::V10)
.Case("{v11}", RISCV::V11)
.Case("{v12}", RISCV::V12)
.Case("{v13}", RISCV::V13)
.Case("{v14}", RISCV::V14)
.Case("{v15}", RISCV::V15)
.Case("{v16}", RISCV::V16)
.Case("{v17}", RISCV::V17)
.Case("{v18}", RISCV::V18)
.Case("{v19}", RISCV::V19)
.Case("{v20}", RISCV::V20)
.Case("{v21}", RISCV::V21)
.Case("{v22}", RISCV::V22)
.Case("{v23}", RISCV::V23)
.Case("{v24}", RISCV::V24)
.Case("{v25}", RISCV::V25)
.Case("{v26}", RISCV::V26)
.Case("{v27}", RISCV::V27)
.Case("{v28}", RISCV::V28)
.Case("{v29}", RISCV::V29)
.Case("{v30}", RISCV::V30)
.Case("{v31}", RISCV::V31)
.Default(RISCV::NoRegister);
if (VReg != RISCV::NoRegister) {
if (TRI->isTypeLegalForClass(RISCV::VMRegClass, VT.SimpleTy))
return std::make_pair(VReg, &RISCV::VMRegClass);
if (TRI->isTypeLegalForClass(RISCV::VRRegClass, VT.SimpleTy))
return std::make_pair(VReg, &RISCV::VRRegClass);
for (const auto *RC :
{&RISCV::VRM2RegClass, &RISCV::VRM4RegClass, &RISCV::VRM8RegClass}) {
if (TRI->isTypeLegalForClass(*RC, VT.SimpleTy)) {
VReg = TRI->getMatchingSuperReg(VReg, RISCV::sub_vrm1_0, RC);
return std::make_pair(VReg, RC);
}
}
}
}
std::pair<Register, const TargetRegisterClass *> Res =
TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
// If we picked one of the Zfinx register classes, remap it to the GPR class.
// FIXME: When Zfinx is supported in CodeGen this will need to take the
// Subtarget into account.
if (Res.second == &RISCV::GPRF16RegClass ||
Res.second == &RISCV::GPRF32RegClass ||
Res.second == &RISCV::GPRPairRegClass)
return std::make_pair(Res.first, &RISCV::GPRRegClass);
return Res;
}
InlineAsm::ConstraintCode
RISCVTargetLowering::getInlineAsmMemConstraint(StringRef ConstraintCode) const {
// Currently only support length 1 constraints.
if (ConstraintCode.size() == 1) {
switch (ConstraintCode[0]) {
case 'A':
return InlineAsm::ConstraintCode::A;
default:
break;
}
}
return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
}
void RISCVTargetLowering::LowerAsmOperandForConstraint(
SDValue Op, StringRef Constraint, std::vector<SDValue> &Ops,
SelectionDAG &DAG) const {
// Currently only support length 1 constraints.
if (Constraint.size() == 1) {
switch (Constraint[0]) {
case 'I':
// Validate & create a 12-bit signed immediate operand.
if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
uint64_t CVal = C->getSExtValue();
if (isInt<12>(CVal))
Ops.push_back(
DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT()));
}
return;
case 'J':
// Validate & create an integer zero operand.
if (isNullConstant(Op))
Ops.push_back(
DAG.getTargetConstant(0, SDLoc(Op), Subtarget.getXLenVT()));
return;
case 'K':
// Validate & create a 5-bit unsigned immediate operand.
if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
uint64_t CVal = C->getZExtValue();
if (isUInt<5>(CVal))
Ops.push_back(
DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT()));
}
return;
case 'S':
TargetLowering::LowerAsmOperandForConstraint(Op, "s", Ops, DAG);
return;
default:
break;
}
}
TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}
Instruction *RISCVTargetLowering::emitLeadingFence(IRBuilderBase &Builder,
Instruction *Inst,
AtomicOrdering Ord) const {
if (Subtarget.hasStdExtZtso()) {
if (isa<LoadInst>(Inst) && Ord == AtomicOrdering::SequentiallyConsistent)
return Builder.CreateFence(Ord);
return nullptr;
}
if (isa<LoadInst>(Inst) && Ord == AtomicOrdering::SequentiallyConsistent)
return Builder.CreateFence(Ord);
if (isa<StoreInst>(Inst) && isReleaseOrStronger(Ord))
return Builder.CreateFence(AtomicOrdering::Release);
return nullptr;
}
Instruction *RISCVTargetLowering::emitTrailingFence(IRBuilderBase &Builder,
Instruction *Inst,
AtomicOrdering Ord) const {
if (Subtarget.hasStdExtZtso()) {
if (isa<StoreInst>(Inst) && Ord == AtomicOrdering::SequentiallyConsistent)
return Builder.CreateFence(Ord);
return nullptr;
}
if (isa<LoadInst>(Inst) && isAcquireOrStronger(Ord))
return Builder.CreateFence(AtomicOrdering::Acquire);
if (Subtarget.enableSeqCstTrailingFence() && isa<StoreInst>(Inst) &&
Ord == AtomicOrdering::SequentiallyConsistent)
return Builder.CreateFence(AtomicOrdering::SequentiallyConsistent);
return nullptr;
}
TargetLowering::AtomicExpansionKind
RISCVTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
// atomicrmw {fadd,fsub} must be expanded to use compare-exchange, as floating
// point operations can't be used in an lr/sc sequence without breaking the
// forward-progress guarantee.
if (AI->isFloatingPointOperation() ||
AI->getOperation() == AtomicRMWInst::UIncWrap ||
AI->getOperation() == AtomicRMWInst::UDecWrap)
return AtomicExpansionKind::CmpXChg;
// Don't expand forced atomics, we want to have __sync libcalls instead.
