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
|
/* SLP - Basic Block Vectorization
Copyright (C) 2007-2022 Free Software Foundation, Inc.
Contributed by Dorit Naishlos <dorit@il.ibm.com>
and Ira Rosen <irar@il.ibm.com>
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#define INCLUDE_ALGORITHM
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "target.h"
#include "rtl.h"
#include "tree.h"
#include "gimple.h"
#include "tree-pass.h"
#include "ssa.h"
#include "optabs-tree.h"
#include "insn-config.h"
#include "recog.h" /* FIXME: for insn_data */
#include "fold-const.h"
#include "stor-layout.h"
#include "gimple-iterator.h"
#include "cfgloop.h"
#include "tree-vectorizer.h"
#include "langhooks.h"
#include "gimple-walk.h"
#include "dbgcnt.h"
#include "tree-vector-builder.h"
#include "vec-perm-indices.h"
#include "gimple-fold.h"
#include "internal-fn.h"
#include "dump-context.h"
#include "cfganal.h"
#include "tree-eh.h"
#include "tree-cfg.h"
#include "alloc-pool.h"
#include "sreal.h"
#include "predict.h"
static bool vect_transform_slp_perm_load_1 (vec_info *, slp_tree,
load_permutation_t &,
const vec<tree> &,
gimple_stmt_iterator *,
poly_uint64, bool, bool,
unsigned *,
unsigned * = nullptr,
bool = false);
static int vectorizable_slp_permutation_1 (vec_info *, gimple_stmt_iterator *,
slp_tree, lane_permutation_t &,
vec<slp_tree> &, bool);
static bool vectorizable_slp_permutation (vec_info *, gimple_stmt_iterator *,
slp_tree, stmt_vector_for_cost *);
static void vect_print_slp_tree (dump_flags_t, dump_location_t, slp_tree);
static object_allocator<_slp_tree> *slp_tree_pool;
static slp_tree slp_first_node;
void
vect_slp_init (void)
{
slp_tree_pool = new object_allocator<_slp_tree> ("SLP nodes");
}
void
vect_slp_fini (void)
{
while (slp_first_node)
delete slp_first_node;
delete slp_tree_pool;
slp_tree_pool = NULL;
}
void *
_slp_tree::operator new (size_t n)
{
gcc_assert (n == sizeof (_slp_tree));
return slp_tree_pool->allocate_raw ();
}
void
_slp_tree::operator delete (void *node, size_t n)
{
gcc_assert (n == sizeof (_slp_tree));
slp_tree_pool->remove_raw (node);
}
/* Initialize a SLP node. */
_slp_tree::_slp_tree ()
{
this->prev_node = NULL;
if (slp_first_node)
slp_first_node->prev_node = this;
this->next_node = slp_first_node;
slp_first_node = this;
SLP_TREE_SCALAR_STMTS (this) = vNULL;
SLP_TREE_SCALAR_OPS (this) = vNULL;
SLP_TREE_VEC_STMTS (this) = vNULL;
SLP_TREE_VEC_DEFS (this) = vNULL;
SLP_TREE_NUMBER_OF_VEC_STMTS (this) = 0;
SLP_TREE_CHILDREN (this) = vNULL;
SLP_TREE_LOAD_PERMUTATION (this) = vNULL;
SLP_TREE_LANE_PERMUTATION (this) = vNULL;
SLP_TREE_DEF_TYPE (this) = vect_uninitialized_def;
SLP_TREE_CODE (this) = ERROR_MARK;
SLP_TREE_VECTYPE (this) = NULL_TREE;
SLP_TREE_REPRESENTATIVE (this) = NULL;
SLP_TREE_REF_COUNT (this) = 1;
this->failed = NULL;
this->max_nunits = 1;
this->lanes = 0;
}
/* Tear down a SLP node. */
_slp_tree::~_slp_tree ()
{
if (this->prev_node)
this->prev_node->next_node = this->next_node;
else
slp_first_node = this->next_node;
if (this->next_node)
this->next_node->prev_node = this->prev_node;
SLP_TREE_CHILDREN (this).release ();
SLP_TREE_SCALAR_STMTS (this).release ();
SLP_TREE_SCALAR_OPS (this).release ();
SLP_TREE_VEC_STMTS (this).release ();
SLP_TREE_VEC_DEFS (this).release ();
SLP_TREE_LOAD_PERMUTATION (this).release ();
SLP_TREE_LANE_PERMUTATION (this).release ();
if (this->failed)
free (failed);
}
/* Recursively free the memory allocated for the SLP tree rooted at NODE. */
void
vect_free_slp_tree (slp_tree node)
{
int i;
slp_tree child;
if (--SLP_TREE_REF_COUNT (node) != 0)
return;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
vect_free_slp_tree (child);
/* If the node defines any SLP only patterns then those patterns are no
longer valid and should be removed. */
stmt_vec_info rep_stmt_info = SLP_TREE_REPRESENTATIVE (node);
if (rep_stmt_info && STMT_VINFO_SLP_VECT_ONLY_PATTERN (rep_stmt_info))
{
stmt_vec_info stmt_info = vect_orig_stmt (rep_stmt_info);
STMT_VINFO_IN_PATTERN_P (stmt_info) = false;
STMT_SLP_TYPE (stmt_info) = STMT_SLP_TYPE (rep_stmt_info);
}
delete node;
}
/* Return a location suitable for dumpings related to the SLP instance. */
dump_user_location_t
_slp_instance::location () const
{
if (!root_stmts.is_empty ())
return root_stmts[0]->stmt;
else
return SLP_TREE_SCALAR_STMTS (root)[0]->stmt;
}
/* Free the memory allocated for the SLP instance. */
void
vect_free_slp_instance (slp_instance instance)
{
vect_free_slp_tree (SLP_INSTANCE_TREE (instance));
SLP_INSTANCE_LOADS (instance).release ();
SLP_INSTANCE_ROOT_STMTS (instance).release ();
instance->subgraph_entries.release ();
instance->cost_vec.release ();
free (instance);
}
/* Create an SLP node for SCALAR_STMTS. */
slp_tree
vect_create_new_slp_node (unsigned nops, tree_code code)
{
slp_tree node = new _slp_tree;
SLP_TREE_SCALAR_STMTS (node) = vNULL;
SLP_TREE_CHILDREN (node).create (nops);
SLP_TREE_DEF_TYPE (node) = vect_internal_def;
SLP_TREE_CODE (node) = code;
return node;
}
/* Create an SLP node for SCALAR_STMTS. */
static slp_tree
vect_create_new_slp_node (slp_tree node,
vec<stmt_vec_info> scalar_stmts, unsigned nops)
{
SLP_TREE_SCALAR_STMTS (node) = scalar_stmts;
SLP_TREE_CHILDREN (node).create (nops);
SLP_TREE_DEF_TYPE (node) = vect_internal_def;
SLP_TREE_REPRESENTATIVE (node) = scalar_stmts[0];
SLP_TREE_LANES (node) = scalar_stmts.length ();
return node;
}
/* Create an SLP node for SCALAR_STMTS. */
static slp_tree
vect_create_new_slp_node (vec<stmt_vec_info> scalar_stmts, unsigned nops)
{
return vect_create_new_slp_node (new _slp_tree, scalar_stmts, nops);
}
/* Create an SLP node for OPS. */
static slp_tree
vect_create_new_slp_node (slp_tree node, vec<tree> ops)
{
SLP_TREE_SCALAR_OPS (node) = ops;
SLP_TREE_DEF_TYPE (node) = vect_external_def;
SLP_TREE_LANES (node) = ops.length ();
return node;
}
/* Create an SLP node for OPS. */
static slp_tree
vect_create_new_slp_node (vec<tree> ops)
{
return vect_create_new_slp_node (new _slp_tree, ops);
}
/* This structure is used in creation of an SLP tree. Each instance
corresponds to the same operand in a group of scalar stmts in an SLP
node. */
typedef struct _slp_oprnd_info
{
/* Def-stmts for the operands. */
vec<stmt_vec_info> def_stmts;
/* Operands. */
vec<tree> ops;
/* Information about the first statement, its vector def-type, type, the
operand itself in case it's constant, and an indication if it's a pattern
stmt. */
tree first_op_type;
enum vect_def_type first_dt;
bool any_pattern;
} *slp_oprnd_info;
/* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each
operand. */
static vec<slp_oprnd_info>
vect_create_oprnd_info (int nops, int group_size)
{
int i;
slp_oprnd_info oprnd_info;
vec<slp_oprnd_info> oprnds_info;
oprnds_info.create (nops);
for (i = 0; i < nops; i++)
{
oprnd_info = XNEW (struct _slp_oprnd_info);
oprnd_info->def_stmts.create (group_size);
oprnd_info->ops.create (group_size);
oprnd_info->first_dt = vect_uninitialized_def;
oprnd_info->first_op_type = NULL_TREE;
oprnd_info->any_pattern = false;
oprnds_info.quick_push (oprnd_info);
}
return oprnds_info;
}
/* Free operands info. */
static void
vect_free_oprnd_info (vec<slp_oprnd_info> &oprnds_info)
{
int i;
slp_oprnd_info oprnd_info;
FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
{
oprnd_info->def_stmts.release ();
oprnd_info->ops.release ();
XDELETE (oprnd_info);
}
oprnds_info.release ();
}
/* Return the execution frequency of NODE (so that a higher value indicates
a "more important" node when optimizing for speed). */
static sreal
vect_slp_node_weight (slp_tree node)
{
stmt_vec_info stmt_info = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (node));
basic_block bb = gimple_bb (stmt_info->stmt);
return bb->count.to_sreal_scale (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count);
}
/* Return true if STMTS contains a pattern statement. */
static bool
vect_contains_pattern_stmt_p (vec<stmt_vec_info> stmts)
{
stmt_vec_info stmt_info;
unsigned int i;
FOR_EACH_VEC_ELT (stmts, i, stmt_info)
if (is_pattern_stmt_p (stmt_info))
return true;
return false;
}
/* Return true when all lanes in the external or constant NODE have
the same value. */
static bool
vect_slp_tree_uniform_p (slp_tree node)
{
gcc_assert (SLP_TREE_DEF_TYPE (node) == vect_constant_def
|| SLP_TREE_DEF_TYPE (node) == vect_external_def);
/* Pre-exsting vectors. */
if (SLP_TREE_SCALAR_OPS (node).is_empty ())
return false;
unsigned i;
tree op, first = NULL_TREE;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
if (!first)
first = op;
else if (!operand_equal_p (first, op, 0))
return false;
return true;
}
/* Find the place of the data-ref in STMT_INFO in the interleaving chain
that starts from FIRST_STMT_INFO. Return -1 if the data-ref is not a part
of the chain. */
int
vect_get_place_in_interleaving_chain (stmt_vec_info stmt_info,
stmt_vec_info first_stmt_info)
{
stmt_vec_info next_stmt_info = first_stmt_info;
int result = 0;
if (first_stmt_info != DR_GROUP_FIRST_ELEMENT (stmt_info))
return -1;
do
{
if (next_stmt_info == stmt_info)
return result;
next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
if (next_stmt_info)
result += DR_GROUP_GAP (next_stmt_info);
}
while (next_stmt_info);
return -1;
}
/* Check whether it is possible to load COUNT elements of type ELT_TYPE
using the method implemented by duplicate_and_interleave. Return true
if so, returning the number of intermediate vectors in *NVECTORS_OUT
(if nonnull) and the type of each intermediate vector in *VECTOR_TYPE_OUT
(if nonnull). */
bool
can_duplicate_and_interleave_p (vec_info *vinfo, unsigned int count,
tree elt_type, unsigned int *nvectors_out,
tree *vector_type_out,
tree *permutes)
{
tree base_vector_type = get_vectype_for_scalar_type (vinfo, elt_type, count);
if (!base_vector_type || !VECTOR_MODE_P (TYPE_MODE (base_vector_type)))
return false;
machine_mode base_vector_mode = TYPE_MODE (base_vector_type);
poly_int64 elt_bytes = count * GET_MODE_UNIT_SIZE (base_vector_mode);
unsigned int nvectors = 1;
for (;;)
{
scalar_int_mode int_mode;
poly_int64 elt_bits = elt_bytes * BITS_PER_UNIT;
if (int_mode_for_size (elt_bits, 1).exists (&int_mode))
{
/* Get the natural vector type for this SLP group size. */
tree int_type = build_nonstandard_integer_type
(GET_MODE_BITSIZE (int_mode), 1);
tree vector_type
= get_vectype_for_scalar_type (vinfo, int_type, count);
if (vector_type
&& VECTOR_MODE_P (TYPE_MODE (vector_type))
&& known_eq (GET_MODE_SIZE (TYPE_MODE (vector_type)),
GET_MODE_SIZE (base_vector_mode)))
{
/* Try fusing consecutive sequences of COUNT / NVECTORS elements
together into elements of type INT_TYPE and using the result
to build NVECTORS vectors. */
poly_uint64 nelts = GET_MODE_NUNITS (TYPE_MODE (vector_type));
vec_perm_builder sel1 (nelts, 2, 3);
vec_perm_builder sel2 (nelts, 2, 3);
poly_int64 half_nelts = exact_div (nelts, 2);
for (unsigned int i = 0; i < 3; ++i)
{
sel1.quick_push (i);
sel1.quick_push (i + nelts);
sel2.quick_push (half_nelts + i);
sel2.quick_push (half_nelts + i + nelts);
}
vec_perm_indices indices1 (sel1, 2, nelts);
vec_perm_indices indices2 (sel2, 2, nelts);
machine_mode vmode = TYPE_MODE (vector_type);
if (can_vec_perm_const_p (vmode, vmode, indices1)
&& can_vec_perm_const_p (vmode, vmode, indices2))
{
if (nvectors_out)
*nvectors_out = nvectors;
if (vector_type_out)
*vector_type_out = vector_type;
if (permutes)
{
permutes[0] = vect_gen_perm_mask_checked (vector_type,
indices1);
permutes[1] = vect_gen_perm_mask_checked (vector_type,
indices2);
}
return true;
}
}
}
if (!multiple_p (elt_bytes, 2, &elt_bytes))
return false;
nvectors *= 2;
}
}
/* Return true if DTA and DTB match. */
static bool
vect_def_types_match (enum vect_def_type dta, enum vect_def_type dtb)
{
return (dta == dtb
|| ((dta == vect_external_def || dta == vect_constant_def)
&& (dtb == vect_external_def || dtb == vect_constant_def)));
}
static const int cond_expr_maps[3][5] = {
{ 4, -1, -2, 1, 2 },
{ 4, -2, -1, 1, 2 },
{ 4, -1, -2, 2, 1 }
};
static const int arg1_map[] = { 1, 1 };
static const int arg2_map[] = { 1, 2 };
static const int arg1_arg4_map[] = { 2, 1, 4 };
static const int op1_op0_map[] = { 2, 1, 0 };
/* For most SLP statements, there is a one-to-one mapping between
gimple arguments and child nodes. If that is not true for STMT,
return an array that contains:
- the number of child nodes, followed by
- for each child node, the index of the argument associated with that node.
The special index -1 is the first operand of an embedded comparison and
the special index -2 is the second operand of an embedded comparison.
SWAP is as for vect_get_and_check_slp_defs. */
static const int *
vect_get_operand_map (const gimple *stmt, unsigned char swap = 0)
{
if (auto assign = dyn_cast<const gassign *> (stmt))
{
if (gimple_assign_rhs_code (assign) == COND_EXPR
&& COMPARISON_CLASS_P (gimple_assign_rhs1 (assign)))
return cond_expr_maps[swap];
if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison
&& swap)
return op1_op0_map;
}
gcc_assert (!swap);
if (auto call = dyn_cast<const gcall *> (stmt))
{
if (gimple_call_internal_p (call))
switch (gimple_call_internal_fn (call))
{
case IFN_MASK_LOAD:
return arg2_map;
case IFN_GATHER_LOAD:
return arg1_map;
case IFN_MASK_GATHER_LOAD:
return arg1_arg4_map;
default:
break;
}
}
return nullptr;
}
/* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that
they are of a valid type and that they match the defs of the first stmt of
the SLP group (stored in OPRNDS_INFO). This function tries to match stmts
by swapping operands of STMTS[STMT_NUM] when possible. Non-zero SWAP
indicates swap is required for cond_expr stmts. Specifically, SWAP
is 1 if STMT is cond and operands of comparison need to be swapped;
SWAP is 2 if STMT is cond and code of comparison needs to be inverted.
If there was a fatal error return -1; if the error could be corrected by
swapping operands of father node of this one, return 1; if everything is
ok return 0. */
static int
vect_get_and_check_slp_defs (vec_info *vinfo, unsigned char swap,
bool *skip_args,
vec<stmt_vec_info> stmts, unsigned stmt_num,
vec<slp_oprnd_info> *oprnds_info)
{
stmt_vec_info stmt_info = stmts[stmt_num];
tree oprnd;
unsigned int i, number_of_oprnds;
enum vect_def_type dt = vect_uninitialized_def;
slp_oprnd_info oprnd_info;
unsigned int commutative_op = -1U;
bool first = stmt_num == 0;
if (!is_a<gcall *> (stmt_info->stmt)
&& !is_a<gassign *> (stmt_info->stmt)
&& !is_a<gphi *> (stmt_info->stmt))
return -1;
number_of_oprnds = gimple_num_args (stmt_info->stmt);
const int *map = vect_get_operand_map (stmt_info->stmt, swap);
if (map)
number_of_oprnds = *map++;
if (gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt))
{
if (gimple_call_internal_p (stmt))
{
internal_fn ifn = gimple_call_internal_fn (stmt);
commutative_op = first_commutative_argument (ifn);
}
}
else if (gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt))
{
if (commutative_tree_code (gimple_assign_rhs_code (stmt)))
commutative_op = 0;
}
bool swapped = (swap != 0);
bool backedge = false;
enum vect_def_type *dts = XALLOCAVEC (enum vect_def_type, number_of_oprnds);
for (i = 0; i < number_of_oprnds; i++)
{
int opno = map ? map[i] : int (i);
if (opno < 0)
oprnd = TREE_OPERAND (gimple_arg (stmt_info->stmt, 0), -1 - opno);
else
{
oprnd = gimple_arg (stmt_info->stmt, opno);
if (gphi *stmt = dyn_cast <gphi *> (stmt_info->stmt))
backedge = dominated_by_p (CDI_DOMINATORS,
gimple_phi_arg_edge (stmt, opno)->src,
gimple_bb (stmt_info->stmt));
}
if (TREE_CODE (oprnd) == VIEW_CONVERT_EXPR)
oprnd = TREE_OPERAND (oprnd, 0);
oprnd_info = (*oprnds_info)[i];
stmt_vec_info def_stmt_info;
if (!vect_is_simple_use (oprnd, vinfo, &dts[i], &def_stmt_info))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: can't analyze def for %T\n",
oprnd);
return -1;
}
if (skip_args[i])
{
oprnd_info->def_stmts.quick_push (NULL);
oprnd_info->ops.quick_push (NULL_TREE);
oprnd_info->first_dt = vect_uninitialized_def;
continue;
}
oprnd_info->def_stmts.quick_push (def_stmt_info);
oprnd_info->ops.quick_push (oprnd);
if (def_stmt_info
&& is_pattern_stmt_p (def_stmt_info))
{
if (STMT_VINFO_RELATED_STMT (vect_orig_stmt (def_stmt_info))
!= def_stmt_info)
oprnd_info->any_pattern = true;
else
/* If we promote this to external use the original stmt def. */
oprnd_info->ops.last ()
= gimple_get_lhs (vect_orig_stmt (def_stmt_info)->stmt);
}
/* If there's a extern def on a backedge make sure we can
code-generate at the region start.
??? This is another case that could be fixed by adjusting
how we split the function but at the moment we'd have conflicting
goals there. */
if (backedge
&& dts[i] == vect_external_def
&& is_a <bb_vec_info> (vinfo)
&& TREE_CODE (oprnd) == SSA_NAME
&& !SSA_NAME_IS_DEFAULT_DEF (oprnd)
&& !dominated_by_p (CDI_DOMINATORS,
as_a <bb_vec_info> (vinfo)->bbs[0],
gimple_bb (SSA_NAME_DEF_STMT (oprnd))))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: extern def %T only defined "
"on backedge\n", oprnd);
return -1;
}
if (first)
{
tree type = TREE_TYPE (oprnd);
dt = dts[i];
if ((dt == vect_constant_def
|| dt == vect_external_def)
&& !GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
&& (TREE_CODE (type) == BOOLEAN_TYPE
|| !can_duplicate_and_interleave_p (vinfo, stmts.length (),
type)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: invalid type of def "
"for variable-length SLP %T\n", oprnd);
return -1;
}
/* For the swapping logic below force vect_reduction_def
for the reduction op in a SLP reduction group. */
if (!STMT_VINFO_DATA_REF (stmt_info)
&& REDUC_GROUP_FIRST_ELEMENT (stmt_info)
&& (int)i == STMT_VINFO_REDUC_IDX (stmt_info)
&& def_stmt_info)
dts[i] = dt = vect_reduction_def;
/* Check the types of the definition. */
switch (dt)
{
case vect_external_def:
case vect_constant_def:
case vect_internal_def:
case vect_reduction_def:
case vect_induction_def:
case vect_nested_cycle:
break;
default:
/* FORNOW: Not supported. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: illegal type of def %T\n",
oprnd);
return -1;
}
oprnd_info->first_dt = dt;
oprnd_info->first_op_type = type;
}
}
if (first)
return 0;
/* Now match the operand definition types to that of the first stmt. */
for (i = 0; i < number_of_oprnds;)
{
if (skip_args[i])
{
++i;
continue;
}
oprnd_info = (*oprnds_info)[i];
dt = dts[i];
stmt_vec_info def_stmt_info = oprnd_info->def_stmts[stmt_num];
oprnd = oprnd_info->ops[stmt_num];
tree type = TREE_TYPE (oprnd);
if (!types_compatible_p (oprnd_info->first_op_type, type))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different operand types\n");
return 1;
}
/* Not first stmt of the group, check that the def-stmt/s match
the def-stmt/s of the first stmt. Allow different definition
types for reduction chains: the first stmt must be a
vect_reduction_def (a phi node), and the rest
end in the reduction chain. */
if ((!vect_def_types_match (oprnd_info->first_dt, dt)
&& !(oprnd_info->first_dt == vect_reduction_def
&& !STMT_VINFO_DATA_REF (stmt_info)
&& REDUC_GROUP_FIRST_ELEMENT (stmt_info)
&& def_stmt_info
&& !STMT_VINFO_DATA_REF (def_stmt_info)
&& (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
== REDUC_GROUP_FIRST_ELEMENT (stmt_info))))
|| (!STMT_VINFO_DATA_REF (stmt_info)
&& REDUC_GROUP_FIRST_ELEMENT (stmt_info)
&& ((!def_stmt_info
|| STMT_VINFO_DATA_REF (def_stmt_info)
|| (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
!= REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
!= (oprnd_info->first_dt != vect_reduction_def))))
{
/* Try swapping operands if we got a mismatch. For BB
vectorization only in case it will clearly improve things. */
if (i == commutative_op && !swapped
&& (!is_a <bb_vec_info> (vinfo)
|| (!vect_def_types_match ((*oprnds_info)[i+1]->first_dt,
dts[i+1])
&& (vect_def_types_match (oprnd_info->first_dt, dts[i+1])
|| vect_def_types_match
((*oprnds_info)[i+1]->first_dt, dts[i])))))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"trying swapped operands\n");
std::swap (dts[i], dts[i+1]);
std::swap ((*oprnds_info)[i]->def_stmts[stmt_num],
(*oprnds_info)[i+1]->def_stmts[stmt_num]);
std::swap ((*oprnds_info)[i]->ops[stmt_num],
(*oprnds_info)[i+1]->ops[stmt_num]);
swapped = true;
continue;
}
if (is_a <bb_vec_info> (vinfo)
&& !oprnd_info->any_pattern)
{
/* Now for commutative ops we should see whether we can
make the other operand matching. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"treating operand as external\n");
oprnd_info->first_dt = dt = vect_external_def;
}
else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different types\n");
return 1;
}
}
/* Make sure to demote the overall operand to external. */
if (dt == vect_external_def)
oprnd_info->first_dt = vect_external_def;
/* For a SLP reduction chain we want to duplicate the reduction to
each of the chain members. That gets us a sane SLP graph (still
the stmts are not 100% correct wrt the initial values). */
else if ((dt == vect_internal_def
|| dt == vect_reduction_def)
&& oprnd_info->first_dt == vect_reduction_def
&& !STMT_VINFO_DATA_REF (stmt_info)
&& REDUC_GROUP_FIRST_ELEMENT (stmt_info)
&& !STMT_VINFO_DATA_REF (def_stmt_info)
&& (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
== REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
{
oprnd_info->def_stmts[stmt_num] = oprnd_info->def_stmts[0];
oprnd_info->ops[stmt_num] = oprnd_info->ops[0];
}
++i;
}
/* Swap operands. */
if (swapped)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"swapped operands to match def types in %G",
stmt_info->stmt);
}
return 0;
}
/* Return true if call statements CALL1 and CALL2 are similar enough
to be combined into the same SLP group. */
bool
compatible_calls_p (gcall *call1, gcall *call2)
{
unsigned int nargs = gimple_call_num_args (call1);
if (nargs != gimple_call_num_args (call2))
return false;
if (gimple_call_combined_fn (call1) != gimple_call_combined_fn (call2))
return false;
if (gimple_call_internal_p (call1))
{
if (!types_compatible_p (TREE_TYPE (gimple_call_lhs (call1)),
TREE_TYPE (gimple_call_lhs (call2))))
return false;
for (unsigned int i = 0; i < nargs; ++i)
if (!types_compatible_p (TREE_TYPE (gimple_call_arg (call1, i)),
TREE_TYPE (gimple_call_arg (call2, i))))
return false;
}
else
{
if (!operand_equal_p (gimple_call_fn (call1),
gimple_call_fn (call2), 0))
return false;
if (gimple_call_fntype (call1) != gimple_call_fntype (call2))
return false;
}
/* Check that any unvectorized arguments are equal. */
if (const int *map = vect_get_operand_map (call1))
{
unsigned int nkept = *map++;
unsigned int mapi = 0;
for (unsigned int i = 0; i < nargs; ++i)
if (mapi < nkept && map[mapi] == int (i))
mapi += 1;
else if (!operand_equal_p (gimple_call_arg (call1, i),
gimple_call_arg (call2, i)))
return false;
}
return true;
}
/* A subroutine of vect_build_slp_tree for checking VECTYPE, which is the
caller's attempt to find the vector type in STMT_INFO with the narrowest
element type. Return true if VECTYPE is nonnull and if it is valid
for STMT_INFO. When returning true, update MAX_NUNITS to reflect the
number of units in VECTYPE. GROUP_SIZE and MAX_NUNITS are as for
vect_build_slp_tree. */
static bool
vect_record_max_nunits (vec_info *vinfo, stmt_vec_info stmt_info,
unsigned int group_size,
tree vectype, poly_uint64 *max_nunits)
{
if (!vectype)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: unsupported data-type in %G\n",
stmt_info->stmt);
/* Fatal mismatch. */
return false;
}
/* If populating the vector type requires unrolling then fail
before adjusting *max_nunits for basic-block vectorization. */
if (is_a <bb_vec_info> (vinfo)
&& !multiple_p (group_size, TYPE_VECTOR_SUBPARTS (vectype)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: unrolling required "
"in basic block SLP\n");
/* Fatal mismatch. */
return false;
}
/* In case of multiple types we need to detect the smallest type. */
vect_update_max_nunits (max_nunits, vectype);
return true;
}
/* Verify if the scalar stmts STMTS are isomorphic, require data
permutation or are of unsupported types of operation. Return
true if they are, otherwise return false and indicate in *MATCHES
which stmts are not isomorphic to the first one. If MATCHES[0]
is false then this indicates the comparison could not be
carried out or the stmts will never be vectorized by SLP.
Note COND_EXPR is possibly isomorphic to another one after swapping its
operands. Set SWAP[i] to 1 if stmt I is COND_EXPR and isomorphic to
the first stmt by swapping the two operands of comparison; set SWAP[i]
to 2 if stmt I is isormorphic to the first stmt by inverting the code
of comparison. Take A1 >= B1 ? X1 : Y1 as an exmple, it can be swapped
to (B1 <= A1 ? X1 : Y1); or be inverted to (A1 < B1) ? Y1 : X1. */
static bool
vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
vec<stmt_vec_info> stmts, unsigned int group_size,
poly_uint64 *max_nunits, bool *matches,
bool *two_operators, tree *node_vectype)
{
unsigned int i;
stmt_vec_info first_stmt_info = stmts[0];
code_helper first_stmt_code = ERROR_MARK;
code_helper alt_stmt_code = ERROR_MARK;
code_helper rhs_code = ERROR_MARK;
code_helper first_cond_code = ERROR_MARK;
tree lhs;
bool need_same_oprnds = false;
tree vectype = NULL_TREE, first_op1 = NULL_TREE;
stmt_vec_info first_load = NULL, prev_first_load = NULL;
bool first_stmt_load_p = false, load_p = false;
bool first_stmt_phi_p = false, phi_p = false;
bool maybe_soft_fail = false;
tree soft_fail_nunits_vectype = NULL_TREE;
/* For every stmt in NODE find its def stmt/s. */
stmt_vec_info stmt_info;
FOR_EACH_VEC_ELT (stmts, i, stmt_info)
{
gimple *stmt = stmt_info->stmt;
swap[i] = 0;
matches[i] = false;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "Build SLP for %G", stmt);
/* Fail to vectorize statements marked as unvectorizable, throw
or are volatile. */
if (!STMT_VINFO_VECTORIZABLE (stmt_info)
|| stmt_can_throw_internal (cfun, stmt)
|| gimple_has_volatile_ops (stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: unvectorizable statement %G",
stmt);
/* ??? For BB vectorization we want to commutate operands in a way
to shuffle all unvectorizable defs into one operand and have
the other still vectorized. The following doesn't reliably
work for this though but it's the easiest we can do here. */
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
lhs = gimple_get_lhs (stmt);
if (lhs == NULL_TREE)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: not GIMPLE_ASSIGN nor "
"GIMPLE_CALL %G", stmt);
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
tree nunits_vectype;
if (!vect_get_vector_types_for_stmt (vinfo, stmt_info, &vectype,
&nunits_vectype, group_size))
{
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
/* Record nunits required but continue analysis, producing matches[]
as if nunits was not an issue. This allows splitting of groups
to happen. */
if (nunits_vectype
&& !vect_record_max_nunits (vinfo, stmt_info, group_size,
nunits_vectype, max_nunits))
{
gcc_assert (is_a <bb_vec_info> (vinfo));
maybe_soft_fail = true;
soft_fail_nunits_vectype = nunits_vectype;
}
gcc_assert (vectype);
gcall *call_stmt = dyn_cast <gcall *> (stmt);
if (call_stmt)
{
combined_fn cfn = gimple_call_combined_fn (call_stmt);
if (cfn != CFN_LAST)
rhs_code = cfn;
else
rhs_code = CALL_EXPR;
if (cfn == CFN_MASK_LOAD
|| cfn == CFN_GATHER_LOAD
|| cfn == CFN_MASK_GATHER_LOAD)
load_p = true;
else if ((internal_fn_p (cfn)
&& !vectorizable_internal_fn_p (as_internal_fn (cfn)))
|| gimple_call_tail_p (call_stmt)
|| gimple_call_noreturn_p (call_stmt)
|| gimple_call_chain (call_stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: unsupported call type %G",
(gimple *) call_stmt);
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
}
else if (gimple_code (stmt) == GIMPLE_PHI)
{
rhs_code = ERROR_MARK;
phi_p = true;
}
else
{
rhs_code = gimple_assign_rhs_code (stmt);
load_p = gimple_vuse (stmt);
}
/* Check the operation. */
if (i == 0)
{
*node_vectype = vectype;
first_stmt_code = rhs_code;
first_stmt_load_p = load_p;
first_stmt_phi_p = phi_p;
/* Shift arguments should be equal in all the packed stmts for a
vector shift with scalar shift operand. */
if (rhs_code == LSHIFT_EXPR || rhs_code == RSHIFT_EXPR
|| rhs_code == LROTATE_EXPR
|| rhs_code == RROTATE_EXPR)
{
/* First see if we have a vector/vector shift. */
if (!directly_supported_p (rhs_code, vectype, optab_vector))
{
/* No vector/vector shift, try for a vector/scalar shift. */
if (!directly_supported_p (rhs_code, vectype, optab_scalar))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: "
"op not supported by target.\n");
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
need_same_oprnds = true;
first_op1 = gimple_assign_rhs2 (stmt);
}
}
else if (rhs_code == WIDEN_LSHIFT_EXPR)
{
need_same_oprnds = true;
first_op1 = gimple_assign_rhs2 (stmt);
}
else if (!load_p
&& rhs_code == BIT_FIELD_REF)
{
tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
if (!is_a <bb_vec_info> (vinfo)
|| TREE_CODE (vec) != SSA_NAME
/* When the element types are not compatible we pun the
source to the target vectype which requires equal size. */
|| ((!VECTOR_TYPE_P (TREE_TYPE (vec))
|| !types_compatible_p (TREE_TYPE (vectype),
TREE_TYPE (TREE_TYPE (vec))))
&& !operand_equal_p (TYPE_SIZE (vectype),
TYPE_SIZE (TREE_TYPE (vec)))))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: "
"BIT_FIELD_REF not supported\n");
/* Fatal mismatch. */
matches[0] = false;
return false;
}
}
else if (rhs_code == CFN_DIV_POW2)
{
need_same_oprnds = true;
first_op1 = gimple_call_arg (call_stmt, 1);
}
}
else
{
if (first_stmt_code != rhs_code
&& alt_stmt_code == ERROR_MARK)
alt_stmt_code = rhs_code;
if ((first_stmt_code != rhs_code
&& (first_stmt_code != IMAGPART_EXPR
|| rhs_code != REALPART_EXPR)
&& (first_stmt_code != REALPART_EXPR
|| rhs_code != IMAGPART_EXPR)
/* Handle mismatches in plus/minus by computing both
and merging the results. */
&& !((first_stmt_code == PLUS_EXPR
|| first_stmt_code == MINUS_EXPR)
&& (alt_stmt_code == PLUS_EXPR
|| alt_stmt_code == MINUS_EXPR)
&& rhs_code == alt_stmt_code)
&& !(first_stmt_code.is_tree_code ()
&& rhs_code.is_tree_code ()
&& (TREE_CODE_CLASS (tree_code (first_stmt_code))
== tcc_comparison)
&& (swap_tree_comparison (tree_code (first_stmt_code))
== tree_code (rhs_code)))
&& !(STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& (first_stmt_code == ARRAY_REF
|| first_stmt_code == BIT_FIELD_REF
|| first_stmt_code == INDIRECT_REF
|| first_stmt_code == COMPONENT_REF
|| first_stmt_code == MEM_REF)
&& (rhs_code == ARRAY_REF
|| rhs_code == BIT_FIELD_REF
|| rhs_code == INDIRECT_REF
|| rhs_code == COMPONENT_REF
|| rhs_code == MEM_REF)))
|| first_stmt_load_p != load_p
|| first_stmt_phi_p != phi_p)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different operation "
"in stmt %G", stmt);
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"original stmt %G", first_stmt_info->stmt);
}
/* Mismatch. */
continue;
}
if (!load_p
&& first_stmt_code == BIT_FIELD_REF
&& (TREE_OPERAND (gimple_assign_rhs1 (first_stmt_info->stmt), 0)
!= TREE_OPERAND (gimple_assign_rhs1 (stmt_info->stmt), 0)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different BIT_FIELD_REF "
"arguments in %G", stmt);
/* Mismatch. */
continue;
}
if (call_stmt && first_stmt_code != CFN_MASK_LOAD)
{
if (!compatible_calls_p (as_a <gcall *> (stmts[0]->stmt),
call_stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different calls in %G",
stmt);
/* Mismatch. */
continue;
}
}
if ((phi_p || gimple_could_trap_p (stmt_info->stmt))
&& (gimple_bb (first_stmt_info->stmt)
!= gimple_bb (stmt_info->stmt)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different BB for PHI "
"or possibly trapping operation in %G", stmt);
/* Mismatch. */
continue;
}
if (need_same_oprnds)
{
tree other_op1 = gimple_arg (stmt, 1);
if (!operand_equal_p (first_op1, other_op1, 0))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different shift "
"arguments in %G", stmt);
/* Mismatch. */
continue;
}
}
if (!types_compatible_p (vectype, *node_vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different vector type "
"in %G", stmt);
/* Mismatch. */
continue;
}
}
/* Grouped store or load. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
{
if (REFERENCE_CLASS_P (lhs))
{
/* Store. */
;
}
else
{
/* Load. */
first_load = DR_GROUP_FIRST_ELEMENT (stmt_info);
if (prev_first_load)
{
/* Check that there are no loads from different interleaving
chains in the same node. */
if (prev_first_load != first_load)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION,
vect_location,
"Build SLP failed: different "
"interleaving chains in one node %G",
stmt);
/* Mismatch. */
continue;
}
}
else
prev_first_load = first_load;
}
} /* Grouped access. */
else
{
if (load_p
&& rhs_code != CFN_GATHER_LOAD
&& rhs_code != CFN_MASK_GATHER_LOAD)
{
/* Not grouped load. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: not grouped load %G", stmt);
/* FORNOW: Not grouped loads are not supported. */
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
/* Not memory operation. */
if (!phi_p
&& rhs_code.is_tree_code ()
&& TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_binary
&& TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_unary
&& TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_expression
&& TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_comparison
&& rhs_code != VIEW_CONVERT_EXPR
&& rhs_code != CALL_EXPR
&& rhs_code != BIT_FIELD_REF)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: operation unsupported %G",
stmt);
if (is_a <bb_vec_info> (vinfo) && i != 0)
continue;
/* Fatal mismatch. */
matches[0] = false;
return false;
}
if (rhs_code == COND_EXPR)
{
tree cond_expr = gimple_assign_rhs1 (stmt);
enum tree_code cond_code = TREE_CODE (cond_expr);
enum tree_code swap_code = ERROR_MARK;
enum tree_code invert_code = ERROR_MARK;
if (i == 0)
first_cond_code = TREE_CODE (cond_expr);
else if (TREE_CODE_CLASS (cond_code) == tcc_comparison)
{
bool honor_nans = HONOR_NANS (TREE_OPERAND (cond_expr, 0));
swap_code = swap_tree_comparison (cond_code);
invert_code = invert_tree_comparison (cond_code, honor_nans);
}
if (first_cond_code == cond_code)
;
/* Isomorphic can be achieved by swapping. */
else if (first_cond_code == swap_code)
swap[i] = 1;
/* Isomorphic can be achieved by inverting. */
else if (first_cond_code == invert_code)
swap[i] = 2;
else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: different"
" operation %G", stmt);
/* Mismatch. */
continue;
}
}
if (rhs_code.is_tree_code ()
&& TREE_CODE_CLASS ((tree_code)rhs_code) == tcc_comparison
&& (swap_tree_comparison ((tree_code)first_stmt_code)
== (tree_code)rhs_code))
swap[i] = 1;
}
matches[i] = true;
}
for (i = 0; i < group_size; ++i)
if (!matches[i])
return false;
/* If we allowed a two-operation SLP node verify the target can cope
with the permute we are going to use. */
if (alt_stmt_code != ERROR_MARK
&& (!alt_stmt_code.is_tree_code ()
|| (TREE_CODE_CLASS (tree_code (alt_stmt_code)) != tcc_reference
&& TREE_CODE_CLASS (tree_code (alt_stmt_code)) != tcc_comparison)))
{
*two_operators = true;
}
if (maybe_soft_fail)
{
unsigned HOST_WIDE_INT const_nunits;
if (!TYPE_VECTOR_SUBPARTS
(soft_fail_nunits_vectype).is_constant (&const_nunits)
|| const_nunits > group_size)
matches[0] = false;
else
{
/* With constant vector elements simulate a mismatch at the
point we need to split. */
unsigned tail = group_size & (const_nunits - 1);
memset (&matches[group_size - tail], 0, sizeof (bool) * tail);
}
return false;
}
return true;
}
/* Traits for the hash_set to record failed SLP builds for a stmt set.
