1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625
7626
7627
7628
7629
7630
7631
7632
7633
7634
7635
7636
7637
7638
7639
7640
7641
7642
7643
7644
7645
7646
7647
7648
7649
7650
7651
7652
7653
7654
7655
7656
7657
7658
7659
7660
7661
7662
7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
7676
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
7747
7748
7749
7750
7751
7752
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
7769
7770
7771
7772
7773
7774
7775
7776
7777
7778
7779
7780
7781
7782
7783
7784
7785
7786
7787
7788
7789
7790
7791
7792
7793
7794
7795
7796
7797
7798
7799
7800
7801
7802
7803
7804
7805
7806
7807
7808
7809
7810
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
7846
7847
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860
7861
7862
7863
7864
7865
7866
7867
7868
7869
7870
7871
7872
7873
7874
7875
7876
7877
7878
7879
7880
7881
7882
7883
7884
7885
7886
7887
7888
7889
7890
7891
7892
7893
7894
7895
7896
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927
7928
7929
7930
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944
7945
7946
7947
7948
7949
7950
7951
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
7976
7977
7978
7979
7980
7981
7982
7983
7984
7985
7986
7987
7988
7989
7990
7991
7992
7993
7994
7995
7996
7997
7998
7999
8000
8001
8002
8003
8004
8005
8006
8007
8008
8009
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
8026
8027
8028
8029
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
8046
8047
8048
8049
8050
8051
8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
8077
8078
8079
8080
8081
8082
8083
8084
8085
8086
8087
8088
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
8113
8114
8115
8116
8117
8118
8119
8120
8121
8122
8123
8124
8125
8126
8127
8128
8129
8130
8131
8132
8133
8134
8135
8136
8137
8138
8139
8140
8141
8142
8143
8144
8145
8146
8147
8148
8149
8150
8151
8152
8153
8154
8155
8156
8157
8158
8159
8160
8161
8162
8163
8164
8165
8166
8167
8168
8169
8170
8171
8172
8173
8174
8175
8176
8177
8178
8179
8180
8181
8182
8183
8184
8185
8186
8187
8188
8189
8190
8191
8192
8193
8194
8195
8196
8197
8198
8199
8200
8201
8202
8203
8204
8205
8206
8207
8208
8209
8210
8211
8212
8213
8214
8215
8216
8217
8218
8219
8220
8221
8222
8223
8224
8225
8226
8227
8228
8229
8230
8231
8232
8233
8234
8235
8236
8237
8238
8239
8240
8241
8242
8243
8244
8245
8246
8247
8248
8249
8250
8251
8252
8253
8254
8255
8256
8257
8258
8259
8260
8261
8262
8263
8264
8265
8266
8267
8268
8269
8270
8271
8272
8273
8274
8275
8276
8277
8278
8279
8280
8281
8282
8283
8284
8285
8286
8287
8288
8289
8290
8291
8292
8293
8294
8295
8296
8297
8298
8299
8300
8301
8302
8303
8304
8305
8306
8307
8308
8309
8310
8311
8312
8313
8314
8315
8316
8317
8318
8319
8320
8321
8322
8323
8324
8325
8326
8327
8328
8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339
8340
8341
8342
8343
8344
8345
8346
8347
8348
8349
8350
8351
8352
8353
8354
8355
8356
8357
8358
8359
8360
8361
8362
8363
8364
8365
8366
8367
8368
8369
8370
8371
8372
8373
8374
8375
8376
8377
8378
8379
8380
8381
8382
8383
8384
8385
8386
8387
8388
8389
8390
8391
8392
8393
8394
8395
8396
8397
8398
8399
8400
8401
8402
8403
8404
8405
8406
8407
8408
8409
8410
8411
8412
8413
8414
8415
8416
8417
8418
8419
8420
8421
8422
8423
8424
8425
8426
8427
8428
8429
8430
8431
8432
8433
8434
8435
8436
8437
8438
8439
8440
8441
8442
8443
8444
8445
8446
8447
8448
8449
8450
8451
8452
8453
8454
8455
8456
8457
8458
8459
8460
8461
8462
8463
8464
8465
8466
8467
8468
8469
8470
8471
8472
8473
8474
8475
8476
8477
8478
8479
8480
8481
8482
8483
8484
8485
8486
8487
8488
8489
8490
8491
8492
8493
8494
8495
8496
8497
8498
8499
8500
8501
8502
8503
8504
8505
8506
8507
8508
8509
8510
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525
8526
8527
8528
8529
8530
8531
8532
8533
8534
8535
8536
8537
8538
8539
8540
8541
8542
8543
8544
8545
8546
8547
8548
8549
8550
8551
8552
8553
8554
8555
8556
8557
8558
8559
8560
8561
8562
8563
8564
8565
8566
8567
8568
8569
8570
8571
8572
8573
8574
8575
8576
8577
8578
8579
8580
8581
8582
8583
8584
8585
8586
8587
8588
8589
8590
8591
8592
8593
8594
8595
8596
8597
8598
8599
8600
8601
8602
8603
8604
8605
8606
8607
8608
8609
8610
8611
8612
8613
8614
8615
8616
8617
8618
8619
8620
8621
8622
8623
8624
8625
8626
8627
8628
8629
8630
8631
8632
8633
8634
8635
8636
8637
8638
8639
8640
8641
8642
8643
8644
8645
8646
8647
8648
8649
8650
8651
8652
8653
8654
8655
8656
8657
8658
8659
8660
8661
8662
8663
8664
8665
8666
8667
8668
8669
8670
8671
8672
8673
8674
8675
8676
8677
8678
8679
8680
8681
8682
8683
8684
8685
8686
8687
8688
8689
8690
8691
8692
8693
8694
8695
8696
8697
8698
8699
8700
8701
8702
8703
8704
8705
8706
8707
8708
8709
8710
8711
8712
8713
8714
8715
8716
8717
8718
8719
8720
8721
8722
8723
8724
8725
8726
8727
8728
8729
8730
8731
8732
8733
8734
8735
8736
8737
8738
8739
8740
8741
8742
8743
8744
8745
8746
8747
8748
8749
8750
8751
8752
8753
8754
8755
8756
8757
8758
8759
8760
8761
8762
8763
8764
8765
8766
8767
8768
8769
8770
8771
8772
8773
8774
8775
8776
8777
8778
8779
8780
8781
8782
8783
8784
8785
8786
8787
8788
8789
8790
8791
8792
8793
8794
8795
8796
8797
8798
8799
8800
8801
8802
8803
8804
8805
8806
8807
8808
8809
8810
8811
8812
8813
8814
8815
8816
8817
8818
8819
8820
8821
8822
8823
8824
8825
8826
8827
8828
8829
8830
8831
8832
8833
8834
8835
8836
8837
8838
8839
8840
8841
8842
8843
8844
8845
8846
8847
8848
8849
8850
8851
8852
8853
8854
8855
8856
8857
8858
8859
8860
8861
8862
8863
8864
8865
8866
8867
8868
8869
8870
8871
8872
8873
8874
8875
8876
8877
8878
8879
8880
8881
8882
8883
8884
8885
8886
8887
8888
8889
8890
8891
8892
8893
8894
8895
8896
8897
8898
8899
8900
8901
8902
8903
8904
8905
8906
8907
8908
8909
8910
8911
8912
8913
8914
8915
8916
8917
8918
8919
8920
8921
8922
8923
8924
8925
8926
8927
8928
8929
8930
8931
8932
8933
8934
8935
8936
8937
8938
8939
8940
8941
8942
8943
8944
8945
8946
8947
8948
8949
8950
8951
8952
8953
8954
8955
8956
8957
8958
8959
8960
8961
8962
8963
8964
8965
8966
8967
8968
8969
8970
8971
8972
8973
8974
8975
8976
8977
8978
8979
8980
8981
8982
8983
8984
8985
8986
8987
8988
8989
8990
8991
8992
8993
8994
8995
8996
8997
8998
8999
9000
9001
9002
9003
9004
9005
9006
9007
9008
9009
9010
9011
9012
9013
9014
9015
9016
9017
9018
9019
9020
9021
9022
9023
9024
9025
9026
9027
9028
9029
9030
9031
9032
9033
9034
9035
9036
9037
9038
9039
9040
9041
9042
9043
9044
9045
9046
9047
9048
9049
9050
9051
9052
9053
9054
9055
9056
9057
9058
9059
9060
9061
9062
9063
9064
9065
9066
9067
9068
9069
9070
9071
9072
9073
9074
9075
9076
9077
9078
9079
9080
9081
9082
9083
9084
9085
9086
9087
9088
9089
9090
9091
9092
9093
9094
9095
9096
9097
9098
9099
9100
9101
9102
9103
9104
9105
9106
9107
9108
9109
9110
9111
9112
9113
9114
9115
9116
9117
9118
9119
9120
9121
9122
9123
9124
9125
9126
9127
9128
9129
9130
9131
9132
9133
9134
9135
9136
9137
9138
9139
9140
9141
9142
9143
9144
9145
9146
9147
9148
9149
9150
9151
9152
9153
9154
9155
9156
9157
9158
9159
9160
9161
9162
9163
9164
9165
9166
9167
9168
9169
9170
9171
9172
9173
9174
9175
9176
9177
9178
9179
9180
9181
9182
9183
9184
9185
9186
9187
9188
9189
9190
9191
9192
9193
9194
9195
9196
9197
9198
9199
9200
9201
9202
9203
9204
9205
9206
9207
9208
9209
9210
9211
9212
9213
9214
9215
9216
9217
9218
9219
9220
9221
9222
9223
9224
9225
9226
9227
9228
9229
9230
9231
9232
9233
9234
9235
9236
9237
9238
9239
9240
9241
9242
9243
9244
9245
9246
9247
9248
9249
9250
9251
9252
9253
9254
9255
9256
9257
9258
9259
9260
9261
9262
9263
9264
9265
9266
9267
9268
9269
9270
9271
9272
9273
9274
9275
9276
9277
9278
9279
9280
9281
9282
9283
9284
9285
9286
9287
9288
9289
9290
9291
9292
9293
9294
9295
9296
9297
9298
9299
9300
9301
9302
9303
9304
9305
9306
9307
9308
9309
9310
9311
9312
9313
9314
9315
9316
9317
9318
9319
9320
9321
9322
9323
9324
9325
9326
9327
9328
9329
9330
9331
9332
9333
9334
9335
9336
9337
9338
9339
9340
9341
9342
9343
9344
9345
9346
9347
9348
9349
9350
9351
9352
9353
9354
9355
9356
9357
9358
9359
9360
9361
9362
9363
9364
9365
9366
9367
9368
9369
9370
9371
9372
9373
9374
9375
9376
9377
9378
9379
9380
9381
9382
9383
9384
9385
9386
9387
9388
9389
9390
9391
9392
9393
9394
9395
9396
9397
9398
9399
9400
9401
9402
9403
9404
9405
9406
9407
9408
9409
9410
9411
9412
9413
9414
9415
9416
9417
9418
9419
9420
9421
9422
9423
9424
9425
9426
9427
9428
9429
9430
9431
9432
9433
9434
9435
9436
9437
9438
9439
9440
9441
9442
9443
9444
9445
9446
9447
9448
9449
9450
9451
9452
9453
9454
9455
9456
9457
9458
9459
9460
9461
9462
9463
9464
9465
9466
9467
9468
9469
9470
9471
9472
9473
9474
9475
9476
9477
9478
9479
9480
9481
9482
9483
9484
9485
9486
9487
9488
9489
9490
9491
9492
9493
9494
9495
9496
9497
9498
9499
9500
9501
9502
9503
9504
9505
9506
9507
9508
9509
9510
9511
9512
9513
9514
9515
9516
9517
9518
9519
9520
9521
9522
9523
9524
9525
9526
9527
9528
9529
9530
9531
9532
9533
9534
9535
9536
9537
9538
9539
9540
9541
9542
9543
9544
9545
9546
9547
9548
9549
9550
9551
9552
9553
9554
9555
9556
9557
9558
9559
9560
9561
9562
9563
9564
9565
9566
9567
9568
9569
9570
9571
9572
9573
9574
9575
9576
9577
9578
9579
9580
9581
9582
9583
9584
9585
9586
9587
9588
9589
9590
9591
9592
9593
9594
9595
9596
9597
9598
9599
9600
9601
9602
9603
9604
9605
9606
9607
9608
9609
9610
9611
9612
9613
9614
9615
9616
9617
9618
9619
9620
9621
9622
9623
9624
9625
9626
9627
9628
9629
9630
9631
9632
9633
9634
9635
9636
9637
9638
9639
9640
9641
9642
9643
9644
9645
9646
9647
9648
9649
9650
9651
9652
9653
9654
9655
9656
9657
9658
9659
9660
9661
9662
9663
9664
9665
9666
9667
9668
9669
9670
9671
9672
9673
9674
9675
9676
9677
9678
9679
9680
9681
9682
9683
9684
9685
9686
9687
9688
9689
9690
9691
9692
9693
9694
9695
9696
9697
9698
9699
9700
9701
9702
9703
9704
9705
9706
9707
9708
9709
9710
9711
9712
9713
9714
9715
9716
9717
9718
9719
9720
9721
9722
9723
9724
9725
9726
9727
9728
9729
9730
9731
9732
9733
9734
9735
9736
9737
9738
9739
9740
9741
9742
9743
9744
9745
9746
9747
9748
9749
9750
9751
9752
9753
9754
9755
9756
9757
9758
9759
9760
9761
9762
9763
9764
9765
9766
9767
9768
9769
9770
9771
9772
9773
9774
9775
9776
9777
9778
9779
9780
9781
9782
9783
9784
9785
9786
9787
9788
9789
9790
9791
9792
9793
9794
9795
9796
9797
9798
9799
9800
9801
9802
9803
9804
9805
9806
9807
9808
9809
9810
9811
9812
9813
9814
9815
9816
9817
9818
9819
9820
9821
9822
9823
9824
9825
9826
9827
9828
9829
9830
9831
9832
9833
9834
9835
9836
9837
9838
9839
9840
9841
9842
9843
9844
9845
9846
9847
9848
9849
9850
9851
9852
9853
9854
9855
9856
9857
9858
9859
9860
9861
9862
9863
9864
9865
9866
9867
9868
9869
9870
9871
9872
9873
9874
9875
9876
9877
9878
9879
9880
9881
9882
9883
9884
9885
9886
9887
9888
9889
9890
9891
9892
9893
9894
9895
9896
9897
9898
9899
9900
9901
9902
9903
9904
9905
9906
9907
9908
9909
9910
9911
9912
9913
9914
9915
9916
9917
9918
9919
9920
9921
9922
9923
9924
9925
9926
9927
9928
9929
9930
9931
9932
9933
9934
9935
9936
9937
9938
9939
9940
9941
9942
9943
9944
9945
9946
9947
9948
9949
9950
9951
9952
9953
9954
9955
9956
9957
9958
9959
9960
9961
9962
9963
9964
9965
9966
9967
9968
9969
9970
9971
9972
9973
9974
9975
9976
9977
9978
9979
9980
9981
9982
9983
9984
9985
9986
9987
9988
9989
9990
9991
9992
9993
9994
9995
9996
9997
9998
9999
10000
10001
10002
10003
10004
10005
10006
10007
10008
10009
10010
10011
10012
10013
10014
10015
10016
10017
10018
10019
10020
10021
10022
10023
10024
10025
10026
10027
10028
10029
10030
10031
10032
10033
10034
10035
10036
10037
10038
10039
10040
10041
10042
10043
10044
10045
10046
10047
10048
10049
10050
10051
10052
10053
10054
10055
10056
10057
10058
10059
10060
10061
10062
10063
10064
10065
10066
10067
10068
10069
10070
10071
10072
10073
10074
10075
10076
10077
10078
10079
10080
10081
10082
10083
10084
10085
10086
10087
10088
10089
10090
10091
10092
10093
10094
10095
10096
10097
10098
10099
10100
10101
10102
10103
10104
10105
10106
10107
10108
10109
10110
10111
10112
10113
10114
10115
10116
10117
10118
10119
10120
10121
10122
10123
10124
10125
10126
10127
10128
10129
10130
10131
10132
10133
10134
10135
10136
10137
10138
10139
10140
10141
10142
10143
10144
10145
10146
10147
10148
10149
10150
10151
10152
10153
10154
10155
10156
10157
10158
10159
10160
10161
10162
10163
10164
10165
10166
10167
10168
10169
10170
10171
10172
10173
10174
10175
10176
10177
10178
10179
10180
10181
10182
10183
10184
10185
10186
10187
10188
10189
10190
10191
10192
10193
10194
10195
10196
10197
10198
10199
10200
10201
10202
10203
10204
10205
10206
10207
10208
10209
10210
10211
10212
10213
10214
10215
10216
10217
10218
10219
10220
10221
10222
10223
10224
10225
10226
10227
10228
10229
10230
10231
10232
10233
10234
10235
10236
10237
10238
10239
10240
10241
10242
10243
10244
10245
10246
10247
10248
10249
10250
10251
10252
10253
10254
10255
10256
10257
10258
10259
10260
10261
10262
10263
10264
10265
10266
10267
10268
10269
10270
10271
10272
10273
10274
10275
10276
10277
10278
10279
10280
10281
10282
10283
10284
10285
10286
10287
10288
10289
10290
10291
10292
10293
10294
10295
10296
10297
10298
10299
10300
10301
10302
10303
10304
10305
10306
10307
10308
10309
10310
10311
10312
10313
10314
10315
10316
10317
10318
10319
10320
10321
10322
10323
10324
10325
10326
10327
10328
10329
10330
10331
10332
10333
10334
10335
10336
10337
10338
10339
10340
10341
10342
10343
10344
10345
10346
10347
10348
10349
10350
10351
10352
10353
10354
10355
10356
10357
10358
10359
10360
10361
10362
10363
10364
10365
10366
10367
10368
10369
10370
10371
10372
10373
10374
10375
10376
10377
10378
10379
10380
10381
10382
10383
10384
10385
10386
10387
10388
10389
10390
10391
10392
10393
10394
10395
10396
10397
10398
10399
10400
10401
10402
10403
10404
10405
10406
10407
10408
10409
10410
10411
10412
10413
10414
10415
10416
10417
10418
10419
10420
10421
10422
10423
10424
10425
10426
10427
10428
10429
10430
10431
10432
10433
10434
10435
10436
10437
10438
10439
10440
10441
10442
10443
10444
10445
10446
10447
10448
10449
10450
10451
10452
10453
10454
10455
10456
10457
10458
10459
10460
10461
10462
10463
10464
10465
10466
10467
10468
10469
10470
10471
10472
10473
10474
10475
10476
10477
10478
10479
10480
10481
10482
10483
10484
10485
10486
10487
10488
10489
10490
10491
10492
10493
10494
10495
10496
10497
10498
10499
10500
10501
10502
10503
10504
10505
10506
10507
10508
10509
10510
10511
10512
10513
10514
10515
10516
10517
10518
10519
10520
10521
10522
10523
10524
10525
10526
10527
10528
10529
10530
10531
10532
10533
10534
10535
10536
10537
10538
10539
10540
10541
10542
10543
10544
10545
10546
10547
10548
10549
10550
10551
10552
10553
10554
10555
10556
10557
10558
10559
10560
10561
10562
10563
10564
10565
10566
10567
10568
10569
10570
10571
10572
10573
10574
10575
10576
10577
10578
10579
10580
10581
10582
10583
10584
10585
10586
10587
10588
10589
10590
10591
10592
10593
10594
10595
10596
10597
10598
10599
10600
10601
10602
10603
10604
10605
10606
10607
10608
10609
10610
10611
10612
10613
10614
10615
10616
10617
10618
10619
10620
10621
10622
10623
10624
10625
10626
10627
10628
10629
10630
10631
10632
10633
10634
10635
10636
10637
10638
10639
10640
10641
10642
10643
10644
10645
10646
10647
10648
10649
10650
10651
10652
10653
10654
10655
10656
10657
10658
10659
10660
10661
10662
10663
10664
10665
10666
10667
10668
10669
10670
10671
10672
10673
10674
10675
10676
10677
10678
10679
10680
10681
10682
10683
10684
10685
10686
10687
10688
10689
10690
10691
10692
10693
10694
10695
10696
10697
10698
10699
10700
10701
10702
10703
10704
10705
10706
10707
10708
10709
10710
10711
10712
10713
10714
10715
10716
10717
10718
10719
10720
10721
10722
10723
10724
10725
10726
10727
10728
10729
10730
10731
10732
10733
10734
10735
10736
10737
10738
10739
10740
10741
10742
10743
10744
10745
10746
10747
10748
10749
10750
10751
10752
10753
10754
10755
10756
10757
10758
10759
10760
10761
10762
10763
10764
10765
10766
10767
10768
10769
10770
10771
10772
10773
10774
10775
10776
10777
10778
10779
10780
10781
10782
10783
10784
10785
10786
10787
10788
10789
10790
10791
10792
10793
10794
10795
10796
10797
10798
10799
10800
10801
10802
10803
10804
10805
10806
10807
10808
10809
10810
10811
10812
10813
10814
10815
10816
10817
10818
10819
10820
10821
10822
10823
10824
10825
10826
10827
10828
10829
10830
10831
10832
10833
10834
10835
10836
10837
10838
10839
10840
10841
10842
10843
10844
10845
10846
10847
10848
10849
10850
10851
10852
10853
10854
10855
10856
10857
10858
10859
10860
10861
10862
10863
10864
10865
10866
10867
10868
10869
10870
10871
10872
10873
10874
10875
10876
10877
10878
10879
10880
10881
10882
10883
10884
10885
10886
10887
10888
10889
10890
10891
10892
10893
10894
10895
10896
10897
10898
10899
10900
10901
10902
10903
10904
10905
10906
10907
10908
10909
10910
10911
10912
10913
10914
10915
10916
10917
10918
10919
10920
10921
10922
10923
10924
10925
10926
10927
10928
10929
10930
10931
10932
10933
10934
10935
10936
10937
10938
10939
10940
10941
10942
10943
10944
10945
10946
10947
10948
10949
10950
10951
10952
10953
10954
10955
10956
10957
10958
10959
10960
10961
10962
10963
10964
10965
10966
10967
10968
10969
10970
10971
10972
10973
10974
10975
10976
10977
10978
10979
10980
10981
10982
10983
10984
10985
10986
10987
10988
10989
10990
10991
10992
10993
10994
10995
10996
10997
10998
10999
11000
11001
11002
11003
11004
11005
11006
11007
11008
11009
11010
11011
11012
11013
11014
11015
11016
11017
11018
11019
11020
11021
11022
11023
11024
11025
11026
11027
11028
11029
11030
11031
11032
11033
11034
11035
11036
11037
11038
11039
11040
11041
11042
11043
11044
11045
11046
11047
11048
11049
11050
11051
11052
11053
11054
11055
11056
11057
11058
11059
11060
11061
11062
11063
11064
11065
11066
11067
11068
11069
11070
11071
11072
11073
11074
11075
11076
11077
11078
11079
11080
11081
11082
11083
11084
11085
11086
11087
11088
11089
11090
11091
11092
11093
11094
11095
11096
11097
11098
11099
11100
11101
11102
11103
11104
11105
11106
11107
11108
11109
11110
11111
11112
11113
11114
11115
11116
11117
11118
11119
11120
11121
11122
11123
11124
11125
11126
11127
11128
11129
11130
11131
11132
11133
11134
11135
11136
11137
11138
11139
11140
11141
11142
11143
11144
11145
11146
11147
11148
11149
11150
11151
11152
11153
11154
11155
11156
11157
11158
11159
11160
11161
11162
11163
11164
11165
11166
11167
11168
11169
11170
11171
11172
11173
11174
11175
11176
11177
11178
11179
11180
11181
11182
11183
11184
11185
11186
11187
11188
11189
11190
11191
11192
11193
11194
11195
11196
11197
11198
11199
11200
11201
11202
11203
11204
11205
11206
11207
11208
11209
11210
11211
11212
11213
11214
11215
11216
11217
11218
11219
11220
11221
11222
11223
11224
11225
11226
11227
11228
11229
11230
11231
11232
11233
11234
11235
11236
11237
11238
11239
11240
11241
11242
11243
11244
11245
11246
11247
11248
11249
11250
11251
11252
11253
11254
11255
11256
11257
11258
11259
11260
11261
11262
11263
11264
11265
11266
11267
11268
11269
11270
11271
11272
11273
11274
11275
11276
11277
11278
11279
11280
11281
11282
11283
11284
11285
11286
11287
11288
11289
11290
11291
11292
11293
11294
11295
11296
11297
11298
11299
11300
11301
11302
11303
11304
11305
11306
11307
11308
11309
11310
11311
11312
11313
11314
11315
11316
11317
11318
11319
11320
11321
11322
11323
11324
11325
11326
11327
11328
11329
11330
11331
11332
11333
11334
11335
11336
11337
11338
11339
11340
11341
11342
11343
11344
11345
11346
11347
11348
11349
11350
11351
11352
11353
11354
11355
11356
11357
11358
11359
11360
11361
11362
11363
11364
11365
11366
11367
11368
11369
11370
11371
11372
11373
11374
11375
11376
11377
11378
11379
11380
11381
11382
11383
11384
11385
11386
11387
11388
11389
11390
11391
11392
11393
11394
11395
11396
11397
11398
11399
11400
11401
11402
11403
11404
11405
11406
11407
11408
11409
11410
11411
11412
11413
11414
11415
11416
11417
11418
11419
11420
11421
11422
11423
11424
11425
11426
11427
11428
11429
11430
11431
11432
11433
11434
11435
11436
11437
11438
11439
11440
11441
11442
11443
11444
11445
11446
11447
11448
11449
11450
11451
11452
11453
11454
11455
11456
11457
11458
11459
11460
11461
11462
11463
11464
11465
11466
11467
11468
11469
11470
11471
11472
11473
11474
11475
11476
11477
11478
11479
11480
11481
11482
11483
11484
11485
11486
11487
11488
11489
11490
11491
11492
11493
11494
11495
11496
11497
11498
11499
11500
11501
11502
11503
11504
11505
11506
11507
11508
11509
11510
11511
11512
11513
11514
11515
11516
11517
11518
11519
11520
11521
11522
11523
11524
11525
11526
11527
11528
11529
11530
11531
11532
11533
11534
11535
11536
11537
11538
11539
11540
11541
11542
11543
11544
11545
11546
11547
11548
11549
11550
11551
11552
11553
11554
11555
11556
11557
11558
11559
11560
11561
11562
11563
11564
11565
11566
11567
11568
11569
11570
11571
11572
11573
11574
11575
11576
11577
11578
11579
11580
11581
11582
11583
11584
11585
11586
11587
11588
11589
11590
11591
11592
11593
11594
11595
11596
11597
11598
11599
11600
11601
11602
11603
11604
11605
11606
11607
11608
11609
11610
11611
11612
11613
11614
11615
11616
11617
11618
11619
11620
11621
11622
11623
11624
11625
11626
11627
11628
11629
11630
11631
11632
11633
11634
11635
11636
11637
11638
11639
11640
11641
11642
11643
11644
11645
11646
11647
11648
11649
11650
11651
11652
11653
11654
11655
11656
11657
11658
11659
11660
11661
11662
11663
11664
11665
11666
11667
11668
11669
11670
11671
11672
11673
11674
11675
11676
11677
11678
11679
11680
11681
11682
11683
11684
11685
11686
11687
11688
11689
11690
11691
11692
11693
11694
11695
11696
11697
11698
11699
11700
11701
11702
11703
11704
11705
11706
11707
11708
11709
11710
11711
11712
11713
11714
11715
11716
11717
11718
11719
11720
11721
11722
11723
11724
11725
11726
11727
11728
11729
11730
11731
11732
11733
11734
11735
11736
11737
11738
11739
11740
11741
11742
11743
11744
11745
11746
11747
11748
11749
11750
11751
11752
11753
11754
11755
11756
11757
11758
11759
11760
11761
11762
11763
11764
11765
11766
11767
11768
11769
11770
11771
11772
11773
11774
11775
11776
11777
11778
11779
11780
11781
11782
11783
11784
11785
11786
11787
11788
11789
11790
11791
11792
11793
11794
11795
11796
11797
11798
11799
11800
11801
11802
11803
11804
11805
11806
11807
11808
11809
11810
11811
11812
11813
11814
11815
11816
11817
11818
11819
11820
11821
11822
11823
11824
11825
11826
11827
11828
11829
11830
11831
11832
11833
11834
11835
11836
11837
11838
11839
11840
11841
11842
11843
11844
11845
11846
11847
11848
11849
11850
11851
11852
11853
11854
11855
11856
11857
11858
11859
11860
11861
11862
11863
11864
11865
11866
11867
11868
11869
11870
11871
11872
11873
11874
11875
11876
11877
11878
11879
11880
11881
11882
11883
11884
11885
11886
11887
11888
11889
11890
11891
11892
11893
11894
11895
11896
11897
11898
11899
11900
11901
11902
11903
11904
11905
11906
11907
11908
11909
11910
11911
11912
11913
11914
11915
11916
11917
11918
11919
11920
11921
11922
11923
11924
11925
11926
11927
11928
11929
11930
11931
11932
11933
11934
11935
11936
11937
11938
11939
11940
11941
11942
11943
11944
11945
11946
11947
11948
11949
11950
11951
11952
11953
11954
11955
11956
11957
11958
11959
11960
11961
11962
11963
11964
11965
11966
11967
11968
11969
11970
11971
11972
11973
11974
11975
11976
11977
11978
11979
11980
11981
11982
11983
11984
11985
11986
11987
11988
11989
11990
11991
11992
11993
11994
11995
11996
11997
11998
11999
12000
12001
12002
12003
12004
12005
12006
12007
12008
12009
12010
12011
12012
12013
12014
12015
12016
12017
12018
12019
12020
12021
12022
12023
12024
12025
12026
12027
12028
12029
12030
12031
12032
12033
12034
12035
12036
12037
12038
12039
12040
12041
12042
12043
12044
12045
12046
12047
12048
12049
12050
12051
12052
12053
12054
12055
12056
12057
12058
12059
12060
12061
12062
12063
12064
12065
12066
12067
12068
12069
12070
12071
12072
12073
12074
12075
12076
12077
12078
12079
12080
12081
12082
12083
12084
12085
12086
12087
12088
12089
12090
12091
12092
12093
12094
12095
12096
12097
12098
12099
12100
12101
12102
12103
12104
12105
12106
12107
12108
12109
12110
12111
12112
12113
12114
12115
12116
12117
12118
12119
12120
12121
12122
12123
12124
12125
12126
12127
12128
12129
12130
12131
12132
12133
12134
12135
12136
12137
12138
12139
12140
12141
12142
12143
12144
12145
12146
12147
12148
12149
12150
12151
12152
12153
12154
12155
12156
12157
12158
12159
12160
12161
12162
12163
12164
12165
12166
12167
12168
12169
12170
12171
12172
12173
12174
12175
12176
12177
12178
12179
12180
12181
12182
12183
12184
12185
12186
12187
12188
12189
12190
12191
12192
12193
12194
12195
12196
12197
12198
12199
12200
12201
12202
12203
12204
12205
12206
12207
12208
12209
12210
12211
12212
12213
12214
12215
12216
12217
12218
12219
12220
12221
12222
12223
12224
12225
12226
12227
12228
12229
12230
12231
12232
12233
12234
12235
12236
12237
12238
12239
12240
12241
12242
12243
12244
12245
12246
12247
12248
12249
12250
12251
12252
12253
12254
12255
12256
12257
12258
12259
12260
12261
12262
12263
12264
12265
12266
12267
12268
12269
12270
12271
12272
12273
12274
12275
12276
12277
12278
12279
12280
12281
12282
12283
12284
12285
12286
12287
12288
12289
12290
12291
12292
12293
12294
12295
12296
12297
12298
12299
12300
12301
12302
12303
12304
12305
12306
12307
12308
12309
12310
12311
12312
12313
12314
12315
12316
12317
12318
12319
12320
12321
12322
12323
12324
12325
12326
12327
12328
12329
12330
12331
12332
12333
12334
12335
12336
12337
12338
12339
12340
12341
12342
12343
12344
12345
12346
12347
12348
12349
12350
12351
12352
12353
12354
12355
12356
12357
12358
12359
12360
12361
12362
12363
12364
12365
12366
12367
12368
12369
12370
12371
12372
12373
12374
12375
12376
12377
12378
12379
12380
12381
12382
12383
12384
12385
12386
12387
12388
12389
12390
12391
12392
12393
12394
12395
12396
12397
12398
12399
12400
12401
12402
12403
12404
12405
12406
12407
12408
12409
12410
12411
12412
12413
12414
12415
12416
12417
12418
12419
12420
12421
12422
12423
12424
12425
12426
12427
12428
12429
12430
12431
12432
12433
12434
12435
12436
12437
12438
12439
12440
12441
12442
12443
12444
12445
12446
12447
12448
12449
12450
12451
12452
12453
12454
12455
12456
12457
12458
12459
12460
12461
12462
12463
12464
12465
12466
12467
12468
12469
12470
12471
12472
12473
12474
12475
12476
12477
12478
12479
12480
12481
12482
12483
12484
12485
12486
12487
12488
12489
12490
12491
12492
12493
12494
12495
12496
12497
12498
12499
12500
12501
12502
12503
12504
12505
12506
12507
12508
12509
12510
12511
12512
12513
12514
12515
12516
12517
12518
12519
12520
12521
12522
12523
12524
12525
12526
12527
12528
12529
12530
12531
12532
12533
12534
12535
12536
12537
12538
12539
12540
12541
12542
12543
12544
12545
12546
12547
12548
12549
12550
12551
12552
12553
12554
12555
12556
12557
12558
12559
12560
12561
12562
12563
12564
12565
12566
12567
12568
12569
12570
12571
12572
12573
12574
12575
12576
12577
12578
12579
12580
12581
12582
12583
12584
12585
12586
12587
12588
12589
12590
12591
12592
12593
12594
12595
12596
12597
12598
12599
12600
12601
12602
12603
12604
12605
12606
12607
12608
12609
12610
12611
12612
12613
12614
12615
12616
12617
12618
12619
12620
12621
12622
12623
12624
12625
12626
12627
12628
12629
12630
12631
12632
12633
12634
12635
12636
12637
12638
12639
12640
12641
12642
12643
12644
12645
12646
12647
12648
12649
12650
12651
12652
12653
12654
12655
12656
12657
12658
12659
12660
12661
12662
12663
12664
12665
12666
12667
12668
12669
12670
12671
12672
12673
12674
12675
12676
12677
12678
12679
12680
12681
12682
12683
12684
12685
12686
12687
12688
12689
12690
12691
12692
12693
12694
12695
12696
12697
12698
12699
12700
12701
12702
12703
12704
12705
12706
12707
12708
12709
12710
12711
12712
12713
12714
12715
12716
12717
12718
12719
12720
12721
12722
12723
12724
12725
12726
12727
12728
12729
12730
12731
12732
12733
12734
12735
12736
12737
12738
12739
12740
12741
12742
12743
12744
12745
12746
12747
12748
12749
12750
12751
12752
12753
12754
12755
12756
12757
12758
12759
12760
12761
12762
12763
12764
12765
12766
12767
12768
12769
12770
12771
12772
12773
12774
12775
12776
12777
12778
12779
12780
12781
12782
12783
12784
12785
12786
12787
12788
12789
12790
12791
12792
12793
12794
12795
12796
12797
12798
12799
12800
12801
12802
12803
12804
12805
12806
12807
12808
12809
12810
12811
12812
12813
12814
12815
12816
12817
12818
12819
12820
12821
12822
12823
12824
12825
12826
12827
12828
12829
12830
12831
12832
12833
12834
12835
12836
12837
12838
12839
12840
12841
12842
12843
12844
12845
12846
12847
12848
12849
12850
12851
12852
12853
12854
12855
12856
12857
12858
12859
12860
12861
12862
12863
12864
12865
12866
12867
12868
12869
12870
12871
12872
12873
12874
12875
12876
12877
12878
12879
12880
12881
12882
12883
12884
12885
12886
12887
12888
12889
12890
12891
12892
12893
12894
12895
12896
12897
12898
12899
12900
12901
12902
12903
12904
12905
12906
12907
12908
12909
12910
12911
12912
12913
12914
12915
12916
12917
12918
12919
12920
12921
12922
12923
12924
12925
12926
12927
12928
12929
12930
12931
12932
12933
12934
12935
12936
12937
12938
12939
12940
12941
12942
12943
12944
12945
12946
12947
12948
12949
12950
12951
12952
12953
12954
12955
12956
12957
12958
12959
12960
12961
12962
12963
12964
12965
12966
12967
12968
12969
12970
12971
12972
12973
12974
12975
12976
12977
12978
12979
12980
12981
12982
12983
12984
12985
12986
12987
12988
12989
12990
12991
12992
12993
12994
12995
12996
12997
12998
12999
13000
13001
13002
13003
13004
13005
13006
13007
13008
13009
13010
13011
13012
13013
13014
13015
13016
13017
13018
13019
13020
13021
13022
13023
13024
13025
13026
13027
13028
13029
13030
13031
13032
13033
13034
13035
13036
13037
13038
13039
13040
13041
13042
13043
13044
13045
13046
13047
13048
13049
13050
13051
13052
13053
13054
13055
13056
13057
13058
13059
13060
13061
13062
13063
13064
13065
13066
13067
13068
13069
13070
13071
13072
13073
13074
13075
13076
13077
13078
13079
13080
13081
13082
13083
13084
13085
13086
13087
13088
13089
13090
13091
13092
13093
13094
13095
13096
13097
13098
13099
13100
13101
13102
13103
13104
13105
13106
13107
13108
13109
13110
13111
13112
13113
13114
13115
13116
13117
13118
13119
13120
13121
13122
13123
13124
13125
13126
13127
13128
13129
13130
13131
13132
13133
13134
13135
13136
13137
13138
13139
13140
13141
13142
13143
13144
13145
13146
13147
13148
13149
13150
13151
13152
13153
13154
13155
13156
13157
13158
13159
13160
13161
13162
13163
13164
13165
13166
13167
13168
13169
13170
13171
13172
13173
13174
13175
13176
13177
13178
13179
13180
13181
13182
13183
13184
13185
13186
13187
13188
13189
13190
13191
13192
13193
13194
13195
13196
13197
13198
13199
13200
13201
13202
13203
13204
13205
13206
13207
13208
13209
13210
13211
13212
13213
13214
13215
13216
13217
13218
13219
13220
13221
13222
13223
13224
13225
|
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
// See the LICENSE file in the project root for more information.
/*
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Linear Scan Register Allocation
a.k.a. LSRA
Preconditions
- All register requirements are expressed in the code stream, either as destination
registers of tree nodes, or as internal registers. These requirements are
expressed in the TreeNodeInfo (gtLsraInfo) on each node, which includes:
- The number of register sources and destinations.
- The register restrictions (candidates) of the target register, both from itself,
as producer of the value (dstCandidates), and from its consuming node (srcCandidates).
Note that the srcCandidates field of TreeNodeInfo refers to the destination register
(not any of its sources).
- The number (internalCount) of registers required, and their register restrictions (internalCandidates).
These are neither inputs nor outputs of the node, but used in the sequence of code generated for the tree.
"Internal registers" are registers used during the code sequence generated for the node.
The register lifetimes must obey the following lifetime model:
- First, any internal registers are defined.
- Next, any source registers are used (and are then freed if they are last use and are not identified as
"delayRegFree").
- Next, the internal registers are used (and are then freed).
- Next, any registers in the kill set for the instruction are killed.
- Next, the destination register(s) are defined (multiple destination registers are only supported on ARM)
- Finally, any "delayRegFree" source registers are freed.
There are several things to note about this order:
- The internal registers will never overlap any use, but they may overlap a destination register.
- Internal registers are never live beyond the node.
- The "delayRegFree" annotation is used for instructions that are only available in a Read-Modify-Write form.
That is, the destination register is one of the sources. In this case, we must not use the same register for
the non-RMW operand as for the destination.
Overview (doLinearScan):
- Walk all blocks, building intervals and RefPositions (buildIntervals)
- Allocate registers (allocateRegisters)
- Annotate nodes with register assignments (resolveRegisters)
- Add move nodes as needed to resolve conflicting register
assignments across non-adjacent edges. (resolveEdges, called from resolveRegisters)
Postconditions:
Tree nodes (GenTree):
- GenTree::gtRegNum (and gtRegPair for ARM) is annotated with the register
assignment for a node. If the node does not require a register, it is
annotated as such (for single registers, gtRegNum = REG_NA; for register
pair type, gtRegPair = REG_PAIR_NONE). For a variable definition or interior
tree node (an "implicit" definition), this is the register to put the result.
For an expression use, this is the place to find the value that has previously
been computed.
- In most cases, this register must satisfy the constraints specified by the TreeNodeInfo.
- In some cases, this is difficult:
- If a lclVar node currently lives in some register, it may not be desirable to move it
(i.e. its current location may be desirable for future uses, e.g. if it's a callee save register,
but needs to be in a specific arg register for a call).
- In other cases there may be conflicts on the restrictions placed by the defining node and the node which
consumes it
- If such a node is constrained to a single fixed register (e.g. an arg register, or a return from a call),
then LSRA is free to annotate the node with a different register. The code generator must issue the appropriate
move.
- However, if such a node is constrained to a set of registers, and its current location does not satisfy that
requirement, LSRA must insert a GT_COPY node between the node and its parent. The gtRegNum on the GT_COPY node
must satisfy the register requirement of the parent.
- GenTree::gtRsvdRegs has a set of registers used for internal temps.
- A tree node is marked GTF_SPILL if the tree node must be spilled by the code generator after it has been
evaluated.
- LSRA currently does not set GTF_SPILLED on such nodes, because it caused problems in the old code generator.
In the new backend perhaps this should change (see also the note below under CodeGen).
- A tree node is marked GTF_SPILLED if it is a lclVar that must be reloaded prior to use.
- The register (gtRegNum) on the node indicates the register to which it must be reloaded.
- For lclVar nodes, since the uses and defs are distinct tree nodes, it is always possible to annotate the node
with the register to which the variable must be reloaded.
- For other nodes, since they represent both the def and use, if the value must be reloaded to a different
register, LSRA must insert a GT_RELOAD node in order to specify the register to which it should be reloaded.
Local variable table (LclVarDsc):
- LclVarDsc::lvRegister is set to true if a local variable has the
same register assignment for its entire lifetime.
- LclVarDsc::lvRegNum / lvOtherReg: these are initialized to their
first value at the end of LSRA (it looks like lvOtherReg isn't?
This is probably a bug (ARM)). Codegen will set them to their current value
as it processes the trees, since a variable can (now) be assigned different
registers over its lifetimes.
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
*/
#include "jitpch.h"
#ifdef _MSC_VER
#pragma hdrstop
#endif
#ifndef LEGACY_BACKEND // This file is ONLY used for the RyuJIT backend that uses the linear scan register allocator
#include "lsra.h"
#ifdef DEBUG
const char* LinearScan::resolveTypeName[] = {"Split", "Join", "Critical", "SharedCritical"};
#endif // DEBUG
/*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XX XX
XX Small Helper functions XX
XX XX
XX XX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
*/
//--------------------------------------------------------------
// lsraAssignRegToTree: Assign the given reg to tree node.
//
// Arguments:
// tree - Gentree node
// reg - register to be assigned
// regIdx - register idx, if tree is a multi-reg call node.
// regIdx will be zero for single-reg result producing tree nodes.
//
// Return Value:
// None
//
void lsraAssignRegToTree(GenTreePtr tree, regNumber reg, unsigned regIdx)
{
if (regIdx == 0)
{
tree->gtRegNum = reg;
}
#if defined(_TARGET_ARM_)
else if (tree->OperIsMultiRegOp())
{
assert(regIdx == 1);
GenTreeMultiRegOp* mul = tree->AsMultiRegOp();
mul->gtOtherReg = reg;
}
else if (tree->OperGet() == GT_COPY)
{
assert(regIdx == 1);
GenTreeCopyOrReload* copy = tree->AsCopyOrReload();
copy->gtOtherRegs[0] = (regNumberSmall)reg;
}
else if (tree->OperIsPutArgSplit())
{
GenTreePutArgSplit* putArg = tree->AsPutArgSplit();
putArg->SetRegNumByIdx(reg, regIdx);
}
#endif // _TARGET_ARM_
else
{
assert(tree->IsMultiRegCall());
GenTreeCall* call = tree->AsCall();
call->SetRegNumByIdx(reg, regIdx);
}
}
//-------------------------------------------------------------
// getWeight: Returns the weight of the RefPosition.
//
// Arguments:
// refPos - ref position
//
// Returns:
// Weight of ref position.
unsigned LinearScan::getWeight(RefPosition* refPos)
{
unsigned weight;
GenTreePtr treeNode = refPos->treeNode;
if (treeNode != nullptr)
{
if (isCandidateLocalRef(treeNode))
{
// Tracked locals: use weighted ref cnt as the weight of the
// ref position.
GenTreeLclVarCommon* lclCommon = treeNode->AsLclVarCommon();
LclVarDsc* varDsc = &(compiler->lvaTable[lclCommon->gtLclNum]);
weight = varDsc->lvRefCntWtd;
if (refPos->getInterval()->isSpilled)
{
// Decrease the weight if the interval has already been spilled.
weight -= BB_UNITY_WEIGHT;
}
}
else
{
// Non-candidate local ref or non-lcl tree node.
// These are considered to have two references in the basic block:
// a def and a use and hence weighted ref count would be 2 times
// the basic block weight in which they appear.
// However, it is generally more harmful to spill tree temps, so we
// double that.
const unsigned TREE_TEMP_REF_COUNT = 2;
const unsigned TREE_TEMP_BOOST_FACTOR = 2;
weight = TREE_TEMP_REF_COUNT * TREE_TEMP_BOOST_FACTOR * blockInfo[refPos->bbNum].weight;
}
}
else
{
// Non-tree node ref positions. These will have a single
// reference in the basic block and hence their weighted
// refcount is equal to the block weight in which they
// appear.
weight = blockInfo[refPos->bbNum].weight;
}
return weight;
}
// allRegs represents a set of registers that can
// be used to allocate the specified type in any point
// in time (more of a 'bank' of registers).
regMaskTP LinearScan::allRegs(RegisterType rt)
{
if (rt == TYP_FLOAT)
{
return availableFloatRegs;
}
else if (rt == TYP_DOUBLE)
{
return availableDoubleRegs;
#ifdef FEATURE_SIMD
// TODO-Cleanup: Add an RBM_ALLSIMD
}
else if (varTypeIsSIMD(rt))
{
return availableDoubleRegs;
#endif // FEATURE_SIMD
}
else
{
return availableIntRegs;
}
}
//--------------------------------------------------------------------------
// allMultiRegCallNodeRegs: represents a set of registers that can be used
// to allocate a multi-reg call node.
//
// Arguments:
// call - Multi-reg call node
//
// Return Value:
// Mask representing the set of available registers for multi-reg call
// node.
//
// Note:
// Multi-reg call node available regs = Bitwise-OR(allregs(GetReturnRegType(i)))
// for all i=0..RetRegCount-1.
regMaskTP LinearScan::allMultiRegCallNodeRegs(GenTreeCall* call)
{
assert(call->HasMultiRegRetVal());
ReturnTypeDesc* retTypeDesc = call->GetReturnTypeDesc();
regMaskTP resultMask = allRegs(retTypeDesc->GetReturnRegType(0));
unsigned count = retTypeDesc->GetReturnRegCount();
for (unsigned i = 1; i < count; ++i)
{
resultMask |= allRegs(retTypeDesc->GetReturnRegType(i));
}
return resultMask;
}
//--------------------------------------------------------------------------
// allRegs: returns the set of registers that can accomodate the type of
// given node.
//
// Arguments:
// tree - GenTree node
//
// Return Value:
// Mask representing the set of available registers for given tree
//
// Note: In case of multi-reg call node, the full set of registers must be
// determined by looking at types of individual return register types.
// In this case, the registers may include registers from different register
// sets and will not be limited to the actual ABI return registers.
regMaskTP LinearScan::allRegs(GenTree* tree)
{
regMaskTP resultMask;
// In case of multi-reg calls, allRegs is defined as
// Bitwise-Or(allRegs(GetReturnRegType(i)) for i=0..ReturnRegCount-1
if (tree->IsMultiRegCall())
{
resultMask = allMultiRegCallNodeRegs(tree->AsCall());
}
else
{
resultMask = allRegs(tree->TypeGet());
}
return resultMask;
}
regMaskTP LinearScan::allSIMDRegs()
{
return availableFloatRegs;
}
//------------------------------------------------------------------------
// internalFloatRegCandidates: Return the set of registers that are appropriate
// for use as internal float registers.
//
// Return Value:
// The set of registers (as a regMaskTP).
//
// Notes:
// compFloatingPointUsed is only required to be set if it is possible that we
// will use floating point callee-save registers.
// It is unlikely, if an internal register is the only use of floating point,
// that it will select a callee-save register. But to be safe, we restrict
// the set of candidates if compFloatingPointUsed is not already set.
regMaskTP LinearScan::internalFloatRegCandidates()
{
if (compiler->compFloatingPointUsed)
{
return allRegs(TYP_FLOAT);
}
else
{
return RBM_FLT_CALLEE_TRASH;
}
}
/*****************************************************************************
* Register types
*****************************************************************************/
template <class T>
RegisterType regType(T type)
{
#ifdef FEATURE_SIMD
if (varTypeIsSIMD(type))
{
return FloatRegisterType;
}
#endif // FEATURE_SIMD
return varTypeIsFloating(TypeGet(type)) ? FloatRegisterType : IntRegisterType;
}
bool useFloatReg(var_types type)
{
return (regType(type) == FloatRegisterType);
}
bool registerTypesEquivalent(RegisterType a, RegisterType b)
{
return varTypeIsIntegralOrI(a) == varTypeIsIntegralOrI(b);
}
bool isSingleRegister(regMaskTP regMask)
{
return (regMask != RBM_NONE && genMaxOneBit(regMask));
}
/*****************************************************************************
* Inline functions for RegRecord
*****************************************************************************/
bool RegRecord::isFree()
{
return ((assignedInterval == nullptr || !assignedInterval->isActive) && !isBusyUntilNextKill);
}
/*****************************************************************************
* Inline functions for LinearScan
*****************************************************************************/
RegRecord* LinearScan::getRegisterRecord(regNumber regNum)
{
assert((unsigned)regNum < ArrLen(physRegs));
return &physRegs[regNum];
}
#ifdef DEBUG
//----------------------------------------------------------------------------
// getConstrainedRegMask: Returns new regMask which is the intersection of
// regMaskActual and regMaskConstraint if the new regMask has at least
// minRegCount registers, otherwise returns regMaskActual.