if (Subtarget.hasForcedAtomics())
return AtomicExpansionKind::None;
unsigned Size = AI->getType()->getPrimitiveSizeInBits();
if (AI->getOperation() == AtomicRMWInst::Nand) {
if (Subtarget.hasStdExtZacas() &&
(Size >= 32 || Subtarget.hasStdExtZabha()))
return AtomicExpansionKind::CmpXChg;
if (Size < 32)
return AtomicExpansionKind::MaskedIntrinsic;
}
if (Size < 32 && !Subtarget.hasStdExtZabha())
return AtomicExpansionKind::MaskedIntrinsic;
return AtomicExpansionKind::None;
}
static Intrinsic::ID
getIntrinsicForMaskedAtomicRMWBinOp(unsigned XLen, AtomicRMWInst::BinOp BinOp) {
if (XLen == 32) {
switch (BinOp) {
default:
llvm_unreachable("Unexpected AtomicRMW BinOp");
case AtomicRMWInst::Xchg:
return Intrinsic::riscv_masked_atomicrmw_xchg_i32;
case AtomicRMWInst::Add:
return Intrinsic::riscv_masked_atomicrmw_add_i32;
case AtomicRMWInst::Sub:
return Intrinsic::riscv_masked_atomicrmw_sub_i32;
case AtomicRMWInst::Nand:
return Intrinsic::riscv_masked_atomicrmw_nand_i32;
case AtomicRMWInst::Max:
return Intrinsic::riscv_masked_atomicrmw_max_i32;
case AtomicRMWInst::Min:
return Intrinsic::riscv_masked_atomicrmw_min_i32;
case AtomicRMWInst::UMax:
return Intrinsic::riscv_masked_atomicrmw_umax_i32;
case AtomicRMWInst::UMin:
return Intrinsic::riscv_masked_atomicrmw_umin_i32;
}
}
if (XLen == 64) {
switch (BinOp) {
default:
llvm_unreachable("Unexpected AtomicRMW BinOp");
case AtomicRMWInst::Xchg:
return Intrinsic::riscv_masked_atomicrmw_xchg_i64;
case AtomicRMWInst::Add:
return Intrinsic::riscv_masked_atomicrmw_add_i64;
case AtomicRMWInst::Sub:
return Intrinsic::riscv_masked_atomicrmw_sub_i64;
case AtomicRMWInst::Nand:
return Intrinsic::riscv_masked_atomicrmw_nand_i64;
case AtomicRMWInst::Max:
return Intrinsic::riscv_masked_atomicrmw_max_i64;
case AtomicRMWInst::Min:
return Intrinsic::riscv_masked_atomicrmw_min_i64;
case AtomicRMWInst::UMax:
return Intrinsic::riscv_masked_atomicrmw_umax_i64;
case AtomicRMWInst::UMin:
return Intrinsic::riscv_masked_atomicrmw_umin_i64;
}
}
llvm_unreachable("Unexpected XLen\n");
}
Value *RISCVTargetLowering::emitMaskedAtomicRMWIntrinsic(
IRBuilderBase &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr,
Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const {
// In the case of an atomicrmw xchg with a constant 0/-1 operand, replace
// the atomic instruction with an AtomicRMWInst::And/Or with appropriate
// mask, as this produces better code than the LR/SC loop emitted by
// int_riscv_masked_atomicrmw_xchg.
if (AI->getOperation() == AtomicRMWInst::Xchg &&
isa<ConstantInt>(AI->getValOperand())) {
ConstantInt *CVal = cast<ConstantInt>(AI->getValOperand());
if (CVal->isZero())
return Builder.CreateAtomicRMW(AtomicRMWInst::And, AlignedAddr,
Builder.CreateNot(Mask, "Inv_Mask"),
AI->getAlign(), Ord);
if (CVal->isMinusOne())
return Builder.CreateAtomicRMW(AtomicRMWInst::Or, AlignedAddr, Mask,
AI->getAlign(), Ord);
}
unsigned XLen = Subtarget.getXLen();
Value *Ordering =
Builder.getIntN(XLen, static_cast<uint64_t>(AI->getOrdering()));
Type *Tys[] = {AlignedAddr->getType()};
Function *LrwOpScwLoop = Intrinsic::getDeclaration(
AI->getModule(),
getIntrinsicForMaskedAtomicRMWBinOp(XLen, AI->getOperation()), Tys);
if (XLen == 64) {
Incr = Builder.CreateSExt(Incr, Builder.getInt64Ty());
Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty());
ShiftAmt = Builder.CreateSExt(ShiftAmt, Builder.getInt64Ty());
}
Value *Result;
// Must pass the shift amount needed to sign extend the loaded value prior
// to performing a signed comparison for min/max. ShiftAmt is the number of
// bits to shift the value into position. Pass XLen-ShiftAmt-ValWidth, which
// is the number of bits to left+right shift the value in order to
// sign-extend.
if (AI->getOperation() == AtomicRMWInst::Min ||
AI->getOperation() == AtomicRMWInst::Max) {
const DataLayout &DL = AI->getModule()->getDataLayout();
unsigned ValWidth =
DL.getTypeStoreSizeInBits(AI->getValOperand()->getType());
Value *SextShamt =
Builder.CreateSub(Builder.getIntN(XLen, XLen - ValWidth), ShiftAmt);
Result = Builder.CreateCall(LrwOpScwLoop,
{AlignedAddr, Incr, Mask, SextShamt, Ordering});
} else {
Result =
Builder.CreateCall(LrwOpScwLoop, {AlignedAddr, Incr, Mask, Ordering});
}
if (XLen == 64)
Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
return Result;
}
TargetLowering::AtomicExpansionKind
RISCVTargetLowering::shouldExpandAtomicCmpXchgInIR(
AtomicCmpXchgInst *CI) const {
// Don't expand forced atomics, we want to have __sync libcalls instead.
if (Subtarget.hasForcedAtomics())
return AtomicExpansionKind::None;
unsigned Size = CI->getCompareOperand()->getType()->getPrimitiveSizeInBits();
if (!(Subtarget.hasStdExtZabha() && Subtarget.hasStdExtZacas()) &&
(Size == 8 || Size == 16))
return AtomicExpansionKind::MaskedIntrinsic;
return AtomicExpansionKind::None;
}
Value *RISCVTargetLowering::emitMaskedAtomicCmpXchgIntrinsic(
IRBuilderBase &Builder, AtomicCmpXchgInst *CI, Value *AlignedAddr,
Value *CmpVal, Value *NewVal, Value *Mask, AtomicOrdering Ord) const {
unsigned XLen = Subtarget.getXLen();
Value *Ordering = Builder.getIntN(XLen, static_cast<uint64_t>(Ord));
Intrinsic::ID CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i32;
if (XLen == 64) {
CmpVal = Builder.CreateSExt(CmpVal, Builder.getInt64Ty());
NewVal = Builder.CreateSExt(NewVal, Builder.getInt64Ty());
Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty());
CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i64;
}
Type *Tys[] = {AlignedAddr->getType()};
Function *MaskedCmpXchg =
Intrinsic::getDeclaration(CI->getModule(), CmpXchgIntrID, Tys);
Value *Result = Builder.CreateCall(
MaskedCmpXchg, {AlignedAddr, CmpVal, NewVal, Mask, Ordering});
if (XLen == 64)
Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
return Result;
}
bool RISCVTargetLowering::shouldRemoveExtendFromGSIndex(SDValue Extend,
EVT DataVT) const {
// We have indexed loads for all supported EEW types. Indices are always
// zero extended.
