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
|
//===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// MachineScheduler schedules machine instructions after phi elimination. It
// preserves LiveIntervals so it can be invoked before register allocation.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "misched"
#include "llvm/CodeGen/MachineScheduler.h"
#include "llvm/ADT/OwningPtr.h"
#include "llvm/ADT/PriorityQueue.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/RegisterClassInfo.h"
#include "llvm/CodeGen/ScheduleDFS.h"
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GraphWriter.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
#include <queue>
using namespace llvm;
namespace llvm {
cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
cl::desc("Force top-down list scheduling"));
cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
cl::desc("Force bottom-up list scheduling"));
}
#ifndef NDEBUG
static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
cl::desc("Pop up a window to show MISched dags after they are processed"));
static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
#else
static bool ViewMISchedDAGs = false;
#endif // NDEBUG
static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
cl::desc("Enable register pressure scheduling."), cl::init(true));
static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
cl::desc("Enable cyclic critical path analysis."), cl::init(true));
static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
cl::desc("Enable load clustering."), cl::init(true));
// Experimental heuristics
static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
cl::desc("Enable scheduling for macro fusion."), cl::init(true));
static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
cl::desc("Verify machine instrs before and after machine scheduling"));
// DAG subtrees must have at least this many nodes.
static const unsigned MinSubtreeSize = 8;
//===----------------------------------------------------------------------===//
// Machine Instruction Scheduling Pass and Registry
//===----------------------------------------------------------------------===//
MachineSchedContext::MachineSchedContext():
MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
RegClassInfo = new RegisterClassInfo();
}
MachineSchedContext::~MachineSchedContext() {
delete RegClassInfo;
}
namespace {
/// MachineScheduler runs after coalescing and before register allocation.
class MachineScheduler : public MachineSchedContext,
public MachineFunctionPass {
public:
MachineScheduler();
virtual void getAnalysisUsage(AnalysisUsage &AU) const;
virtual void releaseMemory() {}
virtual bool runOnMachineFunction(MachineFunction&);
virtual void print(raw_ostream &O, const Module* = 0) const;
static char ID; // Class identification, replacement for typeinfo
protected:
ScheduleDAGInstrs *createMachineScheduler();
};
} // namespace
char MachineScheduler::ID = 0;
char &llvm::MachineSchedulerID = MachineScheduler::ID;
INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
"Machine Instruction Scheduler", false, false)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
INITIALIZE_PASS_END(MachineScheduler, "misched",
"Machine Instruction Scheduler", false, false)
MachineScheduler::MachineScheduler()
: MachineFunctionPass(ID) {
initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
}
void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
AU.addRequiredID(MachineDominatorsID);
AU.addRequired<MachineLoopInfo>();
AU.addRequired<AliasAnalysis>();
AU.addRequired<TargetPassConfig>();
AU.addRequired<SlotIndexes>();
AU.addPreserved<SlotIndexes>();
AU.addRequired<LiveIntervals>();
AU.addPreserved<LiveIntervals>();
MachineFunctionPass::getAnalysisUsage(AU);
}
MachinePassRegistry MachineSchedRegistry::Registry;
/// A dummy default scheduler factory indicates whether the scheduler
/// is overridden on the command line.
static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
return 0;
}
/// MachineSchedOpt allows command line selection of the scheduler.
static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
RegisterPassParser<MachineSchedRegistry> >
MachineSchedOpt("misched",
cl::init(&useDefaultMachineSched), cl::Hidden,
cl::desc("Machine instruction scheduler to use"));
static MachineSchedRegistry
DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
useDefaultMachineSched);
/// Forward declare the standard machine scheduler. This will be used as the
/// default scheduler if the target does not set a default.
static ScheduleDAGInstrs *createGenericSched(MachineSchedContext *C);
/// Decrement this iterator until reaching the top or a non-debug instr.
static MachineBasicBlock::const_iterator
priorNonDebug(MachineBasicBlock::const_iterator I,
MachineBasicBlock::const_iterator Beg) {
assert(I != Beg && "reached the top of the region, cannot decrement");
while (--I != Beg) {
if (!I->isDebugValue())
break;
}
return I;
}
/// Non-const version.
static MachineBasicBlock::iterator
priorNonDebug(MachineBasicBlock::iterator I,
MachineBasicBlock::const_iterator Beg) {
return const_cast<MachineInstr*>(
&*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
}
/// If this iterator is a debug value, increment until reaching the End or a
/// non-debug instruction.
static MachineBasicBlock::const_iterator
nextIfDebug(MachineBasicBlock::const_iterator I,
MachineBasicBlock::const_iterator End) {
for(; I != End; ++I) {
if (!I->isDebugValue())
break;
}
return I;
}
/// Non-const version.
static MachineBasicBlock::iterator
nextIfDebug(MachineBasicBlock::iterator I,
MachineBasicBlock::const_iterator End) {
// Cast the return value to nonconst MachineInstr, then cast to an
// instr_iterator, which does not check for null, finally return a
// bundle_iterator.
return MachineBasicBlock::instr_iterator(
const_cast<MachineInstr*>(
&*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
}
/// Instantiate a ScheduleDAGInstrs that will be owned by the called.
ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() {
// Select the scheduler, or set the default.
MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
if (Ctor != useDefaultMachineSched)
return Ctor(this);
// Get the default scheduler set by the target for this function.
ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this);
if (Scheduler)
return Scheduler;
// Default to GenericScheduler.
return createGenericSched(this);
}
/// Top-level MachineScheduler pass driver.
///
/// Visit blocks in function order. Divide each block into scheduling regions
/// and visit them bottom-up. Visiting regions bottom-up is not required, but is
/// consistent with the DAG builder, which traverses the interior of the
/// scheduling regions bottom-up.
///
/// This design avoids exposing scheduling boundaries to the DAG builder,
/// simplifying the DAG builder's support for "special" target instructions.
/// At the same time the design allows target schedulers to operate across
/// scheduling boundaries, for example to bundle the boudary instructions
/// without reordering them. This creates complexity, because the target
/// scheduler must update the RegionBegin and RegionEnd positions cached by
/// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
/// design would be to split blocks at scheduling boundaries, but LLVM has a
/// general bias against block splitting purely for implementation simplicity.
bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
// Initialize the context of the pass.
MF = &mf;
MLI = &getAnalysis<MachineLoopInfo>();
MDT = &getAnalysis<MachineDominatorTree>();
PassConfig = &getAnalysis<TargetPassConfig>();
AA = &getAnalysis<AliasAnalysis>();
LIS = &getAnalysis<LiveIntervals>();
const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
if (VerifyScheduling) {
DEBUG(LIS->dump());
MF->verify(this, "Before machine scheduling.");
}
RegClassInfo->runOnMachineFunction(*MF);
// Instantiate the selected scheduler for this target, function, and
// optimization level.
OwningPtr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
// Visit all machine basic blocks.
//
// TODO: Visit blocks in global postorder or postorder within the bottom-up
// loop tree. Then we can optionally compute global RegPressure.
for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
MBB != MBBEnd; ++MBB) {
Scheduler->startBlock(MBB);
// Break the block into scheduling regions [I, RegionEnd), and schedule each
// region as soon as it is discovered. RegionEnd points the scheduling
// boundary at the bottom of the region. The DAG does not include RegionEnd,
// but the region does (i.e. the next RegionEnd is above the previous
// RegionBegin). If the current block has no terminator then RegionEnd ==
// MBB->end() for the bottom region.
//
// The Scheduler may insert instructions during either schedule() or
// exitRegion(), even for empty regions. So the local iterators 'I' and
// 'RegionEnd' are invalid across these calls.
unsigned RemainingInstrs = MBB->size();
for(MachineBasicBlock::iterator RegionEnd = MBB->end();
RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
// Avoid decrementing RegionEnd for blocks with no terminator.
if (RegionEnd != MBB->end()
|| TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
--RegionEnd;
// Count the boundary instruction.
--RemainingInstrs;
}
// The next region starts above the previous region. Look backward in the
// instruction stream until we find the nearest boundary.
unsigned NumRegionInstrs = 0;
MachineBasicBlock::iterator I = RegionEnd;
for(;I != MBB->begin(); --I, --RemainingInstrs, ++NumRegionInstrs) {
if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
break;
}
// Notify the scheduler of the region, even if we may skip scheduling
// it. Perhaps it still needs to be bundled.
Scheduler->enterRegion(MBB, I, RegionEnd, NumRegionInstrs);
// Skip empty scheduling regions (0 or 1 schedulable instructions).
if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
// Close the current region. Bundle the terminator if needed.
// This invalidates 'RegionEnd' and 'I'.
Scheduler->exitRegion();
continue;
}
DEBUG(dbgs() << "********** MI Scheduling **********\n");
DEBUG(dbgs() << MF->getName()
<< ":BB#" << MBB->getNumber() << " " << MBB->getName()
<< "\n From: " << *I << " To: ";
if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
else dbgs() << "End";
dbgs() << " RegionInstrs: " << NumRegionInstrs
<< " Remaining: " << RemainingInstrs << "\n");
// Schedule a region: possibly reorder instructions.
// This invalidates 'RegionEnd' and 'I'.
Scheduler->schedule();
// Close the current region.
Scheduler->exitRegion();
// Scheduling has invalidated the current iterator 'I'. Ask the
// scheduler for the top of it's scheduled region.
RegionEnd = Scheduler->begin();
}
assert(RemainingInstrs == 0 && "Instruction count mismatch!");
Scheduler->finishBlock();
}
Scheduler->finalizeSchedule();
DEBUG(LIS->dump());
if (VerifyScheduling)
MF->verify(this, "After machine scheduling.");
return true;
}
void MachineScheduler::print(raw_ostream &O, const Module* m) const {
// unimplemented
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void ReadyQueue::dump() {
dbgs() << Name << ": ";
for (unsigned i = 0, e = Queue.size(); i < e; ++i)
dbgs() << Queue[i]->NodeNum << " ";
dbgs() << "\n";
}
#endif
//===----------------------------------------------------------------------===//
// ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
// preservation.
//===----------------------------------------------------------------------===//
ScheduleDAGMI::~ScheduleDAGMI() {
delete DFSResult;
DeleteContainerPointers(Mutations);
delete SchedImpl;
}
bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
}
bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
if (SuccSU != &ExitSU) {
// Do not use WillCreateCycle, it assumes SD scheduling.
// If Pred is reachable from Succ, then the edge creates a cycle.
if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
return false;
Topo.AddPred(SuccSU, PredDep.getSUnit());
}
SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
// Return true regardless of whether a new edge needed to be inserted.
return true;
}
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
/// NumPredsLeft reaches zero, release the successor node.
///
/// FIXME: Adjust SuccSU height based on MinLatency.
void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
SUnit *SuccSU = SuccEdge->getSUnit();
if (SuccEdge->isWeak()) {
--SuccSU->WeakPredsLeft;
if (SuccEdge->isCluster())
NextClusterSucc = SuccSU;
return;
}
#ifndef NDEBUG
if (SuccSU->NumPredsLeft == 0) {
dbgs() << "*** Scheduling failed! ***\n";
SuccSU->dump(this);
dbgs() << " has been released too many times!\n";
llvm_unreachable(0);
}
#endif
--SuccSU->NumPredsLeft;
if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
SchedImpl->releaseTopNode(SuccSU);
}
/// releaseSuccessors - Call releaseSucc on each of SU's successors.
void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
releaseSucc(SU, &*I);
}
}
/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
/// NumSuccsLeft reaches zero, release the predecessor node.
