1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625
7626
7627
7628
7629
7630
7631
7632
7633
7634
7635
7636
7637
7638
7639
7640
7641
7642
7643
7644
7645
7646
7647
7648
7649
7650
7651
7652
7653
7654
7655
7656
7657
7658
7659
7660
7661
7662
7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
7676
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
7747
7748
7749
7750
7751
7752
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
7769
7770
7771
7772
7773
7774
7775
7776
7777
7778
7779
7780
7781
7782
7783
7784
7785
7786
7787
7788
7789
7790
7791
7792
7793
7794
7795
7796
7797
7798
7799
7800
7801
7802
7803
7804
7805
7806
7807
7808
7809
7810
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
7846
7847
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860
7861
7862
7863
7864
7865
7866
7867
7868
7869
7870
7871
7872
7873
7874
7875
7876
7877
7878
7879
7880
7881
7882
7883
7884
7885
7886
7887
7888
7889
7890
7891
7892
7893
7894
7895
7896
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927
7928
7929
7930
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944
7945
7946
7947
7948
7949
7950
7951
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
7976
7977
7978
7979
7980
7981
7982
7983
7984
7985
7986
7987
7988
7989
7990
7991
7992
7993
7994
7995
7996
7997
7998
7999
8000
8001
8002
8003
8004
8005
8006
8007
8008
8009
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
8026
8027
8028
8029
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
8046
8047
8048
8049
8050
8051
8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
8077
8078
8079
8080
8081
8082
8083
8084
8085
8086
8087
8088
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
8113
8114
8115
8116
8117
8118
8119
8120
8121
8122
8123
8124
8125
8126
8127
8128
8129
8130
8131
8132
8133
8134
8135
8136
8137
8138
8139
8140
8141
8142
8143
8144
8145
8146
8147
8148
8149
8150
8151
8152
8153
8154
8155
8156
8157
8158
8159
8160
8161
8162
8163
8164
8165
8166
8167
8168
8169
8170
8171
8172
8173
8174
8175
8176
8177
8178
8179
8180
8181
8182
8183
8184
8185
8186
8187
8188
8189
8190
8191
8192
8193
8194
8195
8196
8197
8198
8199
8200
8201
8202
8203
8204
8205
8206
8207
8208
8209
8210
8211
8212
8213
8214
8215
8216
8217
8218
8219
8220
8221
8222
8223
8224
8225
8226
8227
8228
8229
8230
8231
8232
8233
8234
8235
8236
8237
8238
8239
8240
8241
8242
8243
8244
8245
8246
8247
8248
8249
8250
8251
8252
8253
8254
8255
8256
8257
8258
8259
8260
8261
8262
8263
8264
8265
8266
8267
8268
8269
8270
8271
8272
8273
8274
8275
8276
8277
8278
8279
8280
8281
8282
8283
8284
8285
8286
8287
8288
8289
8290
8291
8292
8293
8294
8295
8296
8297
8298
8299
8300
8301
8302
8303
8304
8305
8306
8307
8308
8309
8310
8311
8312
8313
8314
8315
8316
8317
8318
8319
8320
8321
8322
8323
8324
8325
8326
8327
8328
8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339
8340
8341
8342
8343
8344
8345
8346
8347
8348
8349
8350
8351
8352
8353
8354
8355
8356
8357
8358
8359
8360
8361
8362
8363
8364
8365
8366
8367
8368
8369
8370
8371
8372
8373
8374
8375
8376
8377
8378
8379
8380
8381
8382
8383
8384
8385
8386
8387
8388
8389
8390
8391
8392
8393
8394
8395
8396
8397
8398
8399
8400
8401
8402
8403
8404
8405
8406
8407
8408
8409
8410
8411
8412
8413
8414
8415
8416
8417
8418
8419
8420
8421
8422
8423
8424
8425
8426
8427
8428
8429
8430
8431
8432
8433
8434
8435
8436
8437
8438
8439
8440
8441
8442
8443
8444
8445
8446
8447
8448
8449
8450
8451
8452
8453
8454
8455
8456
8457
8458
8459
8460
8461
8462
8463
8464
8465
8466
8467
8468
8469
8470
8471
8472
8473
8474
8475
8476
8477
8478
8479
8480
8481
8482
8483
8484
8485
8486
8487
8488
8489
8490
8491
8492
8493
8494
8495
8496
8497
8498
8499
8500
8501
8502
8503
8504
8505
8506
8507
8508
8509
8510
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525
8526
8527
8528
8529
8530
8531
8532
8533
8534
8535
8536
8537
8538
8539
8540
8541
8542
8543
8544
8545
8546
8547
8548
8549
8550
8551
8552
8553
8554
8555
8556
8557
8558
8559
8560
8561
8562
8563
8564
8565
8566
8567
8568
8569
8570
8571
8572
8573
8574
8575
8576
8577
8578
8579
8580
8581
8582
8583
8584
8585
8586
8587
8588
8589
8590
8591
8592
8593
8594
8595
8596
8597
8598
8599
8600
8601
8602
8603
8604
8605
8606
8607
8608
8609
8610
8611
8612
8613
8614
8615
8616
8617
8618
8619
8620
8621
8622
8623
8624
8625
8626
8627
8628
8629
8630
8631
8632
8633
8634
8635
8636
8637
8638
8639
8640
8641
8642
8643
8644
8645
8646
8647
8648
8649
8650
8651
8652
8653
8654
8655
8656
8657
8658
8659
8660
8661
8662
8663
8664
8665
8666
8667
8668
8669
8670
8671
8672
8673
8674
8675
8676
8677
8678
8679
8680
8681
8682
8683
8684
8685
8686
8687
8688
8689
8690
8691
8692
8693
8694
8695
8696
8697
8698
8699
8700
8701
8702
8703
8704
8705
8706
8707
8708
8709
8710
8711
8712
8713
8714
8715
8716
8717
8718
8719
8720
8721
8722
8723
8724
8725
8726
8727
8728
8729
8730
8731
8732
8733
8734
8735
8736
8737
8738
8739
8740
8741
8742
8743
8744
8745
8746
8747
8748
8749
8750
8751
8752
8753
8754
8755
8756
8757
8758
8759
8760
8761
8762
8763
8764
8765
8766
8767
8768
8769
8770
8771
8772
8773
8774
8775
8776
8777
8778
8779
8780
8781
8782
8783
8784
8785
8786
8787
8788
8789
8790
8791
8792
8793
8794
8795
8796
8797
8798
8799
8800
8801
8802
8803
8804
8805
8806
8807
8808
8809
8810
8811
8812
8813
8814
8815
8816
8817
8818
8819
8820
8821
8822
8823
8824
8825
8826
8827
8828
8829
8830
8831
8832
8833
8834
8835
8836
8837
8838
8839
8840
8841
8842
8843
8844
8845
8846
8847
8848
8849
8850
8851
8852
8853
8854
8855
8856
8857
8858
8859
8860
8861
8862
8863
8864
8865
8866
8867
8868
8869
8870
8871
8872
8873
8874
8875
8876
8877
8878
8879
8880
8881
8882
8883
8884
8885
8886
8887
8888
8889
8890
8891
8892
8893
8894
8895
8896
8897
8898
8899
8900
8901
8902
8903
8904
8905
8906
8907
8908
8909
8910
8911
8912
8913
8914
8915
8916
8917
8918
8919
8920
8921
8922
8923
8924
8925
8926
8927
8928
8929
8930
8931
8932
8933
8934
8935
8936
8937
8938
8939
8940
8941
8942
8943
8944
8945
8946
8947
8948
8949
8950
8951
8952
8953
8954
8955
8956
8957
8958
8959
8960
8961
8962
8963
8964
8965
8966
8967
8968
8969
8970
8971
8972
8973
8974
8975
8976
8977
8978
8979
8980
8981
8982
8983
8984
8985
8986
8987
8988
8989
8990
8991
8992
8993
8994
8995
8996
8997
8998
8999
9000
9001
9002
9003
9004
9005
9006
9007
9008
9009
9010
9011
9012
9013
9014
9015
9016
9017
9018
9019
9020
9021
9022
9023
9024
9025
9026
9027
9028
9029
9030
9031
9032
9033
9034
9035
9036
9037
9038
9039
9040
9041
9042
9043
9044
9045
9046
9047
9048
9049
9050
9051
9052
9053
9054
9055
9056
9057
9058
9059
9060
9061
9062
9063
9064
9065
9066
9067
9068
9069
9070
9071
9072
9073
9074
9075
9076
9077
9078
9079
9080
9081
9082
9083
9084
9085
9086
9087
9088
9089
9090
9091
9092
9093
9094
9095
9096
9097
9098
9099
9100
9101
9102
9103
9104
9105
9106
9107
9108
9109
9110
9111
9112
9113
9114
9115
9116
9117
9118
9119
9120
9121
9122
9123
9124
9125
9126
9127
9128
9129
9130
9131
9132
9133
9134
9135
9136
9137
9138
9139
9140
9141
9142
9143
9144
9145
9146
9147
9148
9149
9150
9151
9152
9153
9154
9155
9156
9157
9158
9159
9160
9161
9162
9163
9164
9165
9166
9167
9168
9169
9170
9171
9172
9173
9174
9175
9176
9177
9178
9179
9180
9181
9182
9183
9184
9185
9186
9187
9188
9189
9190
9191
9192
9193
9194
9195
9196
9197
9198
9199
9200
9201
9202
9203
9204
9205
9206
9207
9208
9209
9210
9211
9212
9213
9214
9215
9216
9217
9218
9219
9220
9221
9222
9223
9224
9225
9226
9227
9228
9229
9230
9231
9232
9233
9234
9235
9236
9237
9238
9239
9240
9241
9242
9243
9244
9245
9246
9247
9248
9249
9250
9251
9252
9253
9254
9255
9256
9257
9258
9259
9260
9261
9262
9263
9264
9265
9266
9267
9268
9269
9270
9271
9272
9273
9274
9275
9276
9277
9278
9279
9280
9281
9282
9283
9284
9285
9286
9287
9288
9289
9290
9291
9292
9293
9294
9295
9296
9297
9298
9299
9300
9301
9302
9303
9304
9305
9306
9307
9308
9309
9310
9311
9312
9313
9314
9315
9316
9317
9318
9319
9320
9321
9322
9323
9324
9325
9326
9327
9328
9329
9330
9331
9332
9333
9334
9335
9336
9337
9338
9339
9340
9341
9342
9343
9344
9345
9346
9347
9348
9349
9350
9351
9352
9353
9354
9355
9356
9357
9358
9359
9360
9361
9362
9363
9364
9365
9366
9367
9368
9369
9370
9371
9372
9373
9374
9375
9376
9377
9378
9379
9380
9381
9382
9383
9384
9385
9386
9387
9388
9389
9390
9391
9392
9393
9394
9395
9396
9397
9398
9399
9400
9401
9402
9403
9404
9405
9406
9407
9408
9409
9410
9411
9412
9413
9414
9415
9416
9417
9418
9419
9420
9421
9422
9423
9424
9425
9426
9427
9428
9429
9430
9431
9432
9433
9434
9435
9436
9437
9438
9439
9440
9441
9442
9443
9444
9445
9446
9447
9448
9449
9450
9451
9452
9453
9454
9455
9456
9457
9458
9459
9460
9461
9462
9463
9464
9465
9466
9467
9468
9469
9470
9471
9472
9473
9474
9475
9476
9477
9478
9479
9480
9481
9482
9483
9484
9485
9486
9487
9488
9489
9490
9491
9492
9493
9494
9495
9496
9497
9498
9499
9500
9501
9502
9503
9504
9505
9506
9507
9508
9509
9510
9511
9512
9513
9514
9515
9516
9517
9518
9519
9520
9521
9522
9523
9524
9525
9526
9527
9528
9529
9530
9531
9532
9533
9534
9535
9536
9537
9538
9539
9540
9541
9542
9543
9544
9545
9546
9547
9548
9549
9550
9551
9552
9553
9554
9555
9556
9557
9558
9559
9560
9561
9562
9563
9564
9565
9566
9567
9568
9569
9570
9571
9572
9573
9574
9575
9576
9577
9578
9579
9580
9581
9582
9583
9584
9585
9586
9587
9588
9589
9590
9591
9592
9593
9594
9595
9596
9597
9598
9599
9600
9601
9602
9603
9604
9605
9606
9607
9608
9609
9610
9611
9612
9613
9614
9615
9616
9617
9618
9619
9620
9621
9622
9623
9624
9625
9626
9627
9628
9629
9630
9631
9632
9633
9634
9635
9636
9637
9638
9639
9640
9641
9642
9643
9644
9645
9646
9647
9648
9649
9650
9651
9652
9653
9654
9655
9656
9657
9658
9659
9660
9661
9662
9663
9664
9665
9666
9667
9668
9669
9670
9671
9672
9673
9674
9675
9676
9677
9678
9679
9680
9681
9682
9683
9684
9685
9686
9687
9688
9689
9690
9691
9692
9693
9694
9695
9696
9697
9698
9699
9700
9701
9702
9703
9704
9705
9706
9707
9708
9709
9710
9711
9712
9713
9714
9715
9716
9717
9718
9719
9720
9721
9722
9723
9724
9725
9726
9727
9728
9729
9730
9731
9732
9733
9734
9735
9736
9737
9738
9739
9740
9741
9742
9743
9744
9745
9746
9747
9748
9749
9750
9751
9752
9753
9754
9755
9756
9757
9758
9759
9760
9761
9762
9763
9764
9765
9766
9767
9768
9769
9770
9771
9772
9773
9774
9775
9776
9777
9778
9779
9780
9781
9782
9783
9784
9785
9786
9787
9788
9789
9790
9791
9792
9793
9794
9795
9796
9797
9798
9799
9800
9801
9802
9803
9804
9805
9806
9807
9808
9809
9810
9811
9812
9813
9814
9815
9816
9817
9818
9819
9820
9821
9822
9823
9824
9825
9826
9827
9828
9829
9830
9831
9832
9833
9834
9835
9836
9837
9838
9839
9840
9841
9842
9843
9844
9845
9846
9847
9848
9849
9850
9851
9852
9853
9854
9855
9856
9857
9858
9859
9860
9861
9862
9863
9864
9865
9866
9867
9868
9869
9870
9871
9872
9873
9874
9875
9876
9877
9878
9879
9880
9881
9882
9883
9884
9885
9886
9887
9888
9889
9890
9891
9892
9893
9894
9895
9896
9897
9898
9899
9900
9901
9902
9903
9904
9905
9906
9907
9908
9909
9910
9911
9912
9913
9914
9915
9916
9917
9918
9919
9920
9921
9922
9923
9924
9925
9926
9927
9928
9929
9930
9931
9932
9933
9934
9935
9936
9937
9938
9939
9940
9941
9942
9943
9944
9945
9946
9947
9948
9949
9950
9951
9952
9953
9954
9955
9956
9957
9958
9959
9960
9961
9962
9963
9964
9965
9966
9967
9968
9969
9970
9971
9972
9973
9974
9975
9976
9977
9978
9979
9980
9981
9982
9983
9984
9985
9986
9987
9988
9989
9990
9991
9992
9993
9994
9995
9996
9997
9998
9999
10000
10001
10002
10003
10004
10005
10006
10007
10008
10009
10010
10011
10012
10013
10014
10015
10016
10017
10018
10019
10020
10021
10022
10023
10024
10025
10026
10027
10028
10029
10030
10031
10032
10033
10034
10035
10036
10037
10038
10039
10040
10041
10042
10043
10044
10045
10046
10047
10048
10049
10050
10051
10052
10053
10054
10055
10056
10057
10058
10059
10060
10061
10062
10063
10064
10065
10066
10067
10068
10069
10070
10071
10072
10073
10074
10075
10076
10077
10078
10079
10080
10081
10082
10083
10084
10085
10086
10087
10088
10089
10090
10091
10092
10093
10094
10095
10096
10097
10098
10099
10100
10101
10102
10103
10104
10105
10106
10107
10108
10109
10110
10111
10112
10113
10114
10115
10116
10117
10118
10119
10120
10121
10122
10123
10124
10125
10126
10127
10128
10129
10130
10131
10132
10133
10134
10135
10136
10137
10138
10139
10140
10141
10142
10143
10144
10145
10146
10147
10148
10149
10150
10151
10152
10153
10154
10155
10156
10157
10158
10159
10160
10161
10162
10163
10164
10165
10166
10167
10168
10169
10170
10171
10172
10173
10174
10175
10176
10177
10178
10179
10180
10181
10182
10183
10184
10185
10186
10187
10188
10189
10190
10191
10192
10193
10194
|
/* Perform various loop optimizations, including strength reduction.
Copyright (C) 1987, 88, 89, 91-99, 2000 Free Software Foundation, Inc.
This file is part of GNU CC.
GNU CC 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 2, or (at your option)
any later version.
GNU CC 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 GNU CC; see the file COPYING. If not, write to
the Free Software Foundation, 59 Temple Place - Suite 330,
Boston, MA 02111-1307, USA. */
/* This is the loop optimization pass of the compiler.
It finds invariant computations within loops and moves them
to the beginning of the loop. Then it identifies basic and
general induction variables. Strength reduction is applied to the general
induction variables, and induction variable elimination is applied to
the basic induction variables.
It also finds cases where
a register is set within the loop by zero-extending a narrower value
and changes these to zero the entire register once before the loop
and merely copy the low part within the loop.
Most of the complexity is in heuristics to decide when it is worth
while to do these things. */
#include "config.h"
#include "system.h"
#include "rtl.h"
#include "tm_p.h"
#include "obstack.h"
#include "function.h"
#include "expr.h"
#include "basic-block.h"
#include "insn-config.h"
#include "insn-flags.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "recog.h"
#include "flags.h"
#include "real.h"
#include "loop.h"
#include "except.h"
#include "toplev.h"
/* Vector mapping INSN_UIDs to luids.
The luids are like uids but increase monotonically always.
We use them to see whether a jump comes from outside a given loop. */
int *uid_luid;
/* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
number the insn is contained in. */
struct loop **uid_loop;
/* 1 + largest uid of any insn. */
int max_uid_for_loop;
/* 1 + luid of last insn. */
static int max_luid;
/* Number of loops detected in current function. Used as index to the
next few tables. */
static int max_loop_num;
/* Indexed by register number, contains the number of times the reg
is set during the loop being scanned.
During code motion, a negative value indicates a reg that has been
made a candidate; in particular -2 means that it is an candidate that
we know is equal to a constant and -1 means that it is an candidate
not known equal to a constant.
After code motion, regs moved have 0 (which is accurate now)
while the failed candidates have the original number of times set.
Therefore, at all times, == 0 indicates an invariant register;
< 0 a conditionally invariant one. */
static varray_type set_in_loop;
/* Original value of set_in_loop; same except that this value
is not set negative for a reg whose sets have been made candidates
and not set to 0 for a reg that is moved. */
static varray_type n_times_set;
/* Index by register number, 1 indicates that the register
cannot be moved or strength reduced. */
static varray_type may_not_optimize;
/* Contains the insn in which a register was used if it was used
exactly once; contains const0_rtx if it was used more than once. */
static varray_type reg_single_usage;
/* Nonzero means reg N has already been moved out of one loop.
This reduces the desire to move it out of another. */
static char *moved_once;
/* List of MEMs that are stored in this loop. */
static rtx loop_store_mems;
/* The insn where the first of these was found. */
static rtx first_loop_store_insn;
typedef struct loop_mem_info {
rtx mem; /* The MEM itself. */
rtx reg; /* Corresponding pseudo, if any. */
int optimize; /* Nonzero if we can optimize access to this MEM. */
} loop_mem_info;
/* Array of MEMs that are used (read or written) in this loop, but
cannot be aliased by anything in this loop, except perhaps
themselves. In other words, if loop_mems[i] is altered during the
loop, it is altered by an expression that is rtx_equal_p to it. */
static loop_mem_info *loop_mems;
/* The index of the next available slot in LOOP_MEMS. */
static int loop_mems_idx;
/* The number of elements allocated in LOOP_MEMs. */
static int loop_mems_allocated;
/* Nonzero if we don't know what MEMs were changed in the current
loop. This happens if the loop contains a call (in which case
`loop_info->has_call' will also be set) or if we store into more
than NUM_STORES MEMs. */
static int unknown_address_altered;
/* The above doesn't count any readonly memory locations that are stored.
This does. */
static int unknown_constant_address_altered;
/* Count of movable (i.e. invariant) instructions discovered in the loop. */
static int num_movables;
/* Count of memory write instructions discovered in the loop. */
static int num_mem_sets;
/* Bound on pseudo register number before loop optimization.
A pseudo has valid regscan info if its number is < max_reg_before_loop. */
int max_reg_before_loop;
/* The value to pass to the next call of reg_scan_update. */
static int loop_max_reg;
/* This obstack is used in product_cheap_p to allocate its rtl. It
may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
If we used the same obstack that it did, we would be deallocating
that array. */
static struct obstack temp_obstack;
/* This is where the pointer to the obstack being used for RTL is stored. */
extern struct obstack *rtl_obstack;
#define obstack_chunk_alloc xmalloc
#define obstack_chunk_free free
/* During the analysis of a loop, a chain of `struct movable's
is made to record all the movable insns found.
Then the entire chain can be scanned to decide which to move. */
struct movable
{
rtx insn; /* A movable insn */
rtx set_src; /* The expression this reg is set from. */
rtx set_dest; /* The destination of this SET. */
rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
of any registers used within the LIBCALL. */
int consec; /* Number of consecutive following insns
that must be moved with this one. */
int regno; /* The register it sets */
short lifetime; /* lifetime of that register;
may be adjusted when matching movables
that load the same value are found. */
short savings; /* Number of insns we can move for this reg,
including other movables that force this
or match this one. */
unsigned int cond : 1; /* 1 if only conditionally movable */
unsigned int force : 1; /* 1 means MUST move this insn */
unsigned int global : 1; /* 1 means reg is live outside this loop */
/* If PARTIAL is 1, GLOBAL means something different:
that the reg is live outside the range from where it is set
to the following label. */
unsigned int done : 1; /* 1 inhibits further processing of this */
unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
In particular, moving it does not make it
invariant. */
unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
load SRC, rather than copying INSN. */
unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
first insn of a consecutive sets group. */
unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
enum machine_mode savemode; /* Nonzero means it is a mode for a low part
that we should avoid changing when clearing
the rest of the reg. */
struct movable *match; /* First entry for same value */
struct movable *forces; /* An insn that must be moved if this is */
struct movable *next;
};
static struct movable *the_movables;
FILE *loop_dump_stream;
/* Forward declarations. */
static void verify_dominator PARAMS ((struct loop *));
static void find_and_verify_loops PARAMS ((rtx, struct loops *));
static void mark_loop_jump PARAMS ((rtx, struct loop *));
static void prescan_loop PARAMS ((struct loop *));
static int reg_in_basic_block_p PARAMS ((rtx, rtx));
static int consec_sets_invariant_p PARAMS ((const struct loop *,
rtx, int, rtx));
static int labels_in_range_p PARAMS ((rtx, int));
static void count_one_set PARAMS ((rtx, rtx, varray_type, rtx *));
static void count_loop_regs_set PARAMS ((rtx, rtx, varray_type, varray_type,
int *, int));
static void note_addr_stored PARAMS ((rtx, rtx, void *));
static void note_set_pseudo_multiple_uses PARAMS ((rtx, rtx, void *));
static int loop_reg_used_before_p PARAMS ((const struct loop *, rtx, rtx));
static void scan_loop PARAMS ((struct loop*, int, int));
#if 0
static void replace_call_address PARAMS ((rtx, rtx, rtx));
#endif
static rtx skip_consec_insns PARAMS ((rtx, int));
static int libcall_benefit PARAMS ((rtx));
static void ignore_some_movables PARAMS ((struct movable *));
static void force_movables PARAMS ((struct movable *));
static void combine_movables PARAMS ((struct movable *, int));
static int regs_match_p PARAMS ((rtx, rtx, struct movable *));
static int rtx_equal_for_loop_p PARAMS ((rtx, rtx, struct movable *));
static void add_label_notes PARAMS ((rtx, rtx));
static void move_movables PARAMS ((struct loop *loop, struct movable *,
int, int, int));
static int count_nonfixed_reads PARAMS ((const struct loop *, rtx));
static void strength_reduce PARAMS ((struct loop *, int, int, int));
static void find_single_use_in_loop PARAMS ((rtx, rtx, varray_type));
static int valid_initial_value_p PARAMS ((rtx, rtx, int, rtx));
static void find_mem_givs PARAMS ((const struct loop *, rtx, rtx, int, int));
static void record_biv PARAMS ((struct induction *, rtx, rtx, rtx, rtx, rtx *,
int, int, int));
static void check_final_value PARAMS ((const struct loop *,
struct induction *));
static void record_giv PARAMS ((const struct loop *, struct induction *,
rtx, rtx, rtx, rtx, rtx, int, enum g_types,
int, int, rtx *));
static void update_giv_derive PARAMS ((const struct loop *, rtx));
static int basic_induction_var PARAMS ((const struct loop *, rtx,
enum machine_mode, rtx, rtx,
rtx *, rtx *, rtx **, int *));
static rtx simplify_giv_expr PARAMS ((const struct loop *, rtx, int *));
static int general_induction_var PARAMS ((const struct loop *loop, rtx, rtx *,
rtx *, rtx *, int, int *));
static int consec_sets_giv PARAMS ((const struct loop *, int, rtx,
rtx, rtx, rtx *, rtx *, rtx *));
static int check_dbra_loop PARAMS ((struct loop *, int));
static rtx express_from_1 PARAMS ((rtx, rtx, rtx));
static rtx combine_givs_p PARAMS ((struct induction *, struct induction *));
static void combine_givs PARAMS ((struct iv_class *));
struct recombine_givs_stats;
static int find_life_end PARAMS ((rtx, struct recombine_givs_stats *,
rtx, rtx));
static void recombine_givs PARAMS ((const struct loop *, struct iv_class *,
int));
static int product_cheap_p PARAMS ((rtx, rtx));
static int maybe_eliminate_biv PARAMS ((const struct loop *, struct iv_class *,
int, int, int));
static int maybe_eliminate_biv_1 PARAMS ((const struct loop *, rtx, rtx,
struct iv_class *, int, rtx));
static int last_use_this_basic_block PARAMS ((rtx, rtx));
static void record_initial PARAMS ((rtx, rtx, void *));
static void update_reg_last_use PARAMS ((rtx, rtx));
static rtx next_insn_in_loop PARAMS ((const struct loop *, rtx));
static void load_mems_and_recount_loop_regs_set PARAMS ((const struct loop*,
int *));
static void load_mems PARAMS ((const struct loop *));
static int insert_loop_mem PARAMS ((rtx *, void *));
static int replace_loop_mem PARAMS ((rtx *, void *));
static int replace_loop_reg PARAMS ((rtx *, void *));
static void note_reg_stored PARAMS ((rtx, rtx, void *));
static void try_copy_prop PARAMS ((const struct loop *, rtx, int));
static int replace_label PARAMS ((rtx *, void *));
typedef struct rtx_and_int {
rtx r;
int i;
} rtx_and_int;
typedef struct rtx_pair {
rtx r1;
rtx r2;
} rtx_pair;
/* Nonzero iff INSN is between START and END, inclusive. */
#define INSN_IN_RANGE_P(INSN, START, END) \
(INSN_UID (INSN) < max_uid_for_loop \
&& INSN_LUID (INSN) >= INSN_LUID (START) \
&& INSN_LUID (INSN) <= INSN_LUID (END))
#ifdef HAVE_decrement_and_branch_on_count
/* Test whether BCT applicable and safe. */
static void insert_bct PARAMS ((struct loop *));
/* Auxiliary function that inserts the BCT pattern into the loop. */
static void instrument_loop_bct PARAMS ((rtx, rtx, rtx));
#endif /* HAVE_decrement_and_branch_on_count */
/* Indirect_jump_in_function is computed once per function. */
int indirect_jump_in_function = 0;
static int indirect_jump_in_function_p PARAMS ((rtx));
static int compute_luids PARAMS ((rtx, rtx, int));
static int biv_elimination_giv_has_0_offset PARAMS ((struct induction *,
struct induction *, rtx));
/* Relative gain of eliminating various kinds of operations. */
static int add_cost;
#if 0
static int shift_cost;
static int mult_cost;
#endif
/* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
copy the value of the strength reduced giv to its original register. */
static int copy_cost;
/* Cost of using a register, to normalize the benefits of a giv. */
static int reg_address_cost;
void
init_loop ()
{
char *free_point = (char *) oballoc (1);
rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
#ifdef ADDRESS_COST
reg_address_cost = ADDRESS_COST (reg);
#else
reg_address_cost = rtx_cost (reg, MEM);
#endif
/* We multiply by 2 to reconcile the difference in scale between
these two ways of computing costs. Otherwise the cost of a copy
will be far less than the cost of an add. */
copy_cost = 2 * 2;
/* Free the objects we just allocated. */
obfree (free_point);
/* Initialize the obstack used for rtl in product_cheap_p. */
gcc_obstack_init (&temp_obstack);
}
/* Compute the mapping from uids to luids.
LUIDs are numbers assigned to insns, like uids,
except that luids increase monotonically through the code.
Start at insn START and stop just before END. Assign LUIDs
starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
static int
compute_luids (start, end, prev_luid)
rtx start, end;
int prev_luid;
{
int i;
rtx insn;
for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
{
if (INSN_UID (insn) >= max_uid_for_loop)
continue;
/* Don't assign luids to line-number NOTEs, so that the distance in
luids between two insns is not affected by -g. */
if (GET_CODE (insn) != NOTE
|| NOTE_LINE_NUMBER (insn) <= 0)
uid_luid[INSN_UID (insn)] = ++i;
else
/* Give a line number note the same luid as preceding insn. */
uid_luid[INSN_UID (insn)] = i;
}
return i + 1;
}
/* Entry point of this file. Perform loop optimization
on the current function. F is the first insn of the function
and DUMPFILE is a stream for output of a trace of actions taken
(or 0 if none should be output). */
void
loop_optimize (f, dumpfile, unroll_p, bct_p)
/* f is the first instruction of a chain of insns for one function */
rtx f;
FILE *dumpfile;
int unroll_p, bct_p;
{
register rtx insn;
register int i;
struct loops loops_data;
struct loops *loops = &loops_data;
struct loop_info *loops_info;
loop_dump_stream = dumpfile;
init_recog_no_volatile ();
max_reg_before_loop = max_reg_num ();
loop_max_reg = max_reg_before_loop;
regs_may_share = 0;
/* Count the number of loops. */
max_loop_num = 0;
for (insn = f; insn; insn = NEXT_INSN (insn))
{
if (GET_CODE (insn) == NOTE
&& NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
max_loop_num++;
}
/* Don't waste time if no loops. */
if (max_loop_num == 0)
return;
loops->num = max_loop_num;
moved_once = (char *) xcalloc (max_reg_before_loop, sizeof (char));
/* Get size to use for tables indexed by uids.
Leave some space for labels allocated by find_and_verify_loops. */
max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
uid_luid = (int *) xcalloc (max_uid_for_loop, sizeof (int));
uid_loop = (struct loop **) xcalloc (max_uid_for_loop,
sizeof (struct loop *));
/* Allocate storage for array of loops. */
loops->array = (struct loop *)
xcalloc (loops->num, sizeof (struct loop));
/* Find and process each loop.
First, find them, and record them in order of their beginnings. */
find_and_verify_loops (f, loops);
/* Allocate and initialize auxiliary loop information. */
loops_info = xcalloc (loops->num, sizeof (struct loop_info));
for (i = 0; i < loops->num; i++)
loops->array[i].aux = loops_info + i;
/* Now find all register lifetimes. This must be done after
find_and_verify_loops, because it might reorder the insns in the
function. */
reg_scan (f, max_reg_before_loop, 1);
/* This must occur after reg_scan so that registers created by gcse
will have entries in the register tables.
We could have added a call to reg_scan after gcse_main in toplev.c,
but moving this call to init_alias_analysis is more efficient. */
init_alias_analysis ();
/* See if we went too far. Note that get_max_uid already returns
one more that the maximum uid of all insn. */
if (get_max_uid () > max_uid_for_loop)
abort ();
/* Now reset it to the actual size we need. See above. */
max_uid_for_loop = get_max_uid ();
/* find_and_verify_loops has already called compute_luids, but it
might have rearranged code afterwards, so we need to recompute
the luids now. */
max_luid = compute_luids (f, NULL_RTX, 0);
/* Don't leave gaps in uid_luid for insns that have been
deleted. It is possible that the first or last insn
using some register has been deleted by cross-jumping.
Make sure that uid_luid for that former insn's uid
points to the general area where that insn used to be. */
for (i = 0; i < max_uid_for_loop; i++)
{
uid_luid[0] = uid_luid[i];
if (uid_luid[0] != 0)
break;
}
for (i = 0; i < max_uid_for_loop; i++)
if (uid_luid[i] == 0)
uid_luid[i] = uid_luid[i - 1];
/* If debugging and unrolling loops, we must replicate the tree
nodes corresponding to the BLOCKs inside the loop, so that the
original one to one mapping will remain. We sometimes unroll
loops even when unroll_p is false, so we must always do this when
debugging. */
if (write_symbols != NO_DEBUG)
find_loop_tree_blocks ();
/* Determine if the function has indirect jump. On some systems
this prevents low overhead loop instructions from being used. */
indirect_jump_in_function = indirect_jump_in_function_p (f);
/* Now scan the loops, last ones first, since this means inner ones are done
before outer ones. */
for (i = max_loop_num - 1; i >= 0; i--)
{
struct loop *loop = &loops->array[i];
if (! loop->invalid && loop->end)
scan_loop (loop, unroll_p, bct_p);
}
/* Replicate the BLOCKs. */
if (write_symbols != NO_DEBUG)
unroll_block_trees ();
end_alias_analysis ();
/* Clean up. */
free (moved_once);
free (uid_luid);
free (uid_loop);
free (loops_info);
free (loops->array);
}
/* Returns the next insn, in execution order, after INSN. START and
END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
insn-stream; it is used with loops that are entered near the
bottom. */
static rtx
next_insn_in_loop (loop, insn)
const struct loop *loop;
rtx insn;
{
insn = NEXT_INSN (insn);
if (insn == loop->end)
{
if (loop->top)
/* Go to the top of the loop, and continue there. */
insn = loop->top;
else
/* We're done. */
insn = NULL_RTX;
}
if (insn == loop->scan_start)
/* We're done. */
insn = NULL_RTX;
return insn;
}
/* Optimize one loop described by LOOP. */
/* ??? Could also move memory writes out of loops if the destination address
is invariant, the source is invariant, the memory write is not volatile,
and if we can prove that no read inside the loop can read this address
before the write occurs. If there is a read of this address after the
write, then we can also mark the memory read as invariant. */
static void
scan_loop (loop, unroll_p, bct_p)
struct loop *loop;
int unroll_p, bct_p;
{
register int i;
rtx loop_start = loop->start;
rtx loop_end = loop->end;
/* Additional information about the current loop being processed
that is used to compute the number of loop iterations for loop
unrolling and doloop optimization. */
struct loop_info *loop_info = LOOP_INFO (loop);
rtx p;
/* 1 if we are scanning insns that could be executed zero times. */
int maybe_never = 0;
/* 1 if we are scanning insns that might never be executed
due to a subroutine call which might exit before they are reached. */
int call_passed = 0;
/* Jump insn that enters the loop, or 0 if control drops in. */
rtx loop_entry_jump = 0;
/* Number of insns in the loop. */
int insn_count;
int in_libcall = 0;
int tem;
rtx temp, update_start, update_end;
/* The SET from an insn, if it is the only SET in the insn. */
rtx set, set1;
/* Chain describing insns movable in current loop. */
struct movable *movables = 0;
/* Last element in `movables' -- so we can add elements at the end. */
struct movable *last_movable = 0;
/* Ratio of extra register life span we can justify
for saving an instruction. More if loop doesn't call subroutines
since in that case saving an insn makes more difference
and more registers are available. */
int threshold;
/* Nonzero if we are scanning instructions in a sub-loop. */
int loop_depth = 0;
int nregs;
loop->top = 0;
/* Determine whether this loop starts with a jump down to a test at
the end. This will occur for a small number of loops with a test
that is too complex to duplicate in front of the loop.