//
// Arguments:
// regMaskActual - regMask that needs to be constrained
// regMaskConstraint - regMask constraint that needs to be
// applied to regMaskActual
// minRegCount - Minimum number of regs that should be
// be present in new regMask.
//
// Return Value:
// New regMask that has minRegCount registers after instersection.
// Otherwise returns regMaskActual.
regMaskTP LinearScan::getConstrainedRegMask(regMaskTP regMaskActual, regMaskTP regMaskConstraint, unsigned minRegCount)
{
regMaskTP newMask = regMaskActual & regMaskConstraint;
if (genCountBits(newMask) >= minRegCount)
{
return newMask;
}
return regMaskActual;
}
//------------------------------------------------------------------------
// stressLimitRegs: Given a set of registers, expressed as a register mask, reduce
// them based on the current stress options.
//
// Arguments:
// mask - The current mask of register candidates for a node
//
// Return Value:
// A possibly-modified mask, based on the value of COMPlus_JitStressRegs.
//
// Notes:
// This is the method used to implement the stress options that limit
// the set of registers considered for allocation.
regMaskTP LinearScan::stressLimitRegs(RefPosition* refPosition, regMaskTP mask)
{
if (getStressLimitRegs() != LSRA_LIMIT_NONE)
{
// The refPosition could be null, for example when called
// by getTempRegForResolution().
int minRegCount = (refPosition != nullptr) ? refPosition->minRegCandidateCount : 1;
switch (getStressLimitRegs())
{
case LSRA_LIMIT_CALLEE:
if (!compiler->opts.compDbgEnC)
{
mask = getConstrainedRegMask(mask, RBM_CALLEE_SAVED, minRegCount);
}
break;
case LSRA_LIMIT_CALLER:
{
mask = getConstrainedRegMask(mask, RBM_CALLEE_TRASH, minRegCount);
}
break;
case LSRA_LIMIT_SMALL_SET:
if ((mask & LsraLimitSmallIntSet) != RBM_NONE)
{
mask = getConstrainedRegMask(mask, LsraLimitSmallIntSet, minRegCount);
}
else if ((mask & LsraLimitSmallFPSet) != RBM_NONE)
{
mask = getConstrainedRegMask(mask, LsraLimitSmallFPSet, minRegCount);
}
break;
default:
unreached();
}
if (refPosition != nullptr && refPosition->isFixedRegRef)
{
mask |= refPosition->registerAssignment;
}
}
return mask;
}
#endif // DEBUG
// TODO-Cleanup: Consider adding an overload that takes a varDsc, and can appropriately
// set such fields as isStructField
Interval* LinearScan::newInterval(RegisterType theRegisterType)
{
intervals.emplace_back(theRegisterType, allRegs(theRegisterType));
Interval* newInt = &intervals.back();
#ifdef DEBUG
newInt->intervalIndex = static_cast<unsigned>(intervals.size() - 1);
#endif // DEBUG
DBEXEC(VERBOSE, newInt->dump());
return newInt;
}
RefPosition* LinearScan::newRefPositionRaw(LsraLocation nodeLocation, GenTree* treeNode, RefType refType)
{
refPositions.emplace_back(curBBNum, nodeLocation, treeNode, refType);
RefPosition* newRP = &refPositions.back();
#ifdef DEBUG
newRP->rpNum = static_cast<unsigned>(refPositions.size() - 1);
#endif // DEBUG
return newRP;
}
//------------------------------------------------------------------------
// resolveConflictingDefAndUse: Resolve the situation where we have conflicting def and use
// register requirements on a single-def, single-use interval.
//
// Arguments:
// defRefPosition - The interval definition
// useRefPosition - The (sole) interval use
//
// Return Value:
// None.
//
// Assumptions:
// The two RefPositions are for the same interval, which is a tree-temp.
//
// Notes:
// We require some special handling for the case where the use is a "delayRegFree" case of a fixedReg.
// In that case, if we change the registerAssignment on the useRefPosition, we will lose the fact that,
// even if we assign a different register (and rely on codegen to do the copy), that fixedReg also needs
// to remain busy until the Def register has been allocated. In that case, we don't allow Case 1 or Case 4
// below.
// Here are the cases we consider (in this order):
// 1. If The defRefPosition specifies a single register, and there are no conflicting
// FixedReg uses of it between the def and use, we use that register, and the code generator
// will insert the copy. Note that it cannot be in use because there is a FixedRegRef for the def.
// 2. If the useRefPosition specifies a single register, and it is not in use, and there are no
// conflicting FixedReg uses of it between the def and use, we use that register, and the code generator
// will insert the copy.
// 3. If the defRefPosition specifies a single register (but there are conflicts, as determined
// in 1.), and there are no conflicts with the useRefPosition register (if it's a single register),
/// we set the register requirements on the defRefPosition to the use registers, and the
// code generator will insert a copy on the def. We can't rely on the code generator to put a copy
// on the use if it has multiple possible candidates, as it won't know which one has been allocated.
// 4. If the useRefPosition specifies a single register, and there are no conflicts with the register
// on the defRefPosition, we leave the register requirements on the defRefPosition as-is, and set
// the useRefPosition to the def registers, for similar reasons to case #3.
// 5. If both the defRefPosition and the useRefPosition specify single registers, but both have conflicts,
// We set the candiates on defRefPosition to be all regs of the appropriate type, and since they are
// single registers, codegen can insert the copy.
// 6. Finally, if the RefPositions specify disjoint subsets of the registers (or the use is fixed but
// has a conflict), we must insert a copy. The copy will be inserted before the use if the
// use is not fixed (in the fixed case, the code generator will insert the use).
//
// TODO-CQ: We get bad register allocation in case #3 in the situation where no register is
// available for the lifetime. We end up allocating a register that must be spilled, and it probably
// won't be the register that is actually defined by the target instruction. So, we have to copy it
// and THEN spill it. In this case, we should be using the def requirement. But we need to change
// the interface to this method a bit to make that work (e.g. returning a candidate set to use, but
// leaving the registerAssignment as-is on the def, so that if we find that we need to spill anyway
// we can use the fixed-reg on the def.
//
void LinearScan::resolveConflictingDefAndUse(Interval* interval, RefPosition* defRefPosition)
{
assert(!interval->isLocalVar);
RefPosition* useRefPosition = defRefPosition->nextRefPosition;
regMaskTP defRegAssignment = defRefPosition->registerAssignment;
regMaskTP useRegAssignment = useRefPosition->registerAssignment;
RegRecord* defRegRecord = nullptr;
RegRecord* useRegRecord = nullptr;
regNumber defReg = REG_NA;
regNumber useReg = REG_NA;
bool defRegConflict = false;
bool useRegConflict = false;
// If the useRefPosition is a "delayRegFree", we can't change the registerAssignment
// on it, or we will fail to ensure that the fixedReg is busy at the time the target
// (of the node that uses this interval) is allocated.
bool canChangeUseAssignment = !useRefPosition->isFixedRegRef || !useRefPosition->delayRegFree;
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CONFLICT));
if (!canChangeUseAssignment)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_FIXED_DELAY_USE));
}
if (defRefPosition->isFixedRegRef)
{
defReg = defRefPosition->assignedReg();
defRegRecord = getRegisterRecord(defReg);
if (canChangeUseAssignment)
{
RefPosition* currFixedRegRefPosition = defRegRecord->recentRefPosition;
assert(currFixedRegRefPosition != nullptr &&
currFixedRegRefPosition->nodeLocation == defRefPosition->nodeLocation);
if (currFixedRegRefPosition->nextRefPosition == nullptr ||
currFixedRegRefPosition->nextRefPosition->nodeLocation > useRefPosition->getRefEndLocation())
{
// This is case #1. Use the defRegAssignment
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CASE1));
useRefPosition->registerAssignment = defRegAssignment;
return;
}
else
{
defRegConflict = true;
}
}
}
if (useRefPosition->isFixedRegRef)
{
useReg = useRefPosition->assignedReg();
useRegRecord = getRegisterRecord(useReg);
RefPosition* currFixedRegRefPosition = useRegRecord->recentRefPosition;
// We know that useRefPosition is a fixed use, so the nextRefPosition must not be null.
RefPosition* nextFixedRegRefPosition = useRegRecord->getNextRefPosition();
assert(nextFixedRegRefPosition != nullptr &&
nextFixedRegRefPosition->nodeLocation <= useRefPosition->nodeLocation);
// First, check to see if there are any conflicting FixedReg references between the def and use.
if (nextFixedRegRefPosition->nodeLocation == useRefPosition->nodeLocation)
{
// OK, no conflicting FixedReg references.
// Now, check to see whether it is currently in use.
if (useRegRecord->assignedInterval != nullptr)
{
RefPosition* possiblyConflictingRef = useRegRecord->assignedInterval->recentRefPosition;
LsraLocation possiblyConflictingRefLocation = possiblyConflictingRef->getRefEndLocation();
if (possiblyConflictingRefLocation >= defRefPosition->nodeLocation)
{
useRegConflict = true;
}
}
if (!useRegConflict)
{
// This is case #2. Use the useRegAssignment
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CASE2));
defRefPosition->registerAssignment = useRegAssignment;
return;
}
}
else
{
useRegConflict = true;
}
}
if (defRegRecord != nullptr && !useRegConflict)
{
// This is case #3.
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CASE3));
defRefPosition->registerAssignment = useRegAssignment;
return;
}
if (useRegRecord != nullptr && !defRegConflict && canChangeUseAssignment)
{
// This is case #4.
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CASE4));
useRefPosition->registerAssignment = defRegAssignment;
return;
}
if (defRegRecord != nullptr && useRegRecord != nullptr)
{
// This is case #5.
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CASE5));
RegisterType regType = interval->registerType;
assert((getRegisterType(interval, defRefPosition) == regType) &&
(getRegisterType(interval, useRefPosition) == regType));
regMaskTP candidates = allRegs(regType);
defRefPosition->registerAssignment = candidates;
return;
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DEFUSE_CASE6));
return;
}
//------------------------------------------------------------------------
// conflictingFixedRegReference: Determine whether the current RegRecord has a
// fixed register use that conflicts with 'refPosition'
//
// Arguments:
// refPosition - The RefPosition of interest
//
// Return Value:
// Returns true iff the given RefPosition is NOT a fixed use of this register,
// AND either:
// - there is a RefPosition on this RegRecord at the nodeLocation of the given RefPosition, or
// - the given RefPosition has a delayRegFree, and there is a RefPosition on this RegRecord at
// the nodeLocation just past the given RefPosition.
//
// Assumptions:
// 'refPosition is non-null.
bool RegRecord::conflictingFixedRegReference(RefPosition* refPosition)
{
// Is this a fixed reference of this register? If so, there is no conflict.
if (refPosition->isFixedRefOfRegMask(genRegMask(regNum)))
{
return false;
}
// Otherwise, check for conflicts.
// There is a conflict if:
// 1. There is a recent RefPosition on this RegRecord that is at this location,
// except in the case where it is a special "putarg" that is associated with this interval, OR
// 2. There is an upcoming RefPosition at this location, or at the next location
// if refPosition is a delayed use (i.e. must be kept live through the next/def location).
LsraLocation refLocation = refPosition->nodeLocation;
if (recentRefPosition != nullptr && recentRefPosition->refType != RefTypeKill &&
recentRefPosition->nodeLocation == refLocation &&
(!isBusyUntilNextKill || assignedInterval != refPosition->getInterval()))
{
return true;
}
LsraLocation nextPhysRefLocation = getNextRefLocation();
if (nextPhysRefLocation == refLocation || (refPosition->delayRegFree && nextPhysRefLocation == (refLocation + 1)))
{
return true;
}
return false;
}
void LinearScan::applyCalleeSaveHeuristics(RefPosition* rp)
{
#ifdef _TARGET_AMD64_
if (compiler->opts.compDbgEnC)
{
// We only use RSI and RDI for EnC code, so we don't want to favor callee-save regs.
return;
}
#endif // _TARGET_AMD64_
Interval* theInterval = rp->getInterval();
#ifdef DEBUG
regMaskTP calleeSaveMask = calleeSaveRegs(getRegisterType(theInterval, rp));
if (doReverseCallerCallee())
{
rp->registerAssignment =
getConstrainedRegMask(rp->registerAssignment, calleeSaveMask, rp->minRegCandidateCount);
}
else
#endif // DEBUG
{
// Set preferences so that this register set will be preferred for earlier refs
theInterval->updateRegisterPreferences(rp->registerAssignment);
}
}
void LinearScan::associateRefPosWithInterval(RefPosition* rp)
{
Referenceable* theReferent = rp->referent;
if (theReferent != nullptr)
{
// All RefPositions except the dummy ones at the beginning of blocks
if (rp->isIntervalRef())
{
Interval* theInterval = rp->getInterval();
applyCalleeSaveHeuristics(rp);
if (theInterval->isLocalVar)
{
if (RefTypeIsUse(rp->refType))
{
RefPosition* const prevRP = theInterval->recentRefPosition;
if ((prevRP != nullptr) && (prevRP->bbNum == rp->bbNum))
{
prevRP->lastUse = false;
}
}
rp->lastUse = (rp->refType != RefTypeExpUse) && (rp->refType != RefTypeParamDef) &&
(rp->refType != RefTypeZeroInit) && !extendLifetimes();
}
else if (rp->refType == RefTypeUse)
{
// Ensure that we have consistent def/use on SDSU temps.
// However, there are a couple of cases where this may over-constrain allocation:
// 1. In the case of a non-commutative rmw def (in which the rmw source must be delay-free), or
// 2. In the case where the defining node requires a temp distinct from the target (also a
// delay-free case).
// In those cases, if we propagate a single-register restriction from the consumer to the producer
// the delayed uses will not see a fixed reference in the PhysReg at that position, and may
// incorrectly allocate that register.
// TODO-CQ: This means that we may often require a copy at the use of this node's result.
// This case could be moved to BuildRefPositionsForNode, at the point where the def RefPosition is
// created, causing a RefTypeFixedRef to be added at that location. This, however, results in
// more PhysReg RefPositions (a throughput impact), and a large number of diffs that require
// further analysis to determine benefit.
// See Issue #11274.
RefPosition* prevRefPosition = theInterval->recentRefPosition;
assert(prevRefPosition != nullptr && theInterval->firstRefPosition == prevRefPosition);
// All defs must have a valid treeNode, but we check it below to be conservative.
assert(prevRefPosition->treeNode != nullptr);
regMaskTP prevAssignment = prevRefPosition->registerAssignment;
regMaskTP newAssignment = (prevAssignment & rp->registerAssignment);
if (newAssignment != RBM_NONE)
{
if (!isSingleRegister(newAssignment) ||
(!theInterval->hasNonCommutativeRMWDef && (prevRefPosition->treeNode != nullptr) &&
!prevRefPosition->treeNode->gtLsraInfo.isInternalRegDelayFree))
{
prevRefPosition->registerAssignment = newAssignment;
}
}
else
{
theInterval->hasConflictingDefUse = true;
}
rp->lastUse = true;
}
}
RefPosition* prevRP = theReferent->recentRefPosition;
if (prevRP != nullptr)
{
prevRP->nextRefPosition = rp;
}
else
{
theReferent->firstRefPosition = rp;
}
theReferent->recentRefPosition = rp;
theReferent->lastRefPosition = rp;
}
else
{
assert((rp->refType == RefTypeBB) || (rp->refType == RefTypeKillGCRefs));
}
}
//---------------------------------------------------------------------------
// newRefPosition: allocate and initialize a new RefPosition.
//
// Arguments:
// reg - reg number that identifies RegRecord to be associated
// with this RefPosition
// theLocation - LSRA location of RefPosition
// theRefType - RefPosition type
// theTreeNode - GenTree node for which this RefPosition is created
// mask - Set of valid registers for this RefPosition
// multiRegIdx - register position if this RefPosition corresponds to a
// multi-reg call node.
//
// Return Value:
// a new RefPosition
//
RefPosition* LinearScan::newRefPosition(
regNumber reg, LsraLocation theLocation, RefType theRefType, GenTree* theTreeNode, regMaskTP mask)
{
RefPosition* newRP = newRefPositionRaw(theLocation, theTreeNode, theRefType);
newRP->setReg(getRegisterRecord(reg));
newRP->registerAssignment = mask;
newRP->setMultiRegIdx(0);
newRP->setAllocateIfProfitable(false);
associateRefPosWithInterval(newRP);
DBEXEC(VERBOSE, newRP->dump());
return newRP;
}
//---------------------------------------------------------------------------
// newRefPosition: allocate and initialize a new RefPosition.
//
// Arguments:
// theInterval - interval to which RefPosition is associated with.
// theLocation - LSRA location of RefPosition
// theRefType - RefPosition type
// theTreeNode - GenTree node for which this RefPosition is created
// mask - Set of valid registers for this RefPosition
// multiRegIdx - register position if this RefPosition corresponds to a
// multi-reg call node.
// minRegCount - Minimum number registers that needs to be ensured while
// constraining candidates for this ref position under
// LSRA stress. This is a DEBUG only arg.
//
// Return Value:
// a new RefPosition
//
RefPosition* LinearScan::newRefPosition(Interval* theInterval,
LsraLocation theLocation,
RefType theRefType,
GenTree* theTreeNode,
regMaskTP mask,
unsigned multiRegIdx /* = 0 */
DEBUGARG(unsigned minRegCandidateCount /* = 1 */))
{
#ifdef DEBUG
if (theInterval != nullptr && regType(theInterval->registerType) == FloatRegisterType)
{
// In the case we're using floating point registers we must make sure
// this flag was set previously in the compiler since this will mandate
// whether LSRA will take into consideration FP reg killsets.
assert(compiler->compFloatingPointUsed || ((mask & RBM_FLT_CALLEE_SAVED) == 0));
}
#endif // DEBUG
// If this reference is constrained to a single register (and it's not a dummy
// or Kill reftype already), add a RefTypeFixedReg at this location so that its
// availability can be more accurately determined
bool isFixedRegister = isSingleRegister(mask);
bool insertFixedRef = false;
if (isFixedRegister)
{
// Insert a RefTypeFixedReg for any normal def or use (not ParamDef or BB)
if (theRefType == RefTypeUse || theRefType == RefTypeDef)
{
insertFixedRef = true;
}
}
if (insertFixedRef)
{
regNumber physicalReg = genRegNumFromMask(mask);
RefPosition* pos = newRefPosition(physicalReg, theLocation, RefTypeFixedReg, nullptr, mask);
assert(theInterval != nullptr);
assert((allRegs(theInterval->registerType) & mask) != 0);
}
RefPosition* newRP = newRefPositionRaw(theLocation, theTreeNode, theRefType);
newRP->setInterval(theInterval);
// Spill info
newRP->isFixedRegRef = isFixedRegister;
#ifndef _TARGET_AMD64_
// We don't need this for AMD because the PInvoke method epilog code is explicit
// at register allocation time.
if (theInterval != nullptr && theInterval->isLocalVar && compiler->info.compCallUnmanaged &&
theInterval->varNum == compiler->genReturnLocal)
{
mask &= ~(RBM_PINVOKE_TCB | RBM_PINVOKE_FRAME);
noway_assert(mask != RBM_NONE);
}
#endif // !_TARGET_AMD64_
newRP->registerAssignment = mask;
newRP->setMultiRegIdx(multiRegIdx);
newRP->setAllocateIfProfitable(false);
#ifdef DEBUG
newRP->minRegCandidateCount = minRegCandidateCount;
#endif // DEBUG
associateRefPosWithInterval(newRP);
DBEXEC(VERBOSE, newRP->dump());
return newRP;
}
/*****************************************************************************
* Inline functions for Interval
*****************************************************************************/
RefPosition* Referenceable::getNextRefPosition()
{
if (recentRefPosition == nullptr)
{
return firstRefPosition;
}
else
{
return recentRefPosition->nextRefPosition;
}
}
LsraLocation Referenceable::getNextRefLocation()
{
RefPosition* nextRefPosition = getNextRefPosition();
if (nextRefPosition == nullptr)
{
return MaxLocation;
}
else
{
return nextRefPosition->nodeLocation;
}
}
// Iterate through all the registers of the given type
class RegisterIterator
{
friend class Registers;
public:
RegisterIterator(RegisterType type) : regType(type)
{
if (useFloatReg(regType))
{
currentRegNum = REG_FP_FIRST;
}
else
{
currentRegNum = REG_INT_FIRST;
}
}
protected:
static RegisterIterator Begin(RegisterType regType)
{
return RegisterIterator(regType);
}
static RegisterIterator End(RegisterType regType)
{
RegisterIterator endIter = RegisterIterator(regType);
// This assumes only integer and floating point register types
// if we target a processor with additional register types,
// this would have to change
if (useFloatReg(regType))
{
// This just happens to work for both double & float
endIter.currentRegNum = REG_NEXT(REG_FP_LAST);
}
else
{
endIter.currentRegNum = REG_NEXT(REG_INT_LAST);
}
return endIter;
}
public:
void operator++(int dummy) // int dummy is c++ for "this is postfix ++"
{
currentRegNum = REG_NEXT(currentRegNum);
#ifdef _TARGET_ARM_
if (regType == TYP_DOUBLE)
currentRegNum = REG_NEXT(currentRegNum);
#endif
}
void operator++() // prefix operator++
{
currentRegNum = REG_NEXT(currentRegNum);
#ifdef _TARGET_ARM_
if (regType == TYP_DOUBLE)
currentRegNum = REG_NEXT(currentRegNum);
#endif
}
regNumber operator*()
{
return currentRegNum;
}
bool operator!=(const RegisterIterator& other)
{
return other.currentRegNum != currentRegNum;
}
private:
regNumber currentRegNum;
RegisterType regType;
};
class Registers
{
public:
friend class RegisterIterator;
RegisterType type;
Registers(RegisterType t)
{
type = t;
}
RegisterIterator begin()
{
return RegisterIterator::Begin(type);
}
RegisterIterator end()
{
return RegisterIterator::End(type);
}
};
#ifdef DEBUG
void LinearScan::dumpVarToRegMap(VarToRegMap map)
{
bool anyPrinted = false;
for (unsigned varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++)
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
if (map[varIndex] != REG_STK)
{
printf("V%02u=%s ", varNum, getRegName(map[varIndex]));
anyPrinted = true;
}
}
if (!anyPrinted)
{
printf("none");
}
printf("\n");
}
void LinearScan::dumpInVarToRegMap(BasicBlock* block)
{
printf("Var=Reg beg of BB%02u: ", block->bbNum);
VarToRegMap map = getInVarToRegMap(block->bbNum);
dumpVarToRegMap(map);
}
void LinearScan::dumpOutVarToRegMap(BasicBlock* block)
{
printf("Var=Reg end of BB%02u: ", block->bbNum);
VarToRegMap map = getOutVarToRegMap(block->bbNum);
dumpVarToRegMap(map);
}
#endif // DEBUG
LinearScanInterface* getLinearScanAllocator(Compiler* comp)
{
return new (comp, CMK_LSRA) LinearScan(comp);
}
//------------------------------------------------------------------------
// LSRA constructor
//
// Arguments:
// theCompiler
//
// Notes:
// The constructor takes care of initializing the data structures that are used
// during Lowering, including (in DEBUG) getting the stress environment variables,
// as they may affect the block ordering.
LinearScan::LinearScan(Compiler* theCompiler)
: compiler(theCompiler)
#if MEASURE_MEM_ALLOC
, lsraIAllocator(nullptr)
#endif // MEASURE_MEM_ALLOC
, intervals(LinearScanMemoryAllocatorInterval(theCompiler))
, refPositions(LinearScanMemoryAllocatorRefPosition(theCompiler))
{
#ifdef DEBUG
maxNodeLocation = 0;
activeRefPosition = nullptr;
// Get the value of the environment variable that controls stress for register allocation
lsraStressMask = JitConfig.JitStressRegs();
#if 0
#ifdef DEBUG
if (lsraStressMask != 0)
{
// The code in this #if can be used to debug JitStressRegs issues according to
// method hash. To use, simply set environment variables JitStressRegsHashLo and JitStressRegsHashHi
unsigned methHash = compiler->info.compMethodHash();
char* lostr = getenv("JitStressRegsHashLo");
unsigned methHashLo = 0;
bool dump = false;
if (lostr != nullptr)
{
sscanf_s(lostr, "%x", &methHashLo);
dump = true;
}
char* histr = getenv("JitStressRegsHashHi");
unsigned methHashHi = UINT32_MAX;
if (histr != nullptr)
{
sscanf_s(histr, "%x", &methHashHi);
dump = true;
}
if (methHash < methHashLo || methHash > methHashHi)
{
lsraStressMask = 0;
}
else if (dump == true)
{
printf("JitStressRegs = %x for method %s, hash = 0x%x.\n",
lsraStressMask, compiler->info.compFullName, compiler->info.compMethodHash());
printf(""); // in our logic this causes a flush
}
}
#endif // DEBUG
#endif
dumpTerse = (JitConfig.JitDumpTerseLsra() != 0);
#endif // DEBUG
enregisterLocalVars = ((compiler->opts.compFlags & CLFLG_REGVAR) != 0) && compiler->lvaTrackedCount > 0;
availableIntRegs = (RBM_ALLINT & ~compiler->codeGen->regSet.rsMaskResvd);
#if ETW_EBP_FRAMED
availableIntRegs &= ~RBM_FPBASE;
#endif // ETW_EBP_FRAMED
availableFloatRegs = RBM_ALLFLOAT;
availableDoubleRegs = RBM_ALLDOUBLE;
#ifdef _TARGET_AMD64_
if (compiler->opts.compDbgEnC)
{
// On x64 when the EnC option is set, we always save exactly RBP, RSI and RDI.
// RBP is not available to the register allocator, so RSI and RDI are the only
// callee-save registers available.
availableIntRegs &= ~RBM_CALLEE_SAVED | RBM_RSI | RBM_RDI;
availableFloatRegs &= ~RBM_CALLEE_SAVED;
availableDoubleRegs &= ~RBM_CALLEE_SAVED;
}
#endif // _TARGET_AMD64_
compiler->rpFrameType = FT_NOT_SET;
compiler->rpMustCreateEBPCalled = false;
compiler->codeGen->intRegState.rsIsFloat = false;
compiler->codeGen->floatRegState.rsIsFloat = true;
// Block sequencing (the order in which we schedule).
// Note that we don't initialize the bbVisitedSet until we do the first traversal
// (currently during Lowering's second phase, where it sets the TreeNodeInfo).
// This is so that any blocks that are added during the first phase of Lowering
// are accounted for (and we don't have BasicBlockEpoch issues).
blockSequencingDone = false;
blockSequence = nullptr;
blockSequenceWorkList = nullptr;
curBBSeqNum = 0;
bbSeqCount = 0;
// Information about each block, including predecessor blocks used for variable locations at block entry.
blockInfo = nullptr;
// Populate the register mask table.
// The first two masks in the table are allint/allfloat
// The next N are the masks for each single register.
// After that are the dynamically added ones.
regMaskTable = new (compiler, CMK_LSRA) regMaskTP[numMasks];
regMaskTable[ALLINT_IDX] = allRegs(TYP_INT);
regMaskTable[ALLFLOAT_IDX] = allRegs(TYP_DOUBLE);
regNumber reg;
for (reg = REG_FIRST; reg < REG_COUNT; reg = REG_NEXT(reg))
{
regMaskTable[FIRST_SINGLE_REG_IDX + reg - REG_FIRST] = (reg == REG_STK) ? RBM_NONE : genRegMask(reg);
}
nextFreeMask = FIRST_SINGLE_REG_IDX + REG_COUNT;
noway_assert(nextFreeMask <= numMasks);
}
// Return the reg mask corresponding to the given index.
regMaskTP LinearScan::GetRegMaskForIndex(RegMaskIndex index)
{
assert(index < numMasks);
assert(index < nextFreeMask);
return regMaskTable[index];
}
// Given a reg mask, return the index it corresponds to. If it is not a 'well known' reg mask,
// add it at the end. This method has linear behavior in the worst cases but that is fairly rare.
// Most methods never use any but the well-known masks, and when they do use more
// it is only one or two more.
LinearScan::RegMaskIndex LinearScan::GetIndexForRegMask(regMaskTP mask)
{
RegMaskIndex result;
if (isSingleRegister(mask))
{
result = genRegNumFromMask(mask) + FIRST_SINGLE_REG_IDX;
}
else if (mask == allRegs(TYP_INT))
{
result = ALLINT_IDX;
}
else if (mask == allRegs(TYP_DOUBLE))
{
result = ALLFLOAT_IDX;
}
else
{
for (int i = FIRST_SINGLE_REG_IDX + REG_COUNT; i < nextFreeMask; i++)
{
if (regMaskTable[i] == mask)
{
return i;
}
}
// We only allocate a fixed number of masks. Since we don't reallocate, we will throw a
// noway_assert if we exceed this limit.
noway_assert(nextFreeMask < numMasks);
regMaskTable[nextFreeMask] = mask;
result = nextFreeMask;
nextFreeMask++;
}
assert(mask == regMaskTable[result]);
return result;
}
// We've decided that we can't use a register during register allocation (probably FPBASE),
// but we've already added it to the register masks. Go through the masks and remove it.
void LinearScan::RemoveRegisterFromMasks(regNumber reg)
{
JITDUMP("Removing register %s from LSRA register masks\n", getRegName(reg));
regMaskTP mask = ~genRegMask(reg);
for (int i = 0; i < nextFreeMask; i++)
{
regMaskTable[i] &= mask;
}
JITDUMP("After removing register:\n");
DBEXEC(VERBOSE, dspRegisterMaskTable());
}
#ifdef DEBUG
void LinearScan::dspRegisterMaskTable()
{
printf("LSRA register masks. Total allocated: %d, total used: %d\n", numMasks, nextFreeMask);
for (int i = 0; i < nextFreeMask; i++)
{
printf("%2u: ", i);
dspRegMask(regMaskTable[i]);
printf("\n");
}
}
#endif // DEBUG
//------------------------------------------------------------------------
// getNextCandidateFromWorkList: Get the next candidate for block sequencing
//
// Arguments:
// None.
//
// Return Value:
// The next block to be placed in the sequence.
//
// Notes:
// This method currently always returns the next block in the list, and relies on having
// blocks added to the list only when they are "ready", and on the
// addToBlockSequenceWorkList() method to insert them in the proper order.
// However, a block may be in the list and already selected, if it was subsequently
// encountered as both a flow and layout successor of the most recently selected
// block.
BasicBlock* LinearScan::getNextCandidateFromWorkList()
{
BasicBlockList* nextWorkList = nullptr;
for (BasicBlockList* workList = blockSequenceWorkList; workList != nullptr; workList = nextWorkList)
{
nextWorkList = workList->next;
BasicBlock* candBlock = workList->block;
removeFromBlockSequenceWorkList(workList, nullptr);
if (!isBlockVisited(candBlock))
{
return candBlock;
}
}
return nullptr;
}
//------------------------------------------------------------------------
// setBlockSequence:Determine the block order for register allocation.
//
// Arguments:
// None
//
// Return Value:
// None
//
// Notes:
// On return, the blockSequence array contains the blocks, in the order in which they
// will be allocated.
// This method clears the bbVisitedSet on LinearScan, and when it returns the set
// contains all the bbNums for the block.
// This requires a traversal of the BasicBlocks, and could potentially be
// combined with the first traversal (currently the one in Lowering that sets the
// TreeNodeInfo).
void LinearScan::setBlockSequence()
{
// Reset the "visited" flag on each block.
compiler->EnsureBasicBlockEpoch();
bbVisitedSet = BlockSetOps::MakeEmpty(compiler);
BlockSet readySet(BlockSetOps::MakeEmpty(compiler));
BlockSet predSet(BlockSetOps::MakeEmpty(compiler));
assert(blockSequence == nullptr && bbSeqCount == 0);
blockSequence = new (compiler, CMK_LSRA) BasicBlock*[compiler->fgBBcount];
bbNumMaxBeforeResolution = compiler->fgBBNumMax;
blockInfo = new (compiler, CMK_LSRA) LsraBlockInfo[bbNumMaxBeforeResolution + 1];
assert(blockSequenceWorkList == nullptr);
bool addedInternalBlocks = false;
verifiedAllBBs = false;
hasCriticalEdges = false;
BasicBlock* nextBlock;
// We use a bbNum of 0 for entry RefPositions.
// The other information in blockInfo[0] will never be used.
blockInfo[0].weight = BB_UNITY_WEIGHT;
for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = nextBlock)
{
blockSequence[bbSeqCount] = block;
markBlockVisited(block);
bbSeqCount++;
nextBlock = nullptr;
// Initialize the blockInfo.
// predBBNum will be set later. 0 is never used as a bbNum.
assert(block->bbNum != 0);
blockInfo[block->bbNum].predBBNum = 0;
// We check for critical edges below, but initialize to false.
blockInfo[block->bbNum].hasCriticalInEdge = false;
blockInfo[block->bbNum].hasCriticalOutEdge = false;
blockInfo[block->bbNum].weight = block->getBBWeight(compiler);
#if TRACK_LSRA_STATS
blockInfo[block->bbNum].spillCount = 0;
blockInfo[block->bbNum].copyRegCount = 0;
blockInfo[block->bbNum].resolutionMovCount = 0;
blockInfo[block->bbNum].splitEdgeCount = 0;
#endif // TRACK_LSRA_STATS
if (block->GetUniquePred(compiler) == nullptr)
{
for (flowList* pred = block->bbPreds; pred != nullptr; pred = pred->flNext)
{
BasicBlock* predBlock = pred->flBlock;
if (predBlock->NumSucc(compiler) > 1)
{
blockInfo[block->bbNum].hasCriticalInEdge = true;
hasCriticalEdges = true;
break;
}
else if (predBlock->bbJumpKind == BBJ_SWITCH)
{
assert(!"Switch with single successor");
}
}
}
// Determine which block to schedule next.
// First, update the NORMAL successors of the current block, adding them to the worklist
// according to the desired order. We will handle the EH successors below.
bool checkForCriticalOutEdge = (block->NumSucc(compiler) > 1);
if (!checkForCriticalOutEdge && block->bbJumpKind == BBJ_SWITCH)
{
assert(!"Switch with single successor");
}
const unsigned numSuccs = block->NumSucc(compiler);
for (unsigned succIndex = 0; succIndex < numSuccs; succIndex++)
{
BasicBlock* succ = block->GetSucc(succIndex, compiler);
if (checkForCriticalOutEdge && succ->GetUniquePred(compiler) == nullptr)
{
blockInfo[block->bbNum].hasCriticalOutEdge = true;
hasCriticalEdges = true;
// We can stop checking now.
checkForCriticalOutEdge = false;
}
if (isTraversalLayoutOrder() || isBlockVisited(succ))
{
continue;
}
// We've now seen a predecessor, so add it to the work list and the "readySet".
// It will be inserted in the worklist according to the specified traversal order
// (i.e. pred-first or random, since layout order is handled above).
if (!BlockSetOps::IsMember(compiler, readySet, succ->bbNum))
{
addToBlockSequenceWorkList(readySet, succ, predSet);
BlockSetOps::AddElemD(compiler, readySet, succ->bbNum);
}
}
// For layout order, simply use bbNext
if (isTraversalLayoutOrder())
{
nextBlock = block->bbNext;
continue;
}
while (nextBlock == nullptr)
{
nextBlock = getNextCandidateFromWorkList();
// TODO-Throughput: We would like to bypass this traversal if we know we've handled all
// the blocks - but fgBBcount does not appear to be updated when blocks are removed.
if (nextBlock == nullptr /* && bbSeqCount != compiler->fgBBcount*/ && !verifiedAllBBs)
{
// If we don't encounter all blocks by traversing the regular sucessor links, do a full
// traversal of all the blocks, and add them in layout order.
// This may include:
// - internal-only blocks (in the fgAddCodeList) which may not be in the flow graph
// (these are not even in the bbNext links).
// - blocks that have become unreachable due to optimizations, but that are strongly
// connected (these are not removed)
// - EH blocks
for (Compiler::AddCodeDsc* desc = compiler->fgAddCodeList; desc != nullptr; desc = desc->acdNext)
{
if (!isBlockVisited(block))
{
addToBlockSequenceWorkList(readySet, block, predSet);
BlockSetOps::AddElemD(compiler, readySet, block->bbNum);
}
}
for (BasicBlock* block = compiler->fgFirstBB; block; block = block->bbNext)
{
if (!isBlockVisited(block))
{
addToBlockSequenceWorkList(readySet, block, predSet);
BlockSetOps::AddElemD(compiler, readySet, block->bbNum);
}
}
verifiedAllBBs = true;
}
else
{
break;
}
}
}
blockSequencingDone = true;
#ifdef DEBUG
// Make sure that we've visited all the blocks.
for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = block->bbNext)
{
assert(isBlockVisited(block));
}
JITDUMP("LSRA Block Sequence: ");
int i = 1;
for (BasicBlock *block = startBlockSequence(); block != nullptr; ++i, block = moveToNextBlock())
{
JITDUMP("BB%02u", block->bbNum);
if (block->isMaxBBWeight())
{
JITDUMP("(MAX) ");
}
else
{
JITDUMP("(%6s) ", refCntWtd2str(block->getBBWeight(compiler)));
}
if (i % 10 == 0)
{
JITDUMP("\n ");
}
}
JITDUMP("\n\n");
#endif
}
//------------------------------------------------------------------------
// compareBlocksForSequencing: Compare two basic blocks for sequencing order.
//
// Arguments:
// block1 - the first block for comparison
// block2 - the second block for comparison
// useBlockWeights - whether to use block weights for comparison
//
// Return Value:
// -1 if block1 is preferred.
// 0 if the blocks are equivalent.
// 1 if block2 is preferred.
//
// Notes:
// See addToBlockSequenceWorkList.
int LinearScan::compareBlocksForSequencing(BasicBlock* block1, BasicBlock* block2, bool useBlockWeights)
{
if (useBlockWeights)
{
unsigned weight1 = block1->getBBWeight(compiler);
unsigned weight2 = block2->getBBWeight(compiler);
if (weight1 > weight2)
{
return -1;
}
else if (weight1 < weight2)
{
return 1;
}
}
// If weights are the same prefer LOWER bbnum
if (block1->bbNum < block2->bbNum)
{
return -1;
}
else if (block1->bbNum == block2->bbNum)
{
return 0;
}
else
{
return 1;
}
}
//------------------------------------------------------------------------
// addToBlockSequenceWorkList: Add a BasicBlock to the work list for sequencing.
//
// Arguments:
// sequencedBlockSet - the set of blocks that are already sequenced
// block - the new block to be added
// predSet - the buffer to save predecessors set. A block set allocated by the caller used here as a
// temporary block set for constructing a predecessor set. Allocated by the caller to avoid reallocating a new block
// set with every call to this function
//
// Return Value:
// None.
//
// Notes:
// The first block in the list will be the next one to be sequenced, as soon
// as we encounter a block whose successors have all been sequenced, in pred-first
// order, or the very next block if we are traversing in random order (once implemented).
// This method uses a comparison method to determine the order in which to place
// the blocks in the list. This method queries whether all predecessors of the
// block are sequenced at the time it is added to the list and if so uses block weights
// for inserting the block. A block is never inserted ahead of its predecessors.
// A block at the time of insertion may not have all its predecessors sequenced, in
// which case it will be sequenced based on its block number. Once a block is inserted,
// its priority\order will not be changed later once its remaining predecessors are
// sequenced. This would mean that work list may not be sorted entirely based on
// block weights alone.
//
// Note also that, when random traversal order is implemented, this method
// should insert the blocks into the list in random order, so that we can always
// simply select the first block in the list.
void LinearScan::addToBlockSequenceWorkList(BlockSet sequencedBlockSet, BasicBlock* block, BlockSet& predSet)
{
// The block that is being added is not already sequenced
assert(!BlockSetOps::IsMember(compiler, sequencedBlockSet, block->bbNum));
// Get predSet of block
BlockSetOps::ClearD(compiler, predSet);
flowList* pred;
for (pred = block->bbPreds; pred != nullptr; pred = pred->flNext)
{
BlockSetOps::AddElemD(compiler, predSet, pred->flBlock->bbNum);
}
// If either a rarely run block or all its preds are already sequenced, use block's weight to sequence
bool useBlockWeight = block->isRunRarely() || BlockSetOps::IsSubset(compiler, sequencedBlockSet, predSet);
BasicBlockList* prevNode = nullptr;
BasicBlockList* nextNode = blockSequenceWorkList;
while (nextNode != nullptr)
{
int seqResult;
if (nextNode->block->isRunRarely())
{
// If the block that is yet to be sequenced is a rarely run block, always use block weights for sequencing
seqResult = compareBlocksForSequencing(nextNode->block, block, true);
}
else if (BlockSetOps::IsMember(compiler, predSet, nextNode->block->bbNum))
{
// always prefer unsequenced pred blocks
seqResult = -1;
}
else
{
seqResult = compareBlocksForSequencing(nextNode->block, block, useBlockWeight);
}
if (seqResult > 0)
{
break;
}
prevNode = nextNode;
nextNode = nextNode->next;
}
BasicBlockList* newListNode = new (compiler, CMK_LSRA) BasicBlockList(block, nextNode);
if (prevNode == nullptr)
{
blockSequenceWorkList = newListNode;
}
else
{
prevNode->next = newListNode;
}
}
void LinearScan::removeFromBlockSequenceWorkList(BasicBlockList* listNode, BasicBlockList* prevNode)
{
if (listNode == blockSequenceWorkList)
{
assert(prevNode == nullptr);
blockSequenceWorkList = listNode->next;
}
else
{
assert(prevNode != nullptr && prevNode->next == listNode);
prevNode->next = listNode->next;
}
// TODO-Cleanup: consider merging Compiler::BlockListNode and BasicBlockList
// compiler->FreeBlockListNode(listNode);
}
// Initialize the block order for allocation (called each time a new traversal begins).
BasicBlock* LinearScan::startBlockSequence()
{
if (!blockSequencingDone)
{
setBlockSequence();
}
BasicBlock* curBB = compiler->fgFirstBB;
curBBSeqNum = 0;
curBBNum = curBB->bbNum;
clearVisitedBlocks();
assert(blockSequence[0] == compiler->fgFirstBB);
markBlockVisited(curBB);
return curBB;
}
//------------------------------------------------------------------------
// moveToNextBlock: Move to the next block in order for allocation or resolution.
//
// Arguments:
// None
//
// Return Value:
// The next block.
//
// Notes:
// This method is used when the next block is actually going to be handled.
// It changes curBBNum.
BasicBlock* LinearScan::moveToNextBlock()
{
BasicBlock* nextBlock = getNextBlock();
curBBSeqNum++;
if (nextBlock != nullptr)
{
curBBNum = nextBlock->bbNum;
}
return nextBlock;
}
//------------------------------------------------------------------------
// getNextBlock: Get the next block in order for allocation or resolution.
//
// Arguments:
// None
//
// Return Value:
// The next block.
//
// Notes:
// This method does not actually change the current block - it is used simply
// to determine which block will be next.
BasicBlock* LinearScan::getNextBlock()
{
assert(blockSequencingDone);
unsigned int nextBBSeqNum = curBBSeqNum + 1;
if (nextBBSeqNum < bbSeqCount)
{
return blockSequence[nextBBSeqNum];
}
return nullptr;
}
//------------------------------------------------------------------------
// doLinearScan: The main method for register allocation.
//
// Arguments:
// None
//
// Return Value:
// None.
//
// Assumptions:
// Lowering must have set the NodeInfo (gtLsraInfo) on each node to communicate
// the register requirements.
void LinearScan::doLinearScan()
{
unsigned lsraBlockEpoch = compiler->GetCurBasicBlockEpoch();
splitBBNumToTargetBBNumMap = nullptr;
// This is complicated by the fact that physical registers have refs associated
// with locations where they are killed (e.g. calls), but we don't want to
// count these as being touched.
compiler->codeGen->regSet.rsClearRegsModified();
initMaxSpill();
buildIntervals();
DBEXEC(VERBOSE, TupleStyleDump(LSRA_DUMP_REFPOS));
compiler->EndPhase(PHASE_LINEAR_SCAN_BUILD);
DBEXEC(VERBOSE, lsraDumpIntervals("after buildIntervals"));
clearVisitedBlocks();
initVarRegMaps();
allocateRegisters();
compiler->EndPhase(PHASE_LINEAR_SCAN_ALLOC);
resolveRegisters();
compiler->EndPhase(PHASE_LINEAR_SCAN_RESOLVE);
#if TRACK_LSRA_STATS
if ((JitConfig.DisplayLsraStats() != 0)
#ifdef DEBUG
|| VERBOSE
#endif
)
{
dumpLsraStats(jitstdout);
}
#endif // TRACK_LSRA_STATS
DBEXEC(VERBOSE, TupleStyleDump(LSRA_DUMP_POST));
compiler->compLSRADone = true;
noway_assert(lsraBlockEpoch = compiler->GetCurBasicBlockEpoch());
}
//------------------------------------------------------------------------
// recordVarLocationsAtStartOfBB: Update live-in LclVarDscs with the appropriate
// register location at the start of a block, during codegen.
//
// Arguments:
// bb - the block for which code is about to be generated.
//
// Return Value:
// None.
//
// Assumptions:
// CodeGen will take care of updating the reg masks and the current var liveness,
// after calling this method.
// This is because we need to kill off the dead registers before setting the newly live ones.
void LinearScan::recordVarLocationsAtStartOfBB(BasicBlock* bb)
{
if (!enregisterLocalVars)
{
return;
}
JITDUMP("Recording Var Locations at start of BB%02u\n", bb->bbNum);
VarToRegMap map = getInVarToRegMap(bb->bbNum);
unsigned count = 0;
VarSetOps::AssignNoCopy(compiler, currentLiveVars,
VarSetOps::Intersection(compiler, registerCandidateVars, bb->bbLiveIn));
VarSetOps::Iter iter(compiler, currentLiveVars);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* varDsc = &(compiler->lvaTable[varNum]);
regNumber regNum = getVarReg(map, varIndex);
regNumber oldRegNum = varDsc->lvRegNum;
regNumber newRegNum = regNum;
if (oldRegNum != newRegNum)
{
JITDUMP(" V%02u(%s->%s)", varNum, compiler->compRegVarName(oldRegNum),
compiler->compRegVarName(newRegNum));
varDsc->lvRegNum = newRegNum;
count++;
}
else if (newRegNum != REG_STK)
{
JITDUMP(" V%02u(%s)", varNum, compiler->compRegVarName(newRegNum));
count++;
}
}
if (count == 0)
{
JITDUMP(" <none>\n");
}
JITDUMP("\n");
}
void Interval::setLocalNumber(Compiler* compiler, unsigned lclNum, LinearScan* linScan)
{
LclVarDsc* varDsc = &compiler->lvaTable[lclNum];
assert(varDsc->lvTracked);
assert(varDsc->lvVarIndex < compiler->lvaTrackedCount);
linScan->localVarIntervals[varDsc->lvVarIndex] = this;
assert(linScan->getIntervalForLocalVar(varDsc->lvVarIndex) == this);
this->isLocalVar = true;
this->varNum = lclNum;
}
// identify the candidates which we are not going to enregister due to
// being used in EH in a way we don't want to deal with
// this logic cloned from fgInterBlockLocalVarLiveness
void LinearScan::identifyCandidatesExceptionDataflow()
{
VARSET_TP exceptVars(VarSetOps::MakeEmpty(compiler));
VARSET_TP filterVars(VarSetOps::MakeEmpty(compiler));
VARSET_TP finallyVars(VarSetOps::MakeEmpty(compiler));
BasicBlock* block;
foreach_block(compiler, block)
{
if (block->bbCatchTyp != BBCT_NONE)
{
// live on entry to handler
VarSetOps::UnionD(compiler, exceptVars, block->bbLiveIn);
}
if (block->bbJumpKind == BBJ_EHFILTERRET)
{
// live on exit from filter
VarSetOps::UnionD(compiler, filterVars, block->bbLiveOut);
}
else if (block->bbJumpKind == BBJ_EHFINALLYRET)
{
// live on exit from finally
VarSetOps::UnionD(compiler, finallyVars, block->bbLiveOut);
}
#if FEATURE_EH_FUNCLETS
// Funclets are called and returned from, as such we can only count on the frame
// pointer being restored, and thus everything live in or live out must be on the
// stack
if (block->bbFlags & BBF_FUNCLET_BEG)
{
VarSetOps::UnionD(compiler, exceptVars, block->bbLiveIn);
}
if ((block->bbJumpKind == BBJ_EHFINALLYRET) || (block->bbJumpKind == BBJ_EHFILTERRET) ||
(block->bbJumpKind == BBJ_EHCATCHRET))
{
VarSetOps::UnionD(compiler, exceptVars, block->bbLiveOut);
}
#endif // FEATURE_EH_FUNCLETS
}
// slam them all together (there was really no need to use more than 2 bitvectors here)
VarSetOps::UnionD(compiler, exceptVars, filterVars);
VarSetOps::UnionD(compiler, exceptVars, finallyVars);
/* Mark all pointer variables live on exit from a 'finally'
block as either volatile for non-GC ref types or as
'explicitly initialized' (volatile and must-init) for GC-ref types */
VarSetOps::Iter iter(compiler, exceptVars);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* varDsc = compiler->lvaTable + varNum;
compiler->lvaSetVarDoNotEnregister(varNum DEBUGARG(Compiler::DNER_LiveInOutOfHandler));
if (varTypeIsGC(varDsc))
{
if (VarSetOps::IsMember(compiler, finallyVars, varIndex) && !varDsc->lvIsParam)
{
varDsc->lvMustInit = true;
}
}
}
}
//------------------------------------------------------------------------
// IsContainableMemoryOp: Checks whether this is a memory op that can be contained.