///
/// FIXME: Adjust PredSU height based on MinLatency.
void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
SUnit *PredSU = PredEdge->getSUnit();
if (PredEdge->isWeak()) {
--PredSU->WeakSuccsLeft;
if (PredEdge->isCluster())
NextClusterPred = PredSU;
return;
}
#ifndef NDEBUG
if (PredSU->NumSuccsLeft == 0) {
dbgs() << "*** Scheduling failed! ***\n";
PredSU->dump(this);
dbgs() << " has been released too many times!\n";
llvm_unreachable(0);
}
#endif
--PredSU->NumSuccsLeft;
if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
SchedImpl->releaseBottomNode(PredSU);
}
/// releasePredecessors - Call releasePred on each of SU's predecessors.
void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
releasePred(SU, &*I);
}
}
/// This is normally called from the main scheduler loop but may also be invoked
/// by the scheduling strategy to perform additional code motion.
void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
MachineBasicBlock::iterator InsertPos) {
// Advance RegionBegin if the first instruction moves down.
if (&*RegionBegin == MI)
++RegionBegin;
// Update the instruction stream.
BB->splice(InsertPos, BB, MI);
// Update LiveIntervals
LIS->handleMove(MI, /*UpdateFlags=*/true);
// Recede RegionBegin if an instruction moves above the first.
if (RegionBegin == InsertPos)
RegionBegin = MI;
}
bool ScheduleDAGMI::checkSchedLimit() {
#ifndef NDEBUG
if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
CurrentTop = CurrentBottom;
return false;
}
++NumInstrsScheduled;
#endif
return true;
}
/// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
/// crossing a scheduling boundary. [begin, end) includes all instructions in
/// the region, including the boundary itself and single-instruction regions
/// that don't get scheduled.
void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
MachineBasicBlock::iterator begin,
MachineBasicBlock::iterator end,
unsigned regioninstrs)
{
ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
// For convenience remember the end of the liveness region.
LiveRegionEnd =
(RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
SUPressureDiffs.clear();
SchedImpl->initPolicy(begin, end, regioninstrs);
ShouldTrackPressure = SchedImpl->shouldTrackPressure();
}
// Setup the register pressure trackers for the top scheduled top and bottom
// scheduled regions.
void ScheduleDAGMI::initRegPressure() {
TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
// Close the RPTracker to finalize live ins.
RPTracker.closeRegion();
DEBUG(RPTracker.dump());
// Initialize the live ins and live outs.
TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
// Close one end of the tracker so we can call
// getMaxUpward/DownwardPressureDelta before advancing across any
// instructions. This converts currently live regs into live ins/outs.
TopRPTracker.closeTop();
BotRPTracker.closeBottom();
BotRPTracker.initLiveThru(RPTracker);
if (!BotRPTracker.getLiveThru().empty()) {
TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
DEBUG(dbgs() << "Live Thru: ";
dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
};
// For each live out vreg reduce the pressure change associated with other
// uses of the same vreg below the live-out reaching def.
updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
// Account for liveness generated by the region boundary.
if (LiveRegionEnd != RegionEnd) {
SmallVector<unsigned, 8> LiveUses;
BotRPTracker.recede(&LiveUses);
updatePressureDiffs(LiveUses);
}
assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
// Cache the list of excess pressure sets in this region. This will also track
// the max pressure in the scheduled code for these sets.
RegionCriticalPSets.clear();
const std::vector<unsigned> &RegionPressure =
RPTracker.getPressure().MaxSetPressure;
for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
if (RegionPressure[i] > Limit) {
DEBUG(dbgs() << TRI->getRegPressureSetName(i)
<< " Limit " << Limit
<< " Actual " << RegionPressure[i] << "\n");
RegionCriticalPSets.push_back(PressureChange(i));
}
}
DEBUG(dbgs() << "Excess PSets: ";
for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
dbgs() << TRI->getRegPressureSetName(
RegionCriticalPSets[i].getPSet()) << " ";
dbgs() << "\n");
}
void ScheduleDAGMI::
updateScheduledPressure(const SUnit *SU,
const std::vector<unsigned> &NewMaxPressure) {
const PressureDiff &PDiff = getPressureDiff(SU);
unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
I != E; ++I) {
if (!I->isValid())
break;
unsigned ID = I->getPSet();
while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
++CritIdx;
if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
&& NewMaxPressure[ID] <= INT16_MAX)
RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
}
unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
if (NewMaxPressure[ID] >= Limit - 2) {
DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
<< NewMaxPressure[ID] << " > " << Limit << "(+ "
<< BotRPTracker.getLiveThru()[ID] << " livethru)\n");
}
}
}
/// Update the PressureDiff array for liveness after scheduling this
/// instruction.
void ScheduleDAGMI::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
/// FIXME: Currently assuming single-use physregs.
unsigned Reg = LiveUses[LUIdx];
DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
if (!TRI->isVirtualRegister(Reg))
continue;
// This may be called before CurrentBottom has been initialized. However,
// BotRPTracker must have a valid position. We want the value live into the
// instruction or live out of the block, so ask for the previous
// instruction's live-out.
const LiveInterval &LI = LIS->getInterval(Reg);
VNInfo *VNI;
MachineBasicBlock::const_iterator I =
nextIfDebug(BotRPTracker.getPos(), BB->end());
if (I == BB->end())
VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
else {
LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(I));
VNI = LRQ.valueIn();
}
// RegisterPressureTracker guarantees that readsReg is true for LiveUses.
assert(VNI && "No live value at use.");
for (VReg2UseMap::iterator
UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
SUnit *SU = UI->SU;
DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
<< *SU->getInstr());
// If this use comes before the reaching def, it cannot be a last use, so
// descrease its pressure change.
if (!SU->isScheduled && SU != &ExitSU) {
LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(SU->getInstr()));
if (LRQ.valueIn() == VNI)
getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
}
}
}
}
/// schedule - Called back from MachineScheduler::runOnMachineFunction
/// after setting up the current scheduling region. [RegionBegin, RegionEnd)
/// only includes instructions that have DAG nodes, not scheduling boundaries.
///
/// This is a skeletal driver, with all the functionality pushed into helpers,
/// so that it can be easilly extended by experimental schedulers. Generally,
/// implementing MachineSchedStrategy should be sufficient to implement a new
/// scheduling algorithm. However, if a scheduler further subclasses
/// ScheduleDAGMI then it will want to override this virtual method in order to
/// update any specialized state.
void ScheduleDAGMI::schedule() {
buildDAGWithRegPressure();
Topo.InitDAGTopologicalSorting();
postprocessDAG();
SmallVector<SUnit*, 8> TopRoots, BotRoots;
findRootsAndBiasEdges(TopRoots, BotRoots);
// Initialize the strategy before modifying the DAG.
// This may initialize a DFSResult to be used for queue priority.
SchedImpl->initialize(this);
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
SUnits[su].dumpAll(this));
if (ViewMISchedDAGs) viewGraph();
// Initialize ready queues now that the DAG and priority data are finalized.
initQueues(TopRoots, BotRoots);
bool IsTopNode = false;
while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
assert(!SU->isScheduled && "Node already scheduled");
if (!checkSchedLimit())
break;
scheduleMI(SU, IsTopNode);
updateQueues(SU, IsTopNode);
}
assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
placeDebugValues();
DEBUG({
unsigned BBNum = begin()->getParent()->getNumber();
dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
dumpSchedule();
dbgs() << '\n';
});
}
/// Build the DAG and setup three register pressure trackers.
void ScheduleDAGMI::buildDAGWithRegPressure() {
if (!ShouldTrackPressure) {
RPTracker.reset();
RegionCriticalPSets.clear();
buildSchedGraph(AA);
return;
}
// Initialize the register pressure tracker used by buildSchedGraph.
RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
/*TrackUntiedDefs=*/true);
// Account for liveness generate by the region boundary.
if (LiveRegionEnd != RegionEnd)
RPTracker.recede();
// Build the DAG, and compute current register pressure.
buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
// Initialize top/bottom trackers after computing region pressure.
initRegPressure();
}
/// Apply each ScheduleDAGMutation step in order.
void ScheduleDAGMI::postprocessDAG() {
for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
Mutations[i]->apply(this);
}
}
void ScheduleDAGMI::computeDFSResult() {
if (!DFSResult)
DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
DFSResult->clear();
ScheduledTrees.clear();
DFSResult->resize(SUnits.size());
DFSResult->compute(SUnits);
ScheduledTrees.resize(DFSResult->getNumSubtrees());
}
void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
SmallVectorImpl<SUnit*> &BotRoots) {
for (std::vector<SUnit>::iterator
I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
SUnit *SU = &(*I);
assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
// Order predecessors so DFSResult follows the critical path.
SU->biasCriticalPath();
// A SUnit is ready to top schedule if it has no predecessors.
if (!I->NumPredsLeft)
TopRoots.push_back(SU);
// A SUnit is ready to bottom schedule if it has no successors.
if (!I->NumSuccsLeft)
BotRoots.push_back(SU);
}
ExitSU.biasCriticalPath();
}
/// Compute the max cyclic critical path through the DAG. The scheduling DAG
/// only provides the critical path for single block loops. To handle loops that
/// span blocks, we could use the vreg path latencies provided by
/// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
/// available for use in the scheduler.
///
/// The cyclic path estimation identifies a def-use pair that crosses the back
/// edge and considers the depth and height of the nodes. For example, consider
/// the following instruction sequence where each instruction has unit latency
/// and defines an epomymous virtual register:
///
/// a->b(a,c)->c(b)->d(c)->exit
///
/// The cyclic critical path is a two cycles: b->c->b
/// The acyclic critical path is four cycles: a->b->c->d->exit
/// LiveOutHeight = height(c) = len(c->d->exit) = 2
/// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
/// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
/// LiveInDepth = depth(b) = len(a->b) = 1
///
/// LiveOutDepth - LiveInDepth = 3 - 1 = 2
/// LiveInHeight - LiveOutHeight = 4 - 2 = 2
/// CyclicCriticalPath = min(2, 2) = 2
unsigned ScheduleDAGMI::computeCyclicCriticalPath() {
// This only applies to single block loop.
if (!BB->isSuccessor(BB))
return 0;
unsigned MaxCyclicLatency = 0;
// Visit each live out vreg def to find def/use pairs that cross iterations.
ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs;
for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end();
RI != RE; ++RI) {
unsigned Reg = *RI;
if (!TRI->isVirtualRegister(Reg))
continue;
const LiveInterval &LI = LIS->getInterval(Reg);
const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
if (!DefVNI)
continue;
MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
const SUnit *DefSU = getSUnit(DefMI);
if (!DefSU)
continue;
unsigned LiveOutHeight = DefSU->getHeight();
unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
// Visit all local users of the vreg def.
for (VReg2UseMap::iterator
UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
if (UI->SU == &ExitSU)
continue;
// Only consider uses of the phi.
LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(UI->SU->getInstr()));
if (!LRQ.valueIn()->isPHIDef())
continue;
// Assume that a path spanning two iterations is a cycle, which could
// overestimate in strange cases. This allows cyclic latency to be
// estimated as the minimum slack of the vreg's depth or height.
unsigned CyclicLatency = 0;
if (LiveOutDepth > UI->SU->getDepth())
CyclicLatency = LiveOutDepth - UI->SU->getDepth();
unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency;
if (LiveInHeight > LiveOutHeight) {
if (LiveInHeight - LiveOutHeight < CyclicLatency)
CyclicLatency = LiveInHeight - LiveOutHeight;
}
else
CyclicLatency = 0;
DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
<< UI->SU->NodeNum << ") = " << CyclicLatency << "c\n");
if (CyclicLatency > MaxCyclicLatency)
MaxCyclicLatency = CyclicLatency;
}
}
DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
return MaxCyclicLatency;
}
/// Identify DAG roots and setup scheduler queues.
void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
ArrayRef<SUnit*> BotRoots) {
NextClusterSucc = NULL;
NextClusterPred = NULL;
// Release all DAG roots for scheduling, not including EntrySU/ExitSU.
//
// Nodes with unreleased weak edges can still be roots.
// Release top roots in forward order.
for (SmallVectorImpl<SUnit*>::const_iterator
I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
SchedImpl->releaseTopNode(*I);
}
// Release bottom roots in reverse order so the higher priority nodes appear
// first. This is more natural and slightly more efficient.
for (SmallVectorImpl<SUnit*>::const_reverse_iterator
I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
SchedImpl->releaseBottomNode(*I);
}
releaseSuccessors(&EntrySU);
releasePredecessors(&ExitSU);
SchedImpl->registerRoots();
// Advance past initial DebugValues.
CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
CurrentBottom = RegionEnd;
if (ShouldTrackPressure) {
assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
TopRPTracker.setPos(CurrentTop);
}
}
/// Move an instruction and update register pressure.
void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
// Move the instruction to its new location in the instruction stream.
MachineInstr *MI = SU->getInstr();
if (IsTopNode) {
assert(SU->isTopReady() && "node still has unscheduled dependencies");
if (&*CurrentTop == MI)
CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
else {
moveInstruction(MI, CurrentTop);
TopRPTracker.setPos(MI);
}
if (ShouldTrackPressure) {
// Update top scheduled pressure.
TopRPTracker.advance();
assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
}
}
else {
assert(SU->isBottomReady() && "node still has unscheduled dependencies");
MachineBasicBlock::iterator priorII =
priorNonDebug(CurrentBottom, CurrentTop);
if (&*priorII == MI)
CurrentBottom = priorII;
else {
if (&*CurrentTop == MI) {
CurrentTop = nextIfDebug(++CurrentTop, priorII);
TopRPTracker.setPos(CurrentTop);
}
moveInstruction(MI, CurrentBottom);
CurrentBottom = MI;
}
if (ShouldTrackPressure) {
// Update bottom scheduled pressure.
SmallVector<unsigned, 8> LiveUses;
BotRPTracker.recede(&LiveUses);
assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
updatePressureDiffs(LiveUses);
}
}
}
/// Update scheduler queues after scheduling an instruction.
void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
// Release dependent instructions for scheduling.
if (IsTopNode)
releaseSuccessors(SU);
else
releasePredecessors(SU);
SU->isScheduled = true;
if (DFSResult) {
unsigned SubtreeID = DFSResult->getSubtreeID(SU);
if (!ScheduledTrees.test(SubtreeID)) {
ScheduledTrees.set(SubtreeID);
DFSResult->scheduleTree(SubtreeID);
SchedImpl->scheduleTree(SubtreeID);
}
}
// Notify the scheduling strategy after updating the DAG.
SchedImpl->schedNode(SU, IsTopNode);
}
/// Reinsert any remaining debug_values, just like the PostRA scheduler.
void ScheduleDAGMI::placeDebugValues() {
// If first instruction was a DBG_VALUE then put it back.
if (FirstDbgValue) {
BB->splice(RegionBegin, BB, FirstDbgValue);
RegionBegin = FirstDbgValue;
}
for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
MachineInstr *DbgValue = P.first;
MachineBasicBlock::iterator OrigPrevMI = P.second;
if (&*RegionBegin == DbgValue)
++RegionBegin;
BB->splice(++OrigPrevMI, BB, DbgValue);
if (OrigPrevMI == llvm::prior(RegionEnd))
RegionEnd = DbgValue;
}
DbgValues.clear();
FirstDbgValue = NULL;
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void ScheduleDAGMI::dumpSchedule() const {
for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
if (SUnit *SU = getSUnit(&(*MI)))
SU->dump(this);
else
dbgs() << "Missing SUnit\n";
}
}
#endif
//===----------------------------------------------------------------------===//
// LoadClusterMutation - DAG post-processing to cluster loads.
//===----------------------------------------------------------------------===//
namespace {
/// \brief Post-process the DAG to create cluster edges between neighboring
/// loads.
class LoadClusterMutation : public ScheduleDAGMutation {
struct LoadInfo {
SUnit *SU;
unsigned BaseReg;
unsigned Offset;
LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
: SU(su), BaseReg(reg), Offset(ofs) {}
};
static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
const LoadClusterMutation::LoadInfo &RHS);
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
public:
LoadClusterMutation(const TargetInstrInfo *tii,
const TargetRegisterInfo *tri)
: TII(tii), TRI(tri) {}
virtual void apply(ScheduleDAGMI *DAG);
protected:
void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
};
} // anonymous
bool LoadClusterMutation::LoadInfoLess(
const LoadClusterMutation::LoadInfo &LHS,
const LoadClusterMutation::LoadInfo &RHS) {
if (LHS.BaseReg != RHS.BaseReg)
return LHS.BaseReg < RHS.BaseReg;
return LHS.Offset < RHS.Offset;
}
void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
ScheduleDAGMI *DAG) {
SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
SUnit *SU = Loads[Idx];
unsigned BaseReg;
unsigned Offset;
if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
}
if (LoadRecords.size() < 2)
return;
std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
unsigned ClusterLength = 1;
for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
ClusterLength = 1;
continue;
}
SUnit *SUa = LoadRecords[Idx].SU;
SUnit *SUb = LoadRecords[Idx+1].SU;
if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
&& DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
<< SUb->NodeNum << ")\n");
// Copy successor edges from SUa to SUb. Interleaving computation
// dependent on SUa can prevent load combining due to register reuse.
// Predecessor edges do not need to be copied from SUb to SUa since nearby
// loads should have effectively the same inputs.
for (SUnit::const_succ_iterator
SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
if (SI->getSUnit() == SUb)
continue;
DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
}
++ClusterLength;
}
else
ClusterLength = 1;
}
}
/// \brief Callback from DAG postProcessing to create cluster edges for loads.
void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
// Map DAG NodeNum to store chain ID.
DenseMap<unsigned, unsigned> StoreChainIDs;
// Map each store chain to a set of dependent loads.
SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
SUnit *SU = &DAG->SUnits[Idx];
if (!SU->getInstr()->mayLoad())
continue;
unsigned ChainPredID = DAG->SUnits.size();
for (SUnit::const_pred_iterator
PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
if (PI->isCtrl()) {
ChainPredID = PI->getSUnit()->NodeNum;
break;
}
}
// Check if this chain-like pred has been seen
// before. ChainPredID==MaxNodeID for loads at the top of the schedule.
unsigned NumChains = StoreChainDependents.size();
std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
if (Result.second)
StoreChainDependents.resize(NumChains + 1);
StoreChainDependents[Result.first->second].push_back(SU);
}
// Iterate over the store chains.
for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
}
//===----------------------------------------------------------------------===//
// MacroFusion - DAG post-processing to encourage fusion of macro ops.
//===----------------------------------------------------------------------===//
namespace {
/// \brief Post-process the DAG to create cluster edges between instructions
/// that may be fused by the processor into a single operation.
class MacroFusion : public ScheduleDAGMutation {
const TargetInstrInfo *TII;
public:
MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
virtual void apply(ScheduleDAGMI *DAG);
};
} // anonymous
/// \brief Callback from DAG postProcessing to create cluster edges to encourage
/// fused operations.
void MacroFusion::apply(ScheduleDAGMI *DAG) {
// For now, assume targets can only fuse with the branch.
MachineInstr *Branch = DAG->ExitSU.getInstr();
if (!Branch)
return;
for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
SUnit *SU = &DAG->SUnits[--Idx];
if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
continue;
// Create a single weak edge from SU to ExitSU. The only effect is to cause
// bottom-up scheduling to heavily prioritize the clustered SU. There is no
// need to copy predecessor edges from ExitSU to SU, since top-down
// scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
// of SU, we could create an artificial edge from the deepest root, but it
// hasn't been needed yet.
bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
(void)Success;
assert(Success && "No DAG nodes should be reachable from ExitSU");
DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
break;
}
}
//===----------------------------------------------------------------------===//
// CopyConstrain - DAG post-processing to encourage copy elimination.
//===----------------------------------------------------------------------===//
namespace {
/// \brief Post-process the DAG to create weak edges from all uses of a copy to
/// the one use that defines the copy's source vreg, most likely an induction
/// variable increment.
class CopyConstrain : public ScheduleDAGMutation {
// Transient state.
SlotIndex RegionBeginIdx;
// RegionEndIdx is the slot index of the last non-debug instruction in the
// scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
SlotIndex RegionEndIdx;
public:
CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
virtual void apply(ScheduleDAGMI *DAG);
protected:
void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG);
};
} // anonymous
/// constrainLocalCopy handles two possibilities:
/// 1) Local src:
/// I0: = dst
/// I1: src = ...
/// I2: = dst
/// I3: dst = src (copy)
/// (create pred->succ edges I0->I1, I2->I1)
///
/// 2) Local copy:
/// I0: dst = src (copy)
/// I1: = dst
/// I2: src = ...
/// I3: = dst
/// (create pred->succ edges I1->I2, I3->I2)
///
/// Although the MachineScheduler is currently constrained to single blocks,
/// this algorithm should handle extended blocks. An EBB is a set of
/// contiguously numbered blocks such that the previous block in the EBB is
/// always the single predecessor.
void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG) {
LiveIntervals *LIS = DAG->getLIS();
MachineInstr *Copy = CopySU->getInstr();
// Check for pure vreg copies.
unsigned SrcReg = Copy->getOperand(1).getReg();
if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
return;
unsigned DstReg = Copy->getOperand(0).getReg();
if (!TargetRegisterInfo::isVirtualRegister(DstReg))
return;
// Check if either the dest or source is local. If it's live across a back
// edge, it's not local. Note that if both vregs are live across the back
// edge, we cannot successfully contrain the copy without cyclic scheduling.
unsigned LocalReg = DstReg;
unsigned GlobalReg = SrcReg;
LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
LocalReg = SrcReg;
GlobalReg = DstReg;
LocalLI = &LIS->getInterval(LocalReg);
if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
return;
}
LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
// Find the global segment after the start of the local LI.
LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
// If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
// local live range. We could create edges from other global uses to the local
// start, but the coalescer should have already eliminated these cases, so
// don't bother dealing with it.
if (GlobalSegment == GlobalLI->end())
return;
// If GlobalSegment is killed at the LocalLI->start, the call to find()
// returned the next global segment. But if GlobalSegment overlaps with
// LocalLI->start, then advance to the next segement. If a hole in GlobalLI
// exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
if (GlobalSegment->contains(LocalLI->beginIndex()))
++GlobalSegment;
if (GlobalSegment == GlobalLI->end())
return;
// Check if GlobalLI contains a hole in the vicinity of LocalLI.
if (GlobalSegment != GlobalLI->begin()) {
// Two address defs have no hole.
if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->end,
GlobalSegment->start)) {
return;
}
// If the prior global segment may be defined by the same two-address
// instruction that also defines LocalLI, then can't make a hole here.
if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->start,
LocalLI->beginIndex())) {
return;
}
// If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
// it would be a disconnected component in the live range.
assert(llvm::prior(GlobalSegment)->start < LocalLI->beginIndex() &&
"Disconnected LRG within the scheduling region.");
}
MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
if (!GlobalDef)
return;
SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
if (!GlobalSU)
return;
// GlobalDef is the bottom of the GlobalLI hole. Open the hole by
// constraining the uses of the last local def to precede GlobalDef.
SmallVector<SUnit*,8> LocalUses;
const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
for (SUnit::const_succ_iterator
I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
I != E; ++I) {
if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
continue;
if (I->getSUnit() == GlobalSU)
continue;
if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
return;
LocalUses.push_back(I->getSUnit());
}
// Open the top of the GlobalLI hole by constraining any earlier global uses
// to precede the start of LocalLI.
SmallVector<SUnit*,8> GlobalUses;
MachineInstr *FirstLocalDef =
LIS->getInstructionFromIndex(LocalLI->beginIndex());
SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
for (SUnit::const_pred_iterator
I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
continue;
if (I->getSUnit() == FirstLocalSU)
continue;
if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
return;
GlobalUses.push_back(I->getSUnit());
}
DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
// Add the weak edges.
for (SmallVectorImpl<SUnit*>::const_iterator
I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
<< GlobalSU->NodeNum << ")\n");
DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
}
for (SmallVectorImpl<SUnit*>::const_iterator
I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
<< FirstLocalSU->NodeNum << ")\n");
DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
}
}
/// \brief Callback from DAG postProcessing to create weak edges to encourage
/// copy elimination.
void CopyConstrain::apply(ScheduleDAGMI *DAG) {
MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
if (FirstPos == DAG->end())
return;
RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
RegionEndIdx = DAG->getLIS()->getInstructionIndex(
&*priorNonDebug(DAG->end(), DAG->begin()));
for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
SUnit *SU = &DAG->SUnits[Idx];
if (!SU->getInstr()->isCopy())
continue;
constrainLocalCopy(SU, DAG);
}
}
//===----------------------------------------------------------------------===//
// GenericScheduler - Implementation of the generic MachineSchedStrategy.
//===----------------------------------------------------------------------===//
namespace {
/// GenericScheduler shrinks the unscheduled zone using heuristics to balance
/// the schedule.
class GenericScheduler : public MachineSchedStrategy {
public:
/// Represent the type of SchedCandidate found within a single queue.
/// pickNodeBidirectional depends on these listed by decreasing priority.
enum CandReason {
NoCand, PhysRegCopy, RegExcess, RegCritical, Cluster, Weak, RegMax,
ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
TopDepthReduce, TopPathReduce, NextDefUse, NodeOrder};
#ifndef NDEBUG
static const char *getReasonStr(GenericScheduler::CandReason Reason);
#endif
/// Policy for scheduling the next instruction in the candidate's zone.
struct CandPolicy {
bool ReduceLatency;
unsigned ReduceResIdx;
unsigned DemandResIdx;
CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
};
/// Status of an instruction's critical resource consumption.
struct SchedResourceDelta {
// Count critical resources in the scheduled region required by SU.
unsigned CritResources;
// Count critical resources from another region consumed by SU.
unsigned DemandedResources;
SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
bool operator==(const SchedResourceDelta &RHS) const {
return CritResources == RHS.CritResources
&& DemandedResources == RHS.DemandedResources;
}
bool operator!=(const SchedResourceDelta &RHS) const {
return !operator==(RHS);
}
};
/// Store the state used by GenericScheduler heuristics, required for the
/// lifetime of one invocation of pickNode().
struct SchedCandidate {
CandPolicy Policy;
// The best SUnit candidate.
SUnit *SU;
// The reason for this candidate.
CandReason Reason;
// Set of reasons that apply to multiple candidates.
uint32_t RepeatReasonSet;
// Register pressure values for the best candidate.
RegPressureDelta RPDelta;
// Critical resource consumption of the best candidate.
SchedResourceDelta ResDelta;
SchedCandidate(const CandPolicy &policy)
: Policy(policy), SU(NULL), Reason(NoCand), RepeatReasonSet(0) {}
bool isValid() const { return SU; }
// Copy the status of another candidate without changing policy.
void setBest(SchedCandidate &Best) {
assert(Best.Reason != NoCand && "uninitialized Sched candidate");
SU = Best.SU;
Reason = Best.Reason;
RPDelta = Best.RPDelta;
ResDelta = Best.ResDelta;
}
bool isRepeat(CandReason R) { return RepeatReasonSet & (1 << R); }
void setRepeat(CandReason R) { RepeatReasonSet |= (1 << R); }
void initResourceDelta(const ScheduleDAGMI *DAG,
const TargetSchedModel *SchedModel);
};
/// Summarize the unscheduled region.
struct SchedRemainder {
// Critical path through the DAG in expected latency.
unsigned CriticalPath;
unsigned CyclicCritPath;
// Scaled count of micro-ops left to schedule.
unsigned RemIssueCount;
bool IsAcyclicLatencyLimited;
// Unscheduled resources
SmallVector<unsigned, 16> RemainingCounts;
void reset() {
CriticalPath = 0;
CyclicCritPath = 0;
RemIssueCount = 0;
IsAcyclicLatencyLimited = false;
RemainingCounts.clear();
}
SchedRemainder() { reset(); }
void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
};
/// Each Scheduling boundary is associated with ready queues. It tracks the
/// current cycle in the direction of movement, and maintains the state
/// of "hazards" and other interlocks at the current cycle.
struct SchedBoundary {
ScheduleDAGMI *DAG;
const TargetSchedModel *SchedModel;
SchedRemainder *Rem;
ReadyQueue Available;
ReadyQueue Pending;
bool CheckPending;
// For heuristics, keep a list of the nodes that immediately depend on the
// most recently scheduled node.
SmallPtrSet<const SUnit*, 8> NextSUs;
ScheduleHazardRecognizer *HazardRec;
/// Number of cycles it takes to issue the instructions scheduled in this
/// zone. It is defined as: scheduled-micro-ops / issue-width + stalls.
/// See getStalls().
unsigned CurrCycle;
/// Micro-ops issued in the current cycle
unsigned CurrMOps;
/// MinReadyCycle - Cycle of the soonest available instruction.
unsigned MinReadyCycle;
// The expected latency of the critical path in this scheduled zone.
unsigned ExpectedLatency;
// The latency of dependence chains leading into this zone.
// For each node scheduled bottom-up: DLat = max DLat, N.Depth.
// For each cycle scheduled: DLat -= 1.
unsigned DependentLatency;
/// Count the scheduled (issued) micro-ops that can be retired by
/// time=CurrCycle assuming the first scheduled instr is retired at time=0.
unsigned RetiredMOps;
// Count scheduled resources that have been executed. Resources are
// considered executed if they become ready in the time that it takes to
// saturate any resource including the one in question. Counts are scaled
// for direct comparison with other resources. Counts can be compared with
// MOps * getMicroOpFactor and Latency * getLatencyFactor.
SmallVector<unsigned, 16> ExecutedResCounts;
/// Cache the max count for a single resource.
unsigned MaxExecutedResCount;
// Cache the critical resources ID in this scheduled zone.
unsigned ZoneCritResIdx;
// Is the scheduled region resource limited vs. latency limited.
bool IsResourceLimited;
#ifndef NDEBUG
// Remember the greatest operand latency as an upper bound on the number of
// times we should retry the pending queue because of a hazard.
unsigned MaxObservedLatency;
#endif
void reset() {
// A new HazardRec is created for each DAG and owned by SchedBoundary.
// Destroying and reconstructing it is very expensive though. So keep
// invalid, placeholder HazardRecs.
if (HazardRec && HazardRec->isEnabled()) {
delete HazardRec;
HazardRec = 0;
}
Available.clear();
Pending.clear();
CheckPending = false;
NextSUs.clear();
CurrCycle = 0;
CurrMOps = 0;
MinReadyCycle = UINT_MAX;
ExpectedLatency = 0;
DependentLatency = 0;
RetiredMOps = 0;
MaxExecutedResCount = 0;
ZoneCritResIdx = 0;
IsResourceLimited = false;
#ifndef NDEBUG
MaxObservedLatency = 0;
#endif
// Reserve a zero-count for invalid CritResIdx.
ExecutedResCounts.resize(1);
assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
}
/// Pending queues extend the ready queues with the same ID and the
/// PendingFlag set.
SchedBoundary(unsigned ID, const Twine &Name):
DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
Pending(ID << GenericScheduler::LogMaxQID, Name+".P"),
HazardRec(0) {
reset();
}
~SchedBoundary() { delete HazardRec; }
void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
SchedRemainder *rem);
bool isTop() const {
return Available.getID() == GenericScheduler::TopQID;
}
#ifndef NDEBUG
const char *getResourceName(unsigned PIdx) {
if (!PIdx)
return "MOps";
return SchedModel->getProcResource(PIdx)->Name;
}
#endif
/// Get the number of latency cycles "covered" by the scheduled
/// instructions. This is the larger of the critical path within the zone
/// and the number of cycles required to issue the instructions.
unsigned getScheduledLatency() const {
return std::max(ExpectedLatency, CurrCycle);
}
unsigned getUnscheduledLatency(SUnit *SU) const {
return isTop() ? SU->getHeight() : SU->getDepth();
}
unsigned getResourceCount(unsigned ResIdx) const {
return ExecutedResCounts[ResIdx];
}
/// Get the scaled count of scheduled micro-ops and resources, including
/// executed resources.
unsigned getCriticalCount() const {
if (!ZoneCritResIdx)
return RetiredMOps * SchedModel->getMicroOpFactor();
return getResourceCount(ZoneCritResIdx);
}
/// Get a scaled count for the minimum execution time of the scheduled
/// micro-ops that are ready to execute by getExecutedCount. Notice the
/// feedback loop.
unsigned getExecutedCount() const {
return std::max(CurrCycle * SchedModel->getLatencyFactor(),
MaxExecutedResCount);
}
bool checkHazard(SUnit *SU);
unsigned findMaxLatency(ArrayRef<SUnit*> ReadySUs);
unsigned getOtherResourceCount(unsigned &OtherCritIdx);
void setPolicy(CandPolicy &Policy, SchedBoundary &OtherZone);
void releaseNode(SUnit *SU, unsigned ReadyCycle);
void bumpCycle(unsigned NextCycle);
void incExecutedResources(unsigned PIdx, unsigned Count);
unsigned countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle);
void bumpNode(SUnit *SU);
void releasePending();
void removeReady(SUnit *SU);
SUnit *pickOnlyChoice();
#ifndef NDEBUG
void dumpScheduledState();
#endif
};
private:
const MachineSchedContext *Context;
ScheduleDAGMI *DAG;
const TargetSchedModel *SchedModel;
const TargetRegisterInfo *TRI;
// State of the top and bottom scheduled instruction boundaries.