Note we never remove apart from at destruction time so we do not
need a special value for deleted that differs from empty. */
struct bst_traits
{
typedef vec <stmt_vec_info> value_type;
typedef vec <stmt_vec_info> compare_type;
static inline hashval_t hash (value_type);
static inline bool equal (value_type existing, value_type candidate);
static inline bool is_empty (value_type x) { return !x.exists (); }
static inline bool is_deleted (value_type x) { return !x.exists (); }
static const bool empty_zero_p = true;
static inline void mark_empty (value_type &x) { x.release (); }
static inline void mark_deleted (value_type &x) { x.release (); }
static inline void remove (value_type &x) { x.release (); }
};
inline hashval_t
bst_traits::hash (value_type x)
{
inchash::hash h;
for (unsigned i = 0; i < x.length (); ++i)
h.add_int (gimple_uid (x[i]->stmt));
return h.end ();
}
inline bool
bst_traits::equal (value_type existing, value_type candidate)
{
if (existing.length () != candidate.length ())
return false;
for (unsigned i = 0; i < existing.length (); ++i)
if (existing[i] != candidate[i])
return false;
return true;
}
/* ??? This was std::pair<std::pair<tree_code, vect_def_type>, tree>
but then vec::insert does memmove and that's not compatible with
std::pair. */
struct chain_op_t
{
chain_op_t (tree_code code_, vect_def_type dt_, tree op_)
: code (code_), dt (dt_), op (op_) {}
tree_code code;
vect_def_type dt;
tree op;
};
/* Comparator for sorting associatable chains. */
static int
dt_sort_cmp (const void *op1_, const void *op2_, void *)
{
auto *op1 = (const chain_op_t *) op1_;
auto *op2 = (const chain_op_t *) op2_;
if (op1->dt != op2->dt)
return (int)op1->dt - (int)op2->dt;
return (int)op1->code - (int)op2->code;
}
/* Linearize the associatable expression chain at START with the
associatable operation CODE (where PLUS_EXPR also allows MINUS_EXPR),
filling CHAIN with the result and using WORKLIST as intermediate storage.
CODE_STMT and ALT_CODE_STMT are filled with the first stmt using CODE
or MINUS_EXPR. *CHAIN_STMTS if not NULL is filled with all computation
stmts, starting with START. */
static void
vect_slp_linearize_chain (vec_info *vinfo,
vec<std::pair<tree_code, gimple *> > &worklist,
vec<chain_op_t> &chain,
enum tree_code code, gimple *start,
gimple *&code_stmt, gimple *&alt_code_stmt,
vec<gimple *> *chain_stmts)
{
/* For each lane linearize the addition/subtraction (or other
uniform associatable operation) expression tree. */
worklist.safe_push (std::make_pair (code, start));
while (!worklist.is_empty ())
{
auto entry = worklist.pop ();
gassign *stmt = as_a <gassign *> (entry.second);
enum tree_code in_code = entry.first;
enum tree_code this_code = gimple_assign_rhs_code (stmt);
/* Pick some stmts suitable for SLP_TREE_REPRESENTATIVE. */
if (!code_stmt
&& gimple_assign_rhs_code (stmt) == code)
code_stmt = stmt;
else if (!alt_code_stmt
&& gimple_assign_rhs_code (stmt) == MINUS_EXPR)
alt_code_stmt = stmt;
if (chain_stmts)
chain_stmts->safe_push (stmt);
for (unsigned opnum = 1; opnum <= 2; ++opnum)
{
tree op = gimple_op (stmt, opnum);
vect_def_type dt;
stmt_vec_info def_stmt_info;
bool res = vect_is_simple_use (op, vinfo, &dt, &def_stmt_info);
gcc_assert (res);
if (dt == vect_internal_def
&& is_pattern_stmt_p (def_stmt_info))
op = gimple_get_lhs (def_stmt_info->stmt);
gimple *use_stmt;
use_operand_p use_p;
if (dt == vect_internal_def
&& single_imm_use (op, &use_p, &use_stmt)
&& is_gimple_assign (def_stmt_info->stmt)
&& (gimple_assign_rhs_code (def_stmt_info->stmt) == code
|| (code == PLUS_EXPR
&& (gimple_assign_rhs_code (def_stmt_info->stmt)
== MINUS_EXPR))))
{
tree_code op_def_code = this_code;
if (op_def_code == MINUS_EXPR && opnum == 1)
op_def_code = PLUS_EXPR;
if (in_code == MINUS_EXPR)
op_def_code = op_def_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
worklist.safe_push (std::make_pair (op_def_code,
def_stmt_info->stmt));
}
else
{
tree_code op_def_code = this_code;
if (op_def_code == MINUS_EXPR && opnum == 1)
op_def_code = PLUS_EXPR;
if (in_code == MINUS_EXPR)
op_def_code = op_def_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
chain.safe_push (chain_op_t (op_def_code, dt, op));
}
}
}
}
typedef hash_map <vec <stmt_vec_info>, slp_tree,
simple_hashmap_traits <bst_traits, slp_tree> >
scalar_stmts_to_slp_tree_map_t;
static slp_tree
vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
vec<stmt_vec_info> stmts, unsigned int group_size,
poly_uint64 *max_nunits,
bool *matches, unsigned *limit, unsigned *tree_size,
scalar_stmts_to_slp_tree_map_t *bst_map);
static slp_tree
vect_build_slp_tree (vec_info *vinfo,
vec<stmt_vec_info> stmts, unsigned int group_size,
poly_uint64 *max_nunits,
bool *matches, unsigned *limit, unsigned *tree_size,
scalar_stmts_to_slp_tree_map_t *bst_map)
{
if (slp_tree *leader = bst_map->get (stmts))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "re-using %sSLP tree %p\n",
!(*leader)->failed ? "" : "failed ",
(void *) *leader);
if (!(*leader)->failed)
{
SLP_TREE_REF_COUNT (*leader)++;
vect_update_max_nunits (max_nunits, (*leader)->max_nunits);
stmts.release ();
return *leader;
}
memcpy (matches, (*leader)->failed, sizeof (bool) * group_size);
return NULL;
}
/* Seed the bst_map with a stub node to be filled by vect_build_slp_tree_2
so we can pick up backedge destinations during discovery. */
slp_tree res = new _slp_tree;
SLP_TREE_DEF_TYPE (res) = vect_internal_def;
SLP_TREE_SCALAR_STMTS (res) = stmts;
bst_map->put (stmts.copy (), res);
if (*limit == 0)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"SLP discovery limit exceeded\n");
/* Mark the node invalid so we can detect those when still in use
as backedge destinations. */
SLP_TREE_SCALAR_STMTS (res) = vNULL;
SLP_TREE_DEF_TYPE (res) = vect_uninitialized_def;
res->failed = XNEWVEC (bool, group_size);
memset (res->failed, 0, sizeof (bool) * group_size);
memset (matches, 0, sizeof (bool) * group_size);
return NULL;
}
--*limit;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"starting SLP discovery for node %p\n", (void *) res);
poly_uint64 this_max_nunits = 1;
slp_tree res_ = vect_build_slp_tree_2 (vinfo, res, stmts, group_size,
&this_max_nunits,
matches, limit, tree_size, bst_map);
if (!res_)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"SLP discovery for node %p failed\n", (void *) res);
/* Mark the node invalid so we can detect those when still in use
as backedge destinations. */
SLP_TREE_SCALAR_STMTS (res) = vNULL;
SLP_TREE_DEF_TYPE (res) = vect_uninitialized_def;
res->failed = XNEWVEC (bool, group_size);
if (flag_checking)
{
unsigned i;
for (i = 0; i < group_size; ++i)
if (!matches[i])
break;
gcc_assert (i < group_size);
}
memcpy (res->failed, matches, sizeof (bool) * group_size);
}
else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"SLP discovery for node %p succeeded\n",
(void *) res);
gcc_assert (res_ == res);
res->max_nunits = this_max_nunits;
vect_update_max_nunits (max_nunits, this_max_nunits);
/* Keep a reference for the bst_map use. */
SLP_TREE_REF_COUNT (res)++;
}
return res_;
}
/* Helper for building an associated SLP node chain. */
static void
vect_slp_build_two_operator_nodes (slp_tree perm, tree vectype,
slp_tree op0, slp_tree op1,
stmt_vec_info oper1, stmt_vec_info oper2,
vec<std::pair<unsigned, unsigned> > lperm)
{
unsigned group_size = SLP_TREE_LANES (op1);
slp_tree child1 = new _slp_tree;
SLP_TREE_DEF_TYPE (child1) = vect_internal_def;
SLP_TREE_VECTYPE (child1) = vectype;
SLP_TREE_LANES (child1) = group_size;
SLP_TREE_CHILDREN (child1).create (2);
SLP_TREE_CHILDREN (child1).quick_push (op0);
SLP_TREE_CHILDREN (child1).quick_push (op1);
SLP_TREE_REPRESENTATIVE (child1) = oper1;
slp_tree child2 = new _slp_tree;
SLP_TREE_DEF_TYPE (child2) = vect_internal_def;
SLP_TREE_VECTYPE (child2) = vectype;
SLP_TREE_LANES (child2) = group_size;
SLP_TREE_CHILDREN (child2).create (2);
SLP_TREE_CHILDREN (child2).quick_push (op0);
SLP_TREE_REF_COUNT (op0)++;
SLP_TREE_CHILDREN (child2).quick_push (op1);
SLP_TREE_REF_COUNT (op1)++;
SLP_TREE_REPRESENTATIVE (child2) = oper2;
SLP_TREE_DEF_TYPE (perm) = vect_internal_def;
SLP_TREE_CODE (perm) = VEC_PERM_EXPR;
SLP_TREE_VECTYPE (perm) = vectype;
SLP_TREE_LANES (perm) = group_size;
/* ??? We should set this NULL but that's not expected. */
SLP_TREE_REPRESENTATIVE (perm) = oper1;
SLP_TREE_LANE_PERMUTATION (perm) = lperm;
SLP_TREE_CHILDREN (perm).quick_push (child1);
SLP_TREE_CHILDREN (perm).quick_push (child2);
}
/* Recursively build an SLP tree starting from NODE.
Fail (and return a value not equal to zero) if def-stmts are not
isomorphic, require data permutation or are of unsupported types of
operation. Otherwise, return 0.
The value returned is the depth in the SLP tree where a mismatch
was found. */
static slp_tree
vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
vec<stmt_vec_info> stmts, unsigned int group_size,
poly_uint64 *max_nunits,
bool *matches, unsigned *limit, unsigned *tree_size,
scalar_stmts_to_slp_tree_map_t *bst_map)
{
unsigned nops, i, this_tree_size = 0;
poly_uint64 this_max_nunits = *max_nunits;
matches[0] = false;
stmt_vec_info stmt_info = stmts[0];
if (!is_a<gcall *> (stmt_info->stmt)
&& !is_a<gassign *> (stmt_info->stmt)
&& !is_a<gphi *> (stmt_info->stmt))
return NULL;
nops = gimple_num_args (stmt_info->stmt);
if (const int *map = vect_get_operand_map (stmt_info->stmt))
nops = map[0];
/* If the SLP node is a PHI (induction or reduction), terminate
the recursion. */
bool *skip_args = XALLOCAVEC (bool, nops);
memset (skip_args, 0, sizeof (bool) * nops);
if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
if (gphi *stmt = dyn_cast <gphi *> (stmt_info->stmt))
{
tree scalar_type = TREE_TYPE (PHI_RESULT (stmt));
tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
group_size);
if (!vect_record_max_nunits (vinfo, stmt_info, group_size, vectype,
max_nunits))
return NULL;
vect_def_type def_type = STMT_VINFO_DEF_TYPE (stmt_info);
if (def_type == vect_induction_def)
{
/* Induction PHIs are not cycles but walk the initial
value. Only for inner loops through, for outer loops
we need to pick up the value from the actual PHIs
to more easily support peeling and epilogue vectorization. */
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
if (!nested_in_vect_loop_p (loop, stmt_info))
skip_args[loop_preheader_edge (loop)->dest_idx] = true;
else
loop = loop->inner;
skip_args[loop_latch_edge (loop)->dest_idx] = true;
}
else if (def_type == vect_reduction_def
|| def_type == vect_double_reduction_def
|| def_type == vect_nested_cycle)
{
/* Else def types have to match. */
stmt_vec_info other_info;
bool all_same = true;
FOR_EACH_VEC_ELT (stmts, i, other_info)
{
if (STMT_VINFO_DEF_TYPE (other_info) != def_type)
return NULL;
if (other_info != stmt_info)
all_same = false;
}
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
/* Reduction initial values are not explicitely represented. */
if (!nested_in_vect_loop_p (loop, stmt_info))
skip_args[loop_preheader_edge (loop)->dest_idx] = true;
/* Reduction chain backedge defs are filled manually.
??? Need a better way to identify a SLP reduction chain PHI.
Or a better overall way to SLP match those. */
if (all_same && def_type == vect_reduction_def)
skip_args[loop_latch_edge (loop)->dest_idx] = true;
}
else if (def_type != vect_internal_def)
return NULL;
}
bool two_operators = false;
unsigned char *swap = XALLOCAVEC (unsigned char, group_size);
tree vectype = NULL_TREE;
if (!vect_build_slp_tree_1 (vinfo, swap, stmts, group_size,
&this_max_nunits, matches, &two_operators,
&vectype))
return NULL;
/* If the SLP node is a load, terminate the recursion unless masked. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
{
if (gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt))
gcc_assert (gimple_call_internal_p (stmt, IFN_MASK_LOAD)
|| gimple_call_internal_p (stmt, IFN_GATHER_LOAD)
|| gimple_call_internal_p (stmt, IFN_MASK_GATHER_LOAD));
else
{
*max_nunits = this_max_nunits;
(*tree_size)++;
node = vect_create_new_slp_node (node, stmts, 0);
SLP_TREE_VECTYPE (node) = vectype;
/* And compute the load permutation. Whether it is actually
a permutation depends on the unrolling factor which is
decided later. */
vec<unsigned> load_permutation;
int j;
stmt_vec_info load_info;
load_permutation.create (group_size);
stmt_vec_info first_stmt_info
= DR_GROUP_FIRST_ELEMENT (SLP_TREE_SCALAR_STMTS (node)[0]);
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
{
int load_place = vect_get_place_in_interleaving_chain
(load_info, first_stmt_info);
gcc_assert (load_place != -1);
load_permutation.safe_push (load_place);
}
SLP_TREE_LOAD_PERMUTATION (node) = load_permutation;
return node;
}
}
else if (gimple_assign_single_p (stmt_info->stmt)
&& !gimple_vuse (stmt_info->stmt)
&& gimple_assign_rhs_code (stmt_info->stmt) == BIT_FIELD_REF)
{
/* vect_build_slp_tree_2 determined all BIT_FIELD_REFs reference
the same SSA name vector of a compatible type to vectype. */
vec<std::pair<unsigned, unsigned> > lperm = vNULL;
tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt_info->stmt), 0);
stmt_vec_info estmt_info;
FOR_EACH_VEC_ELT (stmts, i, estmt_info)
{
gassign *estmt = as_a <gassign *> (estmt_info->stmt);
tree bfref = gimple_assign_rhs1 (estmt);
HOST_WIDE_INT lane;
if (!known_eq (bit_field_size (bfref),
tree_to_poly_uint64 (TYPE_SIZE (TREE_TYPE (vectype))))
|| !constant_multiple_p (bit_field_offset (bfref),
bit_field_size (bfref), &lane))
{
lperm.release ();
matches[0] = false;
return NULL;
}
lperm.safe_push (std::make_pair (0, (unsigned)lane));
}
slp_tree vnode = vect_create_new_slp_node (vNULL);
if (operand_equal_p (TYPE_SIZE (vectype), TYPE_SIZE (TREE_TYPE (vec))))
/* ??? We record vectype here but we hide eventually necessary
punning and instead rely on code generation to materialize
VIEW_CONVERT_EXPRs as necessary. We instead should make
this explicit somehow. */
SLP_TREE_VECTYPE (vnode) = vectype;
else
{
/* For different size but compatible elements we can still
use VEC_PERM_EXPR without punning. */
gcc_assert (VECTOR_TYPE_P (TREE_TYPE (vec))
&& types_compatible_p (TREE_TYPE (vectype),
TREE_TYPE (TREE_TYPE (vec))));
SLP_TREE_VECTYPE (vnode) = TREE_TYPE (vec);
}
auto nunits = TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (vnode));
unsigned HOST_WIDE_INT const_nunits;
if (nunits.is_constant (&const_nunits))
SLP_TREE_LANES (vnode) = const_nunits;
SLP_TREE_VEC_DEFS (vnode).safe_push (vec);
/* We are always building a permutation node even if it is an identity
permute to shield the rest of the vectorizer from the odd node
representing an actual vector without any scalar ops.
??? We could hide it completely with making the permute node
external? */
node = vect_create_new_slp_node (node, stmts, 1);
SLP_TREE_CODE (node) = VEC_PERM_EXPR;
SLP_TREE_LANE_PERMUTATION (node) = lperm;
SLP_TREE_VECTYPE (node) = vectype;
SLP_TREE_CHILDREN (node).quick_push (vnode);
return node;
}
/* When discovery reaches an associatable operation see whether we can
improve that to match up lanes in a way superior to the operand
swapping code which at most looks at two defs.
??? For BB vectorization we cannot do the brute-force search
for matching as we can succeed by means of builds from scalars
and have no good way to "cost" one build against another. */
else if (is_a <loop_vec_info> (vinfo)
/* ??? We don't handle !vect_internal_def defs below. */
&& STMT_VINFO_DEF_TYPE (stmt_info) == vect_internal_def
&& is_gimple_assign (stmt_info->stmt)
&& (associative_tree_code (gimple_assign_rhs_code (stmt_info->stmt))
|| gimple_assign_rhs_code (stmt_info->stmt) == MINUS_EXPR)
&& ((FLOAT_TYPE_P (vectype) && flag_associative_math)
|| (INTEGRAL_TYPE_P (TREE_TYPE (vectype))
&& TYPE_OVERFLOW_WRAPS (TREE_TYPE (vectype)))))
{
/* See if we have a chain of (mixed) adds or subtracts or other
associatable ops. */
enum tree_code code = gimple_assign_rhs_code (stmt_info->stmt);
if (code == MINUS_EXPR)
code = PLUS_EXPR;
stmt_vec_info other_op_stmt_info = NULL;
stmt_vec_info op_stmt_info = NULL;
unsigned chain_len = 0;
auto_vec<chain_op_t> chain;
auto_vec<std::pair<tree_code, gimple *> > worklist;
auto_vec<vec<chain_op_t> > chains (group_size);
auto_vec<slp_tree, 4> children;
bool hard_fail = true;
for (unsigned lane = 0; lane < group_size; ++lane)
{
/* For each lane linearize the addition/subtraction (or other
uniform associatable operation) expression tree. */
gimple *op_stmt = NULL, *other_op_stmt = NULL;
vect_slp_linearize_chain (vinfo, worklist, chain, code,
stmts[lane]->stmt, op_stmt, other_op_stmt,
NULL);
if (!op_stmt_info && op_stmt)
op_stmt_info = vinfo->lookup_stmt (op_stmt);
if (!other_op_stmt_info && other_op_stmt)
other_op_stmt_info = vinfo->lookup_stmt (other_op_stmt);
if (chain.length () == 2)
{
/* In a chain of just two elements resort to the regular
operand swapping scheme. If we run into a length
mismatch still hard-FAIL. */
if (chain_len == 0)
hard_fail = false;
else
{
matches[lane] = false;
/* ??? We might want to process the other lanes, but
make sure to not give false matching hints to the
caller for lanes we did not process. */
if (lane != group_size - 1)
matches[0] = false;
}
break;
}
else if (chain_len == 0)
chain_len = chain.length ();
else if (chain.length () != chain_len)
{
/* ??? Here we could slip in magic to compensate with
neutral operands. */
matches[lane] = false;
if (lane != group_size - 1)
matches[0] = false;
break;
}
chains.quick_push (chain.copy ());
chain.truncate (0);
}
if (chains.length () == group_size)
{
/* We cannot yet use SLP_TREE_CODE to communicate the operation. */
if (!op_stmt_info)
{
hard_fail = false;
goto out;
}
/* Now we have a set of chains with the same length. */
/* 1. pre-sort according to def_type and operation. */
for (unsigned lane = 0; lane < group_size; ++lane)
chains[lane].stablesort (dt_sort_cmp, vinfo);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"pre-sorted chains of %s\n",
get_tree_code_name (code));
for (unsigned lane = 0; lane < group_size; ++lane)
{
for (unsigned opnum = 0; opnum < chain_len; ++opnum)
dump_printf (MSG_NOTE, "%s %T ",
get_tree_code_name (chains[lane][opnum].code),
chains[lane][opnum].op);
dump_printf (MSG_NOTE, "\n");
}
}
/* 2. try to build children nodes, associating as necessary. */
for (unsigned n = 0; n < chain_len; ++n)
{
vect_def_type dt = chains[0][n].dt;
unsigned lane;
for (lane = 0; lane < group_size; ++lane)
if (chains[lane][n].dt != dt)
{
if (dt == vect_constant_def
&& chains[lane][n].dt == vect_external_def)
dt = vect_external_def;
else if (dt == vect_external_def
&& chains[lane][n].dt == vect_constant_def)
;
else
break;
}
if (lane != group_size)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"giving up on chain due to mismatched "
"def types\n");
matches[lane] = false;
if (lane != group_size - 1)
matches[0] = false;
goto out;
}
if (dt == vect_constant_def
|| dt == vect_external_def)
{
/* Check whether we can build the invariant. If we can't
we never will be able to. */
tree type = TREE_TYPE (chains[0][n].op);
if (!GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
&& (TREE_CODE (type) == BOOLEAN_TYPE
|| !can_duplicate_and_interleave_p (vinfo, group_size,
type)))
{
matches[0] = false;
goto out;
}
vec<tree> ops;
ops.create (group_size);
for (lane = 0; lane < group_size; ++lane)
ops.quick_push (chains[lane][n].op);
slp_tree child = vect_create_new_slp_node (ops);
SLP_TREE_DEF_TYPE (child) = dt;
children.safe_push (child);
}
else if (dt != vect_internal_def)
{
/* Not sure, we might need sth special.
gcc.dg/vect/pr96854.c,
gfortran.dg/vect/fast-math-pr37021.f90
and gfortran.dg/vect/pr61171.f trigger. */
/* Soft-fail for now. */
hard_fail = false;
goto out;
}
else
{
vec<stmt_vec_info> op_stmts;
op_stmts.create (group_size);
slp_tree child = NULL;
/* Brute-force our way. We have to consider a lane
failing after fixing an earlier fail up in the
SLP discovery recursion. So track the current
permute per lane. */
unsigned *perms = XALLOCAVEC (unsigned, group_size);
memset (perms, 0, sizeof (unsigned) * group_size);
do
{
op_stmts.truncate (0);
for (lane = 0; lane < group_size; ++lane)
op_stmts.quick_push
(vinfo->lookup_def (chains[lane][n].op));
child = vect_build_slp_tree (vinfo, op_stmts,
group_size, &this_max_nunits,
matches, limit,
&this_tree_size, bst_map);
/* ??? We're likely getting too many fatal mismatches
here so maybe we want to ignore them (but then we
have no idea which lanes fatally mismatched). */
if (child || !matches[0])
break;
/* Swap another lane we have not yet matched up into
lanes that did not match. If we run out of
permute possibilities for a lane terminate the
search. */
bool term = false;
for (lane = 1; lane < group_size; ++lane)
if (!matches[lane])
{
if (n + perms[lane] + 1 == chain_len)
{
term = true;
break;
}
std::swap (chains[lane][n],
chains[lane][n + perms[lane] + 1]);
perms[lane]++;
}
if (term)
break;
}
while (1);
if (!child)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"failed to match up op %d\n", n);
op_stmts.release ();
if (lane != group_size - 1)
matches[0] = false;
else
matches[lane] = false;
goto out;
}
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"matched up op %d to\n", n);
vect_print_slp_tree (MSG_NOTE, vect_location, child);
}
children.safe_push (child);
}
}
/* 3. build SLP nodes to combine the chain. */
for (unsigned lane = 0; lane < group_size; ++lane)
if (chains[lane][0].code != code)
{
/* See if there's any alternate all-PLUS entry. */
unsigned n;
for (n = 1; n < chain_len; ++n)
{
for (lane = 0; lane < group_size; ++lane)
if (chains[lane][n].code != code)
break;
if (lane == group_size)
break;
}
if (n != chain_len)
{
/* Swap that in at first position. */
std::swap (children[0], children[n]);
for (lane = 0; lane < group_size; ++lane)
std::swap (chains[lane][0], chains[lane][n]);
}
else
{
/* ??? When this triggers and we end up with two
vect_constant/external_def up-front things break (ICE)
spectacularly finding an insertion place for the
all-constant op. We should have a fully
vect_internal_def operand though(?) so we can swap
that into first place and then prepend the all-zero
constant. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"inserting constant zero to compensate "
"for (partially) negated first "
"operand\n");
chain_len++;
for (lane = 0; lane < group_size; ++lane)
chains[lane].safe_insert
(0, chain_op_t (code, vect_constant_def, NULL_TREE));
vec<tree> zero_ops;
zero_ops.create (group_size);
zero_ops.quick_push (build_zero_cst (TREE_TYPE (vectype)));
for (lane = 1; lane < group_size; ++lane)
zero_ops.quick_push (zero_ops[0]);
slp_tree zero = vect_create_new_slp_node (zero_ops);
SLP_TREE_DEF_TYPE (zero) = vect_constant_def;
children.safe_insert (0, zero);
}
break;
}
for (unsigned i = 1; i < children.length (); ++i)
{
slp_tree op0 = children[i - 1];
slp_tree op1 = children[i];
bool this_two_op = false;
for (unsigned lane = 0; lane < group_size; ++lane)
if (chains[lane][i].code != chains[0][i].code)
{
this_two_op = true;
break;
}
slp_tree child;
if (i == children.length () - 1)
child = vect_create_new_slp_node (node, stmts, 2);
else
child = vect_create_new_slp_node (2, ERROR_MARK);
if (this_two_op)
{
vec<std::pair<unsigned, unsigned> > lperm;
lperm.create (group_size);
for (unsigned lane = 0; lane < group_size; ++lane)
lperm.quick_push (std::make_pair
(chains[lane][i].code != chains[0][i].code, lane));
vect_slp_build_two_operator_nodes (child, vectype, op0, op1,
(chains[0][i].code == code
? op_stmt_info
: other_op_stmt_info),
(chains[0][i].code == code
? other_op_stmt_info
: op_stmt_info),
lperm);
}
else
{
SLP_TREE_DEF_TYPE (child) = vect_internal_def;
SLP_TREE_VECTYPE (child) = vectype;
SLP_TREE_LANES (child) = group_size;
SLP_TREE_CHILDREN (child).quick_push (op0);
SLP_TREE_CHILDREN (child).quick_push (op1);
SLP_TREE_REPRESENTATIVE (child)
= (chains[0][i].code == code
? op_stmt_info : other_op_stmt_info);
}
children[i] = child;
}
*tree_size += this_tree_size + 1;
*max_nunits = this_max_nunits;
while (!chains.is_empty ())
chains.pop ().release ();
return node;
}
out:
while (!children.is_empty ())
vect_free_slp_tree (children.pop ());
while (!chains.is_empty ())
chains.pop ().release ();
/* Hard-fail, otherwise we might run into quadratic processing of the
chains starting one stmt into the chain again. */
if (hard_fail)
return NULL;
/* Fall thru to normal processing. */
}
/* Get at the operands, verifying they are compatible. */
vec<slp_oprnd_info> oprnds_info = vect_create_oprnd_info (nops, group_size);
slp_oprnd_info oprnd_info;
FOR_EACH_VEC_ELT (stmts, i, stmt_info)
{
int res = vect_get_and_check_slp_defs (vinfo, swap[i], skip_args,
stmts, i, &oprnds_info);
if (res != 0)
matches[(res == -1) ? 0 : i] = false;
if (!matches[0])
break;
}
for (i = 0; i < group_size; ++i)
if (!matches[i])
{
vect_free_oprnd_info (oprnds_info);
return NULL;
}
swap = NULL;
auto_vec<slp_tree, 4> children;
stmt_info = stmts[0];
/* Create SLP_TREE nodes for the definition node/s. */
FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
{
slp_tree child;
unsigned int j;
/* We're skipping certain operands from processing, for example
outer loop reduction initial defs. */
if (skip_args[i])
{
children.safe_push (NULL);
continue;
}
if (oprnd_info->first_dt == vect_uninitialized_def)
{
/* COND_EXPR have one too many eventually if the condition
is a SSA name. */
gcc_assert (i == 3 && nops == 4);
continue;
}
if (is_a <bb_vec_info> (vinfo)
&& oprnd_info->first_dt == vect_internal_def
&& !oprnd_info->any_pattern)
{
/* For BB vectorization, if all defs are the same do not
bother to continue the build along the single-lane
graph but use a splat of the scalar value. */
stmt_vec_info first_def = oprnd_info->def_stmts[0];
for (j = 1; j < group_size; ++j)
if (oprnd_info->def_stmts[j] != first_def)
break;
if (j == group_size
/* But avoid doing this for loads where we may be
able to CSE things, unless the stmt is not
vectorizable. */
&& (!STMT_VINFO_VECTORIZABLE (first_def)
|| !gimple_vuse (first_def->stmt)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Using a splat of the uniform operand %G",
first_def->stmt);
oprnd_info->first_dt = vect_external_def;
}
}
if (oprnd_info->first_dt == vect_external_def
|| oprnd_info->first_dt == vect_constant_def)
{
slp_tree invnode = vect_create_new_slp_node (oprnd_info->ops);
SLP_TREE_DEF_TYPE (invnode) = oprnd_info->first_dt;
oprnd_info->ops = vNULL;
children.safe_push (invnode);
continue;
}
if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
group_size, &this_max_nunits,
matches, limit,
&this_tree_size, bst_map)) != NULL)
{
oprnd_info->def_stmts = vNULL;
children.safe_push (child);
continue;
}
/* If the SLP build for operand zero failed and operand zero
and one can be commutated try that for the scalar stmts
that failed the match. */
if (i == 0
/* A first scalar stmt mismatch signals a fatal mismatch. */
&& matches[0]
/* ??? For COND_EXPRs we can swap the comparison operands
as well as the arms under some constraints. */
&& nops == 2
&& oprnds_info[1]->first_dt == vect_internal_def
&& is_gimple_assign (stmt_info->stmt)
/* Swapping operands for reductions breaks assumptions later on. */
&& STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def
&& STMT_VINFO_DEF_TYPE (stmt_info) != vect_double_reduction_def)
{
/* See whether we can swap the matching or the non-matching
stmt operands. */
bool swap_not_matching = true;
do
{
for (j = 0; j < group_size; ++j)
{
if (matches[j] != !swap_not_matching)
continue;
stmt_vec_info stmt_info = stmts[j];
/* Verify if we can swap operands of this stmt. */
gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
if (!stmt
|| !commutative_tree_code (gimple_assign_rhs_code (stmt)))
{
if (!swap_not_matching)
goto fail;
swap_not_matching = false;
break;
}
}
}
while (j != group_size);
/* Swap mismatched definition stmts. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Re-trying with swapped operands of stmts ");
for (j = 0; j < group_size; ++j)
if (matches[j] == !swap_not_matching)
{
std::swap (oprnds_info[0]->def_stmts[j],
oprnds_info[1]->def_stmts[j]);
std::swap (oprnds_info[0]->ops[j],
oprnds_info[1]->ops[j]);
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "%d ", j);
}
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "\n");
/* After swapping some operands we lost track whether an
operand has any pattern defs so be conservative here. */
if (oprnds_info[0]->any_pattern || oprnds_info[1]->any_pattern)
oprnds_info[0]->any_pattern = oprnds_info[1]->any_pattern = true;
/* And try again with scratch 'matches' ... */
bool *tem = XALLOCAVEC (bool, group_size);
if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
group_size, &this_max_nunits,
tem, limit,
&this_tree_size, bst_map)) != NULL)
{
oprnd_info->def_stmts = vNULL;
children.safe_push (child);
continue;
}
}
fail:
/* If the SLP build failed and we analyze a basic-block
simply treat nodes we fail to build as externally defined
(and thus build vectors from the scalar defs).
The cost model will reject outright expensive cases.