//
// Arguments:
// node - the node of interest.
//
// Return value:
// True if this will definitely be a memory reference that could be contained.
//
// Notes:
// This differs from the isMemoryOp() method on GenTree because it checks for
// the case of doNotEnregister local. This won't include locals that
// for some other reason do not become register candidates, nor those that get
// spilled.
// Also, because we usually call this before we redo dataflow, any new lclVars
// introduced after the last dataflow analysis will not yet be marked lvTracked,
// so we don't use that.
//
bool LinearScan::isContainableMemoryOp(GenTree* node)
{
#ifdef _TARGET_XARCH_
if (node->isMemoryOp())
{
return true;
}
if (node->IsLocal())
{
if (!enregisterLocalVars)
{
return true;
}
LclVarDsc* varDsc = &compiler->lvaTable[node->AsLclVar()->gtLclNum];
return varDsc->lvDoNotEnregister;
}
#endif // _TARGET_XARCH_
return false;
}
bool LinearScan::isRegCandidate(LclVarDsc* varDsc)
{
// We shouldn't be called if opt settings do not permit register variables.
assert((compiler->opts.compFlags & CLFLG_REGVAR) != 0);
if (!varDsc->lvTracked)
{
return false;
}
#if !defined(_TARGET_64BIT_)
if (varDsc->lvType == TYP_LONG)
{
// Long variables should not be register candidates.
// Lowering will have split any candidate lclVars into lo/hi vars.
return false;
}
#endif // !defined(_TARGET_64BIT)
// If we have JMP, reg args must be put on the stack
if (compiler->compJmpOpUsed && varDsc->lvIsRegArg)
{
return false;
}
// Don't allocate registers for dependently promoted struct fields
if (compiler->lvaIsFieldOfDependentlyPromotedStruct(varDsc))
{
return false;
}
return true;
}
// Identify locals & compiler temps that are register candidates
// TODO-Cleanup: This was cloned from Compiler::lvaSortByRefCount() in lclvars.cpp in order
// to avoid perturbation, but should be merged.
void LinearScan::identifyCandidates()
{
if (enregisterLocalVars)
{
// Initialize the set of lclVars that are candidates for register allocation.
VarSetOps::AssignNoCopy(compiler, registerCandidateVars, VarSetOps::MakeEmpty(compiler));
// Initialize the sets of lclVars that are used to determine whether, and for which lclVars,
// we need to perform resolution across basic blocks.
// Note that we can't do this in the constructor because the number of tracked lclVars may
// change between the constructor and the actual allocation.
VarSetOps::AssignNoCopy(compiler, resolutionCandidateVars, VarSetOps::MakeEmpty(compiler));
VarSetOps::AssignNoCopy(compiler, splitOrSpilledVars, VarSetOps::MakeEmpty(compiler));
// We set enregisterLocalVars to true only if there are tracked lclVars
assert(compiler->lvaCount != 0);
}
else if (compiler->lvaCount == 0)
{
// Nothing to do. Note that even if enregisterLocalVars is false, we still need to set the
// lvLRACandidate field on all the lclVars to false if we have any.
return;
}
if (compiler->compHndBBtabCount > 0)
{
identifyCandidatesExceptionDataflow();
}
unsigned lclNum;
LclVarDsc* varDsc;
// While we build intervals for the candidate lclVars, we will determine the floating point
// lclVars, if any, to consider for callee-save register preferencing.
// We maintain two sets of FP vars - those that meet the first threshold of weighted ref Count,
// and those that meet the second.
// The first threshold is used for methods that are heuristically deemed either to have light
// fp usage, or other factors that encourage conservative use of callee-save registers, such
// as multiple exits (where there might be an early exit that woudl be excessively penalized by
// lots of prolog/epilog saves & restores).
// The second threshold is used where there are factors deemed to make it more likely that fp
// fp callee save registers will be needed, such as loops or many fp vars.
// We keep two sets of vars, since we collect some of the information to determine which set to
// use as we iterate over the vars.
// When we are generating AVX code on non-Unix (FEATURE_PARTIAL_SIMD_CALLEE_SAVE), we maintain an
// additional set of LargeVectorType vars, and there is a separate threshold defined for those.
// It is assumed that if we encounter these, that we should consider this a "high use" scenario,
// so we don't maintain two sets of these vars.
// This is defined as thresholdLargeVectorRefCntWtd, as we are likely to use the same mechanism
// for vectors on Arm64, though the actual value may differ.
unsigned int floatVarCount = 0;
unsigned int thresholdFPRefCntWtd = 4 * BB_UNITY_WEIGHT;
unsigned int maybeFPRefCntWtd = 2 * BB_UNITY_WEIGHT;
VARSET_TP fpMaybeCandidateVars(VarSetOps::UninitVal());
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
unsigned int largeVectorVarCount = 0;
unsigned int thresholdLargeVectorRefCntWtd = 4 * BB_UNITY_WEIGHT;
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (enregisterLocalVars)
{
VarSetOps::AssignNoCopy(compiler, fpCalleeSaveCandidateVars, VarSetOps::MakeEmpty(compiler));
VarSetOps::AssignNoCopy(compiler, fpMaybeCandidateVars, VarSetOps::MakeEmpty(compiler));
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
VarSetOps::AssignNoCopy(compiler, largeVectorVars, VarSetOps::MakeEmpty(compiler));
VarSetOps::AssignNoCopy(compiler, largeVectorCalleeSaveCandidateVars, VarSetOps::MakeEmpty(compiler));
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
}
#if DOUBLE_ALIGN
unsigned refCntStk = 0;
unsigned refCntReg = 0;
unsigned refCntWtdReg = 0;
unsigned refCntStkParam = 0; // sum of ref counts for all stack based parameters
unsigned refCntWtdStkDbl = 0; // sum of wtd ref counts for stack based doubles
doDoubleAlign = false;
bool checkDoubleAlign = true;
if (compiler->codeGen->isFramePointerRequired() || compiler->opts.MinOpts())
{
checkDoubleAlign = false;
}
else
{
switch (compiler->getCanDoubleAlign())
{
case MUST_DOUBLE_ALIGN:
doDoubleAlign = true;
checkDoubleAlign = false;
break;
case CAN_DOUBLE_ALIGN:
break;
case CANT_DOUBLE_ALIGN:
doDoubleAlign = false;
checkDoubleAlign = false;
break;
default:
unreached();
}
}
#endif // DOUBLE_ALIGN
// Check whether register variables are permitted.
if (!enregisterLocalVars)
{
localVarIntervals = nullptr;
}
else if (compiler->lvaTrackedCount > 0)
{
// initialize mapping from tracked local to interval
localVarIntervals = new (compiler, CMK_LSRA) Interval*[compiler->lvaTrackedCount];
}
INTRACK_STATS(regCandidateVarCount = 0);
for (lclNum = 0, varDsc = compiler->lvaTable; lclNum < compiler->lvaCount; lclNum++, varDsc++)
{
// Initialize all variables to REG_STK
varDsc->lvRegNum = REG_STK;
#ifndef _TARGET_64BIT_
varDsc->lvOtherReg = REG_STK;
#endif // _TARGET_64BIT_
if (!enregisterLocalVars)
{
varDsc->lvLRACandidate = false;
continue;
}
#if DOUBLE_ALIGN
if (checkDoubleAlign)
{
if (varDsc->lvIsParam && !varDsc->lvIsRegArg)
{
refCntStkParam += varDsc->lvRefCnt;
}
else if (!isRegCandidate(varDsc) || varDsc->lvDoNotEnregister)
{
refCntStk += varDsc->lvRefCnt;
if ((varDsc->lvType == TYP_DOUBLE) ||
((varTypeIsStruct(varDsc) && varDsc->lvStructDoubleAlign &&
(compiler->lvaGetPromotionType(varDsc) != Compiler::PROMOTION_TYPE_INDEPENDENT))))
{
refCntWtdStkDbl += varDsc->lvRefCntWtd;
}
}
else
{
refCntReg += varDsc->lvRefCnt;
refCntWtdReg += varDsc->lvRefCntWtd;
}
}
#endif // DOUBLE_ALIGN
/* Track all locals that can be enregistered */
if (!isRegCandidate(varDsc))
{
varDsc->lvLRACandidate = 0;
if (varDsc->lvTracked)
{
localVarIntervals[varDsc->lvVarIndex] = nullptr;
}
continue;
}
assert(varDsc->lvTracked);
varDsc->lvLRACandidate = 1;
// Start with lvRegister as false - set it true only if the variable gets
// the same register assignment throughout
varDsc->lvRegister = false;
/* If the ref count is zero */
if (varDsc->lvRefCnt == 0)
{
/* Zero ref count, make this untracked */
varDsc->lvRefCntWtd = 0;
varDsc->lvLRACandidate = 0;
}
// Variables that are address-exposed are never enregistered, or tracked.
// A struct may be promoted, and a struct that fits in a register may be fully enregistered.
// Pinned variables may not be tracked (a condition of the GCInfo representation)
// or enregistered, on x86 -- it is believed that we can enregister pinned (more properly, "pinning")
// references when using the general GC encoding.
if (varDsc->lvAddrExposed || !varTypeIsEnregisterableStruct(varDsc))
{
varDsc->lvLRACandidate = 0;
#ifdef DEBUG
Compiler::DoNotEnregisterReason dner = Compiler::DNER_AddrExposed;
if (!varDsc->lvAddrExposed)
{
dner = Compiler::DNER_IsStruct;
}
#endif // DEBUG
compiler->lvaSetVarDoNotEnregister(lclNum DEBUGARG(dner));
}
else if (varDsc->lvPinned)
{
varDsc->lvTracked = 0;
#ifdef JIT32_GCENCODER
compiler->lvaSetVarDoNotEnregister(lclNum DEBUGARG(Compiler::DNER_PinningRef));
#endif // JIT32_GCENCODER
}
// Are we not optimizing and we have exception handlers?
// if so mark all args and locals as volatile, so that they
// won't ever get enregistered.
//
if (compiler->opts.MinOpts() && compiler->compHndBBtabCount > 0)
{
compiler->lvaSetVarDoNotEnregister(lclNum DEBUGARG(Compiler::DNER_LiveInOutOfHandler));
}
if (varDsc->lvDoNotEnregister)
{
varDsc->lvLRACandidate = 0;
localVarIntervals[varDsc->lvVarIndex] = nullptr;
continue;
}
var_types type = genActualType(varDsc->TypeGet());
switch (type)
{
#if CPU_HAS_FP_SUPPORT
case TYP_FLOAT:
case TYP_DOUBLE:
if (compiler->opts.compDbgCode)
{
varDsc->lvLRACandidate = 0;
}
#ifdef ARM_SOFTFP
if (varDsc->lvIsParam && varDsc->lvIsRegArg)
{
type = (type == TYP_DOUBLE) ? TYP_LONG : TYP_INT;
}
#endif // ARM_SOFTFP
break;
#endif // CPU_HAS_FP_SUPPORT
case TYP_INT:
case TYP_LONG:
case TYP_REF:
case TYP_BYREF:
break;
#ifdef FEATURE_SIMD
case TYP_SIMD12:
case TYP_SIMD16:
case TYP_SIMD32:
if (varDsc->lvPromoted)
{
varDsc->lvLRACandidate = 0;
}
break;
// TODO-1stClassStructs: Move TYP_SIMD8 up with the other SIMD types, after handling the param issue
// (passing & returning as TYP_LONG).
case TYP_SIMD8:
#endif // FEATURE_SIMD
case TYP_STRUCT:
{
varDsc->lvLRACandidate = 0;
}
break;
case TYP_UNDEF:
case TYP_UNKNOWN:
noway_assert(!"lvType not set correctly");
varDsc->lvType = TYP_INT;
__fallthrough;
default:
varDsc->lvLRACandidate = 0;
}
if (varDsc->lvLRACandidate)
{
Interval* newInt = newInterval(type);
newInt->setLocalNumber(compiler, lclNum, this);
VarSetOps::AddElemD(compiler, registerCandidateVars, varDsc->lvVarIndex);
// we will set this later when we have determined liveness
varDsc->lvMustInit = false;
if (varDsc->lvIsStructField)
{
newInt->isStructField = true;
}
INTRACK_STATS(regCandidateVarCount++);
// We maintain two sets of FP vars - those that meet the first threshold of weighted ref Count,
// and those that meet the second (see the definitions of thresholdFPRefCntWtd and maybeFPRefCntWtd
// above).
CLANG_FORMAT_COMMENT_ANCHOR;
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
// Additionally, when we are generating AVX on non-UNIX amd64, we keep a separate set of the LargeVectorType
// vars.
if (varDsc->lvType == LargeVectorType)
{
largeVectorVarCount++;
VarSetOps::AddElemD(compiler, largeVectorVars, varDsc->lvVarIndex);
unsigned refCntWtd = varDsc->lvRefCntWtd;
if (refCntWtd >= thresholdLargeVectorRefCntWtd)
{
VarSetOps::AddElemD(compiler, largeVectorCalleeSaveCandidateVars, varDsc->lvVarIndex);
}
}
else
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (regType(type) == FloatRegisterType)
{
floatVarCount++;
unsigned refCntWtd = varDsc->lvRefCntWtd;
if (varDsc->lvIsRegArg)
{
// Don't count the initial reference for register params. In those cases,
// using a callee-save causes an extra copy.
refCntWtd -= BB_UNITY_WEIGHT;
}
if (refCntWtd >= thresholdFPRefCntWtd)
{
VarSetOps::AddElemD(compiler, fpCalleeSaveCandidateVars, varDsc->lvVarIndex);
}
else if (refCntWtd >= maybeFPRefCntWtd)
{
VarSetOps::AddElemD(compiler, fpMaybeCandidateVars, varDsc->lvVarIndex);
}
}
}
else
{
localVarIntervals[varDsc->lvVarIndex] = nullptr;
}
}
#if DOUBLE_ALIGN
if (checkDoubleAlign)
{
// TODO-CQ: Fine-tune this:
// In the legacy reg predictor, this runs after allocation, and then demotes any lclVars
// allocated to the frame pointer, which is probably the wrong order.
// However, because it runs after allocation, it can determine the impact of demoting
// the lclVars allocated to the frame pointer.
// => Here, estimate of the EBP refCnt and weighted refCnt is a wild guess.
//
unsigned refCntEBP = refCntReg / 8;
unsigned refCntWtdEBP = refCntWtdReg / 8;
doDoubleAlign =
compiler->shouldDoubleAlign(refCntStk, refCntEBP, refCntWtdEBP, refCntStkParam, refCntWtdStkDbl);
}
#endif // DOUBLE_ALIGN
// The factors we consider to determine which set of fp vars to use as candidates for callee save
// registers current include the number of fp vars, whether there are loops, and whether there are
// multiple exits. These have been selected somewhat empirically, but there is probably room for
// more tuning.
CLANG_FORMAT_COMMENT_ANCHOR;
#ifdef DEBUG
if (VERBOSE)
{
printf("\nFP callee save candidate vars: ");
if (enregisterLocalVars && !VarSetOps::IsEmpty(compiler, fpCalleeSaveCandidateVars))
{
dumpConvertedVarSet(compiler, fpCalleeSaveCandidateVars);
printf("\n");
}
else
{
printf("None\n\n");
}
}
#endif
JITDUMP("floatVarCount = %d; hasLoops = %d, singleExit = %d\n", floatVarCount, compiler->fgHasLoops,
(compiler->fgReturnBlocks == nullptr || compiler->fgReturnBlocks->next == nullptr));
// Determine whether to use the 2nd, more aggressive, threshold for fp callee saves.
if (floatVarCount > 6 && compiler->fgHasLoops &&
(compiler->fgReturnBlocks == nullptr || compiler->fgReturnBlocks->next == nullptr))
{
assert(enregisterLocalVars);
#ifdef DEBUG
if (VERBOSE)
{
printf("Adding additional fp callee save candidates: \n");
if (!VarSetOps::IsEmpty(compiler, fpMaybeCandidateVars))
{
dumpConvertedVarSet(compiler, fpMaybeCandidateVars);
printf("\n");
}
else
{
printf("None\n\n");
}
}
#endif
VarSetOps::UnionD(compiler, fpCalleeSaveCandidateVars, fpMaybeCandidateVars);
}
#ifdef _TARGET_ARM_
#ifdef DEBUG
if (VERBOSE)
{
// Frame layout is only pre-computed for ARM
printf("\nlvaTable after IdentifyCandidates\n");
compiler->lvaTableDump();
}
#endif // DEBUG
#endif // _TARGET_ARM_
}
// TODO-Throughput: This mapping can surely be more efficiently done
void LinearScan::initVarRegMaps()
{
if (!enregisterLocalVars)
{
inVarToRegMaps = nullptr;
outVarToRegMaps = nullptr;
return;
}
assert(compiler->lvaTrackedFixed); // We should have already set this to prevent us from adding any new tracked
// variables.
// The compiler memory allocator requires that the allocation be an
// even multiple of int-sized objects
unsigned int varCount = compiler->lvaTrackedCount;
regMapCount = (unsigned int)roundUp(varCount, sizeof(int));
// Not sure why blocks aren't numbered from zero, but they don't appear to be.
// So, if we want to index by bbNum we have to know the maximum value.
unsigned int bbCount = compiler->fgBBNumMax + 1;
inVarToRegMaps = new (compiler, CMK_LSRA) regNumberSmall*[bbCount];
outVarToRegMaps = new (compiler, CMK_LSRA) regNumberSmall*[bbCount];
if (varCount > 0)
{
// This VarToRegMap is used during the resolution of critical edges.
sharedCriticalVarToRegMap = new (compiler, CMK_LSRA) regNumberSmall[regMapCount];
for (unsigned int i = 0; i < bbCount; i++)
{
VarToRegMap inVarToRegMap = new (compiler, CMK_LSRA) regNumberSmall[regMapCount];
VarToRegMap outVarToRegMap = new (compiler, CMK_LSRA) regNumberSmall[regMapCount];
for (unsigned int j = 0; j < regMapCount; j++)
{
inVarToRegMap[j] = REG_STK;
outVarToRegMap[j] = REG_STK;
}
inVarToRegMaps[i] = inVarToRegMap;
outVarToRegMaps[i] = outVarToRegMap;
}
}
else
{
sharedCriticalVarToRegMap = nullptr;
for (unsigned int i = 0; i < bbCount; i++)
{
inVarToRegMaps[i] = nullptr;
outVarToRegMaps[i] = nullptr;
}
}
}
void LinearScan::setInVarRegForBB(unsigned int bbNum, unsigned int varNum, regNumber reg)
{
assert(enregisterLocalVars);
assert(reg < UCHAR_MAX && varNum < compiler->lvaCount);
inVarToRegMaps[bbNum][compiler->lvaTable[varNum].lvVarIndex] = (regNumberSmall)reg;
}
void LinearScan::setOutVarRegForBB(unsigned int bbNum, unsigned int varNum, regNumber reg)
{
assert(enregisterLocalVars);
assert(reg < UCHAR_MAX && varNum < compiler->lvaCount);
outVarToRegMaps[bbNum][compiler->lvaTable[varNum].lvVarIndex] = (regNumberSmall)reg;
}
LinearScan::SplitEdgeInfo LinearScan::getSplitEdgeInfo(unsigned int bbNum)
{
assert(enregisterLocalVars);
SplitEdgeInfo splitEdgeInfo;
assert(bbNum <= compiler->fgBBNumMax);
assert(bbNum > bbNumMaxBeforeResolution);
assert(splitBBNumToTargetBBNumMap != nullptr);
splitBBNumToTargetBBNumMap->Lookup(bbNum, &splitEdgeInfo);
assert(splitEdgeInfo.toBBNum <= bbNumMaxBeforeResolution);
assert(splitEdgeInfo.fromBBNum <= bbNumMaxBeforeResolution);
return splitEdgeInfo;
}
VarToRegMap LinearScan::getInVarToRegMap(unsigned int bbNum)
{
assert(enregisterLocalVars);
assert(bbNum <= compiler->fgBBNumMax);
// For the blocks inserted to split critical edges, the inVarToRegMap is
// equal to the outVarToRegMap at the "from" block.
if (bbNum > bbNumMaxBeforeResolution)
{
SplitEdgeInfo splitEdgeInfo = getSplitEdgeInfo(bbNum);
unsigned fromBBNum = splitEdgeInfo.fromBBNum;
if (fromBBNum == 0)
{
assert(splitEdgeInfo.toBBNum != 0);
return inVarToRegMaps[splitEdgeInfo.toBBNum];
}
else
{
return outVarToRegMaps[fromBBNum];
}
}
return inVarToRegMaps[bbNum];
}
VarToRegMap LinearScan::getOutVarToRegMap(unsigned int bbNum)
{
assert(enregisterLocalVars);
assert(bbNum <= compiler->fgBBNumMax);
// For the blocks inserted to split critical edges, the outVarToRegMap is
// equal to the inVarToRegMap at the target.
if (bbNum > bbNumMaxBeforeResolution)
{
// If this is an empty block, its in and out maps are both the same.
// We identify this case by setting fromBBNum or toBBNum to 0, and using only the other.
SplitEdgeInfo splitEdgeInfo = getSplitEdgeInfo(bbNum);
unsigned toBBNum = splitEdgeInfo.toBBNum;
if (toBBNum == 0)
{
assert(splitEdgeInfo.fromBBNum != 0);
return outVarToRegMaps[splitEdgeInfo.fromBBNum];
}
else
{
return inVarToRegMaps[toBBNum];
}
}
return outVarToRegMaps[bbNum];
}
//------------------------------------------------------------------------
// setVarReg: Set the register associated with a variable in the given 'bbVarToRegMap'.
//
// Arguments:
// bbVarToRegMap - the map of interest
// trackedVarIndex - the lvVarIndex for the variable
// reg - the register to which it is being mapped
//
// Return Value:
// None
//
void LinearScan::setVarReg(VarToRegMap bbVarToRegMap, unsigned int trackedVarIndex, regNumber reg)
{
assert(trackedVarIndex < compiler->lvaTrackedCount);
regNumberSmall regSmall = (regNumberSmall)reg;
assert((regNumber)regSmall == reg);
bbVarToRegMap[trackedVarIndex] = regSmall;
}
//------------------------------------------------------------------------
// getVarReg: Get the register associated with a variable in the given 'bbVarToRegMap'.
//
// Arguments:
// bbVarToRegMap - the map of interest
// trackedVarIndex - the lvVarIndex for the variable
//
// Return Value:
// The register to which 'trackedVarIndex' is mapped
//
regNumber LinearScan::getVarReg(VarToRegMap bbVarToRegMap, unsigned int trackedVarIndex)
{
assert(enregisterLocalVars);
assert(trackedVarIndex < compiler->lvaTrackedCount);
return (regNumber)bbVarToRegMap[trackedVarIndex];
}
// Initialize the incoming VarToRegMap to the given map values (generally a predecessor of
// the block)
VarToRegMap LinearScan::setInVarToRegMap(unsigned int bbNum, VarToRegMap srcVarToRegMap)
{
assert(enregisterLocalVars);
VarToRegMap inVarToRegMap = inVarToRegMaps[bbNum];
memcpy(inVarToRegMap, srcVarToRegMap, (regMapCount * sizeof(regNumber)));
return inVarToRegMap;
}
// given a tree node
RefType refTypeForLocalRefNode(GenTree* node)
{
assert(node->IsLocal());
// We don't support updates
assert((node->gtFlags & GTF_VAR_USEASG) == 0);
if (node->gtFlags & GTF_VAR_DEF)
{
return RefTypeDef;
}
else
{
return RefTypeUse;
}
}
//------------------------------------------------------------------------
// checkLastUses: Check correctness of last use flags
//
// Arguments:
// The block for which we are checking last uses.
//
// Notes:
// This does a backward walk of the RefPositions, starting from the liveOut set.
// This method was previously used to set the last uses, which were computed by
// liveness, but were not create in some cases of multiple lclVar references in the
// same tree. However, now that last uses are computed as RefPositions are created,
// that is no longer necessary, and this method is simply retained as a check.
// The exception to the check-only behavior is when LSRA_EXTEND_LIFETIMES if set via
// COMPlus_JitStressRegs. In that case, this method is required, because even though
// the RefPositions will not be marked lastUse in that case, we still need to correclty
// mark the last uses on the tree nodes, which is done by this method.
//
#ifdef DEBUG
void LinearScan::checkLastUses(BasicBlock* block)
{
if (VERBOSE)
{
JITDUMP("\n\nCHECKING LAST USES for block %u, liveout=", block->bbNum);
dumpConvertedVarSet(compiler, block->bbLiveOut);
JITDUMP("\n==============================\n");
}
unsigned keepAliveVarNum = BAD_VAR_NUM;
if (compiler->lvaKeepAliveAndReportThis())
{
keepAliveVarNum = compiler->info.compThisArg;
assert(compiler->info.compIsStatic == false);
}
// find which uses are lastUses
// Work backwards starting with live out.
// 'computedLive' is updated to include any exposed use (including those in this
// block that we've already seen). When we encounter a use, if it's
// not in that set, then it's a last use.
VARSET_TP computedLive(VarSetOps::MakeCopy(compiler, block->bbLiveOut));
bool foundDiff = false;
auto currentRefPosition = refPositions.rbegin();
while (currentRefPosition->refType != RefTypeBB)
{
// We should never see ParamDefs or ZeroInits within a basic block.
assert(currentRefPosition->refType != RefTypeParamDef && currentRefPosition->refType != RefTypeZeroInit);
if (currentRefPosition->isIntervalRef() && currentRefPosition->getInterval()->isLocalVar)
{
unsigned varNum = currentRefPosition->getInterval()->varNum;
unsigned varIndex = currentRefPosition->getInterval()->getVarIndex(compiler);
LsraLocation loc = currentRefPosition->nodeLocation;
// We should always have a tree node for a localVar, except for the "special" RefPositions.
GenTreePtr tree = currentRefPosition->treeNode;
assert(tree != nullptr || currentRefPosition->refType == RefTypeExpUse ||
currentRefPosition->refType == RefTypeDummyDef);
if (!VarSetOps::IsMember(compiler, computedLive, varIndex) && varNum != keepAliveVarNum)
{
// There was no exposed use, so this is a "last use" (and we mark it thus even if it's a def)
if (extendLifetimes())
{
// NOTE: this is a bit of a hack. When extending lifetimes, the "last use" bit will be clear.
// This bit, however, would normally be used during resolveLocalRef to set the value of
// GTF_VAR_DEATH on the node for a ref position. If this bit is not set correctly even when
// extending lifetimes, the code generator will assert as it expects to have accurate last
// use information. To avoid these asserts, set the GTF_VAR_DEATH bit here.
// Note also that extendLifetimes() is an LSRA stress mode, so it will only be true for
// Checked or Debug builds, for which this method will be executed.
if (tree != nullptr)
{
tree->gtFlags |= GTF_VAR_DEATH;
}
}
else if (!currentRefPosition->lastUse)
{
JITDUMP("missing expected last use of V%02u @%u\n", compiler->lvaTrackedToVarNum[varIndex], loc);
foundDiff = true;
}
VarSetOps::AddElemD(compiler, computedLive, varIndex);
}
else if (currentRefPosition->lastUse)
{
JITDUMP("unexpected last use of V%02u @%u\n", compiler->lvaTrackedToVarNum[varIndex], loc);
foundDiff = true;
}
else if (extendLifetimes() && tree != nullptr)
{
// NOTE: see the comment above re: the extendLifetimes hack.
tree->gtFlags &= ~GTF_VAR_DEATH;
}
if (currentRefPosition->refType == RefTypeDef || currentRefPosition->refType == RefTypeDummyDef)
{
VarSetOps::RemoveElemD(compiler, computedLive, varIndex);
}
}
assert(currentRefPosition != refPositions.rend());
++currentRefPosition;
}
VARSET_TP liveInNotComputedLive(VarSetOps::Diff(compiler, block->bbLiveIn, computedLive));
VarSetOps::Iter liveInNotComputedLiveIter(compiler, liveInNotComputedLive);
unsigned liveInNotComputedLiveIndex = 0;
while (liveInNotComputedLiveIter.NextElem(&liveInNotComputedLiveIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[liveInNotComputedLiveIndex];
if (compiler->lvaTable[varNum].lvLRACandidate)
{
JITDUMP("BB%02u: V%02u is in LiveIn set, but not computed live.\n", block->bbNum, varNum);
foundDiff = true;
}
}
VarSetOps::DiffD(compiler, computedLive, block->bbLiveIn);
const VARSET_TP& computedLiveNotLiveIn(computedLive); // reuse the buffer.
VarSetOps::Iter computedLiveNotLiveInIter(compiler, computedLiveNotLiveIn);
unsigned computedLiveNotLiveInIndex = 0;
while (computedLiveNotLiveInIter.NextElem(&computedLiveNotLiveInIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[computedLiveNotLiveInIndex];
if (compiler->lvaTable[varNum].lvLRACandidate)
{
JITDUMP("BB%02u: V%02u is computed live, but not in LiveIn set.\n", block->bbNum, varNum);
foundDiff = true;
}
}
assert(!foundDiff);
}
#endif // DEBUG
void LinearScan::addRefsForPhysRegMask(regMaskTP mask, LsraLocation currentLoc, RefType refType, bool isLastUse)
{
for (regNumber reg = REG_FIRST; mask; reg = REG_NEXT(reg), mask >>= 1)
{
if (mask & 1)
{
// This assumes that these are all "special" RefTypes that
// don't need to be recorded on the tree (hence treeNode is nullptr)
RefPosition* pos = newRefPosition(reg, currentLoc, refType, nullptr,
genRegMask(reg)); // This MUST occupy the physical register (obviously)
if (isLastUse)
{
pos->lastUse = true;
}
}
}
}
//------------------------------------------------------------------------
// getKillSetForNode: Return the registers killed by the given tree node.
//
// Arguments:
// compiler - the compiler context to use
// tree - the tree for which the kill set is needed.
//
// Return Value: a register mask of the registers killed
//
regMaskTP LinearScan::getKillSetForNode(GenTree* tree)
{
regMaskTP killMask = RBM_NONE;
switch (tree->OperGet())
{
#ifdef _TARGET_XARCH_
case GT_MUL:
// We use the 128-bit multiply when performing an overflow checking unsigned multiply
//
if (((tree->gtFlags & GTF_UNSIGNED) != 0) && tree->gtOverflowEx())
{
// Both RAX and RDX are killed by the operation
killMask = RBM_RAX | RBM_RDX;
}
break;
case GT_MULHI:
#if defined(_TARGET_X86_) && !defined(LEGACY_BACKEND)
case GT_MUL_LONG:
#endif
killMask = RBM_RAX | RBM_RDX;
break;
case GT_MOD:
case GT_DIV:
case GT_UMOD:
case GT_UDIV:
if (!varTypeIsFloating(tree->TypeGet()))
{
// RDX needs to be killed early, because it must not be used as a source register
// (unlike most cases, where the kill happens AFTER the uses). So for this kill,
// we add the RefPosition at the tree loc (where the uses are located) instead of the
// usual kill location which is the same as the defs at tree loc+1.
// Note that we don't have to add interference for the live vars, because that
// will be done below, and is not sensitive to the precise location.
LsraLocation currentLoc = tree->gtLsraInfo.loc;
assert(currentLoc != 0);
addRefsForPhysRegMask(RBM_RDX, currentLoc, RefTypeKill, true);
// Both RAX and RDX are killed by the operation
killMask = RBM_RAX | RBM_RDX;
}
break;
#endif // _TARGET_XARCH_
case GT_STORE_OBJ:
if (tree->OperIsCopyBlkOp())
{
assert(tree->AsObj()->gtGcPtrCount != 0);
killMask = compiler->compHelperCallKillSet(CORINFO_HELP_ASSIGN_BYREF);
break;
}
__fallthrough;
case GT_STORE_BLK:
case GT_STORE_DYN_BLK:
{
GenTreeBlk* blkNode = tree->AsBlk();
bool isCopyBlk = varTypeIsStruct(blkNode->Data());
switch (blkNode->gtBlkOpKind)
{
case GenTreeBlk::BlkOpKindHelper:
if (isCopyBlk)
{
killMask = compiler->compHelperCallKillSet(CORINFO_HELP_MEMCPY);
}
else
{
killMask = compiler->compHelperCallKillSet(CORINFO_HELP_MEMSET);
}
break;
#ifdef _TARGET_XARCH_
case GenTreeBlk::BlkOpKindRepInstr:
if (isCopyBlk)
{
// rep movs kills RCX, RDI and RSI
killMask = RBM_RCX | RBM_RDI | RBM_RSI;
}
else
{
// rep stos kills RCX and RDI.
// (Note that the Data() node, if not constant, will be assigned to
// RCX, but it's find that this kills it, as the value is not available
// after this node in any case.)
killMask = RBM_RDI | RBM_RCX;
}
break;
#else
case GenTreeBlk::BlkOpKindRepInstr:
#endif
case GenTreeBlk::BlkOpKindUnroll:
case GenTreeBlk::BlkOpKindInvalid:
// for these 'gtBlkOpKind' kinds, we leave 'killMask' = RBM_NONE
break;
}
}
break;
case GT_RETURNTRAP:
killMask = compiler->compHelperCallKillSet(CORINFO_HELP_STOP_FOR_GC);
break;
case GT_CALL:
#ifdef _TARGET_X86_
if (compiler->compFloatingPointUsed)
{
if (tree->TypeGet() == TYP_DOUBLE)
{
needDoubleTmpForFPCall = true;
}
else if (tree->TypeGet() == TYP_FLOAT)
{
needFloatTmpForFPCall = true;
}
}
#endif // _TARGET_X86_
#if defined(_TARGET_X86_) || defined(_TARGET_ARM_)
if (tree->IsHelperCall())
{
GenTreeCall* call = tree->AsCall();
CorInfoHelpFunc helpFunc = compiler->eeGetHelperNum(call->gtCallMethHnd);
killMask = compiler->compHelperCallKillSet(helpFunc);
}
else
#endif // defined(_TARGET_X86_) || defined(_TARGET_ARM_)
{
// if there is no FP used, we can ignore the FP kills
if (compiler->compFloatingPointUsed)
{
killMask = RBM_CALLEE_TRASH;
}
else
{
killMask = RBM_INT_CALLEE_TRASH;
}
#ifdef _TARGET_ARM_
if (tree->AsCall()->IsVirtualStub())
{
killMask |= compiler->virtualStubParamInfo->GetRegMask();
}
#else // !_TARGET_ARM_
// Verify that the special virtual stub call registers are in the kill mask.
// We don't just add them unconditionally to the killMask because for most architectures
// they are already in the RBM_CALLEE_TRASH set,
// and we don't want to introduce extra checks and calls in this hot function.
assert(!tree->AsCall()->IsVirtualStub() || ((killMask & compiler->virtualStubParamInfo->GetRegMask()) ==
compiler->virtualStubParamInfo->GetRegMask()));
#endif
}
break;
case GT_STOREIND:
if (compiler->codeGen->gcInfo.gcIsWriteBarrierAsgNode(tree))
{
killMask = RBM_CALLEE_TRASH_NOGC;
}
break;
#if defined(PROFILING_SUPPORTED)
// If this method requires profiler ELT hook then mark these nodes as killing
// callee trash registers (excluding RAX and XMM0). The reason for this is that
// profiler callback would trash these registers. See vm\amd64\asmhelpers.asm for
// more details.
case GT_RETURN:
if (compiler->compIsProfilerHookNeeded())
{
killMask = compiler->compHelperCallKillSet(CORINFO_HELP_PROF_FCN_LEAVE);
}
break;
case GT_PROF_HOOK:
if (compiler->compIsProfilerHookNeeded())
{
killMask = compiler->compHelperCallKillSet(CORINFO_HELP_PROF_FCN_TAILCALL);
}
break;
#endif // PROFILING_SUPPORTED
default:
// for all other 'tree->OperGet()' kinds, leave 'killMask' = RBM_NONE
break;
}
return killMask;
}
//------------------------------------------------------------------------
// buildKillPositionsForNode:
// Given some tree node add refpositions for all the registers this node kills
//
// Arguments:
// tree - the tree for which kill positions should be generated
// currentLoc - the location at which the kills should be added
//
// Return Value:
// true - kills were inserted
// false - no kills were inserted
//
// Notes:
// The return value is needed because if we have any kills, we need to make sure that
// all defs are located AFTER the kills. On the other hand, if there aren't kills,
// the multiple defs for a regPair are in different locations.
// If we generate any kills, we will mark all currentLiveVars as being preferenced
// to avoid the killed registers. This is somewhat conservative.
bool LinearScan::buildKillPositionsForNode(GenTree* tree, LsraLocation currentLoc)
{
regMaskTP killMask = getKillSetForNode(tree);
bool isCallKill = ((killMask == RBM_INT_CALLEE_TRASH) || (killMask == RBM_CALLEE_TRASH));
if (killMask != RBM_NONE)
{
// The killMask identifies a set of registers that will be used during codegen.
// Mark these as modified here, so when we do final frame layout, we'll know about
// all these registers. This is especially important if killMask contains
// callee-saved registers, which affect the frame size since we need to save/restore them.
// In the case where we have a copyBlk with GC pointers, can need to call the
// CORINFO_HELP_ASSIGN_BYREF helper, which kills callee-saved RSI and RDI, if
// LSRA doesn't assign RSI/RDI, they wouldn't get marked as modified until codegen,
// which is too late.
compiler->codeGen->regSet.rsSetRegsModified(killMask DEBUGARG(dumpTerse));
addRefsForPhysRegMask(killMask, currentLoc, RefTypeKill, true);
// TODO-CQ: It appears to be valuable for both fp and int registers to avoid killing the callee
// save regs on infrequently exectued paths. However, it results in a large number of asmDiffs,
// many of which appear to be regressions (because there is more spill on the infrequently path),
// but are not really because the frequent path becomes smaller. Validating these diffs will need
// to be done before making this change.
// if (!blockSequence[curBBSeqNum]->isRunRarely())
if (enregisterLocalVars)
{
VarSetOps::Iter iter(compiler, currentLiveVars);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* varDsc = compiler->lvaTable + varNum;
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (varDsc->lvType == LargeVectorType)
{
if (!VarSetOps::IsMember(compiler, largeVectorCalleeSaveCandidateVars, varIndex))
{
continue;
}
}
else
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (varTypeIsFloating(varDsc) &&
!VarSetOps::IsMember(compiler, fpCalleeSaveCandidateVars, varIndex))
{
continue;
}
Interval* interval = getIntervalForLocalVar(varIndex);
if (isCallKill)
{
interval->preferCalleeSave = true;
}
regMaskTP newPreferences = allRegs(interval->registerType) & (~killMask);
if (newPreferences != RBM_NONE)
{
interval->updateRegisterPreferences(newPreferences);
}
else
{
// If there are no callee-saved registers, the call could kill all the registers.
// This is a valid state, so in that case assert should not trigger. The RA will spill in order to
// free a register later.
assert(compiler->opts.compDbgEnC || (calleeSaveRegs(varDsc->lvType)) == RBM_NONE);
}
}
}
if (killGCRefs(tree))
{
RefPosition* pos = newRefPosition((Interval*)nullptr, currentLoc, RefTypeKillGCRefs, tree,
(allRegs(TYP_REF) & ~RBM_ARG_REGS));
}
return true;
}
return false;
}
//------------------------------------------------------------------------
// killGCRefs:
// Given some tree node return does it need all GC refs to be spilled from
// callee save registers.
//
// Arguments:
// tree - the tree for which we ask about gc refs.
//
// Return Value:
// true - tree kills GC refs on callee save registers
// false - tree doesn't affect GC refs on callee save registers
bool LinearScan::killGCRefs(GenTree* tree)
{
if (tree->IsCall())
{
if ((tree->gtFlags & GTF_CALL_UNMANAGED) != 0)
{
return true;
}
if (tree->AsCall()->gtCallMethHnd == compiler->eeFindHelper(CORINFO_HELP_JIT_PINVOKE_BEGIN))
{
assert(compiler->opts.ShouldUsePInvokeHelpers());
return true;
}
}
return false;
}
//----------------------------------------------------------------------------
// defineNewInternalTemp: Defines a ref position for an internal temp.
//
// Arguments:
// tree - Gentree node requiring an internal register
// regType - Register type
// currentLoc - Location of the temp Def position
// regMask - register mask of candidates for temp
// minRegCandidateCount - Minimum registers to be ensured in candidate
// set under LSRA stress mode. This is a
// DEBUG only arg.
RefPosition* LinearScan::defineNewInternalTemp(GenTree* tree,
RegisterType regType,
LsraLocation currentLoc,
regMaskTP regMask DEBUGARG(unsigned minRegCandidateCount))
{
Interval* current = newInterval(regType);
current->isInternal = true;
return newRefPosition(current, currentLoc, RefTypeDef, tree, regMask, 0 DEBUG_ARG(minRegCandidateCount));
}
//------------------------------------------------------------------------
// buildInternalRegisterDefsForNode - build Def positions for internal
// registers required for tree node.
//
// Arguments:
// tree - Gentree node that needs internal registers
// currentLoc - Location at which Def positions need to be defined
// temps - in-out array which is populated with ref positions
// created for Def of internal registers
// minRegCandidateCount - Minimum registers to be ensured in candidate
// set of ref positions under LSRA stress. This is
// a DEBUG only arg.
//
// Returns:
// The total number of Def positions created for internal registers of tree node.
int LinearScan::buildInternalRegisterDefsForNode(GenTree* tree,
LsraLocation currentLoc,
RefPosition* temps[] // populates
DEBUGARG(unsigned minRegCandidateCount))
{
int count;
int internalIntCount = tree->gtLsraInfo.internalIntCount;
regMaskTP internalCands = tree->gtLsraInfo.getInternalCandidates(this);
// If the number of internal integer registers required is the same as the number of candidate integer registers in
// the candidate set, then they must be handled as fixed registers.
// (E.g. for the integer registers that floating point arguments must be copied into for a varargs call.)
bool fixedRegs = false;
regMaskTP internalIntCandidates = (internalCands & allRegs(TYP_INT));
if (((int)genCountBits(internalIntCandidates)) == internalIntCount)
{
fixedRegs = true;
}
for (count = 0; count < internalIntCount; count++)
{
regMaskTP internalIntCands = (internalCands & allRegs(TYP_INT));
if (fixedRegs)
{
internalIntCands = genFindLowestBit(internalIntCands);
internalCands &= ~internalIntCands;
}
temps[count] =
defineNewInternalTemp(tree, IntRegisterType, currentLoc, internalIntCands DEBUG_ARG(minRegCandidateCount));
}
int internalFloatCount = tree->gtLsraInfo.internalFloatCount;
for (int i = 0; i < internalFloatCount; i++)
{
regMaskTP internalFPCands = (internalCands & internalFloatRegCandidates());
temps[count++] =
defineNewInternalTemp(tree, FloatRegisterType, currentLoc, internalFPCands DEBUG_ARG(minRegCandidateCount));
}
assert(count < MaxInternalRegisters);
assert(count == (internalIntCount + internalFloatCount));
return count;
}
//------------------------------------------------------------------------
// buildInternalRegisterUsesForNode - adds Use positions for internal
// registers required for tree node.
//
// Arguments:
// tree - Gentree node that needs internal registers
// currentLoc - Location at which Use positions need to be defined
// defs - int array containing Def positions of internal
// registers.
// total - Total number of Def positions in 'defs' array.
// minRegCandidateCount - Minimum registers to be ensured in candidate
// set of ref positions under LSRA stress. This is
// a DEBUG only arg.
//
// Returns:
// Void.
void LinearScan::buildInternalRegisterUsesForNode(GenTree* tree,
LsraLocation currentLoc,
RefPosition* defs[],
int total DEBUGARG(unsigned minRegCandidateCount))
{
assert(total < MaxInternalRegisters);
// defs[] has been populated by buildInternalRegisterDefsForNode
// now just add uses to the defs previously added.
for (int i = 0; i < total; i++)
{
RefPosition* prevRefPosition = defs[i];
assert(prevRefPosition != nullptr);
regMaskTP mask = prevRefPosition->registerAssignment;
if (prevRefPosition->isPhysRegRef)
{
newRefPosition(defs[i]->getReg()->regNum, currentLoc, RefTypeUse, tree, mask);
}
else
{
RefPosition* newest = newRefPosition(defs[i]->getInterval(), currentLoc, RefTypeUse, tree, mask,
0 DEBUG_ARG(minRegCandidateCount));
if (tree->gtLsraInfo.isInternalRegDelayFree)
{
newest->delayRegFree = true;
}
}
}
}
regMaskTP LinearScan::getUseCandidates(GenTree* useNode)
{
TreeNodeInfo info = useNode->gtLsraInfo;
return info.getSrcCandidates(this);
}
regMaskTP LinearScan::getDefCandidates(GenTree* tree)
{
TreeNodeInfo info = tree->gtLsraInfo;
return info.getDstCandidates(this);
}
RegisterType LinearScan::getDefType(GenTree* tree)
{
return tree->TypeGet();
}
//------------------------------------------------------------------------
// LocationInfoListNode: used to store a single `LocationInfo` value for a
// node during `buildIntervals`.
//
// This is the node type for `LocationInfoList` below.
//
class LocationInfoListNode final : public LocationInfo
{
friend class LocationInfoList;
friend class LocationInfoListNodePool;
LocationInfoListNode* m_next; // The next node in the list
public:
LocationInfoListNode(LsraLocation l, Interval* i, GenTree* t, unsigned regIdx = 0) : LocationInfo(l, i, t, regIdx)
{
}
//------------------------------------------------------------------------
// LocationInfoListNode::Next: Returns the next node in the list.
LocationInfoListNode* Next() const
{
return m_next;
}
};
//------------------------------------------------------------------------
// LocationInfoList: used to store a list of `LocationInfo` values for a
// node during `buildIntervals`.
//
// Given an IR node that either directly defines N registers or that is a
// contained node with uses that define a total of N registers, that node
// will map to N `LocationInfo` values. These values are stored as a
// linked list of `LocationInfoListNode` values.
//
class LocationInfoList final
{
friend class LocationInfoListNodePool;
LocationInfoListNode* m_head; // The head of the list
LocationInfoListNode* m_tail; // The tail of the list
public:
LocationInfoList() : m_head(nullptr), m_tail(nullptr)
{
}
LocationInfoList(LocationInfoListNode* node) : m_head(node), m_tail(node)
{
assert(m_head->m_next == nullptr);
}
//------------------------------------------------------------------------
// LocationInfoList::IsEmpty: Returns true if the list is empty.
//
bool IsEmpty() const
{
return m_head == nullptr;
}
//------------------------------------------------------------------------
// LocationInfoList::Begin: Returns the first node in the list.
//
LocationInfoListNode* Begin() const
{
return m_head;
}
//------------------------------------------------------------------------
// LocationInfoList::End: Returns the position after the last node in the
// list. The returned value is suitable for use as
// a sentinel for iteration.
//
LocationInfoListNode* End() const
{
return nullptr;
}
//------------------------------------------------------------------------
// LocationInfoList::Append: Appends a node to the list.
//
// Arguments:
// node - The node to append. Must not be part of an existing list.
//
void Append(LocationInfoListNode* node)
{
assert(node->m_next == nullptr);
if (m_tail == nullptr)
{
assert(m_head == nullptr);
m_head = node;
}
else
{
m_tail->m_next = node;
}
m_tail = node;
}
//------------------------------------------------------------------------
// LocationInfoList::Append: Appends another list to this list.
//
// Arguments:
// other - The list to append.
//
void Append(LocationInfoList other)
{
if (m_tail == nullptr)
{
assert(m_head == nullptr);
m_head = other.m_head;
}
else
{
m_tail->m_next = other.m_head;
}
m_tail = other.m_tail;
}
};
//------------------------------------------------------------------------
// LocationInfoListNodePool: manages a pool of `LocationInfoListNode`
// values to decrease overall memory usage
// during `buildIntervals`.
//
// `buildIntervals` involves creating a list of location info values per
// node that either directly produces a set of registers or that is a
// contained node with register-producing sources. However, these lists
// are short-lived: they are destroyed once the use of the corresponding
// node is processed. As such, there is typically only a small number of
// `LocationInfoListNode` values in use at any given time. Pooling these
// values avoids otherwise frequent allocations.
class LocationInfoListNodePool final
{
LocationInfoListNode* m_freeList;
Compiler* m_compiler;
public:
//------------------------------------------------------------------------
// LocationInfoListNodePool::LocationInfoListNodePool:
// Creates a pool of `LocationInfoListNode` values.
//
// Arguments:
// compiler - The compiler context.
// preallocate - The number of nodes to preallocate.
//
LocationInfoListNodePool(Compiler* compiler, unsigned preallocate = 0) : m_compiler(compiler)
{
if (preallocate > 0)
{
size_t preallocateSize = sizeof(LocationInfoListNode) * preallocate;
auto* preallocatedNodes = reinterpret_cast<LocationInfoListNode*>(compiler->compGetMem(preallocateSize));
LocationInfoListNode* head = preallocatedNodes;
head->m_next = nullptr;
for (unsigned i = 1; i < preallocate; i++)
{
LocationInfoListNode* node = &preallocatedNodes[i];
node->m_next = head;
head = node;
}
m_freeList = head;
}
}
//------------------------------------------------------------------------
// LocationInfoListNodePool::GetNode: Fetches an unused node from the
// pool.
//
// Arguments:
// l - - The `LsraLocation` for the `LocationInfo` value.
// i - The interval for the `LocationInfo` value.
// t - The IR node for the `LocationInfo` value
// regIdx - The register index for the `LocationInfo` value.
//
// Returns:
// A pooled or newly-allocated `LocationInfoListNode`, depending on the
// contents of the pool.
LocationInfoListNode* GetNode(LsraLocation l, Interval* i, GenTree* t, unsigned regIdx = 0)
{
LocationInfoListNode* head = m_freeList;
if (head == nullptr)
{
head = reinterpret_cast<LocationInfoListNode*>(m_compiler->compGetMem(sizeof(LocationInfoListNode)));
}
else
{
m_freeList = head->m_next;
}
head->loc = l;
head->interval = i;
head->treeNode = t;
head->multiRegIdx = regIdx;
head->m_next = nullptr;
return head;
}
//------------------------------------------------------------------------
// LocationInfoListNodePool::ReturnNodes: Returns a list of nodes to the
// pool.