SchedRemainder Rem;
SchedBoundary Top;
SchedBoundary Bot;
MachineSchedPolicy RegionPolicy;
public:
/// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
enum {
TopQID = 1,
BotQID = 2,
LogMaxQID = 2
};
GenericScheduler(const MachineSchedContext *C):
Context(C), DAG(0), SchedModel(0), TRI(0),
Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
virtual void initPolicy(MachineBasicBlock::iterator Begin,
MachineBasicBlock::iterator End,
unsigned NumRegionInstrs);
bool shouldTrackPressure() const { return RegionPolicy.ShouldTrackPressure; }
virtual void initialize(ScheduleDAGMI *dag);
virtual SUnit *pickNode(bool &IsTopNode);
virtual void schedNode(SUnit *SU, bool IsTopNode);
virtual void releaseTopNode(SUnit *SU);
virtual void releaseBottomNode(SUnit *SU);
virtual void registerRoots();
protected:
void checkAcyclicLatency();
void tryCandidate(SchedCandidate &Cand,
SchedCandidate &TryCand,
SchedBoundary &Zone,
const RegPressureTracker &RPTracker,
RegPressureTracker &TempTracker);
SUnit *pickNodeBidirectional(bool &IsTopNode);
void pickNodeFromQueue(SchedBoundary &Zone,
const RegPressureTracker &RPTracker,
SchedCandidate &Candidate);
void reschedulePhysRegCopies(SUnit *SU, bool isTop);
#ifndef NDEBUG
void traceCandidate(const SchedCandidate &Cand);
#endif
};
} // namespace
void GenericScheduler::SchedRemainder::
init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
reset();
if (!SchedModel->hasInstrSchedModel())
return;
RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
for (std::vector<SUnit>::iterator
I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
* SchedModel->getMicroOpFactor();
for (TargetSchedModel::ProcResIter
PI = SchedModel->getWriteProcResBegin(SC),
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
unsigned PIdx = PI->ProcResourceIdx;
unsigned Factor = SchedModel->getResourceFactor(PIdx);
RemainingCounts[PIdx] += (Factor * PI->Cycles);
}
}
}
void GenericScheduler::SchedBoundary::
init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
reset();
DAG = dag;
SchedModel = smodel;
Rem = rem;
if (SchedModel->hasInstrSchedModel())
ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
}
/// Initialize the per-region scheduling policy.
void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
MachineBasicBlock::iterator End,
unsigned NumRegionInstrs) {
const TargetMachine &TM = Context->MF->getTarget();
// Avoid setting up the register pressure tracker for small regions to save
// compile time. As a rough heuristic, only track pressure when the number of
// schedulable instructions exceeds half the integer register file.
unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
TM.getTargetLowering()->getRegClassFor(MVT::i32));
RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
// For generic targets, we default to bottom-up, because it's simpler and more
// compile-time optimizations have been implemented in that direction.
RegionPolicy.OnlyBottomUp = true;
// Allow the subtarget to override default policy.
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
ST.overrideSchedPolicy(RegionPolicy, Begin, End, NumRegionInstrs);
// After subtarget overrides, apply command line options.
if (!EnableRegPressure)
RegionPolicy.ShouldTrackPressure = false;
// Check -misched-topdown/bottomup can force or unforce scheduling direction.
// e.g. -misched-bottomup=false allows scheduling in both directions.
assert((!ForceTopDown || !ForceBottomUp) &&
"-misched-topdown incompatible with -misched-bottomup");
if (ForceBottomUp.getNumOccurrences() > 0) {
RegionPolicy.OnlyBottomUp = ForceBottomUp;
if (RegionPolicy.OnlyBottomUp)
RegionPolicy.OnlyTopDown = false;
}
if (ForceTopDown.getNumOccurrences() > 0) {
RegionPolicy.OnlyTopDown = ForceTopDown;
if (RegionPolicy.OnlyTopDown)
RegionPolicy.OnlyBottomUp = false;
}
}
void GenericScheduler::initialize(ScheduleDAGMI *dag) {
DAG = dag;
SchedModel = DAG->getSchedModel();
TRI = DAG->TRI;
Rem.init(DAG, SchedModel);
Top.init(DAG, SchedModel, &Rem);
Bot.init(DAG, SchedModel, &Rem);
// Initialize resource counts.
// Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
// are disabled, then these HazardRecs will be disabled.
const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
const TargetMachine &TM = DAG->MF.getTarget();
if (!Top.HazardRec) {
Top.HazardRec =
TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
}
if (!Bot.HazardRec) {
Bot.HazardRec =
TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
}
}
void GenericScheduler::releaseTopNode(SUnit *SU) {
if (SU->isScheduled)
return;
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
if (I->isWeak())
continue;
unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
unsigned Latency = I->getLatency();
#ifndef NDEBUG
Top.MaxObservedLatency = std::max(Latency, Top.MaxObservedLatency);
#endif
if (SU->TopReadyCycle < PredReadyCycle + Latency)
SU->TopReadyCycle = PredReadyCycle + Latency;
}
Top.releaseNode(SU, SU->TopReadyCycle);
}
void GenericScheduler::releaseBottomNode(SUnit *SU) {
if (SU->isScheduled)
return;
assert(SU->getInstr() && "Scheduled SUnit must have instr");
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
if (I->isWeak())
continue;
unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
unsigned Latency = I->getLatency();
#ifndef NDEBUG
Bot.MaxObservedLatency = std::max(Latency, Bot.MaxObservedLatency);
#endif
if (SU->BotReadyCycle < SuccReadyCycle + Latency)
SU->BotReadyCycle = SuccReadyCycle + Latency;
}
Bot.releaseNode(SU, SU->BotReadyCycle);
}
/// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
/// critical path by more cycles than it takes to drain the instruction buffer.
/// We estimate an upper bounds on in-flight instructions as:
///
/// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
/// InFlightIterations = AcyclicPath / CyclesPerIteration
/// InFlightResources = InFlightIterations * LoopResources
///
/// TODO: Check execution resources in addition to IssueCount.
void GenericScheduler::checkAcyclicLatency() {
if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
return;
// Scaled number of cycles per loop iteration.
unsigned IterCount =
std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
Rem.RemIssueCount);
// Scaled acyclic critical path.
unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
// InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
unsigned InFlightCount =
(AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
unsigned BufferLimit =
SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
DEBUG(dbgs() << "IssueCycles="
<< Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
<< "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
<< "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
<< " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
<< "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
if (Rem.IsAcyclicLatencyLimited)
dbgs() << " ACYCLIC LATENCY LIMIT\n");
}
void GenericScheduler::registerRoots() {
Rem.CriticalPath = DAG->ExitSU.getDepth();
// Some roots may not feed into ExitSU. Check all of them in case.
for (std::vector<SUnit*>::const_iterator
I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
if ((*I)->getDepth() > Rem.CriticalPath)
Rem.CriticalPath = (*I)->getDepth();
}
DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
if (EnableCyclicPath) {
Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
checkAcyclicLatency();
}
}
/// Does this SU have a hazard within the current instruction group.
///
/// The scheduler supports two modes of hazard recognition. The first is the
/// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
/// supports highly complicated in-order reservation tables
/// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
///
/// The second is a streamlined mechanism that checks for hazards based on
/// simple counters that the scheduler itself maintains. It explicitly checks
/// for instruction dispatch limitations, including the number of micro-ops that
/// can dispatch per cycle.
///
/// TODO: Also check whether the SU must start a new group.
bool GenericScheduler::SchedBoundary::checkHazard(SUnit *SU) {
if (HazardRec->isEnabled())
return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
<< SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
return true;
}
return false;
}
// Find the unscheduled node in ReadySUs with the highest latency.
unsigned GenericScheduler::SchedBoundary::
findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
SUnit *LateSU = 0;
unsigned RemLatency = 0;
for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
I != E; ++I) {
unsigned L = getUnscheduledLatency(*I);
if (L > RemLatency) {
RemLatency = L;
LateSU = *I;
}
}
if (LateSU) {
DEBUG(dbgs() << Available.getName() << " RemLatency SU("
<< LateSU->NodeNum << ") " << RemLatency << "c\n");
}
return RemLatency;
}
// Count resources in this zone and the remaining unscheduled
// instruction. Return the max count, scaled. Set OtherCritIdx to the critical
// resource index, or zero if the zone is issue limited.
unsigned GenericScheduler::SchedBoundary::
getOtherResourceCount(unsigned &OtherCritIdx) {
OtherCritIdx = 0;
if (!SchedModel->hasInstrSchedModel())
return 0;
unsigned OtherCritCount = Rem->RemIssueCount
+ (RetiredMOps * SchedModel->getMicroOpFactor());
DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
<< OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
PIdx != PEnd; ++PIdx) {
unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
if (OtherCount > OtherCritCount) {
OtherCritCount = OtherCount;
OtherCritIdx = PIdx;
}
}
if (OtherCritIdx) {
DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
<< OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
<< " " << getResourceName(OtherCritIdx) << "\n");
}
return OtherCritCount;
}
/// Set the CandPolicy for this zone given the current resources and latencies
/// inside and outside the zone.
void GenericScheduler::SchedBoundary::setPolicy(CandPolicy &Policy,
SchedBoundary &OtherZone) {
// Now that potential stalls have been considered, apply preemptive heuristics
// based on the the total latency and resources inside and outside this
// zone.
// Compute remaining latency. We need this both to determine whether the
// overall schedule has become latency-limited and whether the instructions
// outside this zone are resource or latency limited.
//
// The "dependent" latency is updated incrementally during scheduling as the
// max height/depth of scheduled nodes minus the cycles since it was
// scheduled:
// DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
//
// The "independent" latency is the max ready queue depth:
// ILat = max N.depth for N in Available|Pending
//
// RemainingLatency is the greater of independent and dependent latency.
unsigned RemLatency = DependentLatency;
RemLatency = std::max(RemLatency, findMaxLatency(Available.elements()));
RemLatency = std::max(RemLatency, findMaxLatency(Pending.elements()));
// Compute the critical resource outside the zone.
unsigned OtherCritIdx;
unsigned OtherCount = OtherZone.getOtherResourceCount(OtherCritIdx);
bool OtherResLimited = false;
if (SchedModel->hasInstrSchedModel()) {
unsigned LFactor = SchedModel->getLatencyFactor();
OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
}
if (!OtherResLimited && (RemLatency + CurrCycle > Rem->CriticalPath)) {
Policy.ReduceLatency |= true;
DEBUG(dbgs() << " " << Available.getName() << " RemainingLatency "
<< RemLatency << " + " << CurrCycle << "c > CritPath "
<< Rem->CriticalPath << "\n");
}
// If the same resource is limiting inside and outside the zone, do nothing.
if (ZoneCritResIdx == OtherCritIdx)
return;
DEBUG(
if (IsResourceLimited) {
dbgs() << " " << Available.getName() << " ResourceLimited: "
<< getResourceName(ZoneCritResIdx) << "\n";
}
if (OtherResLimited)
dbgs() << " RemainingLimit: " << getResourceName(OtherCritIdx) << "\n";
if (!IsResourceLimited && !OtherResLimited)
dbgs() << " Latency limited both directions.\n");
if (IsResourceLimited && !Policy.ReduceResIdx)
Policy.ReduceResIdx = ZoneCritResIdx;
if (OtherResLimited)
Policy.DemandResIdx = OtherCritIdx;
}
void GenericScheduler::SchedBoundary::releaseNode(SUnit *SU,
unsigned ReadyCycle) {
if (ReadyCycle < MinReadyCycle)
MinReadyCycle = ReadyCycle;
// Check for interlocks first. For the purpose of other heuristics, an
// instruction that cannot issue appears as if it's not in the ReadyQueue.
bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU))
Pending.push(SU);
else
Available.push(SU);
// Record this node as an immediate dependent of the scheduled node.