return Extend.getOpcode() == ISD::ZERO_EXTEND &&
isTypeLegal(Extend.getValueType()) &&
isTypeLegal(Extend.getOperand(0).getValueType()) &&
Extend.getOperand(0).getValueType().getVectorElementType() != MVT::i1;
}
bool RISCVTargetLowering::shouldConvertFpToSat(unsigned Op, EVT FPVT,
EVT VT) const {
if (!isOperationLegalOrCustom(Op, VT) || !FPVT.isSimple())
return false;
switch (FPVT.getSimpleVT().SimpleTy) {
case MVT::f16:
return Subtarget.hasStdExtZfhmin();
case MVT::f32:
return Subtarget.hasStdExtF();
case MVT::f64:
return Subtarget.hasStdExtD();
default:
return false;
}
}
unsigned RISCVTargetLowering::getJumpTableEncoding() const {
// If we are using the small code model, we can reduce size of jump table
// entry to 4 bytes.
if (Subtarget.is64Bit() && !isPositionIndependent() &&
getTargetMachine().getCodeModel() == CodeModel::Small) {
return MachineJumpTableInfo::EK_Custom32;
}
return TargetLowering::getJumpTableEncoding();
}
const MCExpr *RISCVTargetLowering::LowerCustomJumpTableEntry(
const MachineJumpTableInfo *MJTI, const MachineBasicBlock *MBB,
unsigned uid, MCContext &Ctx) const {
assert(Subtarget.is64Bit() && !isPositionIndependent() &&
getTargetMachine().getCodeModel() == CodeModel::Small);
return MCSymbolRefExpr::create(MBB->getSymbol(), Ctx);
}
bool RISCVTargetLowering::isVScaleKnownToBeAPowerOfTwo() const {
// We define vscale to be VLEN/RVVBitsPerBlock. VLEN is always a power
// of two >= 64, and RVVBitsPerBlock is 64. Thus, vscale must be
// a power of two as well.
// FIXME: This doesn't work for zve32, but that's already broken
// elsewhere for the same reason.
assert(Subtarget.getRealMinVLen() >= 64 && "zve32* unsupported");
static_assert(RISCV::RVVBitsPerBlock == 64,
"RVVBitsPerBlock changed, audit needed");
return true;
}
bool RISCVTargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
SDValue &Offset,
ISD::MemIndexedMode &AM,
SelectionDAG &DAG) const {
// Target does not support indexed loads.
if (!Subtarget.hasVendorXTHeadMemIdx())
return false;
if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
return false;
Base = Op->getOperand(0);
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
int64_t RHSC = RHS->getSExtValue();
if (Op->getOpcode() == ISD::SUB)
RHSC = -(uint64_t)RHSC;
// The constants that can be encoded in the THeadMemIdx instructions
// are of the form (sign_extend(imm5) << imm2).
bool isLegalIndexedOffset = false;
for (unsigned i = 0; i < 4; i++)
if (isInt<5>(RHSC >> i) && ((RHSC % (1LL << i)) == 0)) {
isLegalIndexedOffset = true;
break;
}
if (!isLegalIndexedOffset)
return false;
Offset = Op->getOperand(1);
return true;
}
return false;
}
bool RISCVTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
SDValue &Offset,
ISD::MemIndexedMode &AM,
SelectionDAG &DAG) const {
EVT VT;
SDValue Ptr;
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
VT = LD->getMemoryVT();
Ptr = LD->getBasePtr();
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
VT = ST->getMemoryVT();
Ptr = ST->getBasePtr();
} else
return false;
if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, DAG))
return false;
AM = ISD::PRE_INC;
return true;
}
bool RISCVTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
SDValue &Base,
SDValue &Offset,
ISD::MemIndexedMode &AM,
SelectionDAG &DAG) const {
EVT VT;
SDValue Ptr;
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
VT = LD->getMemoryVT();
Ptr = LD->getBasePtr();
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
VT = ST->getMemoryVT();
Ptr = ST->getBasePtr();
} else
return false;
if (!getIndexedAddressParts(Op, Base, Offset, AM, DAG))
return false;
// Post-indexing updates the base, so it's not a valid transform
// if that's not the same as the load's pointer.
if (Ptr != Base)
return false;
AM = ISD::POST_INC;
return true;
}
bool RISCVTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
EVT VT) const {
EVT SVT = VT.getScalarType();
if (!SVT.isSimple())
return false;
switch (SVT.getSimpleVT().SimpleTy) {
case MVT::f16:
return VT.isVector() ? Subtarget.hasVInstructionsF16()
: Subtarget.hasStdExtZfhOrZhinx();
case MVT::f32:
return Subtarget.hasStdExtFOrZfinx();
case MVT::f64:
return Subtarget.hasStdExtDOrZdinx();
default:
break;
}
return false;
}
ISD::NodeType RISCVTargetLowering::getExtendForAtomicCmpSwapArg() const {
// Zacas will use amocas.w which does not require extension.
return Subtarget.hasStdExtZacas() ? ISD::ANY_EXTEND : ISD::SIGN_EXTEND;
}
Register RISCVTargetLowering::getExceptionPointerRegister(
const Constant *PersonalityFn) const {
return RISCV::X10;
}
Register RISCVTargetLowering::getExceptionSelectorRegister(
const Constant *PersonalityFn) const {
return RISCV::X11;
}
bool RISCVTargetLowering::shouldExtendTypeInLibCall(EVT Type) const {
// Return false to suppress the unnecessary extensions if the LibCall
// arguments or return value is a float narrower than XLEN on a soft FP ABI.
if (Subtarget.isSoftFPABI() && (Type.isFloatingPoint() && !Type.isVector() &&
Type.getSizeInBits() < Subtarget.getXLen()))
return false;
return true;
}
bool RISCVTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const {
if (Subtarget.is64Bit() && Type == MVT::i32)
return true;
return IsSigned;
}
bool RISCVTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT,
SDValue C) const {
// Check integral scalar types.
const bool HasExtMOrZmmul =
Subtarget.hasStdExtM() || Subtarget.hasStdExtZmmul();
if (!VT.isScalarInteger())
return false;
// Omit the optimization if the sub target has the M extension and the data
// size exceeds XLen.
if (HasExtMOrZmmul && VT.getSizeInBits() > Subtarget.getXLen())
return false;
if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) {
// Break the MUL to a SLLI and an ADD/SUB.
const APInt &Imm = ConstNode->getAPIntValue();
if ((Imm + 1).isPowerOf2() || (Imm - 1).isPowerOf2() ||
(1 - Imm).isPowerOf2() || (-1 - Imm).isPowerOf2())
return true;
// Optimize the MUL to (SH*ADD x, (SLLI x, bits)) if Imm is not simm12.
if (Subtarget.hasStdExtZba() && !Imm.isSignedIntN(12) &&
((Imm - 2).isPowerOf2() || (Imm - 4).isPowerOf2() ||
(Imm - 8).isPowerOf2()))
return true;
// Break the MUL to two SLLI instructions and an ADD/SUB, if Imm needs
// a pair of LUI/ADDI.
if (!Imm.isSignedIntN(12) && Imm.countr_zero() < 12 &&
ConstNode->hasOneUse()) {
APInt ImmS = Imm.ashr(Imm.countr_zero());
if ((ImmS + 1).isPowerOf2() || (ImmS - 1).isPowerOf2() ||
(1 - ImmS).isPowerOf2())
return true;
}
}
return false;
}
bool RISCVTargetLowering::isMulAddWithConstProfitable(SDValue AddNode,
SDValue ConstNode) const {
// Let the DAGCombiner decide for vectors.