//
// Arguments:
// list - The list to return.
//
void ReturnNodes(LocationInfoList& list)
{
assert(list.m_head != nullptr);
assert(list.m_tail != nullptr);
LocationInfoListNode* head = m_freeList;
list.m_tail->m_next = head;
m_freeList = list.m_head;
}
};
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
VARSET_VALRET_TP
LinearScan::buildUpperVectorSaveRefPositions(GenTree* tree, LsraLocation currentLoc)
{
assert(enregisterLocalVars);
VARSET_TP liveLargeVectors(VarSetOps::MakeEmpty(compiler));
regMaskTP fpCalleeKillSet = RBM_NONE;
if (!VarSetOps::IsEmpty(compiler, largeVectorVars))
{
// We actually need to find any calls that kill the upper-half of the callee-save vector registers.
// But we will use as a proxy any node that kills floating point registers.
// (Note that some calls are masquerading as other nodes at this point so we can't just check for calls.)
fpCalleeKillSet = getKillSetForNode(tree);
if ((fpCalleeKillSet & RBM_FLT_CALLEE_TRASH) != RBM_NONE)
{
VarSetOps::AssignNoCopy(compiler, liveLargeVectors,
VarSetOps::Intersection(compiler, currentLiveVars, largeVectorVars));
VarSetOps::Iter iter(compiler, liveLargeVectors);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
Interval* varInterval = getIntervalForLocalVar(varIndex);
Interval* tempInterval = newInterval(LargeVectorType);
tempInterval->isInternal = true;
RefPosition* pos =
newRefPosition(tempInterval, currentLoc, RefTypeUpperVectorSaveDef, tree, RBM_FLT_CALLEE_SAVED);
// We are going to save the existing relatedInterval of varInterval on tempInterval, so that we can set
// the tempInterval as the relatedInterval of varInterval, so that we can build the corresponding
// RefTypeUpperVectorSaveUse RefPosition. We will then restore the relatedInterval onto varInterval,
// and set varInterval as the relatedInterval of tempInterval.
tempInterval->relatedInterval = varInterval->relatedInterval;
varInterval->relatedInterval = tempInterval;
}
}
}
return liveLargeVectors;
}
void LinearScan::buildUpperVectorRestoreRefPositions(GenTree* tree,
LsraLocation currentLoc,
VARSET_VALARG_TP liveLargeVectors)
{
assert(enregisterLocalVars);
if (!VarSetOps::IsEmpty(compiler, liveLargeVectors))
{
VarSetOps::Iter iter(compiler, liveLargeVectors);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
Interval* varInterval = getIntervalForLocalVar(varIndex);
Interval* tempInterval = varInterval->relatedInterval;
assert(tempInterval->isInternal == true);
RefPosition* pos =
newRefPosition(tempInterval, currentLoc, RefTypeUpperVectorSaveUse, tree, RBM_FLT_CALLEE_SAVED);
// Restore the relatedInterval onto varInterval, and set varInterval as the relatedInterval
// of tempInterval.
varInterval->relatedInterval = tempInterval->relatedInterval;
tempInterval->relatedInterval = varInterval;
}
}
}
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
#ifdef DEBUG
//------------------------------------------------------------------------
// ComputeOperandDstCount: computes the number of registers defined by a
// node.
//
// For most nodes, this is simple:
// - Nodes that do not produce values (e.g. stores and other void-typed
// nodes) and nodes that immediately use the registers they define
// produce no registers
// - Nodes that are marked as defining N registers define N registers.
//
// For contained nodes, however, things are more complicated: for purposes
// of bookkeeping, a contained node is treated as producing the transitive
// closure of the registers produced by its sources.
//
// Arguments:
// operand - The operand for which to compute a register count.
//
// Returns:
// The number of registers defined by `operand`.
//
static int ComputeOperandDstCount(GenTree* operand)
{
TreeNodeInfo& operandInfo = operand->gtLsraInfo;
if (operandInfo.isLocalDefUse)
{
// Operands that define an unused value do not produce any registers.
return 0;
}
else if (operandInfo.dstCount != 0)
{
// Operands that have a specified number of destination registers consume all of their operands
// and therefore produce exactly that number of registers.
return operandInfo.dstCount;
}
else if (operandInfo.srcCount != 0)
{
// If an operand has no destination registers but does have source registers, it must be a store
// or a compare.
assert(operand->OperIsStore() || operand->OperIsBlkOp() || operand->OperIsPutArgStk() ||
operand->OperIsCompare() || operand->OperIs(GT_CMP, GT_JCMP) || operand->IsSIMDEqualityOrInequality());
return 0;
}
else if (!operand->OperIsFieldListHead() && (operand->OperIsStore() || operand->TypeGet() == TYP_VOID))
{
// Stores and void-typed operands may be encountered when processing call nodes, which contain
// pointers to argument setup stores.
return 0;
}
else if (operand->OperIsPutArgStk())
{
// A PUTARG_STK argument is an operand of a call, but is neither contained, nor does it produce
// a result.
assert(!operand->isContained());
return 0;
}
else
{
// If a field list or non-void-typed operand is not an unused value and does not have source registers,
// that argument is contained within its parent and produces `sum(operand_dst_count)` registers.
int dstCount = 0;
for (GenTree* op : operand->Operands())
{
dstCount += ComputeOperandDstCount(op);
}
return dstCount;
}
}
//------------------------------------------------------------------------
// ComputeAvailableSrcCount: computes the number of registers available as
// sources for a node.
//
// This is simply the sum of the number of registers produced by each
// operand to the node.
//
// Arguments:
// node - The node for which to compute a source count.
//
// Retures:
// The number of registers available as sources for `node`.
//
static int ComputeAvailableSrcCount(GenTree* node)
{
int numSources = 0;
for (GenTree* operand : node->Operands())
{
numSources += ComputeOperandDstCount(operand);
}
return numSources;
}
#endif // DEBUG
static GenTree* GetFirstOperand(GenTree* node)
{
GenTree* firstOperand = nullptr;
node->VisitOperands([&firstOperand](GenTree* operand) -> GenTree::VisitResult {
firstOperand = operand;
return GenTree::VisitResult::Abort;
});
return firstOperand;
}
void LinearScan::buildRefPositionsForNode(GenTree* tree,
BasicBlock* block,
LocationInfoListNodePool& listNodePool,
HashTableBase<GenTree*, LocationInfoList>& operandToLocationInfoMap,
LsraLocation currentLoc)
{
#ifdef _TARGET_ARM_
assert(!isRegPairType(tree->TypeGet()));
#endif // _TARGET_ARM_
// The LIR traversal doesn't visit GT_LIST or GT_ARGPLACE nodes.
// GT_CLS_VAR nodes should have been eliminated by rationalizer.
assert(tree->OperGet() != GT_ARGPLACE);
assert(tree->OperGet() != GT_LIST);
assert(tree->OperGet() != GT_CLS_VAR);
// The LIR traversal visits only the first node in a GT_FIELD_LIST.
assert((tree->OperGet() != GT_FIELD_LIST) || tree->AsFieldList()->IsFieldListHead());
// The set of internal temporary registers used by this node are stored in the
// gtRsvdRegs register mask. Clear it out.
tree->gtRsvdRegs = RBM_NONE;
TreeNodeInfo info = tree->gtLsraInfo;
assert(info.IsValid(this));
int consume = info.srcCount;
int produce = info.dstCount;
#ifdef DEBUG
if (VERBOSE)
{
lsraDispNode(tree, LSRA_DUMP_REFPOS, (produce != 0));
JITDUMP("\n");
if (tree->isContained())
{
JITDUMP("Contained\n");
}
else if (tree->OperIs(GT_LCL_VAR, GT_LCL_FLD) && info.isLocalDefUse)
{
JITDUMP("Unused\n");
}
else
{
JITDUMP(" consume=%d produce=%d\n", consume, produce);
}
if (consume != 0)
{
JITDUMP("at start of tree, map contains: { ");
bool first = true;
for (auto kvp : operandToLocationInfoMap)
{
GenTree* node = kvp.Key();
LocationInfoList defList = kvp.Value();
JITDUMP("%sN%03u. %s -> (", first ? "" : "; ", node->gtSeqNum, GenTree::OpName(node->OperGet()));
for (LocationInfoListNode *def = defList.Begin(), *end = defList.End(); def != end; def = def->Next())
{
JITDUMP("%s%d.N%03u", def == defList.Begin() ? "" : ", ", def->loc, def->treeNode->gtSeqNum);
}
JITDUMP(")");
first = false;
}
JITDUMP(" }\n");
}
}
#endif // DEBUG
assert(((consume == 0) && (produce == 0)) || (ComputeAvailableSrcCount(tree) == consume));
if (tree->OperIs(GT_LCL_VAR, GT_LCL_FLD))
{
LclVarDsc* const varDsc = &compiler->lvaTable[tree->AsLclVarCommon()->gtLclNum];
if (isCandidateVar(varDsc))
{
assert(consume == 0);
// We handle tracked variables differently from non-tracked ones. If it is tracked,
// we simply add a use or def of the tracked variable. Otherwise, for a use we need
// to actually add the appropriate references for loading or storing the variable.
//
// It won't actually get used or defined until the appropriate ancestor tree node
// is processed, unless this is marked "isLocalDefUse" because it is a stack-based argument
// to a call
assert(varDsc->lvTracked);
unsigned varIndex = varDsc->lvVarIndex;
// We have only approximate last-use information at this point. This is because the
// execution order doesn't actually reflect the true order in which the localVars
// are referenced - but the order of the RefPositions will, so we recompute it after
// RefPositions are built.
// Use the old value for setting currentLiveVars - note that we do this with the
// not-quite-correct setting of lastUse. However, this is OK because
// 1) this is only for preferencing, which doesn't require strict correctness, and
// 2) the cases where these out-of-order uses occur should not overlap a kill.
// TODO-Throughput: clean this up once we have the execution order correct. At that point
// we can update currentLiveVars at the same place that we create the RefPosition.
if ((tree->gtFlags & GTF_VAR_DEATH) != 0)
{
VarSetOps::RemoveElemD(compiler, currentLiveVars, varIndex);
}
if (!info.isLocalDefUse && !tree->isContained())
{
assert(produce != 0);
LocationInfoList list(listNodePool.GetNode(currentLoc, getIntervalForLocalVar(varIndex), tree));
bool added = operandToLocationInfoMap.AddOrUpdate(tree, list);
assert(added);
tree->gtLsraInfo.definesAnyRegisters = true;
}
return;
}
}
if (tree->isContained())
{
assert(!info.isLocalDefUse);
assert(consume == 0);
assert(produce == 0);
assert(info.internalIntCount == 0);
assert(info.internalFloatCount == 0);
// Contained nodes map to the concatenated lists of their operands.
LocationInfoList locationInfoList;
tree->VisitOperands([&](GenTree* op) -> GenTree::VisitResult {
if (!op->gtLsraInfo.definesAnyRegisters)
{
assert(ComputeOperandDstCount(op) == 0);
return GenTree::VisitResult::Continue;
}
LocationInfoList operandList;
bool removed = operandToLocationInfoMap.TryRemove(op, &operandList);
assert(removed);
locationInfoList.Append(operandList);
return GenTree::VisitResult::Continue;
});
if (!locationInfoList.IsEmpty())
{
bool added = operandToLocationInfoMap.AddOrUpdate(tree, locationInfoList);
assert(added);
tree->gtLsraInfo.definesAnyRegisters = true;
}
JITDUMP("\n");
return;
}
// Handle the case of local variable assignment
Interval* varDefInterval = nullptr;
RefType defRefType = RefTypeDef;
GenTree* defNode = tree;
// noAdd means the node creates a def but for purposes of map
// management do not add it because data is not flowing up the
// tree but over (as in ASG nodes)
bool noAdd = info.isLocalDefUse;
RefPosition* prevPos = nullptr;
bool isSpecialPutArg = false;
assert(!tree->OperIsAssignment());
if (tree->OperIsLocalStore())
{
GenTreeLclVarCommon* const store = tree->AsLclVarCommon();
assert((consume > 1) || (regType(store->gtOp1->TypeGet()) == regType(store->TypeGet())));
LclVarDsc* varDsc = &compiler->lvaTable[store->gtLclNum];
if (isCandidateVar(varDsc))
{
// We always push the tracked lclVar intervals
assert(varDsc->lvTracked);
unsigned varIndex = varDsc->lvVarIndex;
varDefInterval = getIntervalForLocalVar(varIndex);
defRefType = refTypeForLocalRefNode(tree);
defNode = tree;
if (produce == 0)
{
produce = 1;
noAdd = true;
}
assert(consume <= MAX_RET_REG_COUNT);
if (consume == 1)
{
// Get the location info for the register defined by the first operand.
LocationInfoList operandDefs;
bool found = operandToLocationInfoMap.TryGetValue(GetFirstOperand(tree), &operandDefs);
assert(found);
// Since we only expect to consume one register, we should only have a single register to
// consume.
assert(operandDefs.Begin()->Next() == operandDefs.End());
LocationInfo& operandInfo = *static_cast<LocationInfo*>(operandDefs.Begin());
Interval* srcInterval = operandInfo.interval;
if (srcInterval->relatedInterval == nullptr)
{
// Preference the source to the dest, unless this is a non-last-use localVar.
// Note that the last-use info is not correct, but it is a better approximation than preferencing
// the source to the dest, if the source's lifetime extends beyond the dest.
if (!srcInterval->isLocalVar || (operandInfo.treeNode->gtFlags & GTF_VAR_DEATH) != 0)
{
srcInterval->assignRelatedInterval(varDefInterval);
}
}
else if (!srcInterval->isLocalVar)
{
// Preference the source to dest, if src is not a local var.
srcInterval->assignRelatedInterval(varDefInterval);
}
}
if ((tree->gtFlags & GTF_VAR_DEATH) == 0)
{
VarSetOps::AddElemD(compiler, currentLiveVars, varIndex);
}
}
else if (store->gtOp1->OperIs(GT_BITCAST))
{
store->gtType = store->gtOp1->gtType = store->gtOp1->AsUnOp()->gtOp1->TypeGet();
// Get the location info for the register defined by the first operand.
LocationInfoList operandDefs;
bool found = operandToLocationInfoMap.TryGetValue(GetFirstOperand(store), &operandDefs);
assert(found);
// Since we only expect to consume one register, we should only have a single register to consume.
assert(operandDefs.Begin()->Next() == operandDefs.End());
LocationInfo& operandInfo = *static_cast<LocationInfo*>(operandDefs.Begin());
Interval* srcInterval = operandInfo.interval;
srcInterval->registerType = regType(store->TypeGet());
RefPosition* srcDefPosition = srcInterval->firstRefPosition;
assert(srcDefPosition != nullptr);
assert(srcDefPosition->refType == RefTypeDef);
assert(srcDefPosition->treeNode == store->gtOp1);
srcDefPosition->registerAssignment = allRegs(store->TypeGet());
store->gtOp1->gtLsraInfo.setSrcCandidates(this, allRegs(store->TypeGet()));
}
}
else if (noAdd && produce == 0)
{
// This is the case for dead nodes that occur after
// tree rationalization
// TODO-Cleanup: Identify and remove these dead nodes prior to register allocation.
if (tree->IsMultiRegCall())
{
// In case of multi-reg call node, produce = number of return registers
produce = tree->AsCall()->GetReturnTypeDesc()->GetReturnRegCount();
}
else
{
produce = 1;
}
}
Interval* prefSrcInterval = nullptr;
// If this is a binary operator that will be encoded with 2 operand fields
// (i.e. the target is read-modify-write), preference the dst to op1.
bool hasDelayFreeSrc = tree->gtLsraInfo.hasDelayFreeSrc;
#if defined(DEBUG) && defined(_TARGET_X86_)
// On x86, `LSRA_LIMIT_CALLER` is too restrictive to allow the use of special put args: this stress mode
// leaves only three registers allocatable--eax, ecx, and edx--of which the latter two are also used for the
// first two integral arguments to a call. This can leave us with too few registers to succesfully allocate in
// situations like the following:
//
// t1026 = lclVar ref V52 tmp35 u:3 REG NA <l:$3a1, c:$98d>
//
// /--* t1026 ref
// t1352 = * putarg_reg ref REG NA
//
// t342 = lclVar int V14 loc6 u:4 REG NA $50c
//
// t343 = const int 1 REG NA $41
//
// /--* t342 int
// +--* t343 int
// t344 = * + int REG NA $495
//
// t345 = lclVar int V04 arg4 u:2 REG NA $100
//
// /--* t344 int
// +--* t345 int
// t346 = * % int REG NA $496
//
// /--* t346 int
// t1353 = * putarg_reg int REG NA
//
// t1354 = lclVar ref V52 tmp35 (last use) REG NA
//
// /--* t1354 ref
// t1355 = * lea(b+0) byref REG NA
//
// Here, the first `putarg_reg` would normally be considered a special put arg, which would remove `ecx` from the
// set of allocatable registers, leaving only `eax` and `edx`. The allocator will then fail to allocate a register
// for the def of `t345` if arg4 is not a register candidate: the corresponding ref position will be constrained to
// { `ecx`, `ebx`, `esi`, `edi` }, which `LSRA_LIMIT_CALLER` will further constrain to `ecx`, which will not be
// available due to the special put arg.
const bool supportsSpecialPutArg = getStressLimitRegs() != LSRA_LIMIT_CALLER;
#else
const bool supportsSpecialPutArg = true;
#endif
if (supportsSpecialPutArg && tree->OperGet() == GT_PUTARG_REG && isCandidateLocalRef(tree->gtGetOp1()) &&
(tree->gtGetOp1()->gtFlags & GTF_VAR_DEATH) == 0)
{
// This is the case for a "pass-through" copy of a lclVar. In the case where it is a non-last-use,
// we don't want the def of the copy to kill the lclVar register, if it is assigned the same register
// (which is actually what we hope will happen).
JITDUMP("Setting putarg_reg as a pass-through of a non-last use lclVar\n");
// Get the register information for the first operand of the node.
LocationInfoList operandDefs;
bool found = operandToLocationInfoMap.TryGetValue(GetFirstOperand(tree), &operandDefs);
assert(found);
// Preference the destination to the interval of the first register defined by the first operand.
Interval* srcInterval = operandDefs.Begin()->interval;
assert(srcInterval->isLocalVar);
prefSrcInterval = srcInterval;
isSpecialPutArg = true;
}
RefPosition* internalRefs[MaxInternalRegisters];
#ifdef DEBUG
// Number of registers required for tree node is the sum of
// consume + produce + internalCount. This is the minimum
// set of registers that needs to be ensured in candidate
// set of ref positions created.
unsigned minRegCount = consume + produce + info.internalIntCount + info.internalFloatCount;
#endif // DEBUG
// make intervals for all the 'internal' register requirements for this node
// where internal means additional registers required temporarily
int internalCount = buildInternalRegisterDefsForNode(tree, currentLoc, internalRefs DEBUG_ARG(minRegCount));
// pop all ref'd tree temps
tree->VisitOperands([&](GenTree* operand) -> GenTree::VisitResult {
// Skip operands that do not define any registers, whether directly or indirectly.
if (!operand->gtLsraInfo.definesAnyRegisters)
{
return GenTree::VisitResult::Continue;
}
// Remove the list of registers defined by the current operand from the map. Note that this
// is only correct because tree nodes are singly-used: if this property ever changes (e.g.
// if tree nodes are eventually allowed to be multiply-used), then the removal is only
// correct at the last use.
LocationInfoList operandDefs;
bool removed = operandToLocationInfoMap.TryRemove(operand, &operandDefs);
assert(removed);
assert(!operandDefs.IsEmpty());
#ifdef _TARGET_ARM_
regMaskTP currCandidates = RBM_NONE;
#endif // _TARGET_ARM_
LocationInfoListNode* const operandDefsEnd = operandDefs.End();
for (LocationInfoListNode* operandDefsIterator = operandDefs.Begin(); operandDefsIterator != operandDefsEnd;
operandDefsIterator = operandDefsIterator->Next())
{
LocationInfo& locInfo = *static_cast<LocationInfo*>(operandDefsIterator);
// for interstitial tree temps, a use is always last and end; this is set by default in newRefPosition
GenTree* const useNode = locInfo.treeNode;
assert(useNode != nullptr);
Interval* const i = locInfo.interval;
if (useNode->gtLsraInfo.isTgtPref)
{
prefSrcInterval = i;
}
const bool delayRegFree = (hasDelayFreeSrc && useNode->gtLsraInfo.isDelayFree);
#ifdef DEBUG
// If delayRegFree, then Use will interfere with the destination of
// the consuming node. Therefore, we also need add the kill set of
// consuming node to minRegCount.
//
// For example consider the following IR on x86, where v01 and v02
// are method args coming in ecx and edx respectively.
// GT_DIV(v01, v02)
//
// For GT_DIV minRegCount will be 3 without adding kill set
// of GT_DIV node.
//
// Assume further JitStressRegs=2, which would constrain
// candidates to callee trashable regs { eax, ecx, edx } on
// use positions of v01 and v02. LSRA allocates ecx for v01.
// Use position of v02 cannot be allocated a regs since it
// is marked delay-reg free and {eax,edx} are getting killed
// before the def of GT_DIV. For this reason, minRegCount
// for Use position of v02 also needs to take into account
// of kill set of its consuming node.
unsigned minRegCountForUsePos = minRegCount;
if (delayRegFree && (lsraStressMask != 0))
{
regMaskTP killMask = getKillSetForNode(tree);
if (killMask != RBM_NONE)
{
minRegCountForUsePos += genCountBits(killMask);
}
}
#endif // DEBUG
regMaskTP candidates = getUseCandidates(useNode);
#ifdef _TARGET_ARM_
if (useNode->OperIsPutArgSplit() || useNode->OperIsMultiRegOp())
{
// get i-th candidate, set bits in useCandidates must be in sequential order.
candidates = genFindLowestReg(candidates & ~currCandidates);
currCandidates |= candidates;
}
#endif // _TARGET_ARM_
assert((candidates & allRegs(i->registerType)) != 0);
// For non-localVar uses we record nothing, as nothing needs to be written back to the tree.
GenTree* const refPosNode = i->isLocalVar ? useNode : nullptr;
RefPosition* pos = newRefPosition(i, currentLoc, RefTypeUse, refPosNode, candidates,
locInfo.multiRegIdx DEBUG_ARG(minRegCountForUsePos));
if (delayRegFree)
{
pos->delayRegFree = true;
}
if (useNode->IsRegOptional())
{
pos->setAllocateIfProfitable(true);
}
}
listNodePool.ReturnNodes(operandDefs);
return GenTree::VisitResult::Continue;
});
buildInternalRegisterUsesForNode(tree, currentLoc, internalRefs, internalCount DEBUG_ARG(minRegCount));
RegisterType registerType = getDefType(tree);
regMaskTP candidates = getDefCandidates(tree);
regMaskTP useCandidates = getUseCandidates(tree);
#ifdef DEBUG
if (VERBOSE && produce)
{
printf("Def candidates ");
dumpRegMask(candidates);
printf(", Use candidates ");
dumpRegMask(useCandidates);
printf("\n");
}
#endif // DEBUG
#if defined(_TARGET_AMD64_)
// Multi-reg call node is the only node that could produce multi-reg value
assert(produce <= 1 || (tree->IsMultiRegCall() && produce == MAX_RET_REG_COUNT));
#endif // _TARGET_xxx_
// Add kill positions before adding def positions
buildKillPositionsForNode(tree, currentLoc + 1);
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
VARSET_TP liveLargeVectors(VarSetOps::UninitVal());
if (enregisterLocalVars && (RBM_FLT_CALLEE_SAVED != RBM_NONE))
{
// Build RefPositions for saving any live large vectors.
// This must be done after the kills, so that we know which large vectors are still live.
VarSetOps::AssignNoCopy(compiler, liveLargeVectors, buildUpperVectorSaveRefPositions(tree, currentLoc + 1));
}
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
ReturnTypeDesc* retTypeDesc = nullptr;
bool isMultiRegCall = tree->IsMultiRegCall();
if (isMultiRegCall)
{
retTypeDesc = tree->AsCall()->GetReturnTypeDesc();
assert((int)genCountBits(candidates) == produce);
assert(candidates == retTypeDesc->GetABIReturnRegs());
}
// push defs
LocationInfoList locationInfoList;
LsraLocation defLocation = currentLoc + 1;
for (int i = 0; i < produce; i++)
{
regMaskTP currCandidates = candidates;
Interval* interval = varDefInterval;
// In case of multi-reg call node, registerType is given by
// the type of ith position return register.
if (isMultiRegCall)
{
registerType = retTypeDesc->GetReturnRegType((unsigned)i);
currCandidates = genRegMask(retTypeDesc->GetABIReturnReg(i));
useCandidates = allRegs(registerType);
}
#ifdef _TARGET_ARM_
// If oper is GT_PUTARG_REG, set bits in useCandidates must be in sequential order.
if (tree->OperIsPutArgSplit() || tree->OperIsMultiRegOp())
{
// get i-th candidate
currCandidates = genFindLowestReg(candidates);
candidates &= ~currCandidates;
}
#endif // _TARGET_ARM_
if (interval == nullptr)
{
// Make a new interval
interval = newInterval(registerType);
if (hasDelayFreeSrc)
{
interval->hasNonCommutativeRMWDef = true;
}
else if (tree->OperIsConst())
{
assert(!tree->IsReuseRegVal());
interval->isConstant = true;
}
if ((currCandidates & useCandidates) != RBM_NONE)
{
interval->updateRegisterPreferences(currCandidates & useCandidates);
}
if (isSpecialPutArg)
{
interval->isSpecialPutArg = true;
}
}
else
{
assert(registerTypesEquivalent(interval->registerType, registerType));
}
if (prefSrcInterval != nullptr)
{
interval->assignRelatedIntervalIfUnassigned(prefSrcInterval);
}
// for assignments, we want to create a refposition for the def
// but not push it
if (!noAdd)
{
locationInfoList.Append(listNodePool.GetNode(defLocation, interval, tree, (unsigned)i));
}
RefPosition* pos = newRefPosition(interval, defLocation, defRefType, defNode, currCandidates,
(unsigned)i DEBUG_ARG(minRegCount));
if (info.isLocalDefUse)
{
// This must be an unused value, OR it is a special node for which we allocate
// a target register even though it produces no value.
assert(defNode->IsUnusedValue() || (defNode->gtOper == GT_LOCKADD));
pos->isLocalDefUse = true;
pos->lastUse = true;
}
interval->updateRegisterPreferences(currCandidates);
interval->updateRegisterPreferences(useCandidates);
}
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
// SaveDef position must be at the same location as Def position of call node.
if (enregisterLocalVars)
{
buildUpperVectorRestoreRefPositions(tree, defLocation, liveLargeVectors);
}
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (!locationInfoList.IsEmpty())
{
bool added = operandToLocationInfoMap.AddOrUpdate(tree, locationInfoList);
assert(added);
tree->gtLsraInfo.definesAnyRegisters = true;
}
JITDUMP("\n");
}
// make an interval for each physical register
void LinearScan::buildPhysRegRecords()
{
RegisterType regType = IntRegisterType;
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
RegRecord* curr = &physRegs[reg];
curr->init(reg);
}
}
BasicBlock* getNonEmptyBlock(BasicBlock* block)
{
while (block != nullptr && block->bbTreeList == nullptr)
{
BasicBlock* nextBlock = block->bbNext;
// Note that here we use the version of NumSucc that does not take a compiler.
// That way this doesn't have to take a compiler, or be an instance method, e.g. of LinearScan.
// If we have an empty block, it must have jump type BBJ_NONE or BBJ_ALWAYS, in which
// case we don't need the version that takes a compiler.
assert(block->NumSucc() == 1 && ((block->bbJumpKind == BBJ_ALWAYS) || (block->bbJumpKind == BBJ_NONE)));
// sometimes the first block is empty and ends with an uncond branch
// assert( block->GetSucc(0) == nextBlock);
block = nextBlock;
}
assert(block != nullptr && block->bbTreeList != nullptr);
return block;
}
//------------------------------------------------------------------------
// insertZeroInitRefPositions: Handle lclVars that are live-in to the first block
//
// Notes:
// Prior to calling this method, 'currentLiveVars' must be set to the set of register
// candidate variables that are liveIn to the first block.
// For each register candidate that is live-in to the first block:
// - If it is a GC ref, or if compInitMem is set, a ZeroInit RefPosition will be created.
// - Otherwise, it will be marked as spilled, since it will not be assigned a register
// on entry and will be loaded from memory on the undefined path.
// Note that, when the compInitMem option is not set, we may encounter these on
// paths that are protected by the same condition as an earlier def. However, since
// we don't do the analysis to determine this - and couldn't rely on always identifying
// such cases even if we tried - we must conservatively treat the undefined path as
// being possible. This is a relatively rare case, so the introduced conservatism is
// not expected to warrant the analysis required to determine the best placement of
// an initialization.
//
void LinearScan::insertZeroInitRefPositions()
{
assert(enregisterLocalVars);
#ifdef DEBUG
VARSET_TP expectedLiveVars(VarSetOps::Intersection(compiler, registerCandidateVars, compiler->fgFirstBB->bbLiveIn));
assert(VarSetOps::Equal(compiler, currentLiveVars, expectedLiveVars));
#endif // DEBUG
// insert defs for this, then a block boundary
VarSetOps::Iter iter(compiler, currentLiveVars);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* varDsc = compiler->lvaTable + varNum;
if (!varDsc->lvIsParam && isCandidateVar(varDsc))
{
JITDUMP("V%02u was live in to first block:", varNum);
Interval* interval = getIntervalForLocalVar(varIndex);
if (compiler->info.compInitMem || varTypeIsGC(varDsc->TypeGet()))
{
JITDUMP(" creating ZeroInit\n");
GenTree* firstNode = getNonEmptyBlock(compiler->fgFirstBB)->firstNode();
RefPosition* pos =
newRefPosition(interval, MinLocation, RefTypeZeroInit, firstNode, allRegs(interval->registerType));
varDsc->lvMustInit = true;
}
else
{
setIntervalAsSpilled(interval);
JITDUMP(" marking as spilled\n");
}
}
}
}
#if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
// -----------------------------------------------------------------------
// Sets the register state for an argument of type STRUCT for System V systems.
// See Compiler::raUpdateRegStateForArg(RegState *regState, LclVarDsc *argDsc) in regalloc.cpp
// for how state for argument is updated for unix non-structs and Windows AMD64 structs.
void LinearScan::unixAmd64UpdateRegStateForArg(LclVarDsc* argDsc)
{
assert(varTypeIsStruct(argDsc));
RegState* intRegState = &compiler->codeGen->intRegState;
RegState* floatRegState = &compiler->codeGen->floatRegState;
if ((argDsc->lvArgReg != REG_STK) && (argDsc->lvArgReg != REG_NA))
{
if (genRegMask(argDsc->lvArgReg) & (RBM_ALLFLOAT))
{
assert(genRegMask(argDsc->lvArgReg) & (RBM_FLTARG_REGS));
floatRegState->rsCalleeRegArgMaskLiveIn |= genRegMask(argDsc->lvArgReg);
}
else
{
assert(genRegMask(argDsc->lvArgReg) & (RBM_ARG_REGS));
intRegState->rsCalleeRegArgMaskLiveIn |= genRegMask(argDsc->lvArgReg);
}
}
if ((argDsc->lvOtherArgReg != REG_STK) && (argDsc->lvOtherArgReg != REG_NA))
{
if (genRegMask(argDsc->lvOtherArgReg) & (RBM_ALLFLOAT))
{
assert(genRegMask(argDsc->lvOtherArgReg) & (RBM_FLTARG_REGS));
floatRegState->rsCalleeRegArgMaskLiveIn |= genRegMask(argDsc->lvOtherArgReg);
}
else
{
assert(genRegMask(argDsc->lvOtherArgReg) & (RBM_ARG_REGS));
intRegState->rsCalleeRegArgMaskLiveIn |= genRegMask(argDsc->lvOtherArgReg);
}
}
}
#endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
//------------------------------------------------------------------------
// updateRegStateForArg: Updates rsCalleeRegArgMaskLiveIn for the appropriate
// regState (either compiler->intRegState or compiler->floatRegState),
// with the lvArgReg on "argDsc"
//
// Arguments:
// argDsc - the argument for which the state is to be updated.
//
// Return Value: None
//
// Assumptions:
// The argument is live on entry to the function
// (or is untracked and therefore assumed live)
//
// Notes:
// This relies on a method in regAlloc.cpp that is shared between LSRA
// and regAlloc. It is further abstracted here because regState is updated
// separately for tracked and untracked variables in LSRA.
//
void LinearScan::updateRegStateForArg(LclVarDsc* argDsc)
{
#if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
// For System V AMD64 calls the argDsc can have 2 registers (for structs.)
// Handle them here.
if (varTypeIsStruct(argDsc))
{
unixAmd64UpdateRegStateForArg(argDsc);
}
else
#endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
{
RegState* intRegState = &compiler->codeGen->intRegState;
RegState* floatRegState = &compiler->codeGen->floatRegState;
// In the case of AMD64 we'll still use the floating point registers
// to model the register usage for argument on vararg calls, so
// we will ignore the varargs condition to determine whether we use
// XMM registers or not for setting up the call.
bool isFloat = (isFloatRegType(argDsc->lvType)
#ifndef _TARGET_AMD64_
&& !compiler->info.compIsVarArgs
#endif
&& !compiler->opts.compUseSoftFP);
if (argDsc->lvIsHfaRegArg())
{
isFloat = true;
}
if (isFloat)
{
JITDUMP("Float arg V%02u in reg %s\n", (argDsc - compiler->lvaTable), getRegName(argDsc->lvArgReg));
compiler->raUpdateRegStateForArg(floatRegState, argDsc);
}
else
{
JITDUMP("Int arg V%02u in reg %s\n", (argDsc - compiler->lvaTable), getRegName(argDsc->lvArgReg));
#if FEATURE_MULTIREG_ARGS
if (argDsc->lvOtherArgReg != REG_NA)
{
JITDUMP("(second half) in reg %s\n", getRegName(argDsc->lvOtherArgReg));
}
#endif // FEATURE_MULTIREG_ARGS
compiler->raUpdateRegStateForArg(intRegState, argDsc);
}
}
}
//------------------------------------------------------------------------
// findPredBlockForLiveIn: Determine which block should be used for the register locations of the live-in variables.
//
// Arguments:
// block - The block for which we're selecting a predecesor.
// prevBlock - The previous block in in allocation order.
// pPredBlockIsAllocated - A debug-only argument that indicates whether any of the predecessors have been seen
// in allocation order.
//
// Return Value:
// The selected predecessor.
//
// Assumptions:
// in DEBUG, caller initializes *pPredBlockIsAllocated to false, and it will be set to true if the block
// returned is in fact a predecessor.
//
// Notes:
// This will select a predecessor based on the heuristics obtained by getLsraBlockBoundaryLocations(), which can be
// one of:
// LSRA_BLOCK_BOUNDARY_PRED - Use the register locations of a predecessor block (default)
// LSRA_BLOCK_BOUNDARY_LAYOUT - Use the register locations of the previous block in layout order.
// This is the only case where this actually returns a different block.
// LSRA_BLOCK_BOUNDARY_ROTATE - Rotate the register locations from a predecessor.
// For this case, the block returned is the same as for LSRA_BLOCK_BOUNDARY_PRED, but
// the register locations will be "rotated" to stress the resolution and allocation
// code.
BasicBlock* LinearScan::findPredBlockForLiveIn(BasicBlock* block,
BasicBlock* prevBlock DEBUGARG(bool* pPredBlockIsAllocated))
{
BasicBlock* predBlock = nullptr;
#ifdef DEBUG
assert(*pPredBlockIsAllocated == false);
if (getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_LAYOUT)
{
if (prevBlock != nullptr)
{
predBlock = prevBlock;
}
}
else
#endif // DEBUG
if (block != compiler->fgFirstBB)
{
predBlock = block->GetUniquePred(compiler);
if (predBlock != nullptr)
{
if (isBlockVisited(predBlock))
{
if (predBlock->bbJumpKind == BBJ_COND)
{
// Special handling to improve matching on backedges.
BasicBlock* otherBlock = (block == predBlock->bbNext) ? predBlock->bbJumpDest : predBlock->bbNext;
noway_assert(otherBlock != nullptr);
if (isBlockVisited(otherBlock))
{
// This is the case when we have a conditional branch where one target has already
// been visited. It would be best to use the same incoming regs as that block,
// so that we have less likelihood of having to move registers.
// For example, in determining the block to use for the starting register locations for
// "block" in the following example, we'd like to use the same predecessor for "block"
// as for "otherBlock", so that both successors of predBlock have the same locations, reducing
// the likelihood of needing a split block on a backedge:
//
// otherPred
// |
// otherBlock <-+
// . . . |
// |
// predBlock----+
// |
// block
//
for (flowList* pred = otherBlock->bbPreds; pred != nullptr; pred = pred->flNext)
{
BasicBlock* otherPred = pred->flBlock;
if (otherPred->bbNum == blockInfo[otherBlock->bbNum].predBBNum)
{
predBlock = otherPred;
break;
}
}
}
}
}
else
{
predBlock = nullptr;
}
}
else
{
for (flowList* pred = block->bbPreds; pred != nullptr; pred = pred->flNext)
{
BasicBlock* candidatePredBlock = pred->flBlock;
if (isBlockVisited(candidatePredBlock))
{
if (predBlock == nullptr || predBlock->bbWeight < candidatePredBlock->bbWeight)
{
predBlock = candidatePredBlock;
INDEBUG(*pPredBlockIsAllocated = true;)
}
}
}
}
if (predBlock == nullptr)
{
predBlock = prevBlock;
assert(predBlock != nullptr);
JITDUMP("\n\nNo allocated predecessor; ");
}
}
return predBlock;
}
void LinearScan::buildIntervals()
{
BasicBlock* block;
JITDUMP("\nbuildIntervals ========\n");
// Now build (empty) records for all of the physical registers
buildPhysRegRecords();
#ifdef DEBUG
if (VERBOSE)
{
printf("\n-----------------\n");
printf("LIVENESS:\n");
printf("-----------------\n");
foreach_block(compiler, block)
{
printf("BB%02u use def in out\n", block->bbNum);
dumpConvertedVarSet(compiler, block->bbVarUse);
printf("\n");
dumpConvertedVarSet(compiler, block->bbVarDef);
printf("\n");
dumpConvertedVarSet(compiler, block->bbLiveIn);
printf("\n");
dumpConvertedVarSet(compiler, block->bbLiveOut);
printf("\n");
}
}
#endif // DEBUG
#if DOUBLE_ALIGN
// We will determine whether we should double align the frame during
// identifyCandidates(), but we initially assume that we will not.
doDoubleAlign = false;
#endif
identifyCandidates();
// Figure out if we're going to use a frame pointer. We need to do this before building
// the ref positions, because those objects will embed the frame register in various register masks
// if the frame pointer is not reserved. If we decide to have a frame pointer, setFrameType() will
// remove the frame pointer from the masks.
setFrameType();
DBEXEC(VERBOSE, TupleStyleDump(LSRA_DUMP_PRE));
// second part:
JITDUMP("\nbuildIntervals second part ========\n");
LsraLocation currentLoc = 0;
// TODO-Cleanup: This duplicates prior behavior where entry (ParamDef) RefPositions were
// being assigned the bbNum of the last block traversed in the 2nd phase of Lowering.
// Previously, the block sequencing was done for the (formerly separate) TreeNodeInfoInit pass,
// and the curBBNum was left as the last block sequenced. This block was then used to set the
// weight for the entry (ParamDef) RefPositions. It would be logical to set this to the
// normalized entry weight (compiler->fgCalledCount), but that results in a net regression.
if (!blockSequencingDone)
{
setBlockSequence();
}
curBBNum = blockSequence[bbSeqCount - 1]->bbNum;
// Next, create ParamDef RefPositions for all the tracked parameters, in order of their varIndex.
// Assign these RefPositions to the (nonexistent) BB0.
curBBNum = 0;
LclVarDsc* argDsc;
unsigned int lclNum;
RegState* intRegState = &compiler->codeGen->intRegState;
RegState* floatRegState = &compiler->codeGen->floatRegState;
intRegState->rsCalleeRegArgMaskLiveIn = RBM_NONE;
floatRegState->rsCalleeRegArgMaskLiveIn = RBM_NONE;
for (unsigned int varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++)
{
lclNum = compiler->lvaTrackedToVarNum[varIndex];
argDsc = &(compiler->lvaTable[lclNum]);
if (!argDsc->lvIsParam)
{
continue;
}
// Only reserve a register if the argument is actually used.
// Is it dead on entry? If compJmpOpUsed is true, then the arguments
// have to be kept alive, so we have to consider it as live on entry.
// Use lvRefCnt instead of checking bbLiveIn because if it's volatile we
// won't have done dataflow on it, but it needs to be marked as live-in so
// it will get saved in the prolog.
if (!compiler->compJmpOpUsed && argDsc->lvRefCnt == 0 && !compiler->opts.compDbgCode)
{
continue;
}
if (argDsc->lvIsRegArg)
{
updateRegStateForArg(argDsc);
}
if (isCandidateVar(argDsc))
{
Interval* interval = getIntervalForLocalVar(varIndex);
regMaskTP mask = allRegs(TypeGet(argDsc));
if (argDsc->lvIsRegArg)
{
// Set this interval as currently assigned to that register
regNumber inArgReg = argDsc->lvArgReg;
assert(inArgReg < REG_COUNT);
mask = genRegMask(inArgReg);
assignPhysReg(inArgReg, interval);
}
RefPosition* pos = newRefPosition(interval, MinLocation, RefTypeParamDef, nullptr, mask);
}
else if (varTypeIsStruct(argDsc->lvType))
{
for (unsigned fieldVarNum = argDsc->lvFieldLclStart;
fieldVarNum < argDsc->lvFieldLclStart + argDsc->lvFieldCnt; ++fieldVarNum)
{
LclVarDsc* fieldVarDsc = &(compiler->lvaTable[fieldVarNum]);
if (fieldVarDsc->lvLRACandidate)
{
assert(fieldVarDsc->lvTracked);
Interval* interval = getIntervalForLocalVar(fieldVarDsc->lvVarIndex);
RefPosition* pos =
newRefPosition(interval, MinLocation, RefTypeParamDef, nullptr, allRegs(TypeGet(fieldVarDsc)));
}
}
}
else
{
// We can overwrite the register (i.e. codegen saves it on entry)
assert(argDsc->lvRefCnt == 0 || !argDsc->lvIsRegArg || argDsc->lvDoNotEnregister ||
!argDsc->lvLRACandidate || (varTypeIsFloating(argDsc->TypeGet()) && compiler->opts.compDbgCode));
}
}
// Now set up the reg state for the non-tracked args
// (We do this here because we want to generate the ParamDef RefPositions in tracked
// order, so that loop doesn't hit the non-tracked args)
for (unsigned argNum = 0; argNum < compiler->info.compArgsCount; argNum++, argDsc++)
{
argDsc = &(compiler->lvaTable[argNum]);
if (argDsc->lvPromotedStruct())
{
noway_assert(argDsc->lvFieldCnt == 1); // We only handle one field here
unsigned fieldVarNum = argDsc->lvFieldLclStart;
argDsc = &(compiler->lvaTable[fieldVarNum]);
}
noway_assert(argDsc->lvIsParam);
if (!argDsc->lvTracked && argDsc->lvIsRegArg)
{
updateRegStateForArg(argDsc);
}
}
// If there is a secret stub param, it is also live in
if (compiler->info.compPublishStubParam)
{
intRegState->rsCalleeRegArgMaskLiveIn |= RBM_SECRET_STUB_PARAM;
}
LocationInfoListNodePool listNodePool(compiler, 8);
SmallHashTable<GenTree*, LocationInfoList, 32> operandToLocationInfoMap(compiler);
BasicBlock* predBlock = nullptr;
BasicBlock* prevBlock = nullptr;
// Initialize currentLiveVars to the empty set. We will set it to the current
// live-in at the entry to each block (this will include the incoming args on
// the first block).
VarSetOps::AssignNoCopy(compiler, currentLiveVars, VarSetOps::MakeEmpty(compiler));
for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock())
{
JITDUMP("\nNEW BLOCK BB%02u\n", block->bbNum);
bool predBlockIsAllocated = false;
predBlock = findPredBlockForLiveIn(block, prevBlock DEBUGARG(&predBlockIsAllocated));
if (predBlock)
{
JITDUMP("\n\nSetting BB%02u as the predecessor for determining incoming variable registers of BB%02u\n",
block->bbNum, predBlock->bbNum);
assert(predBlock->bbNum <= bbNumMaxBeforeResolution);
blockInfo[block->bbNum].predBBNum = predBlock->bbNum;
}
if (enregisterLocalVars)
{
VarSetOps::AssignNoCopy(compiler, currentLiveVars,
VarSetOps::Intersection(compiler, registerCandidateVars, block->bbLiveIn));
if (block == compiler->fgFirstBB)
{
insertZeroInitRefPositions();
// The first real location is at 1; 0 is for the entry.
currentLoc = 1;
}
// Any lclVars live-in to a block are resolution candidates.
VarSetOps::UnionD(compiler, resolutionCandidateVars, currentLiveVars);
// Determine if we need any DummyDefs.
// We need DummyDefs for cases where "predBlock" isn't really a predecessor.
// Note that it's possible to have uses of unitialized variables, in which case even the first
// block may require DummyDefs, which we are not currently adding - this means that these variables
// will always be considered to be in memory on entry (and reloaded when the use is encountered).
// TODO-CQ: Consider how best to tune this. Currently, if we create DummyDefs for uninitialized
// variables (which may actually be initialized along the dynamically executed paths, but not
// on all static paths), we wind up with excessive liveranges for some of these variables.
VARSET_TP newLiveIn(VarSetOps::MakeCopy(compiler, currentLiveVars));
if (predBlock)
{
// Compute set difference: newLiveIn = currentLiveVars - predBlock->bbLiveOut
VarSetOps::DiffD(compiler, newLiveIn, predBlock->bbLiveOut);
}
bool needsDummyDefs = (!VarSetOps::IsEmpty(compiler, newLiveIn) && block != compiler->fgFirstBB);
// Create dummy def RefPositions
if (needsDummyDefs)
{
// If we are using locations from a predecessor, we should never require DummyDefs.
assert(!predBlockIsAllocated);
JITDUMP("Creating dummy definitions\n");
VarSetOps::Iter iter(compiler, newLiveIn);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* varDsc = compiler->lvaTable + varNum;
// Add a dummyDef for any candidate vars that are in the "newLiveIn" set.
// If this is the entry block, don't add any incoming parameters (they're handled with ParamDefs).
if (isCandidateVar(varDsc) && (predBlock != nullptr || !varDsc->lvIsParam))
{
Interval* interval = getIntervalForLocalVar(varIndex);
RefPosition* pos = newRefPosition(interval, currentLoc, RefTypeDummyDef, nullptr,
allRegs(interval->registerType));
}
}
JITDUMP("Finished creating dummy definitions\n\n");
}
}
// Add a dummy RefPosition to mark the block boundary.
// Note that we do this AFTER adding the exposed uses above, because the
// register positions for those exposed uses need to be recorded at
// this point.
RefPosition* pos = newRefPosition((Interval*)nullptr, currentLoc, RefTypeBB, nullptr, RBM_NONE);
currentLoc += 2;
JITDUMP("\n");
LIR::Range& blockRange = LIR::AsRange(block);
for (GenTree* node : blockRange.NonPhiNodes())
{
// We increment the number position of each tree node by 2 to simplify the logic when there's the case of
// a tree that implicitly does a dual-definition of temps (the long case). In this case it is easier to
// already have an idle spot to handle a dual-def instead of making some messy adjustments if we only
// increment the number position by one.
CLANG_FORMAT_COMMENT_ANCHOR;
#ifdef DEBUG
node->gtSeqNum = currentLoc;
// In DEBUG, we want to set the gtRegTag to GT_REGTAG_REG, so that subsequent dumps will so the register
// value.
// Although this looks like a no-op it sets the tag.
node->gtRegNum = node->gtRegNum;
#endif
node->gtLsraInfo.Initialize(this, node, currentLoc);
TreeNodeInfoInit(node);
// If the node produces an unused value, mark it as a local def-use
if (node->IsValue() && node->IsUnusedValue())
{
node->gtLsraInfo.isLocalDefUse = true;
node->gtLsraInfo.dstCount = 0;
}
#ifdef DEBUG
if (VERBOSE)
{
compiler->gtDispTree(node, nullptr, nullptr, true);
printf(" +");
node->gtLsraInfo.dump(this);
}
#endif // DEBUG
// Only nodes that produce values should have a non-zero dstCount.
assert((node->gtLsraInfo.dstCount == 0) || node->IsValue());
buildRefPositionsForNode(node, block, listNodePool, operandToLocationInfoMap, currentLoc);
#ifdef DEBUG
if (currentLoc > maxNodeLocation)
{
maxNodeLocation = currentLoc;
}
#endif // DEBUG
currentLoc += 2;
}
// Note: the visited set is cleared in LinearScan::doLinearScan()
markBlockVisited(block);
if (enregisterLocalVars)
{
// Insert exposed uses for a lclVar that is live-out of 'block' but not live-in to the
// next block, or any unvisited successors.
// This will address lclVars that are live on a backedge, as well as those that are kept
// live at a GT_JMP.
//
// Blocks ending with "jmp method" are marked as BBJ_HAS_JMP,
// and jmp call is represented using GT_JMP node which is a leaf node.
// Liveness phase keeps all the arguments of the method live till the end of
// block by adding them to liveout set of the block containing GT_JMP.
//
// The target of a GT_JMP implicitly uses all the current method arguments, however
// there are no actual references to them. This can cause LSRA to assert, because
// the variables are live but it sees no references. In order to correctly model the
// liveness of these arguments, we add dummy exposed uses, in the same manner as for
// backward branches. This will happen automatically via expUseSet.