We search for the first insn or label in the loop, skipping NOTEs.
However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
(because we might have a loop executed only once that contains a
loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
(in case we have a degenerate loop).
Note that if we mistakenly think that a loop is entered at the top
when, in fact, it is entered at the exit test, the only effect will be
slightly poorer optimization. Making the opposite error can generate
incorrect code. Since very few loops now start with a jump to the
exit test, the code here to detect that case is very conservative. */
for (p = NEXT_INSN (loop_start);
p != loop_end
&& GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
&& (GET_CODE (p) != NOTE
|| (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
&& NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
p = NEXT_INSN (p))
;
loop->scan_start = p;
/* Set up variables describing this loop. */
prescan_loop (loop);
threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
/* If loop has a jump before the first label,
the true entry is the target of that jump.
Start scan from there.
But record in LOOP->TOP the place where the end-test jumps
back to so we can scan that after the end of the loop. */
if (GET_CODE (p) == JUMP_INSN)
{
loop_entry_jump = p;
/* Loop entry must be unconditional jump (and not a RETURN) */
if (simplejump_p (p)
&& JUMP_LABEL (p) != 0
/* Check to see whether the jump actually
jumps out of the loop (meaning it's no loop).
This case can happen for things like
do {..} while (0). If this label was generated previously
by loop, we can't tell anything about it and have to reject
the loop. */
&& INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
{
loop->top = next_label (loop->scan_start);
loop->scan_start = JUMP_LABEL (p);
}
}
/* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
as required by loop_reg_used_before_p. So skip such loops. (This
test may never be true, but it's best to play it safe.)
Also, skip loops where we do not start scanning at a label. This
test also rejects loops starting with a JUMP_INSN that failed the
test above. */
if (INSN_UID (loop->scan_start) >= max_uid_for_loop
|| GET_CODE (loop->scan_start) != CODE_LABEL)
{
if (loop_dump_stream)
fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
INSN_UID (loop_start), INSN_UID (loop_end));
return;
}
/* Count number of times each reg is set during this loop.
Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
/* Allocate extra space for REGS that might be created by
load_mems. We allocate a little extra slop as well, in the hopes
that even after the moving of movables creates some new registers
we won't have to reallocate these arrays. However, we do grow
the arrays, if necessary, in load_mems_recount_loop_regs_set. */
nregs = max_reg_num () + loop_mems_idx + 16;
VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
may_not_optimize, reg_single_usage, &insn_count, nregs);
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
{
VARRAY_CHAR (may_not_optimize, i) = 1;
VARRAY_INT (set_in_loop, i) = 1;
}
#ifdef AVOID_CCMODE_COPIES
/* Don't try to move insns which set CC registers if we should not
create CCmode register copies. */
for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
VARRAY_CHAR (may_not_optimize, i) = 1;
#endif
bcopy ((char *) &set_in_loop->data,
(char *) &n_times_set->data, nregs * sizeof (int));
if (loop_dump_stream)
{
fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
if (loop->cont)
fprintf (loop_dump_stream, "Continue at insn %d.\n",
INSN_UID (loop->cont));
}
/* Scan through the loop finding insns that are safe to move.
Set set_in_loop negative for the reg being set, so that
this reg will be considered invariant for subsequent insns.
We consider whether subsequent insns use the reg
in deciding whether it is worth actually moving.
MAYBE_NEVER is nonzero if we have passed a conditional jump insn
and therefore it is possible that the insns we are scanning
would never be executed. At such times, we must make sure
that it is safe to execute the insn once instead of zero times.
When MAYBE_NEVER is 0, all insns will be executed at least once
so that is not a problem. */
for (p = next_insn_in_loop (loop, loop->scan_start);
p != NULL_RTX;
p = next_insn_in_loop (loop, p))
{
if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
&& find_reg_note (p, REG_LIBCALL, NULL_RTX))
in_libcall = 1;
else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
&& find_reg_note (p, REG_RETVAL, NULL_RTX))
in_libcall = 0;
if (GET_CODE (p) == INSN
&& (set = single_set (p))
&& GET_CODE (SET_DEST (set)) == REG
&& ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
{
int tem1 = 0;
int tem2 = 0;
int move_insn = 0;
rtx src = SET_SRC (set);
rtx dependencies = 0;
/* Figure out what to use as a source of this insn. If a REG_EQUIV
note is given or if a REG_EQUAL note with a constant operand is
specified, use it as the source and mark that we should move
this insn by calling emit_move_insn rather that duplicating the
insn.
Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
is present. */
temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
if (temp)
src = XEXP (temp, 0), move_insn = 1;
else
{
temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
if (temp && CONSTANT_P (XEXP (temp, 0)))
src = XEXP (temp, 0), move_insn = 1;
if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
{
src = XEXP (temp, 0);
/* A libcall block can use regs that don't appear in
the equivalent expression. To move the libcall,
we must move those regs too. */
dependencies = libcall_other_reg (p, src);
}
}
/* Don't try to optimize a register that was made
by loop-optimization for an inner loop.
We don't know its life-span, so we can't compute the benefit. */
if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
;
else if (/* The register is used in basic blocks other
than the one where it is set (meaning that
something after this point in the loop might
depend on its value before the set). */
! reg_in_basic_block_p (p, SET_DEST (set))
/* And the set is not guaranteed to be executed one
the loop starts, or the value before the set is
needed before the set occurs...
??? Note we have quadratic behaviour here, mitigated
by the fact that the previous test will often fail for
large loops. Rather than re-scanning the entire loop
each time for register usage, we should build tables
of the register usage and use them here instead. */
&& (maybe_never
|| loop_reg_used_before_p (loop, set, p)))
/* It is unsafe to move the set.
This code used to consider it OK to move a set of a variable
which was not created by the user and not used in an exit test.
That behavior is incorrect and was removed. */
;
else if ((tem = loop_invariant_p (loop, src))
&& (dependencies == 0
|| (tem2 = loop_invariant_p (loop, dependencies)) != 0)
&& (VARRAY_INT (set_in_loop,
REGNO (SET_DEST (set))) == 1
|| (tem1
= consec_sets_invariant_p
(loop, SET_DEST (set),
VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
p)))
/* If the insn can cause a trap (such as divide by zero),
can't move it unless it's guaranteed to be executed
once loop is entered. Even a function call might
prevent the trap insn from being reached
(since it might exit!) */
&& ! ((maybe_never || call_passed)
&& may_trap_p (src)))
{
register struct movable *m;
register int regno = REGNO (SET_DEST (set));
/* A potential lossage is where we have a case where two insns
can be combined as long as they are both in the loop, but
we move one of them outside the loop. For large loops,
this can lose. The most common case of this is the address
of a function being called.
Therefore, if this register is marked as being used exactly
once if we are in a loop with calls (a "large loop"), see if
we can replace the usage of this register with the source
of this SET. If we can, delete this insn.
Don't do this if P has a REG_RETVAL note or if we have
SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
if (loop_info->has_call
&& VARRAY_RTX (reg_single_usage, regno) != 0
&& VARRAY_RTX (reg_single_usage, regno) != const0_rtx
&& REGNO_FIRST_UID (regno) == INSN_UID (p)
&& (REGNO_LAST_UID (regno)
== INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
&& VARRAY_INT (set_in_loop, regno) == 1
&& ! side_effects_p (SET_SRC (set))
&& ! find_reg_note (p, REG_RETVAL, NULL_RTX)
&& (! SMALL_REGISTER_CLASSES
|| (! (GET_CODE (SET_SRC (set)) == REG
&& REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
/* This test is not redundant; SET_SRC (set) might be
a call-clobbered register and the life of REGNO
might span a call. */
&& ! modified_between_p (SET_SRC (set), p,
VARRAY_RTX
(reg_single_usage, regno))
&& no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
&& validate_replace_rtx (SET_DEST (set), SET_SRC (set),
VARRAY_RTX
(reg_single_usage, regno)))
{
/* Replace any usage in a REG_EQUAL note. Must copy the
new source, so that we don't get rtx sharing between the
SET_SOURCE and REG_NOTES of insn p. */
REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
= replace_rtx (REG_NOTES (VARRAY_RTX
(reg_single_usage, regno)),
SET_DEST (set), copy_rtx (SET_SRC (set)));
PUT_CODE (p, NOTE);
NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
NOTE_SOURCE_FILE (p) = 0;
VARRAY_INT (set_in_loop, regno) = 0;
continue;
}
m = (struct movable *) alloca (sizeof (struct movable));
m->next = 0;
m->insn = p;
m->set_src = src;
m->dependencies = dependencies;
m->set_dest = SET_DEST (set);
m->force = 0;
m->consec = VARRAY_INT (set_in_loop,
REGNO (SET_DEST (set))) - 1;
m->done = 0;
m->forces = 0;
m->partial = 0;
m->move_insn = move_insn;
m->move_insn_first = 0;
m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
m->savemode = VOIDmode;
m->regno = regno;
/* Set M->cond if either loop_invariant_p
or consec_sets_invariant_p returned 2
(only conditionally invariant). */
m->cond = ((tem | tem1 | tem2) > 1);
m->global = (uid_luid[REGNO_LAST_UID (regno)]
> INSN_LUID (loop_end)
|| uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
m->match = 0;
m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
- uid_luid[REGNO_FIRST_UID (regno)]);
m->savings = VARRAY_INT (n_times_set, regno);
if (find_reg_note (p, REG_RETVAL, NULL_RTX))
m->savings += libcall_benefit (p);
VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
/* Add M to the end of the chain MOVABLES. */
if (movables == 0)
movables = m;
else
last_movable->next = m;
last_movable = m;
if (m->consec > 0)
{
/* It is possible for the first instruction to have a
REG_EQUAL note but a non-invariant SET_SRC, so we must
remember the status of the first instruction in case
the last instruction doesn't have a REG_EQUAL note. */
m->move_insn_first = m->move_insn;
/* Skip this insn, not checking REG_LIBCALL notes. */
p = next_nonnote_insn (p);
/* Skip the consecutive insns, if there are any. */
p = skip_consec_insns (p, m->consec);
/* Back up to the last insn of the consecutive group. */
p = prev_nonnote_insn (p);
/* We must now reset m->move_insn, m->is_equiv, and possibly
m->set_src to correspond to the effects of all the
insns. */
temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
if (temp)
m->set_src = XEXP (temp, 0), m->move_insn = 1;
else
{
temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
if (temp && CONSTANT_P (XEXP (temp, 0)))
m->set_src = XEXP (temp, 0), m->move_insn = 1;
else
m->move_insn = 0;
}
m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
}
}
/* If this register is always set within a STRICT_LOW_PART
or set to zero, then its high bytes are constant.
So clear them outside the loop and within the loop
just load the low bytes.
We must check that the machine has an instruction to do so.
Also, if the value loaded into the register
depends on the same register, this cannot be done. */
else if (SET_SRC (set) == const0_rtx
&& GET_CODE (NEXT_INSN (p)) == INSN
&& (set1 = single_set (NEXT_INSN (p)))
&& GET_CODE (set1) == SET
&& (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
&& (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
&& (SUBREG_REG (XEXP (SET_DEST (set1), 0))
== SET_DEST (set))
&& !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
{
register int regno = REGNO (SET_DEST (set));
if (VARRAY_INT (set_in_loop, regno) == 2)
{
register struct movable *m;
m = (struct movable *) alloca (sizeof (struct movable));
m->next = 0;
m->insn = p;
m->set_dest = SET_DEST (set);
m->dependencies = 0;
m->force = 0;
m->consec = 0;
m->done = 0;
m->forces = 0;
m->move_insn = 0;
m->move_insn_first = 0;
m->partial = 1;
/* If the insn may not be executed on some cycles,
we can't clear the whole reg; clear just high part.
Not even if the reg is used only within this loop.
Consider this:
while (1)
while (s != t) {
if (foo ()) x = *s;
use (x);
}
Clearing x before the inner loop could clobber a value
being saved from the last time around the outer loop.
However, if the reg is not used outside this loop
and all uses of the register are in the same
basic block as the store, there is no problem.
If this insn was made by loop, we don't know its
INSN_LUID and hence must make a conservative
assumption. */
m->global = (INSN_UID (p) >= max_uid_for_loop
|| (uid_luid[REGNO_LAST_UID (regno)]
> INSN_LUID (loop_end))
|| (uid_luid[REGNO_FIRST_UID (regno)]
< INSN_LUID (p))
|| (labels_in_range_p
(p, uid_luid[REGNO_FIRST_UID (regno)])));
if (maybe_never && m->global)
m->savemode = GET_MODE (SET_SRC (set1));
else
m->savemode = VOIDmode;
m->regno = regno;
m->cond = 0;
m->match = 0;
m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
- uid_luid[REGNO_FIRST_UID (regno)]);
m->savings = 1;
VARRAY_INT (set_in_loop, regno) = -1;
/* Add M to the end of the chain MOVABLES. */
if (movables == 0)
movables = m;
else
last_movable->next = m;
last_movable = m;
}
}
}
/* Past a call insn, we get to insns which might not be executed
because the call might exit. This matters for insns that trap.
Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
so they don't count. */
else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
call_passed = 1;
/* Past a label or a jump, we get to insns for which we
can't count on whether or how many times they will be
executed during each iteration. Therefore, we can
only move out sets of trivial variables
(those not used after the loop). */
/* Similar code appears twice in strength_reduce. */
else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
/* If we enter the loop in the middle, and scan around to the
beginning, don't set maybe_never for that. This must be an
unconditional jump, otherwise the code at the top of the
loop might never be executed. Unconditional jumps are
followed a by barrier then loop end. */
&& ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop->top
&& NEXT_INSN (NEXT_INSN (p)) == loop_end
&& simplejump_p (p)))
maybe_never = 1;
else if (GET_CODE (p) == NOTE)
{
/* At the virtual top of a converted loop, insns are again known to
be executed: logically, the loop begins here even though the exit
code has been duplicated. */
if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
maybe_never = call_passed = 0;
else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
loop_depth++;
else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
loop_depth--;
}
}
/* If one movable subsumes another, ignore that other. */
ignore_some_movables (movables);
/* For each movable insn, see if the reg that it loads
leads when it dies right into another conditionally movable insn.
If so, record that the second insn "forces" the first one,
since the second can be moved only if the first is. */
force_movables (movables);
/* See if there are multiple movable insns that load the same value.
If there are, make all but the first point at the first one
through the `match' field, and add the priorities of them
all together as the priority of the first. */
combine_movables (movables, nregs);
/* Now consider each movable insn to decide whether it is worth moving.
Store 0 in set_in_loop for each reg that is moved.
Generally this increases code size, so do not move moveables when
optimizing for code size. */
if (! optimize_size)
move_movables (loop, movables, threshold, insn_count, nregs);
/* Now candidates that still are negative are those not moved.
Change set_in_loop to indicate that those are not actually invariant. */
for (i = 0; i < nregs; i++)
if (VARRAY_INT (set_in_loop, i) < 0)
VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
/* Now that we've moved some things out of the loop, we might be able to
hoist even more memory references. */
load_mems_and_recount_loop_regs_set (loop, &insn_count);
for (update_start = loop_start;
PREV_INSN (update_start)
&& GET_CODE (PREV_INSN (update_start)) != CODE_LABEL;
update_start = PREV_INSN (update_start))
;
update_end = NEXT_INSN (loop_end);
reg_scan_update (update_start, update_end, loop_max_reg);
loop_max_reg = max_reg_num ();
if (flag_strength_reduce)
{
the_movables = movables;
strength_reduce (loop, insn_count, unroll_p, bct_p);
reg_scan_update (update_start, update_end, loop_max_reg);
loop_max_reg = max_reg_num ();
}
VARRAY_FREE (reg_single_usage);
VARRAY_FREE (set_in_loop);
VARRAY_FREE (n_times_set);
VARRAY_FREE (may_not_optimize);
}
/* Add elements to *OUTPUT to record all the pseudo-regs
mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
void
record_excess_regs (in_this, not_in_this, output)
rtx in_this, not_in_this;
rtx *output;
{
enum rtx_code code;
const char *fmt;
int i;
code = GET_CODE (in_this);
switch (code)
{
case PC:
case CC0:
case CONST_INT:
case CONST_DOUBLE:
case CONST:
case SYMBOL_REF:
case LABEL_REF:
return;
case REG:
if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
&& ! reg_mentioned_p (in_this, not_in_this))
*output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
return;
default:
break;
}
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
int j;
switch (fmt[i])
{
case 'E':
for (j = 0; j < XVECLEN (in_this, i); j++)
record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
break;
case 'e':
record_excess_regs (XEXP (in_this, i), not_in_this, output);
break;
}
}
}
/* Check what regs are referred to in the libcall block ending with INSN,
aside from those mentioned in the equivalent value.
If there are none, return 0.
If there are one or more, return an EXPR_LIST containing all of them. */
rtx
libcall_other_reg (insn, equiv)
rtx insn, equiv;
{
rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
rtx p = XEXP (note, 0);
rtx output = 0;
/* First, find all the regs used in the libcall block
that are not mentioned as inputs to the result. */
while (p != insn)
{
if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
|| GET_CODE (p) == CALL_INSN)
record_excess_regs (PATTERN (p), equiv, &output);
p = NEXT_INSN (p);
}
return output;
}
/* Return 1 if all uses of REG
are between INSN and the end of the basic block. */
static int
reg_in_basic_block_p (insn, reg)
rtx insn, reg;
{
int regno = REGNO (reg);
rtx p;
if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
return 0;
/* Search this basic block for the already recorded last use of the reg. */
for (p = insn; p; p = NEXT_INSN (p))
{
switch (GET_CODE (p))
{
case NOTE:
break;
case INSN:
case CALL_INSN:
/* Ordinary insn: if this is the last use, we win. */
if (REGNO_LAST_UID (regno) == INSN_UID (p))
return 1;
break;
case JUMP_INSN:
/* Jump insn: if this is the last use, we win. */
if (REGNO_LAST_UID (regno) == INSN_UID (p))
return 1;
/* Otherwise, it's the end of the basic block, so we lose. */
return 0;
case CODE_LABEL:
case BARRIER:
/* It's the end of the basic block, so we lose. */
return 0;
default:
break;
}
}
/* The "last use" doesn't follow the "first use"?? */
abort ();
}
/* Compute the benefit of eliminating the insns in the block whose
last insn is LAST. This may be a group of insns used to compute a
value directly or can contain a library call. */
static int
libcall_benefit (last)
rtx last;
{
rtx insn;
int benefit = 0;
for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
insn != last; insn = NEXT_INSN (insn))
{
if (GET_CODE (insn) == CALL_INSN)
benefit += 10; /* Assume at least this many insns in a library
routine. */
else if (GET_CODE (insn) == INSN
&& GET_CODE (PATTERN (insn)) != USE
&& GET_CODE (PATTERN (insn)) != CLOBBER)
benefit++;
}
return benefit;
}
/* Skip COUNT insns from INSN, counting library calls as 1 insn. */
static rtx
skip_consec_insns (insn, count)
rtx insn;
int count;
{
for (; count > 0; count--)
{
rtx temp;
/* If first insn of libcall sequence, skip to end. */
/* Do this at start of loop, since INSN is guaranteed to
be an insn here. */
if (GET_CODE (insn) != NOTE
&& (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
insn = XEXP (temp, 0);
do insn = NEXT_INSN (insn);
while (GET_CODE (insn) == NOTE);
}
return insn;
}
/* Ignore any movable whose insn falls within a libcall
which is part of another movable.
We make use of the fact that the movable for the libcall value
was made later and so appears later on the chain. */
static void
ignore_some_movables (movables)
struct movable *movables;
{
register struct movable *m, *m1;
for (m = movables; m; m = m->next)
{
/* Is this a movable for the value of a libcall? */
rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
if (note)
{
rtx insn;
/* Check for earlier movables inside that range,
and mark them invalid. We cannot use LUIDs here because
insns created by loop.c for prior loops don't have LUIDs.
Rather than reject all such insns from movables, we just
explicitly check each insn in the libcall (since invariant
libcalls aren't that common). */
for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
for (m1 = movables; m1 != m; m1 = m1->next)
if (m1->insn == insn)
m1->done = 1;
}
}
}
/* For each movable insn, see if the reg that it loads
leads when it dies right into another conditionally movable insn.
If so, record that the second insn "forces" the first one,
since the second can be moved only if the first is. */
static void
force_movables (movables)
struct movable *movables;
{
register struct movable *m, *m1;
for (m1 = movables; m1; m1 = m1->next)
/* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
if (!m1->partial && !m1->done)
{
int regno = m1->regno;
for (m = m1->next; m; m = m->next)
/* ??? Could this be a bug? What if CSE caused the
register of M1 to be used after this insn?
Since CSE does not update regno_last_uid,
this insn M->insn might not be where it dies.
But very likely this doesn't matter; what matters is
that M's reg is computed from M1's reg. */
if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
&& !m->done)
break;
if (m != 0 && m->set_src == m1->set_dest
/* If m->consec, m->set_src isn't valid. */
&& m->consec == 0)
m = 0;
/* Increase the priority of the moving the first insn
since it permits the second to be moved as well. */
if (m != 0)
{
m->forces = m1;
m1->lifetime += m->lifetime;
m1->savings += m->savings;
}
}
}
/* Find invariant expressions that are equal and can be combined into
one register. */
static void
combine_movables (movables, nregs)
struct movable *movables;
int nregs;
{
register struct movable *m;
char *matched_regs = (char *) xmalloc (nregs);
enum machine_mode mode;
/* Regs that are set more than once are not allowed to match
or be matched. I'm no longer sure why not. */
/* Perhaps testing m->consec_sets would be more appropriate here? */
for (m = movables; m; m = m->next)
if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
{
register struct movable *m1;
int regno = m->regno;
bzero (matched_regs, nregs);
matched_regs[regno] = 1;
/* We want later insns to match the first one. Don't make the first
one match any later ones. So start this loop at m->next. */
for (m1 = m->next; m1; m1 = m1->next)
if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
/* A reg used outside the loop mustn't be eliminated. */
&& !m1->global
/* A reg used for zero-extending mustn't be eliminated. */
&& !m1->partial
&& (matched_regs[m1->regno]
||
(
/* Can combine regs with different modes loaded from the
same constant only if the modes are the same or
if both are integer modes with M wider or the same
width as M1. The check for integer is redundant, but
safe, since the only case of differing destination
modes with equal sources is when both sources are
VOIDmode, i.e., CONST_INT. */
(GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
|| (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
&& GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
&& (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
>= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
/* See if the source of M1 says it matches M. */
&& ((GET_CODE (m1->set_src) == REG
&& matched_regs[REGNO (m1->set_src)])
|| rtx_equal_for_loop_p (m->set_src, m1->set_src,
movables))))
&& ((m->dependencies == m1->dependencies)
|| rtx_equal_p (m->dependencies, m1->dependencies)))
{
m->lifetime += m1->lifetime;
m->savings += m1->savings;
m1->done = 1;
m1->match = m;
matched_regs[m1->regno] = 1;
}
}
/* Now combine the regs used for zero-extension.
This can be done for those not marked `global'
provided their lives don't overlap. */
for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
mode = GET_MODE_WIDER_MODE (mode))
{
register struct movable *m0 = 0;
/* Combine all the registers for extension from mode MODE.
Don't combine any that are used outside this loop. */
for (m = movables; m; m = m->next)
if (m->partial && ! m->global
&& mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
{
register struct movable *m1;
int first = uid_luid[REGNO_FIRST_UID (m->regno)];
int last = uid_luid[REGNO_LAST_UID (m->regno)];
if (m0 == 0)
{
/* First one: don't check for overlap, just record it. */
m0 = m;
continue;
}
/* Make sure they extend to the same mode.
(Almost always true.) */
if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
continue;
/* We already have one: check for overlap with those
already combined together. */
for (m1 = movables; m1 != m; m1 = m1->next)
if (m1 == m0 || (m1->partial && m1->match == m0))
if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
|| uid_luid[REGNO_LAST_UID (m1->regno)] < first))
goto overlap;
/* No overlap: we can combine this with the others. */
m0->lifetime += m->lifetime;
m0->savings += m->savings;
m->done = 1;
m->match = m0;
overlap: ;
}
}
/* Clean up. */
free (matched_regs);
}
/* Return 1 if regs X and Y will become the same if moved. */
static int
regs_match_p (x, y, movables)
rtx x, y;
struct movable *movables;
{
int xn = REGNO (x);
int yn = REGNO (y);
struct movable *mx, *my;
for (mx = movables; mx; mx = mx->next)
if (mx->regno == xn)
break;
for (my = movables; my; my = my->next)
if (my->regno == yn)
break;
return (mx && my
&& ((mx->match == my->match && mx->match != 0)
|| mx->match == my
|| mx == my->match));
}
/* Return 1 if X and Y are identical-looking rtx's.
This is the Lisp function EQUAL for rtx arguments.
If two registers are matching movables or a movable register and an
equivalent constant, consider them equal. */
static int
rtx_equal_for_loop_p (x, y, movables)
rtx x, y;
struct movable *movables;
{
register int i;
register int j;
register struct movable *m;
register enum rtx_code code;
register const char *fmt;
if (x == y)
return 1;
if (x == 0 || y == 0)
return 0;
code = GET_CODE (x);
/* If we have a register and a constant, they may sometimes be
equal. */
if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
&& CONSTANT_P (y))
{
for (m = movables; m; m = m->next)
if (m->move_insn && m->regno == REGNO (x)
&& rtx_equal_p (m->set_src, y))
return 1;
}
else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
&& CONSTANT_P (x))
{
for (m = movables; m; m = m->next)
if (m->move_insn && m->regno == REGNO (y)
&& rtx_equal_p (m->set_src, x))
return 1;
}
/* Otherwise, rtx's of different codes cannot be equal. */
if (code != GET_CODE (y))
return 0;
/* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
(REG:SI x) and (REG:HI x) are NOT equivalent. */
if (GET_MODE (x) != GET_MODE (y))
return 0;
/* These three types of rtx's can be compared nonrecursively. */
if (code == REG)
return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
if (code == LABEL_REF)
return XEXP (x, 0) == XEXP (y, 0);
if (code == SYMBOL_REF)
return XSTR (x, 0) == XSTR (y, 0);
/* Compare the elements. If any pair of corresponding elements
fail to match, return 0 for the whole things. */
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
switch (fmt[i])
{
case 'w':
if (XWINT (x, i) != XWINT (y, i))
return 0;
break;
case 'i':
if (XINT (x, i) != XINT (y, i))
return 0;
break;
case 'E':
/* Two vectors must have the same length. */
if (XVECLEN (x, i) != XVECLEN (y, i))
return 0;
/* And the corresponding elements must match. */
for (j = 0; j < XVECLEN (x, i); j++)
if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
return 0;
break;
case 'e':
if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
return 0;
break;
case 's':
if (strcmp (XSTR (x, i), XSTR (y, i)))
return 0;
break;
case 'u':
/* These are just backpointers, so they don't matter. */
break;
case '0':
break;
/* It is believed that rtx's at this level will never
contain anything but integers and other rtx's,
except for within LABEL_REFs and SYMBOL_REFs. */
default:
abort ();
}
}
return 1;
}
/* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
insns in INSNS which use the reference. */
static void
add_label_notes (x, insns)
rtx x;
rtx insns;
{
enum rtx_code code = GET_CODE (x);
int i, j;
const char *fmt;
rtx insn;
if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
{
/* This code used to ignore labels that referred to dispatch tables to
avoid flow generating (slighly) worse code.
We no longer ignore such label references (see LABEL_REF handling in
mark_jump_label for additional information). */
for (insn = insns; insn; insn = NEXT_INSN (insn))
if (reg_mentioned_p (XEXP (x, 0), insn))
REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
REG_NOTES (insn));
}
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
add_label_notes (XEXP (x, i), insns);
else if (fmt[i] == 'E')
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
add_label_notes (XVECEXP (x, i, j), insns);
}
}
/* Scan MOVABLES, and move the insns that deserve to be moved.
If two matching movables are combined, replace one reg with the
other throughout. */
static void
move_movables (loop, movables, threshold, insn_count, nregs)
struct loop *loop;
struct movable *movables;
int threshold;
int insn_count;
int nregs;
{
rtx new_start = 0;
register struct movable *m;
register rtx p;
rtx loop_start = loop->start;
rtx loop_end = loop->end;
/* Map of pseudo-register replacements to handle combining
when we move several insns that load the same value
into different pseudo-registers. */
rtx *reg_map = (rtx *) xcalloc (nregs, sizeof (rtx));
char *already_moved = (char *) xcalloc (nregs, sizeof (char));
num_movables = 0;
for (m = movables; m; m = m->next)
{
/* Describe this movable insn. */
if (loop_dump_stream)
{
fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
INSN_UID (m->insn), m->regno, m->lifetime);
if (m->consec > 0)
fprintf (loop_dump_stream, "consec %d, ", m->consec);
if (m->cond)
fprintf (loop_dump_stream, "cond ");
if (m->force)
fprintf (loop_dump_stream, "force ");
if (m->global)
fprintf (loop_dump_stream, "global ");
if (m->done)
fprintf (loop_dump_stream, "done ");
if (m->move_insn)
fprintf (loop_dump_stream, "move-insn ");
if (m->match)
fprintf (loop_dump_stream, "matches %d ",
INSN_UID (m->match->insn));
if (m->forces)
fprintf (loop_dump_stream, "forces %d ",
INSN_UID (m->forces->insn));
}
/* Count movables. Value used in heuristics in strength_reduce. */
num_movables++;
/* Ignore the insn if it's already done (it matched something else).
Otherwise, see if it is now safe to move. */
if (!m->done
&& (! m->cond
|| (1 == loop_invariant_p (loop, m->set_src)
&& (m->dependencies == 0
|| 1 == loop_invariant_p (loop, m->dependencies))
&& (m->consec == 0
|| 1 == consec_sets_invariant_p (loop, m->set_dest,
m->consec + 1,
m->insn))))
&& (! m->forces || m->forces->done))
{
register int regno;
register rtx p;
int savings = m->savings;
/* We have an insn that is safe to move.
Compute its desirability. */
p = m->insn;
regno = m->regno;
if (loop_dump_stream)
fprintf (loop_dump_stream, "savings %d ", savings);
if (moved_once[regno] && loop_dump_stream)
fprintf (loop_dump_stream, "halved since already moved ");
/* An insn MUST be moved if we already moved something else
which is safe only if this one is moved too: that is,
if already_moved[REGNO] is nonzero. */
/* An insn is desirable to move if the new lifetime of the
register is no more than THRESHOLD times the old lifetime.
If it's not desirable, it means the loop is so big
that moving won't speed things up much,
and it is liable to make register usage worse. */
/* It is also desirable to move if it can be moved at no
extra cost because something else was already moved. */
if (already_moved[regno]
|| flag_move_all_movables
|| (threshold * savings * m->lifetime) >=
(moved_once[regno] ? insn_count * 2 : insn_count)
|| (m->forces && m->forces->done
&& VARRAY_INT (n_times_set, m->forces->regno) == 1))
{
int count;
register struct movable *m1;
rtx first = NULL_RTX;
/* Now move the insns that set the reg. */
if (m->partial && m->match)
{
rtx newpat, i1;
rtx r1, r2;
/* Find the end of this chain of matching regs.
Thus, we load each reg in the chain from that one reg.
And that reg is loaded with 0 directly,
since it has ->match == 0. */
for (m1 = m; m1->match; m1 = m1->match);
newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
SET_DEST (PATTERN (m1->insn)));
i1 = emit_insn_before (newpat, loop_start);
/* Mark the moved, invariant reg as being allowed to
share a hard reg with the other matching invariant. */
REG_NOTES (i1) = REG_NOTES (m->insn);
r1 = SET_DEST (PATTERN (m->insn));
r2 = SET_DEST (PATTERN (m1->insn));
regs_may_share
= gen_rtx_EXPR_LIST (VOIDmode, r1,
gen_rtx_EXPR_LIST (VOIDmode, r2,
regs_may_share));
delete_insn (m->insn);
if (new_start == 0)
new_start = i1;
if (loop_dump_stream)
fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
}
/* If we are to re-generate the item being moved with a
new move insn, first delete what we have and then emit
the move insn before the loop. */
else if (m->move_insn)
{
rtx i1, temp;
for (count = m->consec; count >= 0; count--)
{
/* If this is the first insn of a library call sequence,
skip to the end. */
if (GET_CODE (p) != NOTE
&& (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
p = XEXP (temp, 0);
/* If this is the last insn of a libcall sequence, then
delete every insn in the sequence except the last.
The last insn is handled in the normal manner. */
if (GET_CODE (p) != NOTE
&& (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
{
temp = XEXP (temp, 0);
while (temp != p)
temp = delete_insn (temp);
}
temp = p;
p = delete_insn (p);
/* simplify_giv_expr expects that it can walk the insns
at m->insn forwards and see this old sequence we are
tossing here. delete_insn does preserve the next
pointers, but when we skip over a NOTE we must fix
it up. Otherwise that code walks into the non-deleted
insn stream. */
while (p && GET_CODE (p) == NOTE)
p = NEXT_INSN (temp) = NEXT_INSN (p);
}
start_sequence ();
emit_move_insn (m->set_dest, m->set_src);
temp = get_insns ();
end_sequence ();
add_label_notes (m->set_src, temp);
i1 = emit_insns_before (temp, loop_start);
if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
REG_NOTES (i1)
= gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
m->set_src, REG_NOTES (i1));
if (loop_dump_stream)
fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
/* The more regs we move, the less we like moving them. */
threshold -= 3;
}
else
{
for (count = m->consec; count >= 0; count--)
{
rtx i1, temp;
/* If first insn of libcall sequence, skip to end. */
/* Do this at start of loop, since p is guaranteed to
be an insn here. */
if (GET_CODE (p) != NOTE
&& (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
p = XEXP (temp, 0);
/* If last insn of libcall sequence, move all
insns except the last before the loop. The last
insn is handled in the normal manner. */
if (GET_CODE (p) != NOTE
&& (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
{
rtx fn_address = 0;
rtx fn_reg = 0;
rtx fn_address_insn = 0;
first = 0;
for (temp = XEXP (temp, 0); temp != p;
temp = NEXT_INSN (temp))
{
rtx body;
rtx n;
rtx next;
if (GET_CODE (temp) == NOTE)
continue;
body = PATTERN (temp);
/* Find the next insn after TEMP,
not counting USE or NOTE insns. */
for (next = NEXT_INSN (temp); next != p;
next = NEXT_INSN (next))
if (! (GET_CODE (next) == INSN
&& GET_CODE (PATTERN (next)) == USE)
&& GET_CODE (next) != NOTE)
break;
/* If that is the call, this may be the insn
that loads the function address.
Extract the function address from the insn
that loads it into a register.
If this insn was cse'd, we get incorrect code.
So emit a new move insn that copies the
function address into the register that the
call insn will use. flow.c will delete any
redundant stores that we have created. */
if (GET_CODE (next) == CALL_INSN
&& GET_CODE (body) == SET
&& GET_CODE (SET_DEST (body)) == REG
&& (n = find_reg_note (temp, REG_EQUAL,
NULL_RTX)))
{
fn_reg = SET_SRC (body);
if (GET_CODE (fn_reg) != REG)
fn_reg = SET_DEST (body);
fn_address = XEXP (n, 0);
fn_address_insn = temp;
}
/* We have the call insn.