EVT VT = AddNode.getValueType();
if (VT.isVector())
return true;
// Let the DAGCombiner decide for larger types.
if (VT.getScalarSizeInBits() > Subtarget.getXLen())
return true;
// It is worse if c1 is simm12 while c1*c2 is not.
ConstantSDNode *C1Node = cast<ConstantSDNode>(AddNode.getOperand(1));
ConstantSDNode *C2Node = cast<ConstantSDNode>(ConstNode);
const APInt &C1 = C1Node->getAPIntValue();
const APInt &C2 = C2Node->getAPIntValue();
if (C1.isSignedIntN(12) && !(C1 * C2).isSignedIntN(12))
return false;
// Default to true and let the DAGCombiner decide.
return true;
}
bool RISCVTargetLowering::allowsMisalignedMemoryAccesses(
EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
unsigned *Fast) const {
if (!VT.isVector()) {
if (Fast)
*Fast = Subtarget.enableUnalignedScalarMem();
return Subtarget.enableUnalignedScalarMem();
}
// All vector implementations must support element alignment
EVT ElemVT = VT.getVectorElementType();
if (Alignment >= ElemVT.getStoreSize()) {
if (Fast)
*Fast = 1;
return true;
}
// Note: We lower an unmasked unaligned vector access to an equally sized
// e8 element type access. Given this, we effectively support all unmasked
// misaligned accesses. TODO: Work through the codegen implications of
// allowing such accesses to be formed, and considered fast.
if (Fast)
*Fast = Subtarget.enableUnalignedVectorMem();
return Subtarget.enableUnalignedVectorMem();
}
EVT RISCVTargetLowering::getOptimalMemOpType(const MemOp &Op,
const AttributeList &FuncAttributes) const {
if (!Subtarget.hasVInstructions())
return MVT::Other;
if (FuncAttributes.hasFnAttr(Attribute::NoImplicitFloat))
return MVT::Other;
// We use LMUL1 memory operations here for a non-obvious reason. Our caller
// has an expansion threshold, and we want the number of hardware memory
// operations to correspond roughly to that threshold. LMUL>1 operations
// are typically expanded linearly internally, and thus correspond to more
// than one actual memory operation. Note that store merging and load
// combining will typically form larger LMUL operations from the LMUL1
// operations emitted here, and that's okay because combining isn't
// introducing new memory operations; it's just merging existing ones.
const unsigned MinVLenInBytes = Subtarget.getRealMinVLen()/8;
if (Op.size() < MinVLenInBytes)
// TODO: Figure out short memops. For the moment, do the default thing
// which ends up using scalar sequences.
return MVT::Other;
// Prefer i8 for non-zero memset as it allows us to avoid materializing
// a large scalar constant and instead use vmv.v.x/i to do the
// broadcast. For everything else, prefer ELenVT to minimize VL and thus
// maximize the chance we can encode the size in the vsetvli.
MVT ELenVT = MVT::getIntegerVT(Subtarget.getELen());
MVT PreferredVT = (Op.isMemset() && !Op.isZeroMemset()) ? MVT::i8 : ELenVT;
// Do we have sufficient alignment for our preferred VT? If not, revert
// to largest size allowed by our alignment criteria.
if (PreferredVT != MVT::i8 && !Subtarget.enableUnalignedVectorMem()) {
Align RequiredAlign(PreferredVT.getStoreSize());
if (Op.isFixedDstAlign())
RequiredAlign = std::min(RequiredAlign, Op.getDstAlign());
if (Op.isMemcpy())
RequiredAlign = std::min(RequiredAlign, Op.getSrcAlign());
PreferredVT = MVT::getIntegerVT(RequiredAlign.value() * 8);
}
return MVT::getVectorVT(PreferredVT, MinVLenInBytes/PreferredVT.getStoreSize());
}
bool RISCVTargetLowering::splitValueIntoRegisterParts(
SelectionDAG &DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) const {
bool IsABIRegCopy = CC.has_value();
EVT ValueVT = Val.getValueType();
if (IsABIRegCopy && (ValueVT == MVT::f16 || ValueVT == MVT::bf16) &&
PartVT == MVT::f32) {
// Cast the [b]f16 to i16, extend to i32, pad with ones to make a float
// nan, and cast to f32.
Val = DAG.getNode(ISD::BITCAST, DL, MVT::i16, Val);
Val = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Val);
Val = DAG.getNode(ISD::OR, DL, MVT::i32, Val,
DAG.getConstant(0xFFFF0000, DL, MVT::i32));
Val = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
Parts[0] = Val;
return true;
}
if (ValueVT.isScalableVector() && PartVT.isScalableVector()) {
LLVMContext &Context = *DAG.getContext();
EVT ValueEltVT = ValueVT.getVectorElementType();
EVT PartEltVT = PartVT.getVectorElementType();
unsigned ValueVTBitSize = ValueVT.getSizeInBits().getKnownMinValue();
unsigned PartVTBitSize = PartVT.getSizeInBits().getKnownMinValue();
if (PartVTBitSize % ValueVTBitSize == 0) {
assert(PartVTBitSize >= ValueVTBitSize);
// If the element types are different, bitcast to the same element type of
// PartVT first.
// Give an example here, we want copy a <vscale x 1 x i8> value to
// <vscale x 4 x i16>.