//
// Note that a block ending with GT_JMP has no successors and hence the variables
// for which dummy use ref positions are added are arguments of the method.
VARSET_TP expUseSet(VarSetOps::MakeCopy(compiler, block->bbLiveOut));
VarSetOps::IntersectionD(compiler, expUseSet, registerCandidateVars);
BasicBlock* nextBlock = getNextBlock();
if (nextBlock != nullptr)
{
VarSetOps::DiffD(compiler, expUseSet, nextBlock->bbLiveIn);
}
for (BasicBlock* succ : block->GetAllSuccs(compiler))
{
if (VarSetOps::IsEmpty(compiler, expUseSet))
{
break;
}
if (isBlockVisited(succ))
{
continue;
}
VarSetOps::DiffD(compiler, expUseSet, succ->bbLiveIn);
}
if (!VarSetOps::IsEmpty(compiler, expUseSet))
{
JITDUMP("Exposed uses:");
VarSetOps::Iter iter(compiler, expUseSet);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* varDsc = compiler->lvaTable + varNum;
assert(isCandidateVar(varDsc));
Interval* interval = getIntervalForLocalVar(varIndex);
RefPosition* pos =
newRefPosition(interval, currentLoc, RefTypeExpUse, nullptr, allRegs(interval->registerType));
JITDUMP(" V%02u", varNum);
}
JITDUMP("\n");
}
// Clear the "last use" flag on any vars that are live-out from this block.
{
VarSetOps::Iter iter(compiler, block->bbLiveOut);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
LclVarDsc* const varDsc = &compiler->lvaTable[varNum];
if (isCandidateVar(varDsc))
{
RefPosition* const lastRP = getIntervalForLocalVar(varIndex)->lastRefPosition;
if ((lastRP != nullptr) && (lastRP->bbNum == block->bbNum))
{
lastRP->lastUse = false;
}
}
}
}
#ifdef DEBUG
checkLastUses(block);
if (VERBOSE)
{
printf("use: ");
dumpConvertedVarSet(compiler, block->bbVarUse);
printf("\ndef: ");
dumpConvertedVarSet(compiler, block->bbVarDef);
printf("\n");
}
#endif // DEBUG
}
prevBlock = block;
}
if (enregisterLocalVars)
{
if (compiler->lvaKeepAliveAndReportThis())
{
// If we need to KeepAliveAndReportThis, add a dummy exposed use of it at the end
unsigned keepAliveVarNum = compiler->info.compThisArg;
assert(compiler->info.compIsStatic == false);
LclVarDsc* varDsc = compiler->lvaTable + keepAliveVarNum;
if (isCandidateVar(varDsc))
{
JITDUMP("Adding exposed use of this, for lvaKeepAliveAndReportThis\n");
Interval* interval = getIntervalForLocalVar(varDsc->lvVarIndex);
RefPosition* pos =
newRefPosition(interval, currentLoc, RefTypeExpUse, nullptr, allRegs(interval->registerType));
}
}
#ifdef DEBUG
if (getLsraExtendLifeTimes())
{
LclVarDsc* varDsc;
for (lclNum = 0, varDsc = compiler->lvaTable; lclNum < compiler->lvaCount; lclNum++, varDsc++)
{
if (varDsc->lvLRACandidate)
{
JITDUMP("Adding exposed use of V%02u for LsraExtendLifetimes\n", lclNum);
Interval* interval = getIntervalForLocalVar(varDsc->lvVarIndex);
RefPosition* pos =
newRefPosition(interval, currentLoc, RefTypeExpUse, nullptr, allRegs(interval->registerType));
}
}
}
#endif // DEBUG
}
// If the last block has successors, create a RefTypeBB to record
// what's live
if (prevBlock->NumSucc(compiler) > 0)
{
RefPosition* pos = newRefPosition((Interval*)nullptr, currentLoc, RefTypeBB, nullptr, RBM_NONE);
}
#ifdef DEBUG
// Make sure we don't have any blocks that were not visited
foreach_block(compiler, block)
{
assert(isBlockVisited(block));
}
if (VERBOSE)
{
lsraDumpIntervals("BEFORE VALIDATING INTERVALS");
dumpRefPositions("BEFORE VALIDATING INTERVALS");
validateIntervals();
}
#endif // DEBUG
}
#ifdef DEBUG
void LinearScan::dumpVarRefPositions(const char* title)
{
if (enregisterLocalVars)
{
printf("\nVAR REFPOSITIONS %s\n", title);
for (unsigned i = 0; i < compiler->lvaCount; i++)
{
printf("--- V%02u\n", i);
LclVarDsc* varDsc = compiler->lvaTable + i;
if (varDsc->lvIsRegCandidate())
{
Interval* interval = getIntervalForLocalVar(varDsc->lvVarIndex);
for (RefPosition* ref = interval->firstRefPosition; ref != nullptr; ref = ref->nextRefPosition)
{
ref->dump();
}
}
}
printf("\n");
}
}
void LinearScan::validateIntervals()
{
if (enregisterLocalVars)
{
for (unsigned i = 0; i < compiler->lvaTrackedCount; i++)
{
if (!compiler->lvaTable[compiler->lvaTrackedToVarNum[i]].lvLRACandidate)
{
continue;
}
Interval* interval = getIntervalForLocalVar(i);
bool defined = false;
printf("-----------------\n");
for (RefPosition* ref = interval->firstRefPosition; ref != nullptr; ref = ref->nextRefPosition)
{
ref->dump();
RefType refType = ref->refType;
if (!defined && RefTypeIsUse(refType))
{
if (compiler->info.compMethodName != nullptr)
{
printf("%s: ", compiler->info.compMethodName);
}
printf("LocalVar V%02u: undefined use at %u\n", interval->varNum, ref->nodeLocation);
}
// Note that there can be multiple last uses if they are on disjoint paths,
// so we can't really check the lastUse flag
if (ref->lastUse)
{
defined = false;
}
if (RefTypeIsDef(refType))
{
defined = true;
}
}
}
}
}
#endif // DEBUG
// Set the default rpFrameType based upon codeGen->isFramePointerRequired()
// This was lifted from the register predictor
//
void LinearScan::setFrameType()
{
FrameType frameType = FT_NOT_SET;
#if DOUBLE_ALIGN
compiler->codeGen->setDoubleAlign(false);
if (doDoubleAlign)
{
frameType = FT_DOUBLE_ALIGN_FRAME;
compiler->codeGen->setDoubleAlign(true);
}
else
#endif // DOUBLE_ALIGN
if (compiler->codeGen->isFramePointerRequired())
{
frameType = FT_EBP_FRAME;
}
else
{
if (compiler->rpMustCreateEBPCalled == false)
{
#ifdef DEBUG
const char* reason;
#endif // DEBUG
compiler->rpMustCreateEBPCalled = true;
if (compiler->rpMustCreateEBPFrame(INDEBUG(&reason)))
{
JITDUMP("; Decided to create an EBP based frame for ETW stackwalking (%s)\n", reason);
compiler->codeGen->setFrameRequired(true);
}
}
if (compiler->codeGen->isFrameRequired())
{
frameType = FT_EBP_FRAME;
}
else
{
frameType = FT_ESP_FRAME;
}
}
switch (frameType)
{
case FT_ESP_FRAME:
noway_assert(!compiler->codeGen->isFramePointerRequired());
noway_assert(!compiler->codeGen->isFrameRequired());
compiler->codeGen->setFramePointerUsed(false);
break;
case FT_EBP_FRAME:
compiler->codeGen->setFramePointerUsed(true);
break;
#if DOUBLE_ALIGN
case FT_DOUBLE_ALIGN_FRAME:
noway_assert(!compiler->codeGen->isFramePointerRequired());
compiler->codeGen->setFramePointerUsed(false);
break;
#endif // DOUBLE_ALIGN
default:
noway_assert(!"rpFrameType not set correctly!");
break;
}
// If we are using FPBASE as the frame register, we cannot also use it for
// a local var. Note that we may have already added it to the register masks,
// which are computed when the LinearScan class constructor is created, and
// used during lowering. Luckily, the TreeNodeInfo only stores an index to
// the masks stored in the LinearScan class, so we only need to walk the
// unique masks and remove FPBASE.
if (frameType == FT_EBP_FRAME)
{
if ((availableIntRegs & RBM_FPBASE) != 0)
{
RemoveRegisterFromMasks(REG_FPBASE);
// We know that we're already in "read mode" for availableIntRegs. However,
// we need to remove the FPBASE register, so subsequent users (like callers
// to allRegs()) get the right thing. The RemoveRegisterFromMasks() code
// fixes up everything that already took a dependency on the value that was
// previously read, so this completes the picture.
availableIntRegs.OverrideAssign(availableIntRegs & ~RBM_FPBASE);
}
}
compiler->rpFrameType = frameType;
}
// Is the copyReg/moveReg given by this RefPosition still busy at the
// given location?
bool copyOrMoveRegInUse(RefPosition* ref, LsraLocation loc)
{
assert(ref->copyReg || ref->moveReg);
if (ref->getRefEndLocation() >= loc)
{
return true;
}
Interval* interval = ref->getInterval();
RefPosition* nextRef = interval->getNextRefPosition();
if (nextRef != nullptr && nextRef->treeNode == ref->treeNode && nextRef->getRefEndLocation() >= loc)
{
return true;
}
return false;
}
// Determine whether the register represented by "physRegRecord" is available at least
// at the "currentLoc", and if so, return the next location at which it is in use in
// "nextRefLocationPtr"
//
bool LinearScan::registerIsAvailable(RegRecord* physRegRecord,
LsraLocation currentLoc,
LsraLocation* nextRefLocationPtr,
RegisterType regType)
{
*nextRefLocationPtr = MaxLocation;
LsraLocation nextRefLocation = MaxLocation;
regMaskTP regMask = genRegMask(physRegRecord->regNum);
if (physRegRecord->isBusyUntilNextKill)
{
return false;
}
RefPosition* nextPhysReference = physRegRecord->getNextRefPosition();
if (nextPhysReference != nullptr)
{
nextRefLocation = nextPhysReference->nodeLocation;
// if (nextPhysReference->refType == RefTypeFixedReg) nextRefLocation--;
}
else if (!physRegRecord->isCalleeSave)
{
nextRefLocation = MaxLocation - 1;
}
Interval* assignedInterval = physRegRecord->assignedInterval;
if (assignedInterval != nullptr)
{
RefPosition* recentReference = assignedInterval->recentRefPosition;
// The only case where we have an assignedInterval, but recentReference is null
// is where this interval is live at procedure entry (i.e. an arg register), in which
// case it's still live and its assigned register is not available
// (Note that the ParamDef will be recorded as a recentReference when we encounter
// it, but we will be allocating registers, potentially to other incoming parameters,
// as we process the ParamDefs.)
if (recentReference == nullptr)
{
return false;
}
// Is this a copyReg/moveReg? It is if the register assignment doesn't match.
// (the recentReference may not be a copyReg/moveReg, because we could have seen another
// reference since the copyReg/moveReg)
if (!assignedInterval->isAssignedTo(physRegRecord->regNum))
{
// Don't reassign it if it's still in use
if ((recentReference->copyReg || recentReference->moveReg) &&
copyOrMoveRegInUse(recentReference, currentLoc))
{
return false;
}
}
else if (!assignedInterval->isActive && assignedInterval->isConstant)
{
// Treat this as unassigned, i.e. do nothing.
// TODO-CQ: Consider adjusting the heuristics (probably in the caller of this method)
// to avoid reusing these registers.
}
// If this interval isn't active, it's available if it isn't referenced
// at this location (or the previous location, if the recent RefPosition
// is a delayRegFree).
else if (!assignedInterval->isActive &&
(recentReference->refType == RefTypeExpUse || recentReference->getRefEndLocation() < currentLoc))
{
// This interval must have a next reference (otherwise it wouldn't be assigned to this register)
RefPosition* nextReference = recentReference->nextRefPosition;
if (nextReference != nullptr)
{
if (nextReference->nodeLocation < nextRefLocation)
{
nextRefLocation = nextReference->nodeLocation;
}
}
else
{
assert(recentReference->copyReg && recentReference->registerAssignment != regMask);
}
}
else
{
return false;
}
}
if (nextRefLocation < *nextRefLocationPtr)
{
*nextRefLocationPtr = nextRefLocation;
}
#ifdef _TARGET_ARM_
if (regType == TYP_DOUBLE)
{
// Recurse, but check the other half this time (TYP_FLOAT)
if (!registerIsAvailable(findAnotherHalfRegRec(physRegRecord), currentLoc, nextRefLocationPtr, TYP_FLOAT))
return false;
nextRefLocation = *nextRefLocationPtr;
}
#endif // _TARGET_ARM_
return (nextRefLocation >= currentLoc);
}
//------------------------------------------------------------------------
// getRegisterType: Get the RegisterType to use for the given RefPosition
//
// Arguments:
// currentInterval: The interval for the current allocation
// refPosition: The RefPosition of the current Interval for which a register is being allocated
//
// Return Value:
// The RegisterType that should be allocated for this RefPosition
//
// Notes:
// This will nearly always be identical to the registerType of the interval, except in the case
// of SIMD types of 8 bytes (currently only Vector2) when they are passed and returned in integer
// registers, or copied to a return temp.
// This method need only be called in situations where we may be dealing with the register requirements
// of a RefTypeUse RefPosition (i.e. not when we are only looking at the type of an interval, nor when
// we are interested in the "defining" type of the interval). This is because the situation of interest
// only happens at the use (where it must be copied to an integer register).
RegisterType LinearScan::getRegisterType(Interval* currentInterval, RefPosition* refPosition)
{
assert(refPosition->getInterval() == currentInterval);
RegisterType regType = currentInterval->registerType;
regMaskTP candidates = refPosition->registerAssignment;
assert((candidates & allRegs(regType)) != RBM_NONE);
return regType;
}
//------------------------------------------------------------------------
// tryAllocateFreeReg: Find a free register that satisfies the requirements for refPosition,
// and takes into account the preferences for the given Interval
//
// Arguments:
// currentInterval: The interval for the current allocation
// refPosition: The RefPosition of the current Interval for which a register is being allocated
//
// Return Value:
// The regNumber, if any, allocated to the RefPositon. Returns REG_NA if no free register is found.
//
// Notes:
// TODO-CQ: Consider whether we need to use a different order for tree temps than for vars, as
// reg predict does
static const regNumber lsraRegOrder[] = {REG_VAR_ORDER};
const unsigned lsraRegOrderSize = ArrLen(lsraRegOrder);
static const regNumber lsraRegOrderFlt[] = {REG_VAR_ORDER_FLT};
const unsigned lsraRegOrderFltSize = ArrLen(lsraRegOrderFlt);
regNumber LinearScan::tryAllocateFreeReg(Interval* currentInterval, RefPosition* refPosition)
{
regNumber foundReg = REG_NA;
RegisterType regType = getRegisterType(currentInterval, refPosition);
const regNumber* regOrder;
unsigned regOrderSize;
if (useFloatReg(regType))
{
regOrder = lsraRegOrderFlt;
regOrderSize = lsraRegOrderFltSize;
}
else
{
regOrder = lsraRegOrder;
regOrderSize = lsraRegOrderSize;
}
LsraLocation currentLocation = refPosition->nodeLocation;
RefPosition* nextRefPos = refPosition->nextRefPosition;
LsraLocation nextLocation = (nextRefPos == nullptr) ? currentLocation : nextRefPos->nodeLocation;
regMaskTP candidates = refPosition->registerAssignment;
regMaskTP preferences = currentInterval->registerPreferences;
if (RefTypeIsDef(refPosition->refType))
{
if (currentInterval->hasConflictingDefUse)
{
resolveConflictingDefAndUse(currentInterval, refPosition);
candidates = refPosition->registerAssignment;
}
// Otherwise, check for the case of a fixed-reg def of a reg that will be killed before the
// use, or interferes at the point of use (which shouldn't happen, but Lower doesn't mark
// the contained nodes as interfering).
// Note that we may have a ParamDef RefPosition that is marked isFixedRegRef, but which
// has had its registerAssignment changed to no longer be a single register.
else if (refPosition->isFixedRegRef && nextRefPos != nullptr && RefTypeIsUse(nextRefPos->refType) &&
!nextRefPos->isFixedRegRef && genMaxOneBit(refPosition->registerAssignment))
{
regNumber defReg = refPosition->assignedReg();
RegRecord* defRegRecord = getRegisterRecord(defReg);
RefPosition* currFixedRegRefPosition = defRegRecord->recentRefPosition;
assert(currFixedRegRefPosition != nullptr &&
currFixedRegRefPosition->nodeLocation == refPosition->nodeLocation);
// If there is another fixed reference to this register before the use, change the candidates
// on this RefPosition to include that of nextRefPos.
if (currFixedRegRefPosition->nextRefPosition != nullptr &&
currFixedRegRefPosition->nextRefPosition->nodeLocation <= nextRefPos->getRefEndLocation())
{
candidates |= nextRefPos->registerAssignment;
if (preferences == refPosition->registerAssignment)
{
preferences = candidates;
}
}
}
}
preferences &= candidates;
if (preferences == RBM_NONE)
{
preferences = candidates;
}
regMaskTP relatedPreferences = RBM_NONE;
#ifdef DEBUG
candidates = stressLimitRegs(refPosition, candidates);
#endif
bool mustAssignARegister = true;
assert(candidates != RBM_NONE);
// If the related interval has no further references, it is possible that it is a source of the
// node that produces this interval. However, we don't want to use the relatedInterval for preferencing
// if its next reference is not a new definition (as it either is or will become live).
Interval* relatedInterval = currentInterval->relatedInterval;
if (relatedInterval != nullptr)
{
RefPosition* nextRelatedRefPosition = relatedInterval->getNextRefPosition();
if (nextRelatedRefPosition != nullptr)
{
// Don't use the relatedInterval for preferencing if its next reference is not a new definition.
if (!RefTypeIsDef(nextRelatedRefPosition->refType))
{
relatedInterval = nullptr;
}
// Is the relatedInterval simply a copy to another relatedInterval?
else if ((relatedInterval->relatedInterval != nullptr) &&
(nextRelatedRefPosition->nextRefPosition != nullptr) &&
(nextRelatedRefPosition->nextRefPosition->nextRefPosition == nullptr) &&
(nextRelatedRefPosition->nextRefPosition->nodeLocation <
relatedInterval->relatedInterval->getNextRefLocation()))
{
// The current relatedInterval has only two remaining RefPositions, both of which
// occur prior to the next RefPosition for its relatedInterval.
// It is likely a copy.
relatedInterval = relatedInterval->relatedInterval;
}
}
}
if (relatedInterval != nullptr)
{
// If the related interval already has an assigned register, then use that
// as the related preference. We'll take the related
// interval preferences into account in the loop over all the registers.
if (relatedInterval->assignedReg != nullptr)
{
relatedPreferences = genRegMask(relatedInterval->assignedReg->regNum);
}
else
{
relatedPreferences = relatedInterval->registerPreferences;
}
}
bool preferCalleeSave = currentInterval->preferCalleeSave;
// For floating point, we want to be less aggressive about using callee-save registers.
// So in that case, we just need to ensure that the current RefPosition is covered.
RefPosition* rangeEndRefPosition;
RefPosition* lastRefPosition = currentInterval->lastRefPosition;
if (useFloatReg(currentInterval->registerType))
{
rangeEndRefPosition = refPosition;
}
else
{
rangeEndRefPosition = currentInterval->lastRefPosition;
// If we have a relatedInterval that is not currently occupying a register,
// and whose lifetime begins after this one ends,
// we want to try to select a register that will cover its lifetime.
if ((relatedInterval != nullptr) && (relatedInterval->assignedReg == nullptr) &&
(relatedInterval->getNextRefLocation() >= rangeEndRefPosition->nodeLocation))
{
lastRefPosition = relatedInterval->lastRefPosition;
preferCalleeSave = relatedInterval->preferCalleeSave;
}
}
// If this has a delayed use (due to being used in a rmw position of a
// non-commutative operator), its endLocation is delayed until the "def"
// position, which is one location past the use (getRefEndLocation() takes care of this).
LsraLocation rangeEndLocation = rangeEndRefPosition->getRefEndLocation();
LsraLocation lastLocation = lastRefPosition->getRefEndLocation();
regNumber prevReg = REG_NA;
if (currentInterval->assignedReg)
{
bool useAssignedReg = false;
// This was an interval that was previously allocated to the given
// physical register, and we should try to allocate it to that register
// again, if possible and reasonable.
// Use it preemptively (i.e. before checking other available regs)
// only if it is preferred and available.
RegRecord* regRec = currentInterval->assignedReg;
prevReg = regRec->regNum;
regMaskTP prevRegBit = genRegMask(prevReg);
// Is it in the preferred set of regs?
if ((prevRegBit & preferences) != RBM_NONE)
{
// Is it currently available?
LsraLocation nextPhysRefLoc;
if (registerIsAvailable(regRec, currentLocation, &nextPhysRefLoc, currentInterval->registerType))
{
// If the register is next referenced at this location, only use it if
// this has a fixed reg requirement (i.e. this is the reference that caused
// the FixedReg ref to be created)
if (!regRec->conflictingFixedRegReference(refPosition))
{
useAssignedReg = true;
}
}
}
if (useAssignedReg)
{
regNumber foundReg = prevReg;
assignPhysReg(regRec, currentInterval);
refPosition->registerAssignment = genRegMask(foundReg);
return foundReg;
}
else
{
// Don't keep trying to allocate to this register
currentInterval->assignedReg = nullptr;
}
}
RegRecord* availablePhysRegInterval = nullptr;
bool unassignInterval = false;
// Each register will receive a score which is the sum of the scoring criteria below.
// These were selected on the assumption that they will have an impact on the "goodness"
// of a register selection, and have been tuned to a certain extent by observing the impact
// of the ordering on asmDiffs. However, there is probably much more room for tuning,
// and perhaps additional criteria.
//
// These are FLAGS (bits) so that we can easily order them and add them together.
// If the scores are equal, but one covers more of the current interval's range,
// then it wins. Otherwise, the one encountered earlier in the regOrder wins.
enum RegisterScore
{
VALUE_AVAILABLE = 0x40, // It is a constant value that is already in an acceptable register.
COVERS = 0x20, // It is in the interval's preference set and it covers the entire lifetime.
OWN_PREFERENCE = 0x10, // It is in the preference set of this interval.
COVERS_RELATED = 0x08, // It is in the preference set of the related interval and covers the entire lifetime.
RELATED_PREFERENCE = 0x04, // It is in the preference set of the related interval.
CALLER_CALLEE = 0x02, // It is in the right "set" for the interval (caller or callee-save).
UNASSIGNED = 0x01, // It is not currently assigned to an inactive interval.
};
int bestScore = 0;
// Compute the best possible score so we can stop looping early if we find it.
// TODO-Throughput: At some point we may want to short-circuit the computation of each score, but
// probably not until we've tuned the order of these criteria. At that point,
// we'll need to avoid the short-circuit if we've got a stress option to reverse
// the selection.
int bestPossibleScore = COVERS + UNASSIGNED + OWN_PREFERENCE + CALLER_CALLEE;
if (relatedPreferences != RBM_NONE)
{
bestPossibleScore |= RELATED_PREFERENCE + COVERS_RELATED;
}
LsraLocation bestLocation = MinLocation;
// In non-debug builds, this will simply get optimized away
bool reverseSelect = false;
#ifdef DEBUG
reverseSelect = doReverseSelect();
#endif // DEBUG
// An optimization for the common case where there is only one candidate -
// avoid looping over all the other registers
regNumber singleReg = REG_NA;
if (genMaxOneBit(candidates))
{
regOrderSize = 1;
singleReg = genRegNumFromMask(candidates);
regOrder = &singleReg;
}
for (unsigned i = 0; i < regOrderSize && (candidates != RBM_NONE); i++)
{
regNumber regNum = regOrder[i];
regMaskTP candidateBit = genRegMask(regNum);
if (!(candidates & candidateBit))
{
continue;
}
candidates &= ~candidateBit;
RegRecord* physRegRecord = getRegisterRecord(regNum);
int score = 0;
LsraLocation nextPhysRefLocation = MaxLocation;
// By chance, is this register already holding this interval, as a copyReg or having
// been restored as inactive after a kill?
if (physRegRecord->assignedInterval == currentInterval)
{
availablePhysRegInterval = physRegRecord;
unassignInterval = false;
break;
}
// Find the next RefPosition of the physical register
if (!registerIsAvailable(physRegRecord, currentLocation, &nextPhysRefLocation, regType))
{
continue;
}
// If the register is next referenced at this location, only use it if
// this has a fixed reg requirement (i.e. this is the reference that caused
// the FixedReg ref to be created)
if (physRegRecord->conflictingFixedRegReference(refPosition))
{
continue;
}
// If this is a definition of a constant interval, check to see if its value is already in this register.
if (currentInterval->isConstant && RefTypeIsDef(refPosition->refType) &&
(physRegRecord->assignedInterval != nullptr) && physRegRecord->assignedInterval->isConstant)
{
noway_assert(refPosition->treeNode != nullptr);
GenTree* otherTreeNode = physRegRecord->assignedInterval->firstRefPosition->treeNode;
noway_assert(otherTreeNode != nullptr);
if (refPosition->treeNode->OperGet() == otherTreeNode->OperGet())
{
switch (otherTreeNode->OperGet())
{
case GT_CNS_INT:
if ((refPosition->treeNode->AsIntCon()->IconValue() ==
otherTreeNode->AsIntCon()->IconValue()) &&
(varTypeGCtype(refPosition->treeNode) == varTypeGCtype(otherTreeNode)))
{
#ifdef _TARGET_64BIT_
// If the constant is negative, only reuse registers of the same type.
// This is because, on a 64-bit system, we do not sign-extend immediates in registers to
// 64-bits unless they are actually longs, as this requires a longer instruction.
// This doesn't apply to a 32-bit system, on which long values occupy multiple registers.
// (We could sign-extend, but we would have to always sign-extend, because if we reuse more
// than once, we won't have access to the instruction that originally defines the constant).
if ((refPosition->treeNode->TypeGet() == otherTreeNode->TypeGet()) ||
(refPosition->treeNode->AsIntCon()->IconValue() >= 0))
#endif // _TARGET_64BIT_
{
score |= VALUE_AVAILABLE;
}
}
break;
case GT_CNS_DBL:
{
// For floating point constants, the values must be identical, not simply compare
// equal. So we compare the bits.
if (refPosition->treeNode->AsDblCon()->isBitwiseEqual(otherTreeNode->AsDblCon()) &&
(refPosition->treeNode->TypeGet() == otherTreeNode->TypeGet()))
{
score |= VALUE_AVAILABLE;
}
break;
}
default:
// for all other 'otherTreeNode->OperGet()' kinds, we leave 'score' unchanged
break;
}
}
}
// If the nextPhysRefLocation is a fixedRef for the rangeEndRefPosition, increment it so that
// we don't think it isn't covering the live range.
// This doesn't handle the case where earlier RefPositions for this Interval are also
// FixedRefs of this regNum, but at least those are only interesting in the case where those
// are "local last uses" of the Interval - otherwise the liveRange would interfere with the reg.
if (nextPhysRefLocation == rangeEndLocation && rangeEndRefPosition->isFixedRefOfReg(regNum))
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_INCREMENT_RANGE_END, currentInterval, regNum));
nextPhysRefLocation++;
}
if ((candidateBit & preferences) != RBM_NONE)
{
score |= OWN_PREFERENCE;
if (nextPhysRefLocation > rangeEndLocation)
{
score |= COVERS;
}
}
if (relatedInterval != nullptr && (candidateBit & relatedPreferences) != RBM_NONE)
{
score |= RELATED_PREFERENCE;
if (nextPhysRefLocation > relatedInterval->lastRefPosition->nodeLocation)
{
score |= COVERS_RELATED;
}
}
// If we had a fixed-reg def of a reg that will be killed before the use, prefer it to any other registers
// with the same score. (Note that we haven't changed the original registerAssignment on the RefPosition).
// Overload the RELATED_PREFERENCE value.
else if (candidateBit == refPosition->registerAssignment)
{
score |= RELATED_PREFERENCE;
}
if ((preferCalleeSave && physRegRecord->isCalleeSave) || (!preferCalleeSave && !physRegRecord->isCalleeSave))
{
score |= CALLER_CALLEE;
}
// The register is considered unassigned if it has no assignedInterval, OR
// if its next reference is beyond the range of this interval.
if (!isAssigned(physRegRecord, lastLocation ARM_ARG(currentInterval->registerType)))
{
score |= UNASSIGNED;
}
bool foundBetterCandidate = false;
if (score > bestScore)
{
foundBetterCandidate = true;
}
else if (score == bestScore)
{
// Prefer a register that covers the range.
if (bestLocation <= lastLocation)
{
if (nextPhysRefLocation > bestLocation)
{
foundBetterCandidate = true;
}
}
// If both cover the range, prefer a register that is killed sooner (leaving the longer range register
// available). If both cover the range and also getting killed at the same location, prefer the one which
// is same as previous assignment.
else if (nextPhysRefLocation > lastLocation)
{
if (nextPhysRefLocation < bestLocation)
{
foundBetterCandidate = true;
}
else if (nextPhysRefLocation == bestLocation && prevReg == regNum)
{
foundBetterCandidate = true;
}
}
}
#ifdef DEBUG
if (doReverseSelect() && bestScore != 0)
{
foundBetterCandidate = !foundBetterCandidate;
}
#endif // DEBUG
if (foundBetterCandidate)
{
bestLocation = nextPhysRefLocation;
availablePhysRegInterval = physRegRecord;
unassignInterval = true;
bestScore = score;
}
// there is no way we can get a better score so break out
if (!reverseSelect && score == bestPossibleScore && bestLocation == rangeEndLocation + 1)
{
break;
}
}
if (availablePhysRegInterval != nullptr)
{
if (unassignInterval && isAssigned(availablePhysRegInterval ARM_ARG(currentInterval->registerType)))
{
Interval* const intervalToUnassign = availablePhysRegInterval->assignedInterval;
unassignPhysReg(availablePhysRegInterval ARM_ARG(currentInterval->registerType));
if ((bestScore & VALUE_AVAILABLE) != 0 && intervalToUnassign != nullptr)
{
assert(intervalToUnassign->isConstant);
refPosition->treeNode->SetReuseRegVal();
}
// If we considered this "unassigned" because this interval's lifetime ends before
// the next ref, remember it.
else if ((bestScore & UNASSIGNED) != 0 && intervalToUnassign != nullptr)
{
updatePreviousInterval(availablePhysRegInterval, intervalToUnassign, intervalToUnassign->registerType);
}
}
else
{
assert((bestScore & VALUE_AVAILABLE) == 0);
}
assignPhysReg(availablePhysRegInterval, currentInterval);
foundReg = availablePhysRegInterval->regNum;
regMaskTP foundRegMask = genRegMask(foundReg);
refPosition->registerAssignment = foundRegMask;
if (relatedInterval != nullptr)
{
relatedInterval->updateRegisterPreferences(foundRegMask);
}
}
return foundReg;
}
//------------------------------------------------------------------------
// canSpillReg: Determine whether we can spill physRegRecord
//
// Arguments:
// physRegRecord - reg to spill
// refLocation - Location of RefPosition where this register will be spilled
// recentAssignedRefWeight - Weight of recent assigned RefPosition which will be determined in this function
// farthestRefPosWeight - Current farthestRefPosWeight at allocateBusyReg()
//
// Return Value:
// True - if we can spill physRegRecord
// False - otherwise
//
// Note: This helper is designed to be used only from allocateBusyReg() and canSpillDoubleReg()
//
bool LinearScan::canSpillReg(RegRecord* physRegRecord,
LsraLocation refLocation,
unsigned* recentAssignedRefWeight,
unsigned farthestRefPosWeight)
{
assert(physRegRecord->assignedInterval != nullptr);
RefPosition* recentAssignedRef = physRegRecord->assignedInterval->recentRefPosition;
if (recentAssignedRef != nullptr)
{
if (recentAssignedRef->nodeLocation == refLocation)
{
// We can't spill a register that's being used at the current location
return false;
}
// If the current position has the candidate register marked to be delayed,
// check if the previous location is using this register, if that's the case we have to skip
// since we can't spill this register.
if (recentAssignedRef->delayRegFree && (refLocation == recentAssignedRef->nodeLocation + 1))
{
return false;
}
// We don't prefer to spill a register if the weight of recentAssignedRef > weight
// of the spill candidate found so far. We would consider spilling a greater weight
// ref position only if the refPosition being allocated must need a reg.
*recentAssignedRefWeight = getWeight(recentAssignedRef);
if (*recentAssignedRefWeight > farthestRefPosWeight)
{
return false;
}
}
return true;
}
#ifdef _TARGET_ARM_
bool LinearScan::canSpillDoubleReg(RegRecord* physRegRecord,
LsraLocation refLocation,
unsigned* recentAssignedRefWeight,
unsigned farthestRefPosWeight)
{
bool retVal = true;
unsigned weight = BB_ZERO_WEIGHT;
unsigned weight2 = BB_ZERO_WEIGHT;
RegRecord* physRegRecord2 = findAnotherHalfRegRec(physRegRecord);
if (physRegRecord->assignedInterval != nullptr)
retVal &= canSpillReg(physRegRecord, refLocation, &weight, farthestRefPosWeight);
if (physRegRecord2->assignedInterval != nullptr)
retVal &= canSpillReg(physRegRecord2, refLocation, &weight2, farthestRefPosWeight);
if (!(weight == BB_ZERO_WEIGHT && weight2 == BB_ZERO_WEIGHT))
{
// weight and/or weight2 have been updated.
*recentAssignedRefWeight = (weight > weight2) ? weight : weight2;
}
return retVal;
}
#endif
//----------------------------------------------------------------------------
// checkActiveInterval: Test activness of an interval
// and check assertions if the interval is not active
//
// Arguments:
// interval - An interval to be tested
// refLocation - Location where the interval is being tested
//
// Return Value:
// True - iff the interval is active
// False - otherwise
//
// Note: This helper is designed to be used only from checkActiveIntervals()
//
bool LinearScan::checkActiveInterval(Interval* interval, LsraLocation refLocation)
{
if (!interval->isActive)
{
RefPosition* recentAssignedRef = interval->recentRefPosition;
// Note that we may or may not have actually handled the reference yet, so it could either
// be recentAssignedRef, or the next reference.
assert(recentAssignedRef != nullptr);
if (recentAssignedRef->nodeLocation != refLocation)
{
if (recentAssignedRef->nodeLocation + 1 == refLocation)
{
assert(recentAssignedRef->delayRegFree);
}
else
{
RefPosition* nextAssignedRef = recentAssignedRef->nextRefPosition;
#ifdef _TARGET_ARM_
// This function is invoked only from allocateBusyReg() through checkActiveIntervals().
//
// For ARM32, when we try to allocate a double register which consists of two float registers,
// tryAllocateFreeReg() may fail to allocate a double register even if one of two float
// reigsters is assigned to a inactive constant interval.
// https://github.com/dotnet/coreclr/pull/14080
//
// Therefore an inactive constant interval may be encountered here.
assert(nextAssignedRef != nullptr || interval->isConstant);
if (nextAssignedRef != nullptr)
assert(nextAssignedRef->nodeLocation == refLocation ||
(nextAssignedRef->nodeLocation + 1 == refLocation && nextAssignedRef->delayRegFree));
#else // !_TARGET_ARM_
assert(nextAssignedRef != nullptr);
assert(nextAssignedRef->nodeLocation == refLocation ||
(nextAssignedRef->nodeLocation + 1 == refLocation && nextAssignedRef->delayRegFree));
#endif // !_TARGET_ARM_
}
}
return false;
}
return true;
}
//----------------------------------------------------------------------------------------
// checkActiveIntervals: Test activness of a interval assinged to a register
// and check assertions if the interval is not active.
// We look into all intervals of two float registers consisting
// a double regsiter for ARM32.
//
// Arguments:
// physRegRecord - A register
// refLocation - Location where intervsl is being tested
// registerType - Type of regsiter
//
// Return Value:
// True - iff the all intervals are active
// False - otherwise
//
// Note: This helper is designed to be used only from allocateBusyReg()
//
bool LinearScan::checkActiveIntervals(RegRecord* physRegRecord, LsraLocation refLocation, RegisterType registerType)
{
Interval* assignedInterval = physRegRecord->assignedInterval;
#ifdef _TARGET_ARM_
// Check two intervals for a double register in ARM32
Interval* assignedInterval2 = nullptr;
if (registerType == TYP_DOUBLE)
assignedInterval2 = findAnotherHalfRegRec(physRegRecord)->assignedInterval;
// Both intervals should not be nullptr at the same time, becasue we already handle this case before.
assert(!(assignedInterval == nullptr && assignedInterval2 == nullptr));
if (assignedInterval != nullptr && !checkActiveInterval(assignedInterval, refLocation))
return false;
if (assignedInterval2 != nullptr && !checkActiveInterval(assignedInterval2, refLocation))
return false;
return true;
#else
return checkActiveInterval(assignedInterval, refLocation);
#endif
}
#ifdef _TARGET_ARM_
void LinearScan::unassignDoublePhysReg(RegRecord* doubleRegRecord)
{
assert(genIsValidDoubleReg(doubleRegRecord->regNum));
RegRecord* doubleRegRecordLo = doubleRegRecord;
RegRecord* doubleRegRecordHi = findAnotherHalfRegRec(doubleRegRecordLo);
// For a double register, we has following four cases.
// Case 1: doubleRegRecLo is assigned to TYP_DOUBLE interval
// Case 2: doubleRegRecLo and doubleRegRecHi are assigned to different TYP_FLOAT intervals
// Case 3: doubelRegRecLo is assgined to TYP_FLOAT interval and doubleRegRecHi is nullptr
// Case 4: doubleRegRecordLo is nullptr, and doubleRegRecordHi is assigned to a TYP_FLOAT interval
if (doubleRegRecordLo->assignedInterval != nullptr)
{
if (doubleRegRecordLo->assignedInterval->registerType == TYP_DOUBLE)
{
// Case 1: doubleRegRecLo is assigned to TYP_DOUBLE interval
unassignPhysReg(doubleRegRecordLo, doubleRegRecordLo->assignedInterval->recentRefPosition);
}
else
{
// Case 2: doubleRegRecLo and doubleRegRecHi are assigned to different TYP_FLOAT intervals
// Case 3: doubelRegRecLo is assgined to TYP_FLOAT interval and doubleRegRecHi is nullptr
assert(doubleRegRecordLo->assignedInterval->registerType == TYP_FLOAT);
unassignPhysReg(doubleRegRecordLo, doubleRegRecordLo->assignedInterval->recentRefPosition);
if (doubleRegRecordHi != nullptr)
{
if (doubleRegRecordHi->assignedInterval != nullptr)
{
assert(doubleRegRecordHi->assignedInterval->registerType == TYP_FLOAT);
unassignPhysReg(doubleRegRecordHi, doubleRegRecordHi->assignedInterval->recentRefPosition);
}
}
}
}
else
{
// Case 4: doubleRegRecordLo is nullptr, and doubleRegRecordHi is assigned to a TYP_FLOAT interval
assert(doubleRegRecordHi->assignedInterval != nullptr);
assert(doubleRegRecordHi->assignedInterval->registerType == TYP_FLOAT);
unassignPhysReg(doubleRegRecordHi, doubleRegRecordHi->assignedInterval->recentRefPosition);
}
}
#endif // _TARGET_ARM_
//----------------------------------------------------------------------------------------
// isRegInUse: Test whether regRec is being used at the refPosition
//
// Arguments:
// regRec - A register to be tested
// refPosition - RefPosition where regRec is tested
// nextLocation - next RefPosition of interval assigned to regRec
//
// Return Value:
// True - if regRec is beding used
// False - otherwise
//
// Note: This helper is designed to be used only from allocateBusyReg()
//
bool LinearScan::isRegInUse(RegRecord* regRec, RefPosition* refPosition, LsraLocation* nextLocation)
{
Interval* assignedInterval = regRec->assignedInterval;
if (assignedInterval != nullptr)
{
LsraLocation refLocation = refPosition->nodeLocation;
RefPosition* nextRefPosition = assignedInterval->getNextRefPosition();
*nextLocation = assignedInterval->getNextRefLocation();
// We should never spill a register that's occupied by an Interval with its next use at the current
// location.
// Normally this won't occur (unless we actually had more uses in a single node than there are registers),
// because we'll always find something with a later nextLocation, but it can happen in stress when
// we have LSRA_SELECT_NEAREST.
if ((*nextLocation == refLocation) && !refPosition->isFixedRegRef && nextRefPosition->RequiresRegister())
{
return true;
}
}
return false;
}
//------------------------------------------------------------------------
// allocateBusyReg: Find a busy register that satisfies the requirements for refPosition,
// and that can be spilled.
//
// Arguments:
// current The interval for the current allocation
// refPosition The RefPosition of the current Interval for which a register is being allocated
// allocateIfProfitable If true, a reg may not be allocated if all other ref positions currently
// occupying registers are more important than the 'refPosition'.
//
// Return Value:
// The regNumber allocated to the RefPositon. Returns REG_NA if no free register is found.
//
// Note: Currently this routine uses weight and farthest distance of next reference
// to select a ref position for spilling.
// a) if allocateIfProfitable = false
// The ref position chosen for spilling will be the lowest weight
// of all and if there is is more than one ref position with the
// same lowest weight, among them choses the one with farthest
// distance to its next reference.
//
// b) if allocateIfProfitable = true
// The ref position chosen for spilling will not only be lowest weight
// of all but also has a weight lower than 'refPosition'. If there is
// no such ref position, reg will not be allocated.
regNumber LinearScan::allocateBusyReg(Interval* current, RefPosition* refPosition, bool allocateIfProfitable)
{
regNumber foundReg = REG_NA;
RegisterType regType = getRegisterType(current, refPosition);
regMaskTP candidates = refPosition->registerAssignment;
regMaskTP preferences = (current->registerPreferences & candidates);
if (preferences == RBM_NONE)
{
preferences = candidates;
}
if (candidates == RBM_NONE)
{
// This assumes only integer and floating point register types
// if we target a processor with additional register types,
// this would have to change
candidates = allRegs(regType);
}
#ifdef DEBUG
candidates = stressLimitRegs(refPosition, candidates);
#endif // DEBUG
// TODO-CQ: Determine whether/how to take preferences into account in addition to
// prefering the one with the furthest ref position when considering
// a candidate to spill
RegRecord* farthestRefPhysRegRecord = nullptr;
#ifdef _TARGET_ARM_
RegRecord* farthestRefPhysRegRecord2 = nullptr;
#endif
LsraLocation farthestLocation = MinLocation;
LsraLocation refLocation = refPosition->nodeLocation;
unsigned farthestRefPosWeight;
if (allocateIfProfitable)
{
// If allocating a reg is optional, we will consider those ref positions
// whose weight is less than 'refPosition' for spilling.
farthestRefPosWeight = getWeight(refPosition);
}
else
{
// If allocating a reg is a must, we start off with max weight so
// that the first spill candidate will be selected based on
// farthest distance alone. Since we start off with farthestLocation
// initialized to MinLocation, the first available ref position
// will be selected as spill candidate and its weight as the
// fathestRefPosWeight.
farthestRefPosWeight = BB_MAX_WEIGHT;
}
for (regNumber regNum : Registers(regType))
{
regMaskTP candidateBit = genRegMask(regNum);
if (!(candidates & candidateBit))
{
continue;
}
RegRecord* physRegRecord = getRegisterRecord(regNum);
#ifdef _TARGET_ARM_
RegRecord* physRegRecord2 = nullptr;
// For ARM32, let's consider two float registers consisting a double reg together,
// when allocaing a double register.
if (current->registerType == TYP_DOUBLE)
{
assert(genIsValidDoubleReg(regNum));
physRegRecord2 = findAnotherHalfRegRec(physRegRecord);
}
#endif
if (physRegRecord->isBusyUntilNextKill)
{
continue;
}
Interval* assignedInterval = physRegRecord->assignedInterval;
#ifdef _TARGET_ARM_
Interval* assignedInterval2 = (physRegRecord2 == nullptr) ? nullptr : physRegRecord2->assignedInterval;
#endif
// If there is a fixed reference at the same location (and it's not due to this reference),
// don't use it.
if (physRegRecord->conflictingFixedRegReference(refPosition))
{
assert(candidates != candidateBit);
continue;
}
LsraLocation physRegNextLocation = MaxLocation;
if (refPosition->isFixedRefOfRegMask(candidateBit))
{
// Either there is a fixed reference due to this node, or one associated with a
// fixed use fed by a def at this node.
// In either case, we must use this register as it's the only candidate
// TODO-CQ: At the time we allocate a register to a fixed-reg def, if it's not going
// to remain live until the use, we should set the candidates to allRegs(regType)
// to avoid a spill - codegen can then insert the copy.
assert(candidates == candidateBit);
// If a refPosition has a fixed reg as its candidate and is also marked
// as allocateIfProfitable, we should allocate fixed reg only if the
// weight of this ref position is greater than the weight of the ref
// position to which fixed reg is assigned. Such a case would arise
// on x86 under LSRA stress.
if (!allocateIfProfitable)
{
physRegNextLocation = MaxLocation;
farthestRefPosWeight = BB_MAX_WEIGHT;
}
}
else
{
physRegNextLocation = physRegRecord->getNextRefLocation();
// If refPosition requires a fixed register, we should reject all others.
// Otherwise, we will still evaluate all phyRegs though their next location is
// not better than farthestLocation found so far.
//
// TODO: this method should be using an approach similar to tryAllocateFreeReg()
// where it uses a regOrder array to avoid iterating over any but the single
// fixed candidate.
if (refPosition->isFixedRegRef && physRegNextLocation < farthestLocation)
{
continue;
}
}
// If this register is not assigned to an interval, either
// - it has a FixedReg reference at the current location that is not this reference, OR
// - this is the special case of a fixed loReg, where this interval has a use at the same location
// In either case, we cannot use it
CLANG_FORMAT_COMMENT_ANCHOR;
#ifdef _TARGET_ARM_
if (assignedInterval == nullptr && assignedInterval2 == nullptr)
#else
if (assignedInterval == nullptr)
#endif
{
RefPosition* nextPhysRegPosition = physRegRecord->getNextRefPosition();
#ifndef _TARGET_ARM64_
// TODO-Cleanup: Revisit this after Issue #3524 is complete
// On ARM64 the nodeLocation is not always == refLocation, Disabling this assert for now.
assert(nextPhysRegPosition->nodeLocation == refLocation && candidateBit != candidates);
#endif
continue;
}
#ifdef _TARGET_ARM_
RefPosition* recentAssignedRef = (assignedInterval == nullptr) ? nullptr : assignedInterval->recentRefPosition;
RefPosition* recentAssignedRef2 =
(assignedInterval2 == nullptr) ? nullptr : assignedInterval2->recentRefPosition;
#else
RefPosition* recentAssignedRef = assignedInterval->recentRefPosition;
#endif
if (!checkActiveIntervals(physRegRecord, refLocation, current->registerType))
{
continue;
}
// If we have a recentAssignedRef, check that it is going to be OK to spill it
//
// TODO-Review: Under what conditions recentAssginedRef would be null?
unsigned recentAssignedRefWeight = BB_ZERO_WEIGHT;
#ifdef _TARGET_ARM_
if (current->registerType == TYP_DOUBLE)
{
if (!canSpillDoubleReg(physRegRecord, refLocation, &recentAssignedRefWeight, farthestRefPosWeight))
continue;
}
else
#endif
// This if-stmt is associated with the above else
if (!canSpillReg(physRegRecord, refLocation, &recentAssignedRefWeight, farthestRefPosWeight))
{
continue;
}
LsraLocation nextLocation = MinLocation;
if (isRegInUse(physRegRecord, refPosition, &nextLocation))
{
continue;
}
#ifdef _TARGET_ARM_
if (current->registerType == TYP_DOUBLE)
{
LsraLocation nextLocation2 = MinLocation;
if (isRegInUse(physRegRecord2, refPosition, &nextLocation2))
{
continue;
}
nextLocation = (nextLocation > nextLocation2) ? nextLocation : nextLocation2;
}
#endif
if (nextLocation > physRegNextLocation)
{
nextLocation = physRegNextLocation;
}
bool isBetterLocation;
#ifdef DEBUG
if (doSelectNearest() && farthestRefPhysRegRecord != nullptr)
{
isBetterLocation = (nextLocation <= farthestLocation);
}
else
#endif
// This if-stmt is associated with the above else
if (recentAssignedRefWeight < farthestRefPosWeight)
{
isBetterLocation = true;
}
else
{
// This would mean the weight of spill ref position we found so far is equal
// to the weight of the ref position that is being evaluated. In this case
// we prefer to spill ref position whose distance to its next reference is
// the farthest.
assert(recentAssignedRefWeight == farthestRefPosWeight);
// If allocateIfProfitable=true, the first spill candidate selected
// will be based on weight alone. After we have found a spill
// candidate whose weight is less than the 'refPosition', we will
// consider farthest distance when there is a tie in weights.
// This is to ensure that we don't spill a ref position whose
// weight is equal to weight of 'refPosition'.
if (allocateIfProfitable && farthestRefPhysRegRecord == nullptr)
{
isBetterLocation = false;
}
else
{
isBetterLocation = (nextLocation > farthestLocation);
if (nextLocation > farthestLocation)
{
isBetterLocation = true;
}
else if (nextLocation == farthestLocation)
{
// Both weight and distance are equal.
// Prefer that ref position which is marked both reload and
// allocate if profitable. These ref positions don't need
// need to be spilled as they are already in memory and
// codegen considers them as contained memory operands.