If it uses the register we suspect it might,
load it with the correct address directly. */
if (GET_CODE (temp) == CALL_INSN
&& fn_address != 0
&& reg_referenced_p (fn_reg, body))
emit_insn_after (gen_move_insn (fn_reg,
fn_address),
fn_address_insn);
if (GET_CODE (temp) == CALL_INSN)
{
i1 = emit_call_insn_before (body, loop_start);
/* Because the USAGE information potentially
contains objects other than hard registers
we need to copy it. */
if (CALL_INSN_FUNCTION_USAGE (temp))
CALL_INSN_FUNCTION_USAGE (i1)
= copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
}
else
i1 = emit_insn_before (body, loop_start);
if (first == 0)
first = i1;
if (temp == fn_address_insn)
fn_address_insn = i1;
REG_NOTES (i1) = REG_NOTES (temp);
delete_insn (temp);
}
if (new_start == 0)
new_start = first;
}
if (m->savemode != VOIDmode)
{
/* P sets REG to zero; but we should clear only
the bits that are not covered by the mode
m->savemode. */
rtx reg = m->set_dest;
rtx sequence;
rtx tem;
start_sequence ();
tem = expand_binop
(GET_MODE (reg), and_optab, reg,
GEN_INT ((((HOST_WIDE_INT) 1
<< GET_MODE_BITSIZE (m->savemode)))
- 1),
reg, 1, OPTAB_LIB_WIDEN);
if (tem == 0)
abort ();
if (tem != reg)
emit_move_insn (reg, tem);
sequence = gen_sequence ();
end_sequence ();
i1 = emit_insn_before (sequence, loop_start);
}
else if (GET_CODE (p) == CALL_INSN)
{
i1 = emit_call_insn_before (PATTERN (p), loop_start);
/* Because the USAGE information potentially
contains objects other than hard registers
we need to copy it. */
if (CALL_INSN_FUNCTION_USAGE (p))
CALL_INSN_FUNCTION_USAGE (i1)
= copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
}
else if (count == m->consec && m->move_insn_first)
{
/* The SET_SRC might not be invariant, so we must
use the REG_EQUAL note. */
start_sequence ();
emit_move_insn (m->set_dest, m->set_src);
temp = get_insns ();
end_sequence ();
add_label_notes (m->set_src, temp);
i1 = emit_insns_before (temp, loop_start);
if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
REG_NOTES (i1)
= gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
: REG_EQUAL),
m->set_src, REG_NOTES (i1));
}
else
i1 = emit_insn_before (PATTERN (p), loop_start);
if (REG_NOTES (i1) == 0)
{
REG_NOTES (i1) = REG_NOTES (p);
/* If there is a REG_EQUAL note present whose value
is not loop invariant, then delete it, since it
may cause problems with later optimization passes.
It is possible for cse to create such notes
like this as a result of record_jump_cond. */
if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
&& ! loop_invariant_p (loop, XEXP (temp, 0)))
remove_note (i1, temp);
}
if (new_start == 0)
new_start = i1;
if (loop_dump_stream)
fprintf (loop_dump_stream, " moved to %d",
INSN_UID (i1));
/* If library call, now fix the REG_NOTES that contain
insn pointers, namely REG_LIBCALL on FIRST
and REG_RETVAL on I1. */
if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
{
XEXP (temp, 0) = first;
temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
XEXP (temp, 0) = i1;
}
temp = p;
delete_insn (p);
p = NEXT_INSN (p);
/* simplify_giv_expr expects that it can walk the insns
at m->insn forwards and see this old sequence we are
tossing here. delete_insn does preserve the next
pointers, but when we skip over a NOTE we must fix
it up. Otherwise that code walks into the non-deleted
insn stream. */
while (p && GET_CODE (p) == NOTE)
p = NEXT_INSN (temp) = NEXT_INSN (p);
}
/* The more regs we move, the less we like moving them. */
threshold -= 3;
}
/* Any other movable that loads the same register
MUST be moved. */
already_moved[regno] = 1;
/* This reg has been moved out of one loop. */
moved_once[regno] = 1;
/* The reg set here is now invariant. */
if (! m->partial)
VARRAY_INT (set_in_loop, regno) = 0;
m->done = 1;
/* Change the length-of-life info for the register
to say it lives at least the full length of this loop.
This will help guide optimizations in outer loops. */
if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
/* This is the old insn before all the moved insns.
We can't use the moved insn because it is out of range
in uid_luid. Only the old insns have luids. */
REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (loop_end))
REGNO_LAST_UID (regno) = INSN_UID (loop_end);
/* Combine with this moved insn any other matching movables. */
if (! m->partial)
for (m1 = movables; m1; m1 = m1->next)
if (m1->match == m)
{
rtx temp;
/* Schedule the reg loaded by M1
for replacement so that shares the reg of M.
If the modes differ (only possible in restricted
circumstances, make a SUBREG.
Note this assumes that the target dependent files
treat REG and SUBREG equally, including within
GO_IF_LEGITIMATE_ADDRESS and in all the
predicates since we never verify that replacing the
original register with a SUBREG results in a
recognizable insn. */
if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
reg_map[m1->regno] = m->set_dest;
else
reg_map[m1->regno]
= gen_lowpart_common (GET_MODE (m1->set_dest),
m->set_dest);
/* Get rid of the matching insn
and prevent further processing of it. */
m1->done = 1;
/* if library call, delete all insn except last, which
is deleted below */
if ((temp = find_reg_note (m1->insn, REG_RETVAL,
NULL_RTX)))
{
for (temp = XEXP (temp, 0); temp != m1->insn;
temp = NEXT_INSN (temp))
delete_insn (temp);
}
delete_insn (m1->insn);
/* Any other movable that loads the same register
MUST be moved. */
already_moved[m1->regno] = 1;
/* The reg merged here is now invariant,
if the reg it matches is invariant. */
if (! m->partial)
VARRAY_INT (set_in_loop, m1->regno) = 0;
}
}
else if (loop_dump_stream)
fprintf (loop_dump_stream, "not desirable");
}
else if (loop_dump_stream && !m->match)
fprintf (loop_dump_stream, "not safe");
if (loop_dump_stream)
fprintf (loop_dump_stream, "\n");
}
if (new_start == 0)
new_start = loop_start;
/* Go through all the instructions in the loop, making
all the register substitutions scheduled in REG_MAP. */
for (p = new_start; p != loop_end; p = NEXT_INSN (p))
if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
|| GET_CODE (p) == CALL_INSN)
{
replace_regs (PATTERN (p), reg_map, nregs, 0);
replace_regs (REG_NOTES (p), reg_map, nregs, 0);
INSN_CODE (p) = -1;
}
/* Clean up. */
free (reg_map);
free (already_moved);
}
#if 0
/* Scan X and replace the address of any MEM in it with ADDR.
REG is the address that MEM should have before the replacement. */
static void
replace_call_address (x, reg, addr)
rtx x, reg, addr;
{
register enum rtx_code code;
register int i;
register const char *fmt;
if (x == 0)
return;
code = GET_CODE (x);
switch (code)
{
case PC:
case CC0:
case CONST_INT:
case CONST_DOUBLE:
case CONST:
case SYMBOL_REF:
case LABEL_REF:
case REG:
return;
case SET:
/* Short cut for very common case. */
replace_call_address (XEXP (x, 1), reg, addr);
return;
case CALL:
/* Short cut for very common case. */
replace_call_address (XEXP (x, 0), reg, addr);
return;
case MEM:
/* If this MEM uses a reg other than the one we expected,
something is wrong. */
if (XEXP (x, 0) != reg)
abort ();
XEXP (x, 0) = addr;
return;
default:
break;
}
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
replace_call_address (XEXP (x, i), reg, addr);
else if (fmt[i] == 'E')
{
register int j;
for (j = 0; j < XVECLEN (x, i); j++)
replace_call_address (XVECEXP (x, i, j), reg, addr);
}
}
}
#endif
/* Return the number of memory refs to addresses that vary
in the rtx X. */
static int
count_nonfixed_reads (loop, x)
const struct loop *loop;
rtx x;
{
register enum rtx_code code;
register int i;
register const char *fmt;
int value;
if (x == 0)
return 0;
code = GET_CODE (x);
switch (code)
{
case PC:
case CC0:
case CONST_INT:
case CONST_DOUBLE:
case CONST:
case SYMBOL_REF:
case LABEL_REF:
case REG:
return 0;
case MEM:
return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
+ count_nonfixed_reads (loop, XEXP (x, 0)));
default:
break;
}
value = 0;
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
value += count_nonfixed_reads (loop, XEXP (x, i));
if (fmt[i] == 'E')
{
register int j;
for (j = 0; j < XVECLEN (x, i); j++)
value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
}
}
return value;
}
#if 0
/* P is an instruction that sets a register to the result of a ZERO_EXTEND.
Replace it with an instruction to load just the low bytes
if the machine supports such an instruction,
and insert above LOOP_START an instruction to clear the register. */
static void
constant_high_bytes (p, loop_start)
rtx p, loop_start;
{
register rtx new;
register int insn_code_number;
/* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
new
= gen_rtx_SET
(VOIDmode,
gen_rtx_STRICT_LOW_PART
(VOIDmode,
gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
SET_DEST (PATTERN (p)), 0)),
XEXP (SET_SRC (PATTERN (p)), 0));
insn_code_number = recog (new, p);
if (insn_code_number)
{
register int i;
/* Clear destination register before the loop. */
emit_insn_before (gen_rtx_SET (VOIDmode,
SET_DEST (PATTERN (p)), const0_rtx),
loop_start);
/* Inside the loop, just load the low part. */
PATTERN (p) = new;
}
}
#endif
/* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
`has_call', `has_volatile', and `has_tablejump' within LOOP.
Set the global variables `unknown_address_altered',
`unknown_constant_address_altered', and `num_mem_sets'. Also, fill
in the array `loop_mems' and the list `loop_store_mems'. */
static void
prescan_loop (loop)
struct loop *loop;
{
register int level = 1;
rtx insn;
struct loop_info *loop_info = LOOP_INFO (loop);
rtx start = loop->start;
rtx end = loop->end;
/* The label after END. Jumping here is just like falling off the
end of the loop. We use next_nonnote_insn instead of next_label
as a hedge against the (pathological) case where some actual insn
might end up between the two. */
rtx exit_target = next_nonnote_insn (end);
loop_info->has_indirect_jump = indirect_jump_in_function;
loop_info->has_call = 0;
loop_info->has_volatile = 0;
loop_info->has_tablejump = 0;
loop_info->has_multiple_exit_targets = 0;
loop->cont = 0;
loop->vtop = 0;
loop->level = 1;
unknown_address_altered = 0;
unknown_constant_address_altered = 0;
loop_store_mems = NULL_RTX;
first_loop_store_insn = NULL_RTX;
loop_mems_idx = 0;
num_mem_sets = 0;
for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
insn = NEXT_INSN (insn))
{
if (GET_CODE (insn) == NOTE)
{
if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
{
++level;
/* Count number of loops contained in this one. */
loop->level++;
}
else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
{
--level;
if (level == 0)
{
end = insn;
break;
}
}
else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
{
if (level == 1)
loop->cont = insn;
}
else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_VTOP)
{
/* If there is a NOTE_INSN_LOOP_VTOP, then this is a for
or while style loop, with a loop exit test at the
start. Thus, we can assume that the loop condition
was true when the loop was entered. */
if (level == 1)
loop->vtop = insn;
}
}
else if (GET_CODE (insn) == CALL_INSN)
{
if (! CONST_CALL_P (insn))
unknown_address_altered = 1;
loop_info->has_call = 1;
}
else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
{
rtx label1 = NULL_RTX;
rtx label2 = NULL_RTX;
if (volatile_refs_p (PATTERN (insn)))
loop_info->has_volatile = 1;
if (GET_CODE (insn) == JUMP_INSN
&& (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
|| GET_CODE (PATTERN (insn)) == ADDR_VEC))
loop_info->has_tablejump = 1;
note_stores (PATTERN (insn), note_addr_stored, NULL);
if (! first_loop_store_insn && loop_store_mems)
first_loop_store_insn = insn;
if (! loop_info->has_multiple_exit_targets
&& GET_CODE (insn) == JUMP_INSN
&& GET_CODE (PATTERN (insn)) == SET
&& SET_DEST (PATTERN (insn)) == pc_rtx)
{
if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
{
label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
}
else
{
label1 = SET_SRC (PATTERN (insn));
}
do {
if (label1 && label1 != pc_rtx)
{
if (GET_CODE (label1) != LABEL_REF)
{
/* Something tricky. */
loop_info->has_multiple_exit_targets = 1;
break;
}
else if (XEXP (label1, 0) != exit_target
&& LABEL_OUTSIDE_LOOP_P (label1))
{
/* A jump outside the current loop. */
loop_info->has_multiple_exit_targets = 1;
break;
}
}
label1 = label2;
label2 = NULL_RTX;
} while (label1);
}
}
else if (GET_CODE (insn) == RETURN)
loop_info->has_multiple_exit_targets = 1;
}
/* Now, rescan the loop, setting up the LOOP_MEMS array. */
if (/* We can't tell what MEMs are aliased by what. */
! unknown_address_altered
/* An exception thrown by a called function might land us
anywhere. */
&& ! loop_info->has_call
/* We don't want loads for MEMs moved to a location before the
one at which their stack memory becomes allocated. (Note
that this is not a problem for malloc, etc., since those
require actual function calls. */
&& ! current_function_calls_alloca
/* There are ways to leave the loop other than falling off the
end. */
&& ! loop_info->has_multiple_exit_targets)
for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
insn = NEXT_INSN (insn))
for_each_rtx (&insn, insert_loop_mem, 0);
}
/* LOOP->CONT_DOMINATOR is now the last label between the loop start
and the continue note that is a the destination of a (cond)jump after
the continue note. If there is any (cond)jump between the loop start
and what we have so far as LOOP->CONT_DOMINATOR that has a
target between LOOP->DOMINATOR and the continue note, move
LOOP->CONT_DOMINATOR forward to that label; if a jump's
destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
static void
verify_dominator (loop)
struct loop *loop;
{
rtx insn;
if (! loop->cont_dominator)
/* This can happen for an empty loop, e.g. in
gcc.c-torture/compile/920410-2.c */
return;
if (loop->cont_dominator == const0_rtx)
{
loop->cont_dominator = 0;
return;
}
for (insn = loop->start; insn != loop->cont_dominator;
insn = NEXT_INSN (insn))
{
if (GET_CODE (insn) == JUMP_INSN
&& GET_CODE (PATTERN (insn)) != RETURN)
{
rtx label = JUMP_LABEL (insn);
int label_luid;
/* If it is not a jump we can easily understand or for
which we do not have jump target information in the JUMP_LABEL
field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
LOOP->CONT_DOMINATOR. */
if ((! condjump_p (insn)
&& ! condjump_in_parallel_p (insn))
|| label == NULL_RTX)
{
loop->cont_dominator = NULL_RTX;
return;
}
label_luid = INSN_LUID (label);
if (label_luid < INSN_LUID (loop->cont)
&& (label_luid
> INSN_LUID (loop->cont)))
loop->cont_dominator = label;
}
}
}
/* Scan the function looking for loops. Record the start and end of each loop.
Also mark as invalid loops any loops that contain a setjmp or are branched
to from outside the loop. */
static void
find_and_verify_loops (f, loops)
rtx f;
struct loops *loops;
{
rtx insn;
rtx label;
int num_loops;
struct loop *current_loop;
struct loop *next_loop;
struct loop *loop;
num_loops = loops->num;
compute_luids (f, NULL_RTX, 0);
/* If there are jumps to undefined labels,
treat them as jumps out of any/all loops.
This also avoids writing past end of tables when there are no loops. */
uid_loop[0] = NULL;
/* Find boundaries of loops, mark which loops are contained within
loops, and invalidate loops that have setjmp. */
num_loops = 0;
current_loop = NULL;
for (insn = f; insn; insn = NEXT_INSN (insn))
{
if (GET_CODE (insn) == NOTE)
switch (NOTE_LINE_NUMBER (insn))
{
case NOTE_INSN_LOOP_BEG:
next_loop = loops->array + num_loops;
next_loop->num = num_loops;
num_loops++;
next_loop->start = insn;
next_loop->outer = current_loop;
current_loop = next_loop;
break;
case NOTE_INSN_SETJMP:
/* In this case, we must invalidate our current loop and any
enclosing loop. */
for (loop = current_loop; loop; loop = loop->outer)
{
loop->invalid = 1;
if (loop_dump_stream)
fprintf (loop_dump_stream,
"\nLoop at %d ignored due to setjmp.\n",
INSN_UID (loop->start));
}
break;
case NOTE_INSN_LOOP_CONT:
current_loop->cont = insn;
break;
case NOTE_INSN_LOOP_END:
if (! current_loop)
abort ();
current_loop->end = insn;
verify_dominator (current_loop);
current_loop = current_loop->outer;
break;
default:
break;
}
/* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
and NOTE_INSN_LOOP_END notes, update loop->dominator. */
else if (GET_CODE (insn) == JUMP_INSN
&& GET_CODE (PATTERN (insn)) != RETURN
&& current_loop)
{
rtx label = JUMP_LABEL (insn);
if (! condjump_p (insn) && ! condjump_in_parallel_p (insn))
label = NULL_RTX;
loop = current_loop;
do
{
/* First see if we care about this loop. */
if (loop->cont && loop->cont_dominator != const0_rtx)
{
/* If the jump destination is not known, invalidate
loop->const_dominator. */
if (! label)
loop->cont_dominator = const0_rtx;
else
/* Check if the destination is between loop start and
cont. */
if ((INSN_LUID (label)
< INSN_LUID (loop->cont))
&& (INSN_LUID (label)
> INSN_LUID (loop->start))
/* And if there is no later destination already
recorded. */
&& (! loop->cont_dominator
|| (INSN_LUID (label)
> INSN_LUID (loop->cont_dominator))))
loop->cont_dominator = label;
}
loop = loop->outer;
}
while (loop);
}
/* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
enclosing loop, but this doesn't matter. */
uid_loop[INSN_UID (insn)] = current_loop;
}
/* Any loop containing a label used in an initializer must be invalidated,
because it can be jumped into from anywhere. */
for (label = forced_labels; label; label = XEXP (label, 1))
{
for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
loop; loop = loop->outer)
loop->invalid = 1;
}
/* Any loop containing a label used for an exception handler must be
invalidated, because it can be jumped into from anywhere. */
for (label = exception_handler_labels; label; label = XEXP (label, 1))
{
for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
loop; loop = loop->outer)
loop->invalid = 1;
}
/* Now scan all insn's in the function. If any JUMP_INSN branches into a
loop that it is not contained within, that loop is marked invalid.
If any INSN or CALL_INSN uses a label's address, then the loop containing
that label is marked invalid, because it could be jumped into from
anywhere.
Also look for blocks of code ending in an unconditional branch that
exits the loop. If such a block is surrounded by a conditional
branch around the block, move the block elsewhere (see below) and
invert the jump to point to the code block. This may eliminate a
label in our loop and will simplify processing by both us and a
possible second cse pass. */
for (insn = f; insn; insn = NEXT_INSN (insn))
if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
{
struct loop *this_loop = uid_loop[INSN_UID (insn)];
if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
{
rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
if (note)
{
for (loop = uid_loop[INSN_UID (XEXP (note, 0))];
loop; loop = loop->outer)
loop->invalid = 1;
}
}
if (GET_CODE (insn) != JUMP_INSN)
continue;
mark_loop_jump (PATTERN (insn), this_loop);
/* See if this is an unconditional branch outside the loop. */
if (this_loop
&& (GET_CODE (PATTERN (insn)) == RETURN
|| (simplejump_p (insn)
&& (uid_loop[INSN_UID (JUMP_LABEL (insn))]
!= this_loop)))
&& get_max_uid () < max_uid_for_loop)
{
rtx p;
rtx our_next = next_real_insn (insn);
rtx last_insn_to_move = NEXT_INSN (insn);
struct loop *dest_loop;
struct loop *outer_loop = NULL;
/* Go backwards until we reach the start of the loop, a label,
or a JUMP_INSN. */
for (p = PREV_INSN (insn);
GET_CODE (p) != CODE_LABEL
&& ! (GET_CODE (p) == NOTE
&& NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
&& GET_CODE (p) != JUMP_INSN;
p = PREV_INSN (p))
;
/* Check for the case where we have a jump to an inner nested
loop, and do not perform the optimization in that case. */
if (JUMP_LABEL (insn))
{
dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
if (dest_loop)
{
for (outer_loop = dest_loop; outer_loop;
outer_loop = outer_loop->outer)
if (outer_loop == this_loop)
break;
}
}
/* Make sure that the target of P is within the current loop. */
if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
&& uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
outer_loop = this_loop;
/* If we stopped on a JUMP_INSN to the next insn after INSN,
we have a block of code to try to move.
We look backward and then forward from the target of INSN
to find a BARRIER at the same loop depth as the target.
If we find such a BARRIER, we make a new label for the start
of the block, invert the jump in P and point it to that label,
and move the block of code to the spot we found. */
if (! outer_loop
&& GET_CODE (p) == JUMP_INSN
&& JUMP_LABEL (p) != 0
/* Just ignore jumps to labels that were never emitted.
These always indicate compilation errors. */
&& INSN_UID (JUMP_LABEL (p)) != 0
&& condjump_p (p)
&& ! simplejump_p (p)
&& next_real_insn (JUMP_LABEL (p)) == our_next
/* If it's not safe to move the sequence, then we
mustn't try. */
&& insns_safe_to_move_p (p, NEXT_INSN (insn),
&last_insn_to_move))
{
rtx target
= JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
struct loop *target_loop = uid_loop[INSN_UID (target)];
rtx loc, loc2;
for (loc = target; loc; loc = PREV_INSN (loc))
if (GET_CODE (loc) == BARRIER
/* Don't move things inside a tablejump. */
&& ((loc2 = next_nonnote_insn (loc)) == 0
|| GET_CODE (loc2) != CODE_LABEL
|| (loc2 = next_nonnote_insn (loc2)) == 0
|| GET_CODE (loc2) != JUMP_INSN
|| (GET_CODE (PATTERN (loc2)) != ADDR_VEC
&& GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
&& uid_loop[INSN_UID (loc)] == target_loop)
break;
if (loc == 0)
for (loc = target; loc; loc = NEXT_INSN (loc))
if (GET_CODE (loc) == BARRIER
/* Don't move things inside a tablejump. */
&& ((loc2 = next_nonnote_insn (loc)) == 0
|| GET_CODE (loc2) != CODE_LABEL
|| (loc2 = next_nonnote_insn (loc2)) == 0
|| GET_CODE (loc2) != JUMP_INSN
|| (GET_CODE (PATTERN (loc2)) != ADDR_VEC
&& GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
&& uid_loop[INSN_UID (loc)] == target_loop)
break;
if (loc)
{
rtx cond_label = JUMP_LABEL (p);
rtx new_label = get_label_after (p);
/* Ensure our label doesn't go away. */
LABEL_NUSES (cond_label)++;
/* Verify that uid_loop is large enough and that
we can invert P. */
if (invert_jump (p, new_label))
{
rtx q, r;
/* If no suitable BARRIER was found, create a suitable
one before TARGET. Since TARGET is a fall through
path, we'll need to insert an jump around our block
and a add a BARRIER before TARGET.
This creates an extra unconditional jump outside
the loop. However, the benefits of removing rarely
executed instructions from inside the loop usually
outweighs the cost of the extra unconditional jump
outside the loop. */
if (loc == 0)
{
rtx temp;
temp = gen_jump (JUMP_LABEL (insn));
temp = emit_jump_insn_before (temp, target);
JUMP_LABEL (temp) = JUMP_LABEL (insn);
LABEL_NUSES (JUMP_LABEL (insn))++;
loc = emit_barrier_before (target);
}
/* Include the BARRIER after INSN and copy the
block after LOC. */
new_label = squeeze_notes (new_label,
last_insn_to_move);
reorder_insns (new_label, last_insn_to_move, loc);
/* All those insns are now in TARGET_LOOP. */
for (q = new_label;
q != NEXT_INSN (last_insn_to_move);
q = NEXT_INSN (q))
uid_loop[INSN_UID (q)] = target_loop;
/* The label jumped to by INSN is no longer a loop exit.
Unless INSN does not have a label (e.g., it is a
RETURN insn), search loop->exit_labels to find
its label_ref, and remove it. Also turn off
LABEL_OUTSIDE_LOOP_P bit. */
if (JUMP_LABEL (insn))
{
for (q = 0,
r = this_loop->exit_labels;
r; q = r, r = LABEL_NEXTREF (r))
if (XEXP (r, 0) == JUMP_LABEL (insn))
{
LABEL_OUTSIDE_LOOP_P (r) = 0;
if (q)
LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
else
this_loop->exit_labels = LABEL_NEXTREF (r);
break;
}
for (loop = this_loop; loop && loop != target_loop;
loop = loop->outer)
loop->exit_count--;
/* If we didn't find it, then something is
wrong. */
if (! r)
abort ();
}
/* P is now a jump outside the loop, so it must be put
in loop->exit_labels, and marked as such.
The easiest way to do this is to just call
mark_loop_jump again for P. */
mark_loop_jump (PATTERN (p), this_loop);
/* If INSN now jumps to the insn after it,
delete INSN. */
if (JUMP_LABEL (insn) != 0
&& (next_real_insn (JUMP_LABEL (insn))
== next_real_insn (insn)))
delete_insn (insn);
}
/* Continue the loop after where the conditional
branch used to jump, since the only branch insn
in the block (if it still remains) is an inter-loop
branch and hence needs no processing. */
insn = NEXT_INSN (cond_label);
if (--LABEL_NUSES (cond_label) == 0)
delete_insn (cond_label);
/* This loop will be continued with NEXT_INSN (insn). */
insn = PREV_INSN (insn);
}
}
}
}
}
/* If any label in X jumps to a loop different from LOOP_NUM and any of the
loops it is contained in, mark the target loop invalid.
For speed, we assume that X is part of a pattern of a JUMP_INSN. */
static void
mark_loop_jump (x, loop)
rtx x;
struct loop *loop;
{
struct loop *dest_loop;
struct loop *outer_loop;
int i;
switch (GET_CODE (x))
{
case PC:
case USE:
case CLOBBER:
case REG:
case MEM:
case CONST_INT:
case CONST_DOUBLE:
case RETURN:
return;
case CONST:
/* There could be a label reference in here. */
mark_loop_jump (XEXP (x, 0), loop);
return;
case PLUS:
case MINUS:
case MULT:
mark_loop_jump (XEXP (x, 0), loop);
mark_loop_jump (XEXP (x, 1), loop);
return;
case LO_SUM:
/* This may refer to a LABEL_REF or SYMBOL_REF. */
mark_loop_jump (XEXP (x, 1), loop);
return;
case SIGN_EXTEND:
case ZERO_EXTEND:
mark_loop_jump (XEXP (x, 0), loop);
return;
case LABEL_REF:
dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
/* Link together all labels that branch outside the loop. This
is used by final_[bg]iv_value and the loop unrolling code. Also
mark this LABEL_REF so we know that this branch should predict
false. */
/* A check to make sure the label is not in an inner nested loop,
since this does not count as a loop exit. */
if (dest_loop)
{
for (outer_loop = dest_loop; outer_loop;
outer_loop = outer_loop->outer)
if (outer_loop == loop)
break;
}
else
outer_loop = NULL;
if (loop && ! outer_loop)
{
LABEL_OUTSIDE_LOOP_P (x) = 1;
LABEL_NEXTREF (x) = loop->exit_labels;
loop->exit_labels = x;
for (outer_loop = loop;
outer_loop && outer_loop != dest_loop;
outer_loop = outer_loop->outer)
outer_loop->exit_count++;
}
/* If this is inside a loop, but not in the current loop or one enclosed
by it, it invalidates at least one loop. */
if (! dest_loop)
return;
/* We must invalidate every nested loop containing the target of this
label, except those that also contain the jump insn. */
for (; dest_loop; dest_loop = dest_loop->outer)
{
/* Stop when we reach a loop that also contains the jump insn. */
for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
if (dest_loop == outer_loop)
return;
/* If we get here, we know we need to invalidate a loop. */
if (loop_dump_stream && ! dest_loop->invalid)
fprintf (loop_dump_stream,
"\nLoop at %d ignored due to multiple entry points.\n",
INSN_UID (dest_loop->start));
dest_loop->invalid = 1;
}
return;
case SET:
/* If this is not setting pc, ignore. */
if (SET_DEST (x) == pc_rtx)
mark_loop_jump (SET_SRC (x), loop);
return;
case IF_THEN_ELSE:
mark_loop_jump (XEXP (x, 1), loop);
mark_loop_jump (XEXP (x, 2), loop);
return;
case PARALLEL:
case ADDR_VEC:
for (i = 0; i < XVECLEN (x, 0); i++)
mark_loop_jump (XVECEXP (x, 0, i), loop);
return;
case ADDR_DIFF_VEC:
for (i = 0; i < XVECLEN (x, 1); i++)
mark_loop_jump (XVECEXP (x, 1, i), loop);
return;
default:
/* Strictly speaking this is not a jump into the loop, only a possible
jump out of the loop. However, we have no way to link the destination
of this jump onto the list of exit labels. To be safe we mark this
loop and any containing loops as invalid. */
if (loop)
{
for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
{
if (loop_dump_stream && ! outer_loop->invalid)
fprintf (loop_dump_stream,
"\nLoop at %d ignored due to unknown exit jump.\n",
INSN_UID (outer_loop->start));
outer_loop->invalid = 1;
}
}
return;
}
}
/* Return nonzero if there is a label in the range from
insn INSN to and including the insn whose luid is END
INSN must have an assigned luid (i.e., it must not have
been previously created by loop.c). */
static int
labels_in_range_p (insn, end)
rtx insn;
int end;
{
while (insn && INSN_LUID (insn) <= end)
{
if (GET_CODE (insn) == CODE_LABEL)
return 1;
insn = NEXT_INSN (insn);
}
return 0;
}
/* Record that a memory reference X is being set. */
static void
note_addr_stored (x, y, data)
rtx x;
rtx y ATTRIBUTE_UNUSED;
void *data ATTRIBUTE_UNUSED;
{
if (x == 0 || GET_CODE (x) != MEM)
return;
/* Count number of memory writes.
This affects heuristics in strength_reduce. */
num_mem_sets++;
/* BLKmode MEM means all memory is clobbered. */
if (GET_MODE (x) == BLKmode)
{
if (RTX_UNCHANGING_P (x))
unknown_constant_address_altered = 1;
else
unknown_address_altered = 1;
return;
}
loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
}
/* X is a value modified by an INSN that references a biv inside a loop
exit test (ie, X is somehow related to the value of the biv). If X
is a pseudo that is used more than once, then the biv is (effectively)
used more than once. DATA is really an `int *', and is set if the
biv is used more than once. */
static void
note_set_pseudo_multiple_uses (x, y, data)
rtx x;
rtx y ATTRIBUTE_UNUSED;
void *data;
{
if (x == 0)
return;
while (GET_CODE (x) == STRICT_LOW_PART
|| GET_CODE (x) == SIGN_EXTRACT
|| GET_CODE (x) == ZERO_EXTRACT
|| GET_CODE (x) == SUBREG)
x = XEXP (x, 0);
if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
return;
/* If we do not have usage information, or if we know the register
is used more than once, note that fact for check_dbra_loop. */
if (REGNO (x) >= max_reg_before_loop
|| ! VARRAY_RTX (reg_single_usage, REGNO (x))
|| VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
*((int *) data) = 1;
}
/* Return nonzero if the rtx X is invariant over the current loop.
The value is 2 if we refer to something only conditionally invariant.
If `unknown_address_altered' is nonzero, no memory ref is invariant.
Otherwise, a memory ref is invariant if it does not conflict with
anything stored in `loop_store_mems'. */
int
loop_invariant_p (loop, x)
const struct loop *loop;
register rtx x;
{
register int i;
register enum rtx_code code;
register const char *fmt;
int conditional = 0;
rtx mem_list_entry;
if (x == 0)
return 1;
code = GET_CODE (x);
switch (code)
{
case CONST_INT:
case CONST_DOUBLE:
case SYMBOL_REF:
case CONST:
return 1;
case LABEL_REF:
/* A LABEL_REF is normally invariant, however, if we are unrolling
loops, and this label is inside the loop, then it isn't invariant.
This is because each unrolled copy of the loop body will have
a copy of this label. If this was invariant, then an insn loading
the address of this label into a register might get moved outside
the loop, and then each loop body would end up using the same label.
We don't know the loop bounds here though, so just fail for all
labels. */
if (flag_unroll_loops)
return 0;
else
return 1;
case PC:
case CC0:
case UNSPEC_VOLATILE:
return 0;
case REG:
/* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
since the reg might be set by initialization within the loop. */
if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
|| x == arg_pointer_rtx)
&& ! current_function_has_nonlocal_goto)
return 1;
if (LOOP_INFO (loop)->has_call
&& REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
return 0;
if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
return 2;
return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
case MEM:
/* Volatile memory references must be rejected. Do this before
checking for read-only items, so that volatile read-only items
will be rejected also. */
if (MEM_VOLATILE_P (x))
return 0;
/* If we had a subroutine call, any location in memory could
have been clobbered. We used to test here for volatile and
readonly, but true_dependence knows how to do that better
than we do. */
if (RTX_UNCHANGING_P (x)
? unknown_constant_address_altered : unknown_address_altered)
return 0;
/* See if there is any dependence between a store and this load. */
mem_list_entry = loop_store_mems;
while (mem_list_entry)
{
if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
x, rtx_varies_p))
return 0;
mem_list_entry = XEXP (mem_list_entry, 1);
}
/* It's not invalidated by a store in memory
but we must still verify the address is invariant. */
break;
case ASM_OPERANDS:
/* Don't mess with insns declared volatile. */
if (MEM_VOLATILE_P (x))
return 0;
break;
default:
break;
}
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
{
int tem = loop_invariant_p (loop, XEXP (x, i));
if (tem == 0)
return 0;
if (tem == 2)
conditional = 1;
}
else if (fmt[i] == 'E')
{
register int j;
for (j = 0; j < XVECLEN (x, i); j++)
{
int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
if (tem == 0)
return 0;
if (tem == 2)
conditional = 1;
}
}
}
return 1 + conditional;
}
/* Return nonzero if all the insns in the loop that set REG
are INSN and the immediately following insns,
and if each of those insns sets REG in an invariant way
(not counting uses of REG in them).
The value is 2 if some of these insns are only conditionally invariant.
We assume that INSN itself is the first set of REG
and that its source is invariant. */
static int
consec_sets_invariant_p (loop, reg, n_sets, insn)
const struct loop *loop;
int n_sets;
rtx reg, insn;
{
register rtx p = insn;
register int regno = REGNO (reg);
rtx temp;
/* Number of sets we have to insist on finding after INSN. */
int count = n_sets - 1;
int old = VARRAY_INT (set_in_loop, regno);
int value = 0;
int this;
/* If N_SETS hit the limit, we can't rely on its value. */
if (n_sets == 127)
return 0;
VARRAY_INT (set_in_loop, regno) = 0;
while (count > 0)
{
register enum rtx_code code;
rtx set;
p = NEXT_INSN (p);
code = GET_CODE (p);
/* If library call, skip to end of it. */
if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
p = XEXP (temp, 0);
this = 0;
if (code == INSN
&& (set = single_set (p))
&& GET_CODE (SET_DEST (set)) == REG
&& REGNO (SET_DEST (set)) == regno)
{
this = loop_invariant_p (loop, SET_SRC (set));
if (this != 0)
value |= this;
else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
{
/* If this is a libcall, then any invariant REG_EQUAL note is OK.
If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
notes are OK. */
this = (CONSTANT_P (XEXP (temp, 0))
|| (find_reg_note (p, REG_RETVAL, NULL_RTX)
&& loop_invariant_p (loop, XEXP (temp, 0))));
if (this != 0)
value |= this;
}
}
if (this != 0)
count--;
else if (code != NOTE)
{
VARRAY_INT (set_in_loop, regno) = old;
return 0;
}
}
VARRAY_INT (set_in_loop, regno) = old;
/* If loop_invariant_p ever returned 2, we return 2. */
return 1 + (value & 2);
}
#if 0
/* I don't think this condition is sufficient to allow INSN
to be moved, so we no longer test it. */
/* Return 1 if all insns in the basic block of INSN and following INSN
that set REG are invariant according to TABLE. */
static int
all_sets_invariant_p (reg, insn, table)
rtx reg, insn;
short *table;
{
register rtx p = insn;
register int regno = REGNO (reg);
while (1)
{
register enum rtx_code code;
p = NEXT_INSN (p);
code = GET_CODE (p);
if (code == CODE_LABEL || code == JUMP_INSN)
return 1;
if (code == INSN && GET_CODE (PATTERN (p)) == SET
&& GET_CODE (SET_DEST (PATTERN (p))) == REG
&& REGNO (SET_DEST (PATTERN (p))) == regno)
{
if (! loop_invariant_p (loop, SET_SRC (PATTERN (p)), table))
return 0;
}
}
}
#endif /* 0 */
/* Look at all uses (not sets) of registers in X. For each, if it is
the single use, set USAGE[REGNO] to INSN; if there was a previous use in
a different insn, set USAGE[REGNO] to const0_rtx. */
static void
find_single_use_in_loop (insn, x, usage)
rtx insn;
rtx x;
varray_type usage;
{
enum rtx_code code = GET_CODE (x);
const char *fmt = GET_RTX_FORMAT (code);
int i, j;
if (code == REG)
VARRAY_RTX (usage, REGNO (x))
= (VARRAY_RTX (usage, REGNO (x)) != 0
&& VARRAY_RTX (usage, REGNO (x)) != insn)
? const0_rtx : insn;
else if (code == SET)
{
/* Don't count SET_DEST if it is a REG; otherwise count things
in SET_DEST because if a register is partially modified, it won't
show up as a potential movable so we don't care how USAGE is set
for it. */
if (GET_CODE (SET_DEST (x)) != REG)
find_single_use_in_loop (insn, SET_DEST (x), usage);
find_single_use_in_loop (insn, SET_SRC (x), usage);
}
else
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e' && XEXP (x, i) != 0)
find_single_use_in_loop (insn, XEXP (x, i), usage);
else if (fmt[i] == 'E')
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
}
}
/* Count and record any set in X which is contained in INSN. Update
MAY_NOT_MOVE and LAST_SET for any register set in X. */
static void
count_one_set (insn, x, may_not_move, last_set)
rtx insn, x;
varray_type may_not_move;
rtx *last_set;
{
if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
/* Don't move a reg that has an explicit clobber.