??? This doesn't treat cases where permutation ultimatively
fails (or we don't try permutation below). Ideally we'd
even compute a permutation that will end up with the maximum
SLP tree size... */
if (is_a <bb_vec_info> (vinfo)
/* ??? Rejecting patterns this way doesn't work. We'd have to
do extra work to cancel the pattern so the uses see the
scalar version. */
&& !is_pattern_stmt_p (stmt_info)
&& !oprnd_info->any_pattern)
{
/* But if there's a leading vector sized set of matching stmts
fail here so we can split the group. This matches the condition
vect_analyze_slp_instance uses. */
/* ??? We might want to split here and combine the results to support
multiple vector sizes better. */
for (j = 0; j < group_size; ++j)
if (!matches[j])
break;
if (!known_ge (j, TYPE_VECTOR_SUBPARTS (vectype)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Building vector operands from scalars\n");
this_tree_size++;
child = vect_create_new_slp_node (oprnd_info->ops);
children.safe_push (child);
oprnd_info->ops = vNULL;
continue;
}
}
gcc_assert (child == NULL);
FOR_EACH_VEC_ELT (children, j, child)
if (child)
vect_free_slp_tree (child);
vect_free_oprnd_info (oprnds_info);
return NULL;
}
vect_free_oprnd_info (oprnds_info);
/* If we have all children of a child built up from uniform scalars
or does more than one possibly expensive vector construction then
just throw that away, causing it built up from scalars.
The exception is the SLP node for the vector store. */
if (is_a <bb_vec_info> (vinfo)
&& !STMT_VINFO_GROUPED_ACCESS (stmt_info)
/* ??? Rejecting patterns this way doesn't work. We'd have to
do extra work to cancel the pattern so the uses see the
scalar version. */
&& !is_pattern_stmt_p (stmt_info))
{
slp_tree child;
unsigned j;
bool all_uniform_p = true;
unsigned n_vector_builds = 0;
FOR_EACH_VEC_ELT (children, j, child)
{
if (!child)
;
else if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
all_uniform_p = false;
else if (!vect_slp_tree_uniform_p (child))
{
all_uniform_p = false;
if (SLP_TREE_DEF_TYPE (child) == vect_external_def)
n_vector_builds++;
}
}
if (all_uniform_p
|| n_vector_builds > 1
|| (n_vector_builds == children.length ()
&& is_a <gphi *> (stmt_info->stmt)))
{
/* Roll back. */
matches[0] = false;
FOR_EACH_VEC_ELT (children, j, child)
if (child)
vect_free_slp_tree (child);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Building parent vector operands from "
"scalars instead\n");
return NULL;
}
}
*tree_size += this_tree_size + 1;
*max_nunits = this_max_nunits;
if (two_operators)
{
/* ??? We'd likely want to either cache in bst_map sth like
{ a+b, NULL, a+b, NULL } and { NULL, a-b, NULL, a-b } or
the true { a+b, a+b, a+b, a+b } ... but there we don't have
explicit stmts to put in so the keying on 'stmts' doesn't
work (but we have the same issue with nodes that use 'ops'). */
slp_tree one = new _slp_tree;
slp_tree two = new _slp_tree;
SLP_TREE_DEF_TYPE (one) = vect_internal_def;
SLP_TREE_DEF_TYPE (two) = vect_internal_def;
SLP_TREE_VECTYPE (one) = vectype;
SLP_TREE_VECTYPE (two) = vectype;
SLP_TREE_CHILDREN (one).safe_splice (children);
SLP_TREE_CHILDREN (two).safe_splice (children);
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (two), i, child)
SLP_TREE_REF_COUNT (child)++;
/* Here we record the original defs since this
node represents the final lane configuration. */
node = vect_create_new_slp_node (node, stmts, 2);
SLP_TREE_VECTYPE (node) = vectype;
SLP_TREE_CODE (node) = VEC_PERM_EXPR;
SLP_TREE_CHILDREN (node).quick_push (one);
SLP_TREE_CHILDREN (node).quick_push (two);
gassign *stmt = as_a <gassign *> (stmts[0]->stmt);
enum tree_code code0 = gimple_assign_rhs_code (stmt);
enum tree_code ocode = ERROR_MARK;
stmt_vec_info ostmt_info;
unsigned j = 0;
FOR_EACH_VEC_ELT (stmts, i, ostmt_info)
{
gassign *ostmt = as_a <gassign *> (ostmt_info->stmt);
if (gimple_assign_rhs_code (ostmt) != code0)
{
SLP_TREE_LANE_PERMUTATION (node).safe_push (std::make_pair (1, i));
ocode = gimple_assign_rhs_code (ostmt);
j = i;
}
else
SLP_TREE_LANE_PERMUTATION (node).safe_push (std::make_pair (0, i));
}
SLP_TREE_CODE (one) = code0;
SLP_TREE_CODE (two) = ocode;
SLP_TREE_LANES (one) = stmts.length ();
SLP_TREE_LANES (two) = stmts.length ();
SLP_TREE_REPRESENTATIVE (one) = stmts[0];
SLP_TREE_REPRESENTATIVE (two) = stmts[j];
return node;
}
node = vect_create_new_slp_node (node, stmts, nops);
SLP_TREE_VECTYPE (node) = vectype;
SLP_TREE_CHILDREN (node).splice (children);
return node;
}
/* Dump a single SLP tree NODE. */
static void
vect_print_slp_tree (dump_flags_t dump_kind, dump_location_t loc,
slp_tree node)
{
unsigned i, j;
slp_tree child;
stmt_vec_info stmt_info;
tree op;
dump_metadata_t metadata (dump_kind, loc.get_impl_location ());
dump_user_location_t user_loc = loc.get_user_location ();
dump_printf_loc (metadata, user_loc,
"node%s %p (max_nunits=" HOST_WIDE_INT_PRINT_UNSIGNED
", refcnt=%u)",
SLP_TREE_DEF_TYPE (node) == vect_external_def
? " (external)"
: (SLP_TREE_DEF_TYPE (node) == vect_constant_def
? " (constant)"
: ""), (void *) node,
estimated_poly_value (node->max_nunits),
SLP_TREE_REF_COUNT (node));
if (SLP_TREE_VECTYPE (node))
dump_printf (metadata, " %T", SLP_TREE_VECTYPE (node));
dump_printf (metadata, "\n");
if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
{
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
dump_printf_loc (metadata, user_loc, "op: VEC_PERM_EXPR\n");
else
dump_printf_loc (metadata, user_loc, "op template: %G",
SLP_TREE_REPRESENTATIVE (node)->stmt);
}
if (SLP_TREE_SCALAR_STMTS (node).exists ())
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
dump_printf_loc (metadata, user_loc, "\tstmt %u %G", i, stmt_info->stmt);
else
{
dump_printf_loc (metadata, user_loc, "\t{ ");
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
dump_printf (metadata, "%T%s ", op,
i < SLP_TREE_SCALAR_OPS (node).length () - 1 ? "," : "");
dump_printf (metadata, "}\n");
}
if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
{
dump_printf_loc (metadata, user_loc, "\tload permutation {");
FOR_EACH_VEC_ELT (SLP_TREE_LOAD_PERMUTATION (node), i, j)
dump_printf (dump_kind, " %u", j);
dump_printf (dump_kind, " }\n");
}
if (SLP_TREE_LANE_PERMUTATION (node).exists ())
{
dump_printf_loc (metadata, user_loc, "\tlane permutation {");
for (i = 0; i < SLP_TREE_LANE_PERMUTATION (node).length (); ++i)
dump_printf (dump_kind, " %u[%u]",
SLP_TREE_LANE_PERMUTATION (node)[i].first,
SLP_TREE_LANE_PERMUTATION (node)[i].second);
dump_printf (dump_kind, " }\n");
}
if (SLP_TREE_CHILDREN (node).is_empty ())
return;
dump_printf_loc (metadata, user_loc, "\tchildren");
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
dump_printf (dump_kind, " %p", (void *)child);
dump_printf (dump_kind, "\n");
}
DEBUG_FUNCTION void
debug (slp_tree node)
{
debug_dump_context ctx;
vect_print_slp_tree (MSG_NOTE,
dump_location_t::from_location_t (UNKNOWN_LOCATION),
node);
}
/* Recursive helper for the dot producer below. */
static void
dot_slp_tree (FILE *f, slp_tree node, hash_set<slp_tree> &visited)
{
if (visited.add (node))
return;
fprintf (f, "\"%p\" [label=\"", (void *)node);
vect_print_slp_tree (MSG_NOTE,
dump_location_t::from_location_t (UNKNOWN_LOCATION),
node);
fprintf (f, "\"];\n");
for (slp_tree child : SLP_TREE_CHILDREN (node))
fprintf (f, "\"%p\" -> \"%p\";", (void *)node, (void *)child);
for (slp_tree child : SLP_TREE_CHILDREN (node))
if (child)
dot_slp_tree (f, child, visited);
}
DEBUG_FUNCTION void
dot_slp_tree (const char *fname, slp_tree node)
{
FILE *f = fopen (fname, "w");
fprintf (f, "digraph {\n");
fflush (f);
{
debug_dump_context ctx (f);
hash_set<slp_tree> visited;
dot_slp_tree (f, node, visited);
}
fflush (f);
fprintf (f, "}\n");
fclose (f);
}
/* Dump a slp tree NODE using flags specified in DUMP_KIND. */
static void
vect_print_slp_graph (dump_flags_t dump_kind, dump_location_t loc,
slp_tree node, hash_set<slp_tree> &visited)
{
unsigned i;
slp_tree child;
if (visited.add (node))
return;
vect_print_slp_tree (dump_kind, loc, node);
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
vect_print_slp_graph (dump_kind, loc, child, visited);
}
static void
vect_print_slp_graph (dump_flags_t dump_kind, dump_location_t loc,
slp_tree entry)
{
hash_set<slp_tree> visited;
vect_print_slp_graph (dump_kind, loc, entry, visited);
}
/* Mark the tree rooted at NODE with PURE_SLP. */
static void
vect_mark_slp_stmts (slp_tree node, hash_set<slp_tree> &visited)
{
int i;
stmt_vec_info stmt_info;
slp_tree child;
if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
return;
if (visited.add (node))
return;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
STMT_SLP_TYPE (stmt_info) = pure_slp;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
vect_mark_slp_stmts (child, visited);
}
static void
vect_mark_slp_stmts (slp_tree node)
{
hash_set<slp_tree> visited;
vect_mark_slp_stmts (node, visited);
}
/* Mark the statements of the tree rooted at NODE as relevant (vect_used). */
static void
vect_mark_slp_stmts_relevant (slp_tree node, hash_set<slp_tree> &visited)
{
int i;
stmt_vec_info stmt_info;
slp_tree child;
if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
return;
if (visited.add (node))
return;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
{
gcc_assert (!STMT_VINFO_RELEVANT (stmt_info)
|| STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope);
STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope;
}
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
vect_mark_slp_stmts_relevant (child, visited);
}
static void
vect_mark_slp_stmts_relevant (slp_tree node)
{
hash_set<slp_tree> visited;
vect_mark_slp_stmts_relevant (node, visited);
}
/* Gather loads in the SLP graph NODE and populate the INST loads array. */
static void
vect_gather_slp_loads (vec<slp_tree> &loads, slp_tree node,
hash_set<slp_tree> &visited)
{
if (!node || visited.add (node))
return;
if (SLP_TREE_CHILDREN (node).length () == 0)
{
if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
return;
stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
if (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
loads.safe_push (node);
}
else
{
unsigned i;
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
vect_gather_slp_loads (loads, child, visited);
}
}
/* Find the last store in SLP INSTANCE. */
stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree node)
{
stmt_vec_info last = NULL;
stmt_vec_info stmt_vinfo;
for (int i = 0; SLP_TREE_SCALAR_STMTS (node).iterate (i, &stmt_vinfo); i++)
{
stmt_vinfo = vect_orig_stmt (stmt_vinfo);
last = last ? get_later_stmt (stmt_vinfo, last) : stmt_vinfo;
}
return last;
}
/* Find the first stmt in NODE. */
stmt_vec_info
vect_find_first_scalar_stmt_in_slp (slp_tree node)
{
stmt_vec_info first = NULL;
stmt_vec_info stmt_vinfo;
for (int i = 0; SLP_TREE_SCALAR_STMTS (node).iterate (i, &stmt_vinfo); i++)
{
stmt_vinfo = vect_orig_stmt (stmt_vinfo);
if (!first
|| get_later_stmt (stmt_vinfo, first) == first)
first = stmt_vinfo;
}
return first;
}
/* Splits a group of stores, currently beginning at FIRST_VINFO, into
two groups: one (still beginning at FIRST_VINFO) of size GROUP1_SIZE
(also containing the first GROUP1_SIZE stmts, since stores are
consecutive), the second containing the remainder.
Return the first stmt in the second group. */
static stmt_vec_info
vect_split_slp_store_group (stmt_vec_info first_vinfo, unsigned group1_size)
{
gcc_assert (DR_GROUP_FIRST_ELEMENT (first_vinfo) == first_vinfo);
gcc_assert (group1_size > 0);
int group2_size = DR_GROUP_SIZE (first_vinfo) - group1_size;
gcc_assert (group2_size > 0);
DR_GROUP_SIZE (first_vinfo) = group1_size;
stmt_vec_info stmt_info = first_vinfo;
for (unsigned i = group1_size; i > 1; i--)
{
stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info);
gcc_assert (DR_GROUP_GAP (stmt_info) == 1);
}
/* STMT is now the last element of the first group. */
stmt_vec_info group2 = DR_GROUP_NEXT_ELEMENT (stmt_info);
DR_GROUP_NEXT_ELEMENT (stmt_info) = 0;
DR_GROUP_SIZE (group2) = group2_size;
for (stmt_info = group2; stmt_info;
stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info))
{
DR_GROUP_FIRST_ELEMENT (stmt_info) = group2;
gcc_assert (DR_GROUP_GAP (stmt_info) == 1);
}
/* For the second group, the DR_GROUP_GAP is that before the original group,
plus skipping over the first vector. */
DR_GROUP_GAP (group2) = DR_GROUP_GAP (first_vinfo) + group1_size;
/* DR_GROUP_GAP of the first group now has to skip over the second group too. */
DR_GROUP_GAP (first_vinfo) += group2_size;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "Split group into %d and %d\n",
group1_size, group2_size);
return group2;
}
/* Calculate the unrolling factor for an SLP instance with GROUP_SIZE
statements and a vector of NUNITS elements. */
static poly_uint64
calculate_unrolling_factor (poly_uint64 nunits, unsigned int group_size)
{
return exact_div (common_multiple (nunits, group_size), group_size);
}
/* Helper that checks to see if a node is a load node. */
static inline bool
vect_is_slp_load_node (slp_tree root)
{
return SLP_TREE_DEF_TYPE (root) == vect_internal_def
&& STMT_VINFO_GROUPED_ACCESS (SLP_TREE_REPRESENTATIVE (root))
&& DR_IS_READ (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (root)));
}
/* Helper function of optimize_load_redistribution that performs the operation
recursively. */
static slp_tree
optimize_load_redistribution_1 (scalar_stmts_to_slp_tree_map_t *bst_map,
vec_info *vinfo, unsigned int group_size,
hash_map<slp_tree, slp_tree> *load_map,
slp_tree root)
{
if (slp_tree *leader = load_map->get (root))
return *leader;
slp_tree node;
unsigned i;
/* For now, we don't know anything about externals so do not do anything. */
if (!root || SLP_TREE_DEF_TYPE (root) != vect_internal_def)
return NULL;
else if (SLP_TREE_CODE (root) == VEC_PERM_EXPR)
{
/* First convert this node into a load node and add it to the leaves
list and flatten the permute from a lane to a load one. If it's
unneeded it will be elided later. */
vec<stmt_vec_info> stmts;
stmts.create (SLP_TREE_LANES (root));
lane_permutation_t lane_perm = SLP_TREE_LANE_PERMUTATION (root);
for (unsigned j = 0; j < lane_perm.length (); j++)
{
std::pair<unsigned, unsigned> perm = lane_perm[j];
node = SLP_TREE_CHILDREN (root)[perm.first];
if (!vect_is_slp_load_node (node)
|| SLP_TREE_CHILDREN (node).exists ())
{
stmts.release ();
goto next;
}
stmts.quick_push (SLP_TREE_SCALAR_STMTS (node)[perm.second]);
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"converting stmts on permute node %p\n",
(void *) root);
bool *matches = XALLOCAVEC (bool, group_size);
poly_uint64 max_nunits = 1;
unsigned tree_size = 0, limit = 1;
node = vect_build_slp_tree (vinfo, stmts, group_size, &max_nunits,
matches, &limit, &tree_size, bst_map);
if (!node)
stmts.release ();
load_map->put (root, node);
return node;
}
next:
load_map->put (root, NULL);
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i , node)
{
slp_tree value
= optimize_load_redistribution_1 (bst_map, vinfo, group_size, load_map,
node);
if (value)
{
SLP_TREE_REF_COUNT (value)++;
SLP_TREE_CHILDREN (root)[i] = value;
/* ??? We know the original leafs of the replaced nodes will
be referenced by bst_map, only the permutes created by
pattern matching are not. */
if (SLP_TREE_REF_COUNT (node) == 1)
load_map->remove (node);
vect_free_slp_tree (node);
}
}
return NULL;
}
/* Temporary workaround for loads not being CSEd during SLP build. This
function will traverse the SLP tree rooted in ROOT for INSTANCE and find
VEC_PERM nodes that blend vectors from multiple nodes that all read from the
same DR such that the final operation is equal to a permuted load. Such
NODES are then directly converted into LOADS themselves. The nodes are
CSEd using BST_MAP. */
static void
optimize_load_redistribution (scalar_stmts_to_slp_tree_map_t *bst_map,
vec_info *vinfo, unsigned int group_size,
hash_map<slp_tree, slp_tree> *load_map,
slp_tree root)
{
slp_tree node;
unsigned i;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i , node)
{
slp_tree value
= optimize_load_redistribution_1 (bst_map, vinfo, group_size, load_map,
node);
if (value)
{
SLP_TREE_REF_COUNT (value)++;
SLP_TREE_CHILDREN (root)[i] = value;
/* ??? We know the original leafs of the replaced nodes will
be referenced by bst_map, only the permutes created by
pattern matching are not. */
if (SLP_TREE_REF_COUNT (node) == 1)
load_map->remove (node);
vect_free_slp_tree (node);
}
}
}
/* Helper function of vect_match_slp_patterns.
Attempts to match patterns against the slp tree rooted in REF_NODE using
VINFO. Patterns are matched in post-order traversal.
If matching is successful the value in REF_NODE is updated and returned, if
not then it is returned unchanged. */
static bool
vect_match_slp_patterns_2 (slp_tree *ref_node, vec_info *vinfo,
slp_tree_to_load_perm_map_t *perm_cache,
slp_compat_nodes_map_t *compat_cache,
hash_set<slp_tree> *visited)
{
unsigned i;
slp_tree node = *ref_node;
bool found_p = false;
if (!node || visited->add (node))
return false;
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
found_p |= vect_match_slp_patterns_2 (&SLP_TREE_CHILDREN (node)[i],
vinfo, perm_cache, compat_cache,
visited);
for (unsigned x = 0; x < num__slp_patterns; x++)
{
vect_pattern *pattern
= slp_patterns[x] (perm_cache, compat_cache, ref_node);
if (pattern)
{
pattern->build (vinfo);
delete pattern;
found_p = true;
}
}
return found_p;
}
/* Applies pattern matching to the given SLP tree rooted in REF_NODE using
vec_info VINFO.
The modified tree is returned. Patterns are tried in order and multiple
patterns may match. */
static bool
vect_match_slp_patterns (slp_instance instance, vec_info *vinfo,
hash_set<slp_tree> *visited,
slp_tree_to_load_perm_map_t *perm_cache,
slp_compat_nodes_map_t *compat_cache)
{
DUMP_VECT_SCOPE ("vect_match_slp_patterns");
slp_tree *ref_node = &SLP_INSTANCE_TREE (instance);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Analyzing SLP tree %p for patterns\n",
(void *) SLP_INSTANCE_TREE (instance));
return vect_match_slp_patterns_2 (ref_node, vinfo, perm_cache, compat_cache,
visited);
}
/* STMT_INFO is a store group of size GROUP_SIZE that we are considering
splitting into two, with the first split group having size NEW_GROUP_SIZE.
Return true if we could use IFN_STORE_LANES instead and if that appears
to be the better approach. */
static bool
vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
unsigned int group_size,
unsigned int new_group_size)
{
tree scalar_type = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
if (!vectype)
return false;
/* Allow the split if one of the two new groups would operate on full
vectors *within* rather than across one scalar loop iteration.
This is purely a heuristic, but it should work well for group
sizes of 3 and 4, where the possible splits are:
3->2+1: OK if the vector has exactly two elements
4->2+2: Likewise
4->3+1: Less clear-cut. */
if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
|| multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
return false;
return vect_store_lanes_supported (vectype, group_size, false);
}
/* Analyze an SLP instance starting from a group of grouped stores. Call
vect_build_slp_tree to build a tree of packed stmts if possible.
Return FALSE if it's impossible to SLP any stmt in the loop. */
static bool
vect_analyze_slp_instance (vec_info *vinfo,
scalar_stmts_to_slp_tree_map_t *bst_map,
stmt_vec_info stmt_info, slp_instance_kind kind,
unsigned max_tree_size, unsigned *limit);
/* Analyze an SLP instance starting from SCALAR_STMTS which are a group
of KIND. Return true if successful. */
static bool
vect_build_slp_instance (vec_info *vinfo,
slp_instance_kind kind,
vec<stmt_vec_info> &scalar_stmts,
vec<stmt_vec_info> &root_stmt_infos,
unsigned max_tree_size, unsigned *limit,
scalar_stmts_to_slp_tree_map_t *bst_map,
/* ??? We need stmt_info for group splitting. */
stmt_vec_info stmt_info_)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Starting SLP discovery for\n");
for (unsigned i = 0; i < scalar_stmts.length (); ++i)
dump_printf_loc (MSG_NOTE, vect_location,
" %G", scalar_stmts[i]->stmt);
}
/* Build the tree for the SLP instance. */
unsigned int group_size = scalar_stmts.length ();
bool *matches = XALLOCAVEC (bool, group_size);
poly_uint64 max_nunits = 1;
unsigned tree_size = 0;
unsigned i;
slp_tree node = vect_build_slp_tree (vinfo, scalar_stmts, group_size,
&max_nunits, matches, limit,
&tree_size, bst_map);
if (node != NULL)
{
/* Calculate the unrolling factor based on the smallest type. */
poly_uint64 unrolling_factor
= calculate_unrolling_factor (max_nunits, group_size);
if (maybe_ne (unrolling_factor, 1U)
&& is_a <bb_vec_info> (vinfo))
{
unsigned HOST_WIDE_INT const_max_nunits;
if (!max_nunits.is_constant (&const_max_nunits)
|| const_max_nunits > group_size)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Build SLP failed: store group "
"size not a multiple of the vector size "
"in basic block SLP\n");
vect_free_slp_tree (node);
return false;
}
/* Fatal mismatch. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"SLP discovery succeeded but node needs "
"splitting\n");
memset (matches, true, group_size);
matches[group_size / const_max_nunits * const_max_nunits] = false;
vect_free_slp_tree (node);
}
else
{
/* Create a new SLP instance. */
slp_instance new_instance = XNEW (class _slp_instance);
SLP_INSTANCE_TREE (new_instance) = node;
SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor;
SLP_INSTANCE_LOADS (new_instance) = vNULL;
SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
SLP_INSTANCE_KIND (new_instance) = kind;
new_instance->reduc_phis = NULL;
new_instance->cost_vec = vNULL;
new_instance->subgraph_entries = vNULL;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"SLP size %u vs. limit %u.\n",
tree_size, max_tree_size);
/* Fixup SLP reduction chains. */
if (kind == slp_inst_kind_reduc_chain)
{
/* If this is a reduction chain with a conversion in front
amend the SLP tree with a node for that. */
gimple *scalar_def
= vect_orig_stmt (scalar_stmts[group_size - 1])->stmt;
if (STMT_VINFO_DEF_TYPE (scalar_stmts[0]) != vect_reduction_def)
{
/* Get at the conversion stmt - we know it's the single use
of the last stmt of the reduction chain. */
use_operand_p use_p;
bool r = single_imm_use (gimple_assign_lhs (scalar_def),
&use_p, &scalar_def);
gcc_assert (r);
stmt_vec_info next_info = vinfo->lookup_stmt (scalar_def);
next_info = vect_stmt_to_vectorize (next_info);
scalar_stmts = vNULL;
scalar_stmts.create (group_size);
for (unsigned i = 0; i < group_size; ++i)
scalar_stmts.quick_push (next_info);
slp_tree conv = vect_create_new_slp_node (scalar_stmts, 1);
SLP_TREE_VECTYPE (conv) = STMT_VINFO_VECTYPE (next_info);
SLP_TREE_CHILDREN (conv).quick_push (node);
SLP_INSTANCE_TREE (new_instance) = conv;
/* We also have to fake this conversion stmt as SLP reduction
group so we don't have to mess with too much code
elsewhere. */
REDUC_GROUP_FIRST_ELEMENT (next_info) = next_info;
REDUC_GROUP_NEXT_ELEMENT (next_info) = NULL;
}
/* Fill the backedge child of the PHI SLP node. The
general matching code cannot find it because the
scalar code does not reflect how we vectorize the
reduction. */
use_operand_p use_p;
imm_use_iterator imm_iter;
class loop *loop = LOOP_VINFO_LOOP (as_a <loop_vec_info> (vinfo));
FOR_EACH_IMM_USE_FAST (use_p, imm_iter,
gimple_get_lhs (scalar_def))
/* There are exactly two non-debug uses, the reduction
PHI and the loop-closed PHI node. */
if (!is_gimple_debug (USE_STMT (use_p))
&& gimple_bb (USE_STMT (use_p)) == loop->header)
{
auto_vec<stmt_vec_info, 64> phis (group_size);
stmt_vec_info phi_info
= vinfo->lookup_stmt (USE_STMT (use_p));
for (unsigned i = 0; i < group_size; ++i)
phis.quick_push (phi_info);
slp_tree *phi_node = bst_map->get (phis);
unsigned dest_idx = loop_latch_edge (loop)->dest_idx;
SLP_TREE_CHILDREN (*phi_node)[dest_idx]
= SLP_INSTANCE_TREE (new_instance);
SLP_INSTANCE_TREE (new_instance)->refcnt++;
}
}
vinfo->slp_instances.safe_push (new_instance);
/* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
the number of scalar stmts in the root in a few places.
Verify that assumption holds. */
gcc_assert (SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (new_instance))
.length () == group_size);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Final SLP tree for instance %p:\n",
(void *) new_instance);
vect_print_slp_graph (MSG_NOTE, vect_location,
SLP_INSTANCE_TREE (new_instance));
}
return true;
}
}
else
{
/* Failed to SLP. */
/* Free the allocated memory. */
scalar_stmts.release ();
}
stmt_vec_info stmt_info = stmt_info_;
/* Try to break the group up into pieces. */
if (kind == slp_inst_kind_store)
{
/* ??? We could delay all the actual splitting of store-groups
until after SLP discovery of the original group completed.
Then we can recurse to vect_build_slp_instance directly. */
for (i = 0; i < group_size; i++)
if (!matches[i])
break;
/* For basic block SLP, try to break the group up into multiples of
a vector size. */
if (is_a <bb_vec_info> (vinfo)
&& (i > 1 && i < group_size))
{
tree scalar_type
= TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
1 << floor_log2 (i));
unsigned HOST_WIDE_INT const_nunits;
if (vectype
&& TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits))
{
/* Split into two groups at the first vector boundary. */
gcc_assert ((const_nunits & (const_nunits - 1)) == 0);
unsigned group1_size = i & ~(const_nunits - 1);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Splitting SLP group at stmt %u\n", i);
stmt_vec_info rest = vect_split_slp_store_group (stmt_info,
group1_size);
bool res = vect_analyze_slp_instance (vinfo, bst_map, stmt_info,
kind, max_tree_size,
limit);
/* Split the rest at the failure point and possibly
re-analyze the remaining matching part if it has
at least two lanes. */
if (group1_size < i
&& (i + 1 < group_size
|| i - group1_size > 1))
{
stmt_vec_info rest2 = rest;
rest = vect_split_slp_store_group (rest, i - group1_size);
if (i - group1_size > 1)
res |= vect_analyze_slp_instance (vinfo, bst_map, rest2,
kind, max_tree_size,
limit);
}
/* Re-analyze the non-matching tail if it has at least
two lanes. */
if (i + 1 < group_size)
res |= vect_analyze_slp_instance (vinfo, bst_map,
rest, kind, max_tree_size,
limit);
return res;
}
}
/* For loop vectorization split into arbitrary pieces of size > 1. */
if (is_a <loop_vec_info> (vinfo)
&& (i > 1 && i < group_size)
&& !vect_slp_prefer_store_lanes_p (vinfo, stmt_info, group_size, i))
{
unsigned group1_size = i;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Splitting SLP group at stmt %u\n", i);
stmt_vec_info rest = vect_split_slp_store_group (stmt_info,
group1_size);
/* Loop vectorization cannot handle gaps in stores, make sure
the split group appears as strided. */
STMT_VINFO_STRIDED_P (rest) = 1;
DR_GROUP_GAP (rest) = 0;
STMT_VINFO_STRIDED_P (stmt_info) = 1;
DR_GROUP_GAP (stmt_info) = 0;
bool res = vect_analyze_slp_instance (vinfo, bst_map, stmt_info,
kind, max_tree_size, limit);
if (i + 1 < group_size)
res |= vect_analyze_slp_instance (vinfo, bst_map,
rest, kind, max_tree_size, limit);
return res;
}
/* Even though the first vector did not all match, we might be able to SLP
(some) of the remainder. FORNOW ignore this possibility. */
}
/* Failed to SLP. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "SLP discovery failed\n");
return false;
}
/* Analyze an SLP instance starting from a group of grouped stores. Call
vect_build_slp_tree to build a tree of packed stmts if possible.
Return FALSE if it's impossible to SLP any stmt in the loop. */
static bool
vect_analyze_slp_instance (vec_info *vinfo,
scalar_stmts_to_slp_tree_map_t *bst_map,
stmt_vec_info stmt_info,
slp_instance_kind kind,
unsigned max_tree_size, unsigned *limit)
{
unsigned int i;
vec<stmt_vec_info> scalar_stmts;
if (is_a <bb_vec_info> (vinfo))
vect_location = stmt_info->stmt;
stmt_vec_info next_info = stmt_info;
if (kind == slp_inst_kind_store)
{
/* Collect the stores and store them in scalar_stmts. */
scalar_stmts.create (DR_GROUP_SIZE (stmt_info));
while (next_info)
{
scalar_stmts.quick_push (vect_stmt_to_vectorize (next_info));
next_info = DR_GROUP_NEXT_ELEMENT (next_info);
}
}
else if (kind == slp_inst_kind_reduc_chain)
{
/* Collect the reduction stmts and store them in scalar_stmts. */
scalar_stmts.create (REDUC_GROUP_SIZE (stmt_info));
while (next_info)
{
scalar_stmts.quick_push (vect_stmt_to_vectorize (next_info));
next_info = REDUC_GROUP_NEXT_ELEMENT (next_info);
}
/* Mark the first element of the reduction chain as reduction to properly
transform the node. In the reduction analysis phase only the last
element of the chain is marked as reduction. */
STMT_VINFO_DEF_TYPE (stmt_info)
= STMT_VINFO_DEF_TYPE (scalar_stmts.last ());
STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info))
= STMT_VINFO_REDUC_DEF (vect_orig_stmt (scalar_stmts.last ()));
}
else if (kind == slp_inst_kind_ctor)
{
tree rhs = gimple_assign_rhs1 (stmt_info->stmt);
tree val;
scalar_stmts.create (CONSTRUCTOR_NELTS (rhs));
FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (rhs), i, val)
{
stmt_vec_info def_info = vinfo->lookup_def (val);
def_info = vect_stmt_to_vectorize (def_info);
scalar_stmts.quick_push (def_info);
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Analyzing vectorizable constructor: %G\n",
stmt_info->stmt);
}
else if (kind == slp_inst_kind_reduc_group)
{
/* Collect reduction statements. */
const vec<stmt_vec_info> &reductions
= as_a <loop_vec_info> (vinfo)->reductions;
scalar_stmts.create (reductions.length ());
for (i = 0; reductions.iterate (i, &next_info); i++)
if ((STMT_VINFO_RELEVANT_P (next_info)
|| STMT_VINFO_LIVE_P (next_info))
/* ??? Make sure we didn't skip a conversion around a reduction
path. In that case we'd have to reverse engineer that conversion
stmt following the chain using reduc_idx and from the PHI
using reduc_def. */
&& STMT_VINFO_DEF_TYPE (next_info) == vect_reduction_def)
scalar_stmts.quick_push (next_info);
/* If less than two were relevant/live there's nothing to SLP. */
if (scalar_stmts.length () < 2)
return false;
}
else
gcc_unreachable ();
vec<stmt_vec_info> roots = vNULL;
if (kind == slp_inst_kind_ctor)
{
roots.create (1);
roots.quick_push (stmt_info);
}
/* Build the tree for the SLP instance. */
bool res = vect_build_slp_instance (vinfo, kind, scalar_stmts,
roots,
max_tree_size, limit, bst_map,
kind == slp_inst_kind_store
? stmt_info : NULL);
if (!res)
roots.release ();
/* ??? If this is slp_inst_kind_store and the above succeeded here's
where we should do store group splitting. */
return res;
}
/* Check if there are stmts in the loop can be vectorized using SLP. Build SLP
trees of packed scalar stmts if SLP is possible. */
opt_result
vect_analyze_slp (vec_info *vinfo, unsigned max_tree_size)
{
unsigned int i;
stmt_vec_info first_element;
slp_instance instance;
DUMP_VECT_SCOPE ("vect_analyze_slp");
unsigned limit = max_tree_size;
scalar_stmts_to_slp_tree_map_t *bst_map
= new scalar_stmts_to_slp_tree_map_t ();
/* Find SLP sequences starting from groups of grouped stores. */
FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element)
vect_analyze_slp_instance (vinfo, bst_map, first_element,
STMT_VINFO_GROUPED_ACCESS (first_element)
? slp_inst_kind_store : slp_inst_kind_ctor,
max_tree_size, &limit);
if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo))
{
for (unsigned i = 0; i < bb_vinfo->roots.length (); ++i)
{
vect_location = bb_vinfo->roots[i].roots[0]->stmt;
if (vect_build_slp_instance (bb_vinfo, bb_vinfo->roots[i].kind,
bb_vinfo->roots[i].stmts,
bb_vinfo->roots[i].roots,
max_tree_size, &limit, bst_map, NULL))
{
bb_vinfo->roots[i].stmts = vNULL;
bb_vinfo->roots[i].roots = vNULL;
}
}
}
if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
{
/* Find SLP sequences starting from reduction chains. */
FOR_EACH_VEC_ELT (loop_vinfo->reduction_chains, i, first_element)
if (! STMT_VINFO_RELEVANT_P (first_element)
&& ! STMT_VINFO_LIVE_P (first_element))
;
else if (! vect_analyze_slp_instance (vinfo, bst_map, first_element,
slp_inst_kind_reduc_chain,
max_tree_size, &limit))
{
/* Dissolve reduction chain group. */
stmt_vec_info vinfo = first_element;
stmt_vec_info last = NULL;
while (vinfo)
{
stmt_vec_info next = REDUC_GROUP_NEXT_ELEMENT (vinfo);
REDUC_GROUP_FIRST_ELEMENT (vinfo) = NULL;
REDUC_GROUP_NEXT_ELEMENT (vinfo) = NULL;
last = vinfo;
vinfo = next;
}
STMT_VINFO_DEF_TYPE (first_element) = vect_internal_def;
/* It can be still vectorized as part of an SLP reduction. */
loop_vinfo->reductions.safe_push (last);
}
/* Find SLP sequences starting from groups of reductions. */
if (loop_vinfo->reductions.length () > 1)
vect_analyze_slp_instance (vinfo, bst_map, loop_vinfo->reductions[0],
slp_inst_kind_reduc_group, max_tree_size,
&limit);
}
hash_set<slp_tree> visited_patterns;
slp_tree_to_load_perm_map_t perm_cache;
slp_compat_nodes_map_t compat_cache;
/* See if any patterns can be found in the SLP tree. */
bool pattern_found = false;
FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
pattern_found |= vect_match_slp_patterns (instance, vinfo,
&visited_patterns, &perm_cache,
&compat_cache);
/* If any were found optimize permutations of loads. */
if (pattern_found)
{
hash_map<slp_tree, slp_tree> load_map;
FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
{
slp_tree root = SLP_INSTANCE_TREE (instance);
optimize_load_redistribution (bst_map, vinfo, SLP_TREE_LANES (root),
&load_map, root);
}
}
/* The map keeps a reference on SLP nodes built, release that. */
for (scalar_stmts_to_slp_tree_map_t::iterator it = bst_map->begin ();
it != bst_map->end (); ++it)
if ((*it).second)
vect_free_slp_tree ((*it).second);
delete bst_map;
if (pattern_found && dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Pattern matched SLP tree\n");
hash_set<slp_tree> visited;
FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
vect_print_slp_graph (MSG_NOTE, vect_location,
SLP_INSTANCE_TREE (instance), visited);
}
return opt_result::success ();
}
/* Estimates the cost of inserting layout changes into the SLP graph.