// We need to convert <vscale x 1 x i8> to <vscale x 8 x i8> by insert
// subvector, then we can bitcast to <vscale x 4 x i16>.
if (ValueEltVT != PartEltVT) {
if (PartVTBitSize > ValueVTBitSize) {
unsigned Count = PartVTBitSize / ValueEltVT.getFixedSizeInBits();
assert(Count != 0 && "The number of element should not be zero.");
EVT SameEltTypeVT =
EVT::getVectorVT(Context, ValueEltVT, Count, /*IsScalable=*/true);
Val = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, SameEltTypeVT,
DAG.getUNDEF(SameEltTypeVT), Val,
DAG.getVectorIdxConstant(0, DL));
}
Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
} else {
Val =
DAG.getNode(ISD::INSERT_SUBVECTOR, DL, PartVT, DAG.getUNDEF(PartVT),
Val, DAG.getVectorIdxConstant(0, DL));
}
Parts[0] = Val;
return true;
}
}
return false;
}
SDValue RISCVTargetLowering::joinRegisterPartsIntoValue(
SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts, unsigned NumParts,
MVT PartVT, EVT ValueVT, std::optional<CallingConv::ID> CC) const {
bool IsABIRegCopy = CC.has_value();
if (IsABIRegCopy && (ValueVT == MVT::f16 || ValueVT == MVT::bf16) &&
PartVT == MVT::f32) {
SDValue Val = Parts[0];
// Cast the f32 to i32, truncate to i16, and cast back to [b]f16.
Val = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Val);
Val = DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Val);
Val = DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
return Val;
}
if (ValueVT.isScalableVector() && PartVT.isScalableVector()) {
LLVMContext &Context = *DAG.getContext();
SDValue Val = Parts[0];
EVT ValueEltVT = ValueVT.getVectorElementType();
EVT PartEltVT = PartVT.getVectorElementType();
unsigned ValueVTBitSize = ValueVT.getSizeInBits().getKnownMinValue();
unsigned PartVTBitSize = PartVT.getSizeInBits().getKnownMinValue();
if (PartVTBitSize % ValueVTBitSize == 0) {
assert(PartVTBitSize >= ValueVTBitSize);
EVT SameEltTypeVT = ValueVT;
// If the element types are different, convert it to the same element type
// of PartVT.
// Give an example here, we want copy a <vscale x 1 x i8> value from
// <vscale x 4 x i16>.
// We need to convert <vscale x 4 x i16> to <vscale x 8 x i8> first,
// then we can extract <vscale x 1 x i8>.
if (ValueEltVT != PartEltVT) {
unsigned Count = PartVTBitSize / ValueEltVT.getFixedSizeInBits();
assert(Count != 0 && "The number of element should not be zero.");
SameEltTypeVT =
EVT::getVectorVT(Context, ValueEltVT, Count, /*IsScalable=*/true);
Val = DAG.getNode(ISD::BITCAST, DL, SameEltTypeVT, Val);
}
Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
DAG.getVectorIdxConstant(0, DL));
return Val;
}
}
return SDValue();
}
bool RISCVTargetLowering::isIntDivCheap(EVT VT, AttributeList Attr) const {
// When aggressively optimizing for code size, we prefer to use a div
// instruction, as it is usually smaller than the alternative sequence.
// TODO: Add vector division?
bool OptSize = Attr.hasFnAttr(Attribute::MinSize);
return OptSize && !VT.isVector();
}
bool RISCVTargetLowering::preferScalarizeSplat(SDNode *N) const {
// Scalarize zero_ext and sign_ext might stop match to widening instruction in
// some situation.
unsigned Opc = N->getOpcode();
if (Opc == ISD::ZERO_EXTEND || Opc == ISD::SIGN_EXTEND)
return false;
return true;
}
static Value *useTpOffset(IRBuilderBase &IRB, unsigned Offset) {
Module *M = IRB.GetInsertBlock()->getParent()->getParent();
Function *ThreadPointerFunc =
Intrinsic::getDeclaration(M, Intrinsic::thread_pointer);
return IRB.CreateConstGEP1_32(IRB.getInt8Ty(),
IRB.CreateCall(ThreadPointerFunc), Offset);
}
Value *RISCVTargetLowering::getIRStackGuard(IRBuilderBase &IRB) const {
// Fuchsia provides a fixed TLS slot for the stack cookie.
// <zircon/tls.h> defines ZX_TLS_STACK_GUARD_OFFSET with this value.
if (Subtarget.isTargetFuchsia())
return useTpOffset(IRB, -0x10);
return TargetLowering::getIRStackGuard(IRB);
}
bool RISCVTargetLowering::isLegalInterleavedAccessType(
VectorType *VTy, unsigned Factor, Align Alignment, unsigned AddrSpace,
const DataLayout &DL) const {
EVT VT = getValueType(DL, VTy);
// Don't lower vlseg/vsseg for vector types that can't be split.
if (!isTypeLegal(VT))
return false;
if (!isLegalElementTypeForRVV(VT.getScalarType()) ||
!allowsMemoryAccessForAlignment(VTy->getContext(), DL, VT, AddrSpace,
Alignment))
return false;
MVT ContainerVT = VT.getSimpleVT();
if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
if (!Subtarget.useRVVForFixedLengthVectors())
return false;
// Sometimes the interleaved access pass picks up splats as interleaves of
// one element. Don't lower these.
if (FVTy->getNumElements() < 2)
return false;
ContainerVT = getContainerForFixedLengthVector(VT.getSimpleVT());
}
// Need to make sure that EMUL * NFIELDS ≤ 8
auto [LMUL, Fractional] = RISCVVType::decodeVLMUL(getLMUL(ContainerVT));
if (Fractional)
return true;
return Factor * LMUL <= 8;
}
bool RISCVTargetLowering::isLegalStridedLoadStore(EVT DataType,
Align Alignment) const {
if (!Subtarget.hasVInstructions())
return false;
// Only support fixed vectors if we know the minimum vector size.
if (DataType.isFixedLengthVector() && !Subtarget.useRVVForFixedLengthVectors())
return false;
EVT ScalarType = DataType.getScalarType();
if (!isLegalElementTypeForRVV(ScalarType))
return false;
if (!Subtarget.enableUnalignedVectorMem() &&
Alignment < ScalarType.getStoreSize())
return false;
return true;
}
static const Intrinsic::ID FixedVlsegIntrIds[] = {
Intrinsic::riscv_seg2_load, Intrinsic::riscv_seg3_load,
Intrinsic::riscv_seg4_load, Intrinsic::riscv_seg5_load,
Intrinsic::riscv_seg6_load, Intrinsic::riscv_seg7_load,
Intrinsic::riscv_seg8_load};
/// Lower an interleaved load into a vlsegN intrinsic.