NextSUs.insert(SU);
}
/// Move the boundary of scheduled code by one cycle.
void GenericScheduler::SchedBoundary::bumpCycle(unsigned NextCycle) {
if (SchedModel->getMicroOpBufferSize() == 0) {
assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
if (MinReadyCycle > NextCycle)
NextCycle = MinReadyCycle;
}
// Update the current micro-ops, which will issue in the next cycle.
unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
// Decrement DependentLatency based on the next cycle.
if ((NextCycle - CurrCycle) > DependentLatency)
DependentLatency = 0;
else
DependentLatency -= (NextCycle - CurrCycle);
if (!HazardRec->isEnabled()) {
// Bypass HazardRec virtual calls.
CurrCycle = NextCycle;
}
else {
// Bypass getHazardType calls in case of long latency.
for (; CurrCycle != NextCycle; ++CurrCycle) {
if (isTop())
HazardRec->AdvanceCycle();
else
HazardRec->RecedeCycle();
}
}
CheckPending = true;
unsigned LFactor = SchedModel->getLatencyFactor();
IsResourceLimited =
(int)(getCriticalCount() - (getScheduledLatency() * LFactor))
> (int)LFactor;
DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
}
void GenericScheduler::SchedBoundary::incExecutedResources(unsigned PIdx,
unsigned Count) {
ExecutedResCounts[PIdx] += Count;
if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
MaxExecutedResCount = ExecutedResCounts[PIdx];
}
/// Add the given processor resource to this scheduled zone.
///
/// \param Cycles indicates the number of consecutive (non-pipelined) cycles
/// during which this resource is consumed.
///
/// \return the next cycle at which the instruction may execute without
/// oversubscribing resources.
unsigned GenericScheduler::SchedBoundary::
countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle) {
unsigned Factor = SchedModel->getResourceFactor(PIdx);
unsigned Count = Factor * Cycles;
DEBUG(dbgs() << " " << getResourceName(PIdx)
<< " +" << Cycles << "x" << Factor << "u\n");
// Update Executed resources counts.
incExecutedResources(PIdx, Count);
assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
Rem->RemainingCounts[PIdx] -= Count;
// Check if this resource exceeds the current critical resource. If so, it
// becomes the critical resource.
if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
ZoneCritResIdx = PIdx;
DEBUG(dbgs() << " *** Critical resource "
<< getResourceName(PIdx) << ": "
<< getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
}
// TODO: We don't yet model reserved resources. It's not hard though.
return CurrCycle;
}
/// Move the boundary of scheduled code by one SUnit.
void GenericScheduler::SchedBoundary::bumpNode(SUnit *SU) {
// Update the reservation table.
if (HazardRec->isEnabled()) {
if (!isTop() && SU->isCall) {
// Calls are scheduled with their preceding instructions. For bottom-up
// scheduling, clear the pipeline state before emitting.
HazardRec->Reset();
}
HazardRec->EmitInstruction(SU);
}
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
CurrMOps += IncMOps;
// checkHazard prevents scheduling multiple instructions per cycle that exceed
// issue width. However, we commonly reach the maximum. In this case
// opportunistically bump the cycle to avoid uselessly checking everything in
// the readyQ. Furthermore, a single instruction may produce more than one
// cycle's worth of micro-ops.
//
// TODO: Also check if this SU must end a dispatch group.
unsigned NextCycle = CurrCycle;
if (CurrMOps >= SchedModel->getIssueWidth()) {
++NextCycle;
DEBUG(dbgs() << " *** Max MOps " << CurrMOps
<< " at cycle " << CurrCycle << '\n');
}
unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
switch (SchedModel->getMicroOpBufferSize()) {
case 0:
assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
break;
case 1:
if (ReadyCycle > NextCycle) {
NextCycle = ReadyCycle;
DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
}
break;
default:
// We don't currently model the OOO reorder buffer, so consider all
// scheduled MOps to be "retired".
break;
}
RetiredMOps += IncMOps;
// Update resource counts and critical resource.
if (SchedModel->hasInstrSchedModel()) {
unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
Rem->RemIssueCount -= DecRemIssue;
if (ZoneCritResIdx) {
// Scale scheduled micro-ops for comparing with the critical resource.
unsigned ScaledMOps =
RetiredMOps * SchedModel->getMicroOpFactor();
// If scaled micro-ops are now more than the previous critical resource by
// a full cycle, then micro-ops issue becomes critical.
if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
>= (int)SchedModel->getLatencyFactor()) {
ZoneCritResIdx = 0;
DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
<< ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
}
}
for (TargetSchedModel::ProcResIter
PI = SchedModel->getWriteProcResBegin(SC),
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
unsigned RCycle =
countResource(PI->ProcResourceIdx, PI->Cycles, ReadyCycle);
if (RCycle > NextCycle)
NextCycle = RCycle;
}
}
// Update ExpectedLatency and DependentLatency.
unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
if (SU->getDepth() > TopLatency) {
TopLatency = SU->getDepth();
DEBUG(dbgs() << " " << Available.getName()
<< " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
}
if (SU->getHeight() > BotLatency) {
BotLatency = SU->getHeight();
DEBUG(dbgs() << " " << Available.getName()
<< " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
}
// If we stall for any reason, bump the cycle.
if (NextCycle > CurrCycle) {
bumpCycle(NextCycle);
}
else {
// After updating ZoneCritResIdx and ExpectedLatency, check if we're
// resource limited. If a stall occured, bumpCycle does this.
unsigned LFactor = SchedModel->getLatencyFactor();
IsResourceLimited =
(int)(getCriticalCount() - (getScheduledLatency() * LFactor))
> (int)LFactor;
}
DEBUG(dumpScheduledState());
}
/// Release pending ready nodes in to the available queue. This makes them
/// visible to heuristics.
void GenericScheduler::SchedBoundary::releasePending() {
// If the available queue is empty, it is safe to reset MinReadyCycle.
if (Available.empty())
MinReadyCycle = UINT_MAX;
// Check to see if any of the pending instructions are ready to issue. If
// so, add them to the available queue.
bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
SUnit *SU = *(Pending.begin()+i);
unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
if (ReadyCycle < MinReadyCycle)
MinReadyCycle = ReadyCycle;
if (!IsBuffered && ReadyCycle > CurrCycle)
continue;
if (checkHazard(SU))
continue;
Available.push(SU);
Pending.remove(Pending.begin()+i);
--i; --e;
}
DEBUG(if (!Pending.empty()) Pending.dump());
CheckPending = false;
}
/// Remove SU from the ready set for this boundary.
void GenericScheduler::SchedBoundary::removeReady(SUnit *SU) {
if (Available.isInQueue(SU))
Available.remove(Available.find(SU));
else {
assert(Pending.isInQueue(SU) && "bad ready count");
Pending.remove(Pending.find(SU));
}
}
/// If this queue only has one ready candidate, return it. As a side effect,
/// defer any nodes that now hit a hazard, and advance the cycle until at least
/// one node is ready. If multiple instructions are ready, return NULL.
SUnit *GenericScheduler::SchedBoundary::pickOnlyChoice() {
if (CheckPending)
releasePending();
if (CurrMOps > 0) {
// Defer any ready instrs that now have a hazard.
for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
if (checkHazard(*I)) {
Pending.push(*I);
I = Available.remove(I);
continue;
}
++I;
}
}
for (unsigned i = 0; Available.empty(); ++i) {
assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
"permanent hazard"); (void)i;
bumpCycle(CurrCycle + 1);
releasePending();
}
if (Available.size() == 1)
return *Available.begin();
return NULL;
}
#ifndef NDEBUG
// This is useful information to dump after bumpNode.
// Note that the Queue contents are more useful before pickNodeFromQueue.
void GenericScheduler::SchedBoundary::dumpScheduledState() {
unsigned ResFactor;
unsigned ResCount;
if (ZoneCritResIdx) {
ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
ResCount = getResourceCount(ZoneCritResIdx);
}
else {
ResFactor = SchedModel->getMicroOpFactor();
ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
}
unsigned LFactor = SchedModel->getLatencyFactor();
dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
<< " Retired: " << RetiredMOps;
dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
<< ResCount / ResFactor << " " << getResourceName(ZoneCritResIdx)
<< "\n ExpectedLatency: " << ExpectedLatency << "c\n"
<< (IsResourceLimited ? " - Resource" : " - Latency")
<< " limited.\n";
}
#endif
void GenericScheduler::SchedCandidate::
initResourceDelta(const ScheduleDAGMI *DAG,
const TargetSchedModel *SchedModel) {
if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
return;
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
for (TargetSchedModel::ProcResIter
PI = SchedModel->getWriteProcResBegin(SC),
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
if (PI->ProcResourceIdx == Policy.ReduceResIdx)
ResDelta.CritResources += PI->Cycles;
if (PI->ProcResourceIdx == Policy.DemandResIdx)
ResDelta.DemandedResources += PI->Cycles;
}
}
/// Return true if this heuristic determines order.
static bool tryLess(int TryVal, int CandVal,
GenericScheduler::SchedCandidate &TryCand,
GenericScheduler::SchedCandidate &Cand,
GenericScheduler::CandReason Reason) {
if (TryVal < CandVal) {
TryCand.Reason = Reason;
return true;
}
if (TryVal > CandVal) {
if (Cand.Reason > Reason)
Cand.Reason = Reason;
return true;
}
Cand.setRepeat(Reason);
return false;
}
static bool tryGreater(int TryVal, int CandVal,
GenericScheduler::SchedCandidate &TryCand,
GenericScheduler::SchedCandidate &Cand,
GenericScheduler::CandReason Reason) {
if (TryVal > CandVal) {
TryCand.Reason = Reason;
return true;
}
if (TryVal < CandVal) {
if (Cand.Reason > Reason)
Cand.Reason = Reason;
return true;
}
Cand.setRepeat(Reason);
return false;
}
static bool tryPressure(const PressureChange &TryP,
const PressureChange &CandP,
GenericScheduler::SchedCandidate &TryCand,
GenericScheduler::SchedCandidate &Cand,
GenericScheduler::CandReason Reason) {
int TryRank = TryP.getPSetOrMax();
int CandRank = CandP.getPSetOrMax();
// If both candidates affect the same set, go with the smallest increase.
if (TryRank == CandRank) {
return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
Reason);
}
// If one candidate decreases and the other increases, go with it.