It can also say that the insertion is impossible. */
struct slpg_layout_cost
{
slpg_layout_cost () = default;
slpg_layout_cost (sreal, bool);
static slpg_layout_cost impossible () { return { sreal::max (), 0 }; }
bool is_possible () const { return depth != sreal::max (); }
bool operator== (const slpg_layout_cost &) const;
bool operator!= (const slpg_layout_cost &) const;
bool is_better_than (const slpg_layout_cost &, bool) const;
void add_parallel_cost (const slpg_layout_cost &);
void add_serial_cost (const slpg_layout_cost &);
void split (unsigned int);
/* The longest sequence of layout changes needed during any traversal
of the partition dag, weighted by execution frequency.
This is the most important metric when optimizing for speed, since
it helps to ensure that we keep the number of operations on
critical paths to a minimum. */
sreal depth = 0;
/* An estimate of the total number of operations needed. It is weighted by
execution frequency when optimizing for speed but not when optimizing for
size. In order to avoid double-counting, a node with a fanout of N will
distribute 1/N of its total cost to each successor.
This is the most important metric when optimizing for size, since
it helps to keep the total number of operations to a minimum, */
sreal total = 0;
};
/* Construct costs for a node with weight WEIGHT. A higher weight
indicates more frequent execution. IS_FOR_SIZE is true if we are
optimizing for size rather than speed. */
slpg_layout_cost::slpg_layout_cost (sreal weight, bool is_for_size)
: depth (weight), total (is_for_size && weight > 0 ? 1 : weight)
{
}
bool
slpg_layout_cost::operator== (const slpg_layout_cost &other) const
{
return depth == other.depth && total == other.total;
}
bool
slpg_layout_cost::operator!= (const slpg_layout_cost &other) const
{
return !operator== (other);
}
/* Return true if these costs are better than OTHER. IS_FOR_SIZE is
true if we are optimizing for size rather than speed. */
bool
slpg_layout_cost::is_better_than (const slpg_layout_cost &other,
bool is_for_size) const
{
if (is_for_size)
{
if (total != other.total)
return total < other.total;
return depth < other.depth;
}
else
{
if (depth != other.depth)
return depth < other.depth;
return total < other.total;
}
}
/* Increase the costs to account for something with cost INPUT_COST
happening in parallel with the current costs. */
void
slpg_layout_cost::add_parallel_cost (const slpg_layout_cost &input_cost)
{
depth = std::max (depth, input_cost.depth);
total += input_cost.total;
}
/* Increase the costs to account for something with cost INPUT_COST
happening in series with the current costs. */
void
slpg_layout_cost::add_serial_cost (const slpg_layout_cost &other)
{
depth += other.depth;
total += other.total;
}
/* Split the total cost among TIMES successors or predecessors. */
void
slpg_layout_cost::split (unsigned int times)
{
if (times > 1)
total /= times;
}
/* Information about one node in the SLP graph, for use during
vect_optimize_slp_pass. */
struct slpg_vertex
{
slpg_vertex (slp_tree node_) : node (node_) {}
/* The node itself. */
slp_tree node;
/* Which partition the node belongs to, or -1 if none. Nodes outside of
partitions are flexible; they can have whichever layout consumers
want them to have. */
int partition = -1;
/* The number of nodes that directly use the result of this one
(i.e. the number of nodes that count this one as a child). */
unsigned int out_degree = 0;
/* The execution frequency of the node. */
sreal weight = 0;
/* The total execution frequency of all nodes that directly use the
result of this one. */
sreal out_weight = 0;
};
/* Information about one partition of the SLP graph, for use during
vect_optimize_slp_pass. */
struct slpg_partition_info
{
/* The nodes in the partition occupy indices [NODE_BEGIN, NODE_END)
of m_partitioned_nodes. */
unsigned int node_begin = 0;
unsigned int node_end = 0;
/* Which layout we've chosen to use for this partition, or -1 if
we haven't picked one yet. */
int layout = -1;
/* The number of predecessors and successors in the partition dag.
The predecessors always have lower partition numbers and the
successors always have higher partition numbers.
Note that the directions of these edges are not necessarily the
same as in the data flow graph. For example, if an SCC has separate
partitions for an inner loop and an outer loop, the inner loop's
partition will have at least two incoming edges from the outer loop's
partition: one for a live-in value and one for a live-out value.
In data flow terms, one of these edges would also be from the outer loop
to the inner loop, but the other would be in the opposite direction. */
unsigned int in_degree = 0;
unsigned int out_degree = 0;
};
/* Information about the costs of using a particular layout for a
particular partition. It can also say that the combination is
impossible. */
struct slpg_partition_layout_costs
{
bool is_possible () const { return internal_cost.is_possible (); }
void mark_impossible () { internal_cost = slpg_layout_cost::impossible (); }
/* The costs inherited from predecessor partitions. */
slpg_layout_cost in_cost;
/* The inherent cost of the layout within the node itself. For example,
this is nonzero for a load if choosing a particular layout would require
the load to permute the loaded elements. It is nonzero for a
VEC_PERM_EXPR if the permutation cannot be eliminated or converted
to full-vector moves. */
slpg_layout_cost internal_cost;
/* The costs inherited from successor partitions. */
slpg_layout_cost out_cost;
};
/* This class tries to optimize the layout of vectors in order to avoid
unnecessary shuffling. At the moment, the set of possible layouts are
restricted to bijective permutations.
The goal of the pass depends on whether we're optimizing for size or
for speed. When optimizing for size, the goal is to reduce the overall
number of layout changes (including layout changes implied by things
like load permutations). When optimizing for speed, the goal is to
reduce the maximum latency attributable to layout changes on any
non-cyclical path through the data flow graph.
For example, when optimizing a loop nest for speed, we will prefer
to make layout changes outside of a loop rather than inside of a loop,
and will prefer to make layout changes in parallel rather than serially,
even if that increases the overall number of layout changes.
The high-level procedure is:
(1) Build a graph in which edges go from uses (parents) to definitions
(children).
(2) Divide the graph into a dag of strongly-connected components (SCCs).
(3) When optimizing for speed, partition the nodes in each SCC based
on their containing cfg loop. When optimizing for size, treat
each SCC as a single partition.
This gives us a dag of partitions. The goal is now to assign a
layout to each partition.
(4) Construct a set of vector layouts that are worth considering.
Record which nodes must keep their current layout.
(5) Perform a forward walk over the partition dag (from loads to stores)
accumulating the "forward" cost of using each layout. When visiting
each partition, assign a tentative choice of layout to the partition
and use that choice when calculating the cost of using a different
layout in successor partitions.
(6) Perform a backward walk over the partition dag (from stores to loads),
accumulating the "backward" cost of using each layout. When visiting
each partition, make a final choice of layout for that partition based
on the accumulated forward costs (from (5)) and backward costs
(from (6)).
(7) Apply the chosen layouts to the SLP graph.
For example, consider the SLP statements:
S1: a_1 = load
loop:
S2: a_2 = PHI<a_1, a_3>
S3: b_1 = load
S4: a_3 = a_2 + b_1
exit:
S5: a_4 = PHI<a_3>
S6: store a_4
S2 and S4 form an SCC and are part of the same loop. Every other
statement is in a singleton SCC. In this example there is a one-to-one
mapping between SCCs and partitions and the partition dag looks like this;
S1 S3
\ /
S2+S4
|
S5
|
S6
S2, S3 and S4 will have a higher execution frequency than the other
statements, so when optimizing for speed, the goal is to avoid any
layout changes:
- within S3
- within S2+S4
- on the S3->S2+S4 edge
For example, if S3 was originally a reversing load, the goal of the
pass is to make it an unreversed load and change the layout on the
S1->S2+S4 and S2+S4->S5 edges to compensate. (Changing the layout
on S1->S2+S4 and S5->S6 would also be acceptable.)
The difference between SCCs and partitions becomes important if we
add an outer loop:
S1: a_1 = ...
loop1:
S2: a_2 = PHI<a_1, a_6>
S3: b_1 = load
S4: a_3 = a_2 + b_1
loop2:
S5: a_4 = PHI<a_3, a_5>
S6: c_1 = load
S7: a_5 = a_4 + c_1
exit2:
S8: a_6 = PHI<a_5>
S9: store a_6
exit1:
Here, S2, S4, S5, S7 and S8 form a single SCC. However, when optimizing
for speed, we usually do not want restrictions in the outer loop to "infect"
the decision for the inner loop. For example, if an outer-loop node
in the SCC contains a statement with a fixed layout, that should not
prevent the inner loop from using a different layout. Conversely,
the inner loop should not dictate a layout to the outer loop: if the
outer loop does a lot of computation, then it may not be efficient to
do all of that computation in the inner loop's preferred layout.
So when optimizing for speed, we partition the SCC into S2+S4+S8 (outer)
and S5+S7 (inner). We also try to arrange partitions so that:
- the partition for an outer loop comes before the partition for
an inner loop
- if a sibling loop A dominates a sibling loop B, A's partition
comes before B's
This gives the following partition dag for the example above:
S1 S3
\ /
S2+S4+S8 S6
| \\ /
| S5+S7
|
S9
There are two edges from S2+S4+S8 to S5+S7: one for the edge S4->S5 and
one for a reversal of the edge S7->S8.
The backward walk picks a layout for S5+S7 before S2+S4+S8. The choice
for S2+S4+S8 therefore has to balance the cost of using the outer loop's
preferred layout against the cost of changing the layout on entry to the
inner loop (S4->S5) and on exit from the inner loop (S7->S8 reversed).
Although this works well when optimizing for speed, it has the downside
when optimizing for size that the choice of layout for S5+S7 is completely
independent of S9, which lessens the chance of reducing the overall number
of permutations. We therefore do not partition SCCs when optimizing
for size.
To give a concrete example of the difference between optimizing
for size and speed, consider:
a[0] = (b[1] << c[3]) - d[1];
a[1] = (b[0] << c[2]) - d[0];
a[2] = (b[3] << c[1]) - d[3];
a[3] = (b[2] << c[0]) - d[2];
There are three different layouts here: one for a, one for b and d,
and one for c. When optimizing for speed it is better to permute each
of b, c and d into the order required by a, since those permutations
happen in parallel. But when optimizing for size, it is better to:
- permute c into the same order as b
- do the arithmetic
- permute the result into the order required by a
This gives 2 permutations rather than 3. */
class vect_optimize_slp_pass
{
public:
vect_optimize_slp_pass (vec_info *vinfo) : m_vinfo (vinfo) {}
void run ();
private:
/* Graph building. */
struct loop *containing_loop (slp_tree);
bool is_cfg_latch_edge (graph_edge *);
void build_vertices (hash_set<slp_tree> &, slp_tree);
void build_vertices ();
void build_graph ();
/* Partitioning. */
void create_partitions ();
template<typename T> void for_each_partition_edge (unsigned int, T);
/* Layout selection. */
bool is_compatible_layout (slp_tree, unsigned int);
int change_layout_cost (slp_tree, unsigned int, unsigned int);
slpg_partition_layout_costs &partition_layout_costs (unsigned int,
unsigned int);
void change_vec_perm_layout (slp_tree, lane_permutation_t &,
int, unsigned int);
int internal_node_cost (slp_tree, int, unsigned int);
void start_choosing_layouts ();
/* Cost propagation. */
slpg_layout_cost edge_layout_cost (graph_edge *, unsigned int,
unsigned int, unsigned int);
slpg_layout_cost total_in_cost (unsigned int);
slpg_layout_cost forward_cost (graph_edge *, unsigned int, unsigned int);
slpg_layout_cost backward_cost (graph_edge *, unsigned int, unsigned int);
void forward_pass ();
void backward_pass ();
/* Rematerialization. */
slp_tree get_result_with_layout (slp_tree, unsigned int);
void materialize ();
/* Clean-up. */
void remove_redundant_permutations ();
void dump ();
vec_info *m_vinfo;
/* True if we should optimize the graph for size, false if we should
optimize it for speed. (It wouldn't be easy to make this decision
more locally.) */
bool m_optimize_size;
/* A graph of all SLP nodes, with edges leading from uses to definitions.
In other words, a node's predecessors are its slp_tree parents and
a node's successors are its slp_tree children. */
graph *m_slpg = nullptr;
/* The vertices of M_SLPG, indexed by slp_tree::vertex. */
auto_vec<slpg_vertex> m_vertices;
/* The list of all leaves of M_SLPG. such as external definitions, constants,
and loads. */
auto_vec<int> m_leafs;
/* This array has one entry for every vector layout that we're considering.
Element 0 is null and indicates "no change". Other entries describe
permutations that are inherent in the current graph and that we would
like to reverse if possible.
For example, a permutation { 1, 2, 3, 0 } means that something has
effectively been permuted in that way, such as a load group
{ a[1], a[2], a[3], a[0] } (viewed as a permutation of a[0:3]).
We'd then like to apply the reverse permutation { 3, 0, 1, 2 }
in order to put things "back" in order. */
auto_vec<vec<unsigned> > m_perms;
/* A partitioning of the nodes for which a layout must be chosen.
Each partition represents an <SCC, cfg loop> pair; that is,
nodes in different SCCs belong to different partitions, and nodes
within an SCC can be further partitioned according to a containing
cfg loop. Partition <SCC1, L1> comes before <SCC2, L2> if:
- SCC1 != SCC2 and SCC1 is a predecessor of SCC2 in a forward walk
from leaves (such as loads) to roots (such as stores).
- SCC1 == SCC2 and L1's header strictly dominates L2's header. */
auto_vec<slpg_partition_info> m_partitions;
/* The list of all nodes for which a layout must be chosen. Nodes for
partition P come before the nodes for partition P+1. Nodes within a
partition are in reverse postorder. */
auto_vec<unsigned int> m_partitioned_nodes;
/* Index P * num-layouts + L contains the cost of using layout L
for partition P. */
auto_vec<slpg_partition_layout_costs> m_partition_layout_costs;
/* Index N * num-layouts + L, if nonnull, is a node that provides the
original output of node N adjusted to have layout L. */
auto_vec<slp_tree> m_node_layouts;
};
/* Fill the vertices and leafs vector with all nodes in the SLP graph.
Also record whether we should optimize anything for speed rather
than size. */
void
vect_optimize_slp_pass::build_vertices (hash_set<slp_tree> &visited,
slp_tree node)
{
unsigned i;
slp_tree child;
if (visited.add (node))
return;
if (stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node))
{
basic_block bb = gimple_bb (vect_orig_stmt (rep)->stmt);
if (optimize_bb_for_speed_p (bb))
m_optimize_size = false;
}
node->vertex = m_vertices.length ();
m_vertices.safe_push (slpg_vertex (node));
bool leaf = true;
bool force_leaf = false;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
{
leaf = false;
build_vertices (visited, child);
}
else
force_leaf = true;
/* Since SLP discovery works along use-def edges all cycles have an
entry - but there's the exception of cycles where we do not handle
the entry explicitely (but with a NULL SLP node), like some reductions
and inductions. Force those SLP PHIs to act as leafs to make them
backwards reachable. */
if (leaf || force_leaf)
m_leafs.safe_push (node->vertex);
}
/* Fill the vertices and leafs vector with all nodes in the SLP graph. */
void
vect_optimize_slp_pass::build_vertices ()
{
hash_set<slp_tree> visited;
unsigned i;
slp_instance instance;
FOR_EACH_VEC_ELT (m_vinfo->slp_instances, i, instance)
build_vertices (visited, SLP_INSTANCE_TREE (instance));
}
/* Apply (reverse) bijectite PERM to VEC. */
template <class T>
static void
vect_slp_permute (vec<unsigned> perm,
vec<T> &vec, bool reverse)
{
auto_vec<T, 64> saved;
saved.create (vec.length ());
for (unsigned i = 0; i < vec.length (); ++i)
saved.quick_push (vec[i]);
if (reverse)
{
for (unsigned i = 0; i < vec.length (); ++i)
vec[perm[i]] = saved[i];
for (unsigned i = 0; i < vec.length (); ++i)
gcc_assert (vec[perm[i]] == saved[i]);
}
else
{
for (unsigned i = 0; i < vec.length (); ++i)
vec[i] = saved[perm[i]];
for (unsigned i = 0; i < vec.length (); ++i)
gcc_assert (vec[i] == saved[perm[i]]);
}
}
/* Return the cfg loop that contains NODE. */
struct loop *
vect_optimize_slp_pass::containing_loop (slp_tree node)
{
stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node);
if (!rep)
return ENTRY_BLOCK_PTR_FOR_FN (cfun)->loop_father;
return gimple_bb (vect_orig_stmt (rep)->stmt)->loop_father;
}
/* Return true if UD (an edge from a use to a definition) is associated
with a loop latch edge in the cfg. */
bool
vect_optimize_slp_pass::is_cfg_latch_edge (graph_edge *ud)
{
slp_tree use = m_vertices[ud->src].node;
slp_tree def = m_vertices[ud->dest].node;
if (SLP_TREE_DEF_TYPE (use) != vect_internal_def
|| SLP_TREE_DEF_TYPE (def) != vect_internal_def)
return false;
stmt_vec_info use_rep = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (use));
return (is_a<gphi *> (use_rep->stmt)
&& bb_loop_header_p (gimple_bb (use_rep->stmt))
&& containing_loop (def) == containing_loop (use));
}
/* Build the graph. Mark edges that correspond to cfg loop latch edges with
a nonnull data field. */
void
vect_optimize_slp_pass::build_graph ()
{
m_optimize_size = true;
build_vertices ();
m_slpg = new_graph (m_vertices.length ());
for (slpg_vertex &v : m_vertices)
for (slp_tree child : SLP_TREE_CHILDREN (v.node))
if (child)
{
graph_edge *ud = add_edge (m_slpg, v.node->vertex, child->vertex);
if (is_cfg_latch_edge (ud))
ud->data = this;
}
}
/* Return true if E corresponds to a loop latch edge in the cfg. */
static bool
skip_cfg_latch_edges (graph_edge *e)
{
return e->data;
}
/* Create the node partitions. */
void
vect_optimize_slp_pass::create_partitions ()
{
/* Calculate a postorder of the graph, ignoring edges that correspond
to natural latch edges in the cfg. Reading the vector from the end
to the beginning gives the reverse postorder. */
auto_vec<int> initial_rpo;
graphds_dfs (m_slpg, &m_leafs[0], m_leafs.length (), &initial_rpo,
false, NULL, skip_cfg_latch_edges);
gcc_assert (initial_rpo.length () == m_vertices.length ());
/* Calculate the strongly connected components of the graph. */
auto_vec<int> scc_grouping;
unsigned int num_sccs = graphds_scc (m_slpg, NULL, NULL, &scc_grouping);
/* Create a new index order in which all nodes from the same SCC are
consecutive. Use scc_pos to record the index of the first node in
each SCC. */
auto_vec<unsigned int> scc_pos (num_sccs);
int last_component = -1;
unsigned int node_count = 0;
for (unsigned int node_i : scc_grouping)
{
if (last_component != m_slpg->vertices[node_i].component)
{
last_component = m_slpg->vertices[node_i].component;
gcc_assert (last_component == int (scc_pos.length ()));
scc_pos.quick_push (node_count);
}
node_count += 1;
}
gcc_assert (node_count == initial_rpo.length ()
&& last_component + 1 == int (num_sccs));
/* Use m_partitioned_nodes to group nodes into SCC order, with the nodes
inside each SCC following the RPO we calculated above. The fact that
we ignored natural latch edges when calculating the RPO should ensure
that, for natural loop nests:
- the first node that we encounter in a cfg loop is the loop header phi
- the loop header phis are in dominance order
Arranging for this is an optimization (see below) rather than a
correctness issue. Unnatural loops with a tangled mess of backedges
will still work correctly, but might give poorer results.
Also update scc_pos so that it gives 1 + the index of the last node
in the SCC. */
m_partitioned_nodes.safe_grow (node_count);
for (unsigned int old_i = initial_rpo.length (); old_i-- > 0;)
{
unsigned int node_i = initial_rpo[old_i];
unsigned int new_i = scc_pos[m_slpg->vertices[node_i].component]++;
m_partitioned_nodes[new_i] = node_i;
}
/* When optimizing for speed, partition each SCC based on the containing
cfg loop. The order we constructed above should ensure that, for natural
cfg loops, we'll create sub-SCC partitions for outer loops before
the corresponding sub-SCC partitions for inner loops. Similarly,
when one sibling loop A dominates another sibling loop B, we should
create a sub-SCC partition for A before a sub-SCC partition for B.
As above, nothing depends for correctness on whether this achieves
a natural nesting, but we should get better results when it does. */
m_partitions.reserve (m_vertices.length ());
unsigned int next_partition_i = 0;
hash_map<struct loop *, int> loop_partitions;
unsigned int rpo_begin = 0;
unsigned int num_partitioned_nodes = 0;
for (unsigned int rpo_end : scc_pos)
{
loop_partitions.empty ();
unsigned int partition_i = next_partition_i;
for (unsigned int rpo_i = rpo_begin; rpo_i < rpo_end; ++rpo_i)
{
/* Handle externals and constants optimistically throughout.
But treat existing vectors as fixed since we do not handle
permuting them. */
unsigned int node_i = m_partitioned_nodes[rpo_i];
auto &vertex = m_vertices[node_i];
if ((SLP_TREE_DEF_TYPE (vertex.node) == vect_external_def
&& !SLP_TREE_VEC_DEFS (vertex.node).exists ())
|| SLP_TREE_DEF_TYPE (vertex.node) == vect_constant_def)
vertex.partition = -1;
else
{
bool existed;
if (m_optimize_size)
existed = next_partition_i > partition_i;
else
{
struct loop *loop = containing_loop (vertex.node);
auto &entry = loop_partitions.get_or_insert (loop, &existed);
if (!existed)
entry = next_partition_i;
partition_i = entry;
}
if (!existed)
{
m_partitions.quick_push (slpg_partition_info ());
next_partition_i += 1;
}
vertex.partition = partition_i;
num_partitioned_nodes += 1;
m_partitions[partition_i].node_end += 1;
}
}
rpo_begin = rpo_end;
}
/* Assign ranges of consecutive node indices to each partition,
in partition order. Start with node_end being the same as
node_begin so that the next loop can use it as a counter. */
unsigned int node_begin = 0;
for (auto &partition : m_partitions)
{
partition.node_begin = node_begin;
node_begin += partition.node_end;
partition.node_end = partition.node_begin;
}
gcc_assert (node_begin == num_partitioned_nodes);
/* Finally build the list of nodes in partition order. */
m_partitioned_nodes.truncate (num_partitioned_nodes);
for (unsigned int node_i = 0; node_i < m_vertices.length (); ++node_i)
{
int partition_i = m_vertices[node_i].partition;
if (partition_i >= 0)
{
unsigned int order_i = m_partitions[partition_i].node_end++;
m_partitioned_nodes[order_i] = node_i;
}
}
}
/* Look for edges from earlier partitions into node NODE_I and edges from
node NODE_I into later partitions. Call:
FN (ud, other_node_i)
for each such use-to-def edge ud, where other_node_i is the node at the
other end of the edge. */
template<typename T>
void
vect_optimize_slp_pass::for_each_partition_edge (unsigned int node_i, T fn)
{
int partition_i = m_vertices[node_i].partition;
for (graph_edge *pred = m_slpg->vertices[node_i].pred;
pred; pred = pred->pred_next)
{
int src_partition_i = m_vertices[pred->src].partition;
if (src_partition_i >= 0 && src_partition_i != partition_i)
fn (pred, pred->src);
}
for (graph_edge *succ = m_slpg->vertices[node_i].succ;
succ; succ = succ->succ_next)
{
int dest_partition_i = m_vertices[succ->dest].partition;
if (dest_partition_i >= 0 && dest_partition_i != partition_i)
fn (succ, succ->dest);
}
}
/* Return true if layout LAYOUT_I is compatible with the number of SLP lanes
that NODE would operate on. This test is independent of NODE's actual
operation. */
bool
vect_optimize_slp_pass::is_compatible_layout (slp_tree node,
unsigned int layout_i)
{
if (layout_i == 0)
return true;
if (SLP_TREE_LANES (node) != m_perms[layout_i].length ())
return false;
return true;
}
/* Return the cost (in arbtirary units) of going from layout FROM_LAYOUT_I
to layout TO_LAYOUT_I for a node like NODE. Return -1 if either of the
layouts is incompatible with NODE or if the change is not possible for
some other reason.
The properties taken from NODE include the number of lanes and the
vector type. The actual operation doesn't matter. */
int
vect_optimize_slp_pass::change_layout_cost (slp_tree node,
unsigned int from_layout_i,
unsigned int to_layout_i)
{
if (!is_compatible_layout (node, from_layout_i)
|| !is_compatible_layout (node, to_layout_i))
return -1;
if (from_layout_i == to_layout_i)
return 0;
auto_vec<slp_tree, 1> children (1);
children.quick_push (node);
auto_lane_permutation_t perm (SLP_TREE_LANES (node));
if (from_layout_i > 0)
for (unsigned int i : m_perms[from_layout_i])
perm.quick_push ({ 0, i });
else
for (unsigned int i = 0; i < SLP_TREE_LANES (node); ++i)
perm.quick_push ({ 0, i });
if (to_layout_i > 0)
vect_slp_permute (m_perms[to_layout_i], perm, true);
auto count = vectorizable_slp_permutation_1 (m_vinfo, nullptr, node, perm,
children, false);
if (count >= 0)
return MAX (count, 1);
/* ??? In principle we could try changing via layout 0, giving two
layout changes rather than 1. Doing that would require
corresponding support in get_result_with_layout. */
return -1;
}
/* Return the costs of assigning layout LAYOUT_I to partition PARTITION_I. */
inline slpg_partition_layout_costs &
vect_optimize_slp_pass::partition_layout_costs (unsigned int partition_i,
unsigned int layout_i)
{
return m_partition_layout_costs[partition_i * m_perms.length () + layout_i];
}
/* Change PERM in one of two ways:
- if IN_LAYOUT_I < 0, accept input operand I in the layout that has been
chosen for child I of NODE.
- if IN_LAYOUT >= 0, accept all inputs operands with that layout.
In both cases, arrange for the output to have layout OUT_LAYOUT_I */
void
vect_optimize_slp_pass::
change_vec_perm_layout (slp_tree node, lane_permutation_t &perm,
int in_layout_i, unsigned int out_layout_i)
{
for (auto &entry : perm)
{
int this_in_layout_i = in_layout_i;
if (this_in_layout_i < 0)
{
slp_tree in_node = SLP_TREE_CHILDREN (node)[entry.first];
unsigned int in_partition_i = m_vertices[in_node->vertex].partition;
this_in_layout_i = m_partitions[in_partition_i].layout;
}
if (this_in_layout_i > 0)
entry.second = m_perms[this_in_layout_i][entry.second];
}
if (out_layout_i > 0)
vect_slp_permute (m_perms[out_layout_i], perm, true);
}
/* Check whether the target allows NODE to be rearranged so that the node's
output has layout OUT_LAYOUT_I. Return the cost of the change if so,
in the same arbitrary units as for change_layout_cost. Return -1 otherwise.
If NODE is a VEC_PERM_EXPR and IN_LAYOUT_I < 0, also check whether
NODE can adapt to the layout changes that have (perhaps provisionally)
been chosen for NODE's children, so that no extra permutations are
needed on either the input or the output of NODE.
If NODE is a VEC_PERM_EXPR and IN_LAYOUT_I >= 0, instead assume
that all inputs will be forced into layout IN_LAYOUT_I beforehand.
IN_LAYOUT_I has no meaning for other types of node.
Keeping the node as-is is always valid. If the target doesn't appear to
support the node as-is then layout 0 has a high and arbitrary cost instead
of being invalid. On the one hand, this ensures that every node has at
least one valid layout, avoiding what would otherwise be an awkward
special case. On the other, it still encourages the pass to change
an invalid pre-existing layout choice into a valid one. */
int
vect_optimize_slp_pass::internal_node_cost (slp_tree node, int in_layout_i,
unsigned int out_layout_i)
{
const int fallback_cost = 100;
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
{
auto_lane_permutation_t tmp_perm;
tmp_perm.safe_splice (SLP_TREE_LANE_PERMUTATION (node));
/* Check that the child nodes support the chosen layout. Checking
the first child is enough, since any second child would have the
same shape. */
if (in_layout_i > 0
&& !is_compatible_layout (SLP_TREE_CHILDREN (node)[0], in_layout_i))
return -1;
change_vec_perm_layout (node, tmp_perm, in_layout_i, out_layout_i);
int count = vectorizable_slp_permutation_1 (m_vinfo, nullptr,
node, tmp_perm,
SLP_TREE_CHILDREN (node),
false);
if (count < 0)
{
if (in_layout_i == 0 && out_layout_i == 0)
return fallback_cost;
return -1;
}
/* We currently have no way of telling whether the new layout is cheaper
or more expensive than the old one. But at least in principle,
it should be worth making zero permutations (whole-vector shuffles)
cheaper than real permutations, in case the pass is able to remove
the latter. */
return count == 0 ? 0 : 1;
}
stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node);
if (rep
&& STMT_VINFO_DATA_REF (rep)
&& DR_IS_READ (STMT_VINFO_DATA_REF (rep)))
{
auto_load_permutation_t tmp_perm;
tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));
if (out_layout_i > 0)
vect_slp_permute (m_perms[out_layout_i], tmp_perm, true);
poly_uint64 vf = 1;
if (auto loop_vinfo = dyn_cast<loop_vec_info> (m_vinfo))
vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
unsigned int n_perms;
if (!vect_transform_slp_perm_load_1 (m_vinfo, node, tmp_perm, vNULL,
nullptr, vf, true, false, &n_perms))
{
if (out_layout_i == 0)
return fallback_cost;
return -1;
}
/* See the comment above the corresponding VEC_PERM_EXPR handling. */
return n_perms == 0 ? 0 : 1;
}
return 0;
}
/* Decide which element layouts we should consider using. Calculate the
weights associated with inserting layout changes on partition edges.
Also mark partitions that cannot change layout, by setting their
layout to zero. */
void
vect_optimize_slp_pass::start_choosing_layouts ()
{
/* Used to assign unique permutation indices. */
using perm_hash = unbounded_hashmap_traits<
vec_free_hash_base<int_hash_base<unsigned>>,
int_hash<int, -1, -2>
>;
hash_map<vec<unsigned>, int, perm_hash> layout_ids;
/* Layout 0 is "no change". */
m_perms.safe_push (vNULL);
/* Create layouts from existing permutations. */
auto_load_permutation_t tmp_perm;
for (unsigned int node_i : m_partitioned_nodes)
{
/* Leafs also double as entries to the reverse graph. Allow the
layout of those to be changed. */
auto &vertex = m_vertices[node_i];
auto &partition = m_partitions[vertex.partition];
if (!m_slpg->vertices[node_i].succ)
partition.layout = 0;
/* Loads and VEC_PERM_EXPRs are the only things generating permutes. */
slp_tree node = vertex.node;
stmt_vec_info dr_stmt = SLP_TREE_REPRESENTATIVE (node);
slp_tree child;
unsigned HOST_WIDE_INT imin, imax = 0;
bool any_permute = false;
tmp_perm.truncate (0);
if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
{
/* If splitting out a SLP_TREE_LANE_PERMUTATION can make the node
unpermuted, record a layout that reverses this permutation. */
gcc_assert (partition.layout == 0);
if (!STMT_VINFO_GROUPED_ACCESS (dr_stmt))
continue;
dr_stmt = DR_GROUP_FIRST_ELEMENT (dr_stmt);
imin = DR_GROUP_SIZE (dr_stmt) + 1;
tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));
}
else if (SLP_TREE_CODE (node) == VEC_PERM_EXPR
&& SLP_TREE_CHILDREN (node).length () == 1
&& (child = SLP_TREE_CHILDREN (node)[0])
&& (TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (child))
.is_constant (&imin)))
{
/* If the child has the same vector size as this node,
reversing the permutation can make the permutation a no-op.
In other cases it can change a true permutation into a
full-vector extract. */
tmp_perm.reserve (SLP_TREE_LANES (node));
for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
tmp_perm.quick_push (SLP_TREE_LANE_PERMUTATION (node)[j].second);
}
else
continue;
for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
{
unsigned idx = tmp_perm[j];
imin = MIN (imin, idx);
imax = MAX (imax, idx);
if (idx - tmp_perm[0] != j)
any_permute = true;
}
/* If the span doesn't match we'd disrupt VF computation, avoid
that for now. */
if (imax - imin + 1 != SLP_TREE_LANES (node))
continue;
/* If there's no permute no need to split one out. In this case
we can consider turning a load into a permuted load, if that
turns out to be cheaper than alternatives. */
if (!any_permute)
{
partition.layout = -1;
continue;
}
/* For now only handle true permutes, like
vect_attempt_slp_rearrange_stmts did. This allows us to be lazy
when permuting constants and invariants keeping the permute
bijective. */
auto_sbitmap load_index (SLP_TREE_LANES (node));
bitmap_clear (load_index);
for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
bitmap_set_bit (load_index, tmp_perm[j] - imin);
unsigned j;
for (j = 0; j < SLP_TREE_LANES (node); ++j)
if (!bitmap_bit_p (load_index, j))
break;
if (j != SLP_TREE_LANES (node))
continue;
vec<unsigned> perm = vNULL;
perm.safe_grow (SLP_TREE_LANES (node), true);
for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
perm[j] = tmp_perm[j] - imin;
if (int (m_perms.length ()) >= param_vect_max_layout_candidates)
{
/* Continue to use existing layouts, but don't add any more. */
int *entry = layout_ids.get (perm);
partition.layout = entry ? *entry : 0;
perm.release ();
}
else
{
bool existed;
int &layout_i = layout_ids.get_or_insert (perm, &existed);
if (existed)
perm.release ();
else
{
layout_i = m_perms.length ();
m_perms.safe_push (perm);
}
partition.layout = layout_i;
}
}
/* Initially assume that every layout is possible and has zero cost
in every partition. */
m_partition_layout_costs.safe_grow_cleared (m_partitions.length ()
* m_perms.length ());
/* We have to mark outgoing permutations facing non-reduction graph
entries that are not represented as to be materialized. */
for (slp_instance instance : m_vinfo->slp_instances)
if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_ctor)
{
unsigned int node_i = SLP_INSTANCE_TREE (instance)->vertex;
m_partitions[m_vertices[node_i].partition].layout = 0;
}
/* Check which layouts each node and partition can handle. Calculate the
weights associated with inserting layout changes on edges. */
for (unsigned int node_i : m_partitioned_nodes)
{
auto &vertex = m_vertices[node_i];
auto &partition = m_partitions[vertex.partition];
slp_tree node = vertex.node;
if (stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node))
{
vertex.weight = vect_slp_node_weight (node);
/* We do not handle stores with a permutation, so all
incoming permutations must have been materialized. */
if (STMT_VINFO_DATA_REF (rep)
&& DR_IS_WRITE (STMT_VINFO_DATA_REF (rep)))
/* ??? We're forcing materialization in place
of the child here, we'd need special handling
in materialization to leave layout -1 here. */
partition.layout = 0;
/* We cannot change the layout of an operation that is
not independent on lanes. Note this is an explicit
negative list since that's much shorter than the respective
positive one but it's critical to keep maintaining it. */
if (is_gimple_call (STMT_VINFO_STMT (rep)))
switch (gimple_call_combined_fn (STMT_VINFO_STMT (rep)))
{
case CFN_COMPLEX_ADD_ROT90:
case CFN_COMPLEX_ADD_ROT270:
case CFN_COMPLEX_MUL:
case CFN_COMPLEX_MUL_CONJ:
case CFN_VEC_ADDSUB:
case CFN_VEC_FMADDSUB:
case CFN_VEC_FMSUBADD:
partition.layout = 0;
default:;
}
}
auto process_edge = [&](graph_edge *ud, unsigned int other_node_i)
{
auto &other_vertex = m_vertices[other_node_i];
/* Count the number of edges from earlier partitions and the number
of edges to later partitions. */
if (other_vertex.partition < vertex.partition)
partition.in_degree += 1;
else
partition.out_degree += 1;
/* If the current node uses the result of OTHER_NODE_I, accumulate
the effects of that. */
if (ud->src == int (node_i))
{
other_vertex.out_weight += vertex.weight;
other_vertex.out_degree += 1;
}
};
for_each_partition_edge (node_i, process_edge);
}
}
/* Return the incoming costs for node NODE_I, assuming that each input keeps
its current (provisional) choice of layout. The inputs do not necessarily
have the same layout as each other. */
slpg_layout_cost
vect_optimize_slp_pass::total_in_cost (unsigned int node_i)
{
auto &vertex = m_vertices[node_i];
slpg_layout_cost cost;
auto add_cost = [&](graph_edge *, unsigned int other_node_i)
{
auto &other_vertex = m_vertices[other_node_i];
if (other_vertex.partition < vertex.partition)
{
auto &other_partition = m_partitions[other_vertex.partition];
auto &other_costs = partition_layout_costs (other_vertex.partition,
other_partition.layout);
slpg_layout_cost this_cost = other_costs.in_cost;
this_cost.add_serial_cost (other_costs.internal_cost);
this_cost.split (other_partition.out_degree);
cost.add_parallel_cost (this_cost);
}
};
for_each_partition_edge (node_i, add_cost);
return cost;
}
/* Return the cost of switching between layout LAYOUT1_I (at node NODE1_I)
and layout LAYOUT2_I on cross-partition use-to-def edge UD. Return
slpg_layout_cost::impossible () if the change isn't possible. */
slpg_layout_cost
vect_optimize_slp_pass::
edge_layout_cost (graph_edge *ud, unsigned int node1_i, unsigned int layout1_i,
unsigned int layout2_i)
{
auto &def_vertex = m_vertices[ud->dest];
auto &use_vertex = m_vertices[ud->src];
auto def_layout_i = ud->dest == int (node1_i) ? layout1_i : layout2_i;
auto use_layout_i = ud->dest == int (node1_i) ? layout2_i : layout1_i;
auto factor = change_layout_cost (def_vertex.node, def_layout_i,
use_layout_i);
if (factor < 0)
return slpg_layout_cost::impossible ();
/* We have a choice of putting the layout change at the site of the
definition or at the site of the use. Prefer the former when
optimizing for size or when the execution frequency of the
definition is no greater than the combined execution frequencies of
the uses. When putting the layout change at the site of the definition,
divvy up the cost among all consumers. */
if (m_optimize_size || def_vertex.weight <= def_vertex.out_weight)
{
slpg_layout_cost cost = { def_vertex.weight * factor, m_optimize_size };
cost.split (def_vertex.out_degree);
return cost;
}
return { use_vertex.weight * factor, m_optimize_size };
}
/* UD represents a use-def link between FROM_NODE_I and a node in a later
partition; FROM_NODE_I could be the definition node or the use node.