///
/// E.g. Lower an interleaved load (Factor = 2):
/// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
/// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
/// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
///
/// Into:
/// %ld2 = { <4 x i32>, <4 x i32> } call llvm.riscv.seg2.load.v4i32.p0.i64(
/// %ptr, i64 4)
/// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
/// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
bool RISCVTargetLowering::lowerInterleavedLoad(
LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
ArrayRef<unsigned> Indices, unsigned Factor) const {
IRBuilder<> Builder(LI);
auto *VTy = cast<FixedVectorType>(Shuffles[0]->getType());
if (!isLegalInterleavedAccessType(VTy, Factor, LI->getAlign(),
LI->getPointerAddressSpace(),
LI->getModule()->getDataLayout()))
return false;
auto *XLenTy = Type::getIntNTy(LI->getContext(), Subtarget.getXLen());
Function *VlsegNFunc =
Intrinsic::getDeclaration(LI->getModule(), FixedVlsegIntrIds[Factor - 2],
{VTy, LI->getPointerOperandType(), XLenTy});
Value *VL = ConstantInt::get(XLenTy, VTy->getNumElements());
CallInst *VlsegN =
Builder.CreateCall(VlsegNFunc, {LI->getPointerOperand(), VL});
for (unsigned i = 0; i < Shuffles.size(); i++) {
Value *SubVec = Builder.CreateExtractValue(VlsegN, Indices[i]);
Shuffles[i]->replaceAllUsesWith(SubVec);
}
return true;
}
static const Intrinsic::ID FixedVssegIntrIds[] = {
Intrinsic::riscv_seg2_store, Intrinsic::riscv_seg3_store,
Intrinsic::riscv_seg4_store, Intrinsic::riscv_seg5_store,
Intrinsic::riscv_seg6_store, Intrinsic::riscv_seg7_store,
Intrinsic::riscv_seg8_store};
/// Lower an interleaved store into a vssegN intrinsic.
///
/// E.g. Lower an interleaved store (Factor = 3):
/// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
/// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
/// store <12 x i32> %i.vec, <12 x i32>* %ptr
///
/// Into:
/// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
/// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
/// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
/// call void llvm.riscv.seg3.store.v4i32.p0.i64(%sub.v0, %sub.v1, %sub.v2,
/// %ptr, i32 4)
///
/// Note that the new shufflevectors will be removed and we'll only generate one
/// vsseg3 instruction in CodeGen.
bool RISCVTargetLowering::lowerInterleavedStore(StoreInst *SI,
ShuffleVectorInst *SVI,
unsigned Factor) const {
IRBuilder<> Builder(SI);
auto *ShuffleVTy = cast<FixedVectorType>(SVI->getType());
// Given SVI : <n*factor x ty>, then VTy : <n x ty>
auto *VTy = FixedVectorType::get(ShuffleVTy->getElementType(),
ShuffleVTy->getNumElements() / Factor);
if (!isLegalInterleavedAccessType(VTy, Factor, SI->getAlign(),
SI->getPointerAddressSpace(),
SI->getModule()->getDataLayout()))
return false;
auto *XLenTy = Type::getIntNTy(SI->getContext(), Subtarget.getXLen());
Function *VssegNFunc =
Intrinsic::getDeclaration(SI->getModule(), FixedVssegIntrIds[Factor - 2],
{VTy, SI->getPointerOperandType(), XLenTy});
auto Mask = SVI->getShuffleMask();
SmallVector<Value *, 10> Ops;
for (unsigned i = 0; i < Factor; i++) {
Value *Shuffle = Builder.CreateShuffleVector(
SVI->getOperand(0), SVI->getOperand(1),
createSequentialMask(Mask[i], VTy->getNumElements(), 0));
Ops.push_back(Shuffle);
}
// This VL should be OK (should be executable in one vsseg instruction,
// potentially under larger LMULs) because we checked that the fixed vector
// type fits in isLegalInterleavedAccessType
Value *VL = ConstantInt::get(XLenTy, VTy->getNumElements());
Ops.append({SI->getPointerOperand(), VL});
Builder.CreateCall(VssegNFunc, Ops);
return true;
}
bool RISCVTargetLowering::lowerDeinterleaveIntrinsicToLoad(IntrinsicInst *DI,
LoadInst *LI) const {
assert(LI->isSimple());
IRBuilder<> Builder(LI);
// Only deinterleave2 supported at present.
if (DI->getIntrinsicID() != Intrinsic::experimental_vector_deinterleave2)
return false;
unsigned Factor = 2;
VectorType *VTy = cast<VectorType>(DI->getOperand(0)->getType());
VectorType *ResVTy = cast<VectorType>(DI->getType()->getContainedType(0));
if (!isLegalInterleavedAccessType(ResVTy, Factor, LI->getAlign(),
LI->getPointerAddressSpace(),
LI->getModule()->getDataLayout()))
return false;
Function *VlsegNFunc;
Value *VL;
Type *XLenTy = Type::getIntNTy(LI->getContext(), Subtarget.getXLen());
SmallVector<Value *, 10> Ops;
if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
VlsegNFunc = Intrinsic::getDeclaration(
LI->getModule(), FixedVlsegIntrIds[Factor - 2],
{ResVTy, LI->getPointerOperandType(), XLenTy});
VL = ConstantInt::get(XLenTy, FVTy->getNumElements());
} else {
static const Intrinsic::ID IntrIds[] = {
Intrinsic::riscv_vlseg2, Intrinsic::riscv_vlseg3,
Intrinsic::riscv_vlseg4, Intrinsic::riscv_vlseg5,
Intrinsic::riscv_vlseg6, Intrinsic::riscv_vlseg7,
Intrinsic::riscv_vlseg8};
VlsegNFunc = Intrinsic::getDeclaration(LI->getModule(), IntrIds[Factor - 2],
{ResVTy, XLenTy});
VL = Constant::getAllOnesValue(XLenTy);
Ops.append(Factor, PoisonValue::get(ResVTy));
}
Ops.append({LI->getPointerOperand(), VL});
Value *Vlseg = Builder.CreateCall(VlsegNFunc, Ops);
DI->replaceAllUsesWith(Vlseg);
return true;
}
bool RISCVTargetLowering::lowerInterleaveIntrinsicToStore(IntrinsicInst *II,
StoreInst *SI) const {
assert(SI->isSimple());
IRBuilder<> Builder(SI);
// Only interleave2 supported at present.