It's not worth the pain to try to do it correctly. */
VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
{
rtx dest = SET_DEST (x);
while (GET_CODE (dest) == SUBREG
|| GET_CODE (dest) == ZERO_EXTRACT
|| GET_CODE (dest) == SIGN_EXTRACT
|| GET_CODE (dest) == STRICT_LOW_PART)
dest = XEXP (dest, 0);
if (GET_CODE (dest) == REG)
{
register int regno = REGNO (dest);
/* If this is the first setting of this reg
in current basic block, and it was set before,
it must be set in two basic blocks, so it cannot
be moved out of the loop. */
if (VARRAY_INT (set_in_loop, regno) > 0
&& last_set[regno] == 0)
VARRAY_CHAR (may_not_move, regno) = 1;
/* If this is not first setting in current basic block,
see if reg was used in between previous one and this.
If so, neither one can be moved. */
if (last_set[regno] != 0
&& reg_used_between_p (dest, last_set[regno], insn))
VARRAY_CHAR (may_not_move, regno) = 1;
if (VARRAY_INT (set_in_loop, regno) < 127)
++VARRAY_INT (set_in_loop, regno);
last_set[regno] = insn;
}
}
}
/* Increment SET_IN_LOOP at the index of each register
that is modified by an insn between FROM and TO.
If the value of an element of SET_IN_LOOP becomes 127 or more,
stop incrementing it, to avoid overflow.
Store in SINGLE_USAGE[I] the single insn in which register I is
used, if it is only used once. Otherwise, it is set to 0 (for no
uses) or const0_rtx for more than one use. This parameter may be zero,
in which case this processing is not done.
Store in *COUNT_PTR the number of actual instruction
in the loop. We use this to decide what is worth moving out. */
/* last_set[n] is nonzero iff reg n has been set in the current basic block.
In that case, it is the insn that last set reg n. */
static void
count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
register rtx from, to;
varray_type may_not_move;
varray_type single_usage;
int *count_ptr;
int nregs;
{
register rtx *last_set = (rtx *) xcalloc (nregs, sizeof (rtx));
register rtx insn;
register int count = 0;
for (insn = from; insn != to; insn = NEXT_INSN (insn))
{
if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
{
++count;
/* Record registers that have exactly one use. */
find_single_use_in_loop (insn, PATTERN (insn), single_usage);
/* Include uses in REG_EQUAL notes. */
if (REG_NOTES (insn))
find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
if (GET_CODE (PATTERN (insn)) == SET
|| GET_CODE (PATTERN (insn)) == CLOBBER)
count_one_set (insn, PATTERN (insn), may_not_move, last_set);
else if (GET_CODE (PATTERN (insn)) == PARALLEL)
{
register int i;
for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
may_not_move, last_set);
}
}
if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
bzero ((char *) last_set, nregs * sizeof (rtx));
}
*count_ptr = count;
/* Clean up. */
free (last_set);
}
/* Given a loop that is bounded by LOOP->START and LOOP->END and that
is entered at LOOP->SCAN_START, return 1 if the register set in SET
contained in insn INSN is used by any insn that precedes INSN in
cyclic order starting from the loop entry point.
We don't want to use INSN_LUID here because if we restrict INSN to those
that have a valid INSN_LUID, it means we cannot move an invariant out
from an inner loop past two loops. */
static int
loop_reg_used_before_p (loop, set, insn)
const struct loop *loop;
rtx set, insn;
{
rtx reg = SET_DEST (set);
rtx p;
/* Scan forward checking for register usage. If we hit INSN, we
are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
{
if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
&& reg_overlap_mentioned_p (reg, PATTERN (p)))
return 1;
if (p == loop->end)
p = loop->start;
}
return 0;
}
/* A "basic induction variable" or biv is a pseudo reg that is set
(within this loop) only by incrementing or decrementing it. */
/* A "general induction variable" or giv is a pseudo reg whose
value is a linear function of a biv. */
/* Bivs are recognized by `basic_induction_var';
Givs by `general_induction_var'. */
/* Indexed by register number, indicates whether or not register is an
induction variable, and if so what type. */
varray_type reg_iv_type;
/* Indexed by register number, contains pointer to `struct induction'
if register is an induction variable. This holds general info for
all induction variables. */
varray_type reg_iv_info;
/* Indexed by register number, contains pointer to `struct iv_class'
if register is a basic induction variable. This holds info describing
the class (a related group) of induction variables that the biv belongs
to. */
struct iv_class **reg_biv_class;
/* The head of a list which links together (via the next field)
every iv class for the current loop. */
struct iv_class *loop_iv_list;
/* Givs made from biv increments are always splittable for loop unrolling.
Since there is no regscan info for them, we have to keep track of them
separately. */
int first_increment_giv, last_increment_giv;
/* Communication with routines called via `note_stores'. */
static rtx note_insn;
/* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
static rtx addr_placeholder;
/* ??? Unfinished optimizations, and possible future optimizations,
for the strength reduction code. */
/* ??? The interaction of biv elimination, and recognition of 'constant'
bivs, may cause problems. */
/* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
performance problems.
Perhaps don't eliminate things that can be combined with an addressing
mode. Find all givs that have the same biv, mult_val, and add_val;
then for each giv, check to see if its only use dies in a following
memory address. If so, generate a new memory address and check to see
if it is valid. If it is valid, then store the modified memory address,
otherwise, mark the giv as not done so that it will get its own iv. */
/* ??? Could try to optimize branches when it is known that a biv is always
positive. */
/* ??? When replace a biv in a compare insn, we should replace with closest
giv so that an optimized branch can still be recognized by the combiner,
e.g. the VAX acb insn. */
/* ??? Many of the checks involving uid_luid could be simplified if regscan
was rerun in loop_optimize whenever a register was added or moved.
Also, some of the optimizations could be a little less conservative. */
/* Perform strength reduction and induction variable elimination.
Pseudo registers created during this function will be beyond the last
valid index in several tables including n_times_set and regno_last_uid.
This does not cause a problem here, because the added registers cannot be
givs outside of their loop, and hence will never be reconsidered.
But scan_loop must check regnos to make sure they are in bounds. */
static void
strength_reduce (loop, insn_count, unroll_p, bct_p)
struct loop *loop;
int insn_count;
int unroll_p, bct_p ATTRIBUTE_UNUSED;
{
rtx p;
rtx set;
rtx inc_val;
rtx mult_val;
rtx dest_reg;
rtx *location;
/* This is 1 if current insn is not executed at least once for every loop
iteration. */
int not_every_iteration = 0;
/* This is 1 if current insn may be executed more than once for every
loop iteration. */
int maybe_multiple = 0;
/* This is 1 if we have past a branch back to the top of the loop
(aka a loop latch). */
int past_loop_latch = 0;
/* Temporary list pointers for traversing loop_iv_list. */
struct iv_class *bl, **backbl;
struct loop_info *loop_info = LOOP_INFO (loop);
/* Ratio of extra register life span we can justify
for saving an instruction. More if loop doesn't call subroutines
since in that case saving an insn makes more difference
and more registers are available. */
/* ??? could set this to last value of threshold in move_movables */
int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
/* Map of pseudo-register replacements. */
rtx *reg_map = NULL;
int reg_map_size;
int call_seen;
rtx test;
rtx end_insert_before;
int loop_depth = 0;
int n_extra_increment;
int unrolled_insn_copies = 0;
rtx loop_start = loop->start;
rtx loop_end = loop->end;
rtx loop_scan_start = loop->scan_start;
rtx loop_top = loop->top;
rtx loop_cont = loop->cont;
/* If loop_scan_start points to the loop exit test, we have to be wary of
subversive use of gotos inside expression statements. */
if (prev_nonnote_insn (loop_scan_start) != prev_nonnote_insn (loop_start))
maybe_multiple = back_branch_in_range_p (loop, loop_scan_start);
VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
reg_biv_class = (struct iv_class **)
xcalloc (max_reg_before_loop, sizeof (struct iv_class *));
loop_iv_list = 0;
addr_placeholder = gen_reg_rtx (Pmode);
/* Save insn immediately after the loop_end. Insns inserted after loop_end
must be put before this insn, so that they will appear in the right
order (i.e. loop order).
If loop_end is the end of the current function, then emit a
NOTE_INSN_DELETED after loop_end and set end_insert_before to the
dummy note insn. */
if (NEXT_INSN (loop_end) != 0)
end_insert_before = NEXT_INSN (loop_end);
else
end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
/* Scan through loop to find all possible bivs. */
for (p = next_insn_in_loop (loop, loop_scan_start);
p != NULL_RTX;
p = next_insn_in_loop (loop, p))
{
if (GET_CODE (p) == INSN
&& (set = single_set (p))
&& GET_CODE (SET_DEST (set)) == REG)
{
dest_reg = SET_DEST (set);
if (REGNO (dest_reg) < max_reg_before_loop
&& REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
&& REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
{
int multi_insn_incr = 0;
if (basic_induction_var (loop, SET_SRC (set),
GET_MODE (SET_SRC (set)),
dest_reg, p, &inc_val, &mult_val,
&location, &multi_insn_incr))
{
/* It is a possible basic induction variable.
Create and initialize an induction structure for it. */
struct induction *v
= (struct induction *) alloca (sizeof (struct induction));
record_biv (v, p, dest_reg, inc_val, mult_val, location,
not_every_iteration, maybe_multiple,
multi_insn_incr);
REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
}
else if (REGNO (dest_reg) < max_reg_before_loop)
REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
}
}
/* Past CODE_LABEL, we get to insns that may be executed multiple
times. The only way we can be sure that they can't is if every
jump insn between here and the end of the loop either
returns, exits the loop, is a jump to a location that is still
behind the label, or is a jump to the loop start. */
if (GET_CODE (p) == CODE_LABEL)
{
rtx insn = p;
maybe_multiple = 0;
while (1)
{
insn = NEXT_INSN (insn);
if (insn == loop_scan_start)
break;
if (insn == loop_end)
{
if (loop_top != 0)
insn = loop_top;
else
break;
if (insn == loop_scan_start)
break;
}
if (GET_CODE (insn) == JUMP_INSN
&& GET_CODE (PATTERN (insn)) != RETURN
&& (! condjump_p (insn)
|| (JUMP_LABEL (insn) != 0
&& JUMP_LABEL (insn) != loop_scan_start
&& ! loop_insn_first_p (p, JUMP_LABEL (insn)))))
{
maybe_multiple = 1;
break;
}
}
}
/* Past a jump, we get to insns for which we can't count
on whether they will be executed during each iteration. */
/* This code appears twice in strength_reduce. There is also similar
code in scan_loop. */
if (GET_CODE (p) == JUMP_INSN
/* If we enter the loop in the middle, and scan around to the
beginning, don't set not_every_iteration for that.
This can be any kind of jump, since we want to know if insns
will be executed if the loop is executed. */
&& ! (JUMP_LABEL (p) == loop_top
&& ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
|| (NEXT_INSN (p) == loop_end && condjump_p (p)))))
{
rtx label = 0;
/* If this is a jump outside the loop, then it also doesn't
matter. Check to see if the target of this branch is on the
loop->exits_labels list. */
for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
if (XEXP (label, 0) == JUMP_LABEL (p))
break;
if (! label)
not_every_iteration = 1;
}
else if (GET_CODE (p) == NOTE)
{
/* At the virtual top of a converted loop, insns are again known to
be executed each iteration: logically, the loop begins here
even though the exit code has been duplicated.
Insns are also again known to be executed each iteration at
the LOOP_CONT note. */
if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
|| NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
&& loop_depth == 0)
not_every_iteration = 0;
else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
loop_depth++;
else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
loop_depth--;
}
/* Note if we pass a loop latch. If we do, then we can not clear
NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
a loop since a jump before the last CODE_LABEL may have started
a new loop iteration.
Note that LOOP_TOP is only set for rotated loops and we need
this check for all loops, so compare against the CODE_LABEL
which immediately follows LOOP_START. */
if (GET_CODE (p) == JUMP_INSN
&& JUMP_LABEL (p) == NEXT_INSN (loop_start))
past_loop_latch = 1;
/* Unlike in the code motion pass where MAYBE_NEVER indicates that
an insn may never be executed, NOT_EVERY_ITERATION indicates whether
or not an insn is known to be executed each iteration of the
loop, whether or not any iterations are known to occur.
Therefore, if we have just passed a label and have no more labels
between here and the test insn of the loop, and we have not passed
a jump to the top of the loop, then we know these insns will be
executed each iteration. */
if (not_every_iteration
&& ! past_loop_latch
&& GET_CODE (p) == CODE_LABEL
&& no_labels_between_p (p, loop_end)
&& loop_insn_first_p (p, loop_cont))
not_every_iteration = 0;
}
/* Scan loop_iv_list to remove all regs that proved not to be bivs.
Make a sanity check against n_times_set. */
for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
{
int fail = 0;
if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
/* Above happens if register modified by subreg, etc. */
/* Make sure it is not recognized as a basic induction var: */
|| VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
/* If never incremented, it is invariant that we decided not to
move. So leave it alone. */
|| ! bl->incremented)
fail = 1;
else if (bl->biv_count > 1)
{
/* ??? If we have multiple increments for this BIV, and any of
them take multiple insns to perform the increment, drop the
BIV, since the bit below that converts the extra increments
into GIVs can't handle the multiple insn increment. */
struct induction *v;
for (v = bl->biv; v ; v = v->next_iv)
if (v->multi_insn_incr)
fail = 1;
}
if (fail)
{
if (loop_dump_stream)
fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
bl->regno,
(REG_IV_TYPE (bl->regno) != BASIC_INDUCT
? "not induction variable"
: (! bl->incremented ? "never incremented"
: "count error")));
REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
*backbl = bl->next;
}
else
{
backbl = &bl->next;
if (loop_dump_stream)
fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
}
}
/* Exit if there are no bivs. */
if (! loop_iv_list)
{
/* Can still unroll the loop anyways, but indicate that there is no
strength reduction info available. */
if (unroll_p)
unroll_loop (loop, insn_count, end_insert_before, 0);
goto egress;
}
/* Find initial value for each biv by searching backwards from loop_start,
halting at first label. Also record any test condition. */
call_seen = 0;
for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
{
note_insn = p;
if (GET_CODE (p) == CALL_INSN)
call_seen = 1;
if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
|| GET_CODE (p) == CALL_INSN)
note_stores (PATTERN (p), record_initial, NULL);
/* Record any test of a biv that branches around the loop if no store
between it and the start of loop. We only care about tests with
constants and registers and only certain of those. */
if (GET_CODE (p) == JUMP_INSN
&& JUMP_LABEL (p) != 0
&& next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
&& (test = get_condition_for_loop (loop, p)) != 0
&& GET_CODE (XEXP (test, 0)) == REG
&& REGNO (XEXP (test, 0)) < max_reg_before_loop
&& (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
&& valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
&& bl->init_insn == 0)
{
/* If an NE test, we have an initial value! */
if (GET_CODE (test) == NE)
{
bl->init_insn = p;
bl->init_set = gen_rtx_SET (VOIDmode,
XEXP (test, 0), XEXP (test, 1));
}
else
bl->initial_test = test;
}
}
/* Look at the each biv and see if we can say anything better about its
initial value from any initializing insns set up above. (This is done
in two passes to avoid missing SETs in a PARALLEL.) */
for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
{
rtx src;
rtx note;
if (! bl->init_insn)
continue;
/* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
is a constant, use the value of that. */
if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
&& CONSTANT_P (XEXP (note, 0)))
|| ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
&& CONSTANT_P (XEXP (note, 0))))
src = XEXP (note, 0);
else
src = SET_SRC (bl->init_set);
if (loop_dump_stream)
fprintf (loop_dump_stream,
"Biv %d initialized at insn %d: initial value ",
bl->regno, INSN_UID (bl->init_insn));
if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
|| GET_MODE (src) == VOIDmode)
&& valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
{
bl->initial_value = src;
if (loop_dump_stream)
{
if (GET_CODE (src) == CONST_INT)
{
fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
fputc ('\n', loop_dump_stream);
}
else
{
print_rtl (loop_dump_stream, src);
fprintf (loop_dump_stream, "\n");
}
}
}
else
{
struct iv_class *bl2 = 0;
rtx increment = NULL_RTX;
/* Biv initial value is not a simple move. If it is the sum of
another biv and a constant, check if both bivs are incremented
in lockstep. Then we are actually looking at a giv.
For simplicity, we only handle the case where there is but a
single increment, and the register is not used elsewhere. */
if (bl->biv_count == 1
&& bl->regno < max_reg_before_loop
&& uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
&& GET_CODE (src) == PLUS
&& GET_CODE (XEXP (src, 0)) == REG
&& CONSTANT_P (XEXP (src, 1))
&& ((increment = biv_total_increment (bl)) != NULL_RTX))
{
int regno = REGNO (XEXP (src, 0));
for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
if (bl2->regno == regno)
break;
}
/* Now, can we transform this biv into a giv? */
if (bl2
&& bl2->biv_count == 1
&& rtx_equal_p (increment, biv_total_increment (bl2))
/* init_insn is only set to insns that are before loop_start
without any intervening labels. */
&& ! reg_set_between_p (bl2->biv->src_reg,
PREV_INSN (bl->init_insn), loop_start)
/* The register from BL2 must be set before the register from
BL is set, or we must be able to move the latter set after
the former set. Currently there can't be any labels
in-between when biv_total_increment returns nonzero both times
but we test it here in case some day some real cfg analysis
gets used to set always_computable. */
&& (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
: (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
bl2->biv->insn)
&& no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
&& validate_change (bl->biv->insn,
&SET_SRC (single_set (bl->biv->insn)),
copy_rtx (src), 0))
{
rtx dominator = loop->cont_dominator;
rtx giv = bl->biv->src_reg;
rtx giv_insn = bl->biv->insn;
rtx after_giv = NEXT_INSN (giv_insn);
if (loop_dump_stream)
fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
/* Let this giv be discovered by the generic code. */
REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
reg_biv_class[bl->regno] = (struct iv_class *) NULL_PTR;
/* We can get better optimization if we can move the giv setting
before the first giv use. */
if (dominator
&& ! loop_insn_first_p (dominator, loop_scan_start)
&& ! reg_set_between_p (bl2->biv->src_reg, loop_start,
dominator)
&& ! reg_used_between_p (giv, loop_start, dominator)
&& ! reg_used_between_p (giv, giv_insn, loop_end))
{
rtx p;
rtx next;
for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
{
if ((GET_RTX_CLASS (GET_CODE (next)) == 'i'
&& (reg_mentioned_p (giv, PATTERN (next))
|| reg_set_p (bl2->biv->src_reg, next)))
|| GET_CODE (next) == JUMP_INSN)
break;
#ifdef HAVE_cc0
if (GET_RTX_CLASS (GET_CODE (next)) != 'i'
|| ! sets_cc0_p (PATTERN (next)))
#endif
dominator = next;
}
if (loop_dump_stream)
fprintf (loop_dump_stream, "move after insn %d\n",
INSN_UID (dominator));
/* Avoid problems with luids by actually moving the insn
and adjusting all luids in the range. */
reorder_insns (giv_insn, giv_insn, dominator);
for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
p = PREV_INSN (p);
compute_luids (giv_insn, after_giv, INSN_LUID (p));
/* If the only purpose of the init insn is to initialize
this giv, delete it. */
if (single_set (bl->init_insn)
&& ! reg_used_between_p (giv, bl->init_insn, loop_start))
delete_insn (bl->init_insn);
}
else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
{
rtx p = PREV_INSN (giv_insn);
while (INSN_UID (p) >= max_uid_for_loop)
p = PREV_INSN (p);
reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
compute_luids (after_giv, NEXT_INSN (giv_insn),
INSN_LUID (p));
}
/* Remove this biv from the chain. */
if (bl->next)
{
/* We move the following giv from *bl->next into *bl.
We have to update reg_biv_class for that moved biv
to point to its new address. */
*bl = *bl->next;
reg_biv_class[bl->regno] = bl;
}
else
{
*backbl = 0;
break;
}
}
/* If we can't make it a giv,
let biv keep initial value of "itself". */
else if (loop_dump_stream)
fprintf (loop_dump_stream, "is complex\n");
}
}
/* If a biv is unconditionally incremented several times in a row, convert
all but the last increment into a giv. */
/* Get an upper bound for the number of registers
we might have after all bivs have been processed. */
first_increment_giv = max_reg_num ();
for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
n_extra_increment += bl->biv_count - 1;
/* If the loop contains volatile memory references do not allow any
replacements to take place, since this could loose the volatile
markers. */
if (n_extra_increment && ! loop_info->has_volatile)
{
int nregs = first_increment_giv + n_extra_increment;
/* Reallocate reg_iv_type and reg_iv_info. */
VARRAY_GROW (reg_iv_type, nregs);
VARRAY_GROW (reg_iv_info, nregs);
for (bl = loop_iv_list; bl; bl = bl->next)
{
struct induction **vp, *v, *next;
int biv_dead_after_loop = 0;
/* The biv increments lists are in reverse order. Fix this
first. */
for (v = bl->biv, bl->biv = 0; v; v = next)
{
next = v->next_iv;
v->next_iv = bl->biv;
bl->biv = v;
}
/* We must guard against the case that an early exit between v->insn
and next->insn leaves the biv live after the loop, since that
would mean that we'd be missing an increment for the final
value. The following test to set biv_dead_after_loop is like
the first part of the test to set bl->eliminable.
We don't check here if we can calculate the final value, since
this can't succeed if we already know that there is a jump
between v->insn and next->insn, yet next->always_executed is
set and next->maybe_multiple is cleared. Such a combination
implies that the jump destination is outside the loop.
If we want to make this check more sophisticated, we should
check each branch between v->insn and next->insn individually
to see if the biv is dead at its destination. */
if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
&& bl->init_insn
&& INSN_UID (bl->init_insn) < max_uid_for_loop
&& (uid_luid[REGNO_FIRST_UID (bl->regno)]
>= INSN_LUID (bl->init_insn))
#ifdef HAVE_decrement_and_branch_until_zero
&& ! bl->nonneg
#endif
&& ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
biv_dead_after_loop = 1;
for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
{
HOST_WIDE_INT offset;
rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
int old_regno, new_regno;
if (! v->always_executed
|| v->maybe_multiple
|| GET_CODE (v->add_val) != CONST_INT
|| ! next->always_executed
|| next->maybe_multiple
|| ! CONSTANT_P (next->add_val)
|| v->mult_val != const1_rtx
|| next->mult_val != const1_rtx
|| ! (biv_dead_after_loop
|| no_jumps_between_p (v->insn, next->insn)))
{
vp = &v->next_iv;
continue;
}
offset = INTVAL (v->add_val);
set = single_set (v->insn);
add_val = plus_constant (next->add_val, offset);
old_reg = v->dest_reg;
dest_reg = gen_reg_rtx (v->mode);
/* Unlike reg_iv_type / reg_iv_info, the other three arrays
have been allocated with some slop space, so we may not
actually need to reallocate them. If we do, the following
if statement will be executed just once in this loop. */
if ((unsigned) max_reg_num () > n_times_set->num_elements)
{
/* Grow all the remaining arrays. */
VARRAY_GROW (set_in_loop, nregs);
VARRAY_GROW (n_times_set, nregs);
VARRAY_GROW (may_not_optimize, nregs);
VARRAY_GROW (reg_single_usage, nregs);
}
if (! validate_change (next->insn, next->location, add_val, 0))
{
vp = &v->next_iv;
continue;
}
/* Here we can try to eliminate the increment by combining
it into the uses. */
/* Set last_use_insn so that we can check against it. */
for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
p != next->insn;
p = next_insn_in_loop (loop, p))
{
if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
continue;
if (reg_mentioned_p (old_reg, PATTERN (p)))
{
last_use_insn = p;
}
}
/* If we can't get the LUIDs for the insns, we can't
calculate the lifetime. This is likely from unrolling
of an inner loop, so there is little point in making this
a DEST_REG giv anyways. */
if (INSN_UID (v->insn) >= max_uid_for_loop
|| INSN_UID (last_use_insn) >= max_uid_for_loop
|| ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
{
/* Change the increment at NEXT back to what it was. */
if (! validate_change (next->insn, next->location,
next->add_val, 0))
abort ();
vp = &v->next_iv;
continue;
}
next->add_val = add_val;
v->dest_reg = dest_reg;
v->giv_type = DEST_REG;
v->location = &SET_SRC (set);
v->cant_derive = 0;
v->combined_with = 0;
v->maybe_dead = 0;
v->derive_adjustment = 0;
v->same = 0;
v->ignore = 0;
v->new_reg = 0;
v->final_value = 0;
v->same_insn = 0;
v->auto_inc_opt = 0;
v->unrolled = 0;
v->shared = 0;
v->derived_from = 0;
v->always_computable = 1;
v->always_executed = 1;
v->replaceable = 1;
v->no_const_addval = 0;
old_regno = REGNO (old_reg);
new_regno = REGNO (dest_reg);
VARRAY_INT (set_in_loop, old_regno)--;
VARRAY_INT (set_in_loop, new_regno) = 1;
VARRAY_INT (n_times_set, old_regno)--;
VARRAY_INT (n_times_set, new_regno) = 1;
VARRAY_CHAR (may_not_optimize, new_regno) = 0;
REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
REG_IV_INFO (new_regno) = v;
/* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
it must be replaced. */
note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
/* Remove the increment from the list of biv increments,
and record it as a giv. */
*vp = next;
bl->biv_count--;
v->next_iv = bl->giv;
bl->giv = v;
bl->giv_count++;
v->benefit = rtx_cost (SET_SRC (set), SET);
bl->total_benefit += v->benefit;
/* Now replace the biv with DEST_REG in all insns between
the replaced increment and the next increment, and
remember the last insn that needed a replacement. */
for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
p != next->insn;
p = next_insn_in_loop (loop, p))
{
rtx note;
if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
continue;
if (reg_mentioned_p (old_reg, PATTERN (p)))
{
last_use_insn = p;
if (! validate_replace_rtx (old_reg, dest_reg, p))
abort ();
}
for (note = REG_NOTES (p); note; note = XEXP (note, 1))
{
if (GET_CODE (note) == EXPR_LIST)
XEXP (note, 0)
= replace_rtx (XEXP (note, 0), old_reg, dest_reg);
}
}
v->last_use = last_use_insn;
v->lifetime = INSN_LUID (last_use_insn) - INSN_LUID (v->insn);
/* If the lifetime is zero, it means that this register is really
a dead store. So mark this as a giv that can be ignored.
This will not prevent the biv from being eliminated. */
if (v->lifetime == 0)
v->ignore = 1;
if (loop_dump_stream)
fprintf (loop_dump_stream,
"Increment %d of biv %d converted to giv %d.\n",
INSN_UID (v->insn), old_regno, new_regno);
}
}
}
last_increment_giv = max_reg_num () - 1;
/* Search the loop for general induction variables. */
/* A register is a giv if: it is only set once, it is a function of a
biv and a constant (or invariant), and it is not a biv. */
not_every_iteration = 0;
loop_depth = 0;
maybe_multiple = 0;
p = loop_scan_start;
while (1)
{
p = NEXT_INSN (p);
/* At end of a straight-in loop, we are done.
At end of a loop entered at the bottom, scan the top. */
if (p == loop_scan_start)
break;
if (p == loop_end)
{
if (loop_top != 0)
p = loop_top;
else
break;
if (p == loop_scan_start)
break;
}
/* Look for a general induction variable in a register. */
if (GET_CODE (p) == INSN
&& (set = single_set (p))
&& GET_CODE (SET_DEST (set)) == REG
&& ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
{
rtx src_reg;
rtx add_val;
rtx mult_val;
int benefit;
rtx regnote = 0;
rtx last_consec_insn;
dest_reg = SET_DEST (set);
if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
continue;
if (/* SET_SRC is a giv. */
(general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
&mult_val, 0, &benefit)
/* Equivalent expression is a giv. */
|| ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
&& general_induction_var (loop, XEXP (regnote, 0), &src_reg,
&add_val, &mult_val, 0,
&benefit)))
/* Don't try to handle any regs made by loop optimization.
We have nothing on them in regno_first_uid, etc. */
&& REGNO (dest_reg) < max_reg_before_loop
/* Don't recognize a BASIC_INDUCT_VAR here. */
&& dest_reg != src_reg
/* This must be the only place where the register is set. */
&& (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
/* or all sets must be consecutive and make a giv. */
|| (benefit = consec_sets_giv (loop, benefit, p,
src_reg, dest_reg,
&add_val, &mult_val,
&last_consec_insn))))
{
struct induction *v
= (struct induction *) alloca (sizeof (struct induction));
/* If this is a library call, increase benefit. */
if (find_reg_note (p, REG_RETVAL, NULL_RTX))
benefit += libcall_benefit (p);
/* Skip the consecutive insns, if there are any. */
if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
p = last_consec_insn;
record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
benefit, DEST_REG, not_every_iteration,
maybe_multiple, NULL_PTR);
}
}
#ifndef DONT_REDUCE_ADDR
/* Look for givs which are memory addresses. */
/* This resulted in worse code on a VAX 8600. I wonder if it
still does. */
if (GET_CODE (p) == INSN)
find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
maybe_multiple);
#endif
/* Update the status of whether giv can derive other givs. This can
change when we pass a label or an insn that updates a biv. */
if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
|| GET_CODE (p) == CODE_LABEL)
update_giv_derive (loop, p);
/* Past CODE_LABEL, we get to insns that may be executed multiple
times. The only way we can be sure that they can't is if every
every jump insn between here and the end of the loop either
returns, exits the loop, is a forward jump, or is a jump
to the loop start. */
if (GET_CODE (p) == CODE_LABEL)
{
rtx insn = p;
maybe_multiple = 0;
while (1)
{
insn = NEXT_INSN (insn);
if (insn == loop_scan_start)
break;
if (insn == loop_end)
{
if (loop_top != 0)
insn = loop_top;
else
break;
if (insn == loop_scan_start)
break;
}
if (GET_CODE (insn) == JUMP_INSN
&& GET_CODE (PATTERN (insn)) != RETURN
&& (! condjump_p (insn)
|| (JUMP_LABEL (insn) != 0
&& JUMP_LABEL (insn) != loop_scan_start
&& (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
|| INSN_UID (insn) >= max_uid_for_loop
|| (INSN_LUID (JUMP_LABEL (insn))
< INSN_LUID (insn))))))
{
maybe_multiple = 1;
break;
}
}
}
/* Past a jump, we get to insns for which we can't count
on whether they will be executed during each iteration. */
/* This code appears twice in strength_reduce. There is also similar
code in scan_loop. */
if (GET_CODE (p) == JUMP_INSN
/* If we enter the loop in the middle, and scan around to the
beginning, don't set not_every_iteration for that.
This can be any kind of jump, since we want to know if insns
will be executed if the loop is executed. */
&& ! (JUMP_LABEL (p) == loop_top
&& ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
|| (NEXT_INSN (p) == loop_end && condjump_p (p)))))
{
rtx label = 0;
/* If this is a jump outside the loop, then it also doesn't
matter. Check to see if the target of this branch is on the
loop->exits_labels list. */
for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
if (XEXP (label, 0) == JUMP_LABEL (p))
break;
if (! label)
not_every_iteration = 1;
}
else if (GET_CODE (p) == NOTE)
{
/* At the virtual top of a converted loop, insns are again known to
be executed each iteration: logically, the loop begins here
even though the exit code has been duplicated.
Insns are also again known to be executed each iteration at
the LOOP_CONT note. */
if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
|| NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
&& loop_depth == 0)
not_every_iteration = 0;
else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
loop_depth++;
else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
loop_depth--;
}
/* Unlike in the code motion pass where MAYBE_NEVER indicates that
an insn may never be executed, NOT_EVERY_ITERATION indicates whether
or not an insn is known to be executed each iteration of the
loop, whether or not any iterations are known to occur.
Therefore, if we have just passed a label and have no more labels
between here and the test insn of the loop, we know these insns
will be executed each iteration. */
if (not_every_iteration && GET_CODE (p) == CODE_LABEL
&& no_labels_between_p (p, loop_end)
&& loop_insn_first_p (p, loop_cont))
not_every_iteration = 0;
}
/* Try to calculate and save the number of loop iterations. This is
set to zero if the actual number can not be calculated. This must
be called after all giv's have been identified, since otherwise it may
fail if the iteration variable is a giv. */
loop_iterations (loop);
/* Now for each giv for which we still don't know whether or not it is
replaceable, check to see if it is replaceable because its final value
can be calculated. This must be done after loop_iterations is called,
so that final_giv_value will work correctly. */
for (bl = loop_iv_list; bl; bl = bl->next)
{
struct induction *v;
for (v = bl->giv; v; v = v->next_iv)
if (! v->replaceable && ! v->not_replaceable)
check_final_value (loop, v);
}
/* Try to prove that the loop counter variable (if any) is always
nonnegative; if so, record that fact with a REG_NONNEG note
so that "decrement and branch until zero" insn can be used. */
check_dbra_loop (loop, insn_count);
/* Create reg_map to hold substitutions for replaceable giv regs.
Some givs might have been made from biv increments, so look at
reg_iv_type for a suitable size. */
reg_map_size = reg_iv_type->num_elements;
reg_map = (rtx *) xcalloc (reg_map_size, sizeof (rtx));
/* Examine each iv class for feasibility of strength reduction/induction
variable elimination. */
for (bl = loop_iv_list; bl; bl = bl->next)
{
struct induction *v;
int benefit;
int all_reduced;
rtx final_value = 0;
unsigned int nregs;
/* Test whether it will be possible to eliminate this biv
provided all givs are reduced. This is possible if either
the reg is not used outside the loop, or we can compute
what its final value will be.
For architectures with a decrement_and_branch_until_zero insn,
don't do this if we put a REG_NONNEG note on the endtest for
this biv. */
/* Compare against bl->init_insn rather than loop_start.
We aren't concerned with any uses of the biv between
init_insn and loop_start since these won't be affected
by the value of the biv elsewhere in the function, so
long as init_insn doesn't use the biv itself.
March 14, 1989 -- self@bayes.arc.nasa.gov */
if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
&& bl->init_insn
&& INSN_UID (bl->init_insn) < max_uid_for_loop
&& uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
#ifdef HAVE_decrement_and_branch_until_zero
&& ! bl->nonneg
#endif
&& ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
|| ((final_value = final_biv_value (loop, bl))
#ifdef HAVE_decrement_and_branch_until_zero
&& ! bl->nonneg
#endif
))
bl->eliminable = maybe_eliminate_biv (loop, bl, 0, threshold,
insn_count);
else
{
if (loop_dump_stream)
{
fprintf (loop_dump_stream,
"Cannot eliminate biv %d.\n",
bl->regno);
fprintf (loop_dump_stream,
"First use: insn %d, last use: insn %d.\n",
REGNO_FIRST_UID (bl->regno),
REGNO_LAST_UID (bl->regno));
}
}
/* Combine all giv's for this iv_class. */
combine_givs (bl);
/* This will be true at the end, if all givs which depend on this
biv have been strength reduced.