The node at the other end of the link wants to use layout TO_LAYOUT_I.
Return the cost of any necessary fix-ups on edge UD, or return
slpg_layout_cost::impossible () if the change isn't possible.
At this point, FROM_NODE_I's partition has chosen the cheapest
layout based on the information available so far, but this choice
is only provisional. */
slpg_layout_cost
vect_optimize_slp_pass::forward_cost (graph_edge *ud, unsigned int from_node_i,
unsigned int to_layout_i)
{
auto &from_vertex = m_vertices[from_node_i];
unsigned int from_partition_i = from_vertex.partition;
slpg_partition_info &from_partition = m_partitions[from_partition_i];
gcc_assert (from_partition.layout >= 0);
/* First calculate the cost on the assumption that FROM_PARTITION sticks
with its current layout preference. */
slpg_layout_cost cost = slpg_layout_cost::impossible ();
auto edge_cost = edge_layout_cost (ud, from_node_i,
from_partition.layout, to_layout_i);
if (edge_cost.is_possible ())
{
auto &from_costs = partition_layout_costs (from_partition_i,
from_partition.layout);
cost = from_costs.in_cost;
cost.add_serial_cost (from_costs.internal_cost);
cost.split (from_partition.out_degree);
cost.add_serial_cost (edge_cost);
}
/* Take the minimum of that cost and the cost that applies if
FROM_PARTITION instead switches to TO_LAYOUT_I. */
auto &direct_layout_costs = partition_layout_costs (from_partition_i,
to_layout_i);
if (direct_layout_costs.is_possible ())
{
slpg_layout_cost direct_cost = direct_layout_costs.in_cost;
direct_cost.add_serial_cost (direct_layout_costs.internal_cost);
direct_cost.split (from_partition.out_degree);
if (!cost.is_possible ()
|| direct_cost.is_better_than (cost, m_optimize_size))
cost = direct_cost;
}
return cost;
}
/* UD represents a use-def link between TO_NODE_I and a node in an earlier
partition; TO_NODE_I could be the definition node or the use node.
The node at the other end of the link wants to use layout FROM_LAYOUT_I;
return the cost of any necessary fix-ups on edge UD, or
slpg_layout_cost::impossible () if the choice cannot be made.
At this point, TO_NODE_I's partition has a fixed choice of layout. */
slpg_layout_cost
vect_optimize_slp_pass::backward_cost (graph_edge *ud, unsigned int to_node_i,
unsigned int from_layout_i)
{
auto &to_vertex = m_vertices[to_node_i];
unsigned int to_partition_i = to_vertex.partition;
slpg_partition_info &to_partition = m_partitions[to_partition_i];
gcc_assert (to_partition.layout >= 0);
/* If TO_NODE_I is a VEC_PERM_EXPR consumer, see whether it can be
adjusted for this input having layout FROM_LAYOUT_I. Assume that
any other inputs keep their current choice of layout. */
auto &to_costs = partition_layout_costs (to_partition_i,
to_partition.layout);
if (ud->src == int (to_node_i)
&& SLP_TREE_CODE (to_vertex.node) == VEC_PERM_EXPR)
{
auto &from_partition = m_partitions[m_vertices[ud->dest].partition];
auto old_layout = from_partition.layout;
from_partition.layout = from_layout_i;
int factor = internal_node_cost (to_vertex.node, -1,
to_partition.layout);
from_partition.layout = old_layout;
if (factor >= 0)
{
slpg_layout_cost cost = to_costs.out_cost;
cost.add_serial_cost ({ to_vertex.weight * factor,
m_optimize_size });
cost.split (to_partition.in_degree);
return cost;
}
}
/* Compute the cost if we insert any necessary layout change on edge UD. */
auto edge_cost = edge_layout_cost (ud, to_node_i,
to_partition.layout, from_layout_i);
if (edge_cost.is_possible ())
{
slpg_layout_cost cost = to_costs.out_cost;
cost.add_serial_cost (to_costs.internal_cost);
cost.split (to_partition.in_degree);
cost.add_serial_cost (edge_cost);
return cost;
}
return slpg_layout_cost::impossible ();
}
/* Make a forward pass through the partitions, accumulating input costs.
Make a tentative (provisional) choice of layout for each partition,
ensuring that this choice still allows later partitions to keep
their original layout. */
void
vect_optimize_slp_pass::forward_pass ()
{
for (unsigned int partition_i = 0; partition_i < m_partitions.length ();
++partition_i)
{
auto &partition = m_partitions[partition_i];
/* If the partition consists of a single VEC_PERM_EXPR, precompute
the incoming cost that would apply if every predecessor partition
keeps its current layout. This is used within the loop below. */
slpg_layout_cost in_cost;
slp_tree single_node = nullptr;
if (partition.node_end == partition.node_begin + 1)
{
unsigned int node_i = m_partitioned_nodes[partition.node_begin];
single_node = m_vertices[node_i].node;
if (SLP_TREE_CODE (single_node) == VEC_PERM_EXPR)
in_cost = total_in_cost (node_i);
}
/* Go through the possible layouts. Decide which ones are valid
for this partition and record which of the valid layouts has
the lowest cost. */
unsigned int min_layout_i = 0;
slpg_layout_cost min_layout_cost = slpg_layout_cost::impossible ();
for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
{
auto &layout_costs = partition_layout_costs (partition_i, layout_i);
if (!layout_costs.is_possible ())
continue;
/* If the recorded layout is already 0 then the layout cannot
change. */
if (partition.layout == 0 && layout_i != 0)
{
layout_costs.mark_impossible ();
continue;
}
bool is_possible = true;
for (unsigned int order_i = partition.node_begin;
order_i < partition.node_end; ++order_i)
{
unsigned int node_i = m_partitioned_nodes[order_i];
auto &vertex = m_vertices[node_i];
/* Reject the layout if it is individually incompatible
with any node in the partition. */
if (!is_compatible_layout (vertex.node, layout_i))
{
is_possible = false;
break;
}
auto add_cost = [&](graph_edge *ud, unsigned int other_node_i)
{
auto &other_vertex = m_vertices[other_node_i];
if (other_vertex.partition < vertex.partition)
{
/* Accumulate the incoming costs from earlier
partitions, plus the cost of any layout changes
on UD itself. */
auto cost = forward_cost (ud, other_node_i, layout_i);
if (!cost.is_possible ())
is_possible = false;
else
layout_costs.in_cost.add_parallel_cost (cost);
}
else
/* Reject the layout if it would make layout 0 impossible
for later partitions. This amounts to testing that the
target supports reversing the layout change on edges
to later partitions.
In principle, it might be possible to push a layout
change all the way down a graph, so that it never
needs to be reversed and so that the target doesn't
need to support the reverse operation. But it would
be awkward to bail out if we hit a partition that
does not support the new layout, especially since
we are not dealing with a lattice. */
is_possible &= edge_layout_cost (ud, other_node_i, 0,
layout_i).is_possible ();
};
for_each_partition_edge (node_i, add_cost);
/* Accumulate the cost of using LAYOUT_I within NODE,
both for the inputs and the outputs. */
int factor = internal_node_cost (vertex.node, layout_i,
layout_i);
if (factor < 0)
{
is_possible = false;
break;
}
else if (factor)
layout_costs.internal_cost.add_serial_cost
({ vertex.weight * factor, m_optimize_size });
}
if (!is_possible)
{
layout_costs.mark_impossible ();
continue;
}
/* Combine the incoming and partition-internal costs. */
slpg_layout_cost combined_cost = layout_costs.in_cost;
combined_cost.add_serial_cost (layout_costs.internal_cost);
/* If this partition consists of a single VEC_PERM_EXPR, see
if the VEC_PERM_EXPR can be changed to support output layout
LAYOUT_I while keeping all the provisional choices of input
layout. */
if (single_node
&& SLP_TREE_CODE (single_node) == VEC_PERM_EXPR)
{
int factor = internal_node_cost (single_node, -1, layout_i);
if (factor >= 0)
{
auto weight = m_vertices[single_node->vertex].weight;
slpg_layout_cost internal_cost
= { weight * factor, m_optimize_size };
slpg_layout_cost alt_cost = in_cost;
alt_cost.add_serial_cost (internal_cost);
if (alt_cost.is_better_than (combined_cost, m_optimize_size))
{
combined_cost = alt_cost;
layout_costs.in_cost = in_cost;
layout_costs.internal_cost = internal_cost;
}
}
}
/* Record the layout with the lowest cost. Prefer layout 0 in
the event of a tie between it and another layout. */
if (!min_layout_cost.is_possible ()
|| combined_cost.is_better_than (min_layout_cost,
m_optimize_size))
{
min_layout_i = layout_i;
min_layout_cost = combined_cost;
}
}
/* This loop's handling of earlier partitions should ensure that
choosing the original layout for the current partition is no
less valid than it was in the original graph, even with the
provisional layout choices for those earlier partitions. */
gcc_assert (min_layout_cost.is_possible ());
partition.layout = min_layout_i;
}
}
/* Make a backward pass through the partitions, accumulating output costs.
Make a final choice of layout for each partition. */
void
vect_optimize_slp_pass::backward_pass ()
{
for (unsigned int partition_i = m_partitions.length (); partition_i-- > 0;)
{
auto &partition = m_partitions[partition_i];
unsigned int min_layout_i = 0;
slpg_layout_cost min_layout_cost = slpg_layout_cost::impossible ();
for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
{
auto &layout_costs = partition_layout_costs (partition_i, layout_i);
if (!layout_costs.is_possible ())
continue;
/* Accumulate the costs from successor partitions. */
bool is_possible = true;
for (unsigned int order_i = partition.node_begin;
order_i < partition.node_end; ++order_i)
{
unsigned int node_i = m_partitioned_nodes[order_i];
auto &vertex = m_vertices[node_i];
auto add_cost = [&](graph_edge *ud, unsigned int other_node_i)
{
auto &other_vertex = m_vertices[other_node_i];
auto &other_partition = m_partitions[other_vertex.partition];
if (other_vertex.partition > vertex.partition)
{
/* Accumulate the incoming costs from later
partitions, plus the cost of any layout changes
on UD itself. */
auto cost = backward_cost (ud, other_node_i, layout_i);
if (!cost.is_possible ())
is_possible = false;
else
layout_costs.out_cost.add_parallel_cost (cost);
}
else
/* Make sure that earlier partitions can (if necessary
or beneficial) keep the layout that they chose in
the forward pass. This ensures that there is at
least one valid choice of layout. */
is_possible &= edge_layout_cost (ud, other_node_i,
other_partition.layout,
layout_i).is_possible ();
};
for_each_partition_edge (node_i, add_cost);
}
if (!is_possible)
{
layout_costs.mark_impossible ();
continue;
}
/* Locally combine the costs from the forward and backward passes.
(This combined cost is not passed on, since that would lead
to double counting.) */
slpg_layout_cost combined_cost = layout_costs.in_cost;
combined_cost.add_serial_cost (layout_costs.internal_cost);
combined_cost.add_serial_cost (layout_costs.out_cost);
/* Record the layout with the lowest cost. Prefer layout 0 in
the event of a tie between it and another layout. */
if (!min_layout_cost.is_possible ()
|| combined_cost.is_better_than (min_layout_cost,
m_optimize_size))
{
min_layout_i = layout_i;
min_layout_cost = combined_cost;
}
}
gcc_assert (min_layout_cost.is_possible ());
partition.layout = min_layout_i;
}
}
/* Return a node that applies layout TO_LAYOUT_I to the original form of NODE.
NODE already has the layout that was selected for its partition. */
slp_tree
vect_optimize_slp_pass::get_result_with_layout (slp_tree node,
unsigned int to_layout_i)
{
unsigned int result_i = node->vertex * m_perms.length () + to_layout_i;
slp_tree result = m_node_layouts[result_i];
if (result)
return result;
if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
|| SLP_TREE_DEF_TYPE (node) == vect_external_def)
{
/* If the vector is uniform or unchanged, there's nothing to do. */
if (to_layout_i == 0 || vect_slp_tree_uniform_p (node))
result = node;
else
{
auto scalar_ops = SLP_TREE_SCALAR_OPS (node).copy ();
result = vect_create_new_slp_node (scalar_ops);
vect_slp_permute (m_perms[to_layout_i], scalar_ops, true);
}
}
else
{
unsigned int partition_i = m_vertices[node->vertex].partition;
unsigned int from_layout_i = m_partitions[partition_i].layout;
if (from_layout_i == to_layout_i)
return node;
/* If NODE is itself a VEC_PERM_EXPR, try to create a parallel
permutation instead of a serial one. Leave the new permutation
in TMP_PERM on success. */
auto_lane_permutation_t tmp_perm;
unsigned int num_inputs = 1;
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
{
tmp_perm.safe_splice (SLP_TREE_LANE_PERMUTATION (node));
if (from_layout_i != 0)
vect_slp_permute (m_perms[from_layout_i], tmp_perm, false);
if (to_layout_i != 0)
vect_slp_permute (m_perms[to_layout_i], tmp_perm, true);
if (vectorizable_slp_permutation_1 (m_vinfo, nullptr, node,
tmp_perm,
SLP_TREE_CHILDREN (node),
false) >= 0)
num_inputs = SLP_TREE_CHILDREN (node).length ();
else
tmp_perm.truncate (0);
}
if (dump_enabled_p ())
{
if (tmp_perm.length () > 0)
dump_printf_loc (MSG_NOTE, vect_location,
"duplicating permutation node %p with"
" layout %d\n",
(void *) node, to_layout_i);
else
dump_printf_loc (MSG_NOTE, vect_location,
"inserting permutation node in place of %p\n",
(void *) node);
}
unsigned int num_lanes = SLP_TREE_LANES (node);
result = vect_create_new_slp_node (num_inputs, VEC_PERM_EXPR);
if (SLP_TREE_SCALAR_STMTS (node).length ())
{
auto &stmts = SLP_TREE_SCALAR_STMTS (result);
stmts.safe_splice (SLP_TREE_SCALAR_STMTS (result));
if (from_layout_i != 0)
vect_slp_permute (m_perms[from_layout_i], stmts, false);
if (to_layout_i != 0)
vect_slp_permute (m_perms[to_layout_i], stmts, true);
}
SLP_TREE_REPRESENTATIVE (result) = SLP_TREE_REPRESENTATIVE (node);
SLP_TREE_LANES (result) = num_lanes;
SLP_TREE_VECTYPE (result) = SLP_TREE_VECTYPE (node);
result->vertex = -1;
auto &lane_perm = SLP_TREE_LANE_PERMUTATION (result);
if (tmp_perm.length ())
{
lane_perm.safe_splice (tmp_perm);
SLP_TREE_CHILDREN (result).safe_splice (SLP_TREE_CHILDREN (node));
}
else
{
lane_perm.create (num_lanes);
for (unsigned j = 0; j < num_lanes; ++j)
lane_perm.quick_push ({ 0, j });
if (from_layout_i != 0)
vect_slp_permute (m_perms[from_layout_i], lane_perm, false);
if (to_layout_i != 0)
vect_slp_permute (m_perms[to_layout_i], lane_perm, true);
SLP_TREE_CHILDREN (result).safe_push (node);
}
for (slp_tree child : SLP_TREE_CHILDREN (result))
child->refcnt++;
}
m_node_layouts[result_i] = result;
return result;
}
/* Apply the chosen vector layouts to the SLP graph. */
void
vect_optimize_slp_pass::materialize ()
{
/* We no longer need the costs, so avoid having two O(N * P) arrays
live at the same time. */
m_partition_layout_costs.release ();
m_node_layouts.safe_grow_cleared (m_vertices.length () * m_perms.length ());
auto_sbitmap fully_folded (m_vertices.length ());
bitmap_clear (fully_folded);
for (unsigned int node_i : m_partitioned_nodes)
{
auto &vertex = m_vertices[node_i];
slp_tree node = vertex.node;
int layout_i = m_partitions[vertex.partition].layout;
gcc_assert (layout_i >= 0);
/* Rearrange the scalar statements to match the chosen layout. */
if (layout_i > 0)
vect_slp_permute (m_perms[layout_i],
SLP_TREE_SCALAR_STMTS (node), true);
/* Update load and lane permutations. */
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
{
/* First try to absorb the input vector layouts. If that fails,
force the inputs to have layout LAYOUT_I too. We checked that
that was possible before deciding to use nonzero output layouts.
(Note that at this stage we don't really have any guarantee that
the target supports the original VEC_PERM_EXPR.) */
auto &perm = SLP_TREE_LANE_PERMUTATION (node);
auto_lane_permutation_t tmp_perm;
tmp_perm.safe_splice (perm);
change_vec_perm_layout (node, tmp_perm, -1, layout_i);
if (vectorizable_slp_permutation_1 (m_vinfo, nullptr, node,
tmp_perm,
SLP_TREE_CHILDREN (node),
false) >= 0)
{
if (dump_enabled_p ()
&& !std::equal (tmp_perm.begin (), tmp_perm.end (),
perm.begin ()))
dump_printf_loc (MSG_NOTE, vect_location,
"absorbing input layouts into %p\n",
(void *) node);
std::copy (tmp_perm.begin (), tmp_perm.end (), perm.begin ());
bitmap_set_bit (fully_folded, node_i);
}
else
{
/* Not MSG_MISSED because it would make no sense to users. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"failed to absorb input layouts into %p\n",
(void *) node);
change_vec_perm_layout (nullptr, perm, layout_i, layout_i);
}
}
else
{
gcc_assert (!SLP_TREE_LANE_PERMUTATION (node).exists ());
auto &load_perm = SLP_TREE_LOAD_PERMUTATION (node);
if (layout_i > 0)
/* ??? When we handle non-bijective permutes the idea
is that we can force the load-permutation to be
{ min, min + 1, min + 2, ... max }. But then the
scalar defs might no longer match the lane content
which means wrong-code with live lane vectorization.
So we possibly have to have NULL entries for those. */
vect_slp_permute (m_perms[layout_i], load_perm, true);
}
}
/* Do this before any nodes disappear, since it involves a walk
over the leaves. */
remove_redundant_permutations ();
/* Replace each child with a correctly laid-out version. */
for (unsigned int node_i : m_partitioned_nodes)
{
/* Skip nodes that have already been handled above. */
if (bitmap_bit_p (fully_folded, node_i))
continue;
auto &vertex = m_vertices[node_i];
int in_layout_i = m_partitions[vertex.partition].layout;
gcc_assert (in_layout_i >= 0);
unsigned j;
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (vertex.node), j, child)
{
if (!child)
continue;
slp_tree new_child = get_result_with_layout (child, in_layout_i);
if (new_child != child)
{
vect_free_slp_tree (child);
SLP_TREE_CHILDREN (vertex.node)[j] = new_child;
new_child->refcnt += 1;
}
}
}
}
/* Elide load permutations that are not necessary. Such permutations might
be pre-existing, rather than created by the layout optimizations. */
void
vect_optimize_slp_pass::remove_redundant_permutations ()
{
for (unsigned int node_i : m_leafs)
{
slp_tree node = m_vertices[node_i].node;
if (!SLP_TREE_LOAD_PERMUTATION (node).exists ())
continue;
/* In basic block vectorization we allow any subchain of an interleaving
chain.
FORNOW: not in loop SLP because of realignment complications. */
if (is_a <bb_vec_info> (m_vinfo))
{
bool subchain_p = true;
stmt_vec_info next_load_info = NULL;
stmt_vec_info load_info;
unsigned j;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
{
if (j != 0
&& (next_load_info != load_info
|| DR_GROUP_GAP (load_info) != 1))
{
subchain_p = false;
break;
}
next_load_info = DR_GROUP_NEXT_ELEMENT (load_info);
}
if (subchain_p)
{
SLP_TREE_LOAD_PERMUTATION (node).release ();
continue;
}
}
else
{
loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
stmt_vec_info load_info;
bool this_load_permuted = false;
unsigned j;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
if (SLP_TREE_LOAD_PERMUTATION (node)[j] != j)
{
this_load_permuted = true;
break;
}
stmt_vec_info first_stmt_info
= DR_GROUP_FIRST_ELEMENT (SLP_TREE_SCALAR_STMTS (node)[0]);
if (!this_load_permuted
/* The load requires permutation when unrolling exposes
a gap either because the group is larger than the SLP
group-size or because there is a gap between the groups. */
&& (known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1U)
|| ((SLP_TREE_LANES (node) == DR_GROUP_SIZE (first_stmt_info))
&& DR_GROUP_GAP (first_stmt_info) == 0)))
{
SLP_TREE_LOAD_PERMUTATION (node).release ();
continue;
}
}
}
}
/* Print the partition graph and layout information to the dump file. */
void
vect_optimize_slp_pass::dump ()
{
dump_printf_loc (MSG_NOTE, vect_location,
"SLP optimize permutations:\n");
for (unsigned int layout_i = 1; layout_i < m_perms.length (); ++layout_i)
{
dump_printf_loc (MSG_NOTE, vect_location, " %d: { ", layout_i);
const char *sep = "";
for (unsigned int idx : m_perms[layout_i])
{
dump_printf (MSG_NOTE, "%s%d", sep, idx);
sep = ", ";
}
dump_printf (MSG_NOTE, " }\n");
}
dump_printf_loc (MSG_NOTE, vect_location,
"SLP optimize partitions:\n");
for (unsigned int partition_i = 0; partition_i < m_partitions.length ();
++partition_i)
{
auto &partition = m_partitions[partition_i];
dump_printf_loc (MSG_NOTE, vect_location, " -------------\n");
dump_printf_loc (MSG_NOTE, vect_location,
" partition %d (layout %d):\n",
partition_i, partition.layout);
dump_printf_loc (MSG_NOTE, vect_location, " nodes:\n");
for (unsigned int order_i = partition.node_begin;
order_i < partition.node_end; ++order_i)
{
auto &vertex = m_vertices[m_partitioned_nodes[order_i]];
dump_printf_loc (MSG_NOTE, vect_location, " - %p:\n",
(void *) vertex.node);
dump_printf_loc (MSG_NOTE, vect_location,
" weight: %f\n",
vertex.weight.to_double ());
if (vertex.out_degree)
dump_printf_loc (MSG_NOTE, vect_location,
" out weight: %f (degree %d)\n",
vertex.out_weight.to_double (),
vertex.out_degree);
if (SLP_TREE_CODE (vertex.node) == VEC_PERM_EXPR)
dump_printf_loc (MSG_NOTE, vect_location,
" op: VEC_PERM_EXPR\n");
else if (auto rep = SLP_TREE_REPRESENTATIVE (vertex.node))
dump_printf_loc (MSG_NOTE, vect_location,
" op template: %G", rep->stmt);
}
dump_printf_loc (MSG_NOTE, vect_location, " edges:\n");
for (unsigned int order_i = partition.node_begin;
order_i < partition.node_end; ++order_i)
{
unsigned int node_i = m_partitioned_nodes[order_i];
auto &vertex = m_vertices[node_i];
auto print_edge = [&](graph_edge *, unsigned int other_node_i)
{
auto &other_vertex = m_vertices[other_node_i];
if (other_vertex.partition < vertex.partition)
dump_printf_loc (MSG_NOTE, vect_location,
" - %p [%d] --> %p\n",
(void *) other_vertex.node,
other_vertex.partition,
(void *) vertex.node);
else
dump_printf_loc (MSG_NOTE, vect_location,
" - %p --> [%d] %p\n",
(void *) vertex.node,
other_vertex.partition,
(void *) other_vertex.node);
};
for_each_partition_edge (node_i, print_edge);
}
for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
{
auto &layout_costs = partition_layout_costs (partition_i, layout_i);
if (layout_costs.is_possible ())
{
dump_printf_loc (MSG_NOTE, vect_location,
" layout %d:%s\n", layout_i,
partition.layout == int (layout_i)
? " (*)" : "");
slpg_layout_cost combined_cost = layout_costs.in_cost;
combined_cost.add_serial_cost (layout_costs.internal_cost);
combined_cost.add_serial_cost (layout_costs.out_cost);
#define TEMPLATE "{depth: %f, total: %f}"
dump_printf_loc (MSG_NOTE, vect_location,
" " TEMPLATE "\n",
layout_costs.in_cost.depth.to_double (),
layout_costs.in_cost.total.to_double ());
dump_printf_loc (MSG_NOTE, vect_location,
" + " TEMPLATE "\n",
layout_costs.internal_cost.depth.to_double (),
layout_costs.internal_cost.total.to_double ());
dump_printf_loc (MSG_NOTE, vect_location,
" + " TEMPLATE "\n",
layout_costs.out_cost.depth.to_double (),
layout_costs.out_cost.total.to_double ());
dump_printf_loc (MSG_NOTE, vect_location,
" = " TEMPLATE "\n",
combined_cost.depth.to_double (),
combined_cost.total.to_double ());
#undef TEMPLATE
}
else
dump_printf_loc (MSG_NOTE, vect_location,
" layout %d: rejected\n", layout_i);
}
}
}
/* Main entry point for the SLP graph optimization pass. */
void
vect_optimize_slp_pass::run ()
{
build_graph ();
create_partitions ();
start_choosing_layouts ();
if (m_perms.length () > 1)
{
forward_pass ();
backward_pass ();
if (dump_enabled_p ())
dump ();
materialize ();
while (!m_perms.is_empty ())
m_perms.pop ().release ();
}
else
remove_redundant_permutations ();
free_graph (m_slpg);
}
/* Optimize the SLP graph of VINFO. */
void
vect_optimize_slp (vec_info *vinfo)
{
if (vinfo->slp_instances.is_empty ())
return;
vect_optimize_slp_pass (vinfo).run ();
}
/* Gather loads reachable from the individual SLP graph entries. */
void
vect_gather_slp_loads (vec_info *vinfo)
{
unsigned i;
slp_instance instance;
FOR_EACH_VEC_ELT (vinfo->slp_instances, i, instance)
{
hash_set<slp_tree> visited;
vect_gather_slp_loads (SLP_INSTANCE_LOADS (instance),
SLP_INSTANCE_TREE (instance), visited);
}
}
/* For each possible SLP instance decide whether to SLP it and calculate overall
unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at
least one instance. */
bool
vect_make_slp_decision (loop_vec_info loop_vinfo)
{
unsigned int i;
poly_uint64 unrolling_factor = 1;
const vec<slp_instance> &slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
slp_instance instance;
int decided_to_slp = 0;
DUMP_VECT_SCOPE ("vect_make_slp_decision");
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
/* FORNOW: SLP if you can. */
/* All unroll factors have the form:
GET_MODE_SIZE (vinfo->vector_mode) * X
for some rational X, so they must have a common multiple. */
unrolling_factor
= force_common_multiple (unrolling_factor,
SLP_INSTANCE_UNROLLING_FACTOR (instance));
/* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we
call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and
loop-based vectorization. Such stmts will be marked as HYBRID. */
vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance));
decided_to_slp++;
}
LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo) = unrolling_factor;
if (decided_to_slp && dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Decided to SLP %d instances. Unrolling factor ",
decided_to_slp);
dump_dec (MSG_NOTE, unrolling_factor);
dump_printf (MSG_NOTE, "\n");
}
return (decided_to_slp > 0);
}
/* Private data for vect_detect_hybrid_slp. */
struct vdhs_data
{
loop_vec_info loop_vinfo;
vec<stmt_vec_info> *worklist;
};
/* Walker for walk_gimple_op. */
static tree
vect_detect_hybrid_slp (tree *tp, int *, void *data)
{
walk_stmt_info *wi = (walk_stmt_info *)data;
vdhs_data *dat = (vdhs_data *)wi->info;
if (wi->is_lhs)
return NULL_TREE;
stmt_vec_info def_stmt_info = dat->loop_vinfo->lookup_def (*tp);
if (!def_stmt_info)
return NULL_TREE;
def_stmt_info = vect_stmt_to_vectorize (def_stmt_info);
if (PURE_SLP_STMT (def_stmt_info))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "marking hybrid: %G",
def_stmt_info->stmt);
STMT_SLP_TYPE (def_stmt_info) = hybrid;
dat->worklist->safe_push (def_stmt_info);
}
return NULL_TREE;
}
/* Look if STMT_INFO is consumed by SLP indirectly and mark it pure_slp
if so, otherwise pushing it to WORKLIST. */
static void
maybe_push_to_hybrid_worklist (vec_info *vinfo,
vec<stmt_vec_info> &worklist,
stmt_vec_info stmt_info)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Processing hybrid candidate : %G", stmt_info->stmt);
stmt_vec_info orig_info = vect_orig_stmt (stmt_info);
imm_use_iterator iter2;
ssa_op_iter iter1;
use_operand_p use_p;
def_operand_p def_p;
bool any_def = false;
FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_info->stmt, iter1, SSA_OP_DEF)
{
any_def = true;
FOR_EACH_IMM_USE_FAST (use_p, iter2, DEF_FROM_PTR (def_p))
{
if (is_gimple_debug (USE_STMT (use_p)))
continue;
stmt_vec_info use_info = vinfo->lookup_stmt (USE_STMT (use_p));
/* An out-of loop use means this is a loop_vect sink. */
if (!use_info)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Found loop_vect sink: %G", stmt_info->stmt);
worklist.safe_push (stmt_info);
return;
}
else if (!STMT_SLP_TYPE (vect_stmt_to_vectorize (use_info)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Found loop_vect use: %G", use_info->stmt);
worklist.safe_push (stmt_info);
return;
}
}
}
/* No def means this is a loo_vect sink. */
if (!any_def)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Found loop_vect sink: %G", stmt_info->stmt);
worklist.safe_push (stmt_info);
return;
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Marked SLP consumed stmt pure: %G", stmt_info->stmt);
STMT_SLP_TYPE (stmt_info) = pure_slp;
}
/* Find stmts that must be both vectorized and SLPed. */
void
vect_detect_hybrid_slp (loop_vec_info loop_vinfo)
{
DUMP_VECT_SCOPE ("vect_detect_hybrid_slp");
/* All stmts participating in SLP are marked pure_slp, all other
stmts are loop_vect.
First collect all loop_vect stmts into a worklist.
SLP patterns cause not all original scalar stmts to appear in
SLP_TREE_SCALAR_STMTS and thus not all of them are marked pure_slp.
Rectify this here and do a backward walk over the IL only considering
stmts as loop_vect when they are used by a loop_vect stmt and otherwise
mark them as pure_slp. */
auto_vec<stmt_vec_info> worklist;
for (int i = LOOP_VINFO_LOOP (loop_vinfo)->num_nodes - 1; i >= 0; --i)
{
basic_block bb = LOOP_VINFO_BBS (loop_vinfo)[i];
for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
gsi_next (&gsi))
{
gphi *phi = gsi.phi ();
stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (phi);
if (!STMT_SLP_TYPE (stmt_info) && STMT_VINFO_RELEVANT (stmt_info))
maybe_push_to_hybrid_worklist (loop_vinfo,
worklist, stmt_info);
}
for (gimple_stmt_iterator gsi = gsi_last_bb (bb); !gsi_end_p (gsi);
gsi_prev (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (is_gimple_debug (stmt))
continue;
stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (stmt);
if (STMT_VINFO_IN_PATTERN_P (stmt_info))
{
for (gimple_stmt_iterator gsi2
= gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
!gsi_end_p (gsi2); gsi_next (&gsi2))
{
stmt_vec_info patt_info
= loop_vinfo->lookup_stmt (gsi_stmt (gsi2));
if (!STMT_SLP_TYPE (patt_info)
&& STMT_VINFO_RELEVANT (patt_info))
maybe_push_to_hybrid_worklist (loop_vinfo,
worklist, patt_info);
}
stmt_info = STMT_VINFO_RELATED_STMT (stmt_info);
}
if (!STMT_SLP_TYPE (stmt_info) && STMT_VINFO_RELEVANT (stmt_info))
maybe_push_to_hybrid_worklist (loop_vinfo,
worklist, stmt_info);
}
}
/* Now we have a worklist of non-SLP stmts, follow use->def chains and
mark any SLP vectorized stmt as hybrid.