if (II->getIntrinsicID() != Intrinsic::experimental_vector_interleave2)
return false;
unsigned Factor = 2;
VectorType *VTy = cast<VectorType>(II->getType());
VectorType *InVTy = cast<VectorType>(II->getOperand(0)->getType());
if (!isLegalInterleavedAccessType(InVTy, Factor, SI->getAlign(),
SI->getPointerAddressSpace(),
SI->getModule()->getDataLayout()))
return false;
Function *VssegNFunc;
Value *VL;
Type *XLenTy = Type::getIntNTy(SI->getContext(), Subtarget.getXLen());
if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
VssegNFunc = Intrinsic::getDeclaration(
SI->getModule(), FixedVssegIntrIds[Factor - 2],
{InVTy, SI->getPointerOperandType(), XLenTy});
VL = ConstantInt::get(XLenTy, FVTy->getNumElements());
} else {
static const Intrinsic::ID IntrIds[] = {
Intrinsic::riscv_vsseg2, Intrinsic::riscv_vsseg3,
Intrinsic::riscv_vsseg4, Intrinsic::riscv_vsseg5,
Intrinsic::riscv_vsseg6, Intrinsic::riscv_vsseg7,
Intrinsic::riscv_vsseg8};
VssegNFunc = Intrinsic::getDeclaration(SI->getModule(), IntrIds[Factor - 2],
{InVTy, XLenTy});
VL = Constant::getAllOnesValue(XLenTy);
}
Builder.CreateCall(VssegNFunc, {II->getOperand(0), II->getOperand(1),
SI->getPointerOperand(), VL});
return true;
}
MachineInstr *
RISCVTargetLowering::EmitKCFICheck(MachineBasicBlock &MBB,
MachineBasicBlock::instr_iterator &MBBI,
const TargetInstrInfo *TII) const {
assert(MBBI->isCall() && MBBI->getCFIType() &&
"Invalid call instruction for a KCFI check");
assert(is_contained({RISCV::PseudoCALLIndirect, RISCV::PseudoTAILIndirect},
MBBI->getOpcode()));
MachineOperand &Target = MBBI->getOperand(0);
Target.setIsRenamable(false);
return BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(RISCV::KCFI_CHECK))
.addReg(Target.getReg())
.addImm(MBBI->getCFIType())
.getInstr();
}
#define GET_REGISTER_MATCHER
#include "RISCVGenAsmMatcher.inc"
Register
RISCVTargetLowering::getRegisterByName(const char *RegName, LLT VT,
const MachineFunction &MF) const {
Register Reg = MatchRegisterAltName(RegName);
if (Reg == RISCV::NoRegister)
Reg = MatchRegisterName(RegName);
if (Reg == RISCV::NoRegister)
report_fatal_error(
Twine("Invalid register name \"" + StringRef(RegName) + "\"."));
BitVector ReservedRegs = Subtarget.getRegisterInfo()->getReservedRegs(MF);
if (!ReservedRegs.test(Reg) && !Subtarget.isRegisterReservedByUser(Reg))
report_fatal_error(Twine("Trying to obtain non-reserved register \"" +
StringRef(RegName) + "\"."));
return Reg;
}
MachineMemOperand::Flags
RISCVTargetLowering::getTargetMMOFlags(const Instruction &I) const {
const MDNode *NontemporalInfo = I.getMetadata(LLVMContext::MD_nontemporal);
if (NontemporalInfo == nullptr)
return MachineMemOperand::MONone;
// 1 for default value work as __RISCV_NTLH_ALL
// 2 -> __RISCV_NTLH_INNERMOST_PRIVATE
// 3 -> __RISCV_NTLH_ALL_PRIVATE
// 4 -> __RISCV_NTLH_INNERMOST_SHARED
// 5 -> __RISCV_NTLH_ALL
int NontemporalLevel = 5;
const MDNode *RISCVNontemporalInfo =
I.getMetadata("riscv-nontemporal-domain");
if (RISCVNontemporalInfo != nullptr)
NontemporalLevel =
cast<ConstantInt>(
cast<ConstantAsMetadata>(RISCVNontemporalInfo->getOperand(0))
->getValue())
->getZExtValue();
assert((1 <= NontemporalLevel && NontemporalLevel <= 5) &&
"RISC-V target doesn't support this non-temporal domain.");
NontemporalLevel -= 2;
MachineMemOperand::Flags Flags = MachineMemOperand::MONone;
if (NontemporalLevel & 0b1)
Flags |= MONontemporalBit0;
if (NontemporalLevel & 0b10)
Flags |= MONontemporalBit1;
return Flags;
}
MachineMemOperand::Flags
RISCVTargetLowering::getTargetMMOFlags(const MemSDNode &Node) const {
MachineMemOperand::Flags NodeFlags = Node.getMemOperand()->getFlags();
MachineMemOperand::Flags TargetFlags = MachineMemOperand::MONone;
TargetFlags |= (NodeFlags & MONontemporalBit0);
TargetFlags |= (NodeFlags & MONontemporalBit1);
return TargetFlags;
}
bool RISCVTargetLowering::areTwoSDNodeTargetMMOFlagsMergeable(
const MemSDNode &NodeX, const MemSDNode &NodeY) const {
return getTargetMMOFlags(NodeX) == getTargetMMOFlags(NodeY);
}
bool RISCVTargetLowering::isCtpopFast(EVT VT) const {
if (VT.isScalableVector())
return isTypeLegal(VT) && Subtarget.hasStdExtZvbb();
if (VT.isFixedLengthVector() && Subtarget.hasStdExtZvbb())
return true;
return Subtarget.hasStdExtZbb() &&
(VT == MVT::i32 || VT == MVT::i64 || VT.isFixedLengthVector());
}
unsigned RISCVTargetLowering::getCustomCtpopCost(EVT VT,
ISD::CondCode Cond) const {
return isCtpopFast(VT) ? 0 : 1;
}
bool RISCVTargetLowering::fallBackToDAGISel(const Instruction &Inst) const {
// GISel support is in progress or complete for these opcodes.
unsigned Op = Inst.getOpcode();
if (Op == Instruction::Add || Op == Instruction::Sub ||
Op == Instruction::And || Op == Instruction::Or ||
Op == Instruction::Xor || Op == Instruction::InsertElement ||
Op == Instruction::ShuffleVector || Op == Instruction::Load)
return false;
if (Inst.getType()->isScalableTy())
return true;
for (unsigned i = 0; i < Inst.getNumOperands(); ++i)
if (Inst.getOperand(i)->getType()->isScalableTy() &&
!isa<ReturnInst>(&Inst))
return true;
if (const AllocaInst *AI = dyn_cast<AllocaInst>(&Inst)) {
if (AI->getAllocatedType()->isScalableTy())
return true;
}
return false;
}
SDValue
RISCVTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
SelectionDAG &DAG,
SmallVectorImpl<SDNode *> &Created) const {
AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
if (isIntDivCheap(N->getValueType(0), Attr))
return SDValue(N, 0); // Lower SDIV as SDIV
// Only perform this transform if short forward branch opt is supported.
if (!Subtarget.hasShortForwardBranchOpt())
return SDValue();
EVT VT = N->getValueType(0);
if (!(VT == MVT::i32 || (VT == MVT::i64 && Subtarget.is64Bit())))
return SDValue();
// Ensure 2**k-1 < 2048 so that we can just emit a single addi/addiw.