CLANG_FORMAT_COMMENT_ANCHOR;
#ifdef _TARGET_ARM
// TODO-CQ-ARM: Just conservatively "and" two condition. We may implement better condision later.
isBetterLocation = true;
if (recentAssignedRef != nullptr)
isBetterLocation &= (recentAssignedRef->reload && recentAssignedRef->AllocateIfProfitable());
if (recentAssignedRef2 != nullptr)
isBetterLocation &= (recentAssignedRef2->reload && recentAssignedRef2->AllocateIfProfitable());
#else
isBetterLocation = (recentAssignedRef != nullptr) && recentAssignedRef->reload &&
recentAssignedRef->AllocateIfProfitable();
#endif
}
else
{
isBetterLocation = false;
}
}
}
if (isBetterLocation)
{
farthestLocation = nextLocation;
farthestRefPhysRegRecord = physRegRecord;
#ifdef _TARGET_ARM_
farthestRefPhysRegRecord2 = physRegRecord2;
#endif
farthestRefPosWeight = recentAssignedRefWeight;
}
}
#if DEBUG
if (allocateIfProfitable)
{
// There may not be a spill candidate or if one is found
// its weight must be less than the weight of 'refPosition'
assert((farthestRefPhysRegRecord == nullptr) || (farthestRefPosWeight < getWeight(refPosition)));
}
else
{
// Must have found a spill candidate.
assert(farthestRefPhysRegRecord != nullptr);
if ((farthestLocation == refLocation) && !refPosition->isFixedRegRef)
{
#ifdef _TARGET_ARM_
Interval* assignedInterval =
(farthestRefPhysRegRecord == nullptr) ? nullptr : farthestRefPhysRegRecord->assignedInterval;
Interval* assignedInterval2 =
(farthestRefPhysRegRecord2 == nullptr) ? nullptr : farthestRefPhysRegRecord2->assignedInterval;
RefPosition* nextRefPosition =
(assignedInterval == nullptr) ? nullptr : assignedInterval->getNextRefPosition();
RefPosition* nextRefPosition2 =
(assignedInterval2 == nullptr) ? nullptr : assignedInterval2->getNextRefPosition();
if (nextRefPosition != nullptr)
{
if (nextRefPosition2 != nullptr)
{
assert(!nextRefPosition->RequiresRegister() || !nextRefPosition2->RequiresRegister());
}
else
{
assert(!nextRefPosition->RequiresRegister());
}
}
else
{
assert(nextRefPosition2 != nullptr && !nextRefPosition2->RequiresRegister());
}
#else // !_TARGET_ARM_
Interval* assignedInterval = farthestRefPhysRegRecord->assignedInterval;
RefPosition* nextRefPosition = assignedInterval->getNextRefPosition();
assert(!nextRefPosition->RequiresRegister());
#endif // !_TARGET_ARM_
}
else
{
assert(farthestLocation > refLocation || refPosition->isFixedRegRef);
}
}
#endif // DEBUG
if (farthestRefPhysRegRecord != nullptr)
{
foundReg = farthestRefPhysRegRecord->regNum;
#ifdef _TARGET_ARM_
if (current->registerType == TYP_DOUBLE)
{
assert(genIsValidDoubleReg(foundReg));
unassignDoublePhysReg(farthestRefPhysRegRecord);
}
else
#endif
{
unassignPhysReg(farthestRefPhysRegRecord, farthestRefPhysRegRecord->assignedInterval->recentRefPosition);
}
assignPhysReg(farthestRefPhysRegRecord, current);
refPosition->registerAssignment = genRegMask(foundReg);
}
else
{
foundReg = REG_NA;
refPosition->registerAssignment = RBM_NONE;
}
return foundReg;
}
// Grab a register to use to copy and then immediately use.
// This is called only for localVar intervals that already have a register
// assignment that is not compatible with the current RefPosition.
// This is not like regular assignment, because we don't want to change
// any preferences or existing register assignments.
// Prefer a free register that's got the earliest next use.
// Otherwise, spill something with the farthest next use
//
regNumber LinearScan::assignCopyReg(RefPosition* refPosition)
{
Interval* currentInterval = refPosition->getInterval();
assert(currentInterval != nullptr);
assert(currentInterval->isActive);
bool foundFreeReg = false;
RegRecord* bestPhysReg = nullptr;
LsraLocation bestLocation = MinLocation;
regMaskTP candidates = refPosition->registerAssignment;
// Save the relatedInterval, if any, so that it doesn't get modified during allocation.
Interval* savedRelatedInterval = currentInterval->relatedInterval;
currentInterval->relatedInterval = nullptr;
// We don't want really want to change the default assignment,
// so 1) pretend this isn't active, and 2) remember the old reg
regNumber oldPhysReg = currentInterval->physReg;
RegRecord* oldRegRecord = currentInterval->assignedReg;
assert(oldRegRecord->regNum == oldPhysReg);
currentInterval->isActive = false;
regNumber allocatedReg = tryAllocateFreeReg(currentInterval, refPosition);
if (allocatedReg == REG_NA)
{
allocatedReg = allocateBusyReg(currentInterval, refPosition, false);
}
// Now restore the old info
currentInterval->relatedInterval = savedRelatedInterval;
currentInterval->physReg = oldPhysReg;
currentInterval->assignedReg = oldRegRecord;
currentInterval->isActive = true;
refPosition->copyReg = true;
return allocatedReg;
}
//------------------------------------------------------------------------
// isAssigned: This is the function to check if the given RegRecord has an assignedInterval
// regardless of lastLocation.
// So it would be call isAssigned() with Maxlocation value.
//
// Arguments:
// regRec - The RegRecord to check that it is assigned.
// newRegType - There are elements to judge according to the upcoming register type.
//
// Return Value:
// Returns true if the given RegRecord has an assignedInterval.
//
// Notes:
// There is the case to check if the RegRecord has an assignedInterval regardless of Lastlocation.
//
bool LinearScan::isAssigned(RegRecord* regRec ARM_ARG(RegisterType newRegType))
{
return isAssigned(regRec, MaxLocation ARM_ARG(newRegType));
}
//------------------------------------------------------------------------
// isAssigned: Check whether the given RegRecord has an assignedInterval
// that has a reference prior to the given location.
//
// Arguments:
// regRec - The RegRecord of interest
// lastLocation - The LsraLocation up to which we want to check
// newRegType - The `RegisterType` of interval we want to check
// (this is for the purposes of checking the other half of a TYP_DOUBLE RegRecord)
//
// Return value:
// Returns true if the given RegRecord (and its other half, if TYP_DOUBLE) has an assignedInterval
// that is referenced prior to the given location
//
// Notes:
// The register is not considered to be assigned if it has no assignedInterval, or that Interval's
// next reference is beyond lastLocation
//
bool LinearScan::isAssigned(RegRecord* regRec, LsraLocation lastLocation ARM_ARG(RegisterType newRegType))
{
Interval* assignedInterval = regRec->assignedInterval;
if ((assignedInterval == nullptr) || assignedInterval->getNextRefLocation() > lastLocation)
{
#ifdef _TARGET_ARM_
if (newRegType == TYP_DOUBLE)
{
RegRecord* anotherRegRec = findAnotherHalfRegRec(regRec);
if ((anotherRegRec->assignedInterval == nullptr) ||
(anotherRegRec->assignedInterval->getNextRefLocation() > lastLocation))
{
// In case the newRegType is a double register,
// the score would be set UNASSIGNED if another register is also not set.
return false;
}
}
else
#endif
{
return false;
}
}
return true;
}
// Check if the interval is already assigned and if it is then unassign the physical record
// then set the assignedInterval to 'interval'
//
void LinearScan::checkAndAssignInterval(RegRecord* regRec, Interval* interval)
{
if (regRec->assignedInterval != nullptr && regRec->assignedInterval != interval)
{
// This is allocated to another interval. Either it is inactive, or it was allocated as a
// copyReg and is therefore not the "assignedReg" of the other interval. In the latter case,
// we simply unassign it - in the former case we need to set the physReg on the interval to
// REG_NA to indicate that it is no longer in that register.
// The lack of checking for this case resulted in an assert in the retail version of System.dll,
// in method SerialStream.GetDcbFlag.
// Note that we can't check for the copyReg case, because we may have seen a more recent
// RefPosition for the Interval that was NOT a copyReg.
if (regRec->assignedInterval->assignedReg == regRec)
{
assert(regRec->assignedInterval->isActive == false);
regRec->assignedInterval->physReg = REG_NA;
}
unassignPhysReg(regRec->regNum);
}
updateAssignedInterval(regRec, interval, interval->registerType);
}
// Assign the given physical register interval to the given interval
void LinearScan::assignPhysReg(RegRecord* regRec, Interval* interval)
{
regMaskTP assignedRegMask = genRegMask(regRec->regNum);
compiler->codeGen->regSet.rsSetRegsModified(assignedRegMask DEBUGARG(dumpTerse));
checkAndAssignInterval(regRec, interval);
interval->assignedReg = regRec;
interval->physReg = regRec->regNum;
interval->isActive = true;
if (interval->isLocalVar)
{
// Prefer this register for future references
interval->updateRegisterPreferences(assignedRegMask);
}
}
//------------------------------------------------------------------------
// setIntervalAsSplit: Set this Interval as being split
//
// Arguments:
// interval - The Interval which is being split
//
// Return Value:
// None.
//
// Notes:
// The given Interval will be marked as split, and it will be added to the
// set of splitOrSpilledVars.
//
// Assumptions:
// "interval" must be a lclVar interval, as tree temps are never split.
// This is asserted in the call to getVarIndex().
//
void LinearScan::setIntervalAsSplit(Interval* interval)
{
if (interval->isLocalVar)
{
unsigned varIndex = interval->getVarIndex(compiler);
if (!interval->isSplit)
{
VarSetOps::AddElemD(compiler, splitOrSpilledVars, varIndex);
}
else
{
assert(VarSetOps::IsMember(compiler, splitOrSpilledVars, varIndex));
}
}
interval->isSplit = true;
}
//------------------------------------------------------------------------
// setIntervalAsSpilled: Set this Interval as being spilled
//
// Arguments:
// interval - The Interval which is being spilled
//
// Return Value:
// None.
//
// Notes:
// The given Interval will be marked as spilled, and it will be added
// to the set of splitOrSpilledVars.
//
void LinearScan::setIntervalAsSpilled(Interval* interval)
{
if (interval->isLocalVar)
{
unsigned varIndex = interval->getVarIndex(compiler);
if (!interval->isSpilled)
{
VarSetOps::AddElemD(compiler, splitOrSpilledVars, varIndex);
}
else
{
assert(VarSetOps::IsMember(compiler, splitOrSpilledVars, varIndex));
}
}
interval->isSpilled = true;
}
//------------------------------------------------------------------------
// spill: Spill this Interval between "fromRefPosition" and "toRefPosition"
//
// Arguments:
// fromRefPosition - The RefPosition at which the Interval is to be spilled
// toRefPosition - The RefPosition at which it must be reloaded
//
// Return Value:
// None.
//
// Assumptions:
// fromRefPosition and toRefPosition must not be null
//
void LinearScan::spillInterval(Interval* interval, RefPosition* fromRefPosition, RefPosition* toRefPosition)
{
assert(fromRefPosition != nullptr && toRefPosition != nullptr);
assert(fromRefPosition->getInterval() == interval && toRefPosition->getInterval() == interval);
assert(fromRefPosition->nextRefPosition == toRefPosition);
if (!fromRefPosition->lastUse)
{
// If not allocated a register, Lcl var def/use ref positions even if reg optional
// should be marked as spillAfter.
if (!fromRefPosition->RequiresRegister() && !(interval->isLocalVar && fromRefPosition->IsActualRef()))
{
fromRefPosition->registerAssignment = RBM_NONE;
}
else
{
fromRefPosition->spillAfter = true;
}
}
assert(toRefPosition != nullptr);
#ifdef DEBUG
if (VERBOSE)
{
dumpLsraAllocationEvent(LSRA_EVENT_SPILL, interval);
}
#endif // DEBUG
INTRACK_STATS(updateLsraStat(LSRA_STAT_SPILL, fromRefPosition->bbNum));
interval->isActive = false;
setIntervalAsSpilled(interval);
// If fromRefPosition occurs before the beginning of this block, mark this as living in the stack
// on entry to this block.
if (fromRefPosition->nodeLocation <= curBBStartLocation)
{
// This must be a lclVar interval
assert(interval->isLocalVar);
setInVarRegForBB(curBBNum, interval->varNum, REG_STK);
}
}
//------------------------------------------------------------------------
// unassignPhysRegNoSpill: Unassign the given physical register record from
// an active interval, without spilling.
//
// Arguments:
// regRec - the RegRecord to be unasssigned
//
// Return Value:
// None.
//
// Assumptions:
// The assignedInterval must not be null, and must be active.
//
// Notes:
// This method is used to unassign a register when an interval needs to be moved to a
// different register, but not (yet) spilled.
void LinearScan::unassignPhysRegNoSpill(RegRecord* regRec)
{
Interval* assignedInterval = regRec->assignedInterval;
assert(assignedInterval != nullptr && assignedInterval->isActive);
assignedInterval->isActive = false;
unassignPhysReg(regRec, nullptr);
assignedInterval->isActive = true;
}
//------------------------------------------------------------------------
// checkAndClearInterval: Clear the assignedInterval for the given
// physical register record
//
// Arguments:
// regRec - the physical RegRecord to be unasssigned
// spillRefPosition - The RefPosition at which the assignedInterval is to be spilled
// or nullptr if we aren't spilling
//
// Return Value:
// None.
//
// Assumptions:
// see unassignPhysReg
//
void LinearScan::checkAndClearInterval(RegRecord* regRec, RefPosition* spillRefPosition)
{
Interval* assignedInterval = regRec->assignedInterval;
assert(assignedInterval != nullptr);
regNumber thisRegNum = regRec->regNum;
if (spillRefPosition == nullptr)
{
// Note that we can't assert for the copyReg case
//
if (assignedInterval->physReg == thisRegNum)
{
assert(assignedInterval->isActive == false);
}
}
else
{
assert(spillRefPosition->getInterval() == assignedInterval);
}
updateAssignedInterval(regRec, nullptr, assignedInterval->registerType);
}
//------------------------------------------------------------------------
// unassignPhysReg: Unassign the given physical register record, and spill the
// assignedInterval at the given spillRefPosition, if any.
//
// Arguments:
// regRec - The RegRecord to be unasssigned
// newRegType - The RegisterType of interval that would be assigned
//
// Return Value:
// None.
//
// Notes:
// On ARM architecture, Intervals have to be unassigned considering
// with the register type of interval that would be assigned.
//
void LinearScan::unassignPhysReg(RegRecord* regRec ARM_ARG(RegisterType newRegType))
{
RegRecord* regRecToUnassign = regRec;
#ifdef _TARGET_ARM_
RegRecord* anotherRegRec = nullptr;
if ((regRecToUnassign->assignedInterval != nullptr) &&
(regRecToUnassign->assignedInterval->registerType == TYP_DOUBLE))
{
// If the register type of interval(being unassigned or new) is TYP_DOUBLE,
// It should have to be valid double register (even register)
if (!genIsValidDoubleReg(regRecToUnassign->regNum))
{
regRecToUnassign = findAnotherHalfRegRec(regRec);
}
}
else
{
if (newRegType == TYP_DOUBLE)
{
anotherRegRec = findAnotherHalfRegRec(regRecToUnassign);
}
}
#endif
if (regRecToUnassign->assignedInterval != nullptr)
{
unassignPhysReg(regRecToUnassign, regRecToUnassign->assignedInterval->recentRefPosition);
}
#ifdef _TARGET_ARM_
if ((anotherRegRec != nullptr) && (anotherRegRec->assignedInterval != nullptr))
{
unassignPhysReg(anotherRegRec, anotherRegRec->assignedInterval->recentRefPosition);
}
#endif
}
//------------------------------------------------------------------------
// unassignPhysReg: Unassign the given physical register record, and spill the
// assignedInterval at the given spillRefPosition, if any.
//
// Arguments:
// regRec - the RegRecord to be unasssigned
// spillRefPosition - The RefPosition at which the assignedInterval is to be spilled
//
// Return Value:
// None.
//
// Assumptions:
// The assignedInterval must not be null.
// If spillRefPosition is null, the assignedInterval must be inactive, or not currently
// assigned to this register (e.g. this is a copyReg for that Interval).
// Otherwise, spillRefPosition must be associated with the assignedInterval.
//
void LinearScan::unassignPhysReg(RegRecord* regRec, RefPosition* spillRefPosition)
{
Interval* assignedInterval = regRec->assignedInterval;
assert(assignedInterval != nullptr);
regNumber thisRegNum = regRec->regNum;
#ifdef _TARGET_ARM_
RegRecord* anotherRegRec = nullptr;
// Prepare second half RegRecord of a double register for TYP_DOUBLE
if (assignedInterval->registerType == TYP_DOUBLE)
{
assert(isFloatRegType(regRec->registerType));
anotherRegRec = findAnotherHalfRegRec(regRec);
// Both two RegRecords should have been assigned to the same interval.
assert(assignedInterval == anotherRegRec->assignedInterval);
}
#endif // _TARGET_ARM_
checkAndClearInterval(regRec, spillRefPosition);
#ifdef _TARGET_ARM_
if (assignedInterval->registerType == TYP_DOUBLE)
{
// Both two RegRecords should have been unassigned together.
assert(regRec->assignedInterval == nullptr);
assert(anotherRegRec->assignedInterval == nullptr);
}
#endif // _TARGET_ARM_
#ifdef DEBUG
if (VERBOSE && !dumpTerse)
{
printf("unassigning %s: ", getRegName(regRec->regNum));
assignedInterval->dump();
printf("\n");
}
#endif // DEBUG
RefPosition* nextRefPosition = nullptr;
if (spillRefPosition != nullptr)
{
nextRefPosition = spillRefPosition->nextRefPosition;
}
if (assignedInterval->physReg != REG_NA && assignedInterval->physReg != thisRegNum)
{
// This must have been a temporary copy reg, but we can't assert that because there
// may have been intervening RefPositions that were not copyRegs.
// reg->assignedInterval has already been set to nullptr by checkAndClearInterval()
assert(regRec->assignedInterval == nullptr);
return;
}
regNumber victimAssignedReg = assignedInterval->physReg;
assignedInterval->physReg = REG_NA;
bool spill = assignedInterval->isActive && nextRefPosition != nullptr;
if (spill)
{
// If this is an active interval, it must have a recentRefPosition,
// otherwise it would not be active
assert(spillRefPosition != nullptr);
#if 0
// TODO-CQ: Enable this and insert an explicit GT_COPY (otherwise there's no way to communicate
// to codegen that we want the copyReg to be the new home location).
// If the last reference was a copyReg, and we're spilling the register
// it was copied from, then make the copyReg the new primary location
// if possible
if (spillRefPosition->copyReg)
{
regNumber copyFromRegNum = victimAssignedReg;
regNumber copyRegNum = genRegNumFromMask(spillRefPosition->registerAssignment);
if (copyFromRegNum == thisRegNum &&
getRegisterRecord(copyRegNum)->assignedInterval == assignedInterval)
{
assert(copyRegNum != thisRegNum);
assignedInterval->physReg = copyRegNum;
assignedInterval->assignedReg = this->getRegisterRecord(copyRegNum);
return;
}
}
#endif // 0
#ifdef DEBUG
// With JitStressRegs == 0x80 (LSRA_EXTEND_LIFETIMES), we may have a RefPosition
// that is not marked lastUse even though the treeNode is a lastUse. In that case
// we must not mark it for spill because the register will have been immediately freed
// after use. While we could conceivably add special handling for this case in codegen,
// it would be messy and undesirably cause the "bleeding" of LSRA stress modes outside
// of LSRA.
if (extendLifetimes() && assignedInterval->isLocalVar && RefTypeIsUse(spillRefPosition->refType) &&
spillRefPosition->treeNode != nullptr && (spillRefPosition->treeNode->gtFlags & GTF_VAR_DEATH) != 0)
{
dumpLsraAllocationEvent(LSRA_EVENT_SPILL_EXTENDED_LIFETIME, assignedInterval);
assignedInterval->isActive = false;
spill = false;
// If the spillRefPosition occurs before the beginning of this block, it will have
// been marked as living in this register on entry to this block, but we now need
// to mark this as living on the stack.
if (spillRefPosition->nodeLocation <= curBBStartLocation)
{
setInVarRegForBB(curBBNum, assignedInterval->varNum, REG_STK);
if (spillRefPosition->nextRefPosition != nullptr)
{
setIntervalAsSpilled(assignedInterval);
}
}
else
{
// Otherwise, we need to mark spillRefPosition as lastUse, or the interval
// will remain active beyond its allocated range during the resolution phase.
spillRefPosition->lastUse = true;
}
}
else
#endif // DEBUG
{
spillInterval(assignedInterval, spillRefPosition, nextRefPosition);
}
}
// Maintain the association with the interval, if it has more references.
// Or, if we "remembered" an interval assigned to this register, restore it.
if (nextRefPosition != nullptr)
{
assignedInterval->assignedReg = regRec;
}
else if (canRestorePreviousInterval(regRec, assignedInterval))
{
regRec->assignedInterval = regRec->previousInterval;
regRec->previousInterval = nullptr;
#ifdef _TARGET_ARM_
// Note:
// We can not use updateAssignedInterval() and updatePreviousInterval() here,
// because regRec may not be a even-numbered float register.
// Update second half RegRecord of a double register for TYP_DOUBLE
if (regRec->assignedInterval->registerType == TYP_DOUBLE)
{
RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(regRec);
anotherHalfRegRec->assignedInterval = regRec->assignedInterval;
anotherHalfRegRec->previousInterval = nullptr;
}
#endif // _TARGET_ARM_
#ifdef DEBUG
if (spill)
{
dumpLsraAllocationEvent(LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL_AFTER_SPILL, regRec->assignedInterval,
thisRegNum);
}
else
{
dumpLsraAllocationEvent(LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL, regRec->assignedInterval, thisRegNum);
}
#endif // DEBUG
}
else
{
updateAssignedInterval(regRec, nullptr, assignedInterval->registerType);
updatePreviousInterval(regRec, nullptr, assignedInterval->registerType);
}
}
//------------------------------------------------------------------------
// spillGCRefs: Spill any GC-type intervals that are currently in registers.a
//
// Arguments:
// killRefPosition - The RefPosition for the kill
//
// Return Value:
// None.
//
void LinearScan::spillGCRefs(RefPosition* killRefPosition)
{
// For each physical register that can hold a GC type,
// if it is occupied by an interval of a GC type, spill that interval.
regMaskTP candidateRegs = killRefPosition->registerAssignment;
while (candidateRegs != RBM_NONE)
{
regMaskTP nextRegBit = genFindLowestBit(candidateRegs);
candidateRegs &= ~nextRegBit;
regNumber nextReg = genRegNumFromMask(nextRegBit);
RegRecord* regRecord = getRegisterRecord(nextReg);
Interval* assignedInterval = regRecord->assignedInterval;
if (assignedInterval == nullptr || (assignedInterval->isActive == false) ||
!varTypeIsGC(assignedInterval->registerType))
{
continue;
}
unassignPhysReg(regRecord, assignedInterval->recentRefPosition);
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_DONE_KILL_GC_REFS, nullptr, REG_NA, nullptr));
}
//------------------------------------------------------------------------
// processBlockEndAllocation: Update var locations after 'currentBlock' has been allocated
//
// Arguments:
// currentBlock - the BasicBlock we have just finished allocating registers for
//
// Return Value:
// None
//
// Notes:
// Calls processBlockEndLocations() to set the outVarToRegMap, then gets the next block,
// and sets the inVarToRegMap appropriately.
void LinearScan::processBlockEndAllocation(BasicBlock* currentBlock)
{
assert(currentBlock != nullptr);
if (enregisterLocalVars)
{
processBlockEndLocations(currentBlock);
}
markBlockVisited(currentBlock);
// Get the next block to allocate.
// When the last block in the method has successors, there will be a final "RefTypeBB" to
// ensure that we get the varToRegMap set appropriately, but in that case we don't need
// to worry about "nextBlock".
BasicBlock* nextBlock = getNextBlock();
if (nextBlock != nullptr)
{
processBlockStartLocations(nextBlock, true);
}
}
//------------------------------------------------------------------------
// rotateBlockStartLocation: When in the LSRA_BLOCK_BOUNDARY_ROTATE stress mode, attempt to
// "rotate" the register assignment for a localVar to the next higher
// register that is available.
//
// Arguments:
// interval - the Interval for the variable whose register is getting rotated
// targetReg - its register assignment from the predecessor block being used for live-in
// availableRegs - registers available for use
//
// Return Value:
// The new register to use.
#ifdef DEBUG
regNumber LinearScan::rotateBlockStartLocation(Interval* interval, regNumber targetReg, regMaskTP availableRegs)
{
if (targetReg != REG_STK && getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_ROTATE)
{
// If we're rotating the register locations at block boundaries, try to use
// the next higher register number of the appropriate register type.
regMaskTP candidateRegs = allRegs(interval->registerType) & availableRegs;
regNumber firstReg = REG_NA;
regNumber newReg = REG_NA;
while (candidateRegs != RBM_NONE)
{
regMaskTP nextRegBit = genFindLowestBit(candidateRegs);
candidateRegs &= ~nextRegBit;
regNumber nextReg = genRegNumFromMask(nextRegBit);
if (nextReg > targetReg)
{
newReg = nextReg;
break;
}
else if (firstReg == REG_NA)
{
firstReg = nextReg;
}
}
if (newReg == REG_NA)
{
assert(firstReg != REG_NA);
newReg = firstReg;
}
targetReg = newReg;
}
return targetReg;
}
#endif // DEBUG
#ifdef _TARGET_ARM_
//--------------------------------------------------------------------------------------
// isSecondHalfReg: Test if recRec is second half of double register
// which is assigned to an interval.
//
// Arguments:
// regRec - a register to be tested
// interval - an interval which is assigned to some register
//
// Assumptions:
// None
//
// Return Value:
// True only if regRec is second half of assignedReg in interval
//
bool LinearScan::isSecondHalfReg(RegRecord* regRec, Interval* interval)
{
RegRecord* assignedReg = interval->assignedReg;
if (assignedReg != nullptr && interval->registerType == TYP_DOUBLE)
{
// interval should have been allocated to a valid double register
assert(genIsValidDoubleReg(assignedReg->regNum));
// Find a second half RegRecord of double register
regNumber firstRegNum = assignedReg->regNum;
regNumber secondRegNum = REG_NEXT(firstRegNum);
assert(genIsValidFloatReg(secondRegNum) && !genIsValidDoubleReg(secondRegNum));
RegRecord* secondRegRec = getRegisterRecord(secondRegNum);
return secondRegRec == regRec;
}
return false;
}
//------------------------------------------------------------------------------------------
// findAnotherHalfRegRec: Find another half RegRecord which forms same ARM32 double register
//
// Arguments:
// regRec - A float RegRecord
//
// Assumptions:
// None
//
// Return Value:
// A RegRecord which forms same double register with regRec
//
RegRecord* LinearScan::findAnotherHalfRegRec(RegRecord* regRec)
{
regNumber anotherHalfRegNum;
RegRecord* anotherHalfRegRec;
assert(genIsValidFloatReg(regRec->regNum));
// Find another half register for TYP_DOUBLE interval,
// following same logic in canRestorePreviousInterval().
if (genIsValidDoubleReg(regRec->regNum))
{
anotherHalfRegNum = REG_NEXT(regRec->regNum);
assert(!genIsValidDoubleReg(anotherHalfRegNum));
}
else
{
anotherHalfRegNum = REG_PREV(regRec->regNum);
assert(genIsValidDoubleReg(anotherHalfRegNum));
}
anotherHalfRegRec = getRegisterRecord(anotherHalfRegNum);
return anotherHalfRegRec;
}
#endif
//--------------------------------------------------------------------------------------
// canRestorePreviousInterval: Test if we can restore previous interval
//
// Arguments:
// regRec - a register which contains previous interval to be restored
// assignedInterval - an interval just unassigned
//
// Assumptions:
// None
//
// Return Value:
// True only if previous interval of regRec can be restored
//
bool LinearScan::canRestorePreviousInterval(RegRecord* regRec, Interval* assignedInterval)
{
bool retVal =
(regRec->previousInterval != nullptr && regRec->previousInterval != assignedInterval &&
regRec->previousInterval->assignedReg == regRec && regRec->previousInterval->getNextRefPosition() != nullptr);
#ifdef _TARGET_ARM_
if (retVal && regRec->previousInterval->registerType == TYP_DOUBLE)
{
RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(regRec);
retVal = retVal && anotherHalfRegRec->assignedInterval == nullptr;
}
#endif
return retVal;
}
bool LinearScan::isAssignedToInterval(Interval* interval, RegRecord* regRec)
{
bool isAssigned = (interval->assignedReg == regRec);
#ifdef _TARGET_ARM_
isAssigned |= isSecondHalfReg(regRec, interval);
#endif
return isAssigned;
}
//------------------------------------------------------------------------
// processBlockStartLocations: Update var locations on entry to 'currentBlock' and clear constant
// registers.
//
// Arguments:
// currentBlock - the BasicBlock we are about to allocate registers for
// allocationPass - true if we are currently allocating registers (versus writing them back)
//
// Return Value:
// None
//
// Notes:
// During the allocation pass, we use the outVarToRegMap of the selected predecessor to
// determine the lclVar locations for the inVarToRegMap.
// During the resolution (write-back) pass, we only modify the inVarToRegMap in cases where
// a lclVar was spilled after the block had been completed.
void LinearScan::processBlockStartLocations(BasicBlock* currentBlock, bool allocationPass)
{
// If we have no register candidates we should only call this method during allocation.
assert(enregisterLocalVars || allocationPass);
if (!enregisterLocalVars)
{
// Just clear any constant registers and return.
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
RegRecord* physRegRecord = getRegisterRecord(reg);
Interval* assignedInterval = physRegRecord->assignedInterval;
if (assignedInterval != nullptr)
{
assert(assignedInterval->isConstant);
physRegRecord->assignedInterval = nullptr;
}
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_START_BB, nullptr, REG_NA, currentBlock));
return;
}
unsigned predBBNum = blockInfo[currentBlock->bbNum].predBBNum;
VarToRegMap predVarToRegMap = getOutVarToRegMap(predBBNum);
VarToRegMap inVarToRegMap = getInVarToRegMap(currentBlock->bbNum);
bool hasCriticalInEdge = blockInfo[currentBlock->bbNum].hasCriticalInEdge;
VarSetOps::AssignNoCopy(compiler, currentLiveVars,
VarSetOps::Intersection(compiler, registerCandidateVars, currentBlock->bbLiveIn));
#ifdef DEBUG
if (getLsraExtendLifeTimes())
{
VarSetOps::AssignNoCopy(compiler, currentLiveVars, registerCandidateVars);
}
// If we are rotating register assignments at block boundaries, we want to make the
// inactive registers available for the rotation.
regMaskTP inactiveRegs = RBM_NONE;
#endif // DEBUG
regMaskTP liveRegs = RBM_NONE;
VarSetOps::Iter iter(compiler, currentLiveVars);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
if (!compiler->lvaTable[varNum].lvLRACandidate)
{
continue;
}
regNumber targetReg;
Interval* interval = getIntervalForLocalVar(varIndex);
RefPosition* nextRefPosition = interval->getNextRefPosition();
assert(nextRefPosition != nullptr);
if (allocationPass)
{
targetReg = getVarReg(predVarToRegMap, varIndex);
#ifdef DEBUG
regNumber newTargetReg = rotateBlockStartLocation(interval, targetReg, (~liveRegs | inactiveRegs));
if (newTargetReg != targetReg)
{
targetReg = newTargetReg;
setIntervalAsSplit(interval);
}
#endif // DEBUG
setVarReg(inVarToRegMap, varIndex, targetReg);
}
else // !allocationPass (i.e. resolution/write-back pass)
{
targetReg = getVarReg(inVarToRegMap, varIndex);
// There are four cases that we need to consider during the resolution pass:
// 1. This variable had a register allocated initially, and it was not spilled in the RefPosition
// that feeds this block. In this case, both targetReg and predVarToRegMap[varIndex] will be targetReg.
// 2. This variable had not been spilled prior to the end of predBB, but was later spilled, so
// predVarToRegMap[varIndex] will be REG_STK, but targetReg is its former allocated value.
// In this case, we will normally change it to REG_STK. We will update its "spilled" status when we
// encounter it in resolveLocalRef().
// 2a. If the next RefPosition is marked as a copyReg, we need to retain the allocated register. This is
// because the copyReg RefPosition will not have recorded the "home" register, yet downstream
// RefPositions rely on the correct "home" register.
// 3. This variable was spilled before we reached the end of predBB. In this case, both targetReg and
// predVarToRegMap[varIndex] will be REG_STK, and the next RefPosition will have been marked
// as reload during allocation time if necessary (note that by the time we actually reach the next
// RefPosition, we may be using a different predecessor, at which it is still in a register).
// 4. This variable was spilled during the allocation of this block, so targetReg is REG_STK
// (because we set inVarToRegMap at the time we spilled it), but predVarToRegMap[varIndex]
// is not REG_STK. We retain the REG_STK value in the inVarToRegMap.
if (targetReg != REG_STK)
{
if (getVarReg(predVarToRegMap, varIndex) != REG_STK)
{
// Case #1 above.
assert(getVarReg(predVarToRegMap, varIndex) == targetReg ||
getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_ROTATE);
}
else if (!nextRefPosition->copyReg)
{
// case #2 above.
setVarReg(inVarToRegMap, varIndex, REG_STK);
targetReg = REG_STK;
}
// Else case 2a. - retain targetReg.
}
// Else case #3 or #4, we retain targetReg and nothing further to do or assert.
}
if (interval->physReg == targetReg)
{
if (interval->isActive)
{
assert(targetReg != REG_STK);
assert(interval->assignedReg != nullptr && interval->assignedReg->regNum == targetReg &&
interval->assignedReg->assignedInterval == interval);
liveRegs |= genRegMask(targetReg);
continue;
}
}
else if (interval->physReg != REG_NA)
{
// This can happen if we are using the locations from a basic block other than the
// immediately preceding one - where the variable was in a different location.
if (targetReg != REG_STK)
{
// Unassign it from the register (it will get a new register below).
if (interval->assignedReg != nullptr && interval->assignedReg->assignedInterval == interval)
{
interval->isActive = false;
unassignPhysReg(getRegisterRecord(interval->physReg), nullptr);
}
else
{
// This interval was live in this register the last time we saw a reference to it,
// but has since been displaced.
interval->physReg = REG_NA;
}
}
else if (allocationPass)
{
// Keep the register assignment - if another var has it, it will get unassigned.
// Otherwise, resolution will fix it up later, and it will be more
// likely to match other assignments this way.
interval->isActive = true;
liveRegs |= genRegMask(interval->physReg);
INDEBUG(inactiveRegs |= genRegMask(interval->physReg));
setVarReg(inVarToRegMap, varIndex, interval->physReg);
}
else
{
interval->physReg = REG_NA;
}
}
if (targetReg != REG_STK)
{
RegRecord* targetRegRecord = getRegisterRecord(targetReg);
liveRegs |= genRegMask(targetReg);
if (!interval->isActive)
{
interval->isActive = true;
interval->physReg = targetReg;
interval->assignedReg = targetRegRecord;
}
Interval* assignedInterval = targetRegRecord->assignedInterval;
if (assignedInterval != interval)
{
// Is there another interval currently assigned to this register? If so unassign it.
if (assignedInterval != nullptr)
{
if (isAssignedToInterval(assignedInterval, targetRegRecord))
{
regNumber assignedRegNum = assignedInterval->assignedReg->regNum;
// If the interval is active, it will be set to active when we reach its new
// register assignment (which we must not yet have done, or it wouldn't still be
// assigned to this register).
assignedInterval->isActive = false;
unassignPhysReg(assignedInterval->assignedReg, nullptr);
if (allocationPass && assignedInterval->isLocalVar &&
inVarToRegMap[assignedInterval->getVarIndex(compiler)] == assignedRegNum)
{
inVarToRegMap[assignedInterval->getVarIndex(compiler)] = REG_STK;
}
}
else
{
// This interval is no longer assigned to this register.
updateAssignedInterval(targetRegRecord, nullptr, assignedInterval->registerType);
}
}
assignPhysReg(targetRegRecord, interval);
}
if (interval->recentRefPosition != nullptr && !interval->recentRefPosition->copyReg &&
interval->recentRefPosition->registerAssignment != genRegMask(targetReg))
{
interval->getNextRefPosition()->outOfOrder = true;
}
}
}
// Unassign any registers that are no longer live.
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
if ((liveRegs & genRegMask(reg)) == 0)
{
RegRecord* physRegRecord = getRegisterRecord(reg);
Interval* assignedInterval = physRegRecord->assignedInterval;
if (assignedInterval != nullptr)
{
assert(assignedInterval->isLocalVar || assignedInterval->isConstant);
if (!assignedInterval->isConstant && assignedInterval->assignedReg == physRegRecord)
{
assignedInterval->isActive = false;
if (assignedInterval->getNextRefPosition() == nullptr)
{
unassignPhysReg(physRegRecord, nullptr);
}
inVarToRegMap[assignedInterval->getVarIndex(compiler)] = REG_STK;
}
else
{
// This interval may still be active, but was in another register in an
// intervening block.
updateAssignedInterval(physRegRecord, nullptr, assignedInterval->registerType);
}
#ifdef _TARGET_ARM_
if (assignedInterval->registerType == TYP_DOUBLE)
{
// Skip next float register, because we already addressed a double register
assert(genIsValidDoubleReg(reg));
reg = REG_NEXT(reg);
}
#endif // _TARGET_ARM_
}
}
#ifdef _TARGET_ARM_
else
{
RegRecord* physRegRecord = getRegisterRecord(reg);
Interval* assignedInterval = physRegRecord->assignedInterval;
if (assignedInterval != nullptr && assignedInterval->registerType == TYP_DOUBLE)
{
// Skip next float register, because we already addressed a double register
assert(genIsValidDoubleReg(reg));
reg = REG_NEXT(reg);
}
}
#endif // _TARGET_ARM_
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_START_BB, nullptr, REG_NA, currentBlock));
}
//------------------------------------------------------------------------
// processBlockEndLocations: Record the variables occupying registers after completing the current block.
//
// Arguments:
// currentBlock - the block we have just completed.
//
// Return Value:
// None
//
// Notes:
// This must be called both during the allocation and resolution (write-back) phases.
// This is because we need to have the outVarToRegMap locations in order to set the locations
// at successor blocks during allocation time, but if lclVars are spilled after a block has been
// completed, we need to record the REG_STK location for those variables at resolution time.
void LinearScan::processBlockEndLocations(BasicBlock* currentBlock)
{
assert(currentBlock != nullptr && currentBlock->bbNum == curBBNum);
VarToRegMap outVarToRegMap = getOutVarToRegMap(curBBNum);
VarSetOps::AssignNoCopy(compiler, currentLiveVars,
VarSetOps::Intersection(compiler, registerCandidateVars, currentBlock->bbLiveOut));
#ifdef DEBUG
if (getLsraExtendLifeTimes())
{
VarSetOps::Assign(compiler, currentLiveVars, registerCandidateVars);
}
#endif // DEBUG
regMaskTP liveRegs = RBM_NONE;
VarSetOps::Iter iter(compiler, currentLiveVars);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
Interval* interval = getIntervalForLocalVar(varIndex);
if (interval->isActive)
{
assert(interval->physReg != REG_NA && interval->physReg != REG_STK);
setVarReg(outVarToRegMap, varIndex, interval->physReg);
}
else
{
outVarToRegMap[varIndex] = REG_STK;
}
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_END_BB));
}
#ifdef DEBUG
void LinearScan::dumpRefPositions(const char* str)
{
printf("------------\n");
printf("REFPOSITIONS %s: \n", str);
printf("------------\n");
for (auto& refPos : refPositions)
{
refPos.dump();
}
}
#endif // DEBUG
bool LinearScan::registerIsFree(regNumber regNum, RegisterType regType)
{
RegRecord* physRegRecord = getRegisterRecord(regNum);
bool isFree = physRegRecord->isFree();
#ifdef _TARGET_ARM_
if (isFree && regType == TYP_DOUBLE)
{
isFree = getRegisterRecord(REG_NEXT(regNum))->isFree();
}
#endif // _TARGET_ARM_
return isFree;
}
//------------------------------------------------------------------------
// LinearScan::freeRegister: Make a register available for use
//
// Arguments:
// physRegRecord - the RegRecord for the register to be freed.
//
// Return Value:
// None.
//
// Assumptions:
// None.
// It may be that the RegRecord has already been freed, e.g. due to a kill,
// in which case this method has no effect.
//
// Notes:
// If there is currently an Interval assigned to this register, and it has
// more references (i.e. this is a local last-use, but more uses and/or
// defs remain), it will remain assigned to the physRegRecord. However, since
// it is marked inactive, the register will be available, albeit less desirable
// to allocate.
void LinearScan::freeRegister(RegRecord* physRegRecord)
{
Interval* assignedInterval = physRegRecord->assignedInterval;
// It may have already been freed by a "Kill"
if (assignedInterval != nullptr)
{
assignedInterval->isActive = false;
// If this is a constant node, that we may encounter again (e.g. constant),
// don't unassign it until we need the register.
if (!assignedInterval->isConstant)
{
RefPosition* nextRefPosition = assignedInterval->getNextRefPosition();
// Unassign the register only if there are no more RefPositions, or the next
// one is a def. Note that the latter condition doesn't actually ensure that
// there aren't subsequent uses that could be reached by a def in the assigned
// register, but is merely a heuristic to avoid tying up the register (or using
// it when it's non-optimal). A better alternative would be to use SSA, so that
// we wouldn't unnecessarily link separate live ranges to the same register.
if (nextRefPosition == nullptr || RefTypeIsDef(nextRefPosition->refType))
{
#ifdef _TARGET_ARM_
assert((assignedInterval->registerType != TYP_DOUBLE) || genIsValidDoubleReg(physRegRecord->regNum));
#endif // _TARGET_ARM_
unassignPhysReg(physRegRecord, nullptr);
}
}
}
}
void LinearScan::freeRegisters(regMaskTP regsToFree)
{
if (regsToFree == RBM_NONE)
{
return;
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_FREE_REGS));
while (regsToFree != RBM_NONE)
{
regMaskTP nextRegBit = genFindLowestBit(regsToFree);
regsToFree &= ~nextRegBit;
regNumber nextReg = genRegNumFromMask(nextRegBit);
freeRegister(getRegisterRecord(nextReg));
}
}
// Actual register allocation, accomplished by iterating over all of the previously
// constructed Intervals
// Loosely based on raAssignVars()
//
void LinearScan::allocateRegisters()
{
JITDUMP("*************** In LinearScan::allocateRegisters()\n");
DBEXEC(VERBOSE, lsraDumpIntervals("before allocateRegisters"));
// at start, nothing is active except for register args
for (auto& interval : intervals)
{
Interval* currentInterval = &interval;
currentInterval->recentRefPosition = nullptr;
currentInterval->isActive = false;
if (currentInterval->isLocalVar)
{
LclVarDsc* varDsc = currentInterval->getLocalVar(compiler);
if (varDsc->lvIsRegArg && currentInterval->firstRefPosition != nullptr)
{
currentInterval->isActive = true;
}
}
}
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
getRegisterRecord(reg)->recentRefPosition = nullptr;
getRegisterRecord(reg)->isActive = false;
}
#ifdef DEBUG
regNumber lastAllocatedReg = REG_NA;
if (VERBOSE)
{
dumpRefPositions("BEFORE ALLOCATION");
dumpVarRefPositions("BEFORE ALLOCATION");
printf("\n\nAllocating Registers\n"
"--------------------\n");
if (dumpTerse)
{
dumpRegRecordHeader();
// Now print an empty indent
printf(indentFormat, "");
}
}
#endif // DEBUG
BasicBlock* currentBlock = nullptr;
LsraLocation prevLocation = MinLocation;
regMaskTP regsToFree = RBM_NONE;
regMaskTP delayRegsToFree = RBM_NONE;
// This is the most recent RefPosition for which a register was allocated
// - currently only used for DEBUG but maintained in non-debug, for clarity of code
// (and will be optimized away because in non-debug spillAlways() unconditionally returns false)
RefPosition* lastAllocatedRefPosition = nullptr;
bool handledBlockEnd = false;
for (auto& refPosition : refPositions)
{
RefPosition* currentRefPosition = &refPosition;
#ifdef DEBUG
// Set the activeRefPosition to null until we're done with any boundary handling.
activeRefPosition = nullptr;
if (VERBOSE)
{
if (dumpTerse)
{
// We're really dumping the RegRecords "after" the previous RefPosition, but it's more convenient
// to do this here, since there are a number of "continue"s in this loop.
dumpRegRecords();
}
else
{
printf("\n");
}
}
#endif // DEBUG
// This is the previousRefPosition of the current Referent, if any
RefPosition* previousRefPosition = nullptr;
Interval* currentInterval = nullptr;
Referenceable* currentReferent = nullptr;
bool isInternalRef = false;
RefType refType = currentRefPosition->refType;
currentReferent = currentRefPosition->referent;
if (spillAlways() && lastAllocatedRefPosition != nullptr && !lastAllocatedRefPosition->isPhysRegRef &&
!lastAllocatedRefPosition->getInterval()->isInternal &&
(RefTypeIsDef(lastAllocatedRefPosition->refType) || lastAllocatedRefPosition->getInterval()->isLocalVar))
{
assert(lastAllocatedRefPosition->registerAssignment != RBM_NONE);
RegRecord* regRecord = lastAllocatedRefPosition->getInterval()->assignedReg;
unassignPhysReg(regRecord, lastAllocatedRefPosition);
// Now set lastAllocatedRefPosition to null, so that we don't try to spill it again
lastAllocatedRefPosition = nullptr;
}
// We wait to free any registers until we've completed all the
// uses for the current node.
// This avoids reusing registers too soon.
// We free before the last true def (after all the uses & internal
// registers), and then again at the beginning of the next node.
// This is made easier by assigning two LsraLocations per node - one
// for all the uses, internal registers & all but the last def, and
// another for the final def (if any).
LsraLocation currentLocation = currentRefPosition->nodeLocation;
if ((regsToFree | delayRegsToFree) != RBM_NONE)
{
bool doFreeRegs = false;
// Free at a new location, or at a basic block boundary
if (currentLocation > prevLocation || refType == RefTypeBB)
{
doFreeRegs = true;
}
if (doFreeRegs)
{
freeRegisters(regsToFree);
regsToFree = delayRegsToFree;
delayRegsToFree = RBM_NONE;
}
}
prevLocation = currentLocation;
// get previous refposition, then current refpos is the new previous
if (currentReferent != nullptr)
{
previousRefPosition = currentReferent->recentRefPosition;
currentReferent->recentRefPosition = currentRefPosition;
}
else
{
assert((refType == RefTypeBB) || (refType == RefTypeKillGCRefs));
}
// For the purposes of register resolution, we handle the DummyDefs before
// the block boundary - so the RefTypeBB is after all the DummyDefs.
// However, for the purposes of allocation, we want to handle the block
// boundary first, so that we can free any registers occupied by lclVars
// that aren't live in the next block and make them available for the
// DummyDefs.
if (!handledBlockEnd && (refType == RefTypeBB || refType == RefTypeDummyDef))
{
// Free any delayed regs (now in regsToFree) before processing the block boundary
freeRegisters(regsToFree);
regsToFree = RBM_NONE;
handledBlockEnd = true;
curBBStartLocation = currentRefPosition->nodeLocation;
if (currentBlock == nullptr)
{
currentBlock = startBlockSequence();
}
else
{
processBlockEndAllocation(currentBlock);
currentBlock = moveToNextBlock();
}
#ifdef DEBUG
if (VERBOSE && currentBlock != nullptr && !dumpTerse)
{
currentBlock->dspBlockHeader(compiler);
printf("\n");
}
#endif // DEBUG
}
#ifdef DEBUG
activeRefPosition = currentRefPosition;
if (VERBOSE)
{
if (dumpTerse)
{
dumpRefPositionShort(currentRefPosition, currentBlock);
}
else
{
currentRefPosition->dump();
}
}
#endif // DEBUG
if (refType == RefTypeBB)
{
handledBlockEnd = false;
continue;
}
if (refType == RefTypeKillGCRefs)
{
spillGCRefs(currentRefPosition);
continue;
}
// If this is a FixedReg, disassociate any inactive constant interval from this register.
// Otherwise, do nothing.
if (refType == RefTypeFixedReg)
{
RegRecord* regRecord = currentRefPosition->getReg();
Interval* assignedInterval = regRecord->assignedInterval;
if (assignedInterval != nullptr && !assignedInterval->isActive && assignedInterval->isConstant)
{
regRecord->assignedInterval = nullptr;
#ifdef _TARGET_ARM_
// Update overlapping floating point register for TYP_DOUBLE
if (assignedInterval->registerType == TYP_DOUBLE)
{
regRecord = getRegisterRecord(REG_NEXT(regRecord->regNum));
assignedInterval = regRecord->assignedInterval;
assert(assignedInterval != nullptr && !assignedInterval->isActive && assignedInterval->isConstant);
regRecord->assignedInterval = nullptr;
}
#endif
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_FIXED_REG, nullptr, currentRefPosition->assignedReg()));
continue;
}
// If this is an exposed use, do nothing - this is merely a placeholder to attempt to
// ensure that a register is allocated for the full lifetime. The resolution logic
// will take care of moving to the appropriate register if needed.
if (refType == RefTypeExpUse)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_EXP_USE));
continue;
}
regNumber assignedRegister = REG_NA;
if (currentRefPosition->isIntervalRef())
{
currentInterval = currentRefPosition->getInterval();
assignedRegister = currentInterval->physReg;
#if DEBUG
if (VERBOSE && !dumpTerse)
{
currentInterval->dump();
}
#endif // DEBUG
// Identify the special cases where we decide up-front not to allocate
bool allocate = true;
bool didDump = false;
if (refType == RefTypeParamDef || refType == RefTypeZeroInit)
{
// For a ParamDef with a weighted refCount less than unity, don't enregister it at entry.
// TODO-CQ: Consider doing this only for stack parameters, since otherwise we may be needlessly
// inserting a store.