??? We're visiting def stmts N times (once for each non-SLP and
once for each hybrid-SLP use). */
walk_stmt_info wi;
vdhs_data dat;
dat.worklist = &worklist;
dat.loop_vinfo = loop_vinfo;
memset (&wi, 0, sizeof (wi));
wi.info = (void *)&dat;
while (!worklist.is_empty ())
{
stmt_vec_info stmt_info = worklist.pop ();
/* Since SSA operands are not set up for pattern stmts we need
to use walk_gimple_op. */
wi.is_lhs = 0;
walk_gimple_op (stmt_info->stmt, vect_detect_hybrid_slp, &wi);
}
}
/* Initialize a bb_vec_info struct for the statements in BBS basic blocks. */
_bb_vec_info::_bb_vec_info (vec<basic_block> _bbs, vec_info_shared *shared)
: vec_info (vec_info::bb, shared),
bbs (_bbs),
roots (vNULL)
{
for (unsigned i = 0; i < bbs.length (); ++i)
{
if (i != 0)
for (gphi_iterator si = gsi_start_phis (bbs[i]); !gsi_end_p (si);
gsi_next (&si))
{
gphi *phi = si.phi ();
gimple_set_uid (phi, 0);
add_stmt (phi);
}
for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]);
!gsi_end_p (gsi); gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
gimple_set_uid (stmt, 0);
if (is_gimple_debug (stmt))
continue;
add_stmt (stmt);
}
}
}
/* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the
stmts in the basic block. */
_bb_vec_info::~_bb_vec_info ()
{
/* Reset region marker. */
for (unsigned i = 0; i < bbs.length (); ++i)
{
if (i != 0)
for (gphi_iterator si = gsi_start_phis (bbs[i]); !gsi_end_p (si);
gsi_next (&si))
{
gphi *phi = si.phi ();
gimple_set_uid (phi, -1);
}
for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]);
!gsi_end_p (gsi); gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
gimple_set_uid (stmt, -1);
}
}
for (unsigned i = 0; i < roots.length (); ++i)
{
roots[i].stmts.release ();
roots[i].roots.release ();
}
roots.release ();
}
/* Subroutine of vect_slp_analyze_node_operations. Handle the root of NODE,
given then that child nodes have already been processed, and that
their def types currently match their SLP node's def type. */
static bool
vect_slp_analyze_node_operations_1 (vec_info *vinfo, slp_tree node,
slp_instance node_instance,
stmt_vector_for_cost *cost_vec)
{
stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
/* Calculate the number of vector statements to be created for the
scalar stmts in this node. For SLP reductions it is equal to the
number of vector statements in the children (which has already been
calculated by the recursive call). Otherwise it is the number of
scalar elements in one scalar iteration (DR_GROUP_SIZE) multiplied by
VF divided by the number of elements in a vector. */
if (!STMT_VINFO_DATA_REF (stmt_info)
&& REDUC_GROUP_FIRST_ELEMENT (stmt_info))
{
for (unsigned i = 0; i < SLP_TREE_CHILDREN (node).length (); ++i)
if (SLP_TREE_DEF_TYPE (SLP_TREE_CHILDREN (node)[i]) == vect_internal_def)
{
SLP_TREE_NUMBER_OF_VEC_STMTS (node)
= SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_CHILDREN (node)[i]);
break;
}
}
else
{
poly_uint64 vf;
if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
vf = loop_vinfo->vectorization_factor;
else
vf = 1;
unsigned int group_size = SLP_TREE_LANES (node);
tree vectype = SLP_TREE_VECTYPE (node);
SLP_TREE_NUMBER_OF_VEC_STMTS (node)
= vect_get_num_vectors (vf * group_size, vectype);
}
/* Handle purely internal nodes. */
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
return vectorizable_slp_permutation (vinfo, NULL, node, cost_vec);
gcc_assert (STMT_SLP_TYPE (stmt_info) != loop_vect);
bool dummy;
return vect_analyze_stmt (vinfo, stmt_info, &dummy,
node, node_instance, cost_vec);
}
/* Try to build NODE from scalars, returning true on success.
NODE_INSTANCE is the SLP instance that contains NODE. */
static bool
vect_slp_convert_to_external (vec_info *vinfo, slp_tree node,
slp_instance node_instance)
{
stmt_vec_info stmt_info;
unsigned int i;
if (!is_a <bb_vec_info> (vinfo)
|| node == SLP_INSTANCE_TREE (node_instance)
|| !SLP_TREE_SCALAR_STMTS (node).exists ()
|| vect_contains_pattern_stmt_p (SLP_TREE_SCALAR_STMTS (node))
/* Force the mask use to be built from scalars instead. */
|| VECTOR_BOOLEAN_TYPE_P (SLP_TREE_VECTYPE (node)))
return false;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Building vector operands of %p from scalars instead\n",
(void *) node);
/* Don't remove and free the child nodes here, since they could be
referenced by other structures. The analysis and scheduling phases
(need to) ignore child nodes of anything that isn't vect_internal_def. */
unsigned int group_size = SLP_TREE_LANES (node);
SLP_TREE_DEF_TYPE (node) = vect_external_def;
/* Invariants get their vector type from the uses. */
SLP_TREE_VECTYPE (node) = NULL_TREE;
SLP_TREE_SCALAR_OPS (node).safe_grow (group_size, true);
SLP_TREE_LOAD_PERMUTATION (node).release ();
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
{
tree lhs = gimple_get_lhs (vect_orig_stmt (stmt_info)->stmt);
SLP_TREE_SCALAR_OPS (node)[i] = lhs;
}
return true;
}
/* Return true if all elements of the slice are the same. */
bool
vect_scalar_ops_slice::all_same_p () const
{
for (unsigned int i = 1; i < length; ++i)
if (!operand_equal_p (op (0), op (i)))
return false;
return true;
}
hashval_t
vect_scalar_ops_slice_hash::hash (const value_type &s)
{
hashval_t hash = 0;
for (unsigned i = 0; i < s.length; ++i)
hash = iterative_hash_expr (s.op (i), hash);
return hash;
}
bool
vect_scalar_ops_slice_hash::equal (const value_type &s1,
const compare_type &s2)
{
if (s1.length != s2.length)
return false;
for (unsigned i = 0; i < s1.length; ++i)
if (!operand_equal_p (s1.op (i), s2.op (i)))
return false;
return true;
}
/* Compute the prologue cost for invariant or constant operands represented
by NODE. */
static void
vect_prologue_cost_for_slp (slp_tree node,
stmt_vector_for_cost *cost_vec)
{
/* There's a special case of an existing vector, that costs nothing. */
if (SLP_TREE_SCALAR_OPS (node).length () == 0
&& !SLP_TREE_VEC_DEFS (node).is_empty ())
return;
/* Without looking at the actual initializer a vector of
constants can be implemented as load from the constant pool.
When all elements are the same we can use a splat. */
tree vectype = SLP_TREE_VECTYPE (node);
unsigned group_size = SLP_TREE_SCALAR_OPS (node).length ();
unsigned HOST_WIDE_INT const_nunits;
unsigned nelt_limit;
auto ops = &SLP_TREE_SCALAR_OPS (node);
auto_vec<unsigned int> starts (SLP_TREE_NUMBER_OF_VEC_STMTS (node));
if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits)
&& ! multiple_p (const_nunits, group_size))
{
nelt_limit = const_nunits;
hash_set<vect_scalar_ops_slice_hash> vector_ops;
for (unsigned int i = 0; i < SLP_TREE_NUMBER_OF_VEC_STMTS (node); ++i)
if (!vector_ops.add ({ ops, i * const_nunits, const_nunits }))
starts.quick_push (i * const_nunits);
}
else
{
/* If either the vector has variable length or the vectors
are composed of repeated whole groups we only need to
cost construction once. All vectors will be the same. */
nelt_limit = group_size;
starts.quick_push (0);
}
/* ??? We're just tracking whether vectors in a single node are the same.
Ideally we'd do something more global. */
for (unsigned int start : starts)
{
vect_cost_for_stmt kind;
if (SLP_TREE_DEF_TYPE (node) == vect_constant_def)
kind = vector_load;
else if (vect_scalar_ops_slice { ops, start, nelt_limit }.all_same_p ())
kind = scalar_to_vec;
else
kind = vec_construct;
record_stmt_cost (cost_vec, 1, kind, node, vectype, 0, vect_prologue);
}
}
/* Analyze statements contained in SLP tree NODE after recursively analyzing
the subtree. NODE_INSTANCE contains NODE and VINFO contains INSTANCE.
Return true if the operations are supported. */
static bool
vect_slp_analyze_node_operations (vec_info *vinfo, slp_tree node,
slp_instance node_instance,
hash_set<slp_tree> &visited_set,
vec<slp_tree> &visited_vec,
stmt_vector_for_cost *cost_vec)
{
int i, j;
slp_tree child;
/* Assume we can code-generate all invariants. */
if (!node
|| SLP_TREE_DEF_TYPE (node) == vect_constant_def
|| SLP_TREE_DEF_TYPE (node) == vect_external_def)
return true;
if (SLP_TREE_DEF_TYPE (node) == vect_uninitialized_def)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Failed cyclic SLP reference in %p\n", (void *) node);
return false;
}
gcc_assert (SLP_TREE_DEF_TYPE (node) == vect_internal_def);
/* If we already analyzed the exact same set of scalar stmts we're done.
We share the generated vector stmts for those. */
if (visited_set.add (node))
return true;
visited_vec.safe_push (node);
bool res = true;
unsigned visited_rec_start = visited_vec.length ();
unsigned cost_vec_rec_start = cost_vec->length ();
bool seen_non_constant_child = false;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
{
res = vect_slp_analyze_node_operations (vinfo, child, node_instance,
visited_set, visited_vec,
cost_vec);
if (!res)
break;
if (child && SLP_TREE_DEF_TYPE (child) != vect_constant_def)
seen_non_constant_child = true;
}
/* We're having difficulties scheduling nodes with just constant
operands and no scalar stmts since we then cannot compute a stmt
insertion place. */
if (!seen_non_constant_child && SLP_TREE_SCALAR_STMTS (node).is_empty ())
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Cannot vectorize all-constant op node %p\n",
(void *) node);
res = false;
}
if (res)
res = vect_slp_analyze_node_operations_1 (vinfo, node, node_instance,
cost_vec);
/* If analysis failed we have to pop all recursive visited nodes
plus ourselves. */
if (!res)
{
while (visited_vec.length () >= visited_rec_start)
visited_set.remove (visited_vec.pop ());
cost_vec->truncate (cost_vec_rec_start);
}
/* When the node can be vectorized cost invariant nodes it references.
This is not done in DFS order to allow the refering node
vectorizable_* calls to nail down the invariant nodes vector type
and possibly unshare it if it needs a different vector type than
other referrers. */
if (res)
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child)
if (child
&& (SLP_TREE_DEF_TYPE (child) == vect_constant_def
|| SLP_TREE_DEF_TYPE (child) == vect_external_def)
/* Perform usual caching, note code-generation still
code-gens these nodes multiple times but we expect
to CSE them later. */
&& !visited_set.add (child))
{
visited_vec.safe_push (child);
/* ??? After auditing more code paths make a "default"
and push the vector type from NODE to all children
if it is not already set. */
/* Compute the number of vectors to be generated. */
tree vector_type = SLP_TREE_VECTYPE (child);
if (!vector_type)
{
/* For shifts with a scalar argument we don't need
to cost or code-generate anything.
??? Represent this more explicitely. */
gcc_assert ((STMT_VINFO_TYPE (SLP_TREE_REPRESENTATIVE (node))
== shift_vec_info_type)
&& j == 1);
continue;
}
unsigned group_size = SLP_TREE_LANES (child);
poly_uint64 vf = 1;
if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
vf = loop_vinfo->vectorization_factor;
SLP_TREE_NUMBER_OF_VEC_STMTS (child)
= vect_get_num_vectors (vf * group_size, vector_type);
/* And cost them. */
vect_prologue_cost_for_slp (child, cost_vec);
}
/* If this node or any of its children can't be vectorized, try pruning
the tree here rather than felling the whole thing. */
if (!res && vect_slp_convert_to_external (vinfo, node, node_instance))
{
/* We'll need to revisit this for invariant costing and number
of vectorized stmt setting. */
res = true;
}
return res;
}
/* Mark lanes of NODE that are live outside of the basic-block vectorized
region and that can be vectorized using vectorizable_live_operation
with STMT_VINFO_LIVE_P. Not handled live operations will cause the
scalar code computing it to be retained. */
static void
vect_bb_slp_mark_live_stmts (bb_vec_info bb_vinfo, slp_tree node,
slp_instance instance,
stmt_vector_for_cost *cost_vec,
hash_set<stmt_vec_info> &svisited,
hash_set<slp_tree> &visited)
{
if (visited.add (node))
return;
unsigned i;
stmt_vec_info stmt_info;
stmt_vec_info last_stmt = vect_find_last_scalar_stmt_in_slp (node);
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
{
if (svisited.contains (stmt_info))
continue;
stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
if (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)
&& STMT_VINFO_RELATED_STMT (orig_stmt_info) != stmt_info)
/* Only the pattern root stmt computes the original scalar value. */
continue;
bool mark_visited = true;
gimple *orig_stmt = orig_stmt_info->stmt;
ssa_op_iter op_iter;
def_operand_p def_p;
FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_stmt, op_iter, SSA_OP_DEF)
{
imm_use_iterator use_iter;
gimple *use_stmt;
stmt_vec_info use_stmt_info;
FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p))
if (!is_gimple_debug (use_stmt))
{
use_stmt_info = bb_vinfo->lookup_stmt (use_stmt);
if (!use_stmt_info
|| !PURE_SLP_STMT (vect_stmt_to_vectorize (use_stmt_info)))
{
STMT_VINFO_LIVE_P (stmt_info) = true;
if (vectorizable_live_operation (bb_vinfo, stmt_info,
NULL, node, instance, i,
false, cost_vec))
/* ??? So we know we can vectorize the live stmt
from one SLP node. If we cannot do so from all
or none consistently we'd have to record which
SLP node (and lane) we want to use for the live
operation. So make sure we can code-generate
from all nodes. */
mark_visited = false;
else
STMT_VINFO_LIVE_P (stmt_info) = false;
break;
}
}
/* We have to verify whether we can insert the lane extract
before all uses. The following is a conservative approximation.
We cannot put this into vectorizable_live_operation because
iterating over all use stmts from inside a FOR_EACH_IMM_USE_STMT
doesn't work.
Note that while the fact that we emit code for loads at the
first load should make this a non-problem leafs we construct
from scalars are vectorized after the last scalar def.
??? If we'd actually compute the insert location during
analysis we could use sth less conservative than the last
scalar stmt in the node for the dominance check. */
/* ??? What remains is "live" uses in vector CTORs in the same
SLP graph which is where those uses can end up code-generated
right after their definition instead of close to their original
use. But that would restrict us to code-generate lane-extracts
from the latest stmt in a node. So we compensate for this
during code-generation, simply not replacing uses for those
hopefully rare cases. */
if (STMT_VINFO_LIVE_P (stmt_info))
FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p))
if (!is_gimple_debug (use_stmt)
&& (!(use_stmt_info = bb_vinfo->lookup_stmt (use_stmt))
|| !PURE_SLP_STMT (vect_stmt_to_vectorize (use_stmt_info)))
&& !vect_stmt_dominates_stmt_p (last_stmt->stmt, use_stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Cannot determine insertion place for "
"lane extract\n");
STMT_VINFO_LIVE_P (stmt_info) = false;
mark_visited = true;
}
}
if (mark_visited)
svisited.add (stmt_info);
}
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
vect_bb_slp_mark_live_stmts (bb_vinfo, child, instance,
cost_vec, svisited, visited);
}
/* Determine whether we can vectorize the reduction epilogue for INSTANCE. */
static bool
vectorizable_bb_reduc_epilogue (slp_instance instance,
stmt_vector_for_cost *cost_vec)
{
gassign *stmt = as_a <gassign *> (instance->root_stmts[0]->stmt);
enum tree_code reduc_code = gimple_assign_rhs_code (stmt);
if (reduc_code == MINUS_EXPR)
reduc_code = PLUS_EXPR;
internal_fn reduc_fn;
tree vectype = SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (instance));
if (!vectype
|| !reduction_fn_for_scalar_code (reduc_code, &reduc_fn)
|| reduc_fn == IFN_LAST
|| !direct_internal_fn_supported_p (reduc_fn, vectype, OPTIMIZE_FOR_BOTH)
|| !useless_type_conversion_p (TREE_TYPE (gimple_assign_lhs (stmt)),
TREE_TYPE (vectype)))
return false;
/* There's no way to cost a horizontal vector reduction via REDUC_FN so
cost log2 vector operations plus shuffles and one extraction. */
unsigned steps = floor_log2 (vect_nunits_for_cost (vectype));
record_stmt_cost (cost_vec, steps, vector_stmt, instance->root_stmts[0],
vectype, 0, vect_body);
record_stmt_cost (cost_vec, steps, vec_perm, instance->root_stmts[0],
vectype, 0, vect_body);
record_stmt_cost (cost_vec, 1, vec_to_scalar, instance->root_stmts[0],
vectype, 0, vect_body);
return true;
}
/* Prune from ROOTS all stmts that are computed as part of lanes of NODE
and recurse to children. */
static void
vect_slp_prune_covered_roots (slp_tree node, hash_set<stmt_vec_info> &roots,
hash_set<slp_tree> &visited)
{
if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
|| visited.add (node))
return;
stmt_vec_info stmt;
unsigned i;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt)
roots.remove (vect_orig_stmt (stmt));
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
vect_slp_prune_covered_roots (child, roots, visited);
}
/* Analyze statements in SLP instances of VINFO. Return true if the
operations are supported. */
bool
vect_slp_analyze_operations (vec_info *vinfo)
{
slp_instance instance;
int i;
DUMP_VECT_SCOPE ("vect_slp_analyze_operations");
hash_set<slp_tree> visited;
for (i = 0; vinfo->slp_instances.iterate (i, &instance); )
{
auto_vec<slp_tree> visited_vec;
stmt_vector_for_cost cost_vec;
cost_vec.create (2);
if (is_a <bb_vec_info> (vinfo))
vect_location = instance->location ();
if (!vect_slp_analyze_node_operations (vinfo,
SLP_INSTANCE_TREE (instance),
instance, visited, visited_vec,
&cost_vec)
/* CTOR instances require vectorized defs for the SLP tree root. */
|| (SLP_INSTANCE_KIND (instance) == slp_inst_kind_ctor
&& (SLP_TREE_DEF_TYPE (SLP_INSTANCE_TREE (instance))
!= vect_internal_def
/* Make sure we vectorized with the expected type. */
|| !useless_type_conversion_p
(TREE_TYPE (TREE_TYPE (gimple_assign_rhs1
(instance->root_stmts[0]->stmt))),
TREE_TYPE (SLP_TREE_VECTYPE
(SLP_INSTANCE_TREE (instance))))))
/* Check we can vectorize the reduction. */
|| (SLP_INSTANCE_KIND (instance) == slp_inst_kind_bb_reduc
&& !vectorizable_bb_reduc_epilogue (instance, &cost_vec)))
{
slp_tree node = SLP_INSTANCE_TREE (instance);
stmt_vec_info stmt_info;
if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
stmt_info = SLP_INSTANCE_ROOT_STMTS (instance)[0];
else
stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"removing SLP instance operations starting from: %G",
stmt_info->stmt);
vect_free_slp_instance (instance);
vinfo->slp_instances.ordered_remove (i);
cost_vec.release ();
while (!visited_vec.is_empty ())
visited.remove (visited_vec.pop ());
}
else
{
i++;
if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
{
add_stmt_costs (loop_vinfo->vector_costs, &cost_vec);
cost_vec.release ();
}
else
/* For BB vectorization remember the SLP graph entry
cost for later. */
instance->cost_vec = cost_vec;
}
}
/* Now look for SLP instances with a root that are covered by other
instances and remove them. */
hash_set<stmt_vec_info> roots;
for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
roots.add (SLP_INSTANCE_ROOT_STMTS (instance)[0]);
if (!roots.is_empty ())
{
visited.empty ();
for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
vect_slp_prune_covered_roots (SLP_INSTANCE_TREE (instance), roots,
visited);
for (i = 0; vinfo->slp_instances.iterate (i, &instance); )
if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ()
&& !roots.contains (SLP_INSTANCE_ROOT_STMTS (instance)[0]))
{
stmt_vec_info root = SLP_INSTANCE_ROOT_STMTS (instance)[0];
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"removing SLP instance operations starting "
"from: %G", root->stmt);
vect_free_slp_instance (instance);
vinfo->slp_instances.ordered_remove (i);
}
else
++i;
}
/* Compute vectorizable live stmts. */
if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo))
{
hash_set<stmt_vec_info> svisited;
hash_set<slp_tree> visited;
for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
{
vect_location = instance->location ();
vect_bb_slp_mark_live_stmts (bb_vinfo, SLP_INSTANCE_TREE (instance),
instance, &instance->cost_vec, svisited,
visited);
}
}
return !vinfo->slp_instances.is_empty ();
}
/* Get the SLP instance leader from INSTANCE_LEADER thereby transitively
closing the eventual chain. */
static slp_instance
get_ultimate_leader (slp_instance instance,
hash_map<slp_instance, slp_instance> &instance_leader)
{
auto_vec<slp_instance *, 8> chain;
slp_instance *tem;
while (*(tem = instance_leader.get (instance)) != instance)
{
chain.safe_push (tem);
instance = *tem;
}
while (!chain.is_empty ())
*chain.pop () = instance;
return instance;
}
namespace {
/* Subroutine of vect_bb_partition_graph_r. Map KEY to INSTANCE in
KEY_TO_INSTANCE, making INSTANCE the leader of any previous mapping
for KEY. Return true if KEY was already in KEY_TO_INSTANCE.
INSTANCE_LEADER is as for get_ultimate_leader. */
template<typename T>
bool
vect_map_to_instance (slp_instance instance, T key,
hash_map<T, slp_instance> &key_to_instance,
hash_map<slp_instance, slp_instance> &instance_leader)
{
bool existed_p;
slp_instance &key_instance = key_to_instance.get_or_insert (key, &existed_p);
if (!existed_p)
;
else if (key_instance != instance)
{
/* If we're running into a previously marked key make us the
leader of the current ultimate leader. This keeps the
leader chain acyclic and works even when the current instance
connects two previously independent graph parts. */
slp_instance key_leader
= get_ultimate_leader (key_instance, instance_leader);
if (key_leader != instance)
instance_leader.put (key_leader, instance);
}
key_instance = instance;
return existed_p;
}
}
/* Worker of vect_bb_partition_graph, recurse on NODE. */
static void
vect_bb_partition_graph_r (bb_vec_info bb_vinfo,
slp_instance instance, slp_tree node,
hash_map<stmt_vec_info, slp_instance> &stmt_to_instance,
hash_map<slp_tree, slp_instance> &node_to_instance,
hash_map<slp_instance, slp_instance> &instance_leader)
{
stmt_vec_info stmt_info;
unsigned i;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
vect_map_to_instance (instance, stmt_info, stmt_to_instance,
instance_leader);
if (vect_map_to_instance (instance, node, node_to_instance,
instance_leader))
return;
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
vect_bb_partition_graph_r (bb_vinfo, instance, child, stmt_to_instance,
node_to_instance, instance_leader);
}
/* Partition the SLP graph into pieces that can be costed independently. */
static void
vect_bb_partition_graph (bb_vec_info bb_vinfo)
{
DUMP_VECT_SCOPE ("vect_bb_partition_graph");
/* First walk the SLP graph assigning each involved scalar stmt a
corresponding SLP graph entry and upon visiting a previously
marked stmt, make the stmts leader the current SLP graph entry. */
hash_map<stmt_vec_info, slp_instance> stmt_to_instance;
hash_map<slp_tree, slp_instance> node_to_instance;
hash_map<slp_instance, slp_instance> instance_leader;
slp_instance instance;
for (unsigned i = 0; bb_vinfo->slp_instances.iterate (i, &instance); ++i)
{
instance_leader.put (instance, instance);
vect_bb_partition_graph_r (bb_vinfo,
instance, SLP_INSTANCE_TREE (instance),
stmt_to_instance, node_to_instance,
instance_leader);
}
/* Then collect entries to each independent subgraph. */
for (unsigned i = 0; bb_vinfo->slp_instances.iterate (i, &instance); ++i)
{
slp_instance leader = get_ultimate_leader (instance, instance_leader);
leader->subgraph_entries.safe_push (instance);
if (dump_enabled_p ()
&& leader != instance)
dump_printf_loc (MSG_NOTE, vect_location,
"instance %p is leader of %p\n",
(void *) leader, (void *) instance);
}
}
/* Compute the set of scalar stmts participating in internal and external
nodes. */
static void
vect_slp_gather_vectorized_scalar_stmts (vec_info *vinfo, slp_tree node,
hash_set<slp_tree> &visited,
hash_set<stmt_vec_info> &vstmts,
hash_set<stmt_vec_info> &estmts)
{
int i;
stmt_vec_info stmt_info;
slp_tree child;
if (visited.add (node))
return;
if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
{
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
vstmts.add (stmt_info);
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (child)
vect_slp_gather_vectorized_scalar_stmts (vinfo, child, visited,
vstmts, estmts);
}
else
for (tree def : SLP_TREE_SCALAR_OPS (node))
{
stmt_vec_info def_stmt = vinfo->lookup_def (def);
if (def_stmt)
estmts.add (def_stmt);
}
}
/* Compute the scalar cost of the SLP node NODE and its children
and return it. Do not account defs that are marked in LIFE and
update LIFE according to uses of NODE. */
static void
vect_bb_slp_scalar_cost (vec_info *vinfo,
slp_tree node, vec<bool, va_heap> *life,
stmt_vector_for_cost *cost_vec,
hash_set<stmt_vec_info> &vectorized_scalar_stmts,
hash_set<slp_tree> &visited)
{
unsigned i;
stmt_vec_info stmt_info;
slp_tree child;
if (visited.add (node))
return;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
{
ssa_op_iter op_iter;
def_operand_p def_p;
if ((*life)[i])
continue;
stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
gimple *orig_stmt = orig_stmt_info->stmt;
/* If there is a non-vectorized use of the defs then the scalar
stmt is kept live in which case we do not account it or any
required defs in the SLP children in the scalar cost. This
way we make the vectorization more costly when compared to
the scalar cost. */
if (!STMT_VINFO_LIVE_P (stmt_info))
{
auto_vec<gimple *, 8> worklist;
hash_set<gimple *> *worklist_visited = NULL;
worklist.quick_push (orig_stmt);
do
{
gimple *work_stmt = worklist.pop ();
FOR_EACH_PHI_OR_STMT_DEF (def_p, work_stmt, op_iter, SSA_OP_DEF)
{
imm_use_iterator use_iter;
gimple *use_stmt;
FOR_EACH_IMM_USE_STMT (use_stmt, use_iter,
DEF_FROM_PTR (def_p))
if (!is_gimple_debug (use_stmt))
{
stmt_vec_info use_stmt_info
= vinfo->lookup_stmt (use_stmt);
if (!use_stmt_info
|| !vectorized_scalar_stmts.contains (use_stmt_info))
{
if (use_stmt_info
&& STMT_VINFO_IN_PATTERN_P (use_stmt_info))
{
/* For stmts participating in patterns we have
to check its uses recursively. */
if (!worklist_visited)
worklist_visited = new hash_set<gimple *> ();
if (!worklist_visited->add (use_stmt))
worklist.safe_push (use_stmt);
continue;
}
(*life)[i] = true;
goto next_lane;
}
}
}
}
while (!worklist.is_empty ());
next_lane:
if (worklist_visited)
delete worklist_visited;
if ((*life)[i])
continue;
}
/* Count scalar stmts only once. */
if (gimple_visited_p (orig_stmt))
continue;
gimple_set_visited (orig_stmt, true);
vect_cost_for_stmt kind;
if (STMT_VINFO_DATA_REF (orig_stmt_info))
{
if (DR_IS_READ (STMT_VINFO_DATA_REF (orig_stmt_info)))
kind = scalar_load;
else
kind = scalar_store;
}
else if (vect_nop_conversion_p (orig_stmt_info))
continue;
/* For single-argument PHIs assume coalescing which means zero cost
for the scalar and the vector PHIs. This avoids artificially
favoring the vector path (but may pessimize it in some cases). */
else if (is_a <gphi *> (orig_stmt_info->stmt)
&& gimple_phi_num_args
(as_a <gphi *> (orig_stmt_info->stmt)) == 1)
continue;
else
kind = scalar_stmt;
record_stmt_cost (cost_vec, 1, kind, orig_stmt_info,
SLP_TREE_VECTYPE (node), 0, vect_body);
}
auto_vec<bool, 20> subtree_life;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
{
if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
{
/* Do not directly pass LIFE to the recursive call, copy it to
confine changes in the callee to the current child/subtree. */
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
{
subtree_life.safe_grow_cleared (SLP_TREE_LANES (child), true);
for (unsigned j = 0;
j < SLP_TREE_LANE_PERMUTATION (node).length (); ++j)
{
auto perm = SLP_TREE_LANE_PERMUTATION (node)[j];
if (perm.first == i)
subtree_life[perm.second] = (*life)[j];
}
}
else
{
gcc_assert (SLP_TREE_LANES (node) == SLP_TREE_LANES (child));
subtree_life.safe_splice (*life);
}
vect_bb_slp_scalar_cost (vinfo, child, &subtree_life, cost_vec,
vectorized_scalar_stmts, visited);
subtree_life.truncate (0);
}
}
}
/* Comparator for the loop-index sorted cost vectors. */
static int
li_cost_vec_cmp (const void *a_, const void *b_)
{
auto *a = (const std::pair<unsigned, stmt_info_for_cost *> *)a_;
auto *b = (const std::pair<unsigned, stmt_info_for_cost *> *)b_;
if (a->first < b->first)
return -1;
else if (a->first == b->first)
return 0;
return 1;
}
/* Check if vectorization of the basic block is profitable for the
subgraph denoted by SLP_INSTANCES. */
static bool
vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo,
vec<slp_instance> slp_instances,
loop_p orig_loop)
{
slp_instance instance;
int i;
unsigned int vec_inside_cost = 0, vec_outside_cost = 0, scalar_cost = 0;
unsigned int vec_prologue_cost = 0, vec_epilogue_cost = 0;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location, "Costing subgraph: \n");
hash_set<slp_tree> visited;
FOR_EACH_VEC_ELT (slp_instances, i, instance)
vect_print_slp_graph (MSG_NOTE, vect_location,
SLP_INSTANCE_TREE (instance), visited);
}
/* Compute the set of scalar stmts we know will go away 'locally' when
vectorizing. This used to be tracked with just PURE_SLP_STMT but that's
not accurate for nodes promoted extern late or for scalar stmts that
are used both in extern defs and in vectorized defs. */
hash_set<stmt_vec_info> vectorized_scalar_stmts;
hash_set<stmt_vec_info> scalar_stmts_in_externs;
hash_set<slp_tree> visited;
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
vect_slp_gather_vectorized_scalar_stmts (bb_vinfo,
SLP_INSTANCE_TREE (instance),
visited,
vectorized_scalar_stmts,
scalar_stmts_in_externs);
for (stmt_vec_info rstmt : SLP_INSTANCE_ROOT_STMTS (instance))
vectorized_scalar_stmts.add (rstmt);
}
/* Scalar stmts used as defs in external nodes need to be preseved, so
remove them from vectorized_scalar_stmts. */
for (stmt_vec_info stmt : scalar_stmts_in_externs)
vectorized_scalar_stmts.remove (stmt);
/* Calculate scalar cost and sum the cost for the vector stmts
previously collected. */
stmt_vector_for_cost scalar_costs = vNULL;
stmt_vector_for_cost vector_costs = vNULL;
visited.empty ();
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
auto_vec<bool, 20> life;
life.safe_grow_cleared (SLP_TREE_LANES (SLP_INSTANCE_TREE (instance)),
true);
if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
record_stmt_cost (&scalar_costs,
SLP_INSTANCE_ROOT_STMTS (instance).length (),
scalar_stmt,
SLP_INSTANCE_ROOT_STMTS (instance)[0], 0, vect_body);
vect_bb_slp_scalar_cost (bb_vinfo,
SLP_INSTANCE_TREE (instance),
&life, &scalar_costs, vectorized_scalar_stmts,
visited);
vector_costs.safe_splice (instance->cost_vec);
instance->cost_vec.release ();
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n");
/* When costing non-loop vectorization we need to consider each covered
loop independently and make sure vectorization is profitable. For
now we assume a loop may be not entered or executed an arbitrary
number of iterations (??? static information can provide more
precise info here) which means we can simply cost each containing
loops stmts separately. */
/* First produce cost vectors sorted by loop index. */
auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
li_scalar_costs (scalar_costs.length ());
auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
li_vector_costs (vector_costs.length ());
stmt_info_for_cost *cost;
FOR_EACH_VEC_ELT (scalar_costs, i, cost)
{
unsigned l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
li_scalar_costs.quick_push (std::make_pair (l, cost));
}
/* Use a random used loop as fallback in case the first vector_costs
entry does not have a stmt_info associated with it. */
unsigned l = li_scalar_costs[0].first;
FOR_EACH_VEC_ELT (vector_costs, i, cost)
{
/* We inherit from the previous COST, invariants, externals and
extracts immediately follow the cost for the related stmt. */
if (cost->stmt_info)
l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
li_vector_costs.quick_push (std::make_pair (l, cost));
}
li_scalar_costs.qsort (li_cost_vec_cmp);
li_vector_costs.qsort (li_cost_vec_cmp);
/* Now cost the portions individually. */
unsigned vi = 0;
unsigned si = 0;
bool profitable = true;
while (si < li_scalar_costs.length ()
&& vi < li_vector_costs.length ())
{
unsigned sl = li_scalar_costs[si].first;
unsigned vl = li_vector_costs[vi].first;
if (sl != vl)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Scalar %d and vector %d loop part do not "
"match up, skipping scalar part\n", sl, vl);
/* Skip the scalar part, assuming zero cost on the vector side. */
do
{
si++;
}
while (si < li_scalar_costs.length ()
&& li_scalar_costs[si].first == sl);
continue;
}
class vector_costs *scalar_target_cost_data = init_cost (bb_vinfo, true);
do
{
add_stmt_cost (scalar_target_cost_data, li_scalar_costs[si].second);
si++;
}
while (si < li_scalar_costs.length ()
&& li_scalar_costs[si].first == sl);
unsigned dummy;
finish_cost (scalar_target_cost_data, nullptr,
&dummy, &scalar_cost, &dummy);
/* Complete the target-specific vector cost calculation. */
class vector_costs *vect_target_cost_data = init_cost (bb_vinfo, false);
do
{
add_stmt_cost (vect_target_cost_data, li_vector_costs[vi].second);
vi++;
}
while (vi < li_vector_costs.length ()
&& li_vector_costs[vi].first == vl);
finish_cost (vect_target_cost_data, scalar_target_cost_data,
&vec_prologue_cost, &vec_inside_cost, &vec_epilogue_cost);
delete scalar_target_cost_data;
delete vect_target_cost_data;
vec_outside_cost = vec_prologue_cost + vec_epilogue_cost;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Cost model analysis for part in loop %d:\n", sl);
dump_printf (MSG_NOTE, " Vector cost: %d\n",
vec_inside_cost + vec_outside_cost);
dump_printf (MSG_NOTE, " Scalar cost: %d\n", scalar_cost);
}
/* Vectorization is profitable if its cost is more than the cost of scalar
version. Note that we err on the vector side for equal cost because
the cost estimate is otherwise quite pessimistic (constant uses are
free on the scalar side but cost a load on the vector side for
example). */
if (vec_outside_cost + vec_inside_cost > scalar_cost)
{
profitable = false;
break;
}
}
if (profitable && vi < li_vector_costs.length ())
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Excess vector cost for part in loop %d:\n",
li_vector_costs[vi].first);
profitable = false;
}
/* Unset visited flag. This is delayed when the subgraph is profitable
and we process the loop for remaining unvectorized if-converted code. */
if (!orig_loop || !profitable)
FOR_EACH_VEC_ELT (scalar_costs, i, cost)
gimple_set_visited (cost->stmt_info->stmt, false);
scalar_costs.release ();
vector_costs.release ();
return profitable;
}
/* qsort comparator for lane defs. */
static int
vld_cmp (const void *a_, const void *b_)
{
auto *a = (const std::pair<unsigned, tree> *)a_;
auto *b = (const std::pair<unsigned, tree> *)b_;
return a->first - b->first;
}
/* Return true if USE_STMT is a vector lane insert into VEC and set
*THIS_LANE to the lane number that is set. */
static bool
vect_slp_is_lane_insert (gimple *use_stmt, tree vec, unsigned *this_lane)
{
gassign *use_ass = dyn_cast <gassign *> (use_stmt);
if (!use_ass
|| gimple_assign_rhs_code (use_ass) != BIT_INSERT_EXPR
|| (vec
? gimple_assign_rhs1 (use_ass) != vec
: ((vec = gimple_assign_rhs1 (use_ass)), false))
|| !useless_type_conversion_p (TREE_TYPE (TREE_TYPE (vec)),
TREE_TYPE (gimple_assign_rhs2 (use_ass)))
|| !constant_multiple_p
(tree_to_poly_uint64 (gimple_assign_rhs3 (use_ass)),
tree_to_poly_uint64 (TYPE_SIZE (TREE_TYPE (TREE_TYPE (vec)))),
this_lane))
return false;
return true;
}
/* Find any vectorizable constructors and add them to the grouped_store
array. */
static void
vect_slp_check_for_constructors (bb_vec_info bb_vinfo)
{
for (unsigned i = 0; i < bb_vinfo->bbs.length (); ++i)
for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[i]);
!gsi_end_p (gsi); gsi_next (&gsi))
{
gassign *assign = dyn_cast<gassign *> (gsi_stmt (gsi));
if (!assign)
continue;
tree rhs = gimple_assign_rhs1 (assign);
enum tree_code code = gimple_assign_rhs_code (assign);
use_operand_p use_p;
gimple *use_stmt;
if (code == CONSTRUCTOR)
{
if (!VECTOR_TYPE_P (TREE_TYPE (rhs))
|| maybe_ne (TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)),
CONSTRUCTOR_NELTS (rhs))
|| VECTOR_TYPE_P (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
|| uniform_vector_p (rhs))
continue;
unsigned j;
tree val;
FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (rhs), j, val)
if (TREE_CODE (val) != SSA_NAME
|| !bb_vinfo->lookup_def (val))
break;
if (j != CONSTRUCTOR_NELTS (rhs))
continue;
stmt_vec_info stmt_info = bb_vinfo->lookup_stmt (assign);
BB_VINFO_GROUPED_STORES (bb_vinfo).safe_push (stmt_info);
}
else if (code == BIT_INSERT_EXPR
&& VECTOR_TYPE_P (TREE_TYPE (rhs))
&& TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)).is_constant ()
&& TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)).to_constant () > 1
&& integer_zerop (gimple_assign_rhs3 (assign))
&& useless_type_conversion_p
(TREE_TYPE (TREE_TYPE (rhs)),
TREE_TYPE (gimple_assign_rhs2 (assign)))
&& bb_vinfo->lookup_def (gimple_assign_rhs2 (assign)))
{
/* We start to match on insert to lane zero but since the
inserts need not be ordered we'd have to search both
the def and the use chains. */
tree vectype = TREE_TYPE (rhs);
unsigned nlanes = TYPE_VECTOR_SUBPARTS (vectype).to_constant ();
auto_vec<std::pair<unsigned, tree> > lane_defs (nlanes);
auto_sbitmap lanes (nlanes);
bitmap_clear (lanes);
bitmap_set_bit (lanes, 0);
tree def = gimple_assign_lhs (assign);
lane_defs.quick_push
(std::make_pair (0, gimple_assign_rhs2 (assign)));
unsigned lanes_found = 1;
/* Start with the use chains, the last stmt will be the root. */
stmt_vec_info last = bb_vinfo->lookup_stmt (assign);
vec<stmt_vec_info> roots = vNULL;
roots.safe_push (last);
do
{
use_operand_p use_p;
gimple *use_stmt;
if (!single_imm_use (def, &use_p, &use_stmt))
break;
unsigned this_lane;
if (!bb_vinfo->lookup_stmt (use_stmt)
|| !vect_slp_is_lane_insert (use_stmt, def, &this_lane)
|| !bb_vinfo->lookup_def (gimple_assign_rhs2 (use_stmt)))
break;
if (bitmap_bit_p (lanes, this_lane))
break;
lanes_found++;
bitmap_set_bit (lanes, this_lane);
gassign *use_ass = as_a <gassign *> (use_stmt);
lane_defs.quick_push (std::make_pair
(this_lane, gimple_assign_rhs2 (use_ass)));
last = bb_vinfo->lookup_stmt (use_ass);
roots.safe_push (last);
def = gimple_assign_lhs (use_ass);
}
while (lanes_found < nlanes);
if (roots.length () > 1)
std::swap(roots[0], roots[roots.length () - 1]);
if (lanes_found < nlanes)
{
/* Now search the def chain. */
def = gimple_assign_rhs1 (assign);
do
{
if (TREE_CODE (def) != SSA_NAME
|| !has_single_use (def))
break;
gimple *def_stmt = SSA_NAME_DEF_STMT (def);
unsigned this_lane;
if (!bb_vinfo->lookup_stmt (def_stmt)
|| !vect_slp_is_lane_insert (def_stmt,
NULL_TREE, &this_lane)
|| !bb_vinfo->lookup_def (gimple_assign_rhs2 (def_stmt)))
break;
if (bitmap_bit_p (lanes, this_lane))
break;
lanes_found++;
bitmap_set_bit (lanes, this_lane);
lane_defs.quick_push (std::make_pair
(this_lane,
gimple_assign_rhs2 (def_stmt)));
roots.safe_push (bb_vinfo->lookup_stmt (def_stmt));
def = gimple_assign_rhs1 (def_stmt);
}
while (lanes_found < nlanes);
}
if (lanes_found == nlanes)
{
/* Sort lane_defs after the lane index and register the root. */
lane_defs.qsort (vld_cmp);
vec<stmt_vec_info> stmts;
stmts.create (nlanes);
for (unsigned i = 0; i < nlanes; ++i)
stmts.quick_push (bb_vinfo->lookup_def (lane_defs[i].second));
bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_ctor,
stmts, roots));
}
else
roots.release ();
}
else if (!VECTOR_TYPE_P (TREE_TYPE (rhs))
&& (associative_tree_code (code) || code == MINUS_EXPR)
/* ??? The flag_associative_math and TYPE_OVERFLOW_WRAPS
checks pessimize a two-element reduction. PR54400.