if (Divisor.sgt(2048) || Divisor.slt(-2048))
return SDValue();
return TargetLowering::buildSDIVPow2WithCMov(N, Divisor, DAG, Created);
}
bool RISCVTargetLowering::shouldFoldSelectWithSingleBitTest(
EVT VT, const APInt &AndMask) const {
if (Subtarget.hasStdExtZicond() || Subtarget.hasVendorXVentanaCondOps())
return !Subtarget.hasStdExtZbs() && AndMask.ugt(1024);
return TargetLowering::shouldFoldSelectWithSingleBitTest(VT, AndMask);
}
unsigned RISCVTargetLowering::getMinimumJumpTableEntries() const {
return Subtarget.getMinimumJumpTableEntries();
}
// Handle single arg such as return value.
template <typename Arg>
void RVVArgDispatcher::constructArgInfos(ArrayRef<Arg> ArgList) {
// This lambda determines whether an array of types are constructed by
// homogeneous vector types.
auto isHomogeneousScalableVectorType = [](ArrayRef<Arg> ArgList) {
// First, extract the first element in the argument type.
auto It = ArgList.begin();
MVT FirstArgRegType = It->VT;
// Return if there is no return or the type needs split.
if (It == ArgList.end() || It->Flags.isSplit())
return false;
++It;
// Return if this argument type contains only 1 element, or it's not a
// vector type.
if (It == ArgList.end() || !FirstArgRegType.isScalableVector())
return false;
// Second, check if the following elements in this argument type are all the
// same.
for (; It != ArgList.end(); ++It)
if (It->Flags.isSplit() || It->VT != FirstArgRegType)
return false;
return true;
};
if (isHomogeneousScalableVectorType(ArgList)) {
// Handle as tuple type
RVVArgInfos.push_back({(unsigned)ArgList.size(), ArgList[0].VT, false});
} else {
// Handle as normal vector type
bool FirstVMaskAssigned = false;
for (const auto &OutArg : ArgList) {
MVT RegisterVT = OutArg.VT;
// Skip non-RVV register type
if (!RegisterVT.isVector())
continue;
if (RegisterVT.isFixedLengthVector())
RegisterVT = TLI->getContainerForFixedLengthVector(RegisterVT);
if (!FirstVMaskAssigned && RegisterVT.getVectorElementType() == MVT::i1) {
RVVArgInfos.push_back({1, RegisterVT, true});
FirstVMaskAssigned = true;
continue;
}
RVVArgInfos.push_back({1, RegisterVT, false});
}
}
}
// Handle multiple args.
template <>
void RVVArgDispatcher::constructArgInfos<Type *>(ArrayRef<Type *> TypeList) {
const DataLayout &DL = MF->getDataLayout();
const Function &F = MF->getFunction();
LLVMContext &Context = F.getContext();
bool FirstVMaskAssigned = false;
for (Type *Ty : TypeList) {
StructType *STy = dyn_cast<StructType>(Ty);
if (STy && STy->containsHomogeneousScalableVectorTypes()) {
Type *ElemTy = STy->getTypeAtIndex(0U);
EVT VT = TLI->getValueType(DL, ElemTy);
MVT RegisterVT =
TLI->getRegisterTypeForCallingConv(Context, F.getCallingConv(), VT);
unsigned NumRegs =
TLI->getNumRegistersForCallingConv(Context, F.getCallingConv(), VT);
RVVArgInfos.push_back(
{NumRegs * STy->getNumElements(), RegisterVT, false});
} else {
SmallVector<EVT, 4> ValueVTs;
ComputeValueVTs(*TLI, DL, Ty, ValueVTs);
for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
++Value) {
EVT VT = ValueVTs[Value];
MVT RegisterVT =
TLI->getRegisterTypeForCallingConv(Context, F.getCallingConv(), VT);
unsigned NumRegs =
TLI->getNumRegistersForCallingConv(Context, F.getCallingConv(), VT);
// Skip non-RVV register type
if (!RegisterVT.isVector())
continue;
if (RegisterVT.isFixedLengthVector())
RegisterVT = TLI->getContainerForFixedLengthVector(RegisterVT);
if (!FirstVMaskAssigned &&
RegisterVT.getVectorElementType() == MVT::i1) {
RVVArgInfos.push_back({1, RegisterVT, true});
FirstVMaskAssigned = true;
--NumRegs;
}
RVVArgInfos.insert(RVVArgInfos.end(), NumRegs, {1, RegisterVT, false});
}
}
}
}
void RVVArgDispatcher::allocatePhysReg(unsigned NF, unsigned LMul,
unsigned StartReg) {
assert((StartReg % LMul) == 0 &&
"Start register number should be multiple of lmul");
const MCPhysReg *VRArrays;
switch (LMul) {
default:
report_fatal_error("Invalid lmul");
case 1:
VRArrays = ArgVRs;
break;
case 2:
VRArrays = ArgVRM2s;
break;
case 4:
VRArrays = ArgVRM4s;
break;
case 8:
VRArrays = ArgVRM8s;
break;
}
for (unsigned i = 0; i < NF; ++i)
if (StartReg)
AllocatedPhysRegs.push_back(VRArrays[(StartReg - 8) / LMul + i]);
else
AllocatedPhysRegs.push_back(MCPhysReg());
}
/// This function determines if each RVV argument is passed by register, if the
/// argument can be assigned to a VR, then give it a specific register.
/// Otherwise, assign the argument to 0 which is a invalid MCPhysReg.
void RVVArgDispatcher::compute() {
uint32_t AssignedMap = 0;
auto allocate = [&](const RVVArgInfo &ArgInfo) {
// Allocate first vector mask argument to V0.
if (ArgInfo.FirstVMask) {
AllocatedPhysRegs.push_back(RISCV::V0);
return;
}
unsigned RegsNeeded = divideCeil(
ArgInfo.VT.getSizeInBits().getKnownMinValue(), RISCV::RVVBitsPerBlock);
unsigned TotalRegsNeeded = ArgInfo.NF * RegsNeeded;
for (unsigned StartReg = 0; StartReg + TotalRegsNeeded <= NumArgVRs;
StartReg += RegsNeeded) {
uint32_t Map = ((1 << TotalRegsNeeded) - 1) << StartReg;
if ((AssignedMap & Map) == 0) {
allocatePhysReg(ArgInfo.NF, RegsNeeded, StartReg + 8);
AssignedMap |= Map;
return;
}
}
allocatePhysReg(ArgInfo.NF, RegsNeeded, 0);
};
for (unsigned i = 0; i < RVVArgInfos.size(); ++i)
allocate(RVVArgInfos[i]);
}
MCPhysReg RVVArgDispatcher::getNextPhysReg() {
assert(CurIdx < AllocatedPhysRegs.size() && "Index out of range");
return AllocatedPhysRegs[CurIdx++];
}
namespace llvm::RISCVVIntrinsicsTable {
#define GET_RISCVVIntrinsicsTable_IMPL
#include "RISCVGenSearchableTables.inc"
} // namespace llvm::RISCVVIntrinsicsTable
|