We can't (currently) eliminate the biv unless this is so. */
all_reduced = 1;
/* Check each giv in this class to see if we will benefit by reducing
it. Skip giv's combined with others. */
for (v = bl->giv; v; v = v->next_iv)
{
struct induction *tv;
if (v->ignore || v->same)
continue;
benefit = v->benefit;
/* Reduce benefit if not replaceable, since we will insert
a move-insn to replace the insn that calculates this giv.
Don't do this unless the giv is a user variable, since it
will often be marked non-replaceable because of the duplication
of the exit code outside the loop. In such a case, the copies
we insert are dead and will be deleted. So they don't have
a cost. Similar situations exist. */
/* ??? The new final_[bg]iv_value code does a much better job
of finding replaceable giv's, and hence this code may no longer
be necessary. */
if (! v->replaceable && ! bl->eliminable
&& REG_USERVAR_P (v->dest_reg))
benefit -= copy_cost;
/* Decrease the benefit to count the add-insns that we will
insert to increment the reduced reg for the giv. */
benefit -= add_cost * bl->biv_count;
/* Decide whether to strength-reduce this giv or to leave the code
unchanged (recompute it from the biv each time it is used).
This decision can be made independently for each giv. */
#ifdef AUTO_INC_DEC
/* Attempt to guess whether autoincrement will handle some of the
new add insns; if so, increase BENEFIT (undo the subtraction of
add_cost that was done above). */
if (v->giv_type == DEST_ADDR
&& GET_CODE (v->mult_val) == CONST_INT)
{
if (HAVE_POST_INCREMENT
&& INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
benefit += add_cost * bl->biv_count;
else if (HAVE_PRE_INCREMENT
&& INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
benefit += add_cost * bl->biv_count;
else if (HAVE_POST_DECREMENT
&& -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
benefit += add_cost * bl->biv_count;
else if (HAVE_PRE_DECREMENT
&& -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
benefit += add_cost * bl->biv_count;
}
#endif
/* If an insn is not to be strength reduced, then set its ignore
flag, and clear all_reduced. */
/* A giv that depends on a reversed biv must be reduced if it is
used after the loop exit, otherwise, it would have the wrong
value after the loop exit. To make it simple, just reduce all
of such giv's whether or not we know they are used after the loop
exit. */
if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
&& ! bl->reversed )
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"giv of insn %d not worth while, %d vs %d.\n",
INSN_UID (v->insn),
v->lifetime * threshold * benefit, insn_count);
v->ignore = 1;
all_reduced = 0;
}
else
{
/* Check that we can increment the reduced giv without a
multiply insn. If not, reject it. */
for (tv = bl->biv; tv; tv = tv->next_iv)
if (tv->mult_val == const1_rtx
&& ! product_cheap_p (tv->add_val, v->mult_val))
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"giv of insn %d: would need a multiply.\n",
INSN_UID (v->insn));
v->ignore = 1;
all_reduced = 0;
break;
}
}
}
/* Check for givs whose first use is their definition and whose
last use is the definition of another giv. If so, it is likely
dead and should not be used to derive another giv nor to
eliminate a biv. */
for (v = bl->giv; v; v = v->next_iv)
{
if (v->ignore
|| (v->same && v->same->ignore))
continue;
if (v->last_use)
{
struct induction *v1;
for (v1 = bl->giv; v1; v1 = v1->next_iv)
if (v->last_use == v1->insn)
v->maybe_dead = 1;
}
else if (v->giv_type == DEST_REG
&& REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
{
struct induction *v1;
for (v1 = bl->giv; v1; v1 = v1->next_iv)
if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
v->maybe_dead = 1;
}
}
/* Now that we know which givs will be reduced, try to rearrange the
combinations to reduce register pressure.
recombine_givs calls find_life_end, which needs reg_iv_type and
reg_iv_info to be valid for all pseudos. We do the necessary
reallocation here since it allows to check if there are still
more bivs to process. */
nregs = max_reg_num ();
if (nregs > reg_iv_type->num_elements)
{
/* If there are still more bivs to process, allocate some slack
space so that we're not constantly reallocating these arrays. */
if (bl->next)
nregs += nregs / 4;
/* Reallocate reg_iv_type and reg_iv_info. */
VARRAY_GROW (reg_iv_type, nregs);
VARRAY_GROW (reg_iv_info, nregs);
}
recombine_givs (loop, bl, unroll_p);
/* Reduce each giv that we decided to reduce. */
for (v = bl->giv; v; v = v->next_iv)
{
struct induction *tv;
if (! v->ignore && v->same == 0)
{
int auto_inc_opt = 0;
/* If the code for derived givs immediately below has already
allocated a new_reg, we must keep it. */
if (! v->new_reg)
v->new_reg = gen_reg_rtx (v->mode);
if (v->derived_from)
{
struct induction *d = v->derived_from;
/* In case d->dest_reg is not replaceable, we have
to replace it in v->insn now. */
if (! d->new_reg)
d->new_reg = gen_reg_rtx (d->mode);
PATTERN (v->insn)
= replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
PATTERN (v->insn)
= replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
/* For each place where the biv is incremented, add an
insn to set the new, reduced reg for the giv.
We used to do this only for biv_count != 1, but
this fails when there is a giv after a single biv
increment, e.g. when the last giv was expressed as
pre-decrement. */
for (tv = bl->biv; tv; tv = tv->next_iv)
{
/* We always emit reduced giv increments before the
biv increment when bl->biv_count != 1. So by
emitting the add insns for derived givs after the
biv increment, they pick up the updated value of
the reduced giv.
If the reduced giv is processed with
auto_inc_opt == 1, then it is incremented earlier
than the biv, hence we'll still pick up the right
value.
If it's processed with auto_inc_opt == -1,
that implies that the biv increment is before the
first reduced giv's use. The derived giv's lifetime
is after the reduced giv's lifetime, hence in this
case, the biv increment doesn't matter. */
emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
}
continue;
}
#ifdef AUTO_INC_DEC
/* If the target has auto-increment addressing modes, and
this is an address giv, then try to put the increment
immediately after its use, so that flow can create an
auto-increment addressing mode. */
if (v->giv_type == DEST_ADDR && bl->biv_count == 1
&& bl->biv->always_executed && ! bl->biv->maybe_multiple
/* We don't handle reversed biv's because bl->biv->insn
does not have a valid INSN_LUID. */
&& ! bl->reversed
&& v->always_executed && ! v->maybe_multiple
&& INSN_UID (v->insn) < max_uid_for_loop)
{
/* If other giv's have been combined with this one, then
this will work only if all uses of the other giv's occur
before this giv's insn. This is difficult to check.
We simplify this by looking for the common case where
there is one DEST_REG giv, and this giv's insn is the
last use of the dest_reg of that DEST_REG giv. If the
increment occurs after the address giv, then we can
perform the optimization. (Otherwise, the increment
would have to go before other_giv, and we would not be
able to combine it with the address giv to get an
auto-inc address.) */
if (v->combined_with)
{
struct induction *other_giv = 0;
for (tv = bl->giv; tv; tv = tv->next_iv)
if (tv->same == v)
{
if (other_giv)
break;
else
other_giv = tv;
}
if (! tv && other_giv
&& REGNO (other_giv->dest_reg) < max_reg_before_loop
&& (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
== INSN_UID (v->insn))
&& INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
auto_inc_opt = 1;
}
/* Check for case where increment is before the address
giv. Do this test in "loop order". */
else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
&& (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
|| (INSN_LUID (bl->biv->insn)
> INSN_LUID (loop_scan_start))))
|| (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
&& (INSN_LUID (loop_scan_start)
< INSN_LUID (bl->biv->insn))))
auto_inc_opt = -1;
else
auto_inc_opt = 1;
#ifdef HAVE_cc0
{
rtx prev;
/* We can't put an insn immediately after one setting
cc0, or immediately before one using cc0. */
if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
|| (auto_inc_opt == -1
&& (prev = prev_nonnote_insn (v->insn)) != 0
&& GET_RTX_CLASS (GET_CODE (prev)) == 'i'
&& sets_cc0_p (PATTERN (prev))))
auto_inc_opt = 0;
}
#endif
if (auto_inc_opt)
v->auto_inc_opt = 1;
}
#endif
/* For each place where the biv is incremented, add an insn
to increment the new, reduced reg for the giv. */
for (tv = bl->biv; tv; tv = tv->next_iv)
{
rtx insert_before;
if (! auto_inc_opt)
insert_before = tv->insn;
else if (auto_inc_opt == 1)
insert_before = NEXT_INSN (v->insn);
else
insert_before = v->insn;
if (tv->mult_val == const1_rtx)
emit_iv_add_mult (tv->add_val, v->mult_val,
v->new_reg, v->new_reg, insert_before);
else /* tv->mult_val == const0_rtx */
/* A multiply is acceptable here
since this is presumed to be seldom executed. */
emit_iv_add_mult (tv->add_val, v->mult_val,
v->add_val, v->new_reg, insert_before);
}
/* Add code at loop start to initialize giv's reduced reg. */
emit_iv_add_mult (bl->initial_value, v->mult_val,
v->add_val, v->new_reg, loop_start);
}
}
/* Rescan all givs. If a giv is the same as a giv not reduced, mark it
as not reduced.
For each giv register that can be reduced now: if replaceable,
substitute reduced reg wherever the old giv occurs;
else add new move insn "giv_reg = reduced_reg". */
for (v = bl->giv; v; v = v->next_iv)
{
if (v->same && v->same->ignore)
v->ignore = 1;
if (v->ignore)
continue;
/* Update expression if this was combined, in case other giv was
replaced. */
if (v->same)
v->new_reg = replace_rtx (v->new_reg,
v->same->dest_reg, v->same->new_reg);
if (v->giv_type == DEST_ADDR)
/* Store reduced reg as the address in the memref where we found
this giv. */
validate_change (v->insn, v->location, v->new_reg, 0);
else if (v->replaceable)
{
reg_map[REGNO (v->dest_reg)] = v->new_reg;
#if 0
/* I can no longer duplicate the original problem. Perhaps
this is unnecessary now? */
/* Replaceable; it isn't strictly necessary to delete the old
insn and emit a new one, because v->dest_reg is now dead.
However, especially when unrolling loops, the special
handling for (set REG0 REG1) in the second cse pass may
make v->dest_reg live again. To avoid this problem, emit
an insn to set the original giv reg from the reduced giv.
We can not delete the original insn, since it may be part
of a LIBCALL, and the code in flow that eliminates dead
libcalls will fail if it is deleted. */
emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
v->insn);
#endif
}
else
{
/* Not replaceable; emit an insn to set the original giv reg from
the reduced giv, same as above. */
emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
v->insn);
}
/* When a loop is reversed, givs which depend on the reversed
biv, and which are live outside the loop, must be set to their
correct final value. This insn is only needed if the giv is
not replaceable. The correct final value is the same as the
value that the giv starts the reversed loop with. */
if (bl->reversed && ! v->replaceable)
emit_iv_add_mult (bl->initial_value, v->mult_val,
v->add_val, v->dest_reg, end_insert_before);
else if (v->final_value)
{
rtx insert_before;
/* If the loop has multiple exits, emit the insn before the
loop to ensure that it will always be executed no matter
how the loop exits. Otherwise, emit the insn after the loop,
since this is slightly more efficient. */
if (loop->exit_count)
insert_before = loop_start;
else
insert_before = end_insert_before;
emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
insert_before);
#if 0
/* If the insn to set the final value of the giv was emitted
before the loop, then we must delete the insn inside the loop
that sets it. If this is a LIBCALL, then we must delete
every insn in the libcall. Note, however, that
final_giv_value will only succeed when there are multiple
exits if the giv is dead at each exit, hence it does not
matter that the original insn remains because it is dead
anyways. */
/* Delete the insn inside the loop that sets the giv since
the giv is now set before (or after) the loop. */
delete_insn (v->insn);
#endif
}
if (loop_dump_stream)
{
fprintf (loop_dump_stream, "giv at %d reduced to ",
INSN_UID (v->insn));
print_rtl (loop_dump_stream, v->new_reg);
fprintf (loop_dump_stream, "\n");
}
}
/* All the givs based on the biv bl have been reduced if they
merit it. */
/* For each giv not marked as maybe dead that has been combined with a
second giv, clear any "maybe dead" mark on that second giv.
v->new_reg will either be or refer to the register of the giv it
combined with.
Doing this clearing avoids problems in biv elimination where a
giv's new_reg is a complex value that can't be put in the insn but
the giv combined with (with a reg as new_reg) is marked maybe_dead.
Since the register will be used in either case, we'd prefer it be
used from the simpler giv. */
for (v = bl->giv; v; v = v->next_iv)
if (! v->maybe_dead && v->same)
v->same->maybe_dead = 0;
/* Try to eliminate the biv, if it is a candidate.
This won't work if ! all_reduced,
since the givs we planned to use might not have been reduced.
We have to be careful that we didn't initially think we could eliminate
this biv because of a giv that we now think may be dead and shouldn't
be used as a biv replacement.
Also, there is the possibility that we may have a giv that looks
like it can be used to eliminate a biv, but the resulting insn
isn't valid. This can happen, for example, on the 88k, where a
JUMP_INSN can compare a register only with zero. Attempts to
replace it with a compare with a constant will fail.
Note that in cases where this call fails, we may have replaced some
of the occurrences of the biv with a giv, but no harm was done in
doing so in the rare cases where it can occur. */
if (all_reduced == 1 && bl->eliminable
&& maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
{
/* ?? If we created a new test to bypass the loop entirely,
or otherwise drop straight in, based on this test, then
we might want to rewrite it also. This way some later
pass has more hope of removing the initialization of this
biv entirely. */
/* If final_value != 0, then the biv may be used after loop end
and we must emit an insn to set it just in case.
Reversed bivs already have an insn after the loop setting their
value, so we don't need another one. We can't calculate the
proper final value for such a biv here anyways. */
if (final_value != 0 && ! bl->reversed)
{
rtx insert_before;
/* If the loop has multiple exits, emit the insn before the
loop to ensure that it will always be executed no matter
how the loop exits. Otherwise, emit the insn after the
loop, since this is slightly more efficient. */
if (loop->exit_count)
insert_before = loop_start;
else
insert_before = end_insert_before;
emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
end_insert_before);
}
#if 0
/* Delete all of the instructions inside the loop which set
the biv, as they are all dead. If is safe to delete them,
because an insn setting a biv will never be part of a libcall. */
/* However, deleting them will invalidate the regno_last_uid info,
so keeping them around is more convenient. Final_biv_value
will only succeed when there are multiple exits if the biv
is dead at each exit, hence it does not matter that the original
insn remains, because it is dead anyways. */
for (v = bl->biv; v; v = v->next_iv)
delete_insn (v->insn);
#endif
if (loop_dump_stream)
fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
bl->regno);
}
}
/* Go through all the instructions in the loop, making all the
register substitutions scheduled in REG_MAP. */
for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
|| GET_CODE (p) == CALL_INSN)
{
replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
INSN_CODE (p) = -1;
}
if (loop_info->n_iterations > 0)
{
/* When we completely unroll a loop we will likely not need the increment
of the loop BIV and we will not need the conditional branch at the
end of the loop. */
unrolled_insn_copies = insn_count - 2;
#ifdef HAVE_cc0
/* When we completely unroll a loop on a HAVE_cc0 machine we will not
need the comparison before the conditional branch at the end of the
loop. */
unrolled_insn_copies -= 1;
#endif
/* We'll need one copy for each loop iteration. */
unrolled_insn_copies *= loop_info->n_iterations;
/* A little slop to account for the ability to remove initialization
code, better CSE, and other secondary benefits of completely
unrolling some loops. */
unrolled_insn_copies -= 1;
/* Clamp the value. */
if (unrolled_insn_copies < 0)
unrolled_insn_copies = 0;
}
/* Unroll loops from within strength reduction so that we can use the
induction variable information that strength_reduce has already
collected. Always unroll loops that would be as small or smaller
unrolled than when rolled. */
if (unroll_p
|| (loop_info->n_iterations > 0
&& unrolled_insn_copies <= insn_count))
unroll_loop (loop, insn_count, end_insert_before, 1);
#ifdef HAVE_decrement_and_branch_on_count
/* Instrument the loop with BCT insn. */
if (HAVE_decrement_and_branch_on_count && bct_p
&& flag_branch_on_count_reg)
insert_bct (loop);
#endif /* HAVE_decrement_and_branch_on_count */
if (loop_dump_stream)
fprintf (loop_dump_stream, "\n");
egress:
VARRAY_FREE (reg_iv_type);
VARRAY_FREE (reg_iv_info);
free (reg_biv_class);
if (reg_map)
free (reg_map);
}
/* Return 1 if X is a valid source for an initial value (or as value being
compared against in an initial test).
X must be either a register or constant and must not be clobbered between
the current insn and the start of the loop.
INSN is the insn containing X. */
static int
valid_initial_value_p (x, insn, call_seen, loop_start)
rtx x;
rtx insn;
int call_seen;
rtx loop_start;
{
if (CONSTANT_P (x))
return 1;
/* Only consider pseudos we know about initialized in insns whose luids
we know. */
if (GET_CODE (x) != REG
|| REGNO (x) >= max_reg_before_loop)
return 0;
/* Don't use call-clobbered registers across a call which clobbers it. On
some machines, don't use any hard registers at all. */
if (REGNO (x) < FIRST_PSEUDO_REGISTER
&& (SMALL_REGISTER_CLASSES
|| (call_used_regs[REGNO (x)] && call_seen)))
return 0;
/* Don't use registers that have been clobbered before the start of the
loop. */
if (reg_set_between_p (x, insn, loop_start))
return 0;
return 1;
}
/* Scan X for memory refs and check each memory address
as a possible giv. INSN is the insn whose pattern X comes from.
NOT_EVERY_ITERATION is 1 if the insn might not be executed during
every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
more thanonce in each loop iteration. */
static void
find_mem_givs (loop, x, insn, not_every_iteration, maybe_multiple)
const struct loop *loop;
rtx x;
rtx insn;
int not_every_iteration, maybe_multiple;
{
register int i, j;
register enum rtx_code code;
register const char *fmt;
if (x == 0)
return;
code = GET_CODE (x);
switch (code)
{
case REG:
case CONST_INT:
case CONST:
case CONST_DOUBLE:
case SYMBOL_REF:
case LABEL_REF:
case PC:
case CC0:
case ADDR_VEC:
case ADDR_DIFF_VEC:
case USE:
case CLOBBER:
return;
case MEM:
{
rtx src_reg;
rtx add_val;
rtx mult_val;
int benefit;
/* This code used to disable creating GIVs with mult_val == 1 and
add_val == 0. However, this leads to lost optimizations when
it comes time to combine a set of related DEST_ADDR GIVs, since
this one would not be seen. */
if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
&mult_val, 1, &benefit))
{
/* Found one; record it. */
struct induction *v
= (struct induction *) oballoc (sizeof (struct induction));
record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
add_val, benefit, DEST_ADDR, not_every_iteration,
maybe_multiple, &XEXP (x, 0));
v->mem_mode = GET_MODE (x);
}
}
return;
default:
break;
}
/* Recursively scan the subexpressions for other mem refs. */
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
if (fmt[i] == 'e')
find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
maybe_multiple);
else if (fmt[i] == 'E')
for (j = 0; j < XVECLEN (x, i); j++)
find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
maybe_multiple);
}
/* Fill in the data about one biv update.
V is the `struct induction' in which we record the biv. (It is
allocated by the caller, with alloca.)
INSN is the insn that sets it.
DEST_REG is the biv's reg.
MULT_VAL is const1_rtx if the biv is being incremented here, in which case
INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
being set to INC_VAL.
NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
can be executed more than once per iteration. If MAYBE_MULTIPLE
and NOT_EVERY_ITERATION are both zero, we know that the biv update is
executed exactly once per iteration. */
static void
record_biv (v, insn, dest_reg, inc_val, mult_val, location,
not_every_iteration, maybe_multiple, multi_insn_incr)
struct induction *v;
rtx insn;
rtx dest_reg;
rtx inc_val;
rtx mult_val;
rtx *location;
int not_every_iteration;
int maybe_multiple;
int multi_insn_incr;
{
struct iv_class *bl;
v->insn = insn;
v->src_reg = dest_reg;
v->dest_reg = dest_reg;
v->mult_val = mult_val;
v->add_val = inc_val;
v->location = location;
v->mode = GET_MODE (dest_reg);
v->always_computable = ! not_every_iteration;
v->always_executed = ! not_every_iteration;
v->maybe_multiple = maybe_multiple;
v->multi_insn_incr = multi_insn_incr;
/* Add this to the reg's iv_class, creating a class
if this is the first incrementation of the reg. */
bl = reg_biv_class[REGNO (dest_reg)];
if (bl == 0)
{
/* Create and initialize new iv_class. */
bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
bl->regno = REGNO (dest_reg);
bl->biv = 0;
bl->giv = 0;
bl->biv_count = 0;
bl->giv_count = 0;
/* Set initial value to the reg itself. */
bl->initial_value = dest_reg;
/* We haven't seen the initializing insn yet */
bl->init_insn = 0;
bl->init_set = 0;
bl->initial_test = 0;
bl->incremented = 0;
bl->eliminable = 0;
bl->nonneg = 0;
bl->reversed = 0;
bl->total_benefit = 0;
/* Add this class to loop_iv_list. */
bl->next = loop_iv_list;
loop_iv_list = bl;
/* Put it in the array of biv register classes. */
reg_biv_class[REGNO (dest_reg)] = bl;
}
/* Update IV_CLASS entry for this biv. */
v->next_iv = bl->biv;
bl->biv = v;
bl->biv_count++;
if (mult_val == const1_rtx)
bl->incremented = 1;
if (loop_dump_stream)
{
fprintf (loop_dump_stream,
"Insn %d: possible biv, reg %d,",
INSN_UID (insn), REGNO (dest_reg));
if (GET_CODE (inc_val) == CONST_INT)
{
fprintf (loop_dump_stream, " const =");
fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
fputc ('\n', loop_dump_stream);
}
else
{
fprintf (loop_dump_stream, " const = ");
print_rtl (loop_dump_stream, inc_val);
fprintf (loop_dump_stream, "\n");
}
}
}
/* Fill in the data about one giv.
V is the `struct induction' in which we record the giv. (It is
allocated by the caller, with alloca.)
INSN is the insn that sets it.
BENEFIT estimates the savings from deleting this insn.
TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
into a register or is used as a memory address.
SRC_REG is the biv reg which the giv is computed from.
DEST_REG is the giv's reg (if the giv is stored in a reg).
MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
LOCATION points to the place where this giv's value appears in INSN. */
static void
record_giv (loop, v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
type, not_every_iteration, maybe_multiple, location)
const struct loop *loop;
struct induction *v;
rtx insn;
rtx src_reg;
rtx dest_reg;
rtx mult_val, add_val;
int benefit;
enum g_types type;
int not_every_iteration, maybe_multiple;
rtx *location;
{
struct induction *b;
struct iv_class *bl;
rtx set = single_set (insn);
v->insn = insn;
v->src_reg = src_reg;
v->giv_type = type;
v->dest_reg = dest_reg;
v->mult_val = mult_val;
v->add_val = add_val;
v->benefit = benefit;
v->location = location;
v->cant_derive = 0;
v->combined_with = 0;
v->maybe_multiple = maybe_multiple;
v->multi_insn_incr = 0;
v->maybe_dead = 0;
v->derive_adjustment = 0;
v->same = 0;
v->ignore = 0;
v->new_reg = 0;
v->final_value = 0;
v->same_insn = 0;
v->auto_inc_opt = 0;
v->unrolled = 0;
v->shared = 0;
v->derived_from = 0;
v->last_use = 0;
/* The v->always_computable field is used in update_giv_derive, to
determine whether a giv can be used to derive another giv. For a
DEST_REG giv, INSN computes a new value for the giv, so its value
isn't computable if INSN insn't executed every iteration.
However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
it does not compute a new value. Hence the value is always computable
regardless of whether INSN is executed each iteration. */
if (type == DEST_ADDR)
v->always_computable = 1;
else
v->always_computable = ! not_every_iteration;
v->always_executed = ! not_every_iteration;
if (type == DEST_ADDR)
{
v->mode = GET_MODE (*location);
v->lifetime = 1;
}
else /* type == DEST_REG */
{
v->mode = GET_MODE (SET_DEST (set));
v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
- uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
/* If the lifetime is zero, it means that this register is
really a dead store. So mark this as a giv that can be
ignored. This will not prevent the biv from being eliminated. */
if (v->lifetime == 0)
v->ignore = 1;
REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
REG_IV_INFO (REGNO (dest_reg)) = v;
}
/* Add the giv to the class of givs computed from one biv. */
bl = reg_biv_class[REGNO (src_reg)];
if (bl)
{
v->next_iv = bl->giv;
bl->giv = v;
/* Don't count DEST_ADDR. This is supposed to count the number of
insns that calculate givs. */
if (type == DEST_REG)
bl->giv_count++;
bl->total_benefit += benefit;
}
else
/* Fatal error, biv missing for this giv? */
abort ();
if (type == DEST_ADDR)
v->replaceable = 1;
else
{
/* The giv can be replaced outright by the reduced register only if all
of the following conditions are true:
- the insn that sets the giv is always executed on any iteration
on which the giv is used at all
(there are two ways to deduce this:
either the insn is executed on every iteration,
or all uses follow that insn in the same basic block),
- the giv is not used outside the loop
- no assignments to the biv occur during the giv's lifetime. */
if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
/* Previous line always fails if INSN was moved by loop opt. */
&& uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
< INSN_LUID (loop->end)
&& (! not_every_iteration
|| last_use_this_basic_block (dest_reg, insn)))
{
/* Now check that there are no assignments to the biv within the
giv's lifetime. This requires two separate checks. */
/* Check each biv update, and fail if any are between the first
and last use of the giv.
If this loop contains an inner loop that was unrolled, then
the insn modifying the biv may have been emitted by the loop
unrolling code, and hence does not have a valid luid. Just
mark the biv as not replaceable in this case. It is not very
useful as a biv, because it is used in two different loops.
It is very unlikely that we would be able to optimize the giv
using this biv anyways. */
v->replaceable = 1;
for (b = bl->biv; b; b = b->next_iv)
{
if (INSN_UID (b->insn) >= max_uid_for_loop
|| ((uid_luid[INSN_UID (b->insn)]
>= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
&& (uid_luid[INSN_UID (b->insn)]
<= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
{
v->replaceable = 0;
v->not_replaceable = 1;
break;
}
}
/* If there are any backwards branches that go from after the
biv update to before it, then this giv is not replaceable. */
if (v->replaceable)
for (b = bl->biv; b; b = b->next_iv)
if (back_branch_in_range_p (loop, b->insn))
{
v->replaceable = 0;
v->not_replaceable = 1;
break;
}
}
else
{
/* May still be replaceable, we don't have enough info here to
decide. */
v->replaceable = 0;
v->not_replaceable = 0;
}
}
/* Record whether the add_val contains a const_int, for later use by
combine_givs. */
{
rtx tem = add_val;
v->no_const_addval = 1;
if (tem == const0_rtx)
;
else if (GET_CODE (tem) == CONST_INT)
v->no_const_addval = 0;
else if (GET_CODE (tem) == PLUS)
{
while (1)
{
if (GET_CODE (XEXP (tem, 0)) == PLUS)
tem = XEXP (tem, 0);
else if (GET_CODE (XEXP (tem, 1)) == PLUS)
tem = XEXP (tem, 1);
else
break;
}
if (GET_CODE (XEXP (tem, 1)) == CONST_INT)
v->no_const_addval = 0;
}
}
if (loop_dump_stream)
{
if (type == DEST_REG)
fprintf (loop_dump_stream, "Insn %d: giv reg %d",
INSN_UID (insn), REGNO (dest_reg));
else
fprintf (loop_dump_stream, "Insn %d: dest address",
INSN_UID (insn));
fprintf (loop_dump_stream, " src reg %d benefit %d",
REGNO (src_reg), v->benefit);
fprintf (loop_dump_stream, " lifetime %d",
v->lifetime);
if (v->replaceable)
fprintf (loop_dump_stream, " replaceable");
if (v->no_const_addval)
fprintf (loop_dump_stream, " ncav");
if (GET_CODE (mult_val) == CONST_INT)
{
fprintf (loop_dump_stream, " mult ");
fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
}
else
{
fprintf (loop_dump_stream, " mult ");
print_rtl (loop_dump_stream, mult_val);
}
if (GET_CODE (add_val) == CONST_INT)
{
fprintf (loop_dump_stream, " add ");
fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
}
else
{
fprintf (loop_dump_stream, " add ");
print_rtl (loop_dump_stream, add_val);
}
}
if (loop_dump_stream)
fprintf (loop_dump_stream, "\n");
}
/* All this does is determine whether a giv can be made replaceable because
its final value can be calculated. This code can not be part of record_giv
above, because final_giv_value requires that the number of loop iterations
be known, and that can not be accurately calculated until after all givs
have been identified. */
static void
check_final_value (loop, v)
const struct loop *loop;
struct induction *v;
{
struct iv_class *bl;
rtx final_value = 0;
bl = reg_biv_class[REGNO (v->src_reg)];
/* DEST_ADDR givs will never reach here, because they are always marked
replaceable above in record_giv. */
/* The giv can be replaced outright by the reduced register only if all
of the following conditions are true:
- the insn that sets the giv is always executed on any iteration
on which the giv is used at all
(there are two ways to deduce this:
either the insn is executed on every iteration,
or all uses follow that insn in the same basic block),
- its final value can be calculated (this condition is different
than the one above in record_giv)
- no assignments to the biv occur during the giv's lifetime. */
#if 0
/* This is only called now when replaceable is known to be false. */
/* Clear replaceable, so that it won't confuse final_giv_value. */
v->replaceable = 0;
#endif
if ((final_value = final_giv_value (loop, v))
&& (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
{
int biv_increment_seen = 0;
rtx p = v->insn;
rtx last_giv_use;
v->replaceable = 1;
/* When trying to determine whether or not a biv increment occurs
during the lifetime of the giv, we can ignore uses of the variable
outside the loop because final_value is true. Hence we can not
use regno_last_uid and regno_first_uid as above in record_giv. */
/* Search the loop to determine whether any assignments to the
biv occur during the giv's lifetime. Start with the insn
that sets the giv, and search around the loop until we come
back to that insn again.
Also fail if there is a jump within the giv's lifetime that jumps
to somewhere outside the lifetime but still within the loop. This
catches spaghetti code where the execution order is not linear, and
hence the above test fails. Here we assume that the giv lifetime
does not extend from one iteration of the loop to the next, so as
to make the test easier. Since the lifetime isn't known yet,
this requires two loops. See also record_giv above. */
last_giv_use = v->insn;
while (1)
{
p = NEXT_INSN (p);
if (p == loop->end)
p = NEXT_INSN (loop->start);
if (p == v->insn)
break;
if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
|| GET_CODE (p) == CALL_INSN)
{
if (biv_increment_seen)
{
if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
{
v->replaceable = 0;
v->not_replaceable = 1;
break;
}
}
else if (reg_set_p (v->src_reg, PATTERN (p)))
biv_increment_seen = 1;
else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
last_giv_use = p;
}
}
/* Now that the lifetime of the giv is known, check for branches
from within the lifetime to outside the lifetime if it is still
replaceable. */
if (v->replaceable)
{
p = v->insn;
while (1)
{
p = NEXT_INSN (p);
if (p == loop->end)
p = NEXT_INSN (loop->start);
if (p == last_giv_use)
break;
if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
&& LABEL_NAME (JUMP_LABEL (p))
&& ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
&& loop_insn_first_p (loop->start, JUMP_LABEL (p)))
|| (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
&& loop_insn_first_p (JUMP_LABEL (p), loop->end))))
{
v->replaceable = 0;
v->not_replaceable = 1;
if (loop_dump_stream)
fprintf (loop_dump_stream,
"Found branch outside giv lifetime.\n");
break;
}
}
}
/* If it is replaceable, then save the final value. */
if (v->replaceable)
v->final_value = final_value;
}
if (loop_dump_stream && v->replaceable)
fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
INSN_UID (v->insn), REGNO (v->dest_reg));
}
/* Update the status of whether a giv can derive other givs.
We need to do something special if there is or may be an update to the biv
between the time the giv is defined and the time it is used to derive
another giv.
In addition, a giv that is only conditionally set is not allowed to
derive another giv once a label has been passed.
The cases we look at are when a label or an update to a biv is passed. */
static void
update_giv_derive (loop, p)
const struct loop *loop;
rtx p;
{
struct iv_class *bl;
struct induction *biv, *giv;
rtx tem;
int dummy;
/* Search all IV classes, then all bivs, and finally all givs.
There are three cases we are concerned with. First we have the situation
of a giv that is only updated conditionally. In that case, it may not
derive any givs after a label is passed.
The second case is when a biv update occurs, or may occur, after the
definition of a giv. For certain biv updates (see below) that are
known to occur between the giv definition and use, we can adjust the
giv definition. For others, or when the biv update is conditional,
we must prevent the giv from deriving any other givs. There are two
sub-cases within this case.
If this is a label, we are concerned with any biv update that is done
conditionally, since it may be done after the giv is defined followed by
a branch here (actually, we need to pass both a jump and a label, but
this extra tracking doesn't seem worth it).
If this is a jump, we are concerned about any biv update that may be
executed multiple times. We are actually only concerned about
backward jumps, but it is probably not worth performing the test
on the jump again here.
If this is a biv update, we must adjust the giv status to show that a
subsequent biv update was performed. If this adjustment cannot be done,
the giv cannot derive further givs. */
for (bl = loop_iv_list; bl; bl = bl->next)
for (biv = bl->biv; biv; biv = biv->next_iv)
if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
|| biv->insn == p)
{
for (giv = bl->giv; giv; giv = giv->next_iv)
{
/* If cant_derive is already true, there is no point in
checking all of these conditions again. */
if (giv->cant_derive)
continue;
/* If this giv is conditionally set and we have passed a label,
it cannot derive anything. */
if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
giv->cant_derive = 1;
/* Skip givs that have mult_val == 0, since
they are really invariants. Also skip those that are
replaceable, since we know their lifetime doesn't contain
any biv update. */
else if (giv->mult_val == const0_rtx || giv->replaceable)
continue;
/* The only way we can allow this giv to derive another
is if this is a biv increment and we can form the product
of biv->add_val and giv->mult_val. In this case, we will
be able to compute a compensation. */
else if (biv->insn == p)
{
tem = 0;
if (biv->mult_val == const1_rtx)
tem = simplify_giv_expr (loop,
gen_rtx_MULT (giv->mode,
biv->add_val,
giv->mult_val),
&dummy);
if (tem && giv->derive_adjustment)
tem = simplify_giv_expr
(loop,
gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
&dummy);
if (tem)
giv->derive_adjustment = tem;
else
giv->cant_derive = 1;
}
else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
|| (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
giv->cant_derive = 1;
}
}
}
/* Check whether an insn is an increment legitimate for a basic induction var.
X is the source of insn P, or a part of it.
MODE is the mode in which X should be interpreted.
DEST_REG is the putative biv, also the destination of the insn.
We accept patterns of these forms:
REG = REG + INVARIANT (includes REG = REG - CONSTANT)
REG = INVARIANT + REG
If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
store the additive term into *INC_VAL, and store the place where
we found the additive term into *LOCATION.
If X is an assignment of an invariant into DEST_REG, we set
*MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
We also want to detect a BIV when it corresponds to a variable
whose mode was promoted via PROMOTED_MODE. In that case, an increment
of the variable may be a PLUS that adds a SUBREG of that variable to
an invariant and then sign- or zero-extends the result of the PLUS
into the variable.
Most GIVs in such cases will be in the promoted mode, since that is the
probably the natural computation mode (and almost certainly the mode
used for addresses) on the machine. So we view the pseudo-reg containing
the variable as the BIV, as if it were simply incremented.
Note that treating the entire pseudo as a BIV will result in making
simple increments to any GIVs based on it. However, if the variable
overflows in its declared mode but not its promoted mode, the result will
be incorrect. This is acceptable if the variable is signed, since
overflows in such cases are undefined, but not if it is unsigned, since
those overflows are defined. So we only check for SIGN_EXTEND and
not ZERO_EXTEND.