??? In-order reduction could be handled if we only
traverse one operand chain in vect_slp_linearize_chain. */
&& ((FLOAT_TYPE_P (TREE_TYPE (rhs)) && flag_associative_math)
|| (INTEGRAL_TYPE_P (TREE_TYPE (rhs))
&& TYPE_OVERFLOW_WRAPS (TREE_TYPE (rhs))))
/* Ops with constants at the tail can be stripped here. */
&& TREE_CODE (rhs) == SSA_NAME
&& TREE_CODE (gimple_assign_rhs2 (assign)) == SSA_NAME
/* Should be the chain end. */
&& (!single_imm_use (gimple_assign_lhs (assign),
&use_p, &use_stmt)
|| !is_gimple_assign (use_stmt)
|| (gimple_assign_rhs_code (use_stmt) != code
&& ((code != PLUS_EXPR && code != MINUS_EXPR)
|| (gimple_assign_rhs_code (use_stmt)
!= (code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR))))))
{
/* We start the match at the end of a possible association
chain. */
auto_vec<chain_op_t> chain;
auto_vec<std::pair<tree_code, gimple *> > worklist;
auto_vec<gimple *> chain_stmts;
gimple *code_stmt = NULL, *alt_code_stmt = NULL;
if (code == MINUS_EXPR)
code = PLUS_EXPR;
internal_fn reduc_fn;
if (!reduction_fn_for_scalar_code (code, &reduc_fn)
|| reduc_fn == IFN_LAST)
continue;
vect_slp_linearize_chain (bb_vinfo, worklist, chain, code, assign,
/* ??? */
code_stmt, alt_code_stmt, &chain_stmts);
if (chain.length () > 1)
{
/* Sort the chain according to def_type and operation. */
chain.sort (dt_sort_cmp, bb_vinfo);
/* ??? Now we'd want to strip externals and constants
but record those to be handled in the epilogue. */
/* ??? For now do not allow mixing ops or externs/constants. */
bool invalid = false;
for (unsigned i = 0; i < chain.length (); ++i)
if (chain[i].dt != vect_internal_def
|| chain[i].code != code)
invalid = true;
if (!invalid)
{
vec<stmt_vec_info> stmts;
stmts.create (chain.length ());
for (unsigned i = 0; i < chain.length (); ++i)
stmts.quick_push (bb_vinfo->lookup_def (chain[i].op));
vec<stmt_vec_info> roots;
roots.create (chain_stmts.length ());
for (unsigned i = 0; i < chain_stmts.length (); ++i)
roots.quick_push (bb_vinfo->lookup_stmt (chain_stmts[i]));
bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_bb_reduc,
stmts, roots));
}
}
}
}
}
/* Walk the grouped store chains and replace entries with their
pattern variant if any. */
static void
vect_fixup_store_groups_with_patterns (vec_info *vinfo)
{
stmt_vec_info first_element;
unsigned i;
FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element)
{
/* We also have CTORs in this array. */
if (!STMT_VINFO_GROUPED_ACCESS (first_element))
continue;
if (STMT_VINFO_IN_PATTERN_P (first_element))
{
stmt_vec_info orig = first_element;
first_element = STMT_VINFO_RELATED_STMT (first_element);
DR_GROUP_FIRST_ELEMENT (first_element) = first_element;
DR_GROUP_SIZE (first_element) = DR_GROUP_SIZE (orig);
DR_GROUP_GAP (first_element) = DR_GROUP_GAP (orig);
DR_GROUP_NEXT_ELEMENT (first_element) = DR_GROUP_NEXT_ELEMENT (orig);
vinfo->grouped_stores[i] = first_element;
}
stmt_vec_info prev = first_element;
while (DR_GROUP_NEXT_ELEMENT (prev))
{
stmt_vec_info elt = DR_GROUP_NEXT_ELEMENT (prev);
if (STMT_VINFO_IN_PATTERN_P (elt))
{
stmt_vec_info orig = elt;
elt = STMT_VINFO_RELATED_STMT (elt);
DR_GROUP_NEXT_ELEMENT (prev) = elt;
DR_GROUP_GAP (elt) = DR_GROUP_GAP (orig);
DR_GROUP_NEXT_ELEMENT (elt) = DR_GROUP_NEXT_ELEMENT (orig);
}
DR_GROUP_FIRST_ELEMENT (elt) = first_element;
prev = elt;
}
}
}
/* Check if the region described by BB_VINFO can be vectorized, returning
true if so. When returning false, set FATAL to true if the same failure
would prevent vectorization at other vector sizes, false if it is still
worth trying other sizes. N_STMTS is the number of statements in the
region. */
static bool
vect_slp_analyze_bb_1 (bb_vec_info bb_vinfo, int n_stmts, bool &fatal,
vec<int> *dataref_groups)
{
DUMP_VECT_SCOPE ("vect_slp_analyze_bb");
slp_instance instance;
int i;
poly_uint64 min_vf = 2;
/* The first group of checks is independent of the vector size. */
fatal = true;
/* Analyze the data references. */
if (!vect_analyze_data_refs (bb_vinfo, &min_vf, NULL))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: unhandled data-ref in basic "
"block.\n");
return false;
}
if (!vect_analyze_data_ref_accesses (bb_vinfo, dataref_groups))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: unhandled data access in "
"basic block.\n");
return false;
}
vect_slp_check_for_constructors (bb_vinfo);
/* If there are no grouped stores and no constructors in the region
there is no need to continue with pattern recog as vect_analyze_slp
will fail anyway. */
if (bb_vinfo->grouped_stores.is_empty ()
&& bb_vinfo->roots.is_empty ())
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: no grouped stores in "
"basic block.\n");
return false;
}
/* While the rest of the analysis below depends on it in some way. */
fatal = false;
vect_pattern_recog (bb_vinfo);
/* Update store groups from pattern processing. */
vect_fixup_store_groups_with_patterns (bb_vinfo);
/* Check the SLP opportunities in the basic block, analyze and build SLP
trees. */
if (!vect_analyze_slp (bb_vinfo, n_stmts))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Failed to SLP the basic block.\n");
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: failed to find SLP opportunities "
"in basic block.\n");
}
return false;
}
/* Optimize permutations. */
vect_optimize_slp (bb_vinfo);
/* Gather the loads reachable from the SLP graph entries. */
vect_gather_slp_loads (bb_vinfo);
vect_record_base_alignments (bb_vinfo);
/* Analyze and verify the alignment of data references and the
dependence in the SLP instances. */
for (i = 0; BB_VINFO_SLP_INSTANCES (bb_vinfo).iterate (i, &instance); )
{
vect_location = instance->location ();
if (! vect_slp_analyze_instance_alignment (bb_vinfo, instance)
|| ! vect_slp_analyze_instance_dependence (bb_vinfo, instance))
{
slp_tree node = SLP_INSTANCE_TREE (instance);
stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"removing SLP instance operations starting from: %G",
stmt_info->stmt);
vect_free_slp_instance (instance);
BB_VINFO_SLP_INSTANCES (bb_vinfo).ordered_remove (i);
continue;
}
/* Mark all the statements that we want to vectorize as pure SLP and
relevant. */
vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance));
vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance));
unsigned j;
stmt_vec_info root;
/* Likewise consider instance root stmts as vectorized. */
FOR_EACH_VEC_ELT (SLP_INSTANCE_ROOT_STMTS (instance), j, root)
STMT_SLP_TYPE (root) = pure_slp;
i++;
}
if (! BB_VINFO_SLP_INSTANCES (bb_vinfo).length ())
return false;
if (!vect_slp_analyze_operations (bb_vinfo))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: bad operation in basic block.\n");
return false;
}
vect_bb_partition_graph (bb_vinfo);
return true;
}
/* Subroutine of vect_slp_bb. Try to vectorize the statements for all
basic blocks in BBS, returning true on success.
The region has N_STMTS statements and has the datarefs given by DATAREFS. */
static bool
vect_slp_region (vec<basic_block> bbs, vec<data_reference_p> datarefs,
vec<int> *dataref_groups, unsigned int n_stmts,
loop_p orig_loop)
{
bb_vec_info bb_vinfo;
auto_vector_modes vector_modes;
/* Autodetect first vector size we try. */
machine_mode next_vector_mode = VOIDmode;
targetm.vectorize.autovectorize_vector_modes (&vector_modes, false);
unsigned int mode_i = 0;
vec_info_shared shared;
machine_mode autodetected_vector_mode = VOIDmode;
while (1)
{
bool vectorized = false;
bool fatal = false;
bb_vinfo = new _bb_vec_info (bbs, &shared);
bool first_time_p = shared.datarefs.is_empty ();
BB_VINFO_DATAREFS (bb_vinfo) = datarefs;
if (first_time_p)
bb_vinfo->shared->save_datarefs ();
else
bb_vinfo->shared->check_datarefs ();
bb_vinfo->vector_mode = next_vector_mode;
if (vect_slp_analyze_bb_1 (bb_vinfo, n_stmts, fatal, dataref_groups))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"***** Analysis succeeded with vector mode"
" %s\n", GET_MODE_NAME (bb_vinfo->vector_mode));
dump_printf_loc (MSG_NOTE, vect_location, "SLPing BB part\n");
}
bb_vinfo->shared->check_datarefs ();
auto_vec<slp_instance> profitable_subgraphs;
for (slp_instance instance : BB_VINFO_SLP_INSTANCES (bb_vinfo))
{
if (instance->subgraph_entries.is_empty ())
continue;
vect_location = instance->location ();
if (!unlimited_cost_model (NULL)
&& !vect_bb_vectorization_profitable_p
(bb_vinfo, instance->subgraph_entries, orig_loop))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: vectorization is not "
"profitable.\n");
continue;
}
if (!dbg_cnt (vect_slp))
continue;
profitable_subgraphs.safe_push (instance);
}
/* When we're vectorizing an if-converted loop body make sure
we vectorized all if-converted code. */
if (!profitable_subgraphs.is_empty ()
&& orig_loop)
{
gcc_assert (bb_vinfo->bbs.length () == 1);
for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[0]);
!gsi_end_p (gsi); gsi_next (&gsi))
{
/* The costing above left us with DCEable vectorized scalar
stmts having the visited flag set on profitable
subgraphs. Do the delayed clearing of the flag here. */
if (gimple_visited_p (gsi_stmt (gsi)))
{
gimple_set_visited (gsi_stmt (gsi), false);
continue;
}
if (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED)
continue;
if (gassign *ass = dyn_cast <gassign *> (gsi_stmt (gsi)))
if (gimple_assign_rhs_code (ass) == COND_EXPR)
{
if (!profitable_subgraphs.is_empty ()
&& dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"not profitable because of "
"unprofitable if-converted scalar "
"code\n");
profitable_subgraphs.truncate (0);
}
}
}
/* Finally schedule the profitable subgraphs. */
for (slp_instance instance : profitable_subgraphs)
{
if (!vectorized && dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Basic block will be vectorized "
"using SLP\n");
vectorized = true;
vect_schedule_slp (bb_vinfo, instance->subgraph_entries);
unsigned HOST_WIDE_INT bytes;
if (dump_enabled_p ())
{
if (GET_MODE_SIZE
(bb_vinfo->vector_mode).is_constant (&bytes))
dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
"basic block part vectorized using %wu "
"byte vectors\n", bytes);
else
dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
"basic block part vectorized using "
"variable length vectors\n");
}
}
}
else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"***** Analysis failed with vector mode %s\n",
GET_MODE_NAME (bb_vinfo->vector_mode));
}
if (mode_i == 0)
autodetected_vector_mode = bb_vinfo->vector_mode;
if (!fatal)
while (mode_i < vector_modes.length ()
&& vect_chooses_same_modes_p (bb_vinfo, vector_modes[mode_i]))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"***** The result for vector mode %s would"
" be the same\n",
GET_MODE_NAME (vector_modes[mode_i]));
mode_i += 1;
}
delete bb_vinfo;
if (mode_i < vector_modes.length ()
&& VECTOR_MODE_P (autodetected_vector_mode)
&& (related_vector_mode (vector_modes[mode_i],
GET_MODE_INNER (autodetected_vector_mode))
== autodetected_vector_mode)
&& (related_vector_mode (autodetected_vector_mode,
GET_MODE_INNER (vector_modes[mode_i]))
== vector_modes[mode_i]))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"***** Skipping vector mode %s, which would"
" repeat the analysis for %s\n",
GET_MODE_NAME (vector_modes[mode_i]),
GET_MODE_NAME (autodetected_vector_mode));
mode_i += 1;
}
if (vectorized
|| mode_i == vector_modes.length ()
|| autodetected_vector_mode == VOIDmode
/* If vect_slp_analyze_bb_1 signaled that analysis for all
vector sizes will fail do not bother iterating. */
|| fatal)
return vectorized;
/* Try the next biggest vector size. */
next_vector_mode = vector_modes[mode_i++];
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"***** Re-trying analysis with vector mode %s\n",
GET_MODE_NAME (next_vector_mode));
}
}
/* Main entry for the BB vectorizer. Analyze and transform BBS, returns
true if anything in the basic-block was vectorized. */
static bool
vect_slp_bbs (const vec<basic_block> &bbs, loop_p orig_loop)
{
vec<data_reference_p> datarefs = vNULL;
auto_vec<int> dataref_groups;
int insns = 0;
int current_group = 0;
for (unsigned i = 0; i < bbs.length (); i++)
{
basic_block bb = bbs[i];
for (gimple_stmt_iterator gsi = gsi_after_labels (bb); !gsi_end_p (gsi);
gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (is_gimple_debug (stmt))
continue;
insns++;
if (gimple_location (stmt) != UNKNOWN_LOCATION)
vect_location = stmt;
if (!vect_find_stmt_data_reference (NULL, stmt, &datarefs,
&dataref_groups, current_group))
++current_group;
}
/* New BBs always start a new DR group. */
++current_group;
}
return vect_slp_region (bbs, datarefs, &dataref_groups, insns, orig_loop);
}
/* Special entry for the BB vectorizer. Analyze and transform a single
if-converted BB with ORIG_LOOPs body being the not if-converted
representation. Returns true if anything in the basic-block was
vectorized. */
bool
vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop)
{
auto_vec<basic_block> bbs;
bbs.safe_push (bb);
return vect_slp_bbs (bbs, orig_loop);
}
/* Main entry for the BB vectorizer. Analyze and transform BB, returns
true if anything in the basic-block was vectorized. */
bool
vect_slp_function (function *fun)
{
bool r = false;
int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fun));
unsigned n = pre_and_rev_post_order_compute_fn (fun, NULL, rpo, false);
/* For the moment split the function into pieces to avoid making
the iteration on the vector mode moot. Split at points we know
to not handle well which is CFG merges (SLP discovery doesn't
handle non-loop-header PHIs) and loop exits. Since pattern
recog requires reverse iteration to visit uses before defs
simply chop RPO into pieces. */
auto_vec<basic_block> bbs;
for (unsigned i = 0; i < n; i++)
{
basic_block bb = BASIC_BLOCK_FOR_FN (fun, rpo[i]);
bool split = false;
/* Split when a BB is not dominated by the first block. */
if (!bbs.is_empty ()
&& !dominated_by_p (CDI_DOMINATORS, bb, bbs[0]))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"splitting region at dominance boundary bb%d\n",
bb->index);
split = true;
}
/* Split when the loop determined by the first block
is exited. This is because we eventually insert
invariants at region begin. */
else if (!bbs.is_empty ()
&& bbs[0]->loop_father != bb->loop_father
&& !flow_loop_nested_p (bbs[0]->loop_father, bb->loop_father))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"splitting region at loop %d exit at bb%d\n",
bbs[0]->loop_father->num, bb->index);
split = true;
}
if (split && !bbs.is_empty ())
{
r |= vect_slp_bbs (bbs, NULL);
bbs.truncate (0);
bbs.quick_push (bb);
}
else
bbs.safe_push (bb);
/* When we have a stmt ending this block and defining a
value we have to insert on edges when inserting after it for
a vector containing its definition. Avoid this for now. */
if (gimple *last = last_stmt (bb))
if (gimple_get_lhs (last)
&& is_ctrl_altering_stmt (last))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"splitting region at control altering "
"definition %G", last);
r |= vect_slp_bbs (bbs, NULL);
bbs.truncate (0);
}
}
if (!bbs.is_empty ())
r |= vect_slp_bbs (bbs, NULL);
free (rpo);
return r;
}
/* Build a variable-length vector in which the elements in ELTS are repeated
to a fill NRESULTS vectors of type VECTOR_TYPE. Store the vectors in
RESULTS and add any new instructions to SEQ.
The approach we use is:
(1) Find a vector mode VM with integer elements of mode IM.
(2) Replace ELTS[0:NELTS] with ELTS'[0:NELTS'], where each element of
ELTS' has mode IM. This involves creating NELTS' VIEW_CONVERT_EXPRs
from small vectors to IM.
(3) Duplicate each ELTS'[I] into a vector of mode VM.
(4) Use a tree of interleaving VEC_PERM_EXPRs to create VMs with the
correct byte contents.
(5) Use VIEW_CONVERT_EXPR to cast the final VMs to the required type.
We try to find the largest IM for which this sequence works, in order
to cut down on the number of interleaves. */
void
duplicate_and_interleave (vec_info *vinfo, gimple_seq *seq, tree vector_type,
const vec<tree> &elts, unsigned int nresults,
vec<tree> &results)
{
unsigned int nelts = elts.length ();
tree element_type = TREE_TYPE (vector_type);
/* (1) Find a vector mode VM with integer elements of mode IM. */
unsigned int nvectors = 1;
tree new_vector_type;
tree permutes[2];
if (!can_duplicate_and_interleave_p (vinfo, nelts, element_type,
&nvectors, &new_vector_type,
permutes))
gcc_unreachable ();
/* Get a vector type that holds ELTS[0:NELTS/NELTS']. */
unsigned int partial_nelts = nelts / nvectors;
tree partial_vector_type = build_vector_type (element_type, partial_nelts);
tree_vector_builder partial_elts;
auto_vec<tree, 32> pieces (nvectors * 2);
pieces.quick_grow_cleared (nvectors * 2);
for (unsigned int i = 0; i < nvectors; ++i)
{
/* (2) Replace ELTS[0:NELTS] with ELTS'[0:NELTS'], where each element of
ELTS' has mode IM. */
partial_elts.new_vector (partial_vector_type, partial_nelts, 1);
for (unsigned int j = 0; j < partial_nelts; ++j)
partial_elts.quick_push (elts[i * partial_nelts + j]);
tree t = gimple_build_vector (seq, &partial_elts);
t = gimple_build (seq, VIEW_CONVERT_EXPR,
TREE_TYPE (new_vector_type), t);
/* (3) Duplicate each ELTS'[I] into a vector of mode VM. */
pieces[i] = gimple_build_vector_from_val (seq, new_vector_type, t);
}
/* (4) Use a tree of VEC_PERM_EXPRs to create a single VM with the
correct byte contents.
Conceptually, we need to repeat the following operation log2(nvectors)
times, where hi_start = nvectors / 2:
out[i * 2] = VEC_PERM_EXPR (in[i], in[i + hi_start], lo_permute);
out[i * 2 + 1] = VEC_PERM_EXPR (in[i], in[i + hi_start], hi_permute);
However, if each input repeats every N elements and the VF is
a multiple of N * 2, the HI result is the same as the LO result.
This will be true for the first N1 iterations of the outer loop,
followed by N2 iterations for which both the LO and HI results
are needed. I.e.:
N1 + N2 = log2(nvectors)
Each "N1 iteration" doubles the number of redundant vectors and the
effect of the process as a whole is to have a sequence of nvectors/2**N1
vectors that repeats 2**N1 times. Rather than generate these redundant
vectors, we halve the number of vectors for each N1 iteration. */
unsigned int in_start = 0;
unsigned int out_start = nvectors;
unsigned int new_nvectors = nvectors;
for (unsigned int in_repeat = 1; in_repeat < nvectors; in_repeat *= 2)
{
unsigned int hi_start = new_nvectors / 2;
unsigned int out_i = 0;
for (unsigned int in_i = 0; in_i < new_nvectors; ++in_i)
{
if ((in_i & 1) != 0
&& multiple_p (TYPE_VECTOR_SUBPARTS (new_vector_type),
2 * in_repeat))
continue;
tree output = make_ssa_name (new_vector_type);
tree input1 = pieces[in_start + (in_i / 2)];
tree input2 = pieces[in_start + (in_i / 2) + hi_start];
gassign *stmt = gimple_build_assign (output, VEC_PERM_EXPR,
input1, input2,
permutes[in_i & 1]);
gimple_seq_add_stmt (seq, stmt);
pieces[out_start + out_i] = output;
out_i += 1;
}
std::swap (in_start, out_start);
new_nvectors = out_i;
}
/* (5) Use VIEW_CONVERT_EXPR to cast the final VM to the required type. */
results.reserve (nresults);
for (unsigned int i = 0; i < nresults; ++i)
if (i < new_nvectors)
results.quick_push (gimple_build (seq, VIEW_CONVERT_EXPR, vector_type,
pieces[in_start + i]));
else
results.quick_push (results[i - new_nvectors]);
}
/* For constant and loop invariant defs in OP_NODE this function creates
vector defs that will be used in the vectorized stmts and stores them
to SLP_TREE_VEC_DEFS of OP_NODE. */
static void
vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
{
unsigned HOST_WIDE_INT nunits;
tree vec_cst;
unsigned j, number_of_places_left_in_vector;
tree vector_type;
tree vop;
int group_size = op_node->ops.length ();
unsigned int vec_num, i;
unsigned number_of_copies = 1;
bool constant_p;
gimple_seq ctor_seq = NULL;
auto_vec<tree, 16> permute_results;
/* We always want SLP_TREE_VECTYPE (op_node) here correctly set. */
vector_type = SLP_TREE_VECTYPE (op_node);
unsigned int number_of_vectors = SLP_TREE_NUMBER_OF_VEC_STMTS (op_node);
SLP_TREE_VEC_DEFS (op_node).create (number_of_vectors);
auto_vec<tree> voprnds (number_of_vectors);
/* NUMBER_OF_COPIES is the number of times we need to use the same values in
created vectors. It is greater than 1 if unrolling is performed.
For example, we have two scalar operands, s1 and s2 (e.g., group of
strided accesses of size two), while NUNITS is four (i.e., four scalars
of this type can be packed in a vector). The output vector will contain
two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
will be 2).
If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
containing the operands.
For example, NUNITS is four as before, and the group size is 8
(s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
{s5, s6, s7, s8}. */
/* When using duplicate_and_interleave, we just need one element for
each scalar statement. */
if (!TYPE_VECTOR_SUBPARTS (vector_type).is_constant (&nunits))
nunits = group_size;
number_of_copies = nunits * number_of_vectors / group_size;
number_of_places_left_in_vector = nunits;
constant_p = true;
tree_vector_builder elts (vector_type, nunits, 1);
elts.quick_grow (nunits);
stmt_vec_info insert_after = NULL;
for (j = 0; j < number_of_copies; j++)
{
tree op;
for (i = group_size - 1; op_node->ops.iterate (i, &op); i--)
{
/* Create 'vect_ = {op0,op1,...,opn}'. */
number_of_places_left_in_vector--;
tree orig_op = op;
if (!types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op)))
{
if (CONSTANT_CLASS_P (op))
{
if (VECTOR_BOOLEAN_TYPE_P (vector_type))
{
/* Can't use VIEW_CONVERT_EXPR for booleans because
of possibly different sizes of scalar value and
vector element. */
if (integer_zerop (op))
op = build_int_cst (TREE_TYPE (vector_type), 0);
else if (integer_onep (op))
op = build_all_ones_cst (TREE_TYPE (vector_type));
else
gcc_unreachable ();
}
else
op = fold_unary (VIEW_CONVERT_EXPR,
TREE_TYPE (vector_type), op);
gcc_assert (op && CONSTANT_CLASS_P (op));
}
else
{
tree new_temp = make_ssa_name (TREE_TYPE (vector_type));
gimple *init_stmt;
if (VECTOR_BOOLEAN_TYPE_P (vector_type))
{
tree true_val
= build_all_ones_cst (TREE_TYPE (vector_type));
tree false_val
= build_zero_cst (TREE_TYPE (vector_type));
gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op)));
init_stmt = gimple_build_assign (new_temp, COND_EXPR,
op, true_val,
false_val);
}
else
{
op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type),
op);
init_stmt
= gimple_build_assign (new_temp, VIEW_CONVERT_EXPR,
op);
}
gimple_seq_add_stmt (&ctor_seq, init_stmt);
op = new_temp;
}
}
elts[number_of_places_left_in_vector] = op;
if (!CONSTANT_CLASS_P (op))
constant_p = false;
/* For BB vectorization we have to compute an insert location
when a def is inside the analyzed region since we cannot
simply insert at the BB start in this case. */
stmt_vec_info opdef;
if (TREE_CODE (orig_op) == SSA_NAME
&& !SSA_NAME_IS_DEFAULT_DEF (orig_op)
&& is_a <bb_vec_info> (vinfo)
&& (opdef = vinfo->lookup_def (orig_op)))
{
if (!insert_after)
insert_after = opdef;
else
insert_after = get_later_stmt (insert_after, opdef);
}
if (number_of_places_left_in_vector == 0)
{
if (constant_p
? multiple_p (TYPE_VECTOR_SUBPARTS (vector_type), nunits)
: known_eq (TYPE_VECTOR_SUBPARTS (vector_type), nunits))
vec_cst = gimple_build_vector (&ctor_seq, &elts);
else
{
if (permute_results.is_empty ())
duplicate_and_interleave (vinfo, &ctor_seq, vector_type,
elts, number_of_vectors,
permute_results);
vec_cst = permute_results[number_of_vectors - j - 1];
}
if (!gimple_seq_empty_p (ctor_seq))
{
if (insert_after)
{
gimple_stmt_iterator gsi;
if (gimple_code (insert_after->stmt) == GIMPLE_PHI)
{
gsi = gsi_after_labels (gimple_bb (insert_after->stmt));
gsi_insert_seq_before (&gsi, ctor_seq,
GSI_CONTINUE_LINKING);
}
else if (!stmt_ends_bb_p (insert_after->stmt))
{
gsi = gsi_for_stmt (insert_after->stmt);
gsi_insert_seq_after (&gsi, ctor_seq,
GSI_CONTINUE_LINKING);
}
else
{
/* When we want to insert after a def where the
defining stmt throws then insert on the fallthru
edge. */
edge e = find_fallthru_edge
(gimple_bb (insert_after->stmt)->succs);
basic_block new_bb
= gsi_insert_seq_on_edge_immediate (e, ctor_seq);
gcc_assert (!new_bb);
}
}
else
vinfo->insert_seq_on_entry (NULL, ctor_seq);
ctor_seq = NULL;
}
voprnds.quick_push (vec_cst);
insert_after = NULL;
number_of_places_left_in_vector = nunits;
constant_p = true;
elts.new_vector (vector_type, nunits, 1);
elts.quick_grow (nunits);
}
}
}
/* Since the vectors are created in the reverse order, we should invert
them. */
vec_num = voprnds.length ();
for (j = vec_num; j != 0; j--)
{
vop = voprnds[j - 1];
SLP_TREE_VEC_DEFS (op_node).quick_push (vop);
}
/* In case that VF is greater than the unrolling factor needed for the SLP
group of stmts, NUMBER_OF_VECTORS to be created is greater than
NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
to replicate the vectors. */
while (number_of_vectors > SLP_TREE_VEC_DEFS (op_node).length ())
for (i = 0; SLP_TREE_VEC_DEFS (op_node).iterate (i, &vop) && i < vec_num;
i++)
SLP_TREE_VEC_DEFS (op_node).quick_push (vop);
}
/* Get the Ith vectorized definition from SLP_NODE. */
tree
vect_get_slp_vect_def (slp_tree slp_node, unsigned i)
{
if (SLP_TREE_VEC_STMTS (slp_node).exists ())
return gimple_get_lhs (SLP_TREE_VEC_STMTS (slp_node)[i]);
else
return SLP_TREE_VEC_DEFS (slp_node)[i];
}
/* Get the vectorized definitions of SLP_NODE in *VEC_DEFS. */
void
vect_get_slp_defs (slp_tree slp_node, vec<tree> *vec_defs)
{
vec_defs->create (SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node));
if (SLP_TREE_DEF_TYPE (slp_node) == vect_internal_def)
{
unsigned j;
gimple *vec_def_stmt;
FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (slp_node), j, vec_def_stmt)
vec_defs->quick_push (gimple_get_lhs (vec_def_stmt));
}
else
vec_defs->splice (SLP_TREE_VEC_DEFS (slp_node));
}
/* Get N vectorized definitions for SLP_NODE. */
void
vect_get_slp_defs (vec_info *,
slp_tree slp_node, vec<vec<tree> > *vec_oprnds, unsigned n)
{
if (n == -1U)
n = SLP_TREE_CHILDREN (slp_node).length ();
for (unsigned i = 0; i < n; ++i)
{
slp_tree child = SLP_TREE_CHILDREN (slp_node)[i];
vec<tree> vec_defs = vNULL;
vect_get_slp_defs (child, &vec_defs);
vec_oprnds->quick_push (vec_defs);
}
}
/* A subroutine of vect_transform_slp_perm_load with two extra arguments:
- PERM gives the permutation that the caller wants to use for NODE,
which might be different from SLP_LOAD_PERMUTATION.
- DUMP_P controls whether the function dumps information. */
static bool
vect_transform_slp_perm_load_1 (vec_info *vinfo, slp_tree node,
load_permutation_t &perm,
const vec<tree> &dr_chain,
gimple_stmt_iterator *gsi, poly_uint64 vf,
bool analyze_only, bool dump_p,
unsigned *n_perms, unsigned int *n_loads,
bool dce_chain)
{
stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
int vec_index = 0;
tree vectype = SLP_TREE_VECTYPE (node);
unsigned int group_size = SLP_TREE_SCALAR_STMTS (node).length ();
unsigned int mask_element;
machine_mode mode;
if (!STMT_VINFO_GROUPED_ACCESS (stmt_info))
return false;
stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
mode = TYPE_MODE (vectype);
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
/* Initialize the vect stmts of NODE to properly insert the generated
stmts later. */
if (! analyze_only)
for (unsigned i = SLP_TREE_VEC_STMTS (node).length ();
i < SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++)
SLP_TREE_VEC_STMTS (node).quick_push (NULL);
/* Generate permutation masks for every NODE. Number of masks for each NODE
is equal to GROUP_SIZE.
E.g., we have a group of three nodes with three loads from the same
location in each node, and the vector size is 4. I.e., we have a
a0b0c0a1b1c1... sequence and we need to create the following vectors:
for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3
for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3
...
The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}.
The last mask is illegal since we assume two operands for permute
operation, and the mask element values can't be outside that range.
Hence, the last mask must be converted into {2,5,5,5}.