LclVarDsc* varDsc = currentInterval->getLocalVar(compiler);
assert(varDsc != nullptr);
if (refType == RefTypeParamDef && varDsc->lvRefCntWtd <= BB_UNITY_WEIGHT)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_ENTRY_REG_ALLOCATED, currentInterval));
didDump = true;
allocate = false;
setIntervalAsSpilled(currentInterval);
}
// If it has no actual references, mark it as "lastUse"; since they're not actually part
// of any flow they won't have been marked during dataflow. Otherwise, if we allocate a
// register we won't unassign it.
else if (currentRefPosition->nextRefPosition == nullptr)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_ZERO_REF, currentInterval));
currentRefPosition->lastUse = true;
}
}
#ifdef FEATURE_SIMD
else if (refType == RefTypeUpperVectorSaveDef || refType == RefTypeUpperVectorSaveUse)
{
Interval* lclVarInterval = currentInterval->relatedInterval;
if (lclVarInterval->physReg == REG_NA)
{
allocate = false;
}
}
#endif // FEATURE_SIMD
if (allocate == false)
{
if (assignedRegister != REG_NA)
{
unassignPhysReg(getRegisterRecord(assignedRegister), currentRefPosition);
}
else if (!didDump)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval));
didDump = true;
}
currentRefPosition->registerAssignment = RBM_NONE;
continue;
}
if (currentInterval->isSpecialPutArg)
{
assert(!currentInterval->isLocalVar);
Interval* srcInterval = currentInterval->relatedInterval;
assert(srcInterval->isLocalVar);
if (refType == RefTypeDef)
{
assert(srcInterval->recentRefPosition->nodeLocation == currentLocation - 1);
RegRecord* physRegRecord = srcInterval->assignedReg;
// For a putarg_reg to be special, its next use location has to be the same
// as fixed reg's next kill location. Otherwise, if source lcl var's next use
// is after the kill of fixed reg but before putarg_reg's next use, fixed reg's
// kill would lead to spill of source but not the putarg_reg if it were treated
// as special.
if (srcInterval->isActive &&
genRegMask(srcInterval->physReg) == currentRefPosition->registerAssignment &&
currentInterval->getNextRefLocation() == physRegRecord->getNextRefLocation())
{
assert(physRegRecord->regNum == srcInterval->physReg);
// Special putarg_reg acts as a pass-thru since both source lcl var
// and putarg_reg have the same register allocated. Physical reg
// record of reg continue to point to source lcl var's interval
// instead of to putarg_reg's interval. So if a spill of reg
// allocated to source lcl var happens, to reallocate to another
// tree node, before its use at call node it will lead to spill of
// lcl var instead of putarg_reg since physical reg record is pointing
// to lcl var's interval. As a result, arg reg would get trashed leading
// to bad codegen. The assumption here is that source lcl var of a
// special putarg_reg doesn't get spilled and re-allocated prior to
// its use at the call node. This is ensured by marking physical reg
// record as busy until next kill.
physRegRecord->isBusyUntilNextKill = true;
}
else
{
currentInterval->isSpecialPutArg = false;
}
}
// If this is still a SpecialPutArg, continue;
if (currentInterval->isSpecialPutArg)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_SPECIAL_PUTARG, currentInterval,
currentRefPosition->assignedReg()));
continue;
}
}
if (assignedRegister == REG_NA && RefTypeIsUse(refType))
{
currentRefPosition->reload = true;
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_RELOAD, currentInterval, assignedRegister));
}
}
regMaskTP assignedRegBit = RBM_NONE;
bool isInRegister = false;
if (assignedRegister != REG_NA)
{
isInRegister = true;
assignedRegBit = genRegMask(assignedRegister);
if (!currentInterval->isActive)
{
// If this is a use, it must have started the block on the stack, but the register
// was available for use so we kept the association.
if (RefTypeIsUse(refType))
{
assert(enregisterLocalVars);
assert(inVarToRegMaps[curBBNum][currentInterval->getVarIndex(compiler)] == REG_STK &&
previousRefPosition->nodeLocation <= curBBStartLocation);
isInRegister = false;
}
else
{
currentInterval->isActive = true;
}
}
assert(currentInterval->assignedReg != nullptr &&
currentInterval->assignedReg->regNum == assignedRegister &&
currentInterval->assignedReg->assignedInterval == currentInterval);
}
// If this is a physical register, we unconditionally assign it to itself!
if (currentRefPosition->isPhysRegRef)
{
RegRecord* currentReg = currentRefPosition->getReg();
Interval* assignedInterval = currentReg->assignedInterval;
if (assignedInterval != nullptr)
{
unassignPhysReg(currentReg, assignedInterval->recentRefPosition);
}
currentReg->isActive = true;
assignedRegister = currentReg->regNum;
assignedRegBit = genRegMask(assignedRegister);
if (refType == RefTypeKill)
{
currentReg->isBusyUntilNextKill = false;
}
}
else if (previousRefPosition != nullptr)
{
assert(previousRefPosition->nextRefPosition == currentRefPosition);
assert(assignedRegister == REG_NA || assignedRegBit == previousRefPosition->registerAssignment ||
currentRefPosition->outOfOrder || previousRefPosition->copyReg ||
previousRefPosition->refType == RefTypeExpUse || currentRefPosition->refType == RefTypeDummyDef);
}
else if (assignedRegister != REG_NA)
{
// Handle the case where this is a preassigned register (i.e. parameter).
// We don't want to actually use the preassigned register if it's not
// going to cover the lifetime - but we had to preallocate it to ensure
// that it remained live.
// TODO-CQ: At some point we may want to refine the analysis here, in case
// it might be beneficial to keep it in this reg for PART of the lifetime
if (currentInterval->isLocalVar)
{
regMaskTP preferences = currentInterval->registerPreferences;
bool keepAssignment = true;
bool matchesPreferences = (preferences & genRegMask(assignedRegister)) != RBM_NONE;
// Will the assigned register cover the lifetime? If not, does it at least
// meet the preferences for the next RefPosition?
RegRecord* physRegRecord = getRegisterRecord(currentInterval->physReg);
RefPosition* nextPhysRegRefPos = physRegRecord->getNextRefPosition();
if (nextPhysRegRefPos != nullptr &&
nextPhysRegRefPos->nodeLocation <= currentInterval->lastRefPosition->nodeLocation)
{
// Check to see if the existing assignment matches the preferences (e.g. callee save registers)
// and ensure that the next use of this localVar does not occur after the nextPhysRegRefPos
// There must be a next RefPosition, because we know that the Interval extends beyond the
// nextPhysRegRefPos.
RefPosition* nextLclVarRefPos = currentRefPosition->nextRefPosition;
assert(nextLclVarRefPos != nullptr);
if (!matchesPreferences || nextPhysRegRefPos->nodeLocation < nextLclVarRefPos->nodeLocation ||
physRegRecord->conflictingFixedRegReference(nextLclVarRefPos))
{
keepAssignment = false;
}
}
else if (refType == RefTypeParamDef && !matchesPreferences)
{
// Don't use the register, even if available, if it doesn't match the preferences.
// Note that this case is only for ParamDefs, for which we haven't yet taken preferences
// into account (we've just automatically got the initial location). In other cases,
// we would already have put it in a preferenced register, if it was available.
// TODO-CQ: Consider expanding this to check availability - that would duplicate
// code here, but otherwise we may wind up in this register anyway.
keepAssignment = false;
}
if (keepAssignment == false)
{
currentRefPosition->registerAssignment = allRegs(currentInterval->registerType);
unassignPhysRegNoSpill(physRegRecord);
// If the preferences are currently set to just this register, reset them to allRegs
// of the appropriate type (just as we just reset the registerAssignment for this
// RefPosition.
// Otherwise, simply remove this register from the preferences, if it's there.
if (currentInterval->registerPreferences == assignedRegBit)
{
currentInterval->registerPreferences = currentRefPosition->registerAssignment;
}
else
{
currentInterval->registerPreferences &= ~assignedRegBit;
}
assignedRegister = REG_NA;
assignedRegBit = RBM_NONE;
}
}
}
if (assignedRegister != REG_NA)
{
// If there is a conflicting fixed reference, insert a copy.
RegRecord* physRegRecord = getRegisterRecord(assignedRegister);
if (physRegRecord->conflictingFixedRegReference(currentRefPosition))
{
// We may have already reassigned the register to the conflicting reference.
// If not, we need to unassign this interval.
if (physRegRecord->assignedInterval == currentInterval)
{
unassignPhysRegNoSpill(physRegRecord);
}
currentRefPosition->moveReg = true;
assignedRegister = REG_NA;
setIntervalAsSplit(currentInterval);
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_MOVE_REG, currentInterval, assignedRegister));
}
else if ((genRegMask(assignedRegister) & currentRefPosition->registerAssignment) != 0)
{
currentRefPosition->registerAssignment = assignedRegBit;
if (!currentReferent->isActive)
{
// If we've got an exposed use at the top of a block, the
// interval might not have been active. Otherwise if it's a use,
// the interval must be active.
if (refType == RefTypeDummyDef)
{
currentReferent->isActive = true;
assert(getRegisterRecord(assignedRegister)->assignedInterval == currentInterval);
}
else
{
currentRefPosition->reload = true;
}
}
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, currentInterval, assignedRegister));
}
else
{
assert(currentInterval != nullptr);
// It's already in a register, but not one we need.
if (!RefTypeIsDef(currentRefPosition->refType))
{
regNumber copyReg = assignCopyReg(currentRefPosition);
assert(copyReg != REG_NA);
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_COPY_REG, currentInterval, copyReg));
lastAllocatedRefPosition = currentRefPosition;
if (currentRefPosition->lastUse)
{
if (currentRefPosition->delayRegFree)
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE_DELAYED, currentInterval,
assignedRegister));
delayRegsToFree |= (genRegMask(assignedRegister) | currentRefPosition->registerAssignment);
}
else
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE, currentInterval, assignedRegister));
regsToFree |= (genRegMask(assignedRegister) | currentRefPosition->registerAssignment);
}
}
// If this is a tree temp (non-localVar) interval, we will need an explicit move.
if (!currentInterval->isLocalVar)
{
currentRefPosition->moveReg = true;
currentRefPosition->copyReg = false;
}
continue;
}
else
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NEEDS_NEW_REG, nullptr, assignedRegister));
regsToFree |= genRegMask(assignedRegister);
// We want a new register, but we don't want this to be considered a spill.
assignedRegister = REG_NA;
if (physRegRecord->assignedInterval == currentInterval)
{
unassignPhysRegNoSpill(physRegRecord);
}
}
}
}
if (assignedRegister == REG_NA)
{
bool allocateReg = true;
if (currentRefPosition->AllocateIfProfitable())
{
// We can avoid allocating a register if it is a the last use requiring a reload.
if (currentRefPosition->lastUse && currentRefPosition->reload)
{
allocateReg = false;
}
#ifdef DEBUG
// Under stress mode, don't attempt to allocate a reg to
// reg optional ref position.
if (allocateReg && regOptionalNoAlloc())
{
allocateReg = false;
}
#endif
}
if (allocateReg)
{
// Try to allocate a register
assignedRegister = tryAllocateFreeReg(currentInterval, currentRefPosition);
}
// If no register was found, and if the currentRefPosition must have a register,
// then find a register to spill
if (assignedRegister == REG_NA)
{
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (refType == RefTypeUpperVectorSaveDef)
{
// TODO-CQ: Determine whether copying to two integer callee-save registers would be profitable.
// SaveDef position occurs after the Use of args and at the same location as Kill/Def
// positions of a call node. But SaveDef position cannot use any of the arg regs as
// they are needed for call node.
currentRefPosition->registerAssignment =
(allRegs(TYP_FLOAT) & RBM_FLT_CALLEE_TRASH & ~RBM_FLTARG_REGS);
assignedRegister = tryAllocateFreeReg(currentInterval, currentRefPosition);
// There MUST be caller-save registers available, because they have all just been killed.
// Amd64 Windows: xmm4-xmm5 are guaranteed to be available as xmm0-xmm3 are used for passing args.
// Amd64 Unix: xmm8-xmm15 are guaranteed to be avilable as xmm0-xmm7 are used for passing args.
// X86 RyuJIT Windows: xmm4-xmm7 are guanrateed to be available.
assert(assignedRegister != REG_NA);
// Now, spill it.
// Note:
// i) The reason we have to spill is that SaveDef position is allocated after the Kill positions
// of the call node are processed. Since callee-trash registers are killed by call node
// we explicity spill and unassign the register.
// ii) These will look a bit backward in the dump, but it's a pain to dump the alloc before the
// spill).
unassignPhysReg(getRegisterRecord(assignedRegister), currentRefPosition);
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, currentInterval, assignedRegister));
// Now set assignedRegister to REG_NA again so that we don't re-activate it.
assignedRegister = REG_NA;
}
else
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (currentRefPosition->RequiresRegister() || currentRefPosition->AllocateIfProfitable())
{
if (allocateReg)
{
assignedRegister = allocateBusyReg(currentInterval, currentRefPosition,
currentRefPosition->AllocateIfProfitable());
}
if (assignedRegister != REG_NA)
{
INDEBUG(
dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_SPILLED_REG, currentInterval, assignedRegister));
}
else
{
// This can happen only for those ref positions that are to be allocated
// only if profitable.
noway_assert(currentRefPosition->AllocateIfProfitable());
currentRefPosition->registerAssignment = RBM_NONE;
currentRefPosition->reload = false;
setIntervalAsSpilled(currentInterval);
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval));
}
}
else
{
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval));
currentRefPosition->registerAssignment = RBM_NONE;
currentInterval->isActive = false;
setIntervalAsSpilled(currentInterval);
}
}
#ifdef DEBUG
else
{
if (VERBOSE)
{
if (currentInterval->isConstant && (currentRefPosition->treeNode != nullptr) &&
currentRefPosition->treeNode->IsReuseRegVal())
{
dumpLsraAllocationEvent(LSRA_EVENT_REUSE_REG, nullptr, assignedRegister, currentBlock);
}
else
{
dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, nullptr, assignedRegister, currentBlock);
}
}
}
#endif // DEBUG
if (refType == RefTypeDummyDef && assignedRegister != REG_NA)
{
setInVarRegForBB(curBBNum, currentInterval->varNum, assignedRegister);
}
// If we allocated a register, and this is a use of a spilled value,
// it should have been marked for reload above.
if (assignedRegister != REG_NA && RefTypeIsUse(refType) && !isInRegister)
{
assert(currentRefPosition->reload);
}
}
// If we allocated a register, record it
if (currentInterval != nullptr && assignedRegister != REG_NA)
{
assignedRegBit = genRegMask(assignedRegister);
currentRefPosition->registerAssignment = assignedRegBit;
currentInterval->physReg = assignedRegister;
regsToFree &= ~assignedRegBit; // we'll set it again later if it's dead
// If this interval is dead, free the register.
// The interval could be dead if this is a user variable, or if the
// node is being evaluated for side effects, or a call whose result
// is not used, etc.
if (currentRefPosition->lastUse || currentRefPosition->nextRefPosition == nullptr)
{
assert(currentRefPosition->isIntervalRef());
if (refType != RefTypeExpUse && currentRefPosition->nextRefPosition == nullptr)
{
if (currentRefPosition->delayRegFree)
{
delayRegsToFree |= assignedRegBit;
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE_DELAYED));
}
else
{
regsToFree |= assignedRegBit;
INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE));
}
}
else
{
currentInterval->isActive = false;
}
}
lastAllocatedRefPosition = currentRefPosition;
}
}
// Free registers to clear associated intervals for resolution phase
CLANG_FORMAT_COMMENT_ANCHOR;
#ifdef DEBUG
if (getLsraExtendLifeTimes())
{
// If we have extended lifetimes, we need to make sure all the registers are freed.
for (int regNumIndex = 0; regNumIndex <= REG_FP_LAST; regNumIndex++)
{
RegRecord& regRecord = physRegs[regNumIndex];
Interval* interval = regRecord.assignedInterval;
if (interval != nullptr)
{
interval->isActive = false;
unassignPhysReg(®Record, nullptr);
}
}
}
else
#endif // DEBUG
{
freeRegisters(regsToFree | delayRegsToFree);
}
#ifdef DEBUG
if (VERBOSE)
{
if (dumpTerse)
{
// Dump the RegRecords after the last RefPosition is handled.
dumpRegRecords();
printf("\n");
}
dumpRefPositions("AFTER ALLOCATION");
dumpVarRefPositions("AFTER ALLOCATION");
// Dump the intervals that remain active
printf("Active intervals at end of allocation:\n");
// We COULD just reuse the intervalIter from above, but ArrayListIterator doesn't
// provide a Reset function (!) - we'll probably replace this so don't bother
// adding it
for (auto& interval : intervals)
{
if (interval.isActive)
{
printf("Active ");
interval.dump();
}
}
printf("\n");
}
#endif // DEBUG
}
//-----------------------------------------------------------------------------
// updateAssignedInterval: Update assigned interval of register.
//
// Arguments:
// reg - register to be updated
// interval - interval to be assigned
// regType - regsiter type
//
// Return Value:
// None
//
// Assumptions:
// For ARM32, when "regType" is TYP_DOUBLE, "reg" should be a even-numbered
// float register, i.e. lower half of double register.
//
// Note:
// For ARM32, two float registers consisting a double register are updated
// together when "regType" is TYP_DOUBLE.
//
void LinearScan::updateAssignedInterval(RegRecord* reg, Interval* interval, RegisterType regType)
{
reg->assignedInterval = interval;
#ifdef _TARGET_ARM_
// Update overlapping floating point register for TYP_DOUBLE
if (regType == TYP_DOUBLE)
{
RegRecord* anotherHalfReg = findAnotherHalfRegRec(reg);
anotherHalfReg->assignedInterval = interval;
}
#endif
}
//-----------------------------------------------------------------------------
// updatePreviousInterval: Update previous interval of register.
//
// Arguments:
// reg - register to be updated
// interval - interval to be assigned
// regType - regsiter type
//
// Return Value:
// None
//
// Assumptions:
// For ARM32, when "regType" is TYP_DOUBLE, "reg" should be a even-numbered
// float register, i.e. lower half of double register.
//
// Note:
// For ARM32, two float registers consisting a double register are updated
// together when "regType" is TYP_DOUBLE.
//
void LinearScan::updatePreviousInterval(RegRecord* reg, Interval* interval, RegisterType regType)
{
reg->previousInterval = interval;
#ifdef _TARGET_ARM_
// Update overlapping floating point register for TYP_DOUBLE
if (regType == TYP_DOUBLE)
{
RegRecord* anotherHalfReg = findAnotherHalfRegRec(reg);
anotherHalfReg->previousInterval = interval;
}
#endif
}
// LinearScan::resolveLocalRef
// Description:
// Update the graph for a local reference.
// Also, track the register (if any) that is currently occupied.
// Arguments:
// treeNode: The lclVar that's being resolved
// currentRefPosition: the RefPosition associated with the treeNode
//
// Details:
// This method is called for each local reference, during the resolveRegisters
// phase of LSRA. It is responsible for keeping the following in sync:
// - varDsc->lvRegNum (and lvOtherReg) contain the unique register location.
// If it is not in the same register through its lifetime, it is set to REG_STK.
// - interval->physReg is set to the assigned register
// (i.e. at the code location which is currently being handled by resolveRegisters())
// - interval->isActive is true iff the interval is live and occupying a register
// - interval->isSpilled should have already been set to true if the interval is EVER spilled
// - interval->isSplit is set to true if the interval does not occupy the same
// register throughout the method
// - RegRecord->assignedInterval points to the interval which currently occupies
// the register
// - For each lclVar node:
// - gtRegNum/gtRegPair is set to the currently allocated register(s).
// - GTF_SPILLED is set on a use if it must be reloaded prior to use.
// - GTF_SPILL is set if it must be spilled after use.
//
// A copyReg is an ugly case where the variable must be in a specific (fixed) register,
// but it currently resides elsewhere. The register allocator must track the use of the
// fixed register, but it marks the lclVar node with the register it currently lives in
// and the code generator does the necessary move.
//
// Before beginning, the varDsc for each parameter must be set to its initial location.
//
// NICE: Consider tracking whether an Interval is always in the same location (register/stack)
// in which case it will require no resolution.
//
void LinearScan::resolveLocalRef(BasicBlock* block, GenTreePtr treeNode, RefPosition* currentRefPosition)
{
assert((block == nullptr) == (treeNode == nullptr));
assert(enregisterLocalVars);
// Is this a tracked local? Or just a register allocated for loading
// a non-tracked one?
Interval* interval = currentRefPosition->getInterval();
if (!interval->isLocalVar)
{
return;
}
interval->recentRefPosition = currentRefPosition;
LclVarDsc* varDsc = interval->getLocalVar(compiler);
// NOTE: we set the GTF_VAR_DEATH flag here unless we are extending lifetimes, in which case we write
// this bit in checkLastUses. This is a bit of a hack, but is necessary because codegen requires
// accurate last use info that is not reflected in the lastUse bit on ref positions when we are extending
// lifetimes. See also the comments in checkLastUses.
if ((treeNode != nullptr) && !extendLifetimes())
{
if (currentRefPosition->lastUse)
{
treeNode->gtFlags |= GTF_VAR_DEATH;
}
else
{
treeNode->gtFlags &= ~GTF_VAR_DEATH;
}
}
if (currentRefPosition->registerAssignment == RBM_NONE)
{
assert(!currentRefPosition->RequiresRegister());
assert(interval->isSpilled);
varDsc->lvRegNum = REG_STK;
if (interval->assignedReg != nullptr && interval->assignedReg->assignedInterval == interval)
{
updateAssignedInterval(interval->assignedReg, nullptr, interval->registerType);
}
interval->assignedReg = nullptr;
interval->physReg = REG_NA;
if (treeNode != nullptr)
{
treeNode->SetContained();
}
return;
}
// In most cases, assigned and home registers will be the same
// The exception is the copyReg case, where we've assigned a register
// for a specific purpose, but will be keeping the register assignment
regNumber assignedReg = currentRefPosition->assignedReg();
regNumber homeReg = assignedReg;
// Undo any previous association with a physical register, UNLESS this
// is a copyReg
if (!currentRefPosition->copyReg)
{
regNumber oldAssignedReg = interval->physReg;
if (oldAssignedReg != REG_NA && assignedReg != oldAssignedReg)
{
RegRecord* oldRegRecord = getRegisterRecord(oldAssignedReg);
if (oldRegRecord->assignedInterval == interval)
{
updateAssignedInterval(oldRegRecord, nullptr, interval->registerType);
}
}
}
if (currentRefPosition->refType == RefTypeUse && !currentRefPosition->reload)
{
// Was this spilled after our predecessor was scheduled?
if (interval->physReg == REG_NA)
{
assert(inVarToRegMaps[curBBNum][varDsc->lvVarIndex] == REG_STK);
currentRefPosition->reload = true;
}
}
bool reload = currentRefPosition->reload;
bool spillAfter = currentRefPosition->spillAfter;
// In the reload case we either:
// - Set the register to REG_STK if it will be referenced only from the home location, or
// - Set the register to the assigned register and set GTF_SPILLED if it must be loaded into a register.
if (reload)
{
assert(currentRefPosition->refType != RefTypeDef);
assert(interval->isSpilled);
varDsc->lvRegNum = REG_STK;
if (!spillAfter)
{
interval->physReg = assignedReg;
}
// If there is no treeNode, this must be a RefTypeExpUse, in
// which case we did the reload already
if (treeNode != nullptr)
{
treeNode->gtFlags |= GTF_SPILLED;
if (spillAfter)
{
if (currentRefPosition->AllocateIfProfitable())
{
// This is a use of lclVar that is flagged as reg-optional
// by lower/codegen and marked for both reload and spillAfter.
// In this case we can avoid unnecessary reload and spill
// by setting reg on lclVar to REG_STK and reg on tree node
// to REG_NA. Codegen will generate the code by considering
// it as a contained memory operand.
//
// Note that varDsc->lvRegNum is already to REG_STK above.
interval->physReg = REG_NA;
treeNode->gtRegNum = REG_NA;
treeNode->gtFlags &= ~GTF_SPILLED;
treeNode->SetContained();
}
else
{
treeNode->gtFlags |= GTF_SPILL;
}
}
}
else
{
assert(currentRefPosition->refType == RefTypeExpUse);
}
}
else if (spillAfter && !RefTypeIsUse(currentRefPosition->refType))
{
// In the case of a pure def, don't bother spilling - just assign it to the
// stack. However, we need to remember that it was spilled.
assert(interval->isSpilled);
varDsc->lvRegNum = REG_STK;
interval->physReg = REG_NA;
if (treeNode != nullptr)
{
treeNode->gtRegNum = REG_NA;
}
}
else
{
// Not reload and Not pure-def that's spillAfter
if (currentRefPosition->copyReg || currentRefPosition->moveReg)
{
// For a copyReg or moveReg, we have two cases:
// - In the first case, we have a fixedReg - i.e. a register which the code
// generator is constrained to use.
// The code generator will generate the appropriate move to meet the requirement.
// - In the second case, we were forced to use a different register because of
// interference (or JitStressRegs).
// In this case, we generate a GT_COPY.
// In either case, we annotate the treeNode with the register in which the value
// currently lives. For moveReg, the homeReg is the new register (as assigned above).
// But for copyReg, the homeReg remains unchanged.
assert(treeNode != nullptr);
treeNode->gtRegNum = interval->physReg;
if (currentRefPosition->copyReg)
{
homeReg = interval->physReg;
}
else
{
assert(interval->isSplit);
interval->physReg = assignedReg;
}
if (!currentRefPosition->isFixedRegRef || currentRefPosition->moveReg)
{
// This is the second case, where we need to generate a copy
insertCopyOrReload(block, treeNode, currentRefPosition->getMultiRegIdx(), currentRefPosition);
}
}
else
{
interval->physReg = assignedReg;
if (!interval->isSpilled && !interval->isSplit)
{
if (varDsc->lvRegNum != REG_STK)
{
// If the register assignments don't match, then this interval is split.
if (varDsc->lvRegNum != assignedReg)
{
setIntervalAsSplit(interval);
varDsc->lvRegNum = REG_STK;
}
}
else
{
varDsc->lvRegNum = assignedReg;
}
}
}
if (spillAfter)
{
if (treeNode != nullptr)
{
treeNode->gtFlags |= GTF_SPILL;
}
assert(interval->isSpilled);
interval->physReg = REG_NA;
varDsc->lvRegNum = REG_STK;
}
}
// Update the physRegRecord for the register, so that we know what vars are in
// regs at the block boundaries
RegRecord* physRegRecord = getRegisterRecord(homeReg);
if (spillAfter || currentRefPosition->lastUse)
{
interval->isActive = false;
interval->assignedReg = nullptr;
interval->physReg = REG_NA;
updateAssignedInterval(physRegRecord, nullptr, interval->registerType);
}
else
{
interval->isActive = true;
interval->assignedReg = physRegRecord;
updateAssignedInterval(physRegRecord, interval, interval->registerType);
}
}
void LinearScan::writeRegisters(RefPosition* currentRefPosition, GenTree* tree)
{
lsraAssignRegToTree(tree, currentRefPosition->assignedReg(), currentRefPosition->getMultiRegIdx());
}
//------------------------------------------------------------------------
// insertCopyOrReload: Insert a copy in the case where a tree node value must be moved
// to a different register at the point of use (GT_COPY), or it is reloaded to a different register
// than the one it was spilled from (GT_RELOAD).
//
// Arguments:
// block - basic block in which GT_COPY/GT_RELOAD is inserted.
// tree - This is the node to copy or reload.
// Insert copy or reload node between this node and its parent.
// multiRegIdx - register position of tree node for which copy or reload is needed.
// refPosition - The RefPosition at which copy or reload will take place.
//
// Notes:
// The GT_COPY or GT_RELOAD will be inserted in the proper spot in execution order where the reload is to occur.
//
// For example, for this tree (numbers are execution order, lower is earlier and higher is later):
//
// +---------+----------+
// | GT_ADD (3) |
// +---------+----------+
// |
// / \
// / \
// / \
// +-------------------+ +----------------------+
// | x (1) | "tree" | y (2) |
// +-------------------+ +----------------------+
//
// generate this tree:
//
// +---------+----------+
// | GT_ADD (4) |
// +---------+----------+
// |
// / \
// / \
// / \
// +-------------------+ +----------------------+
// | GT_RELOAD (3) | | y (2) |
// +-------------------+ +----------------------+
// |
// +-------------------+
// | x (1) | "tree"
// +-------------------+
//
// Note in particular that the GT_RELOAD node gets inserted in execution order immediately before the parent of "tree",
// which seems a bit weird since normally a node's parent (in this case, the parent of "x", GT_RELOAD in the "after"
// picture) immediately follows all of its children (that is, normally the execution ordering is postorder).
// The ordering must be this weird "out of normal order" way because the "x" node is being spilled, probably
// because the expression in the tree represented above by "y" has high register requirements. We don't want
// to reload immediately, of course. So we put GT_RELOAD where the reload should actually happen.
//
// Note that GT_RELOAD is required when we reload to a different register than the one we spilled to. It can also be
// used if we reload to the same register. Normally, though, in that case we just mark the node with GTF_SPILLED,
// and the unspilling code automatically reuses the same register, and does the reload when it notices that flag
// when considering a node's operands.
//
void LinearScan::insertCopyOrReload(BasicBlock* block, GenTreePtr tree, unsigned multiRegIdx, RefPosition* refPosition)
{
LIR::Range& blockRange = LIR::AsRange(block);
LIR::Use treeUse;
bool foundUse = blockRange.TryGetUse(tree, &treeUse);
assert(foundUse);
GenTree* parent = treeUse.User();
genTreeOps oper;
if (refPosition->reload)
{
oper = GT_RELOAD;
}
else
{
oper = GT_COPY;
#if TRACK_LSRA_STATS
updateLsraStat(LSRA_STAT_COPY_REG, block->bbNum);
#endif
}
// If the parent is a reload/copy node, then tree must be a multi-reg call node
// that has already had one of its registers spilled. This is Because multi-reg
// call node is the only node whose RefTypeDef positions get independently
// spilled or reloaded. It is possible that one of its RefTypeDef position got
// spilled and the next use of it requires it to be in a different register.
//
// In this case set the ith position reg of reload/copy node to the reg allocated
// for copy/reload refPosition. Essentially a copy/reload node will have a reg
// for each multi-reg position of its child. If there is a valid reg in ith
// position of GT_COPY or GT_RELOAD node then the corresponding result of its
// child needs to be copied or reloaded to that reg.
if (parent->IsCopyOrReload())
{
noway_assert(parent->OperGet() == oper);
noway_assert(tree->IsMultiRegCall());
GenTreeCall* call = tree->AsCall();
GenTreeCopyOrReload* copyOrReload = parent->AsCopyOrReload();
noway_assert(copyOrReload->GetRegNumByIdx(multiRegIdx) == REG_NA);
copyOrReload->SetRegNumByIdx(refPosition->assignedReg(), multiRegIdx);
}
else
{
// Create the new node, with "tree" as its only child.
var_types treeType = tree->TypeGet();
GenTreeCopyOrReload* newNode = new (compiler, oper) GenTreeCopyOrReload(oper, treeType, tree);
assert(refPosition->registerAssignment != RBM_NONE);
newNode->SetRegNumByIdx(refPosition->assignedReg(), multiRegIdx);
newNode->gtLsraInfo.isLsraAdded = true;
newNode->gtLsraInfo.isLocalDefUse = false;
if (refPosition->copyReg)
{
// This is a TEMPORARY copy
assert(isCandidateLocalRef(tree));
newNode->gtFlags |= GTF_VAR_DEATH;
}
// Insert the copy/reload after the spilled node and replace the use of the original node with a use
// of the copy/reload.
blockRange.InsertAfter(tree, newNode);
treeUse.ReplaceWith(compiler, newNode);
}
}
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
//------------------------------------------------------------------------
// insertUpperVectorSaveAndReload: Insert code to save and restore the upper half of a vector that lives
// in a callee-save register at the point of a kill (the upper half is
// not preserved).
//
// Arguments:
// tree - This is the node around which we will insert the Save & Reload.
// It will be a call or some node that turns into a call.
// refPosition - The RefTypeUpperVectorSaveDef RefPosition.
//
void LinearScan::insertUpperVectorSaveAndReload(GenTreePtr tree, RefPosition* refPosition, BasicBlock* block)
{
Interval* lclVarInterval = refPosition->getInterval()->relatedInterval;
assert(lclVarInterval->isLocalVar == true);
LclVarDsc* varDsc = compiler->lvaTable + lclVarInterval->varNum;
assert(varDsc->lvType == LargeVectorType);
regNumber lclVarReg = lclVarInterval->physReg;
if (lclVarReg == REG_NA)
{
return;
}
assert((genRegMask(lclVarReg) & RBM_FLT_CALLEE_SAVED) != RBM_NONE);
regNumber spillReg = refPosition->assignedReg();
bool spillToMem = refPosition->spillAfter;
LIR::Range& blockRange = LIR::AsRange(block);
// First, insert the save before the call.
GenTreePtr saveLcl = compiler->gtNewLclvNode(lclVarInterval->varNum, LargeVectorType);
saveLcl->gtLsraInfo.isLsraAdded = true;
saveLcl->gtRegNum = lclVarReg;
saveLcl->gtLsraInfo.isLocalDefUse = false;
GenTreeSIMD* simdNode =
new (compiler, GT_SIMD) GenTreeSIMD(LargeVectorSaveType, saveLcl, nullptr, SIMDIntrinsicUpperSave,
varDsc->lvBaseType, genTypeSize(LargeVectorType));
simdNode->gtLsraInfo.isLsraAdded = true;
simdNode->gtRegNum = spillReg;
if (spillToMem)
{
simdNode->gtFlags |= GTF_SPILL;
}
blockRange.InsertBefore(tree, LIR::SeqTree(compiler, simdNode));
// Now insert the restore after the call.
GenTreePtr restoreLcl = compiler->gtNewLclvNode(lclVarInterval->varNum, LargeVectorType);
restoreLcl->gtLsraInfo.isLsraAdded = true;
restoreLcl->gtRegNum = lclVarReg;
restoreLcl->gtLsraInfo.isLocalDefUse = false;
simdNode = new (compiler, GT_SIMD)
GenTreeSIMD(LargeVectorType, restoreLcl, nullptr, SIMDIntrinsicUpperRestore, varDsc->lvBaseType, 32);
simdNode->gtLsraInfo.isLsraAdded = true;
simdNode->gtRegNum = spillReg;
if (spillToMem)
{
simdNode->gtFlags |= GTF_SPILLED;
}
blockRange.InsertAfter(tree, LIR::SeqTree(compiler, simdNode));
}
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
//------------------------------------------------------------------------
// initMaxSpill: Initializes the LinearScan members used to track the max number
// of concurrent spills. This is needed so that we can set the
// fields in Compiler, so that the code generator, in turn can
// allocate the right number of spill locations.
//
// Arguments:
// None.
//
// Return Value:
// None.
//
// Assumptions:
// This is called before any calls to updateMaxSpill().
void LinearScan::initMaxSpill()
{
needDoubleTmpForFPCall = false;
needFloatTmpForFPCall = false;
for (int i = 0; i < TYP_COUNT; i++)
{
maxSpill[i] = 0;
currentSpill[i] = 0;
}
}
//------------------------------------------------------------------------
// recordMaxSpill: Sets the fields in Compiler for the max number of concurrent spills.
// (See the comment on initMaxSpill.)
//
// Arguments:
// None.
//
// Return Value:
// None.
//
// Assumptions:
// This is called after updateMaxSpill() has been called for all "real"
// RefPositions.
void LinearScan::recordMaxSpill()
{
// Note: due to the temp normalization process (see tmpNormalizeType)
// only a few types should actually be seen here.
JITDUMP("Recording the maximum number of concurrent spills:\n");
#ifdef _TARGET_X86_
var_types returnType = compiler->tmpNormalizeType(compiler->info.compRetType);
if (needDoubleTmpForFPCall || (returnType == TYP_DOUBLE))
{
JITDUMP("Adding a spill temp for moving a double call/return value between xmm reg and x87 stack.\n");
maxSpill[TYP_DOUBLE] += 1;
}
if (needFloatTmpForFPCall || (returnType == TYP_FLOAT))
{
JITDUMP("Adding a spill temp for moving a float call/return value between xmm reg and x87 stack.\n");
maxSpill[TYP_FLOAT] += 1;
}
#endif // _TARGET_X86_
for (int i = 0; i < TYP_COUNT; i++)
{
if (var_types(i) != compiler->tmpNormalizeType(var_types(i)))
{
// Only normalized types should have anything in the maxSpill array.
// We assume here that if type 'i' does not normalize to itself, then
// nothing else normalizes to 'i', either.
assert(maxSpill[i] == 0);
}
if (maxSpill[i] != 0)
{
JITDUMP(" %s: %d\n", varTypeName(var_types(i)), maxSpill[i]);
compiler->tmpPreAllocateTemps(var_types(i), maxSpill[i]);
}
}
JITDUMP("\n");
}
//------------------------------------------------------------------------
// updateMaxSpill: Update the maximum number of concurrent spills
//
// Arguments:
// refPosition - the current RefPosition being handled
//
// Return Value:
// None.
//
// Assumptions:
// The RefPosition has an associated interval (getInterval() will
// otherwise assert).
//
// Notes:
// This is called for each "real" RefPosition during the writeback
// phase of LSRA. It keeps track of how many concurrently-live
// spills there are, and the largest number seen so far.
void LinearScan::updateMaxSpill(RefPosition* refPosition)
{
RefType refType = refPosition->refType;
if (refPosition->spillAfter || refPosition->reload ||
(refPosition->AllocateIfProfitable() && refPosition->assignedReg() == REG_NA))
{
Interval* interval = refPosition->getInterval();
if (!interval->isLocalVar)
{
// The tmp allocation logic 'normalizes' types to a small number of
// types that need distinct stack locations from each other.
// Those types are currently gc refs, byrefs, <= 4 byte non-GC items,
// 8-byte non-GC items, and 16-byte or 32-byte SIMD vectors.
// LSRA is agnostic to those choices but needs
// to know what they are here.
var_types typ;
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if ((refType == RefTypeUpperVectorSaveDef) || (refType == RefTypeUpperVectorSaveUse))
{
typ = LargeVectorSaveType;
}
else
#endif // !FEATURE_PARTIAL_SIMD_CALLEE_SAVE
{
GenTreePtr treeNode = refPosition->treeNode;
if (treeNode == nullptr)
{
assert(RefTypeIsUse(refType));
treeNode = interval->firstRefPosition->treeNode;
}
assert(treeNode != nullptr);
// In case of multi-reg call nodes, we need to use the type
// of the return register given by multiRegIdx of the refposition.
if (treeNode->IsMultiRegCall())
{
ReturnTypeDesc* retTypeDesc = treeNode->AsCall()->GetReturnTypeDesc();
typ = retTypeDesc->GetReturnRegType(refPosition->getMultiRegIdx());
}
#ifdef _TARGET_ARM_
else if (treeNode->OperIsPutArgSplit())
{
typ = treeNode->AsPutArgSplit()->GetRegType(refPosition->getMultiRegIdx());
}
else if (treeNode->OperIsPutArgReg())
{
// For double arg regs, the type is changed to long since they must be passed via `r0-r3`.
// However when they get spilled, they should be treated as separated int registers.
var_types typNode = treeNode->TypeGet();
typ = (typNode == TYP_LONG) ? TYP_INT : typNode;
}
#endif // _TARGET_ARM_
else
{
typ = treeNode->TypeGet();
}
typ = compiler->tmpNormalizeType(typ);
}
if (refPosition->spillAfter && !refPosition->reload)
{
currentSpill[typ]++;
if (currentSpill[typ] > maxSpill[typ])
{
maxSpill[typ] = currentSpill[typ];
}
}
else if (refPosition->reload)
{
assert(currentSpill[typ] > 0);
currentSpill[typ]--;
}
else if (refPosition->AllocateIfProfitable() && refPosition->assignedReg() == REG_NA)
{
// A spill temp not getting reloaded into a reg because it is
// marked as allocate if profitable and getting used from its
// memory location. To properly account max spill for typ we
// decrement spill count.
assert(RefTypeIsUse(refType));
assert(currentSpill[typ] > 0);
currentSpill[typ]--;
}
JITDUMP(" Max spill for %s is %d\n", varTypeName(typ), maxSpill[typ]);
}
}
}
// This is the final phase of register allocation. It writes the register assignments to
// the tree, and performs resolution across joins and backedges.
//
void LinearScan::resolveRegisters()
{
// Iterate over the tree and the RefPositions in lockstep
// - annotate the tree with register assignments by setting gtRegNum or gtRegPair (for longs)
// on the tree node
// - track globally-live var locations
// - add resolution points at split/merge/critical points as needed
// Need to use the same traversal order as the one that assigns the location numbers.
// Dummy RefPositions have been added at any split, join or critical edge, at the
// point where resolution may be required. These are located:
// - for a split, at the top of the non-adjacent block
// - for a join, at the bottom of the non-adjacent joining block
// - for a critical edge, at the top of the target block of each critical
// edge.
// Note that a target block may have multiple incoming critical or split edges
//
// These RefPositions record the expected location of the Interval at that point.
// At each branch, we identify the location of each liveOut interval, and check
// against the RefPositions at the target.
BasicBlock* block;
LsraLocation currentLocation = MinLocation;
// Clear register assignments - these will be reestablished as lclVar defs (including RefTypeParamDefs)
// are encountered.
if (enregisterLocalVars)
{
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
RegRecord* physRegRecord = getRegisterRecord(reg);
Interval* assignedInterval = physRegRecord->assignedInterval;
if (assignedInterval != nullptr)
{
assignedInterval->assignedReg = nullptr;
assignedInterval->physReg = REG_NA;
}
physRegRecord->assignedInterval = nullptr;
physRegRecord->recentRefPosition = nullptr;
}
// Clear "recentRefPosition" for lclVar intervals
for (unsigned varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++)
{
if (localVarIntervals[varIndex] != nullptr)
{
localVarIntervals[varIndex]->recentRefPosition = nullptr;
localVarIntervals[varIndex]->isActive = false;
}
else
{
assert(compiler->lvaTable[compiler->lvaTrackedToVarNum[varIndex]].lvLRACandidate == false);
}
}
}
// handle incoming arguments and special temps
auto currentRefPosition = refPositions.begin();
if (enregisterLocalVars)
{
VarToRegMap entryVarToRegMap = inVarToRegMaps[compiler->fgFirstBB->bbNum];
while (currentRefPosition != refPositions.end() &&
(currentRefPosition->refType == RefTypeParamDef || currentRefPosition->refType == RefTypeZeroInit))
{
Interval* interval = currentRefPosition->getInterval();
assert(interval != nullptr && interval->isLocalVar);
resolveLocalRef(nullptr, nullptr, currentRefPosition);
regNumber reg = REG_STK;
int varIndex = interval->getVarIndex(compiler);
if (!currentRefPosition->spillAfter && currentRefPosition->registerAssignment != RBM_NONE)
{
reg = currentRefPosition->assignedReg();
}
else
{
reg = REG_STK;
interval->isActive = false;
}
setVarReg(entryVarToRegMap, varIndex, reg);
++currentRefPosition;
}
}
else
{
assert(currentRefPosition == refPositions.end() ||
(currentRefPosition->refType != RefTypeParamDef && currentRefPosition->refType != RefTypeZeroInit));
}
BasicBlock* insertionBlock = compiler->fgFirstBB;
GenTreePtr insertionPoint = LIR::AsRange(insertionBlock).FirstNonPhiNode();
// write back assignments
for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock())
{
assert(curBBNum == block->bbNum);
if (enregisterLocalVars)
{
// Record the var locations at the start of this block.
// (If it's fgFirstBB, we've already done that above, see entryVarToRegMap)
curBBStartLocation = currentRefPosition->nodeLocation;
if (block != compiler->fgFirstBB)
{
processBlockStartLocations(block, false);
}
// Handle the DummyDefs, updating the incoming var location.
for (; currentRefPosition != refPositions.end() && currentRefPosition->refType == RefTypeDummyDef;
++currentRefPosition)
{
assert(currentRefPosition->isIntervalRef());
// Don't mark dummy defs as reload
currentRefPosition->reload = false;
resolveLocalRef(nullptr, nullptr, currentRefPosition);
regNumber reg;
if (currentRefPosition->registerAssignment != RBM_NONE)
{
reg = currentRefPosition->assignedReg();
}
else
{
reg = REG_STK;
currentRefPosition->getInterval()->isActive = false;
}
setInVarRegForBB(curBBNum, currentRefPosition->getInterval()->varNum, reg);
}
}
// The next RefPosition should be for the block. Move past it.
assert(currentRefPosition != refPositions.end());
assert(currentRefPosition->refType == RefTypeBB);
++currentRefPosition;
// Handle the RefPositions for the block
for (; currentRefPosition != refPositions.end() && currentRefPosition->refType != RefTypeBB &&
currentRefPosition->refType != RefTypeDummyDef;
++currentRefPosition)
{
currentLocation = currentRefPosition->nodeLocation;
// Ensure that the spill & copy info is valid.
// First, if it's reload, it must not be copyReg or moveReg
assert(!currentRefPosition->reload || (!currentRefPosition->copyReg && !currentRefPosition->moveReg));
// If it's copyReg it must not be moveReg, and vice-versa
assert(!currentRefPosition->copyReg || !currentRefPosition->moveReg);
switch (currentRefPosition->refType)
{
#ifdef FEATURE_SIMD
case RefTypeUpperVectorSaveUse:
case RefTypeUpperVectorSaveDef:
#endif // FEATURE_SIMD
case RefTypeUse:
case RefTypeDef:
// These are the ones we're interested in
break;
case RefTypeKill:
case RefTypeFixedReg:
// These require no handling at resolution time
assert(currentRefPosition->referent != nullptr);
currentRefPosition->referent->recentRefPosition = currentRefPosition;
continue;
case RefTypeExpUse:
// Ignore the ExpUse cases - a RefTypeExpUse would only exist if the
// variable is dead at the entry to the next block. So we'll mark
// it as in its current location and resolution will take care of any
// mismatch.
assert(getNextBlock() == nullptr ||
!VarSetOps::IsMember(compiler, getNextBlock()->bbLiveIn,
currentRefPosition->getInterval()->getVarIndex(compiler)));
currentRefPosition->referent->recentRefPosition = currentRefPosition;
continue;
case RefTypeKillGCRefs:
// No action to take at resolution time, and no interval to update recentRefPosition for.
continue;
case RefTypeDummyDef:
case RefTypeParamDef:
case RefTypeZeroInit:
// Should have handled all of these already
default:
unreached();
break;
}
updateMaxSpill(currentRefPosition);
GenTree* treeNode = currentRefPosition->treeNode;
#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE
if (currentRefPosition->refType == RefTypeUpperVectorSaveDef)
{
// The treeNode must be a call, and this must be a RefPosition for a LargeVectorType LocalVar.
// If the LocalVar is in a callee-save register, we are going to spill its upper half around the call.
// If we have allocated a register to spill it to, we will use that; otherwise, we will spill it
// to the stack. We can use as a temp register any non-arg caller-save register.
noway_assert(treeNode != nullptr);
currentRefPosition->referent->recentRefPosition = currentRefPosition;
insertUpperVectorSaveAndReload(treeNode, currentRefPosition, block);
}
else if (currentRefPosition->refType == RefTypeUpperVectorSaveUse)
{
continue;
}
#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE
// Most uses won't actually need to be recorded (they're on the def).
// In those cases, treeNode will be nullptr.
if (treeNode == nullptr)
{
// This is either a use, a dead def, or a field of a struct
Interval* interval = currentRefPosition->getInterval();
assert(currentRefPosition->refType == RefTypeUse ||
currentRefPosition->registerAssignment == RBM_NONE || interval->isStructField);
// TODO-Review: Need to handle the case where any of the struct fields
// are reloaded/spilled at this use
assert(!interval->isStructField ||
(currentRefPosition->reload == false && currentRefPosition->spillAfter == false));
if (interval->isLocalVar && !interval->isStructField)
{
LclVarDsc* varDsc = interval->getLocalVar(compiler);
// This must be a dead definition. We need to mark the lclVar
// so that it's not considered a candidate for lvRegister, as
// this dead def will have to go to the stack.
assert(currentRefPosition->refType == RefTypeDef);
varDsc->lvRegNum = REG_STK;
}
continue;
}
LsraLocation loc = treeNode->gtLsraInfo.loc;
assert(treeNode->IsLocal() || currentLocation == loc || currentLocation == loc + 1);
if (currentRefPosition->isIntervalRef() && currentRefPosition->getInterval()->isInternal)
{
treeNode->gtRsvdRegs |= currentRefPosition->registerAssignment;
}
else
{
writeRegisters(currentRefPosition, treeNode);
if (treeNode->IsLocal() && currentRefPosition->getInterval()->isLocalVar)
{
resolveLocalRef(block, treeNode, currentRefPosition);
}
// Mark spill locations on temps
// (local vars are handled in resolveLocalRef, above)
// Note that the tree node will be changed from GTF_SPILL to GTF_SPILLED
// in codegen, taking care of the "reload" case for temps
else if (currentRefPosition->spillAfter || (currentRefPosition->nextRefPosition != nullptr &&
currentRefPosition->nextRefPosition->moveReg))
{
if (treeNode != nullptr && currentRefPosition->isIntervalRef())
{
if (currentRefPosition->spillAfter)
{
treeNode->gtFlags |= GTF_SPILL;
// If this is a constant interval that is reusing a pre-existing value, we actually need
// to generate the value at this point in order to spill it.
if (treeNode->IsReuseRegVal())
{
treeNode->ResetReuseRegVal();
}
// In case of multi-reg call node, also set spill flag on the
// register specified by multi-reg index of current RefPosition.