If we cannot find a biv, we return 0. */
static int
basic_induction_var (loop, x, mode, dest_reg, p, inc_val, mult_val,
location, multi_insn_incr)
const struct loop *loop;
register rtx x;
enum machine_mode mode;
rtx dest_reg;
rtx p;
rtx *inc_val;
rtx *mult_val;
rtx **location;
int *multi_insn_incr;
{
register enum rtx_code code;
rtx *argp, arg;
rtx insn, set = 0;
code = GET_CODE (x);
*location = NULL;
switch (code)
{
case PLUS:
if (rtx_equal_p (XEXP (x, 0), dest_reg)
|| (GET_CODE (XEXP (x, 0)) == SUBREG
&& SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
&& SUBREG_REG (XEXP (x, 0)) == dest_reg))
{
argp = &XEXP (x, 1);
}
else if (rtx_equal_p (XEXP (x, 1), dest_reg)
|| (GET_CODE (XEXP (x, 1)) == SUBREG
&& SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
&& SUBREG_REG (XEXP (x, 1)) == dest_reg))
{
argp = &XEXP (x, 0);
}
else
return 0;
arg = *argp;
if (loop_invariant_p (loop, arg) != 1)
return 0;
*inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
*mult_val = const1_rtx;
*location = argp;
return 1;
case SUBREG:
/* If this is a SUBREG for a promoted variable, check the inner
value. */
if (SUBREG_PROMOTED_VAR_P (x))
return basic_induction_var (loop, SUBREG_REG (x),
GET_MODE (SUBREG_REG (x)),
dest_reg, p, inc_val, mult_val, location,
multi_insn_incr);
return 0;
case REG:
/* If this register is assigned in a previous insn, look at its
source, but don't go outside the loop or past a label. */
insn = p;
while (1)
{
do {
insn = PREV_INSN (insn);
} while (insn && GET_CODE (insn) == NOTE
&& NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
if (!insn)
break;
set = single_set (insn);
if (set == 0)
break;
if ((SET_DEST (set) == x
|| (GET_CODE (SET_DEST (set)) == SUBREG
&& (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
<= UNITS_PER_WORD)
&& (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
== MODE_INT)
&& SUBREG_REG (SET_DEST (set)) == x))
&& basic_induction_var (loop, SET_SRC (set),
(GET_MODE (SET_SRC (set)) == VOIDmode
? GET_MODE (x)
: GET_MODE (SET_SRC (set))),
dest_reg, insn,
inc_val, mult_val, location,
multi_insn_incr))
{
*multi_insn_incr = 1;
return 1;
}
}
/* ... fall through ... */
/* Can accept constant setting of biv only when inside inner most loop.
Otherwise, a biv of an inner loop may be incorrectly recognized
as a biv of the outer loop,
causing code to be moved INTO the inner loop. */
case MEM:
if (loop_invariant_p (loop, x) != 1)
return 0;
case CONST_INT:
case SYMBOL_REF:
case CONST:
/* convert_modes aborts if we try to convert to or from CCmode, so just
exclude that case. It is very unlikely that a condition code value
would be a useful iterator anyways. */
if (loop->level == 1
&& GET_MODE_CLASS (mode) != MODE_CC
&& GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
{
/* Possible bug here? Perhaps we don't know the mode of X. */
*inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
*mult_val = const0_rtx;
return 1;
}
else
return 0;
case SIGN_EXTEND:
return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
dest_reg, p, inc_val, mult_val, location,
multi_insn_incr);
case ASHIFTRT:
/* Similar, since this can be a sign extension. */
for (insn = PREV_INSN (p);
(insn && GET_CODE (insn) == NOTE
&& NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
insn = PREV_INSN (insn))
;
if (insn)
set = single_set (insn);
if (set && SET_DEST (set) == XEXP (x, 0)
&& GET_CODE (XEXP (x, 1)) == CONST_INT
&& INTVAL (XEXP (x, 1)) >= 0
&& GET_CODE (SET_SRC (set)) == ASHIFT
&& XEXP (x, 1) == XEXP (SET_SRC (set), 1)
&& basic_induction_var (loop, XEXP (SET_SRC (set), 0),
GET_MODE (XEXP (x, 0)),
dest_reg, insn, inc_val, mult_val,
location, multi_insn_incr))
{
*multi_insn_incr = 1;
return 1;
}
return 0;
default:
return 0;
}
}
/* A general induction variable (giv) is any quantity that is a linear
function of a basic induction variable,
i.e. giv = biv * mult_val + add_val.
The coefficients can be any loop invariant quantity.
A giv need not be computed directly from the biv;
it can be computed by way of other givs. */
/* Determine whether X computes a giv.
If it does, return a nonzero value
which is the benefit from eliminating the computation of X;
set *SRC_REG to the register of the biv that it is computed from;
set *ADD_VAL and *MULT_VAL to the coefficients,
such that the value of X is biv * mult + add; */
static int
general_induction_var (loop, x, src_reg, add_val, mult_val, is_addr, pbenefit)
const struct loop *loop;
rtx x;
rtx *src_reg;
rtx *add_val;
rtx *mult_val;
int is_addr;
int *pbenefit;
{
rtx orig_x = x;
char *storage;
/* If this is an invariant, forget it, it isn't a giv. */
if (loop_invariant_p (loop, x) == 1)
return 0;
/* See if the expression could be a giv and get its form.
Mark our place on the obstack in case we don't find a giv. */
storage = (char *) oballoc (0);
*pbenefit = 0;
x = simplify_giv_expr (loop, x, pbenefit);
if (x == 0)
{
obfree (storage);
return 0;
}
switch (GET_CODE (x))
{
case USE:
case CONST_INT:
/* Since this is now an invariant and wasn't before, it must be a giv
with MULT_VAL == 0. It doesn't matter which BIV we associate this
with. */
*src_reg = loop_iv_list->biv->dest_reg;
*mult_val = const0_rtx;
*add_val = x;
break;
case REG:
/* This is equivalent to a BIV. */
*src_reg = x;
*mult_val = const1_rtx;
*add_val = const0_rtx;
break;
case PLUS:
/* Either (plus (biv) (invar)) or
(plus (mult (biv) (invar_1)) (invar_2)). */
if (GET_CODE (XEXP (x, 0)) == MULT)
{
*src_reg = XEXP (XEXP (x, 0), 0);
*mult_val = XEXP (XEXP (x, 0), 1);
}
else
{
*src_reg = XEXP (x, 0);
*mult_val = const1_rtx;
}
*add_val = XEXP (x, 1);
break;
case MULT:
/* ADD_VAL is zero. */
*src_reg = XEXP (x, 0);
*mult_val = XEXP (x, 1);
*add_val = const0_rtx;
break;
default:
abort ();
}
/* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
unless they are CONST_INT). */
if (GET_CODE (*add_val) == USE)
*add_val = XEXP (*add_val, 0);
if (GET_CODE (*mult_val) == USE)
*mult_val = XEXP (*mult_val, 0);
if (is_addr)
{
#ifdef ADDRESS_COST
*pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
#else
*pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
#endif
}
else
*pbenefit += rtx_cost (orig_x, SET);
/* Always return true if this is a giv so it will be detected as such,
even if the benefit is zero or negative. This allows elimination
of bivs that might otherwise not be eliminated. */
return 1;
}
/* Given an expression, X, try to form it as a linear function of a biv.
We will canonicalize it to be of the form
(plus (mult (BIV) (invar_1))
(invar_2))
with possible degeneracies.
The invariant expressions must each be of a form that can be used as a
machine operand. We surround then with a USE rtx (a hack, but localized
and certainly unambiguous!) if not a CONST_INT for simplicity in this
routine; it is the caller's responsibility to strip them.
If no such canonicalization is possible (i.e., two biv's are used or an
expression that is neither invariant nor a biv or giv), this routine
returns 0.
For a non-zero return, the result will have a code of CONST_INT, USE,
REG (for a BIV), PLUS, or MULT. No other codes will occur.
*BENEFIT will be incremented by the benefit of any sub-giv encountered. */
static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
static rtx sge_plus_constant PARAMS ((rtx, rtx));
static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
static rtx
simplify_giv_expr (loop, x, benefit)
const struct loop *loop;
rtx x;
int *benefit;
{
enum machine_mode mode = GET_MODE (x);
rtx arg0, arg1;
rtx tem;
/* If this is not an integer mode, or if we cannot do arithmetic in this
mode, this can't be a giv. */
if (mode != VOIDmode
&& (GET_MODE_CLASS (mode) != MODE_INT
|| GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
return NULL_RTX;
switch (GET_CODE (x))
{
case PLUS:
arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
if (arg0 == 0 || arg1 == 0)
return NULL_RTX;
/* Put constant last, CONST_INT last if both constant. */
if ((GET_CODE (arg0) == USE
|| GET_CODE (arg0) == CONST_INT)
&& ! ((GET_CODE (arg0) == USE
&& GET_CODE (arg1) == USE)
|| GET_CODE (arg1) == CONST_INT))
tem = arg0, arg0 = arg1, arg1 = tem;
/* Handle addition of zero, then addition of an invariant. */
if (arg1 == const0_rtx)
return arg0;
else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
switch (GET_CODE (arg0))
{
case CONST_INT:
case USE:
/* Adding two invariants must result in an invariant, so enclose
addition operation inside a USE and return it. */
if (GET_CODE (arg0) == USE)
arg0 = XEXP (arg0, 0);
if (GET_CODE (arg1) == USE)
arg1 = XEXP (arg1, 0);
if (GET_CODE (arg0) == CONST_INT)
tem = arg0, arg0 = arg1, arg1 = tem;
if (GET_CODE (arg1) == CONST_INT)
tem = sge_plus_constant (arg0, arg1);
else
tem = sge_plus (mode, arg0, arg1);
if (GET_CODE (tem) != CONST_INT)
tem = gen_rtx_USE (mode, tem);
return tem;
case REG:
case MULT:
/* biv + invar or mult + invar. Return sum. */
return gen_rtx_PLUS (mode, arg0, arg1);
case PLUS:
/* (a + invar_1) + invar_2. Associate. */
return
simplify_giv_expr (loop,
gen_rtx_PLUS (mode,
XEXP (arg0, 0),
gen_rtx_PLUS (mode,
XEXP (arg0, 1),
arg1)),
benefit);
default:
abort ();
}
/* Each argument must be either REG, PLUS, or MULT. Convert REG to
MULT to reduce cases. */
if (GET_CODE (arg0) == REG)
arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
if (GET_CODE (arg1) == REG)
arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
/* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
Recurse to associate the second PLUS. */
if (GET_CODE (arg1) == MULT)
tem = arg0, arg0 = arg1, arg1 = tem;
if (GET_CODE (arg1) == PLUS)
return
simplify_giv_expr (loop,
gen_rtx_PLUS (mode,
gen_rtx_PLUS (mode, arg0,
XEXP (arg1, 0)),
XEXP (arg1, 1)),
benefit);
/* Now must have MULT + MULT. Distribute if same biv, else not giv. */
if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
return NULL_RTX;
if (!rtx_equal_p (arg0, arg1))
return NULL_RTX;
return simplify_giv_expr (loop,
gen_rtx_MULT (mode,
XEXP (arg0, 0),
gen_rtx_PLUS (mode,
XEXP (arg0, 1),
XEXP (arg1, 1))),
benefit);
case MINUS:
/* Handle "a - b" as "a + b * (-1)". */
return simplify_giv_expr (loop,
gen_rtx_PLUS (mode,
XEXP (x, 0),
gen_rtx_MULT (mode,
XEXP (x, 1),
constm1_rtx)),
benefit);
case MULT:
arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
if (arg0 == 0 || arg1 == 0)
return NULL_RTX;
/* Put constant last, CONST_INT last if both constant. */
if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
&& GET_CODE (arg1) != CONST_INT)
tem = arg0, arg0 = arg1, arg1 = tem;
/* If second argument is not now constant, not giv. */
if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
return NULL_RTX;
/* Handle multiply by 0 or 1. */
if (arg1 == const0_rtx)
return const0_rtx;
else if (arg1 == const1_rtx)
return arg0;
switch (GET_CODE (arg0))
{
case REG:
/* biv * invar. Done. */
return gen_rtx_MULT (mode, arg0, arg1);
case CONST_INT:
/* Product of two constants. */
return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
case USE:
/* invar * invar. It is a giv, but very few of these will
actually pay off, so limit to simple registers. */
if (GET_CODE (arg1) != CONST_INT)
return NULL_RTX;
arg0 = XEXP (arg0, 0);
if (GET_CODE (arg0) == REG)
tem = gen_rtx_MULT (mode, arg0, arg1);
else if (GET_CODE (arg0) == MULT
&& GET_CODE (XEXP (arg0, 0)) == REG
&& GET_CODE (XEXP (arg0, 1)) == CONST_INT)
{
tem = gen_rtx_MULT (mode, XEXP (arg0, 0),
GEN_INT (INTVAL (XEXP (arg0, 1))
* INTVAL (arg1)));
}
else
return NULL_RTX;
return gen_rtx_USE (mode, tem);
case MULT:
/* (a * invar_1) * invar_2. Associate. */
return simplify_giv_expr (loop,
gen_rtx_MULT (mode,
XEXP (arg0, 0),
gen_rtx_MULT (mode,
XEXP (arg0, 1),
arg1)),
benefit);
case PLUS:
/* (a + invar_1) * invar_2. Distribute. */
return simplify_giv_expr (loop,
gen_rtx_PLUS (mode,
gen_rtx_MULT (mode,
XEXP (arg0, 0),
arg1),
gen_rtx_MULT (mode,
XEXP (arg0, 1),
arg1)),
benefit);
default:
abort ();
}
case ASHIFT:
/* Shift by constant is multiply by power of two. */
if (GET_CODE (XEXP (x, 1)) != CONST_INT)
return 0;
return
simplify_giv_expr (loop,
gen_rtx_MULT (mode,
XEXP (x, 0),
GEN_INT ((HOST_WIDE_INT) 1
<< INTVAL (XEXP (x, 1)))),
benefit);
case NEG:
/* "-a" is "a * (-1)" */
return simplify_giv_expr (loop,
gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
benefit);
case NOT:
/* "~a" is "-a - 1". Silly, but easy. */
return simplify_giv_expr (loop,
gen_rtx_MINUS (mode,
gen_rtx_NEG (mode, XEXP (x, 0)),
const1_rtx),
benefit);
case USE:
/* Already in proper form for invariant. */
return x;
case REG:
/* If this is a new register, we can't deal with it. */
if (REGNO (x) >= max_reg_before_loop)
return 0;
/* Check for biv or giv. */
switch (REG_IV_TYPE (REGNO (x)))
{
case BASIC_INDUCT:
return x;
case GENERAL_INDUCT:
{
struct induction *v = REG_IV_INFO (REGNO (x));
/* Form expression from giv and add benefit. Ensure this giv
can derive another and subtract any needed adjustment if so. */
*benefit += v->benefit;
if (v->cant_derive)
return 0;
tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
v->src_reg, v->mult_val),
v->add_val);
if (v->derive_adjustment)
tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
return simplify_giv_expr (loop, tem, benefit);
}
default:
/* If it isn't an induction variable, and it is invariant, we
may be able to simplify things further by looking through
the bits we just moved outside the loop. */
if (loop_invariant_p (loop, x) == 1)
{
struct movable *m;
for (m = the_movables; m ; m = m->next)
if (rtx_equal_p (x, m->set_dest))
{
/* Ok, we found a match. Substitute and simplify. */
/* If we match another movable, we must use that, as
this one is going away. */
if (m->match)
return simplify_giv_expr (loop, m->match->set_dest,
benefit);
/* If consec is non-zero, this is a member of a group of
instructions that were moved together. We handle this
case only to the point of seeking to the last insn and
looking for a REG_EQUAL. Fail if we don't find one. */
if (m->consec != 0)
{
int i = m->consec;
tem = m->insn;
do { tem = NEXT_INSN (tem); } while (--i > 0);
tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
if (tem)
tem = XEXP (tem, 0);
}
else
{
tem = single_set (m->insn);
if (tem)
tem = SET_SRC (tem);
}
if (tem)
{
/* What we are most interested in is pointer
arithmetic on invariants -- only take
patterns we may be able to do something with. */
if (GET_CODE (tem) == PLUS
|| GET_CODE (tem) == MULT
|| GET_CODE (tem) == ASHIFT
|| GET_CODE (tem) == CONST_INT
|| GET_CODE (tem) == SYMBOL_REF)
{
tem = simplify_giv_expr (loop, tem, benefit);
if (tem)
return tem;
}
else if (GET_CODE (tem) == CONST
&& GET_CODE (XEXP (tem, 0)) == PLUS
&& GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
&& GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
{
tem = simplify_giv_expr (loop, XEXP (tem, 0),
benefit);
if (tem)
return tem;
}
}
break;
}
}
break;
}
/* Fall through to general case. */
default:
/* If invariant, return as USE (unless CONST_INT).
Otherwise, not giv. */
if (GET_CODE (x) == USE)
x = XEXP (x, 0);
if (loop_invariant_p (loop, x) == 1)
{
if (GET_CODE (x) == CONST_INT)
return x;
if (GET_CODE (x) == CONST
&& GET_CODE (XEXP (x, 0)) == PLUS
&& GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
&& GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
x = XEXP (x, 0);
return gen_rtx_USE (mode, x);
}
else
return 0;
}
}
/* This routine folds invariants such that there is only ever one
CONST_INT in the summation. It is only used by simplify_giv_expr. */
static rtx
sge_plus_constant (x, c)
rtx x, c;
{
if (GET_CODE (x) == CONST_INT)
return GEN_INT (INTVAL (x) + INTVAL (c));
else if (GET_CODE (x) != PLUS)
return gen_rtx_PLUS (GET_MODE (x), x, c);
else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
{
return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
}
else if (GET_CODE (XEXP (x, 0)) == PLUS
|| GET_CODE (XEXP (x, 1)) != PLUS)
{
return gen_rtx_PLUS (GET_MODE (x),
sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
}
else
{
return gen_rtx_PLUS (GET_MODE (x),
sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
}
}
static rtx
sge_plus (mode, x, y)
enum machine_mode mode;
rtx x, y;
{
while (GET_CODE (y) == PLUS)
{
rtx a = XEXP (y, 0);
if (GET_CODE (a) == CONST_INT)
x = sge_plus_constant (x, a);
else
x = gen_rtx_PLUS (mode, x, a);
y = XEXP (y, 1);
}
if (GET_CODE (y) == CONST_INT)
x = sge_plus_constant (x, y);
else
x = gen_rtx_PLUS (mode, x, y);
return x;
}
/* Help detect a giv that is calculated by several consecutive insns;
for example,
giv = biv * M
giv = giv + A
The caller has already identified the first insn P as having a giv as dest;
we check that all other insns that set the same register follow
immediately after P, that they alter nothing else,
and that the result of the last is still a giv.
The value is 0 if the reg set in P is not really a giv.
Otherwise, the value is the amount gained by eliminating
all the consecutive insns that compute the value.
FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
The coefficients of the ultimate giv value are stored in
*MULT_VAL and *ADD_VAL. */
static int
consec_sets_giv (loop, first_benefit, p, src_reg, dest_reg,
add_val, mult_val, last_consec_insn)
const struct loop *loop;
int first_benefit;
rtx p;
rtx src_reg;
rtx dest_reg;
rtx *add_val;
rtx *mult_val;
rtx *last_consec_insn;
{
int count;
enum rtx_code code;
int benefit;
rtx temp;
rtx set;
/* Indicate that this is a giv so that we can update the value produced in
each insn of the multi-insn sequence.
This induction structure will be used only by the call to
general_induction_var below, so we can allocate it on our stack.
If this is a giv, our caller will replace the induct var entry with
a new induction structure. */
struct induction *v
= (struct induction *) alloca (sizeof (struct induction));
v->src_reg = src_reg;
v->mult_val = *mult_val;
v->add_val = *add_val;
v->benefit = first_benefit;
v->cant_derive = 0;
v->derive_adjustment = 0;
REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
REG_IV_INFO (REGNO (dest_reg)) = v;
count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
while (count > 0)
{
p = NEXT_INSN (p);
code = GET_CODE (p);
/* If libcall, skip to end of call sequence. */
if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
p = XEXP (temp, 0);
if (code == INSN
&& (set = single_set (p))
&& GET_CODE (SET_DEST (set)) == REG
&& SET_DEST (set) == dest_reg
&& (general_induction_var (loop, SET_SRC (set), &src_reg,
add_val, mult_val, 0, &benefit)
/* Giv created by equivalent expression. */
|| ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
&& general_induction_var (loop, XEXP (temp, 0), &src_reg,
add_val, mult_val, 0, &benefit)))
&& src_reg == v->src_reg)
{
if (find_reg_note (p, REG_RETVAL, NULL_RTX))
benefit += libcall_benefit (p);
count--;
v->mult_val = *mult_val;
v->add_val = *add_val;
v->benefit = benefit;
}
else if (code != NOTE)
{
/* Allow insns that set something other than this giv to a
constant. Such insns are needed on machines which cannot
include long constants and should not disqualify a giv. */
if (code == INSN
&& (set = single_set (p))
&& SET_DEST (set) != dest_reg
&& CONSTANT_P (SET_SRC (set)))
continue;
REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
return 0;
}
}
*last_consec_insn = p;
return v->benefit;
}
/* Return an rtx, if any, that expresses giv G2 as a function of the register
represented by G1. If no such expression can be found, or it is clear that
it cannot possibly be a valid address, 0 is returned.
To perform the computation, we note that
G1 = x * v + a and
G2 = y * v + b
where `v' is the biv.
So G2 = (y/b) * G1 + (b - a*y/x).
Note that MULT = y/x.
Update: A and B are now allowed to be additive expressions such that
B contains all variables in A. That is, computing B-A will not require
subtracting variables. */
static rtx
express_from_1 (a, b, mult)
rtx a, b, mult;
{
/* If MULT is zero, then A*MULT is zero, and our expression is B. */
if (mult == const0_rtx)
return b;
/* If MULT is not 1, we cannot handle A with non-constants, since we
would then be required to subtract multiples of the registers in A.
This is theoretically possible, and may even apply to some Fortran
constructs, but it is a lot of work and we do not attempt it here. */
if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
return NULL_RTX;
/* In general these structures are sorted top to bottom (down the PLUS
chain), but not left to right across the PLUS. If B is a higher
order giv than A, we can strip one level and recurse. If A is higher
order, we'll eventually bail out, but won't know that until the end.
If they are the same, we'll strip one level around this loop. */
while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
{
rtx ra, rb, oa, ob, tmp;
ra = XEXP (a, 0), oa = XEXP (a, 1);
if (GET_CODE (ra) == PLUS)
tmp = ra, ra = oa, oa = tmp;
rb = XEXP (b, 0), ob = XEXP (b, 1);
if (GET_CODE (rb) == PLUS)
tmp = rb, rb = ob, ob = tmp;
if (rtx_equal_p (ra, rb))
/* We matched: remove one reg completely. */
a = oa, b = ob;
else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
/* An alternate match. */
a = oa, b = rb;
else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
/* An alternate match. */
a = ra, b = ob;
else
{
/* Indicates an extra register in B. Strip one level from B and
recurse, hoping B was the higher order expression. */
ob = express_from_1 (a, ob, mult);
if (ob == NULL_RTX)
return NULL_RTX;
return gen_rtx_PLUS (GET_MODE (b), rb, ob);
}
}
/* Here we are at the last level of A, go through the cases hoping to
get rid of everything but a constant. */
if (GET_CODE (a) == PLUS)
{
rtx ra, oa;
ra = XEXP (a, 0), oa = XEXP (a, 1);
if (rtx_equal_p (oa, b))
oa = ra;
else if (!rtx_equal_p (ra, b))
return NULL_RTX;
if (GET_CODE (oa) != CONST_INT)
return NULL_RTX;
return GEN_INT (-INTVAL (oa) * INTVAL (mult));
}
else if (GET_CODE (a) == CONST_INT)
{
return plus_constant (b, -INTVAL (a) * INTVAL (mult));
}
else if (GET_CODE (b) == PLUS)
{
if (rtx_equal_p (a, XEXP (b, 0)))
return XEXP (b, 1);
else if (rtx_equal_p (a, XEXP (b, 1)))
return XEXP (b, 0);
else
return NULL_RTX;
}
else if (rtx_equal_p (a, b))
return const0_rtx;
return NULL_RTX;
}
rtx
express_from (g1, g2)
struct induction *g1, *g2;
{
rtx mult, add;
/* The value that G1 will be multiplied by must be a constant integer. Also,
the only chance we have of getting a valid address is if b*c/a (see above
for notation) is also an integer. */
if (GET_CODE (g1->mult_val) == CONST_INT
&& GET_CODE (g2->mult_val) == CONST_INT)
{
if (g1->mult_val == const0_rtx
|| INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
return NULL_RTX;
mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
}
else if (rtx_equal_p (g1->mult_val, g2->mult_val))
mult = const1_rtx;
else
{
/* ??? Find out if the one is a multiple of the other? */
return NULL_RTX;
}
add = express_from_1 (g1->add_val, g2->add_val, mult);
if (add == NULL_RTX)
{
/* Failed. If we've got a multiplication factor between G1 and G2,
scale G1's addend and try again. */
if (INTVAL (mult) > 1)
{
rtx g1_add_val = g1->add_val;
if (GET_CODE (g1_add_val) == MULT
&& GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
{
HOST_WIDE_INT m;
m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
XEXP (g1_add_val, 0), GEN_INT (m));
}
else
{
g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
mult);
}
add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
}
}
if (add == NULL_RTX)
return NULL_RTX;
/* Form simplified final result. */
if (mult == const0_rtx)
return add;
else if (mult == const1_rtx)
mult = g1->dest_reg;
else
mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
if (add == const0_rtx)
return mult;
else
{
if (GET_CODE (add) == PLUS
&& CONSTANT_P (XEXP (add, 1)))
{
rtx tem = XEXP (add, 1);
mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
add = tem;
}
return gen_rtx_PLUS (g2->mode, mult, add);
}
}
/* Return an rtx, if any, that expresses giv G2 as a function of the register
represented by G1. This indicates that G2 should be combined with G1 and
that G2 can use (either directly or via an address expression) a register
used to represent G1. */
static rtx
combine_givs_p (g1, g2)
struct induction *g1, *g2;
{
rtx tem = express_from (g1, g2);
/* If these givs are identical, they can be combined. We use the results
of express_from because the addends are not in a canonical form, so
rtx_equal_p is a weaker test. */
/* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
combination to be the other way round. */
if (tem == g1->dest_reg
&& (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
{
return g1->dest_reg;
}
/* If G2 can be expressed as a function of G1 and that function is valid
as an address and no more expensive than using a register for G2,
the expression of G2 in terms of G1 can be used. */
if (tem != NULL_RTX
&& g2->giv_type == DEST_ADDR
&& memory_address_p (g2->mem_mode, tem)
/* ??? Looses, especially with -fforce-addr, where *g2->location
will always be a register, and so anything more complicated
gets discarded. */
#if 0
#ifdef ADDRESS_COST
&& ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
#else
&& rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
#endif
#endif
)
{
return tem;
}
return NULL_RTX;
}
struct combine_givs_stats
{
int giv_number;
int total_benefit;
};
static int
cmp_combine_givs_stats (xp, yp)
const PTR xp;
const PTR yp;
{
const struct combine_givs_stats * const x =
(const struct combine_givs_stats *) xp;
const struct combine_givs_stats * const y =
(const struct combine_givs_stats *) yp;
int d;
d = y->total_benefit - x->total_benefit;
/* Stabilize the sort. */
if (!d)
d = x->giv_number - y->giv_number;
return d;
}
/* Check all pairs of givs for iv_class BL and see if any can be combined with
any other. If so, point SAME to the giv combined with and set NEW_REG to
be an expression (in terms of the other giv's DEST_REG) equivalent to the
giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
static void
combine_givs (bl)
struct iv_class *bl;
{
/* Additional benefit to add for being combined multiple times. */
const int extra_benefit = 3;
struct induction *g1, *g2, **giv_array;
int i, j, k, giv_count;
struct combine_givs_stats *stats;
rtx *can_combine;
/* Count givs, because bl->giv_count is incorrect here. */
giv_count = 0;
for (g1 = bl->giv; g1; g1 = g1->next_iv)
if (!g1->ignore)
giv_count++;
giv_array
= (struct induction **) alloca (giv_count * sizeof (struct induction *));
i = 0;
for (g1 = bl->giv; g1; g1 = g1->next_iv)
if (!g1->ignore)
giv_array[i++] = g1;
stats = (struct combine_givs_stats *) xcalloc (giv_count, sizeof (*stats));
can_combine = (rtx *) xcalloc (giv_count, giv_count * sizeof(rtx));
for (i = 0; i < giv_count; i++)
{
int this_benefit;
rtx single_use;
g1 = giv_array[i];
stats[i].giv_number = i;
/* If a DEST_REG GIV is used only once, do not allow it to combine
with anything, for in doing so we will gain nothing that cannot
be had by simply letting the GIV with which we would have combined
to be reduced on its own. The losage shows up in particular with
DEST_ADDR targets on hosts with reg+reg addressing, though it can
be seen elsewhere as well. */
if (g1->giv_type == DEST_REG
&& (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
&& single_use != const0_rtx)
continue;
this_benefit = g1->benefit;
/* Add an additional weight for zero addends. */
if (g1->no_const_addval)
this_benefit += 1;
for (j = 0; j < giv_count; j++)
{
rtx this_combine;
g2 = giv_array[j];
if (g1 != g2
&& (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
{
can_combine[i*giv_count + j] = this_combine;
this_benefit += g2->benefit + extra_benefit;
}
}
stats[i].total_benefit = this_benefit;
}
/* Iterate, combining until we can't. */
restart:
qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
if (loop_dump_stream)
{
fprintf (loop_dump_stream, "Sorted combine statistics:\n");
for (k = 0; k < giv_count; k++)
{
g1 = giv_array[stats[k].giv_number];
if (!g1->combined_with && !g1->same)
fprintf (loop_dump_stream, " {%d, %d}",
INSN_UID (giv_array[stats[k].giv_number]->insn),
stats[k].total_benefit);
}
putc ('\n', loop_dump_stream);
}
for (k = 0; k < giv_count; k++)
{
int g1_add_benefit = 0;
i = stats[k].giv_number;
g1 = giv_array[i];
/* If it has already been combined, skip. */
if (g1->combined_with || g1->same)
continue;
for (j = 0; j < giv_count; j++)
{
g2 = giv_array[j];
if (g1 != g2 && can_combine[i*giv_count + j]
/* If it has already been combined, skip. */
&& ! g2->same && ! g2->combined_with)
{
int l;
g2->new_reg = can_combine[i*giv_count + j];
g2->same = g1;
g1->combined_with++;
g1->lifetime += g2->lifetime;
g1_add_benefit += g2->benefit;
/* ??? The new final_[bg]iv_value code does a much better job
of finding replaceable giv's, and hence this code may no
longer be necessary. */
if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
g1_add_benefit -= copy_cost;
/* To help optimize the next set of combinations, remove
this giv from the benefits of other potential mates. */
for (l = 0; l < giv_count; ++l)
{
int m = stats[l].giv_number;
if (can_combine[m*giv_count + j])
stats[l].total_benefit -= g2->benefit + extra_benefit;
}
if (loop_dump_stream)
fprintf (loop_dump_stream,
"giv at %d combined with giv at %d\n",
INSN_UID (g2->insn), INSN_UID (g1->insn));
}
}
/* To help optimize the next set of combinations, remove
this giv from the benefits of other potential mates. */
if (g1->combined_with)
{
for (j = 0; j < giv_count; ++j)
{
int m = stats[j].giv_number;
if (can_combine[m*giv_count + i])
stats[j].total_benefit -= g1->benefit + extra_benefit;
}
g1->benefit += g1_add_benefit;
/* We've finished with this giv, and everything it touched.
Restart the combination so that proper weights for the
rest of the givs are properly taken into account. */
/* ??? Ideally we would compact the arrays at this point, so
as to not cover old ground. But sanely compacting
can_combine is tricky. */
goto restart;
}
}
/* Clean up. */
free (stats);
free (can_combine);
}
struct recombine_givs_stats
{
int giv_number;
int start_luid, end_luid;
};
/* Used below as comparison function for qsort. We want a ascending luid
when scanning the array starting at the end, thus the arguments are
used in reverse. */
static int
cmp_recombine_givs_stats (xp, yp)
const PTR xp;
const PTR yp;
{
const struct recombine_givs_stats * const x =
(const struct recombine_givs_stats *) xp;
const struct recombine_givs_stats * const y =
(const struct recombine_givs_stats *) yp;
int d;
d = y->start_luid - x->start_luid;
/* Stabilize the sort. */
if (!d)
d = y->giv_number - x->giv_number;
return d;
}
/* Scan X, which is a part of INSN, for the end of life of a giv. Also
look for the start of life of a giv where the start has not been seen
yet to unlock the search for the end of its life.
Only consider givs that belong to BIV.
Return the total number of lifetime ends that have been found. */
static int
find_life_end (x, stats, insn, biv)
rtx x, insn, biv;
struct recombine_givs_stats *stats;
{
enum rtx_code code;
const char *fmt;
int i, j;
int retval;
code = GET_CODE (x);
switch (code)
{
case SET:
{
rtx reg = SET_DEST (x);
if (GET_CODE (reg) == REG)
{
int regno = REGNO (reg);
struct induction *v = REG_IV_INFO (regno);
if (REG_IV_TYPE (regno) == GENERAL_INDUCT
&& ! v->ignore
&& v->src_reg == biv
&& stats[v->ix].end_luid <= 0)
{
/* If we see a 0 here for end_luid, it means that we have
scanned the entire loop without finding any use at all.
We must not predicate this code on a start_luid match
since that would make the test fail for givs that have
been hoisted out of inner loops. */
if (stats[v->ix].end_luid == 0)
{
stats[v->ix].end_luid = stats[v->ix].start_luid;
return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
}
else if (stats[v->ix].start_luid == INSN_LUID (insn))
stats[v->ix].end_luid = 0;
}
return find_life_end (SET_SRC (x), stats, insn, biv);
}
break;
}
case REG:
{
int regno = REGNO (x);
struct induction *v = REG_IV_INFO (regno);
if (REG_IV_TYPE (regno) == GENERAL_INDUCT
&& ! v->ignore
&& v->src_reg == biv
&& stats[v->ix].end_luid == 0)
{
while (INSN_UID (insn) >= max_uid_for_loop)
insn = NEXT_INSN (insn);
stats[v->ix].end_luid = INSN_LUID (insn);
return 1;
}
return 0;
}
case LABEL_REF:
case CONST_DOUBLE:
case CONST_INT:
case CONST:
return 0;
default:
break;
}
fmt = GET_RTX_FORMAT (code);
retval = 0;
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
retval += find_life_end (XEXP (x, i), stats, insn, biv);
else if (fmt[i] == 'E')
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
}
return retval;
}
/* For each giv that has been combined with another, look if
we can combine it with the most recently used one instead.
This tends to shorten giv lifetimes, and helps the next step:
try to derive givs from other givs. */
static void
recombine_givs (loop, bl, unroll_p)
const struct loop *loop;
struct iv_class *bl;
int unroll_p;
{
struct induction *v, **giv_array, *last_giv;
struct recombine_givs_stats *stats;
int giv_count;
int i, rescan;
int ends_need_computing;
for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
{
if (! v->ignore)
giv_count++;
}
giv_array
= (struct induction **) xmalloc (giv_count * sizeof (struct induction *));
stats = (struct recombine_givs_stats *) xmalloc (giv_count * sizeof *stats);
/* Initialize stats and set up the ix field for each giv in stats to name
the corresponding index into stats. */
for (i = 0, v = bl->giv; v; v = v->next_iv)
{
rtx p;
if (v->ignore)
continue;
giv_array[i] = v;
stats[i].giv_number = i;
/* If this giv has been hoisted out of an inner loop, use the luid of
the previous insn. */
for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
p = PREV_INSN (p);
stats[i].start_luid = INSN_LUID (p);
i++;
}
qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
/* Set up the ix field for each giv in stats to name
the corresponding index into stats, and
do the actual most-recently-used recombination. */
for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
{
v = giv_array[stats[i].giv_number];
v->ix = i;
if (v->same)
{
struct induction *old_same = v->same;
rtx new_combine;
/* combine_givs_p actually says if we can make this transformation.