For the first two permutations we need the first and the second input
vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation
we need the second and the third vectors: {b1,c1,a2,b2} and
{c2,a3,b3,c3}. */
int vect_stmts_counter = 0;
unsigned int index = 0;
int first_vec_index = -1;
int second_vec_index = -1;
bool noop_p = true;
*n_perms = 0;
vec_perm_builder mask;
unsigned int nelts_to_build;
unsigned int nvectors_per_build;
unsigned int in_nlanes;
bool repeating_p = (group_size == DR_GROUP_SIZE (stmt_info)
&& multiple_p (nunits, group_size));
if (repeating_p)
{
/* A single vector contains a whole number of copies of the node, so:
(a) all permutes can use the same mask; and
(b) the permutes only need a single vector input. */
mask.new_vector (nunits, group_size, 3);
nelts_to_build = mask.encoded_nelts ();
nvectors_per_build = SLP_TREE_VEC_STMTS (node).length ();
in_nlanes = DR_GROUP_SIZE (stmt_info) * 3;
}
else
{
/* We need to construct a separate mask for each vector statement. */
unsigned HOST_WIDE_INT const_nunits, const_vf;
if (!nunits.is_constant (&const_nunits)
|| !vf.is_constant (&const_vf))
return false;
mask.new_vector (const_nunits, const_nunits, 1);
nelts_to_build = const_vf * group_size;
nvectors_per_build = 1;
in_nlanes = const_vf * DR_GROUP_SIZE (stmt_info);
}
auto_sbitmap used_in_lanes (in_nlanes);
bitmap_clear (used_in_lanes);
auto_bitmap used_defs;
unsigned int count = mask.encoded_nelts ();
mask.quick_grow (count);
vec_perm_indices indices;
for (unsigned int j = 0; j < nelts_to_build; j++)
{
unsigned int iter_num = j / group_size;
unsigned int stmt_num = j % group_size;
unsigned int i = (iter_num * DR_GROUP_SIZE (stmt_info) + perm[stmt_num]);
bitmap_set_bit (used_in_lanes, i);
if (repeating_p)
{
first_vec_index = 0;
mask_element = i;
}
else
{
/* Enforced before the loop when !repeating_p. */
unsigned int const_nunits = nunits.to_constant ();
vec_index = i / const_nunits;
mask_element = i % const_nunits;
if (vec_index == first_vec_index
|| first_vec_index == -1)
{
first_vec_index = vec_index;
}
else if (vec_index == second_vec_index
|| second_vec_index == -1)
{
second_vec_index = vec_index;
mask_element += const_nunits;
}
else
{
if (dump_p)
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"permutation requires at "
"least three vectors %G",
stmt_info->stmt);
gcc_assert (analyze_only);
return false;
}
gcc_assert (mask_element < 2 * const_nunits);
}
if (mask_element != index)
noop_p = false;
mask[index++] = mask_element;
if (index == count && !noop_p)
{
indices.new_vector (mask, second_vec_index == -1 ? 1 : 2, nunits);
if (!can_vec_perm_const_p (mode, mode, indices))
{
if (dump_p)
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION,
vect_location,
"unsupported vect permute { ");
for (i = 0; i < count; ++i)
{
dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]);
dump_printf (MSG_MISSED_OPTIMIZATION, " ");
}
dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
}
gcc_assert (analyze_only);
return false;
}
++*n_perms;
}
if (index == count)
{
if (!analyze_only)
{
tree mask_vec = NULL_TREE;
if (! noop_p)
mask_vec = vect_gen_perm_mask_checked (vectype, indices);
if (second_vec_index == -1)
second_vec_index = first_vec_index;
for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
{
/* Generate the permute statement if necessary. */
tree first_vec = dr_chain[first_vec_index + ri];
tree second_vec = dr_chain[second_vec_index + ri];
gimple *perm_stmt;
if (! noop_p)
{
gassign *stmt = as_a <gassign *> (stmt_info->stmt);
tree perm_dest
= vect_create_destination_var (gimple_assign_lhs (stmt),
vectype);
perm_dest = make_ssa_name (perm_dest);
perm_stmt
= gimple_build_assign (perm_dest, VEC_PERM_EXPR,
first_vec, second_vec,
mask_vec);
vect_finish_stmt_generation (vinfo, stmt_info, perm_stmt,
gsi);
if (dce_chain)
{
bitmap_set_bit (used_defs, first_vec_index + ri);
bitmap_set_bit (used_defs, second_vec_index + ri);
}
}
else
{
/* If mask was NULL_TREE generate the requested
identity transform. */
perm_stmt = SSA_NAME_DEF_STMT (first_vec);
if (dce_chain)
bitmap_set_bit (used_defs, first_vec_index + ri);
}
/* Store the vector statement in NODE. */
SLP_TREE_VEC_STMTS (node)[vect_stmts_counter++] = perm_stmt;
}
}
index = 0;
first_vec_index = -1;
second_vec_index = -1;
noop_p = true;
}
}
if (n_loads)
{
if (repeating_p)
*n_loads = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
else
{
/* Enforced above when !repeating_p. */
unsigned int const_nunits = nunits.to_constant ();
*n_loads = 0;
bool load_seen = false;
for (unsigned i = 0; i < in_nlanes; ++i)
{
if (i % const_nunits == 0)
{
if (load_seen)
*n_loads += 1;
load_seen = false;
}
if (bitmap_bit_p (used_in_lanes, i))
load_seen = true;
}
if (load_seen)
*n_loads += 1;
}
}
if (dce_chain)
for (unsigned i = 0; i < dr_chain.length (); ++i)
if (!bitmap_bit_p (used_defs, i))
{
gimple *stmt = SSA_NAME_DEF_STMT (dr_chain[i]);
gimple_stmt_iterator rgsi = gsi_for_stmt (stmt);
gsi_remove (&rgsi, true);
release_defs (stmt);
}
return true;
}
/* Generate vector permute statements from a list of loads in DR_CHAIN.
If ANALYZE_ONLY is TRUE, only check that it is possible to create valid
permute statements for the SLP node NODE. Store the number of vector
permute instructions in *N_PERMS and the number of vector load
instructions in *N_LOADS. If DCE_CHAIN is true, remove all definitions
that were not needed. */
bool
vect_transform_slp_perm_load (vec_info *vinfo,
slp_tree node, const vec<tree> &dr_chain,
gimple_stmt_iterator *gsi, poly_uint64 vf,
bool analyze_only, unsigned *n_perms,
unsigned int *n_loads, bool dce_chain)
{
return vect_transform_slp_perm_load_1 (vinfo, node,
SLP_TREE_LOAD_PERMUTATION (node),
dr_chain, gsi, vf, analyze_only,
dump_enabled_p (), n_perms, n_loads,
dce_chain);
}
/* Produce the next vector result for SLP permutation NODE by adding a vector
statement at GSI. If MASK_VEC is nonnull, add:
<new SSA name> = VEC_PERM_EXPR <FIRST_DEF, SECOND_DEF, MASK_VEC>
otherwise add:
<new SSA name> = FIRST_DEF. */
static void
vect_add_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
slp_tree node, tree first_def, tree second_def,
tree mask_vec)
{
tree vectype = SLP_TREE_VECTYPE (node);
/* ??? We SLP match existing vector element extracts but
allow punning which we need to re-instantiate at uses
but have no good way of explicitly representing. */
if (operand_equal_p (TYPE_SIZE (TREE_TYPE (first_def)), TYPE_SIZE (vectype))
&& !types_compatible_p (TREE_TYPE (first_def), vectype))
{
gassign *conv_stmt
= gimple_build_assign (make_ssa_name (vectype),
build1 (VIEW_CONVERT_EXPR, vectype, first_def));
vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
first_def = gimple_assign_lhs (conv_stmt);
}
gassign *perm_stmt;
tree perm_dest = make_ssa_name (vectype);
if (mask_vec)
{
if (operand_equal_p (TYPE_SIZE (TREE_TYPE (first_def)),
TYPE_SIZE (vectype))
&& !types_compatible_p (TREE_TYPE (second_def), vectype))
{
gassign *conv_stmt
= gimple_build_assign (make_ssa_name (vectype),
build1 (VIEW_CONVERT_EXPR,
vectype, second_def));
vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
second_def = gimple_assign_lhs (conv_stmt);
}
perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
first_def, second_def,
mask_vec);
}
else if (!types_compatible_p (TREE_TYPE (first_def), vectype))
{
/* For identity permutes we still need to handle the case
of lowpart extracts or concats. */
unsigned HOST_WIDE_INT c;
auto first_def_nunits
= TYPE_VECTOR_SUBPARTS (TREE_TYPE (first_def));
if (known_le (TYPE_VECTOR_SUBPARTS (vectype), first_def_nunits))
{
tree lowpart = build3 (BIT_FIELD_REF, vectype, first_def,
TYPE_SIZE (vectype), bitsize_zero_node);
perm_stmt = gimple_build_assign (perm_dest, lowpart);
}
else if (constant_multiple_p (TYPE_VECTOR_SUBPARTS (vectype),
first_def_nunits, &c) && c == 2)
{
tree ctor = build_constructor_va (vectype, 2, NULL_TREE, first_def,
NULL_TREE, second_def);
perm_stmt = gimple_build_assign (perm_dest, ctor);
}
else
gcc_unreachable ();
}
else
{
/* We need a copy here in case the def was external. */
perm_stmt = gimple_build_assign (perm_dest, first_def);
}
vect_finish_stmt_generation (vinfo, NULL, perm_stmt, gsi);
/* Store the vector statement in NODE. */
SLP_TREE_VEC_STMTS (node).quick_push (perm_stmt);
}
/* Subroutine of vectorizable_slp_permutation. Check whether the target
can perform permutation PERM on the (1 or 2) input nodes in CHILDREN.
If GSI is nonnull, emit the permutation there.
When GSI is null, the only purpose of NODE is to give properties
of the result, such as the vector type and number of SLP lanes.
The node does not need to be a VEC_PERM_EXPR.
If the target supports the operation, return the number of individual
VEC_PERM_EXPRs needed, otherwise return -1. Print information to the
dump file if DUMP_P is true. */
static int
vectorizable_slp_permutation_1 (vec_info *vinfo, gimple_stmt_iterator *gsi,
slp_tree node, lane_permutation_t &perm,
vec<slp_tree> &children, bool dump_p)
{
tree vectype = SLP_TREE_VECTYPE (node);
/* ??? We currently only support all same vector input types
while the SLP IL should really do a concat + select and thus accept
arbitrary mismatches. */
slp_tree child;
unsigned i;
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
bool repeating_p = multiple_p (nunits, SLP_TREE_LANES (node));
tree op_vectype = NULL_TREE;
FOR_EACH_VEC_ELT (children, i, child)
if (SLP_TREE_VECTYPE (child))
{
op_vectype = SLP_TREE_VECTYPE (child);
break;
}
if (!op_vectype)
op_vectype = vectype;
FOR_EACH_VEC_ELT (children, i, child)
{
if ((SLP_TREE_DEF_TYPE (child) != vect_internal_def
&& !vect_maybe_update_slp_op_vectype (child, op_vectype))
|| !types_compatible_p (SLP_TREE_VECTYPE (child), op_vectype)
|| !types_compatible_p (TREE_TYPE (vectype), TREE_TYPE (op_vectype)))
{
if (dump_p)
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Unsupported vector types in lane permutation\n");
return -1;
}
if (SLP_TREE_LANES (child) != SLP_TREE_LANES (node))
repeating_p = false;
}
gcc_assert (perm.length () == SLP_TREE_LANES (node));
if (dump_p)
{
dump_printf_loc (MSG_NOTE, vect_location,
"vectorizing permutation");
for (unsigned i = 0; i < perm.length (); ++i)
dump_printf (MSG_NOTE, " op%u[%u]", perm[i].first, perm[i].second);
if (repeating_p)
dump_printf (MSG_NOTE, " (repeat %d)\n", SLP_TREE_LANES (node));
dump_printf (MSG_NOTE, "\n");
}
/* REPEATING_P is true if every output vector is guaranteed to use the
same permute vector. We can handle that case for both variable-length
and constant-length vectors, but we only handle other cases for
constant-length vectors.
Set:
- NPATTERNS and NELTS_PER_PATTERN to the encoding of the permute
mask vector that we want to build.
- NCOPIES to the number of copies of PERM that we need in order
to build the necessary permute mask vectors.
- NOUTPUTS_PER_MASK to the number of output vectors we want to create
for each permute mask vector. This is only relevant when GSI is
nonnull. */
uint64_t npatterns;
unsigned nelts_per_pattern;
uint64_t ncopies;
unsigned noutputs_per_mask;
if (repeating_p)
{
/* We need a single permute mask vector that has the form:
{ X1, ..., Xn, X1 + n, ..., Xn + n, X1 + 2n, ..., Xn + 2n, ... }
In other words, the original n-element permute in PERM is
"unrolled" to fill a full vector. The stepped vector encoding
that we use for permutes requires 3n elements. */
npatterns = SLP_TREE_LANES (node);
nelts_per_pattern = ncopies = 3;
noutputs_per_mask = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
}
else
{
/* Calculate every element of every permute mask vector explicitly,
instead of relying on the pattern described above. */
if (!nunits.is_constant (&npatterns))
return -1;
nelts_per_pattern = ncopies = 1;
if (loop_vec_info linfo = dyn_cast <loop_vec_info> (vinfo))
if (!LOOP_VINFO_VECT_FACTOR (linfo).is_constant (&ncopies))
return -1;
noutputs_per_mask = 1;
}
unsigned olanes = ncopies * SLP_TREE_LANES (node);
gcc_assert (repeating_p || multiple_p (olanes, nunits));
/* Compute the { { SLP operand, vector index}, lane } permutation sequence
from the { SLP operand, scalar lane } permutation as recorded in the
SLP node as intermediate step. This part should already work
with SLP children with arbitrary number of lanes. */
auto_vec<std::pair<std::pair<unsigned, unsigned>, unsigned> > vperm;
auto_vec<unsigned> active_lane;
vperm.create (olanes);
active_lane.safe_grow_cleared (children.length (), true);
for (unsigned i = 0; i < ncopies; ++i)
{
for (unsigned pi = 0; pi < perm.length (); ++pi)
{
std::pair<unsigned, unsigned> p = perm[pi];
tree vtype = SLP_TREE_VECTYPE (children[p.first]);
if (repeating_p)
vperm.quick_push ({{p.first, 0}, p.second + active_lane[p.first]});
else
{
/* We checked above that the vectors are constant-length. */
unsigned vnunits = TYPE_VECTOR_SUBPARTS (vtype).to_constant ();
unsigned vi = (active_lane[p.first] + p.second) / vnunits;
unsigned vl = (active_lane[p.first] + p.second) % vnunits;
vperm.quick_push ({{p.first, vi}, vl});
}
}
/* Advance to the next group. */
for (unsigned j = 0; j < children.length (); ++j)
active_lane[j] += SLP_TREE_LANES (children[j]);
}
if (dump_p)
{
dump_printf_loc (MSG_NOTE, vect_location,
"vectorizing permutation");
for (unsigned i = 0; i < perm.length (); ++i)
dump_printf (MSG_NOTE, " op%u[%u]", perm[i].first, perm[i].second);
if (repeating_p)
dump_printf (MSG_NOTE, " (repeat %d)\n", SLP_TREE_LANES (node));
dump_printf (MSG_NOTE, "\n");
dump_printf_loc (MSG_NOTE, vect_location, "as");
for (unsigned i = 0; i < vperm.length (); ++i)
{
if (i != 0
&& (repeating_p
? multiple_p (i, npatterns)
: multiple_p (i, TYPE_VECTOR_SUBPARTS (vectype))))
dump_printf (MSG_NOTE, ",");
dump_printf (MSG_NOTE, " vops%u[%u][%u]",
vperm[i].first.first, vperm[i].first.second,
vperm[i].second);
}
dump_printf (MSG_NOTE, "\n");
}
/* We can only handle two-vector permutes, everything else should
be lowered on the SLP level. The following is closely inspired
by vect_transform_slp_perm_load and is supposed to eventually
replace it.
??? As intermediate step do code-gen in the SLP tree representation
somehow? */
std::pair<unsigned, unsigned> first_vec = std::make_pair (-1U, -1U);
std::pair<unsigned, unsigned> second_vec = std::make_pair (-1U, -1U);
unsigned int index = 0;
poly_uint64 mask_element;
vec_perm_builder mask;
mask.new_vector (nunits, npatterns, nelts_per_pattern);
unsigned int count = mask.encoded_nelts ();
mask.quick_grow (count);
vec_perm_indices indices;
unsigned nperms = 0;
for (unsigned i = 0; i < vperm.length (); ++i)
{
mask_element = vperm[i].second;
if (first_vec.first == -1U
|| first_vec == vperm[i].first)
first_vec = vperm[i].first;
else if (second_vec.first == -1U
|| second_vec == vperm[i].first)
{
second_vec = vperm[i].first;
mask_element += nunits;
}
else
{
if (dump_p)
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"permutation requires at "
"least three vectors\n");
gcc_assert (!gsi);
return -1;
}
mask[index++] = mask_element;
if (index == count)
{
indices.new_vector (mask, second_vec.first == -1U ? 1 : 2,
TYPE_VECTOR_SUBPARTS (op_vectype));
bool identity_p = indices.series_p (0, 1, 0, 1);
machine_mode vmode = TYPE_MODE (vectype);
machine_mode op_vmode = TYPE_MODE (op_vectype);
unsigned HOST_WIDE_INT c;
if ((!identity_p
&& !can_vec_perm_const_p (vmode, op_vmode, indices))
|| (identity_p
&& !known_le (nunits,
TYPE_VECTOR_SUBPARTS (op_vectype))
&& (!constant_multiple_p (nunits,
TYPE_VECTOR_SUBPARTS (op_vectype),
&c) || c != 2)))
{
if (dump_p)
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION,
vect_location,
"unsupported vect permute { ");
for (i = 0; i < count; ++i)
{
dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]);
dump_printf (MSG_MISSED_OPTIMIZATION, " ");
}
dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
}
gcc_assert (!gsi);
return -1;
}
if (!identity_p)
nperms++;
if (gsi)
{
if (second_vec.first == -1U)
second_vec = first_vec;
slp_tree
first_node = children[first_vec.first],
second_node = children[second_vec.first];
tree mask_vec = NULL_TREE;
if (!identity_p)
mask_vec = vect_gen_perm_mask_checked (vectype, indices);
for (unsigned int vi = 0; vi < noutputs_per_mask; ++vi)
{
tree first_def
= vect_get_slp_vect_def (first_node,
first_vec.second + vi);
tree second_def
= vect_get_slp_vect_def (second_node,
second_vec.second + vi);
vect_add_slp_permutation (vinfo, gsi, node, first_def,
second_def, mask_vec);
}
}
index = 0;
first_vec = std::make_pair (-1U, -1U);
second_vec = std::make_pair (-1U, -1U);
}
}
return nperms;
}
/* Vectorize the SLP permutations in NODE as specified
in SLP_TREE_LANE_PERMUTATION which is a vector of pairs of SLP
child number and lane number.
Interleaving of two two-lane two-child SLP subtrees (not supported):
[ { 0, 0 }, { 1, 0 }, { 0, 1 }, { 1, 1 } ]
A blend of two four-lane two-child SLP subtrees:
[ { 0, 0 }, { 1, 1 }, { 0, 2 }, { 1, 3 } ]
Highpart of a four-lane one-child SLP subtree (not supported):
[ { 0, 2 }, { 0, 3 } ]
Where currently only a subset is supported by code generating below. */
static bool
vectorizable_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
slp_tree node, stmt_vector_for_cost *cost_vec)
{
tree vectype = SLP_TREE_VECTYPE (node);
lane_permutation_t &perm = SLP_TREE_LANE_PERMUTATION (node);
int nperms = vectorizable_slp_permutation_1 (vinfo, gsi, node, perm,
SLP_TREE_CHILDREN (node),
dump_enabled_p ());
if (nperms < 0)
return false;
if (!gsi)
record_stmt_cost (cost_vec, nperms, vec_perm, node, vectype, 0, vect_body);
return true;
}
/* Vectorize SLP NODE. */
static void
vect_schedule_slp_node (vec_info *vinfo,
slp_tree node, slp_instance instance)
{
gimple_stmt_iterator si;
int i;
slp_tree child;
/* For existing vectors there's nothing to do. */
if (SLP_TREE_VEC_DEFS (node).exists ())
return;
gcc_assert (SLP_TREE_VEC_STMTS (node).is_empty ());
/* Vectorize externals and constants. */
if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
|| SLP_TREE_DEF_TYPE (node) == vect_external_def)
{
/* ??? vectorizable_shift can end up using a scalar operand which is
currently denoted as !SLP_TREE_VECTYPE. No need to vectorize the
node in this case. */
if (!SLP_TREE_VECTYPE (node))
return;
vect_create_constant_vectors (vinfo, node);
return;
}
stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
gcc_assert (SLP_TREE_NUMBER_OF_VEC_STMTS (node) != 0);
SLP_TREE_VEC_STMTS (node).create (SLP_TREE_NUMBER_OF_VEC_STMTS (node));
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"------>vectorizing SLP node starting from: %G",
stmt_info->stmt);
if (STMT_VINFO_DATA_REF (stmt_info)
&& SLP_TREE_CODE (node) != VEC_PERM_EXPR)
{
/* Vectorized loads go before the first scalar load to make it
ready early, vectorized stores go before the last scalar
stmt which is where all uses are ready. */
stmt_vec_info last_stmt_info = NULL;
if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
last_stmt_info = vect_find_first_scalar_stmt_in_slp (node);
else /* DR_IS_WRITE */
last_stmt_info = vect_find_last_scalar_stmt_in_slp (node);
si = gsi_for_stmt (last_stmt_info->stmt);
}
else if ((STMT_VINFO_TYPE (stmt_info) == cycle_phi_info_type
|| STMT_VINFO_TYPE (stmt_info) == induc_vec_info_type
|| STMT_VINFO_TYPE (stmt_info) == phi_info_type)
&& SLP_TREE_CODE (node) != VEC_PERM_EXPR)
{
/* For PHI node vectorization we do not use the insertion iterator. */
si = gsi_none ();
}
else
{
/* Emit other stmts after the children vectorized defs which is
earliest possible. */
gimple *last_stmt = NULL;
bool seen_vector_def = false;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
{
/* For fold-left reductions we are retaining the scalar
reduction PHI but we still have SLP_TREE_NUM_VEC_STMTS
set so the representation isn't perfect. Resort to the
last scalar def here. */
if (SLP_TREE_VEC_STMTS (child).is_empty ())
{
gcc_assert (STMT_VINFO_TYPE (SLP_TREE_REPRESENTATIVE (child))
== cycle_phi_info_type);
gphi *phi = as_a <gphi *>
(vect_find_last_scalar_stmt_in_slp (child)->stmt);
if (!last_stmt
|| vect_stmt_dominates_stmt_p (last_stmt, phi))
last_stmt = phi;
}
/* We are emitting all vectorized stmts in the same place and
the last one is the last.
??? Unless we have a load permutation applied and that
figures to re-use an earlier generated load. */
unsigned j;
gimple *vstmt;
FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (child), j, vstmt)
if (!last_stmt
|| vect_stmt_dominates_stmt_p (last_stmt, vstmt))
last_stmt = vstmt;
}
else if (!SLP_TREE_VECTYPE (child))
{
/* For externals we use unvectorized at all scalar defs. */
unsigned j;
tree def;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (child), j, def)
if (TREE_CODE (def) == SSA_NAME
&& !SSA_NAME_IS_DEFAULT_DEF (def))
{
gimple *stmt = SSA_NAME_DEF_STMT (def);
if (!last_stmt
|| vect_stmt_dominates_stmt_p (last_stmt, stmt))
last_stmt = stmt;
}
}
else
{
/* For externals we have to look at all defs since their
insertion place is decided per vector. But beware
of pre-existing vectors where we need to make sure
we do not insert before the region boundary. */
if (SLP_TREE_SCALAR_OPS (child).is_empty ()
&& !vinfo->lookup_def (SLP_TREE_VEC_DEFS (child)[0]))
seen_vector_def = true;
else
{
unsigned j;
tree vdef;
FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (child), j, vdef)
if (TREE_CODE (vdef) == SSA_NAME
&& !SSA_NAME_IS_DEFAULT_DEF (vdef))
{
gimple *vstmt = SSA_NAME_DEF_STMT (vdef);
if (!last_stmt
|| vect_stmt_dominates_stmt_p (last_stmt, vstmt))
last_stmt = vstmt;
}
}
}
/* This can happen when all children are pre-existing vectors or
constants. */
if (!last_stmt)
last_stmt = vect_find_first_scalar_stmt_in_slp (node)->stmt;
if (!last_stmt)
{
gcc_assert (seen_vector_def);
si = gsi_after_labels (as_a <bb_vec_info> (vinfo)->bbs[0]);
}
else if (is_ctrl_altering_stmt (last_stmt))
{
/* We split regions to vectorize at control altering stmts
with a definition so this must be an external which
we can insert at the start of the region. */
si = gsi_after_labels (as_a <bb_vec_info> (vinfo)->bbs[0]);
}
else if (is_a <bb_vec_info> (vinfo)
&& gimple_bb (last_stmt) != gimple_bb (stmt_info->stmt)
&& gimple_could_trap_p (stmt_info->stmt))
{
/* We've constrained possibly trapping operations to all come
from the same basic-block, if vectorized defs would allow earlier
scheduling still force vectorized stmts to the original block.
This is only necessary for BB vectorization since for loop vect
all operations are in a single BB and scalar stmt based
placement doesn't play well with epilogue vectorization. */
gcc_assert (dominated_by_p (CDI_DOMINATORS,
gimple_bb (stmt_info->stmt),
gimple_bb (last_stmt)));
si = gsi_after_labels (gimple_bb (stmt_info->stmt));
}
else if (is_a <gphi *> (last_stmt))
si = gsi_after_labels (gimple_bb (last_stmt));
else
{
si = gsi_for_stmt (last_stmt);
gsi_next (&si);
}
}
bool done_p = false;
/* Handle purely internal nodes. */
if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
{
/* ??? the transform kind is stored to STMT_VINFO_TYPE which might
be shared with different SLP nodes (but usually it's the same
operation apart from the case the stmt is only there for denoting
the actual scalar lane defs ...). So do not call vect_transform_stmt
but open-code it here (partly). */
bool done = vectorizable_slp_permutation (vinfo, &si, node, NULL);
gcc_assert (done);
done_p = true;
}
if (!done_p)
vect_transform_stmt (vinfo, stmt_info, &si, node, instance);
}
/* Replace scalar calls from SLP node NODE with setting of their lhs to zero.
For loop vectorization this is done in vectorizable_call, but for SLP
it needs to be deferred until end of vect_schedule_slp, because multiple
SLP instances may refer to the same scalar stmt. */
static void
vect_remove_slp_scalar_calls (vec_info *vinfo,
slp_tree node, hash_set<slp_tree> &visited)
{
gimple *new_stmt;
gimple_stmt_iterator gsi;
int i;
slp_tree child;
tree lhs;
stmt_vec_info stmt_info;
if (!node || SLP_TREE_DEF_TYPE (node) != vect_internal_def)
return;
if (visited.add (node))
return;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
vect_remove_slp_scalar_calls (vinfo, child, visited);
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
{
gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt);
if (!stmt || gimple_bb (stmt) == NULL)
continue;
if (is_pattern_stmt_p (stmt_info)
|| !PURE_SLP_STMT (stmt_info))
continue;
lhs = gimple_call_lhs (stmt);
new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs)));
gsi = gsi_for_stmt (stmt);
vinfo->replace_stmt (&gsi, stmt_info, new_stmt);
SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt;
}
}
static void
vect_remove_slp_scalar_calls (vec_info *vinfo, slp_tree node)
{
hash_set<slp_tree> visited;
vect_remove_slp_scalar_calls (vinfo, node, visited);
}
/* Vectorize the instance root. */
void
vectorize_slp_instance_root_stmt (slp_tree node, slp_instance instance)
{
gassign *rstmt = NULL;
if (instance->kind == slp_inst_kind_ctor)
{
if (SLP_TREE_NUMBER_OF_VEC_STMTS (node) == 1)
{
gimple *child_stmt = SLP_TREE_VEC_STMTS (node)[0];
tree vect_lhs = gimple_get_lhs (child_stmt);
tree root_lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
if (!useless_type_conversion_p (TREE_TYPE (root_lhs),
TREE_TYPE (vect_lhs)))
vect_lhs = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (root_lhs),
vect_lhs);
rstmt = gimple_build_assign (root_lhs, vect_lhs);
}
else if (SLP_TREE_NUMBER_OF_VEC_STMTS (node) > 1)
{
int nelts = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
gimple *child_stmt;
int j;
vec<constructor_elt, va_gc> *v;
vec_alloc (v, nelts);
/* A CTOR can handle V16HI composition from VNx8HI so we
do not need to convert vector elements if the types
do not match. */
FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (node), j, child_stmt)
CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
gimple_get_lhs (child_stmt));
tree lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
tree rtype
= TREE_TYPE (gimple_assign_rhs1 (instance->root_stmts[0]->stmt));
tree r_constructor = build_constructor (rtype, v);
rstmt = gimple_build_assign (lhs, r_constructor);
}
}
else if (instance->kind == slp_inst_kind_bb_reduc)
{
/* Largely inspired by reduction chain epilogue handling in
vect_create_epilog_for_reduction. */
vec<tree> vec_defs = vNULL;
vect_get_slp_defs (node, &vec_defs);
enum tree_code reduc_code
= gimple_assign_rhs_code (instance->root_stmts[0]->stmt);
/* ??? We actually have to reflect signs somewhere. */
if (reduc_code == MINUS_EXPR)
reduc_code = PLUS_EXPR;
gimple_seq epilogue = NULL;
/* We may end up with more than one vector result, reduce them
to one vector. */
tree vec_def = vec_defs[0];
for (unsigned i = 1; i < vec_defs.length (); ++i)
vec_def = gimple_build (&epilogue, reduc_code, TREE_TYPE (vec_def),
vec_def, vec_defs[i]);
vec_defs.release ();
/* ??? Support other schemes than direct internal fn. */
internal_fn reduc_fn;
if (!reduction_fn_for_scalar_code (reduc_code, &reduc_fn)
|| reduc_fn == IFN_LAST)
gcc_unreachable ();
tree scalar_def = gimple_build (&epilogue, as_combined_fn (reduc_fn),
TREE_TYPE (TREE_TYPE (vec_def)), vec_def);
gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
gsi_insert_seq_before (&rgsi, epilogue, GSI_SAME_STMT);
gimple_assign_set_rhs_from_tree (&rgsi, scalar_def);
update_stmt (gsi_stmt (rgsi));
return;
}
else
gcc_unreachable ();
gcc_assert (rstmt);
gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
gsi_replace (&rgsi, rstmt, true);
}
struct slp_scc_info
{
bool on_stack;
int dfs;
int lowlink;
};
/* Schedule the SLP INSTANCE doing a DFS walk and collecting SCCs. */
static void
vect_schedule_scc (vec_info *vinfo, slp_tree node, slp_instance instance,
hash_map<slp_tree, slp_scc_info> &scc_info,
int &maxdfs, vec<slp_tree> &stack)
{
bool existed_p;
slp_scc_info *info = &scc_info.get_or_insert (node, &existed_p);
gcc_assert (!existed_p);
info->dfs = maxdfs;
info->lowlink = maxdfs;
maxdfs++;
/* Leaf. */
if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
{
info->on_stack = false;
vect_schedule_slp_node (vinfo, node, instance);
return;
}
info->on_stack = true;
stack.safe_push (node);
unsigned i;
slp_tree child;
/* DFS recurse. */
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
{
if (!child)
continue;
slp_scc_info *child_info = scc_info.get (child);
if (!child_info)
{
vect_schedule_scc (vinfo, child, instance, scc_info, maxdfs, stack);
/* Recursion might have re-allocated the node. */
info = scc_info.get (node);
child_info = scc_info.get (child);
info->lowlink = MIN (info->lowlink, child_info->lowlink);
}
else if (child_info->on_stack)
info->lowlink = MIN (info->lowlink, child_info->dfs);
}
if (info->lowlink != info->dfs)
return;
auto_vec<slp_tree, 4> phis_to_fixup;
/* Singleton. */
if (stack.last () == node)
{
stack.pop ();
info->on_stack = false;
vect_schedule_slp_node (vinfo, node, instance);
if (SLP_TREE_CODE (node) != VEC_PERM_EXPR
&& is_a <gphi *> (SLP_TREE_REPRESENTATIVE (node)->stmt))
phis_to_fixup.quick_push (node);
}
else
{
/* SCC. */
int last_idx = stack.length () - 1;
while (stack[last_idx] != node)
last_idx--;
/* We can break the cycle at PHIs who have at least one child
code generated. Then we could re-start the DFS walk until
all nodes in the SCC are covered (we might have new entries
for only back-reachable nodes). But it's simpler to just
iterate and schedule those that are ready. */
unsigned todo = stack.length () - last_idx;
do
{
for (int idx = stack.length () - 1; idx >= last_idx; --idx)
{
slp_tree entry = stack[idx];
if (!entry)
continue;
bool phi = (SLP_TREE_CODE (entry) != VEC_PERM_EXPR
&& is_a <gphi *> (SLP_TREE_REPRESENTATIVE (entry)->stmt));
bool ready = !phi;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (entry), i, child)
if (!child)
{
gcc_assert (phi);
ready = true;
break;
}
else if (scc_info.get (child)->on_stack)
{
if (!phi)
{
ready = false;
break;
}
}
else
{
if (phi)
{
ready = true;
break;
}
}
if (ready)
{
vect_schedule_slp_node (vinfo, entry, instance);
scc_info.get (entry)->on_stack = false;
stack[idx] = NULL;
todo--;
if (phi)
phis_to_fixup.safe_push (entry);
}
}
}
while (todo != 0);
/* Pop the SCC. */
stack.truncate (last_idx);
}
/* Now fixup the backedge def of the vectorized PHIs in this SCC. */
slp_tree phi_node;
FOR_EACH_VEC_ELT (phis_to_fixup, i, phi_node)
{
gphi *phi = as_a <gphi *> (SLP_TREE_REPRESENTATIVE (phi_node)->stmt);
edge_iterator ei;
edge e;
FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
{
unsigned dest_idx = e->dest_idx;
child = SLP_TREE_CHILDREN (phi_node)[dest_idx];
if (!child || SLP_TREE_DEF_TYPE (child) != vect_internal_def)
continue;
/* Simply fill all args. */
for (unsigned i = 0; i < SLP_TREE_VEC_STMTS (phi_node).length (); ++i)
add_phi_arg (as_a <gphi *> (SLP_TREE_VEC_STMTS (phi_node)[i]),
vect_get_slp_vect_def (child, i),
e, gimple_phi_arg_location (phi, dest_idx));
}
}
}
/* Generate vector code for SLP_INSTANCES in the loop/basic block. */
void
vect_schedule_slp (vec_info *vinfo, const vec<slp_instance> &slp_instances)
{
slp_instance instance;
unsigned int i;
hash_map<slp_tree, slp_scc_info> scc_info;
int maxdfs = 0;
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
slp_tree node = SLP_INSTANCE_TREE (instance);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Vectorizing SLP tree:\n");
/* ??? Dump all? */
if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
dump_printf_loc (MSG_NOTE, vect_location, "Root stmt: %G",
SLP_INSTANCE_ROOT_STMTS (instance)[0]->stmt);
vect_print_slp_graph (MSG_NOTE, vect_location,
SLP_INSTANCE_TREE (instance));
}
/* Schedule the tree of INSTANCE, scheduling SCCs in a way to
have a PHI be the node breaking the cycle. */
auto_vec<slp_tree> stack;
if (!scc_info.get (node))
vect_schedule_scc (vinfo, node, instance, scc_info, maxdfs, stack);
if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
vectorize_slp_instance_root_stmt (node, instance);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vectorizing stmts using SLP.\n");
}
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
slp_tree root = SLP_INSTANCE_TREE (instance);
stmt_vec_info store_info;
unsigned int j;
/* Remove scalar call stmts. Do not do this for basic-block
vectorization as not all uses may be vectorized.
??? Why should this be necessary? DCE should be able to
remove the stmts itself.
??? For BB vectorization we can as well remove scalar
stmts starting from the SLP tree root if they have no
uses. */
if (is_a <loop_vec_info> (vinfo))
vect_remove_slp_scalar_calls (vinfo, root);
/* Remove vectorized stores original scalar stmts. */
for (j = 0; SLP_TREE_SCALAR_STMTS (root).iterate (j, &store_info); j++)
{
if (!STMT_VINFO_DATA_REF (store_info)
|| !DR_IS_WRITE (STMT_VINFO_DATA_REF (store_info)))
break;
store_info = vect_orig_stmt (store_info);
/* Free the attached stmt_vec_info and remove the stmt. */
vinfo->remove_stmt (store_info);
/* Invalidate SLP_TREE_REPRESENTATIVE in case we released it
to not crash in vect_free_slp_tree later. */
if (SLP_TREE_REPRESENTATIVE (root) == store_info)
SLP_TREE_REPRESENTATIVE (root) = NULL;
}
}
}
|