// Invalid candidates have UnitInc==0.
if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
Reason)) {
return true;
}
// If the candidates are decreasing pressure, reverse priority.
if (TryP.getUnitInc() < 0)
std::swap(TryRank, CandRank);
return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
}
static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
}
/// Minimize physical register live ranges. Regalloc wants them adjacent to
/// their physreg def/use.
///
/// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
/// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
/// with the operation that produces or consumes the physreg. We'll do this when
/// regalloc has support for parallel copies.
static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
const MachineInstr *MI = SU->getInstr();
if (!MI->isCopy())
return 0;
unsigned ScheduledOper = isTop ? 1 : 0;
unsigned UnscheduledOper = isTop ? 0 : 1;
// If we have already scheduled the physreg produce/consumer, immediately
// schedule the copy.
if (TargetRegisterInfo::isPhysicalRegister(
MI->getOperand(ScheduledOper).getReg()))
return 1;
// If the physreg is at the boundary, defer it. Otherwise schedule it
// immediately to free the dependent. We can hoist the copy later.
bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
if (TargetRegisterInfo::isPhysicalRegister(
MI->getOperand(UnscheduledOper).getReg()))
return AtBoundary ? -1 : 1;
return 0;
}
static bool tryLatency(GenericScheduler::SchedCandidate &TryCand,
GenericScheduler::SchedCandidate &Cand,
GenericScheduler::SchedBoundary &Zone) {
if (Zone.isTop()) {
if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
TryCand, Cand, GenericScheduler::TopDepthReduce))
return true;
}
if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
TryCand, Cand, GenericScheduler::TopPathReduce))
return true;
}
else {
if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
TryCand, Cand, GenericScheduler::BotHeightReduce))
return true;
}
if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
TryCand, Cand, GenericScheduler::BotPathReduce))
return true;
}
return false;
}
/// Apply a set of heursitics to a new candidate. Heuristics are currently
/// hierarchical. This may be more efficient than a graduated cost model because
/// we don't need to evaluate all aspects of the model for each node in the
/// queue. But it's really done to make the heuristics easier to debug and
/// statistically analyze.
///
/// \param Cand provides the policy and current best candidate.
/// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
/// \param Zone describes the scheduled zone that we are extending.
/// \param RPTracker describes reg pressure within the scheduled zone.
/// \param TempTracker is a scratch pressure tracker to reuse in queries.
void GenericScheduler::tryCandidate(SchedCandidate &Cand,
SchedCandidate &TryCand,
SchedBoundary &Zone,
const RegPressureTracker &RPTracker,
RegPressureTracker &TempTracker) {
if (DAG->isTrackingPressure()) {
// Always initialize TryCand's RPDelta.
if (Zone.isTop()) {
TempTracker.getMaxDownwardPressureDelta(
TryCand.SU->getInstr(),
TryCand.RPDelta,
DAG->getRegionCriticalPSets(),
DAG->getRegPressure().MaxSetPressure);
}
else {
if (VerifyScheduling) {
TempTracker.getMaxUpwardPressureDelta(
TryCand.SU->getInstr(),
&DAG->getPressureDiff(TryCand.SU),
TryCand.RPDelta,
DAG->getRegionCriticalPSets(),
DAG->getRegPressure().MaxSetPressure);
}
else {
RPTracker.getUpwardPressureDelta(
TryCand.SU->getInstr(),
DAG->getPressureDiff(TryCand.SU),
TryCand.RPDelta,
DAG->getRegionCriticalPSets(),
DAG->getRegPressure().MaxSetPressure);
}
}
}
DEBUG(if (TryCand.RPDelta.Excess.isValid())
dbgs() << " SU(" << TryCand.SU->NodeNum << ") "
<< TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet())
<< ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n");
// Initialize the candidate if needed.
if (!Cand.isValid()) {
TryCand.Reason = NodeOrder;
return;
}
if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
biasPhysRegCopy(Cand.SU, Zone.isTop()),
TryCand, Cand, PhysRegCopy))
return;
// Avoid exceeding the target's limit. If signed PSetID is negative, it is
// invalid; convert it to INT_MAX to give it lowest priority.
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
Cand.RPDelta.Excess,
TryCand, Cand, RegExcess))
return;
// Avoid increasing the max critical pressure in the scheduled region.
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
Cand.RPDelta.CriticalMax,
TryCand, Cand, RegCritical))
return;
// For loops that are acyclic path limited, aggressively schedule for latency.
// This can result in very long dependence chains scheduled in sequence, so
// once every cycle (when CurrMOps == 0), switch to normal heuristics.
if (Rem.IsAcyclicLatencyLimited && !Zone.CurrMOps
&& tryLatency(TryCand, Cand, Zone))
return;
// Keep clustered nodes together to encourage downstream peephole
// optimizations which may reduce resource requirements.
//
// This is a best effort to set things up for a post-RA pass. Optimizations
// like generating loads of multiple registers should ideally be done within
// the scheduler pass by combining the loads during DAG postprocessing.
const SUnit *NextClusterSU =
Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
TryCand, Cand, Cluster))
return;
// Weak edges are for clustering and other constraints.
if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
getWeakLeft(Cand.SU, Zone.isTop()),
TryCand, Cand, Weak)) {
return;
}
// Avoid increasing the max pressure of the entire region.
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
Cand.RPDelta.CurrentMax,
TryCand, Cand, RegMax))
return;
// Avoid critical resource consumption and balance the schedule.
TryCand.initResourceDelta(DAG, SchedModel);
if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
TryCand, Cand, ResourceReduce))
return;
if (tryGreater(TryCand.ResDelta.DemandedResources,
Cand.ResDelta.DemandedResources,
TryCand, Cand, ResourceDemand))
return;
// Avoid serializing long latency dependence chains.
// For acyclic path limited loops, latency was already checked above.
if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
&& tryLatency(TryCand, Cand, Zone)) {
return;
}
// Prefer immediate defs/users of the last scheduled instruction. This is a
// local pressure avoidance strategy that also makes the machine code
// readable.
if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
TryCand, Cand, NextDefUse))
return;
// Fall through to original instruction order.
if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
|| (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
TryCand.Reason = NodeOrder;
}
}
#ifndef NDEBUG
const char *GenericScheduler::getReasonStr(
GenericScheduler::CandReason Reason) {
switch (Reason) {
case NoCand: return "NOCAND ";
case PhysRegCopy: return "PREG-COPY";
case RegExcess: return "REG-EXCESS";
case RegCritical: return "REG-CRIT ";
case Cluster: return "CLUSTER ";
case Weak: return "WEAK ";
case RegMax: return "REG-MAX ";
case ResourceReduce: return "RES-REDUCE";
case ResourceDemand: return "RES-DEMAND";
case TopDepthReduce: return "TOP-DEPTH ";
case TopPathReduce: return "TOP-PATH ";
case BotHeightReduce:return "BOT-HEIGHT";
case BotPathReduce: return "BOT-PATH ";
case NextDefUse: return "DEF-USE ";
case NodeOrder: return "ORDER ";
};
llvm_unreachable("Unknown reason!");
}
void GenericScheduler::traceCandidate(const SchedCandidate &Cand) {
PressureChange P;
unsigned ResIdx = 0;
unsigned Latency = 0;
switch (Cand.Reason) {
default:
break;
case RegExcess:
P = Cand.RPDelta.Excess;
break;
case RegCritical:
P = Cand.RPDelta.CriticalMax;
break;
case RegMax:
P = Cand.RPDelta.CurrentMax;
break;
case ResourceReduce:
ResIdx = Cand.Policy.ReduceResIdx;
break;
case ResourceDemand:
ResIdx = Cand.Policy.DemandResIdx;
break;
case TopDepthReduce:
Latency = Cand.SU->getDepth();
break;
case TopPathReduce:
Latency = Cand.SU->getHeight();
break;
case BotHeightReduce:
Latency = Cand.SU->getHeight();
break;
case BotPathReduce:
Latency = Cand.SU->getDepth();
break;
}
dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
if (P.isValid())
dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
<< ":" << P.getUnitInc() << " ";
else
dbgs() << " ";
if (ResIdx)
dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
else
dbgs() << " ";
if (Latency)
dbgs() << " " << Latency << " cycles ";
else
dbgs() << " ";
dbgs() << '\n';
}
#endif
/// Pick the best candidate from the queue.
///
/// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
/// DAG building. To adjust for the current scheduling location we need to
/// maintain the number of vreg uses remaining to be top-scheduled.
void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
const RegPressureTracker &RPTracker,
SchedCandidate &Cand) {
ReadyQueue &Q = Zone.Available;
DEBUG(Q.dump());
// getMaxPressureDelta temporarily modifies the tracker.
RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
SchedCandidate TryCand(Cand.Policy);
TryCand.SU = *I;
tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
if (TryCand.Reason != NoCand) {
// Initialize resource delta if needed in case future heuristics query it.
if (TryCand.ResDelta == SchedResourceDelta())
TryCand.initResourceDelta(DAG, SchedModel);
Cand.setBest(TryCand);
DEBUG(traceCandidate(Cand));
}
}
}
static void tracePick(const GenericScheduler::SchedCandidate &Cand,
bool IsTop) {
DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
<< GenericScheduler::getReasonStr(Cand.Reason) << '\n');
}
/// Pick the best candidate node from either the top or bottom queue.
SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
// Schedule as far as possible in the direction of no choice. This is most
// efficient, but also provides the best heuristics for CriticalPSets.
if (SUnit *SU = Bot.pickOnlyChoice()) {
IsTopNode = false;
DEBUG(dbgs() << "Pick Bot NOCAND\n");
return SU;
}
if (SUnit *SU = Top.pickOnlyChoice()) {
IsTopNode = true;
DEBUG(dbgs() << "Pick Top NOCAND\n");
return SU;
}
CandPolicy NoPolicy;
SchedCandidate BotCand(NoPolicy);
SchedCandidate TopCand(NoPolicy);
Bot.setPolicy(BotCand.Policy, Top);
Top.setPolicy(TopCand.Policy, Bot);
// Prefer bottom scheduling when heuristics are silent.
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
assert(BotCand.Reason != NoCand && "failed to find the first candidate");
// If either Q has a single candidate that provides the least increase in
// Excess pressure, we can immediately schedule from that Q.
//
// RegionCriticalPSets summarizes the pressure within the scheduled region and
// affects picking from either Q. If scheduling in one direction must
// increase pressure for one of the excess PSets, then schedule in that
// direction first to provide more freedom in the other direction.
if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
|| (BotCand.Reason == RegCritical
&& !BotCand.isRepeat(RegCritical)))
{
IsTopNode = false;
tracePick(BotCand, IsTopNode);
return BotCand.SU;
}
// Check if the top Q has a better candidate.
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
assert(TopCand.Reason != NoCand && "failed to find the first candidate");
// Choose the queue with the most important (lowest enum) reason.
if (TopCand.Reason < BotCand.Reason) {
IsTopNode = true;
tracePick(TopCand, IsTopNode);
return TopCand.SU;
}
// Otherwise prefer the bottom candidate, in node order if all else failed.