The other tests are here only to avoid keeping a giv alive
that could otherwise be eliminated. */
if (last_giv
&& ((old_same->maybe_dead && ! old_same->combined_with)
|| ! last_giv->maybe_dead
|| last_giv->combined_with)
&& (new_combine = combine_givs_p (last_giv, v)))
{
old_same->combined_with--;
v->new_reg = new_combine;
v->same = last_giv;
last_giv->combined_with++;
/* No need to update lifetimes / benefits here since we have
already decided what to reduce. */
if (loop_dump_stream)
{
fprintf (loop_dump_stream,
"giv at %d recombined with giv at %d as ",
INSN_UID (v->insn), INSN_UID (last_giv->insn));
print_rtl (loop_dump_stream, v->new_reg);
putc ('\n', loop_dump_stream);
}
continue;
}
v = v->same;
}
else if (v->giv_type != DEST_REG)
continue;
if (! last_giv
|| (last_giv->maybe_dead && ! last_giv->combined_with)
|| ! v->maybe_dead
|| v->combined_with)
last_giv = v;
}
ends_need_computing = 0;
/* For each DEST_REG giv, compute lifetime starts, and try to compute
lifetime ends from regscan info. */
for (i = giv_count - 1; i >= 0; i--)
{
v = giv_array[stats[i].giv_number];
if (v->ignore)
continue;
if (v->giv_type == DEST_ADDR)
{
/* Loop unrolling of an inner loop can even create new DEST_REG
givs. */
rtx p;
for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
p = PREV_INSN (p);
stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
if (p != v->insn)
stats[i].end_luid++;
}
else /* v->giv_type == DEST_REG */
{
if (v->last_use)
{
stats[i].start_luid = INSN_LUID (v->insn);
stats[i].end_luid = INSN_LUID (v->last_use);
}
else if (INSN_UID (v->insn) >= max_uid_for_loop)
{
rtx p;
/* This insn has been created by loop optimization on an inner
loop. We don't have a proper start_luid that will match
when we see the first set. But we do know that there will
be no use before the set, so we can set end_luid to 0 so that
we'll start looking for the last use right away. */
for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
p = PREV_INSN (p);
stats[i].start_luid = INSN_LUID (p);
stats[i].end_luid = 0;
ends_need_computing++;
}
else
{
int regno = REGNO (v->dest_reg);
int count = VARRAY_INT (n_times_set, regno) - 1;
rtx p = v->insn;
/* Find the first insn that sets the giv, so that we can verify
if this giv's lifetime wraps around the loop. We also need
the luid of the first setting insn in order to detect the
last use properly. */
while (count)
{
p = prev_nonnote_insn (p);
if (reg_set_p (v->dest_reg, p))
count--;
}
stats[i].start_luid = INSN_LUID (p);
if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
{
stats[i].end_luid = -1;
ends_need_computing++;
}
else
{
stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
if (stats[i].end_luid > INSN_LUID (loop->end))
{
stats[i].end_luid = -1;
ends_need_computing++;
}
}
}
}
}
/* If the regscan information was unconclusive for one or more DEST_REG
givs, scan the all insn in the loop to find out lifetime ends. */
if (ends_need_computing)
{
rtx biv = bl->biv->src_reg;
rtx p = loop->end;
do
{
if (p == loop->start)
p = loop->end;
p = PREV_INSN (p);
if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
continue;
ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
}
while (ends_need_computing);
}
/* Set start_luid back to the last insn that sets the giv. This allows
more combinations. */
for (i = giv_count - 1; i >= 0; i--)
{
v = giv_array[stats[i].giv_number];
if (v->ignore)
continue;
if (INSN_UID (v->insn) < max_uid_for_loop)
stats[i].start_luid = INSN_LUID (v->insn);
}
/* Now adjust lifetime ends by taking combined givs into account. */
for (i = giv_count - 1; i >= 0; i--)
{
unsigned luid;
int j;
v = giv_array[stats[i].giv_number];
if (v->ignore)
continue;
if (v->same && ! v->same->ignore)
{
j = v->same->ix;
luid = stats[i].start_luid;
/* Use unsigned arithmetic to model loop wrap-around. */
if (luid - stats[j].start_luid
> (unsigned) stats[j].end_luid - stats[j].start_luid)
stats[j].end_luid = luid;
}
}
qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
/* Try to derive DEST_REG givs from previous DEST_REG givs with the
same mult_val and non-overlapping lifetime. This reduces register
pressure.
Once we find a DEST_REG giv that is suitable to derive others from,
we set last_giv to this giv, and try to derive as many other DEST_REG
givs from it without joining overlapping lifetimes. If we then
encounter a DEST_REG giv that we can't derive, we set rescan to the
index for this giv (unless rescan is already set).
When we are finished with the current LAST_GIV (i.e. the inner loop
terminates), we start again with rescan, which then becomes the new
LAST_GIV. */
for (i = giv_count - 1; i >= 0; i = rescan)
{
int life_start = 0, life_end = 0;
for (last_giv = 0, rescan = -1; i >= 0; i--)
{
rtx sum;
v = giv_array[stats[i].giv_number];
if (v->giv_type != DEST_REG || v->derived_from || v->same)
continue;
if (! last_giv)
{
/* Don't use a giv that's likely to be dead to derive
others - that would be likely to keep that giv alive. */
if (! v->maybe_dead || v->combined_with)
{
last_giv = v;
life_start = stats[i].start_luid;
life_end = stats[i].end_luid;
}
continue;
}
/* Use unsigned arithmetic to model loop wrap around. */
if (((unsigned) stats[i].start_luid - life_start
>= (unsigned) life_end - life_start)
&& ((unsigned) stats[i].end_luid - life_start
> (unsigned) life_end - life_start)
/* Check that the giv insn we're about to use for deriving
precedes all uses of that giv. Note that initializing the
derived giv would defeat the purpose of reducing register
pressure.
??? We could arrange to move the insn. */
&& ((unsigned) stats[i].end_luid - INSN_LUID (loop->start)
> (unsigned) stats[i].start_luid - INSN_LUID (loop->start))
&& rtx_equal_p (last_giv->mult_val, v->mult_val)
/* ??? Could handle libcalls, but would need more logic. */
&& ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
/* We would really like to know if for any giv that v
is combined with, v->insn or any intervening biv increment
dominates that combined giv. However, we
don't have this detailed control flow information.
N.B. since last_giv will be reduced, it is valid
anywhere in the loop, so we don't need to check the
validity of last_giv.
We rely here on the fact that v->always_executed implies that
there is no jump to someplace else in the loop before the
giv insn, and hence any insn that is executed before the
giv insn in the loop will have a lower luid. */
&& (v->always_executed || ! v->combined_with)
&& (sum = express_from (last_giv, v))
/* Make sure we don't make the add more expensive. ADD_COST
doesn't take different costs of registers and constants into
account, so compare the cost of the actual SET_SRCs. */
&& (rtx_cost (sum, SET)
<= rtx_cost (SET_SRC (single_set (v->insn)), SET))
/* ??? unroll can't understand anything but reg + const_int
sums. It would be cleaner to fix unroll. */
&& ((GET_CODE (sum) == PLUS
&& GET_CODE (XEXP (sum, 0)) == REG
&& GET_CODE (XEXP (sum, 1)) == CONST_INT)
|| ! unroll_p)
&& validate_change (v->insn, &PATTERN (v->insn),
gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
{
v->derived_from = last_giv;
life_end = stats[i].end_luid;
if (loop_dump_stream)
{
fprintf (loop_dump_stream,
"giv at %d derived from %d as ",
INSN_UID (v->insn), INSN_UID (last_giv->insn));
print_rtl (loop_dump_stream, sum);
putc ('\n', loop_dump_stream);
}
}
else if (rescan < 0)
rescan = i;
}
}
/* Clean up. */
free (giv_array);
free (stats);
}
/* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
void
emit_iv_add_mult (b, m, a, reg, insert_before)
rtx b; /* initial value of basic induction variable */
rtx m; /* multiplicative constant */
rtx a; /* additive constant */
rtx reg; /* destination register */
rtx insert_before;
{
rtx seq;
rtx result;
/* Prevent unexpected sharing of these rtx. */
a = copy_rtx (a);
b = copy_rtx (b);
/* Increase the lifetime of any invariants moved further in code. */
update_reg_last_use (a, insert_before);
update_reg_last_use (b, insert_before);
update_reg_last_use (m, insert_before);
start_sequence ();
result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
if (reg != result)
emit_move_insn (reg, result);
seq = gen_sequence ();
end_sequence ();
emit_insn_before (seq, insert_before);
/* It is entirely possible that the expansion created lots of new
registers. Iterate over the sequence we just created and
record them all. */
if (GET_CODE (seq) == SEQUENCE)
{
int i;
for (i = 0; i < XVECLEN (seq, 0); ++i)
{
rtx set = single_set (XVECEXP (seq, 0, i));
if (set && GET_CODE (SET_DEST (set)) == REG)
record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
}
}
else if (GET_CODE (seq) == SET
&& GET_CODE (SET_DEST (seq)) == REG)
record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
}
/* Test whether A * B can be computed without
an actual multiply insn. Value is 1 if so. */
static int
product_cheap_p (a, b)
rtx a;
rtx b;
{
int i;
rtx tmp;
struct obstack *old_rtl_obstack = rtl_obstack;
char *storage = (char *) obstack_alloc (&temp_obstack, 0);
int win = 1;
/* If only one is constant, make it B. */
if (GET_CODE (a) == CONST_INT)
tmp = a, a = b, b = tmp;
/* If first constant, both constant, so don't need multiply. */
if (GET_CODE (a) == CONST_INT)
return 1;
/* If second not constant, neither is constant, so would need multiply. */
if (GET_CODE (b) != CONST_INT)
return 0;
/* One operand is constant, so might not need multiply insn. Generate the
code for the multiply and see if a call or multiply, or long sequence
of insns is generated. */
rtl_obstack = &temp_obstack;
start_sequence ();
expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
tmp = gen_sequence ();
end_sequence ();
if (GET_CODE (tmp) == SEQUENCE)
{
if (XVEC (tmp, 0) == 0)
win = 1;
else if (XVECLEN (tmp, 0) > 3)
win = 0;
else
for (i = 0; i < XVECLEN (tmp, 0); i++)
{
rtx insn = XVECEXP (tmp, 0, i);
if (GET_CODE (insn) != INSN
|| (GET_CODE (PATTERN (insn)) == SET
&& GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
|| (GET_CODE (PATTERN (insn)) == PARALLEL
&& GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
&& GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
{
win = 0;
break;
}
}
}
else if (GET_CODE (tmp) == SET
&& GET_CODE (SET_SRC (tmp)) == MULT)
win = 0;
else if (GET_CODE (tmp) == PARALLEL
&& GET_CODE (XVECEXP (tmp, 0, 0)) == SET
&& GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
win = 0;
/* Free any storage we obtained in generating this multiply and restore rtl
allocation to its normal obstack. */
obstack_free (&temp_obstack, storage);
rtl_obstack = old_rtl_obstack;
return win;
}
/* Check to see if loop can be terminated by a "decrement and branch until
zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
Also try reversing an increment loop to a decrement loop
to see if the optimization can be performed.
Value is nonzero if optimization was performed. */
/* This is useful even if the architecture doesn't have such an insn,
because it might change a loops which increments from 0 to n to a loop
which decrements from n to 0. A loop that decrements to zero is usually
faster than one that increments from zero. */
/* ??? This could be rewritten to use some of the loop unrolling procedures,
such as approx_final_value, biv_total_increment, loop_iterations, and
final_[bg]iv_value. */
static int
check_dbra_loop (loop, insn_count)
struct loop *loop;
int insn_count;
{
struct iv_class *bl;
rtx reg;
rtx jump_label;
rtx final_value;
rtx start_value;
rtx new_add_val;
rtx comparison;
rtx before_comparison;
rtx p;
rtx jump;
rtx first_compare;
int compare_and_branch;
rtx loop_start = loop->start;
rtx loop_end = loop->end;
struct loop_info *loop_info = LOOP_INFO (loop);
/* If last insn is a conditional branch, and the insn before tests a
register value, try to optimize it. Otherwise, we can't do anything. */
jump = PREV_INSN (loop_end);
comparison = get_condition_for_loop (loop, jump);
if (comparison == 0)
return 0;
/* Try to compute whether the compare/branch at the loop end is one or
two instructions. */
get_condition (jump, &first_compare);
if (first_compare == jump)
compare_and_branch = 1;
else if (first_compare == prev_nonnote_insn (jump))
compare_and_branch = 2;
else
return 0;
/* Check all of the bivs to see if the compare uses one of them.
Skip biv's set more than once because we can't guarantee that
it will be zero on the last iteration. Also skip if the biv is
used between its update and the test insn. */
for (bl = loop_iv_list; bl; bl = bl->next)
{
if (bl->biv_count == 1
&& ! bl->biv->maybe_multiple
&& bl->biv->dest_reg == XEXP (comparison, 0)
&& ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
first_compare))
break;
}
if (! bl)
return 0;
/* Look for the case where the basic induction variable is always
nonnegative, and equals zero on the last iteration.
In this case, add a reg_note REG_NONNEG, which allows the
m68k DBRA instruction to be used. */
if (((GET_CODE (comparison) == GT
&& GET_CODE (XEXP (comparison, 1)) == CONST_INT
&& INTVAL (XEXP (comparison, 1)) == -1)
|| (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
&& GET_CODE (bl->biv->add_val) == CONST_INT
&& INTVAL (bl->biv->add_val) < 0)
{
/* Initial value must be greater than 0,
init_val % -dec_value == 0 to ensure that it equals zero on
the last iteration */
if (GET_CODE (bl->initial_value) == CONST_INT
&& INTVAL (bl->initial_value) > 0
&& (INTVAL (bl->initial_value)
% (-INTVAL (bl->biv->add_val))) == 0)
{
/* register always nonnegative, add REG_NOTE to branch */
REG_NOTES (PREV_INSN (loop_end))
= gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
REG_NOTES (PREV_INSN (loop_end)));
bl->nonneg = 1;
return 1;
}
/* If the decrement is 1 and the value was tested as >= 0 before
the loop, then we can safely optimize. */
for (p = loop_start; p; p = PREV_INSN (p))
{
if (GET_CODE (p) == CODE_LABEL)
break;
if (GET_CODE (p) != JUMP_INSN)
continue;
before_comparison = get_condition_for_loop (loop, p);
if (before_comparison
&& XEXP (before_comparison, 0) == bl->biv->dest_reg
&& GET_CODE (before_comparison) == LT
&& XEXP (before_comparison, 1) == const0_rtx
&& ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
&& INTVAL (bl->biv->add_val) == -1)
{
REG_NOTES (PREV_INSN (loop_end))
= gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
REG_NOTES (PREV_INSN (loop_end)));
bl->nonneg = 1;
return 1;
}
}
}
else if (GET_CODE (bl->biv->add_val) == CONST_INT
&& INTVAL (bl->biv->add_val) > 0)
{
/* Try to change inc to dec, so can apply above optimization. */
/* Can do this if:
all registers modified are induction variables or invariant,
all memory references have non-overlapping addresses
(obviously true if only one write)
allow 2 insns for the compare/jump at the end of the loop. */
/* Also, we must avoid any instructions which use both the reversed
biv and another biv. Such instructions will fail if the loop is
reversed. We meet this condition by requiring that either
no_use_except_counting is true, or else that there is only
one biv. */
int num_nonfixed_reads = 0;
/* 1 if the iteration var is used only to count iterations. */
int no_use_except_counting = 0;
/* 1 if the loop has no memory store, or it has a single memory store
which is reversible. */
int reversible_mem_store = 1;
if (bl->giv_count == 0 && ! loop->exit_count)
{
rtx bivreg = regno_reg_rtx[bl->regno];
/* If there are no givs for this biv, and the only exit is the
fall through at the end of the loop, then
see if perhaps there are no uses except to count. */
no_use_except_counting = 1;
for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
{
rtx set = single_set (p);
if (set && GET_CODE (SET_DEST (set)) == REG
&& REGNO (SET_DEST (set)) == bl->regno)
/* An insn that sets the biv is okay. */
;
else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
|| p == prev_nonnote_insn (loop_end))
&& reg_mentioned_p (bivreg, PATTERN (p)))
{
/* If either of these insns uses the biv and sets a pseudo
that has more than one usage, then the biv has uses
other than counting since it's used to derive a value
that is used more than one time. */
int note_set_pseudo_multiple_uses_retval = 0;
note_stores (PATTERN (p), note_set_pseudo_multiple_uses,
¬e_set_pseudo_multiple_uses_retval);
if (note_set_pseudo_multiple_uses_retval)
{
no_use_except_counting = 0;
break;
}
}
else if (reg_mentioned_p (bivreg, PATTERN (p)))
{
no_use_except_counting = 0;
break;
}
}
}
if (no_use_except_counting)
; /* no need to worry about MEMs. */
else if (num_mem_sets <= 1)
{
for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
/* If the loop has a single store, and the destination address is
invariant, then we can't reverse the loop, because this address
might then have the wrong value at loop exit.
This would work if the source was invariant also, however, in that
case, the insn should have been moved out of the loop. */
if (num_mem_sets == 1)
{
struct induction *v;
reversible_mem_store
= (! unknown_address_altered
&& ! unknown_constant_address_altered
&& ! loop_invariant_p (loop,
XEXP (XEXP (loop_store_mems, 0),
0)));
/* If the store depends on a register that is set after the
store, it depends on the initial value, and is thus not
reversible. */
for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
{
if (v->giv_type == DEST_REG
&& reg_mentioned_p (v->dest_reg,
PATTERN (first_loop_store_insn))
&& loop_insn_first_p (first_loop_store_insn, v->insn))
reversible_mem_store = 0;
}
}
}
else
return 0;
/* This code only acts for innermost loops. Also it simplifies
the memory address check by only reversing loops with
zero or one memory access.
Two memory accesses could involve parts of the same array,
and that can't be reversed.
If the biv is used only for counting, than we don't need to worry
about all these things. */
if ((num_nonfixed_reads <= 1
&& ! loop_info->has_call
&& ! loop_info->has_volatile
&& reversible_mem_store
&& (bl->giv_count + bl->biv_count + num_mem_sets
+ num_movables + compare_and_branch == insn_count)
&& (bl == loop_iv_list && bl->next == 0))
|| no_use_except_counting)
{
rtx tem;
/* Loop can be reversed. */
if (loop_dump_stream)
fprintf (loop_dump_stream, "Can reverse loop\n");
/* Now check other conditions:
The increment must be a constant, as must the initial value,
and the comparison code must be LT.
This test can probably be improved since +/- 1 in the constant
can be obtained by changing LT to LE and vice versa; this is
confusing. */
if (comparison
/* for constants, LE gets turned into LT */
&& (GET_CODE (comparison) == LT
|| (GET_CODE (comparison) == LE
&& no_use_except_counting)))
{
HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
rtx initial_value, comparison_value;
int nonneg = 0;
enum rtx_code cmp_code;
int comparison_const_width;
unsigned HOST_WIDE_INT comparison_sign_mask;
add_val = INTVAL (bl->biv->add_val);
comparison_value = XEXP (comparison, 1);
if (GET_MODE (comparison_value) == VOIDmode)
comparison_const_width
= GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
else
comparison_const_width
= GET_MODE_BITSIZE (GET_MODE (comparison_value));
if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
comparison_const_width = HOST_BITS_PER_WIDE_INT;
comparison_sign_mask
= (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
/* If the comparison value is not a loop invariant, then we
can not reverse this loop.
??? If the insns which initialize the comparison value as
a whole compute an invariant result, then we could move
them out of the loop and proceed with loop reversal. */
if (! loop_invariant_p (loop, comparison_value))
return 0;
if (GET_CODE (comparison_value) == CONST_INT)
comparison_val = INTVAL (comparison_value);
initial_value = bl->initial_value;
/* Normalize the initial value if it is an integer and
has no other use except as a counter. This will allow
a few more loops to be reversed. */
if (no_use_except_counting
&& GET_CODE (comparison_value) == CONST_INT
&& GET_CODE (initial_value) == CONST_INT)
{
comparison_val = comparison_val - INTVAL (bl->initial_value);
/* The code below requires comparison_val to be a multiple
of add_val in order to do the loop reversal, so
round up comparison_val to a multiple of add_val.
Since comparison_value is constant, we know that the
current comparison code is LT. */
comparison_val = comparison_val + add_val - 1;
comparison_val
-= (unsigned HOST_WIDE_INT) comparison_val % add_val;
/* We postpone overflow checks for COMPARISON_VAL here;
even if there is an overflow, we might still be able to
reverse the loop, if converting the loop exit test to
NE is possible. */
initial_value = const0_rtx;
}
/* First check if we can do a vanilla loop reversal. */
if (initial_value == const0_rtx
/* If we have a decrement_and_branch_on_count,
prefer the NE test, since this will allow that
instruction to be generated. Note that we must
use a vanilla loop reversal if the biv is used to
calculate a giv or has a non-counting use. */
#if ! defined (HAVE_decrement_and_branch_until_zero) \
&& defined (HAVE_decrement_and_branch_on_count)
&& (! (add_val == 1 && loop->vtop
&& (bl->biv_count == 0
|| no_use_except_counting)))
#endif
&& GET_CODE (comparison_value) == CONST_INT
/* Now do postponed overflow checks on COMPARISON_VAL. */
&& ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
& comparison_sign_mask))
{
/* Register will always be nonnegative, with value
0 on last iteration */
add_adjust = add_val;
nonneg = 1;
cmp_code = GE;
}
else if (add_val == 1 && loop->vtop
&& (bl->biv_count == 0
|| no_use_except_counting))
{
add_adjust = 0;
cmp_code = NE;
}
else
return 0;
if (GET_CODE (comparison) == LE)
add_adjust -= add_val;
/* If the initial value is not zero, or if the comparison
value is not an exact multiple of the increment, then we
can not reverse this loop. */
if (initial_value == const0_rtx
&& GET_CODE (comparison_value) == CONST_INT)
{
if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
return 0;
}
else
{
if (! no_use_except_counting || add_val != 1)
return 0;
}
final_value = comparison_value;
/* Reset these in case we normalized the initial value
and comparison value above. */
if (GET_CODE (comparison_value) == CONST_INT
&& GET_CODE (initial_value) == CONST_INT)
{
comparison_value = GEN_INT (comparison_val);
final_value
= GEN_INT (comparison_val + INTVAL (bl->initial_value));
}
bl->initial_value = initial_value;
/* Save some info needed to produce the new insns. */
reg = bl->biv->dest_reg;
jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
if (jump_label == pc_rtx)
jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
/* Set start_value; if this is not a CONST_INT, we need
to generate a SUB.
Initialize biv to start_value before loop start.
The old initializing insn will be deleted as a
dead store by flow.c. */
if (initial_value == const0_rtx
&& GET_CODE (comparison_value) == CONST_INT)
{
start_value = GEN_INT (comparison_val - add_adjust);
emit_insn_before (gen_move_insn (reg, start_value),
loop_start);
}
else if (GET_CODE (initial_value) == CONST_INT)
{
rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
enum machine_mode mode = GET_MODE (reg);
enum insn_code icode
= add_optab->handlers[(int) mode].insn_code;
if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
|| ! ((*insn_data[icode].operand[1].predicate)
(comparison_value, mode))
|| ! ((*insn_data[icode].operand[2].predicate)
(offset, mode)))
return 0;
start_value
= gen_rtx_PLUS (mode, comparison_value, offset);
emit_insn_before ((GEN_FCN (icode)
(reg, comparison_value, offset)),
loop_start);
if (GET_CODE (comparison) == LE)
final_value = gen_rtx_PLUS (mode, comparison_value,
GEN_INT (add_val));
}
else if (! add_adjust)
{
enum machine_mode mode = GET_MODE (reg);
enum insn_code icode
= sub_optab->handlers[(int) mode].insn_code;
if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
|| ! ((*insn_data[icode].operand[1].predicate)
(comparison_value, mode))
|| ! ((*insn_data[icode].operand[2].predicate)
(initial_value, mode)))
return 0;
start_value
= gen_rtx_MINUS (mode, comparison_value, initial_value);
emit_insn_before ((GEN_FCN (icode)
(reg, comparison_value, initial_value)),
loop_start);
}
else
/* We could handle the other cases too, but it'll be
better to have a testcase first. */
return 0;
/* We may not have a single insn which can increment a reg, so
create a sequence to hold all the insns from expand_inc. */
start_sequence ();
expand_inc (reg, new_add_val);
tem = gen_sequence ();
end_sequence ();
p = emit_insn_before (tem, bl->biv->insn);
delete_insn (bl->biv->insn);
/* Update biv info to reflect its new status. */
bl->biv->insn = p;
bl->initial_value = start_value;
bl->biv->add_val = new_add_val;
/* Update loop info. */
loop_info->initial_value = reg;
loop_info->initial_equiv_value = reg;
loop_info->final_value = const0_rtx;
loop_info->final_equiv_value = const0_rtx;
loop_info->comparison_value = const0_rtx;
loop_info->comparison_code = cmp_code;
loop_info->increment = new_add_val;
/* Inc LABEL_NUSES so that delete_insn will
not delete the label. */
LABEL_NUSES (XEXP (jump_label, 0)) ++;
/* Emit an insn after the end of the loop to set the biv's
proper exit value if it is used anywhere outside the loop. */
if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
|| ! bl->init_insn
|| REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
emit_insn_after (gen_move_insn (reg, final_value),
loop_end);
/* Delete compare/branch at end of loop. */
delete_insn (PREV_INSN (loop_end));
if (compare_and_branch == 2)
delete_insn (first_compare);
/* Add new compare/branch insn at end of loop. */
start_sequence ();
emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
GET_MODE (reg), 0, 0,
XEXP (jump_label, 0));
tem = gen_sequence ();
end_sequence ();
emit_jump_insn_before (tem, loop_end);
for (tem = PREV_INSN (loop_end);
tem && GET_CODE (tem) != JUMP_INSN;
tem = PREV_INSN (tem))
;
if (tem)
JUMP_LABEL (tem) = XEXP (jump_label, 0);
if (nonneg)
{
if (tem)
{
/* Increment of LABEL_NUSES done above. */
/* Register is now always nonnegative,
so add REG_NONNEG note to the branch. */
REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
REG_NOTES (tem));
}
bl->nonneg = 1;
}
/* No insn may reference both the reversed and another biv or it
will fail (see comment near the top of the loop reversal
code).
Earlier on, we have verified that the biv has no use except
counting, or it is the only biv in this function.
However, the code that computes no_use_except_counting does
not verify reg notes. It's possible to have an insn that
references another biv, and has a REG_EQUAL note with an
expression based on the reversed biv. To avoid this case,
remove all REG_EQUAL notes based on the reversed biv
here. */
for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
{
rtx *pnote;
rtx set = single_set (p);
/* If this is a set of a GIV based on the reversed biv, any
REG_EQUAL notes should still be correct. */
if (! set
|| GET_CODE (SET_DEST (set)) != REG
|| (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
|| REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
|| REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
for (pnote = ®_NOTES (p); *pnote;)
{
if (REG_NOTE_KIND (*pnote) == REG_EQUAL
&& reg_mentioned_p (regno_reg_rtx[bl->regno],
XEXP (*pnote, 0)))
*pnote = XEXP (*pnote, 1);
else
pnote = &XEXP (*pnote, 1);
}
}
/* Mark that this biv has been reversed. Each giv which depends
on this biv, and which is also live past the end of the loop
will have to be fixed up. */
bl->reversed = 1;
if (loop_dump_stream)
{
fprintf (loop_dump_stream, "Reversed loop");
if (bl->nonneg)
fprintf (loop_dump_stream, " and added reg_nonneg\n");
else
fprintf (loop_dump_stream, "\n");
}
return 1;
}
}
}
return 0;
}
/* Verify whether the biv BL appears to be eliminable,
based on the insns in the loop that refer to it.
If ELIMINATE_P is non-zero, actually do the elimination.
THRESHOLD and INSN_COUNT are from loop_optimize and are used to
determine whether invariant insns should be placed inside or at the
start of the loop. */
static int
maybe_eliminate_biv (loop, bl, eliminate_p, threshold, insn_count)
const struct loop *loop;
struct iv_class *bl;
int eliminate_p;
int threshold, insn_count;
{
rtx reg = bl->biv->dest_reg;
rtx loop_start = loop->start;
rtx loop_end = loop->end;
rtx p;
/* Scan all insns in the loop, stopping if we find one that uses the
biv in a way that we cannot eliminate. */
for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
{
enum rtx_code code = GET_CODE (p);
rtx where = threshold >= insn_count ? loop_start : p;
/* If this is a libcall that sets a giv, skip ahead to its end. */
if (GET_RTX_CLASS (code) == 'i')
{
rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
if (note)
{
rtx last = XEXP (note, 0);
rtx set = single_set (last);
if (set && GET_CODE (SET_DEST (set)) == REG)
{
int regno = REGNO (SET_DEST (set));
if (regno < max_reg_before_loop
&& REG_IV_TYPE (regno) == GENERAL_INDUCT
&& REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
p = last;
}
}
}
if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
&& reg_mentioned_p (reg, PATTERN (p))
&& ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
eliminate_p, where))
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"Cannot eliminate biv %d: biv used in insn %d.\n",
bl->regno, INSN_UID (p));
break;
}
}
if (p == loop_end)
{
if (loop_dump_stream)
fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
bl->regno, eliminate_p ? "was" : "can be");
return 1;
}
return 0;
}
/* INSN and REFERENCE are instructions in the same insn chain.
Return non-zero if INSN is first. */
int
loop_insn_first_p (insn, reference)
rtx insn, reference;
{
rtx p, q;
for (p = insn, q = reference; ;)
{
/* Start with test for not first so that INSN == REFERENCE yields not
first. */
if (q == insn || ! p)
return 0;
if (p == reference || ! q)
return 1;
/* Either of P or Q might be a NOTE. Notes have the same LUID as the
previous insn, hence the <= comparison below does not work if
P is a note. */
if (INSN_UID (p) < max_uid_for_loop
&& INSN_UID (q) < max_uid_for_loop
&& GET_CODE (p) != NOTE)
return INSN_LUID (p) <= INSN_LUID (q);
if (INSN_UID (p) >= max_uid_for_loop
|| GET_CODE (p) == NOTE)
p = NEXT_INSN (p);
if (INSN_UID (q) >= max_uid_for_loop)
q = NEXT_INSN (q);
}
}
/* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
the offset that we have to take into account due to auto-increment /
div derivation is zero. */
static int
biv_elimination_giv_has_0_offset (biv, giv, insn)
struct induction *biv, *giv;
rtx insn;
{
/* If the giv V had the auto-inc address optimization applied
to it, and INSN occurs between the giv insn and the biv
insn, then we'd have to adjust the value used here.
This is rare, so we don't bother to make this possible. */
if (giv->auto_inc_opt
&& ((loop_insn_first_p (giv->insn, insn)
&& loop_insn_first_p (insn, biv->insn))
|| (loop_insn_first_p (biv->insn, insn)
&& loop_insn_first_p (insn, giv->insn))))
return 0;
/* If the giv V was derived from another giv, and INSN does
not occur between the giv insn and the biv insn, then we'd
have to adjust the value used here. This is rare, so we don't
bother to make this possible. */
if (giv->derived_from
&& ! (giv->always_executed
&& loop_insn_first_p (giv->insn, insn)
&& loop_insn_first_p (insn, biv->insn)))
return 0;
if (giv->same
&& giv->same->derived_from
&& ! (giv->same->always_executed
&& loop_insn_first_p (giv->same->insn, insn)
&& loop_insn_first_p (insn, biv->insn)))
return 0;
return 1;
}
/* If BL appears in X (part of the pattern of INSN), see if we can
eliminate its use. If so, return 1. If not, return 0.
If BIV does not appear in X, return 1.
If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
where extra insns should be added. Depending on how many items have been
moved out of the loop, it will either be before INSN or at the start of
the loop. */
static int
maybe_eliminate_biv_1 (loop, x, insn, bl, eliminate_p, where)
const struct loop *loop;
rtx x, insn;
struct iv_class *bl;
int eliminate_p;
rtx where;
{
enum rtx_code code = GET_CODE (x);
rtx reg = bl->biv->dest_reg;
enum machine_mode mode = GET_MODE (reg);
struct induction *v;
rtx arg, tem;
#ifdef HAVE_cc0
rtx new;
#endif
int arg_operand;
const char *fmt;
int i, j;
switch (code)
{
case REG:
/* If we haven't already been able to do something with this BIV,
we can't eliminate it. */
if (x == reg)
return 0;
return 1;
case SET:
/* If this sets the BIV, it is not a problem. */
if (SET_DEST (x) == reg)
return 1;
/* If this is an insn that defines a giv, it is also ok because
it will go away when the giv is reduced. */
for (v = bl->giv; v; v = v->next_iv)
if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
return 1;
#ifdef HAVE_cc0
if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
{
/* Can replace with any giv that was reduced and
that has (MULT_VAL != 0) and (ADD_VAL == 0).
Require a constant for MULT_VAL, so we know it's nonzero.
??? We disable this optimization to avoid potential
overflows. */
for (v = bl->giv; v; v = v->next_iv)
if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
&& v->add_val == const0_rtx
&& ! v->ignore && ! v->maybe_dead && v->always_computable
&& v->mode == mode
&& 0)
{
if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
continue;
if (! eliminate_p)
return 1;
/* If the giv has the opposite direction of change,
then reverse the comparison. */
if (INTVAL (v->mult_val) < 0)
new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
const0_rtx, v->new_reg);
else
new = v->new_reg;
/* We can probably test that giv's reduced reg. */
if (validate_change (insn, &SET_SRC (x), new, 0))
return 1;
}
/* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
Require a constant for MULT_VAL, so we know it's nonzero.
??? Do this only if ADD_VAL is a pointer to avoid a potential
overflow problem. */
for (v = bl->giv; v; v = v->next_iv)
if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
&& ! v->ignore && ! v->maybe_dead && v->always_computable
&& v->mode == mode
&& (GET_CODE (v->add_val) == SYMBOL_REF
|| GET_CODE (v->add_val) == LABEL_REF
|| GET_CODE (v->add_val) == CONST
|| (GET_CODE (v->add_val) == REG
&& REGNO_POINTER_FLAG (REGNO (v->add_val)))))
{
if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
continue;
if (! eliminate_p)
return 1;
/* If the giv has the opposite direction of change,
then reverse the comparison. */
if (INTVAL (v->mult_val) < 0)
new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
v->new_reg);
else
new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
copy_rtx (v->add_val));
/* Replace biv with the giv's reduced register. */
update_reg_last_use (v->add_val, insn);
if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
return 1;
/* Insn doesn't support that constant or invariant. Copy it
into a register (it will be a loop invariant.) */
tem = gen_reg_rtx (GET_MODE (v->new_reg));
emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
where);
/* Substitute the new register for its invariant value in
the compare expression. */
XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
return 1;
}
}
#endif
break;
case COMPARE:
case EQ: case NE:
case GT: case GE: case GTU: case GEU:
case LT: case LE: case LTU: case LEU:
/* See if either argument is the biv. */
if (XEXP (x, 0) == reg)
arg = XEXP (x, 1), arg_operand = 1;
else if (XEXP (x, 1) == reg)
arg = XEXP (x, 0), arg_operand = 0;
else
break;
if (CONSTANT_P (arg))
{
/* First try to replace with any giv that has constant positive
mult_val and constant add_val. We might be able to support
negative mult_val, but it seems complex to do it in general. */
for (v = bl->giv; v; v = v->next_iv)
if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
&& (GET_CODE (v->add_val) == SYMBOL_REF
|| GET_CODE (v->add_val) == LABEL_REF
|| GET_CODE (v->add_val) == CONST
|| (GET_CODE (v->add_val) == REG
&& REGNO_POINTER_FLAG (REGNO (v->add_val))))
&& ! v->ignore && ! v->maybe_dead && v->always_computable
&& v->mode == mode)
{
if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
continue;
if (! eliminate_p)
return 1;
/* Replace biv with the giv's reduced reg. */
XEXP (x, 1-arg_operand) = v->new_reg;
/* If all constants are actually constant integers and
the derived constant can be directly placed in the COMPARE,
do so. */
if (GET_CODE (arg) == CONST_INT
&& GET_CODE (v->mult_val) == CONST_INT
&& GET_CODE (v->add_val) == CONST_INT
&& validate_change (insn, &XEXP (x, arg_operand),
GEN_INT (INTVAL (arg)
* INTVAL (v->mult_val)
+ INTVAL (v->add_val)), 0))
return 1;
/* Otherwise, load it into a register. */
tem = gen_reg_rtx (mode);
emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
return 1;
/* If that failed, put back the change we made above. */
XEXP (x, 1-arg_operand) = reg;
}
/* Look for giv with positive constant mult_val and nonconst add_val.
Insert insns to calculate new compare value.
??? Turn this off due to possible overflow. */
for (v = bl->giv; v; v = v->next_iv)
if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
&& ! v->ignore && ! v->maybe_dead && v->always_computable
&& v->mode == mode
&& 0)
{
rtx tem;
if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
continue;
if (! eliminate_p)
return 1;
tem = gen_reg_rtx (mode);
/* Replace biv with giv's reduced register. */
validate_change (insn, &XEXP (x, 1 - arg_operand),
v->new_reg, 1);
/* Compute value to compare against. */
emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
/* Use it in this insn. */
validate_change (insn, &XEXP (x, arg_operand), tem, 1);
if (apply_change_group ())
return 1;
}
}
else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
{
if (loop_invariant_p (loop, arg) == 1)
{
/* Look for giv with constant positive mult_val and nonconst
add_val. Insert insns to compute new compare value.
??? Turn this off due to possible overflow. */
for (v = bl->giv; v; v = v->next_iv)
if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
&& ! v->ignore && ! v->maybe_dead && v->always_computable
&& v->mode == mode
&& 0)
{
rtx tem;
if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
continue;
if (! eliminate_p)
return 1;
tem = gen_reg_rtx (mode);
/* Replace biv with giv's reduced register. */
validate_change (insn, &XEXP (x, 1 - arg_operand),
v->new_reg, 1);
/* Compute value to compare against. */
emit_iv_add_mult (arg, v->mult_val, v->add_val,
tem, where);
validate_change (insn, &XEXP (x, arg_operand), tem, 1);
if (apply_change_group ())
return 1;
}
}
/* This code has problems. Basically, you can't know when
seeing if we will eliminate BL, whether a particular giv
of ARG will be reduced. If it isn't going to be reduced,
we can't eliminate BL. We can try forcing it to be reduced,
but that can generate poor code.
The problem is that the benefit of reducing TV, below should
be increased if BL can actually be eliminated, but this means
we might have to do a topological sort of the order in which
we try to process biv. It doesn't seem worthwhile to do
this sort of thing now. */
#if 0
/* Otherwise the reg compared with had better be a biv. */
if (GET_CODE (arg) != REG
|| REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
return 0;
/* Look for a pair of givs, one for each biv,
with identical coefficients. */
for (v = bl->giv; v; v = v->next_iv)
{
struct induction *tv;
if (v->ignore || v->maybe_dead || v->mode != mode)
continue;
for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
if (! tv->ignore && ! tv->maybe_dead
&& rtx_equal_p (tv->mult_val, v->mult_val)
&& rtx_equal_p (tv->add_val, v->add_val)
&& tv->mode == mode)
{
if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
continue;
if (! eliminate_p)
return 1;
/* Replace biv with its giv's reduced reg. */
XEXP (x, 1-arg_operand) = v->new_reg;
/* Replace other operand with the other giv's
reduced reg. */
XEXP (x, arg_operand) = tv->new_reg;
return 1;
}
}
#endif
}
/* If we get here, the biv can't be eliminated. */
return 0;
case MEM:
/* If this address is a DEST_ADDR giv, it doesn't matter if the
biv is used in it, since it will be replaced. */
for (v = bl->giv; v; v = v->next_iv)
if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
return 1;
break;
default:
break;
}
/* See if any subexpression fails elimination. */
fmt = GET_RTX_FORMAT (code);
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
{
switch (fmt[i])
{
case 'e':
if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
eliminate_p, where))
return 0;
break;
case 'E':
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
eliminate_p, where))
return 0;
break;
}
}
return 1;
}
/* Return nonzero if the last use of REG
is in an insn following INSN in the same basic block. */
static int
last_use_this_basic_block (reg, insn)
rtx reg;
rtx insn;
{
rtx n;
for (n = insn;
n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
n = NEXT_INSN (n))
{
if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
return 1;
}
return 0;
}
/* Called via `note_stores' to record the initial value of a biv. Here we
just record the location of the set and process it later. */
static void
record_initial (dest, set, data)
rtx dest;
rtx set;
void *data ATTRIBUTE_UNUSED;
{
struct iv_class *bl;
if (GET_CODE (dest) != REG
|| REGNO (dest) >= max_reg_before_loop
|| REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
return;
bl = reg_biv_class[REGNO (dest)];
/* If this is the first set found, record it. */
if (bl->init_insn == 0)
{
bl->init_insn = note_insn;
bl->init_set = set;
}
}
/* If any of the registers in X are "old" and currently have a last use earlier
than INSN, update them to have a last use of INSN. Their actual last use
will be the previous insn but it will not have a valid uid_luid so we can't
use it. */
static void
update_reg_last_use (x, insn)
rtx x;
rtx insn;
{
/* Check for the case where INSN does not have a valid luid. In this case,
there is no need to modify the regno_last_uid, as this can only happen
when code is inserted after the loop_end to set a pseudo's final value,
and hence this insn will never be the last use of x. */
if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
&& INSN_UID (insn) < max_uid_for_loop
&& uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
else
{
register int i, j;
register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
update_reg_last_use (XEXP (x, i), insn);
else if (fmt[i] == 'E')
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
update_reg_last_use (XVECEXP (x, i, j), insn);
}
}
}
/* Given an insn INSN and condition COND, return the condition in a
canonical form to simplify testing by callers. Specifically:
(1) The code will always be a comparison operation (EQ, NE, GT, etc.).
(2) Both operands will be machine operands; (cc0) will have been replaced.
(3) If an operand is a constant, it will be the second operand.
(4) (LE x const) will be replaced with (LT x <const+1>) and similarly
for GE, GEU, and LEU.
If the condition cannot be understood, or is an inequality floating-point
comparison which needs to be reversed, 0 will be returned.
If REVERSE is non-zero, then reverse the condition prior to canonizing it.
If EARLIEST is non-zero, it is a pointer to a place where the earliest
insn used in locating the condition was found. If a replacement test
of the condition is desired, it should be placed in front of that
insn and we will be sure that the inputs are still valid. */
rtx
canonicalize_condition (insn, cond, reverse, earliest)
rtx insn;
rtx cond;
int reverse;
rtx *earliest;
{
enum rtx_code code;
rtx prev = insn;
rtx set;
rtx tem;
rtx op0, op1;
int reverse_code = 0;
int did_reverse_condition = 0;
enum machine_mode mode;
code = GET_CODE (cond);
mode = GET_MODE (cond);
op0 = XEXP (cond, 0);
op1 = XEXP (cond, 1);
if (reverse)
{
code = reverse_condition (code);
did_reverse_condition ^= 1;
}
if (earliest)
*earliest = insn;
/* If we are comparing a register with zero, see if the register is set
in the previous insn to a COMPARE or a comparison operation. Perform
the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
in cse.c */
while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
{
/* Set non-zero when we find something of interest. */
rtx x = 0;
#ifdef HAVE_cc0
/* If comparison with cc0, import actual comparison from compare
insn. */
if (op0 == cc0_rtx)
{
if ((prev = prev_nonnote_insn (prev)) == 0
|| GET_CODE (prev) != INSN
|| (set = single_set (prev)) == 0
|| SET_DEST (set) != cc0_rtx)
return 0;
op0 = SET_SRC (set);
op1 = CONST0_RTX (GET_MODE (op0));
if (earliest)
*earliest = prev;
}
#endif
/* If this is a COMPARE, pick up the two things being compared. */
if (GET_CODE (op0) == COMPARE)
{
op1 = XEXP (op0, 1);
op0 = XEXP (op0, 0);
continue;
}
else if (GET_CODE (op0) != REG)
break;
/* Go back to the previous insn. Stop if it is not an INSN. We also
stop if it isn't a single set or if it has a REG_INC note because
we don't want to bother dealing with it. */
if ((prev = prev_nonnote_insn (prev)) == 0
|| GET_CODE (prev) != INSN
|| FIND_REG_INC_NOTE (prev, 0)
|| (set = single_set (prev)) == 0)
break;
/* If this is setting OP0, get what it sets it to if it looks
relevant. */
if (rtx_equal_p (SET_DEST (set), op0))
{
enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
/* ??? We may not combine comparisons done in a CCmode with
comparisons not done in a CCmode. This is to aid targets
like Alpha that have an IEEE compliant EQ instruction, and
a non-IEEE compliant BEQ instruction. The use of CCmode is
actually artificial, simply to prevent the combination, but
should not affect other platforms.
However, we must allow VOIDmode comparisons to match either
CCmode or non-CCmode comparison, because some ports have
modeless comparisons inside branch patterns.
??? This mode check should perhaps look more like the mode check
in simplify_comparison in combine. */
if ((GET_CODE (SET_SRC (set)) == COMPARE
|| (((code == NE
|| (code == LT
&& GET_MODE_CLASS (inner_mode) == MODE_INT
&& (GET_MODE_BITSIZE (inner_mode)
<= HOST_BITS_PER_WIDE_INT)
&& (STORE_FLAG_VALUE
& ((HOST_WIDE_INT) 1
<< (GET_MODE_BITSIZE (inner_mode) - 1))))
#ifdef FLOAT_STORE_FLAG_VALUE
|| (code == LT
&& GET_MODE_CLASS (inner_mode) == MODE_FLOAT
&& (REAL_VALUE_NEGATIVE
(FLOAT_STORE_FLAG_VALUE (inner_mode))))
#endif
))
&& GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
&& (((GET_MODE_CLASS (mode) == MODE_CC)
== (GET_MODE_CLASS (inner_mode) == MODE_CC))
|| mode == VOIDmode || inner_mode == VOIDmode))
x = SET_SRC (set);
else if (((code == EQ
|| (code == GE
&& (GET_MODE_BITSIZE (inner_mode)
<= HOST_BITS_PER_WIDE_INT)
&& GET_MODE_CLASS (inner_mode) == MODE_INT
&& (STORE_FLAG_VALUE
& ((HOST_WIDE_INT) 1
<< (GET_MODE_BITSIZE (inner_mode) - 1))))
#ifdef FLOAT_STORE_FLAG_VALUE
|| (code == GE
&& GET_MODE_CLASS (inner_mode) == MODE_FLOAT
&& (REAL_VALUE_NEGATIVE
(FLOAT_STORE_FLAG_VALUE (inner_mode))))
#endif
))
&& GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
&& (((GET_MODE_CLASS (mode) == MODE_CC)
== (GET_MODE_CLASS (inner_mode) == MODE_CC))
|| mode == VOIDmode || inner_mode == VOIDmode))
{
/* We might have reversed a LT to get a GE here. But this wasn't
actually the comparison of data, so we don't flag that we
have had to reverse the condition. */
did_reverse_condition ^= 1;
reverse_code = 1;
x = SET_SRC (set);
}
else
break;
}
else if (reg_set_p (op0, prev))
/* If this sets OP0, but not directly, we have to give up. */
break;
if (x)
{
if (GET_RTX_CLASS (GET_CODE (x)) == '<')
code = GET_CODE (x);
if (reverse_code)
{
code = reverse_condition (code);
if (code == UNKNOWN)
return 0;
did_reverse_condition ^= 1;
reverse_code = 0;
}
op0 = XEXP (x, 0), op1 = XEXP (x, 1);
if (earliest)
*earliest = prev;
}
}
/* If constant is first, put it last. */
if (CONSTANT_P (op0))
code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
/* If OP0 is the result of a comparison, we weren't able to find what
was really being compared, so fail. */
if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
return 0;
/* Canonicalize any ordered comparison with integers involving equality
if we can do computations in the relevant mode and we do not
overflow. */
if (GET_CODE (op1) == CONST_INT
&& GET_MODE (op0) != VOIDmode
&& GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
{
HOST_WIDE_INT const_val = INTVAL (op1);
unsigned HOST_WIDE_INT uconst_val = const_val;
unsigned HOST_WIDE_INT max_val
= (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
switch (code)
{
case LE:
if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
code = LT, op1 = GEN_INT (const_val + 1);
break;
/* When cross-compiling, const_val might be sign-extended from
BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
case GE:
if ((HOST_WIDE_INT) (const_val & max_val)
!= (((HOST_WIDE_INT) 1
<< (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
code = GT, op1 = GEN_INT (const_val - 1);
break;
case LEU:
if (uconst_val < max_val)
code = LTU, op1 = GEN_INT (uconst_val + 1);
break;
case GEU:
if (uconst_val != 0)
code = GTU, op1 = GEN_INT (uconst_val - 1);
break;
default:
break;
}
}
/* If this was floating-point and we reversed anything other than an
EQ or NE or (UN)ORDERED, return zero. */
if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
&& did_reverse_condition
&& code != NE && code != EQ && code != UNORDERED && code != ORDERED
&& ! flag_fast_math
&& GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
return 0;
#ifdef HAVE_cc0
/* Never return CC0; return zero instead. */
if (op0 == cc0_rtx)
return 0;
#endif
return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
}
/* Given a jump insn JUMP, return the condition that will cause it to branch
to its JUMP_LABEL. If the condition cannot be understood, or is an
inequality floating-point comparison which needs to be reversed, 0 will
be returned.
If EARLIEST is non-zero, it is a pointer to a place where the earliest
insn used in locating the condition was found. If a replacement test
of the condition is desired, it should be placed in front of that
insn and we will be sure that the inputs are still valid. */
rtx
get_condition (jump, earliest)
rtx jump;
rtx *earliest;
{
rtx cond;
int reverse;
/* If this is not a standard conditional jump, we can't parse it. */
if (GET_CODE (jump) != JUMP_INSN
|| ! condjump_p (jump) || simplejump_p (jump))
return 0;
cond = XEXP (SET_SRC (PATTERN (jump)), 0);
/* If this branches to JUMP_LABEL when the condition is false, reverse
the condition. */
reverse
= GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
&& XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump);
return canonicalize_condition (jump, cond, reverse, earliest);
}
/* Similar to above routine, except that we also put an invariant last
unless both operands are invariants. */
rtx
get_condition_for_loop (loop, x)
const struct loop *loop;
rtx x;
{
rtx comparison = get_condition (x, NULL_PTR);
if (comparison == 0
|| ! loop_invariant_p (loop, XEXP (comparison, 0))
|| loop_invariant_p (loop, XEXP (comparison, 1)))
return comparison;
return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
XEXP (comparison, 1), XEXP (comparison, 0));
}
#ifdef HAVE_decrement_and_branch_on_count
/* Instrument loop for insertion of bct instruction. We distinguish between
loops with compile-time bounds and those with run-time bounds.
Information from loop_iterations() is used to compute compile-time bounds.
Run-time bounds should use loop preconditioning, but currently ignored.
*/
static void
insert_bct (loop)
struct loop *loop;
{
unsigned HOST_WIDE_INT n_iterations;
rtx loop_start = loop->start;
rtx loop_end = loop->end;
struct loop_info *loop_info = LOOP_INFO (loop);
int loop_num = loop->num;
#if 0
int increment_direction, compare_direction;
/* If the loop condition is <= or >=, the number of iteration
is 1 more than the range of the bounds of the loop. */
int add_iteration = 0;
enum machine_mode loop_var_mode = word_mode;
#endif
/* It's impossible to instrument a competely unrolled loop. */
if (loop_info->unroll_number == loop_info->n_iterations)
return;
/* Make sure that the count register is not in use. */
if (loop_info->used_count_register)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT instrumentation failed: count register already in use\n",
loop_num);
return;
}
/* Make sure that the function has no indirect jumps. */
if (indirect_jump_in_function)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
loop_num);
return;
}
/* Make sure that the last loop insn is a conditional jump. */
if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
|| ! condjump_p (PREV_INSN (loop_end))
|| simplejump_p (PREV_INSN (loop_end)))
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
loop_num);
return;
}
/* Make sure that the loop does not contain a function call
(the count register might be altered by the called function). */
if (loop_info->has_call)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT instrumentation failed: function call in loop\n",
loop_num);
return;
}
/* Make sure that the loop does not jump via a table.
(the count register might be used to perform the branch on table). */
if (loop_info->has_tablejump)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
loop_num);
return;
}
/* Account for loop unrolling in instrumented iteration count. */
if (loop_info->unroll_number > 1)
n_iterations = loop_info->n_iterations / loop_info->unroll_number;
else
n_iterations = loop_info->n_iterations;
if (n_iterations != 0 && n_iterations < 3)
{
/* Allow an enclosing outer loop to benefit if possible. */
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: Too few iterations to benefit from BCT optimization\n",
loop_num);
return;
}
/* Try to instrument the loop. */
/* Handle the simpler case, where the bounds are known at compile time. */
if (n_iterations > 0)
{
struct loop *outer_loop;
struct loop_info *outer_loop_info;
/* Mark all enclosing loops that they cannot use count register. */
for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
{
outer_loop_info = LOOP_INFO (outer_loop);
outer_loop_info->used_count_register = 1;
}
instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
return;
}
/* Handle the more complex case, that the bounds are NOT known
at compile time. In this case we generate run_time calculation
of the number of iterations. */
if (loop_info->iteration_var == 0)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
loop_num);
return;
}
if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
|| GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
loop_num);
return;
}
/* With runtime bounds, if the compare is of the form '!=' we give up */
if (loop_info->comparison_code == NE)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
loop_num);
return;
}
/* Use common loop preconditioning code instead. */
#if 0
else
{
/* We rely on the existence of run-time guard to ensure that the
loop executes at least once. */
rtx sequence;
rtx iterations_num_reg;
unsigned HOST_WIDE_INT increment_value_abs
= INTVAL (increment) * increment_direction;
/* make sure that the increment is a power of two, otherwise (an
expensive) divide is needed. */
if (exact_log2 (increment_value_abs) == -1)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
"insert_bct: not instrumenting BCT because the increment is not power of 2\n");
return;
}
/* compute the number of iterations */
start_sequence ();
{
rtx temp_reg;
/* Again, the number of iterations is calculated by:
;
; compare-val - initial-val + (increment -1) + additional-iteration
; num_iterations = -----------------------------------------------------------------
; increment
*/
/* ??? Do we have to call copy_rtx here before passing rtx to
expand_binop? */
if (compare_direction > 0)
{
/* <, <= :the loop variable is increasing */
temp_reg = expand_binop (loop_var_mode, sub_optab,
comparison_value, initial_value,
NULL_RTX, 0, OPTAB_LIB_WIDEN);
}
else
{
temp_reg = expand_binop (loop_var_mode, sub_optab,
initial_value, comparison_value,
NULL_RTX, 0, OPTAB_LIB_WIDEN);
}
if (increment_value_abs - 1 + add_iteration != 0)
temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
GEN_INT (increment_value_abs - 1
+ add_iteration),
NULL_RTX, 0, OPTAB_LIB_WIDEN);
if (increment_value_abs != 1)
iterations_num_reg = expand_binop (loop_var_mode, asr_optab,
temp_reg,
GEN_INT (exact_log2 (increment_value_abs)),
NULL_RTX, 0, OPTAB_LIB_WIDEN);
else
iterations_num_reg = temp_reg;
}
sequence = gen_sequence ();
end_sequence ();
emit_insn_before (sequence, loop_start);
instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
}
return;
#endif /* Complex case */
}
/* Instrument loop by inserting a bct in it as follows:
1. A new counter register is created.
2. In the head of the loop the new variable is initialized to the value
passed in the loop_num_iterations parameter.
3. At the end of the loop, comparison of the register with 0 is generated.
The created comparison follows the pattern defined for the
decrement_and_branch_on_count insn, so this insn will be generated.
4. The branch on the old variable are deleted. The compare must remain
because it might be used elsewhere. If the loop-variable or condition
register are used elsewhere, they will be eliminated by flow. */
static void
instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
rtx loop_start, loop_end;
rtx loop_num_iterations;
{
rtx counter_reg;
rtx start_label;
rtx sequence;
if (HAVE_decrement_and_branch_on_count)
{
if (loop_dump_stream)
{
fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
if (GET_CODE (loop_num_iterations) == CONST_INT)
fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
INTVAL (loop_num_iterations));
else
fputs ("runtime", loop_dump_stream);
fputs (" iterations)", loop_dump_stream);
}
/* Discard original jump to continue loop. Original compare result
may still be live, so it cannot be discarded explicitly. */
delete_insn (PREV_INSN (loop_end));
/* Insert the label which will delimit the start of the loop. */
start_label = gen_label_rtx ();
emit_label_after (start_label, loop_start);
/* Insert initialization of the count register into the loop header. */
start_sequence ();
counter_reg = gen_reg_rtx (word_mode);
emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
sequence = gen_sequence ();
end_sequence ();
emit_insn_before (sequence, loop_start);
/* Insert new comparison on the count register instead of the
old one, generating the needed BCT pattern (that will be
later recognized by assembly generation phase). */
emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
start_label),
loop_end);
LABEL_NUSES (start_label)++;
}
}
#endif /* HAVE_decrement_and_branch_on_count */
/* Scan the function and determine whether it has indirect (computed) jumps.
This is taken mostly from flow.c; similar code exists elsewhere
in the compiler. It may be useful to put this into rtlanal.c. */
static int
indirect_jump_in_function_p (start)
rtx start;
{
rtx insn;
for (insn = start; insn; insn = NEXT_INSN (insn))
if (computed_jump_p (insn))
return 1;
return 0;
}
/* Add MEM to the LOOP_MEMS array, if appropriate. See the
documentation for LOOP_MEMS for the definition of `appropriate'.
This function is called from prescan_loop via for_each_rtx. */
static int
insert_loop_mem (mem, data)
rtx *mem;
void *data ATTRIBUTE_UNUSED;
{
int i;
rtx m = *mem;
if (m == NULL_RTX)
return 0;
switch (GET_CODE (m))
{
case MEM:
break;
case CLOBBER:
/* We're not interested in MEMs that are only clobbered. */
return -1;
case CONST_DOUBLE:
/* We're not interested in the MEM associated with a
CONST_DOUBLE, so there's no need to traverse into this. */
return -1;
case EXPR_LIST:
/* We're not interested in any MEMs that only appear in notes. */
return -1;
default:
/* This is not a MEM. */
return 0;
}
/* See if we've already seen this MEM. */
for (i = 0; i < loop_mems_idx; ++i)
if (rtx_equal_p (m, loop_mems[i].mem))
{
if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
/* The modes of the two memory accesses are different. If
this happens, something tricky is going on, and we just
don't optimize accesses to this MEM. */
loop_mems[i].optimize = 0;
return 0;
}
/* Resize the array, if necessary. */
if (loop_mems_idx == loop_mems_allocated)
{
if (loop_mems_allocated != 0)
loop_mems_allocated *= 2;
else
loop_mems_allocated = 32;
loop_mems = (loop_mem_info*)
xrealloc (loop_mems,
loop_mems_allocated * sizeof (loop_mem_info));
}
/* Actually insert the MEM. */
loop_mems[loop_mems_idx].mem = m;
/* We can't hoist this MEM out of the loop if it's a BLKmode MEM
because we can't put it in a register. We still store it in the
table, though, so that if we see the same address later, but in a
non-BLK mode, we'll not think we can optimize it at that point. */
loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
loop_mems[loop_mems_idx].reg = NULL_RTX;
++loop_mems_idx;
return 0;
}
/* Like load_mems, but also ensures that SET_IN_LOOP,
MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
values after load_mems. */
static void
load_mems_and_recount_loop_regs_set (loop, insn_count)
const struct loop *loop;
int *insn_count;
{
int nregs = max_reg_num ();
load_mems (loop);
/* Recalculate set_in_loop and friends since load_mems may have
created new registers. */
if (max_reg_num () > nregs)
{
int i;
int old_nregs;
old_nregs = nregs;
nregs = max_reg_num ();
if ((unsigned) nregs > set_in_loop->num_elements)
{
/* Grow all the arrays. */
VARRAY_GROW (set_in_loop, nregs);
VARRAY_GROW (n_times_set, nregs);
VARRAY_GROW (may_not_optimize, nregs);
VARRAY_GROW (reg_single_usage, nregs);
}
/* Clear the arrays */
bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
may_not_optimize, reg_single_usage,
insn_count, nregs);
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
{
VARRAY_CHAR (may_not_optimize, i) = 1;
VARRAY_INT (set_in_loop, i) = 1;
}
#ifdef AVOID_CCMODE_COPIES
/* Don't try to move insns which set CC registers if we should not
create CCmode register copies. */
for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
VARRAY_CHAR (may_not_optimize, i) = 1;
#endif
/* Set n_times_set for the new registers. */
bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
(char *) (&n_times_set->data.i[0] + old_nregs),
(nregs - old_nregs) * sizeof (int));
}
}
/* Move MEMs into registers for the duration of the loop. */
static void
load_mems (loop)
const struct loop *loop;
{
int maybe_never = 0;
int i;
rtx p;
rtx label = NULL_RTX;
rtx end_label = NULL_RTX;
/* Nonzero if the next instruction may never be executed. */
int next_maybe_never = 0;
int last_max_reg = max_reg_num ();
if (loop_mems_idx == 0)
return;
/* Check to see if it's possible that some instructions in the
loop are never executed. */
for (p = next_insn_in_loop (loop, loop->scan_start);
p != NULL_RTX && ! maybe_never;
p = next_insn_in_loop (loop, p))
{
if (GET_CODE (p) == CODE_LABEL)
maybe_never = 1;
else if (GET_CODE (p) == JUMP_INSN
/* If we enter the loop in the middle, and scan
around to the beginning, don't set maybe_never
for that. This must be an unconditional jump,
otherwise the code at the top of the loop might
never be executed. Unconditional jumps are
followed a by barrier then loop end. */
&& ! (GET_CODE (p) == JUMP_INSN
&& JUMP_LABEL (p) == loop->top
&& NEXT_INSN (NEXT_INSN (p)) == loop->end
&& simplejump_p (p)))
{
if (!condjump_p (p))
/* Something complicated. */
maybe_never = 1;
else
/* If there are any more instructions in the loop, they
might not be reached. */
next_maybe_never = 1;
}
else if (next_maybe_never)
maybe_never = 1;
}
/* Actually move the MEMs. */
for (i = 0; i < loop_mems_idx; ++i)
{
regset_head copies;
int written = 0;
rtx reg;
rtx mem = loop_mems[i].mem;
rtx mem_list_entry;
if (MEM_VOLATILE_P (mem)
|| loop_invariant_p (loop, XEXP (mem, 0)) != 1)
/* There's no telling whether or not MEM is modified. */
loop_mems[i].optimize = 0;
/* Go through the MEMs written to in the loop to see if this
one is aliased by one of them. */
mem_list_entry = loop_store_mems;
while (mem_list_entry)
{
if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
written = 1;
else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
mem, rtx_varies_p))
{
/* MEM is indeed aliased by this store. */
loop_mems[i].optimize = 0;
break;
}
mem_list_entry = XEXP (mem_list_entry, 1);
}
if (flag_float_store && written
&& GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
loop_mems[i].optimize = 0;
/* If this MEM is written to, we must be sure that there
are no reads from another MEM that aliases this one. */
if (loop_mems[i].optimize && written)
{
int j;
for (j = 0; j < loop_mems_idx; ++j)
{
if (j == i)
continue;
else if (true_dependence (mem,
VOIDmode,
loop_mems[j].mem,
rtx_varies_p))
{
/* It's not safe to hoist loop_mems[i] out of
the loop because writes to it might not be
seen by reads from loop_mems[j]. */
loop_mems[i].optimize = 0;
break;
}
}
}
if (maybe_never && may_trap_p (mem))
/* We can't access the MEM outside the loop; it might
cause a trap that wouldn't have happened otherwise. */
loop_mems[i].optimize = 0;
if (!loop_mems[i].optimize)
/* We thought we were going to lift this MEM out of the
loop, but later discovered that we could not. */
continue;
INIT_REG_SET (&copies);
/* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
order to keep scan_loop from moving stores to this MEM
out of the loop just because this REG is neither a
user-variable nor used in the loop test. */
reg = gen_reg_rtx (GET_MODE (mem));
REG_USERVAR_P (reg) = 1;
loop_mems[i].reg = reg;
/* Now, replace all references to the MEM with the
corresponding pesudos. */
maybe_never = 0;
for (p = next_insn_in_loop (loop, loop->scan_start);
p != NULL_RTX;
p = next_insn_in_loop (loop, p))
{
rtx_and_int ri;
rtx set;
if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
{
/* See if this copies the mem into a register that isn't
modified afterwards. We'll try to do copy propagation
a little further on. */
set = single_set (p);
if (set
/* @@@ This test is _way_ too conservative. */
&& ! maybe_never
&& GET_CODE (SET_DEST (set)) == REG
&& REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
&& REGNO (SET_DEST (set)) < last_max_reg
&& VARRAY_INT (n_times_set, REGNO (SET_DEST (set))) == 1
&& rtx_equal_p (SET_SRC (set), loop_mems[i].mem))
SET_REGNO_REG_SET (&copies, REGNO (SET_DEST (set)));
ri.r = p;
ri.i = i;
for_each_rtx (&p, replace_loop_mem, &ri);
}
if (GET_CODE (p) == CODE_LABEL
|| GET_CODE (p) == JUMP_INSN)
maybe_never = 1;
}
if (! apply_change_group ())
/* We couldn't replace all occurrences of the MEM. */
loop_mems[i].optimize = 0;
else
{
int j;
rtx set;
/* Load the memory immediately before LOOP->START, which is
the NOTE_LOOP_BEG. */
set = gen_move_insn (reg, mem);
emit_insn_before (set, loop->start);
if (written)
{
if (label == NULL_RTX)
{
/* We must compute the former
right-after-the-end label before we insert
the new one. */
end_label = next_label (loop->end);
label = gen_label_rtx ();
emit_label_after (label, loop->end);
}
/* Store the memory immediately after END, which is
the NOTE_LOOP_END. */
set = gen_move_insn (copy_rtx (mem), reg);
emit_insn_after (set, label);
}
if (loop_dump_stream)
{
fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
REGNO (reg), (written ? "r/w" : "r/o"));
print_rtl (loop_dump_stream, mem);
fputc ('\n', loop_dump_stream);
}
/* Attempt a bit of copy propagation. This helps untangle the
data flow, and enables {basic,general}_induction_var to find
more bivs/givs. */
EXECUTE_IF_SET_IN_REG_SET
(&copies, FIRST_PSEUDO_REGISTER, j,
{
try_copy_prop (loop, loop_mems[i].reg, j);
});
CLEAR_REG_SET (&copies);
}
}
if (label != NULL_RTX)
{
/* Now, we need to replace all references to the previous exit
label with the new one. */
rtx_pair rr;
rr.r1 = end_label;
rr.r2 = label;
for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
{
for_each_rtx (&p, replace_label, &rr);
/* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
field. This is not handled by for_each_rtx because it doesn't
handle unprinted ('0') fields. We need to update JUMP_LABEL
because the immediately following unroll pass will use it.
replace_label would not work anyways, because that only handles
LABEL_REFs. */
if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
JUMP_LABEL (p) = label;
}
}
}
/* For communication between note_reg_stored and its caller. */
struct note_reg_stored_arg
{
int set_seen;
rtx reg;
};
/* Called via note_stores, record in SET_SEEN whether X, which is written,
is equal to ARG. */
static void
note_reg_stored (x, setter, arg)
rtx x, setter ATTRIBUTE_UNUSED;
void *arg;
{
struct note_reg_stored_arg *t = (struct note_reg_stored_arg *)arg;
if (t->reg == x)
t->set_seen = 1;
}
/* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
There must be exactly one insn that sets this pseudo; it will be
deleted if all replacements succeed and we can prove that the register
is not used after the loop. */
static void
try_copy_prop (loop, replacement, regno)
const struct loop *loop;
rtx replacement;
int regno;
{
/* This is the reg that we are copying from. */
rtx reg_rtx = regno_reg_rtx[regno];
rtx init_insn = 0;
rtx insn;
/* These help keep track of whether we replaced all uses of the reg. */
int replaced_last = 0;
int store_is_first = 0;
for (insn = next_insn_in_loop (loop, loop->scan_start);
insn != NULL_RTX;
insn = next_insn_in_loop (loop, insn))
{
rtx set;
/* Only substitute within one extended basic block from the initializing
insn. */
if (GET_CODE (insn) == CODE_LABEL && init_insn)
break;
if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
continue;
/* Is this the initializing insn? */
set = single_set (insn);
if (set
&& GET_CODE (SET_DEST (set)) == REG
&& REGNO (SET_DEST (set)) == regno)
{
if (init_insn)
abort ();
init_insn = insn;
if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
store_is_first = 1;
}
/* Only substitute after seeing the initializing insn. */
if (init_insn && insn != init_insn)
{
struct note_reg_stored_arg arg;
rtx array[3];
array[0] = reg_rtx;
array[1] = replacement;
array[2] = insn;
for_each_rtx (&insn, replace_loop_reg, array);
if (REGNO_LAST_UID (regno) == INSN_UID (insn))
replaced_last = 1;
/* Stop replacing when REPLACEMENT is modified. */
arg.reg = replacement;
arg.set_seen = 0;
note_stores (PATTERN (insn), note_reg_stored, &arg);
if (arg.set_seen)
break;
}
}
if (! init_insn)
abort ();
if (apply_change_group ())
{
if (loop_dump_stream)
fprintf (loop_dump_stream, " Replaced reg %d", regno);
if (store_is_first && replaced_last)
{
PUT_CODE (init_insn, NOTE);
NOTE_LINE_NUMBER (init_insn) = NOTE_INSN_DELETED;
if (loop_dump_stream)
fprintf (loop_dump_stream, ", deleting init_insn (%d)",
INSN_UID (init_insn));
}
if (loop_dump_stream)
fprintf (loop_dump_stream, ".\n");
}
}
/* Replace MEM with its associated pseudo register. This function is
called from load_mems via for_each_rtx. DATA is actually an
rtx_and_int * describing the instruction currently being scanned
and the MEM we are currently replacing. */
static int
replace_loop_mem (mem, data)
rtx *mem;
void *data;
{
rtx_and_int *ri;
rtx insn;
int i;
rtx m = *mem;
if (m == NULL_RTX)
return 0;
switch (GET_CODE (m))
{
case MEM:
break;
case CONST_DOUBLE:
/* We're not interested in the MEM associated with a
CONST_DOUBLE, so there's no need to traverse into one. */
return -1;
default:
/* This is not a MEM. */
return 0;
}
ri = (rtx_and_int*) data;
i = ri->i;
if (!rtx_equal_p (loop_mems[i].mem, m))
/* This is not the MEM we are currently replacing. */
return 0;
insn = ri->r;
/* Actually replace the MEM. */
validate_change (insn, mem, loop_mems[i].reg, 1);
return 0;
}
/* Replace one register with another. Called through for_each_rtx; PX points
to the rtx being scanned. DATA is actually an array of three rtx's; the
first one is the one to be replaced, and the second one the replacement.
The third one is the current insn. */
static int
replace_loop_reg (px, data)
rtx *px;
void *data;
{
rtx x = *px;
rtx *array = (rtx *)data;
if (x == NULL_RTX)
return 0;
if (x == array[0])
validate_change (array[2], px, array[1], 1);
return 0;
}
/* Replace occurrences of the old exit label for the loop with the new
one. DATA is an rtx_pair containing the old and new labels,
respectively. */
static int
replace_label (x, data)
rtx *x;
void *data;
{
rtx l = *x;
rtx old_label = ((rtx_pair*) data)->r1;
rtx new_label = ((rtx_pair*) data)->r2;
if (l == NULL_RTX)
return 0;
if (GET_CODE (l) != LABEL_REF)
return 0;
if (XEXP (l, 0) != old_label)
return 0;
XEXP (l, 0) = new_label;
++LABEL_NUSES (new_label);
--LABEL_NUSES (old_label);
return 0;
}
|