IsTopNode = false;
tracePick(BotCand, IsTopNode);
return BotCand.SU;
}
/// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
if (DAG->top() == DAG->bottom()) {
assert(Top.Available.empty() && Top.Pending.empty() &&
Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
return NULL;
}
SUnit *SU;
do {
if (RegionPolicy.OnlyTopDown) {
SU = Top.pickOnlyChoice();
if (!SU) {
CandPolicy NoPolicy;
SchedCandidate TopCand(NoPolicy);
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
assert(TopCand.Reason != NoCand && "failed to find a candidate");
tracePick(TopCand, true);
SU = TopCand.SU;
}
IsTopNode = true;
}
else if (RegionPolicy.OnlyBottomUp) {
SU = Bot.pickOnlyChoice();
if (!SU) {
CandPolicy NoPolicy;
SchedCandidate BotCand(NoPolicy);
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
assert(BotCand.Reason != NoCand && "failed to find a candidate");
tracePick(BotCand, false);
SU = BotCand.SU;
}
IsTopNode = false;
}
else {
SU = pickNodeBidirectional(IsTopNode);
}
} while (SU->isScheduled);
if (SU->isTopReady())
Top.removeReady(SU);
if (SU->isBottomReady())
Bot.removeReady(SU);
DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
return SU;
}
void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
MachineBasicBlock::iterator InsertPos = SU->getInstr();
if (!isTop)
++InsertPos;
SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
// Find already scheduled copies with a single physreg dependence and move
// them just above the scheduled instruction.
for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
I != E; ++I) {
if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
continue;
SUnit *DepSU = I->getSUnit();
if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
continue;
MachineInstr *Copy = DepSU->getInstr();
if (!Copy->isCopy())
continue;
DEBUG(dbgs() << " Rescheduling physreg copy ";
I->getSUnit()->dump(DAG));
DAG->moveInstruction(Copy, InsertPos);
}
}
/// Update the scheduler's state after scheduling a node. This is the same node
/// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
/// it's state based on the current cycle before MachineSchedStrategy does.
///
/// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
/// them here. See comments in biasPhysRegCopy.
void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
if (IsTopNode) {
SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.CurrCycle);
Top.bumpNode(SU);
if (SU->hasPhysRegUses)
reschedulePhysRegCopies(SU, true);
}
else {
SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.CurrCycle);
Bot.bumpNode(SU);
if (SU->hasPhysRegDefs)
reschedulePhysRegCopies(SU, false);
}
}
/// Create the standard converging machine scheduler. This will be used as the
/// default scheduler if the target does not set a default.
static ScheduleDAGInstrs *createGenericSched(MachineSchedContext *C) {
ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new GenericScheduler(C));
// Register DAG post-processors.
//
// FIXME: extend the mutation API to allow earlier mutations to instantiate
// data and pass it to later mutations. Have a single mutation that gathers
// the interesting nodes in one pass.
DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI));
if (EnableLoadCluster && DAG->TII->enableClusterLoads())
DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
if (EnableMacroFusion)
DAG->addMutation(new MacroFusion(DAG->TII));
return DAG;
}
static MachineSchedRegistry
GenericSchedRegistry("converge", "Standard converging scheduler.",
createGenericSched);
//===----------------------------------------------------------------------===//
// ILP Scheduler. Currently for experimental analysis of heuristics.
//===----------------------------------------------------------------------===//
namespace {
/// \brief Order nodes by the ILP metric.
struct ILPOrder {
const SchedDFSResult *DFSResult;
const BitVector *ScheduledTrees;
bool MaximizeILP;
ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
/// \brief Apply a less-than relation on node priority.
///
/// (Return true if A comes after B in the Q.)
bool operator()(const SUnit *A, const SUnit *B) const {
unsigned SchedTreeA = DFSResult->getSubtreeID(A);
unsigned SchedTreeB = DFSResult->getSubtreeID(B);
if (SchedTreeA != SchedTreeB) {
// Unscheduled trees have lower priority.
if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
return ScheduledTrees->test(SchedTreeB);
// Trees with shallower connections have have lower priority.
if (DFSResult->getSubtreeLevel(SchedTreeA)
!= DFSResult->getSubtreeLevel(SchedTreeB)) {
return DFSResult->getSubtreeLevel(SchedTreeA)
< DFSResult->getSubtreeLevel(SchedTreeB);
}
}
if (MaximizeILP)
return DFSResult->getILP(A) < DFSResult->getILP(B);
else
return DFSResult->getILP(A) > DFSResult->getILP(B);
}
};
/// \brief Schedule based on the ILP metric.
class ILPScheduler : public MachineSchedStrategy {
ScheduleDAGMI *DAG;
ILPOrder Cmp;
std::vector<SUnit*> ReadyQ;
public:
ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
virtual void initialize(ScheduleDAGMI *dag) {
DAG = dag;
DAG->computeDFSResult();
Cmp.DFSResult = DAG->getDFSResult();
Cmp.ScheduledTrees = &DAG->getScheduledTrees();
ReadyQ.clear();
}
virtual void registerRoots() {
// Restore the heap in ReadyQ with the updated DFS results.
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
}
/// Implement MachineSchedStrategy interface.
/// -----------------------------------------
/// Callback to select the highest priority node from the ready Q.
virtual SUnit *pickNode(bool &IsTopNode) {
if (ReadyQ.empty()) return NULL;
std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
SUnit *SU = ReadyQ.back();
ReadyQ.pop_back();
IsTopNode = false;
DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
<< " ILP: " << DAG->getDFSResult()->getILP(SU)
<< " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
<< DAG->getDFSResult()->getSubtreeLevel(
DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
<< "Scheduling " << *SU->getInstr());
return SU;
}
/// \brief Scheduler callback to notify that a new subtree is scheduled.
virtual void scheduleTree(unsigned SubtreeID) {
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
}
/// Callback after a node is scheduled. Mark a newly scheduled tree, notify
/// DFSResults, and resort the priority Q.
virtual void schedNode(SUnit *SU, bool IsTopNode) {
assert(!IsTopNode && "SchedDFSResult needs bottom-up");
}
virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
virtual void releaseBottomNode(SUnit *SU) {
ReadyQ.push_back(SU);
std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
}
};
} // namespace
static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
return new ScheduleDAGMI(C, new ILPScheduler(true));
}
static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
return new ScheduleDAGMI(C, new ILPScheduler(false));
}
static MachineSchedRegistry ILPMaxRegistry(
"ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
static MachineSchedRegistry ILPMinRegistry(
"ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
//===----------------------------------------------------------------------===//
// Machine Instruction Shuffler for Correctness Testing
//===----------------------------------------------------------------------===//
#ifndef NDEBUG
namespace {
/// Apply a less-than relation on the node order, which corresponds to the
/// instruction order prior to scheduling. IsReverse implements greater-than.
template<bool IsReverse>
struct SUnitOrder {
bool operator()(SUnit *A, SUnit *B) const {
if (IsReverse)
return A->NodeNum > B->NodeNum;
else
return A->NodeNum < B->NodeNum;
}
};
/// Reorder instructions as much as possible.
class InstructionShuffler : public MachineSchedStrategy {
bool IsAlternating;
bool IsTopDown;
// Using a less-than relation (SUnitOrder<false>) for the TopQ priority
// gives nodes with a higher number higher priority causing the latest
// instructions to be scheduled first.
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
TopQ;
// When scheduling bottom-up, use greater-than as the queue priority.
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
BottomQ;
public:
InstructionShuffler(bool alternate, bool topdown)
: IsAlternating(alternate), IsTopDown(topdown) {}
virtual void initialize(ScheduleDAGMI *) {
TopQ.clear();
BottomQ.clear();
}
/// Implement MachineSchedStrategy interface.
/// -----------------------------------------
virtual SUnit *pickNode(bool &IsTopNode) {
SUnit *SU;
if (IsTopDown) {
do {
if (TopQ.empty()) return NULL;
SU = TopQ.top();
TopQ.pop();
} while (SU->isScheduled);
IsTopNode = true;
}
else {
do {
if (BottomQ.empty()) return NULL;
SU = BottomQ.top();
BottomQ.pop();
} while (SU->isScheduled);
IsTopNode = false;
}
if (IsAlternating)
IsTopDown = !IsTopDown;
return SU;
}
virtual void schedNode(SUnit *SU, bool IsTopNode) {}
virtual void releaseTopNode(SUnit *SU) {
TopQ.push(SU);
}
virtual void releaseBottomNode(SUnit *SU) {
BottomQ.push(SU);
}
};
} // namespace
static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
bool Alternate = !ForceTopDown && !ForceBottomUp;
bool TopDown = !ForceBottomUp;
assert((TopDown || !ForceTopDown) &&
"-misched-topdown incompatible with -misched-bottomup");
return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
}
static MachineSchedRegistry ShufflerRegistry(
"shuffle", "Shuffle machine instructions alternating directions",
createInstructionShuffler);
#endif // !NDEBUG
//===----------------------------------------------------------------------===//
// GraphWriter support for ScheduleDAGMI.
//===----------------------------------------------------------------------===//
#ifndef NDEBUG
namespace llvm {
template<> struct GraphTraits<
ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
template<>
struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
static std::string getGraphName(const ScheduleDAG *G) {
return G->MF.getName();
}
static bool renderGraphFromBottomUp() {
return true;
}
static bool isNodeHidden(const SUnit *Node) {
return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
}
static bool hasNodeAddressLabel(const SUnit *Node,
const ScheduleDAG *Graph) {
return false;
}
/// If you want to override the dot attributes printed for a particular
/// edge, override this method.
static std::string getEdgeAttributes(const SUnit *Node,
SUnitIterator EI,
const ScheduleDAG *Graph) {
if (EI.isArtificialDep())
return "color=cyan,style=dashed";
if (EI.isCtrlDep())
return "color=blue,style=dashed";
return "";
}
static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
std::string Str;
raw_string_ostream SS(Str);
const SchedDFSResult *DFS =
static_cast<const ScheduleDAGMI*>(G)->getDFSResult();
SS << "SU:" << SU->NodeNum;
if (DFS)
SS << " I:" << DFS->getNumInstrs(SU);
return SS.str();
}
static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
return G->getGraphNodeLabel(SU);
}
static std::string getNodeAttributes(const SUnit *N,
const ScheduleDAG *Graph) {
std::string Str("shape=Mrecord");
const SchedDFSResult *DFS =
static_cast<const ScheduleDAGMI*>(Graph)->getDFSResult();
if (DFS) {
Str += ",style=filled,fillcolor=\"#";
Str += DOT::getColorString(DFS->getSubtreeID(N));
Str += '"';
}
return Str;
}
};
} // namespace llvm
#endif // NDEBUG
/// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
/// rendered using 'dot'.
///
void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
#ifndef NDEBUG
ViewGraph(this, Name, false, Title);
#else
errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
<< "systems with Graphviz or gv!\n";
#endif // NDEBUG
}
/// Out-of-line implementation with no arguments is handy for gdb.
void ScheduleDAGMI::viewGraph() {
viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());
}
|