// Note that the spill flag on treeNode indicates that one or
// more its allocated registers are in that state.
if (treeNode->IsMultiRegCall())
{
GenTreeCall* call = treeNode->AsCall();
call->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx());
}
#ifdef _TARGET_ARM_
else if (treeNode->OperIsPutArgSplit())
{
GenTreePutArgSplit* splitArg = treeNode->AsPutArgSplit();
splitArg->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx());
}
else if (treeNode->OperIsMultiRegOp())
{
GenTreeMultiRegOp* multiReg = treeNode->AsMultiRegOp();
multiReg->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx());
}
#endif
}
// If the value is reloaded or moved to a different register, we need to insert
// a node to hold the register to which it should be reloaded
RefPosition* nextRefPosition = currentRefPosition->nextRefPosition;
assert(nextRefPosition != nullptr);
if (INDEBUG(alwaysInsertReload() ||)
nextRefPosition->assignedReg() != currentRefPosition->assignedReg())
{
if (nextRefPosition->assignedReg() != REG_NA)
{
insertCopyOrReload(block, treeNode, currentRefPosition->getMultiRegIdx(),
nextRefPosition);
}
else
{
assert(nextRefPosition->AllocateIfProfitable());
// In case of tree temps, if def is spilled and use didn't
// get a register, set a flag on tree node to be treated as
// contained at the point of its use.
if (currentRefPosition->spillAfter && currentRefPosition->refType == RefTypeDef &&
nextRefPosition->refType == RefTypeUse)
{
assert(nextRefPosition->treeNode == nullptr);
treeNode->gtFlags |= GTF_NOREG_AT_USE;
}
}
}
}
// We should never have to "spill after" a temp use, since
// they're single use
else
{
unreached();
}
}
}
}
if (enregisterLocalVars)
{
processBlockEndLocations(block);
}
}
if (enregisterLocalVars)
{
#ifdef DEBUG
if (VERBOSE)
{
printf("-----------------------\n");
printf("RESOLVING BB BOUNDARIES\n");
printf("-----------------------\n");
printf("Resolution Candidates: ");
dumpConvertedVarSet(compiler, resolutionCandidateVars);
printf("\n");
printf("Has %sCritical Edges\n\n", hasCriticalEdges ? "" : "No");
printf("Prior to Resolution\n");
foreach_block(compiler, block)
{
printf("\nBB%02u use def in out\n", block->bbNum);
dumpConvertedVarSet(compiler, block->bbVarUse);
printf("\n");
dumpConvertedVarSet(compiler, block->bbVarDef);
printf("\n");
dumpConvertedVarSet(compiler, block->bbLiveIn);
printf("\n");
dumpConvertedVarSet(compiler, block->bbLiveOut);
printf("\n");
dumpInVarToRegMap(block);
dumpOutVarToRegMap(block);
}
printf("\n\n");
}
#endif // DEBUG
resolveEdges();
// Verify register assignments on variables
unsigned lclNum;
LclVarDsc* varDsc;
for (lclNum = 0, varDsc = compiler->lvaTable; lclNum < compiler->lvaCount; lclNum++, varDsc++)
{
if (!isCandidateVar(varDsc))
{
varDsc->lvRegNum = REG_STK;
}
else
{
Interval* interval = getIntervalForLocalVar(varDsc->lvVarIndex);
// Determine initial position for parameters
if (varDsc->lvIsParam)
{
regMaskTP initialRegMask = interval->firstRefPosition->registerAssignment;
regNumber initialReg = (initialRegMask == RBM_NONE || interval->firstRefPosition->spillAfter)
? REG_STK
: genRegNumFromMask(initialRegMask);
regNumber sourceReg = (varDsc->lvIsRegArg) ? varDsc->lvArgReg : REG_STK;
#ifdef _TARGET_ARM_
if (varTypeIsMultiReg(varDsc))
{
// TODO-ARM-NYI: Map the hi/lo intervals back to lvRegNum and lvOtherReg (these should NYI
// before this)
assert(!"Multi-reg types not yet supported");
}
else
#endif // _TARGET_ARM_
{
varDsc->lvArgInitReg = initialReg;
JITDUMP(" Set V%02u argument initial register to %s\n", lclNum, getRegName(initialReg));
}
// Stack args that are part of dependently-promoted structs should never be register candidates (see
// LinearScan::isRegCandidate).
assert(varDsc->lvIsRegArg || !compiler->lvaIsFieldOfDependentlyPromotedStruct(varDsc));
}
// If lvRegNum is REG_STK, that means that either no register
// was assigned, or (more likely) that the same register was not
// used for all references. In that case, codegen gets the register
// from the tree node.
if (varDsc->lvRegNum == REG_STK || interval->isSpilled || interval->isSplit)
{
// For codegen purposes, we'll set lvRegNum to whatever register
// it's currently in as we go.
// However, we never mark an interval as lvRegister if it has either been spilled
// or split.
varDsc->lvRegister = false;
// Skip any dead defs or exposed uses
// (first use exposed will only occur when there is no explicit initialization)
RefPosition* firstRefPosition = interval->firstRefPosition;
while ((firstRefPosition != nullptr) && (firstRefPosition->refType == RefTypeExpUse))
{
firstRefPosition = firstRefPosition->nextRefPosition;
}
if (firstRefPosition == nullptr)
{
// Dead interval
varDsc->lvLRACandidate = false;
if (varDsc->lvRefCnt == 0)
{
varDsc->lvOnFrame = false;
}
else
{
// We may encounter cases where a lclVar actually has no references, but
// a non-zero refCnt. For safety (in case this is some "hidden" lclVar that we're
// not correctly recognizing), we'll mark those as needing a stack location.
// TODO-Cleanup: Make this an assert if/when we correct the refCnt
// updating.
varDsc->lvOnFrame = true;
}
}
else
{
// If the interval was not spilled, it doesn't need a stack location.
if (!interval->isSpilled)
{
varDsc->lvOnFrame = false;
}
if (firstRefPosition->registerAssignment == RBM_NONE || firstRefPosition->spillAfter)
{
// Either this RefPosition is spilled, or regOptional or it is not a "real" def or use
assert(
firstRefPosition->spillAfter || firstRefPosition->AllocateIfProfitable() ||
(firstRefPosition->refType != RefTypeDef && firstRefPosition->refType != RefTypeUse));
varDsc->lvRegNum = REG_STK;
}
else
{
varDsc->lvRegNum = firstRefPosition->assignedReg();
}
}
}
else
{
{
varDsc->lvRegister = true;
varDsc->lvOnFrame = false;
}
#ifdef DEBUG
regMaskTP registerAssignment = genRegMask(varDsc->lvRegNum);
assert(!interval->isSpilled && !interval->isSplit);
RefPosition* refPosition = interval->firstRefPosition;
assert(refPosition != nullptr);
while (refPosition != nullptr)
{
// All RefPositions must match, except for dead definitions,
// copyReg/moveReg and RefTypeExpUse positions
if (refPosition->registerAssignment != RBM_NONE && !refPosition->copyReg &&
!refPosition->moveReg && refPosition->refType != RefTypeExpUse)
{
assert(refPosition->registerAssignment == registerAssignment);
}
refPosition = refPosition->nextRefPosition;
}
#endif // DEBUG
}
}
}
}
#ifdef DEBUG
if (VERBOSE)
{
printf("Trees after linear scan register allocator (LSRA)\n");
compiler->fgDispBasicBlocks(true);
}
verifyFinalAllocation();
#endif // DEBUG
compiler->raMarkStkVars();
recordMaxSpill();
// TODO-CQ: Review this comment and address as needed.
// Change all unused promoted non-argument struct locals to a non-GC type (in this case TYP_INT)
// so that the gc tracking logic and lvMustInit logic will ignore them.
// Extract the code that does this from raAssignVars, and call it here.
// PRECONDITIONS: Ensure that lvPromoted is set on promoted structs, if and
// only if it is promoted on all paths.
// Call might be something like:
// compiler->BashUnusedStructLocals();
}
//
//------------------------------------------------------------------------
// insertMove: Insert a move of a lclVar with the given lclNum into the given block.
//
// Arguments:
// block - the BasicBlock into which the move will be inserted.
// insertionPoint - the instruction before which to insert the move
// lclNum - the lclNum of the var to be moved
// fromReg - the register from which the var is moving
// toReg - the register to which the var is moving
//
// Return Value:
// None.
//
// Notes:
// If insertionPoint is non-NULL, insert before that instruction;
// otherwise, insert "near" the end (prior to the branch, if any).
// If fromReg or toReg is REG_STK, then move from/to memory, respectively.
void LinearScan::insertMove(
BasicBlock* block, GenTreePtr insertionPoint, unsigned lclNum, regNumber fromReg, regNumber toReg)
{
LclVarDsc* varDsc = compiler->lvaTable + lclNum;
// the lclVar must be a register candidate
assert(isRegCandidate(varDsc));
// One or both MUST be a register
assert(fromReg != REG_STK || toReg != REG_STK);
// They must not be the same register.
assert(fromReg != toReg);
// This var can't be marked lvRegister now
varDsc->lvRegNum = REG_STK;
GenTreePtr src = compiler->gtNewLclvNode(lclNum, varDsc->TypeGet());
src->gtLsraInfo.isLsraAdded = true;
// There are three cases we need to handle:
// - We are loading a lclVar from the stack.
// - We are storing a lclVar to the stack.
// - We are copying a lclVar between registers.
//
// In the first and second cases, the lclVar node will be marked with GTF_SPILLED and GTF_SPILL, respectively.
// It is up to the code generator to ensure that any necessary normalization is done when loading or storing the
// lclVar's value.
//
// In the third case, we generate GT_COPY(GT_LCL_VAR) and type each node with the normalized type of the lclVar.
// This is safe because a lclVar is always normalized once it is in a register.
GenTree* dst = src;
if (fromReg == REG_STK)
{
src->gtFlags |= GTF_SPILLED;
src->gtRegNum = toReg;
}
else if (toReg == REG_STK)
{
src->gtFlags |= GTF_SPILL;
src->gtRegNum = fromReg;
}
else
{
var_types movType = genActualType(varDsc->TypeGet());
src->gtType = movType;
dst = new (compiler, GT_COPY) GenTreeCopyOrReload(GT_COPY, movType, src);
// This is the new home of the lclVar - indicate that by clearing the GTF_VAR_DEATH flag.
// Note that if src is itself a lastUse, this will have no effect.
dst->gtFlags &= ~(GTF_VAR_DEATH);
src->gtRegNum = fromReg;
dst->gtRegNum = toReg;
src->gtLsraInfo.isLocalDefUse = false;
dst->gtLsraInfo.isLsraAdded = true;
}
dst->gtLsraInfo.isLocalDefUse = true;
dst->SetUnusedValue();
LIR::Range treeRange = LIR::SeqTree(compiler, dst);
LIR::Range& blockRange = LIR::AsRange(block);
if (insertionPoint != nullptr)
{
blockRange.InsertBefore(insertionPoint, std::move(treeRange));
}
else
{
// Put the copy at the bottom
// If there's a branch, make an embedded statement that executes just prior to the branch
if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH)
{
noway_assert(!blockRange.IsEmpty());
GenTree* branch = blockRange.LastNode();
assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE ||
branch->OperGet() == GT_SWITCH);
blockRange.InsertBefore(branch, std::move(treeRange));
}
else
{
assert(block->bbJumpKind == BBJ_NONE || block->bbJumpKind == BBJ_ALWAYS);
blockRange.InsertAtEnd(std::move(treeRange));
}
}
}
void LinearScan::insertSwap(
BasicBlock* block, GenTreePtr insertionPoint, unsigned lclNum1, regNumber reg1, unsigned lclNum2, regNumber reg2)
{
#ifdef DEBUG
if (VERBOSE)
{
const char* insertionPointString = "top";
if (insertionPoint == nullptr)
{
insertionPointString = "bottom";
}
printf(" BB%02u %s: swap V%02u in %s with V%02u in %s\n", block->bbNum, insertionPointString, lclNum1,
getRegName(reg1), lclNum2, getRegName(reg2));
}
#endif // DEBUG
LclVarDsc* varDsc1 = compiler->lvaTable + lclNum1;
LclVarDsc* varDsc2 = compiler->lvaTable + lclNum2;
assert(reg1 != REG_STK && reg1 != REG_NA && reg2 != REG_STK && reg2 != REG_NA);
GenTreePtr lcl1 = compiler->gtNewLclvNode(lclNum1, varDsc1->TypeGet());
lcl1->gtLsraInfo.isLsraAdded = true;
lcl1->gtLsraInfo.isLocalDefUse = false;
lcl1->gtRegNum = reg1;
GenTreePtr lcl2 = compiler->gtNewLclvNode(lclNum2, varDsc2->TypeGet());
lcl2->gtLsraInfo.isLsraAdded = true;
lcl2->gtLsraInfo.isLocalDefUse = false;
lcl2->gtRegNum = reg2;
GenTreePtr swap = compiler->gtNewOperNode(GT_SWAP, TYP_VOID, lcl1, lcl2);
swap->gtLsraInfo.isLsraAdded = true;
swap->gtLsraInfo.isLocalDefUse = false;
swap->gtRegNum = REG_NA;
lcl1->gtNext = lcl2;
lcl2->gtPrev = lcl1;
lcl2->gtNext = swap;
swap->gtPrev = lcl2;
LIR::Range swapRange = LIR::SeqTree(compiler, swap);
LIR::Range& blockRange = LIR::AsRange(block);
if (insertionPoint != nullptr)
{
blockRange.InsertBefore(insertionPoint, std::move(swapRange));
}
else
{
// Put the copy at the bottom
// If there's a branch, make an embedded statement that executes just prior to the branch
if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH)
{
noway_assert(!blockRange.IsEmpty());
GenTree* branch = blockRange.LastNode();
assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE ||
branch->OperGet() == GT_SWITCH);
blockRange.InsertBefore(branch, std::move(swapRange));
}
else
{
assert(block->bbJumpKind == BBJ_NONE || block->bbJumpKind == BBJ_ALWAYS);
blockRange.InsertAtEnd(std::move(swapRange));
}
}
}
//------------------------------------------------------------------------
// getTempRegForResolution: Get a free register to use for resolution code.
//
// Arguments:
// fromBlock - The "from" block on the edge being resolved.
// toBlock - The "to"block on the edge
// type - the type of register required
//
// Return Value:
// Returns a register that is free on the given edge, or REG_NA if none is available.
//
// Notes:
// It is up to the caller to check the return value, and to determine whether a register is
// available, and to handle that case appropriately.
// It is also up to the caller to cache the return value, as this is not cheap to compute.
regNumber LinearScan::getTempRegForResolution(BasicBlock* fromBlock, BasicBlock* toBlock, var_types type)
{
// TODO-Throughput: This would be much more efficient if we add RegToVarMaps instead of VarToRegMaps
// and they would be more space-efficient as well.
VarToRegMap fromVarToRegMap = getOutVarToRegMap(fromBlock->bbNum);
VarToRegMap toVarToRegMap = getInVarToRegMap(toBlock->bbNum);
regMaskTP freeRegs = allRegs(type);
#ifdef DEBUG
if (getStressLimitRegs() == LSRA_LIMIT_SMALL_SET)
{
return REG_NA;
}
#endif // DEBUG
INDEBUG(freeRegs = stressLimitRegs(nullptr, freeRegs));
// We are only interested in the variables that are live-in to the "to" block.
VarSetOps::Iter iter(compiler, toBlock->bbLiveIn);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex) && freeRegs != RBM_NONE)
{
regNumber fromReg = getVarReg(fromVarToRegMap, varIndex);
regNumber toReg = getVarReg(toVarToRegMap, varIndex);
assert(fromReg != REG_NA && toReg != REG_NA);
if (fromReg != REG_STK)
{
freeRegs &= ~genRegMask(fromReg);
}
if (toReg != REG_STK)
{
freeRegs &= ~genRegMask(toReg);
}
}
if (freeRegs == RBM_NONE)
{
return REG_NA;
}
else
{
regNumber tempReg = genRegNumFromMask(genFindLowestBit(freeRegs));
return tempReg;
}
}
//------------------------------------------------------------------------
// addResolution: Add a resolution move of the given interval
//
// Arguments:
// block - the BasicBlock into which the move will be inserted.
// insertionPoint - the instruction before which to insert the move
// interval - the interval of the var to be moved
// toReg - the register to which the var is moving
// fromReg - the register from which the var is moving
//
// Return Value:
// None.
//
// Notes:
// For joins, we insert at the bottom (indicated by an insertionPoint
// of nullptr), while for splits we insert at the top.
// This is because for joins 'block' is a pred of the join, while for splits it is a succ.
// For critical edges, this function may be called twice - once to move from
// the source (fromReg), if any, to the stack, in which case toReg will be
// REG_STK, and we insert at the bottom (leave insertionPoint as nullptr).
// The next time, we want to move from the stack to the destination (toReg),
// in which case fromReg will be REG_STK, and we insert at the top.
void LinearScan::addResolution(
BasicBlock* block, GenTreePtr insertionPoint, Interval* interval, regNumber toReg, regNumber fromReg)
{
#ifdef DEBUG
const char* insertionPointString = "top";
#endif // DEBUG
if (insertionPoint == nullptr)
{
#ifdef DEBUG
insertionPointString = "bottom";
#endif // DEBUG
}
JITDUMP(" BB%02u %s: move V%02u from ", block->bbNum, insertionPointString, interval->varNum);
JITDUMP("%s to %s", getRegName(fromReg), getRegName(toReg));
insertMove(block, insertionPoint, interval->varNum, fromReg, toReg);
if (fromReg == REG_STK || toReg == REG_STK)
{
assert(interval->isSpilled);
}
else
{
// We should have already marked this as spilled or split.
assert((interval->isSpilled) || (interval->isSplit));
}
INTRACK_STATS(updateLsraStat(LSRA_STAT_RESOLUTION_MOV, block->bbNum));
}
//------------------------------------------------------------------------
// handleOutgoingCriticalEdges: Performs the necessary resolution on all critical edges that feed out of 'block'
//
// Arguments:
// block - the block with outgoing critical edges.
//
// Return Value:
// None..
//
// Notes:
// For all outgoing critical edges (i.e. any successor of this block which is
// a join edge), if there are any conflicts, split the edge by adding a new block,
// and generate the resolution code into that block.
void LinearScan::handleOutgoingCriticalEdges(BasicBlock* block)
{
VARSET_TP outResolutionSet(VarSetOps::Intersection(compiler, block->bbLiveOut, resolutionCandidateVars));
if (VarSetOps::IsEmpty(compiler, outResolutionSet))
{
return;
}
VARSET_TP sameResolutionSet(VarSetOps::MakeEmpty(compiler));
VARSET_TP sameLivePathsSet(VarSetOps::MakeEmpty(compiler));
VARSET_TP singleTargetSet(VarSetOps::MakeEmpty(compiler));
VARSET_TP diffResolutionSet(VarSetOps::MakeEmpty(compiler));
// Get the outVarToRegMap for this block
VarToRegMap outVarToRegMap = getOutVarToRegMap(block->bbNum);
unsigned succCount = block->NumSucc(compiler);
assert(succCount > 1);
VarToRegMap firstSuccInVarToRegMap = nullptr;
BasicBlock* firstSucc = nullptr;
// First, determine the live regs at the end of this block so that we know what regs are
// available to copy into.
// Note that for this purpose we use the full live-out set, because we must ensure that
// even the registers that remain the same across the edge are preserved correctly.
regMaskTP liveOutRegs = RBM_NONE;
VarSetOps::Iter liveOutIter(compiler, block->bbLiveOut);
unsigned liveOutVarIndex = 0;
while (liveOutIter.NextElem(&liveOutVarIndex))
{
regNumber fromReg = getVarReg(outVarToRegMap, liveOutVarIndex);
if (fromReg != REG_STK)
{
liveOutRegs |= genRegMask(fromReg);
}
}
// Next, if this blocks ends with a switch table, we have to make sure not to copy
// into the registers that it uses.
regMaskTP switchRegs = RBM_NONE;
if (block->bbJumpKind == BBJ_SWITCH)
{
// At this point, Lowering has transformed any non-switch-table blocks into
// cascading ifs.
GenTree* switchTable = LIR::AsRange(block).LastNode();
assert(switchTable != nullptr && switchTable->OperGet() == GT_SWITCH_TABLE);
switchRegs = switchTable->gtRsvdRegs;
GenTree* op1 = switchTable->gtGetOp1();
GenTree* op2 = switchTable->gtGetOp2();
noway_assert(op1 != nullptr && op2 != nullptr);
assert(op1->gtRegNum != REG_NA && op2->gtRegNum != REG_NA);
switchRegs |= genRegMask(op1->gtRegNum);
switchRegs |= genRegMask(op2->gtRegNum);
}
#ifdef _TARGET_ARM64_
// Next, if this blocks ends with a JCMP, we have to make sure not to copy
// into the register that it uses or modify the local variable it must consume
LclVarDsc* jcmpLocalVarDsc = nullptr;
if (block->bbJumpKind == BBJ_COND)
{
GenTree* lastNode = LIR::AsRange(block).LastNode();
if (lastNode->OperIs(GT_JCMP))
{
GenTree* op1 = lastNode->gtGetOp1();
switchRegs |= genRegMask(op1->gtRegNum);
if (op1->IsLocal())
{
GenTreeLclVarCommon* lcl = op1->AsLclVarCommon();
jcmpLocalVarDsc = &compiler->lvaTable[lcl->gtLclNum];
}
}
}
#endif
VarToRegMap sameVarToRegMap = sharedCriticalVarToRegMap;
regMaskTP sameWriteRegs = RBM_NONE;
regMaskTP diffReadRegs = RBM_NONE;
// For each var that may require resolution, classify them as:
// - in the same register at the end of this block and at each target (no resolution needed)
// - in different registers at different targets (resolve separately):
// diffResolutionSet
// - in the same register at each target at which it's live, but different from the end of
// this block. We may be able to resolve these as if it is "join", but only if they do not
// write to any registers that are read by those in the diffResolutionSet:
// sameResolutionSet
VarSetOps::Iter outResolutionSetIter(compiler, outResolutionSet);
unsigned outResolutionSetVarIndex = 0;
while (outResolutionSetIter.NextElem(&outResolutionSetVarIndex))
{
regNumber fromReg = getVarReg(outVarToRegMap, outResolutionSetVarIndex);
bool isMatch = true;
bool isSame = false;
bool maybeSingleTarget = false;
bool maybeSameLivePaths = false;
bool liveOnlyAtSplitEdge = true;
regNumber sameToReg = REG_NA;
for (unsigned succIndex = 0; succIndex < succCount; succIndex++)
{
BasicBlock* succBlock = block->GetSucc(succIndex, compiler);
if (!VarSetOps::IsMember(compiler, succBlock->bbLiveIn, outResolutionSetVarIndex))
{
maybeSameLivePaths = true;
continue;
}
else if (liveOnlyAtSplitEdge)
{
// Is the var live only at those target blocks which are connected by a split edge to this block
liveOnlyAtSplitEdge = ((succBlock->bbPreds->flNext == nullptr) && (succBlock != compiler->fgFirstBB));
}
regNumber toReg = getVarReg(getInVarToRegMap(succBlock->bbNum), outResolutionSetVarIndex);
if (sameToReg == REG_NA)
{
sameToReg = toReg;
continue;
}
if (toReg == sameToReg)
{
continue;
}
sameToReg = REG_NA;
break;
}
// Check for the cases where we can't write to a register.
// We only need to check for these cases if sameToReg is an actual register (not REG_STK).
if (sameToReg != REG_NA && sameToReg != REG_STK)
{
// If there's a path on which this var isn't live, it may use the original value in sameToReg.
// In this case, sameToReg will be in the liveOutRegs of this block.
// Similarly, if sameToReg is in sameWriteRegs, it has already been used (i.e. for a lclVar that's
// live only at another target), and we can't copy another lclVar into that reg in this block.
regMaskTP sameToRegMask = genRegMask(sameToReg);
if (maybeSameLivePaths &&
(((sameToRegMask & liveOutRegs) != RBM_NONE) || ((sameToRegMask & sameWriteRegs) != RBM_NONE)))
{
sameToReg = REG_NA;
}
// If this register is used by a switch table at the end of the block, we can't do the copy
// in this block (since we can't insert it after the switch).
if ((sameToRegMask & switchRegs) != RBM_NONE)
{
sameToReg = REG_NA;
}
#ifdef _TARGET_ARM64_
if (jcmpLocalVarDsc && (jcmpLocalVarDsc->lvVarIndex == outResolutionSetVarIndex))
{
sameToReg = REG_NA;
}
#endif
// If the var is live only at those blocks connected by a split edge and not live-in at some of the
// target blocks, we will resolve it the same way as if it were in diffResolutionSet and resolution
// will be deferred to the handling of split edges, which means copy will only be at those target(s).
//
// Another way to achieve similar resolution for vars live only at split edges is by removing them
// from consideration up-front but it requires that we traverse those edges anyway to account for
// the registers that must note be overwritten.
if (liveOnlyAtSplitEdge && maybeSameLivePaths)
{
sameToReg = REG_NA;
}
}
if (sameToReg == REG_NA)
{
VarSetOps::AddElemD(compiler, diffResolutionSet, outResolutionSetVarIndex);
if (fromReg != REG_STK)
{
diffReadRegs |= genRegMask(fromReg);
}
}
else if (sameToReg != fromReg)
{
VarSetOps::AddElemD(compiler, sameResolutionSet, outResolutionSetVarIndex);
setVarReg(sameVarToRegMap, outResolutionSetVarIndex, sameToReg);
if (sameToReg != REG_STK)
{
sameWriteRegs |= genRegMask(sameToReg);
}
}
}
if (!VarSetOps::IsEmpty(compiler, sameResolutionSet))
{
if ((sameWriteRegs & diffReadRegs) != RBM_NONE)
{
// We cannot split the "same" and "diff" regs if the "same" set writes registers
// that must be read by the "diff" set. (Note that when these are done as a "batch"
// we carefully order them to ensure all the input regs are read before they are
// overwritten.)
VarSetOps::UnionD(compiler, diffResolutionSet, sameResolutionSet);
VarSetOps::ClearD(compiler, sameResolutionSet);
}
else
{
// For any vars in the sameResolutionSet, we can simply add the move at the end of "block".
resolveEdge(block, nullptr, ResolveSharedCritical, sameResolutionSet);
}
}
if (!VarSetOps::IsEmpty(compiler, diffResolutionSet))
{
for (unsigned succIndex = 0; succIndex < succCount; succIndex++)
{
BasicBlock* succBlock = block->GetSucc(succIndex, compiler);
// Any "diffResolutionSet" resolution for a block with no other predecessors will be handled later
// as split resolution.
if ((succBlock->bbPreds->flNext == nullptr) && (succBlock != compiler->fgFirstBB))
{
continue;
}
// Now collect the resolution set for just this edge, if any.
// Check only the vars in diffResolutionSet that are live-in to this successor.
bool needsResolution = false;
VarToRegMap succInVarToRegMap = getInVarToRegMap(succBlock->bbNum);
VARSET_TP edgeResolutionSet(VarSetOps::Intersection(compiler, diffResolutionSet, succBlock->bbLiveIn));
VarSetOps::Iter iter(compiler, edgeResolutionSet);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
regNumber fromReg = getVarReg(outVarToRegMap, varIndex);
regNumber toReg = getVarReg(succInVarToRegMap, varIndex);
if (fromReg == toReg)
{
VarSetOps::RemoveElemD(compiler, edgeResolutionSet, varIndex);
}
}
if (!VarSetOps::IsEmpty(compiler, edgeResolutionSet))
{
resolveEdge(block, succBlock, ResolveCritical, edgeResolutionSet);
}
}
}
}
//------------------------------------------------------------------------
// resolveEdges: Perform resolution across basic block edges
//
// Arguments:
// None.
//
// Return Value:
// None.
//
// Notes:
// Traverse the basic blocks.
// - If this block has a single predecessor that is not the immediately
// preceding block, perform any needed 'split' resolution at the beginning of this block
// - Otherwise if this block has critical incoming edges, handle them.
// - If this block has a single successor that has multiple predecesors, perform any needed
// 'join' resolution at the end of this block.
// Note that a block may have both 'split' or 'critical' incoming edge(s) and 'join' outgoing
// edges.
void LinearScan::resolveEdges()
{
JITDUMP("RESOLVING EDGES\n");
// The resolutionCandidateVars set was initialized with all the lclVars that are live-in to
// any block. We now intersect that set with any lclVars that ever spilled or split.
// If there are no candidates for resoultion, simply return.
VarSetOps::IntersectionD(compiler, resolutionCandidateVars, splitOrSpilledVars);
if (VarSetOps::IsEmpty(compiler, resolutionCandidateVars))
{
return;
}
BasicBlock *block, *prevBlock = nullptr;
// Handle all the critical edges first.
// We will try to avoid resolution across critical edges in cases where all the critical-edge
// targets of a block have the same home. We will then split the edges only for the
// remaining mismatches. We visit the out-edges, as that allows us to share the moves that are
// common among allt he targets.
if (hasCriticalEdges)
{
foreach_block(compiler, block)
{
if (block->bbNum > bbNumMaxBeforeResolution)
{
// This is a new block added during resolution - we don't need to visit these now.
continue;
}
if (blockInfo[block->bbNum].hasCriticalOutEdge)
{
handleOutgoingCriticalEdges(block);
}
prevBlock = block;
}
}
prevBlock = nullptr;
foreach_block(compiler, block)
{
if (block->bbNum > bbNumMaxBeforeResolution)
{
// This is a new block added during resolution - we don't need to visit these now.
continue;
}
unsigned succCount = block->NumSucc(compiler);
flowList* preds = block->bbPreds;
BasicBlock* uniquePredBlock = block->GetUniquePred(compiler);
// First, if this block has a single predecessor,
// we may need resolution at the beginning of this block.
// This may be true even if it's the block we used for starting locations,
// if a variable was spilled.
VARSET_TP inResolutionSet(VarSetOps::Intersection(compiler, block->bbLiveIn, resolutionCandidateVars));
if (!VarSetOps::IsEmpty(compiler, inResolutionSet))
{
if (uniquePredBlock != nullptr)
{
// We may have split edges during critical edge resolution, and in the process split
// a non-critical edge as well.
// It is unlikely that we would ever have more than one of these in sequence (indeed,
// I don't think it's possible), but there's no need to assume that it can't.
while (uniquePredBlock->bbNum > bbNumMaxBeforeResolution)
{
uniquePredBlock = uniquePredBlock->GetUniquePred(compiler);
noway_assert(uniquePredBlock != nullptr);
}
resolveEdge(uniquePredBlock, block, ResolveSplit, inResolutionSet);
}
}
// Finally, if this block has a single successor:
// - and that has at least one other predecessor (otherwise we will do the resolution at the
// top of the successor),
// - and that is not the target of a critical edge (otherwise we've already handled it)
// we may need resolution at the end of this block.
if (succCount == 1)
{
BasicBlock* succBlock = block->GetSucc(0, compiler);
if (succBlock->GetUniquePred(compiler) == nullptr)
{
VARSET_TP outResolutionSet(
VarSetOps::Intersection(compiler, succBlock->bbLiveIn, resolutionCandidateVars));
if (!VarSetOps::IsEmpty(compiler, outResolutionSet))
{
resolveEdge(block, succBlock, ResolveJoin, outResolutionSet);
}
}
}
}
// Now, fixup the mapping for any blocks that were adding for edge splitting.
// See the comment prior to the call to fgSplitEdge() in resolveEdge().
// Note that we could fold this loop in with the checking code below, but that
// would only improve the debug case, and would clutter up the code somewhat.
if (compiler->fgBBNumMax > bbNumMaxBeforeResolution)
{
foreach_block(compiler, block)
{
if (block->bbNum > bbNumMaxBeforeResolution)
{
// There may be multiple blocks inserted when we split. But we must always have exactly
// one path (i.e. all blocks must be single-successor and single-predecessor),
// and only one block along the path may be non-empty.
// Note that we may have a newly-inserted block that is empty, but which connects
// two non-resolution blocks. This happens when an edge is split that requires it.
BasicBlock* succBlock = block;
do
{
succBlock = succBlock->GetUniqueSucc();
noway_assert(succBlock != nullptr);
} while ((succBlock->bbNum > bbNumMaxBeforeResolution) && succBlock->isEmpty());
BasicBlock* predBlock = block;
do
{
predBlock = predBlock->GetUniquePred(compiler);
noway_assert(predBlock != nullptr);
} while ((predBlock->bbNum > bbNumMaxBeforeResolution) && predBlock->isEmpty());
unsigned succBBNum = succBlock->bbNum;
unsigned predBBNum = predBlock->bbNum;
if (block->isEmpty())
{
// For the case of the empty block, find the non-resolution block (succ or pred).
if (predBBNum > bbNumMaxBeforeResolution)
{
assert(succBBNum <= bbNumMaxBeforeResolution);
predBBNum = 0;
}
else
{
succBBNum = 0;
}
}
else
{
assert((succBBNum <= bbNumMaxBeforeResolution) && (predBBNum <= bbNumMaxBeforeResolution));
}
SplitEdgeInfo info = {predBBNum, succBBNum};
getSplitBBNumToTargetBBNumMap()->Set(block->bbNum, info);
}
}
}
#ifdef DEBUG
// Make sure the varToRegMaps match up on all edges.
bool foundMismatch = false;
foreach_block(compiler, block)
{
if (block->isEmpty() && block->bbNum > bbNumMaxBeforeResolution)
{
continue;
}
VarToRegMap toVarToRegMap = getInVarToRegMap(block->bbNum);
for (flowList* pred = block->bbPreds; pred != nullptr; pred = pred->flNext)
{
BasicBlock* predBlock = pred->flBlock;
VarToRegMap fromVarToRegMap = getOutVarToRegMap(predBlock->bbNum);
VarSetOps::Iter iter(compiler, block->bbLiveIn);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
regNumber fromReg = getVarReg(fromVarToRegMap, varIndex);
regNumber toReg = getVarReg(toVarToRegMap, varIndex);
if (fromReg != toReg)
{
if (!foundMismatch)
{
foundMismatch = true;
printf("Found mismatched var locations after resolution!\n");
}
unsigned varNum = compiler->lvaTrackedToVarNum[varIndex];
printf(" V%02u: BB%02u to BB%02u: %s to %s\n", varNum, predBlock->bbNum, block->bbNum,
getRegName(fromReg), getRegName(toReg));
}
}
}
}
assert(!foundMismatch);
#endif
JITDUMP("\n");
}
//------------------------------------------------------------------------
// resolveEdge: Perform the specified type of resolution between two blocks.
//
// Arguments:
// fromBlock - the block from which the edge originates
// toBlock - the block at which the edge terminates
// resolveType - the type of resolution to be performed
// liveSet - the set of tracked lclVar indices which may require resolution
//
// Return Value:
// None.
//
// Assumptions:
// The caller must have performed the analysis to determine the type of the edge.
//
// Notes:
// This method emits the correctly ordered moves necessary to place variables in the
// correct registers across a Split, Join or Critical edge.
// In order to avoid overwriting register values before they have been moved to their
// new home (register/stack), it first does the register-to-stack moves (to free those
// registers), then the register to register moves, ensuring that the target register
// is free before the move, and then finally the stack to register moves.
void LinearScan::resolveEdge(BasicBlock* fromBlock,
BasicBlock* toBlock,
ResolveType resolveType,
VARSET_VALARG_TP liveSet)
{
VarToRegMap fromVarToRegMap = getOutVarToRegMap(fromBlock->bbNum);
VarToRegMap toVarToRegMap;
if (resolveType == ResolveSharedCritical)
{
toVarToRegMap = sharedCriticalVarToRegMap;
}
else
{
toVarToRegMap = getInVarToRegMap(toBlock->bbNum);
}
// The block to which we add the resolution moves depends on the resolveType
BasicBlock* block;
switch (resolveType)
{
case ResolveJoin:
case ResolveSharedCritical:
block = fromBlock;
break;
case ResolveSplit:
block = toBlock;
break;
case ResolveCritical:
// fgSplitEdge may add one or two BasicBlocks. It returns the block that splits
// the edge from 'fromBlock' and 'toBlock', but if it inserts that block right after
// a block with a fall-through it will have to create another block to handle that edge.
// These new blocks can be mapped to existing blocks in order to correctly handle
// the calls to recordVarLocationsAtStartOfBB() from codegen. That mapping is handled
// in resolveEdges(), after all the edge resolution has been done (by calling this
// method for each edge).
block = compiler->fgSplitEdge(fromBlock, toBlock);
// Split edges are counted against fromBlock.
INTRACK_STATS(updateLsraStat(LSRA_STAT_SPLIT_EDGE, fromBlock->bbNum));
break;
default:
unreached();
break;
}
#ifndef _TARGET_XARCH_
// We record tempregs for beginning and end of each block.
// For amd64/x86 we only need a tempReg for float - we'll use xchg for int.
// TODO-Throughput: It would be better to determine the tempRegs on demand, but the code below
// modifies the varToRegMaps so we don't have all the correct registers at the time
// we need to get the tempReg.
regNumber tempRegInt =
(resolveType == ResolveSharedCritical) ? REG_NA : getTempRegForResolution(fromBlock, toBlock, TYP_INT);
#endif // !_TARGET_XARCH_
regNumber tempRegFlt = REG_NA;
if ((compiler->compFloatingPointUsed) && (resolveType != ResolveSharedCritical))
{
#ifdef _TARGET_ARM_
// Let's try to reserve a double register for TYP_FLOAT and TYP_DOUBLE
tempRegFlt = getTempRegForResolution(fromBlock, toBlock, TYP_DOUBLE);
if (tempRegFlt == REG_NA)
{
// If fails, try to reserve a float register for TYP_FLOAT
tempRegFlt = getTempRegForResolution(fromBlock, toBlock, TYP_FLOAT);
}
#else
tempRegFlt = getTempRegForResolution(fromBlock, toBlock, TYP_FLOAT);
#endif
}
regMaskTP targetRegsToDo = RBM_NONE;
regMaskTP targetRegsReady = RBM_NONE;
regMaskTP targetRegsFromStack = RBM_NONE;
// The following arrays capture the location of the registers as they are moved:
// - location[reg] gives the current location of the var that was originally in 'reg'.
// (Note that a var may be moved more than once.)
// - source[reg] gives the original location of the var that needs to be moved to 'reg'.
// For example, if a var is in rax and needs to be moved to rsi, then we would start with:
// location[rax] == rax
// source[rsi] == rax -- this doesn't change
// Then, if for some reason we need to move it temporary to rbx, we would have:
// location[rax] == rbx
// Once we have completed the move, we will have:
// location[rax] == REG_NA
// This indicates that the var originally in rax is now in its target register.
regNumberSmall location[REG_COUNT];
C_ASSERT(sizeof(char) == sizeof(regNumberSmall)); // for memset to work
memset(location, REG_NA, REG_COUNT);
regNumberSmall source[REG_COUNT];
memset(source, REG_NA, REG_COUNT);
// What interval is this register associated with?
// (associated with incoming reg)
Interval* sourceIntervals[REG_COUNT];
memset(&sourceIntervals, 0, sizeof(sourceIntervals));
// Intervals for vars that need to be loaded from the stack
Interval* stackToRegIntervals[REG_COUNT];
memset(&stackToRegIntervals, 0, sizeof(stackToRegIntervals));
// Get the starting insertion point for the "to" resolution
GenTreePtr insertionPoint = nullptr;
if (resolveType == ResolveSplit || resolveType == ResolveCritical)
{
insertionPoint = LIR::AsRange(block).FirstNonPhiNode();
}
// First:
// - Perform all moves from reg to stack (no ordering needed on these)
// - For reg to reg moves, record the current location, associating their
// source location with the target register they need to go into
// - For stack to reg moves (done last, no ordering needed between them)
// record the interval associated with the target reg
// TODO-Throughput: We should be looping over the liveIn and liveOut registers, since
// that will scale better than the live variables
VarSetOps::Iter iter(compiler, liveSet);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
regNumber fromReg = getVarReg(fromVarToRegMap, varIndex);
regNumber toReg = getVarReg(toVarToRegMap, varIndex);
if (fromReg == toReg)
{
continue;
}
// For Critical edges, the location will not change on either side of the edge,
// since we'll add a new block to do the move.
if (resolveType == ResolveSplit)
{
setVarReg(toVarToRegMap, varIndex, fromReg);
}
else if (resolveType == ResolveJoin || resolveType == ResolveSharedCritical)
{
setVarReg(fromVarToRegMap, varIndex, toReg);
}
assert(fromReg < UCHAR_MAX && toReg < UCHAR_MAX);
Interval* interval = getIntervalForLocalVar(varIndex);
if (fromReg == REG_STK)
{
stackToRegIntervals[toReg] = interval;
targetRegsFromStack |= genRegMask(toReg);
}
else if (toReg == REG_STK)
{
// Do the reg to stack moves now
addResolution(block, insertionPoint, interval, REG_STK, fromReg);
JITDUMP(" (%s)\n", resolveTypeName[resolveType]);
}
else
{
location[fromReg] = (regNumberSmall)fromReg;
source[toReg] = (regNumberSmall)fromReg;
sourceIntervals[fromReg] = interval;
targetRegsToDo |= genRegMask(toReg);
}
}
// REGISTER to REGISTER MOVES
// First, find all the ones that are ready to move now
regMaskTP targetCandidates = targetRegsToDo;
while (targetCandidates != RBM_NONE)
{
regMaskTP targetRegMask = genFindLowestBit(targetCandidates);
targetCandidates &= ~targetRegMask;
regNumber targetReg = genRegNumFromMask(targetRegMask);
if (location[targetReg] == REG_NA)
{
targetRegsReady |= targetRegMask;
}
}
// Perform reg to reg moves
while (targetRegsToDo != RBM_NONE)
{
while (targetRegsReady != RBM_NONE)
{
regMaskTP targetRegMask = genFindLowestBit(targetRegsReady);
targetRegsToDo &= ~targetRegMask;
targetRegsReady &= ~targetRegMask;
regNumber targetReg = genRegNumFromMask(targetRegMask);
assert(location[targetReg] != targetReg);
regNumber sourceReg = (regNumber)source[targetReg];
regNumber fromReg = (regNumber)location[sourceReg];
assert(fromReg < UCHAR_MAX && sourceReg < UCHAR_MAX);
Interval* interval = sourceIntervals[sourceReg];
assert(interval != nullptr);
addResolution(block, insertionPoint, interval, targetReg, fromReg);
JITDUMP(" (%s)\n", resolveTypeName[resolveType]);
sourceIntervals[sourceReg] = nullptr;
location[sourceReg] = REG_NA;
// Do we have a free targetReg?
if (fromReg == sourceReg && source[fromReg] != REG_NA)
{
regMaskTP fromRegMask = genRegMask(fromReg);
targetRegsReady |= fromRegMask;
}
}
if (targetRegsToDo != RBM_NONE)
{
regMaskTP targetRegMask = genFindLowestBit(targetRegsToDo);
regNumber targetReg = genRegNumFromMask(targetRegMask);
// Is it already there due to other moves?
// If not, move it to the temp reg, OR swap it with another register
regNumber sourceReg = (regNumber)source[targetReg];
regNumber fromReg = (regNumber)location[sourceReg];
if (targetReg == fromReg)
{
targetRegsToDo &= ~targetRegMask;
}
else
{
regNumber tempReg = REG_NA;
bool useSwap = false;
if (emitter::isFloatReg(targetReg))
{
#ifdef _TARGET_ARM_
if (sourceIntervals[fromReg]->registerType == TYP_DOUBLE)
{
// ARM32 requires a double temp register for TYP_DOUBLE.
// We tried to reserve a double temp register first, but sometimes we can't.
tempReg = genIsValidDoubleReg(tempRegFlt) ? tempRegFlt : REG_NA;
}
else
#endif // _TARGET_ARM_
tempReg = tempRegFlt;
}
#ifdef _TARGET_XARCH_
else
{
useSwap = true;
}
#else // !_TARGET_XARCH_
else
{
tempReg = tempRegInt;
}
#endif // !_TARGET_XARCH_
if (useSwap || tempReg == REG_NA)
{
// First, we have to figure out the destination register for what's currently in fromReg,
// so that we can find its sourceInterval.
regNumber otherTargetReg = REG_NA;
// By chance, is fromReg going where it belongs?
if (location[source[fromReg]] == targetReg)
{
otherTargetReg = fromReg;
// If we can swap, we will be done with otherTargetReg as well.
// Otherwise, we'll spill it to the stack and reload it later.
if (useSwap)
{
regMaskTP fromRegMask = genRegMask(fromReg);
targetRegsToDo &= ~fromRegMask;
}
}
else
{
// Look at the remaining registers from targetRegsToDo (which we expect to be relatively
// small at this point) to find out what's currently in targetReg.
regMaskTP mask = targetRegsToDo;
while (mask != RBM_NONE && otherTargetReg == REG_NA)
{
regMaskTP nextRegMask = genFindLowestBit(mask);
regNumber nextReg = genRegNumFromMask(nextRegMask);
mask &= ~nextRegMask;
if (location[source[nextReg]] == targetReg)
{
otherTargetReg = nextReg;
}
}
}
assert(otherTargetReg != REG_NA);
if (useSwap)
{
// Generate a "swap" of fromReg and targetReg
insertSwap(block, insertionPoint, sourceIntervals[source[otherTargetReg]]->varNum, targetReg,
sourceIntervals[sourceReg]->varNum, fromReg);
location[sourceReg] = REG_NA;
location[source[otherTargetReg]] = (regNumberSmall)fromReg;
INTRACK_STATS(updateLsraStat(LSRA_STAT_RESOLUTION_MOV, block->bbNum));
}
else
{
// Spill "targetReg" to the stack and add its eventual target (otherTargetReg)
// to "targetRegsFromStack", which will be handled below.
// NOTE: This condition is very rare. Setting COMPlus_JitStressRegs=0x203
// has been known to trigger it in JIT SH.
// First, spill "otherInterval" from targetReg to the stack.
Interval* otherInterval = sourceIntervals[source[otherTargetReg]];
setIntervalAsSpilled(otherInterval);
addResolution(block, insertionPoint, otherInterval, REG_STK, targetReg);
JITDUMP(" (%s)\n", resolveTypeName[resolveType]);
location[source[otherTargetReg]] = REG_STK;
// Now, move the interval that is going to targetReg, and add its "fromReg" to
// "targetRegsReady".
addResolution(block, insertionPoint, sourceIntervals[sourceReg], targetReg, fromReg);
JITDUMP(" (%s)\n", resolveTypeName[resolveType]);
location[sourceReg] = REG_NA;
targetRegsReady |= genRegMask(fromReg);
}
targetRegsToDo &= ~targetRegMask;
}
else
{
compiler->codeGen->regSet.rsSetRegsModified(genRegMask(tempReg) DEBUGARG(dumpTerse));
assert(sourceIntervals[targetReg] != nullptr);
addResolution(block, insertionPoint, sourceIntervals[targetReg], tempReg, targetReg);
JITDUMP(" (%s)\n", resolveTypeName[resolveType]);
location[targetReg] = (regNumberSmall)tempReg;
targetRegsReady |= targetRegMask;
}
}
}
}
// Finally, perform stack to reg moves
// All the target regs will be empty at this point
while (targetRegsFromStack != RBM_NONE)
{
regMaskTP targetRegMask = genFindLowestBit(targetRegsFromStack);
targetRegsFromStack &= ~targetRegMask;
regNumber targetReg = genRegNumFromMask(targetRegMask);
Interval* interval = stackToRegIntervals[targetReg];
assert(interval != nullptr);
addResolution(block, insertionPoint, interval, targetReg, REG_STK);
JITDUMP(" (%s)\n", resolveTypeName[resolveType]);
}
}
//------------------------------------------------------------------------
// GetIndirSourceCount: Get the source registers for an indirection that might be contained.
//
// Arguments:
// node - The node of interest
//
// Return Value:
// The number of source registers used by the *parent* of this node.
//
int LinearScan::GetIndirSourceCount(GenTreeIndir* indirTree)
{
GenTree* const addr = indirTree->gtOp1;
if (!addr->isContained())
{
return 1;
}
if (!addr->OperIs(GT_LEA))
{
return 0;
}
GenTreeAddrMode* const addrMode = addr->AsAddrMode();
unsigned srcCount = 0;
if ((addrMode->Base() != nullptr) && !addrMode->Base()->isContained())
{
srcCount++;
}
if (addrMode->Index() != nullptr)
{
// We never have a contained index.
assert(!addrMode->Index()->isContained());
srcCount++;
}
return srcCount;
}
void TreeNodeInfo::Initialize(LinearScan* lsra, GenTree* node, LsraLocation location)
{
regMaskTP dstCandidates;
// if there is a reg indicated on the tree node, use that for dstCandidates
// the exception is the NOP, which sometimes show up around late args.
// TODO-Cleanup: get rid of those NOPs.
if (node->gtRegNum == REG_STK)
{
dstCandidates = RBM_NONE;
}
else if (node->gtRegNum == REG_NA || node->gtOper == GT_NOP)
{
#ifdef ARM_SOFTFP
if (node->OperGet() == GT_PUTARG_REG)
{
dstCandidates = lsra->allRegs(TYP_INT);
}
else
#endif
{
dstCandidates = lsra->allRegs(node->TypeGet());
}
}
else
{
dstCandidates = genRegMask(node->gtRegNum);
}
internalIntCount = 0;
internalFloatCount = 0;
isLocalDefUse = false;
isLsraAdded = false;
definesAnyRegisters = false;
setDstCandidates(lsra, dstCandidates);
srcCandsIndex = dstCandsIndex;
setInternalCandidates(lsra, lsra->allRegs(TYP_INT));
loc = location;
#ifdef DEBUG
isInitialized = true;
#endif
assert(IsValid(lsra));
}
regMaskTP TreeNodeInfo::getSrcCandidates(LinearScan* lsra)
{
return lsra->GetRegMaskForIndex(srcCandsIndex);
}
void TreeNodeInfo::setSrcCandidates(LinearScan* lsra, regMaskTP mask)
{
LinearScan::RegMaskIndex i = lsra->GetIndexForRegMask(mask);
assert(FitsIn<unsigned char>(i));
srcCandsIndex = (unsigned char)i;
}
regMaskTP TreeNodeInfo::getDstCandidates(LinearScan* lsra)
{
return lsra->GetRegMaskForIndex(dstCandsIndex);
}
void TreeNodeInfo::setDstCandidates(LinearScan* lsra, regMaskTP mask)
{
LinearScan::RegMaskIndex i = lsra->GetIndexForRegMask(mask);
assert(FitsIn<unsigned char>(i));
dstCandsIndex = (unsigned char)i;
}
regMaskTP TreeNodeInfo::getInternalCandidates(LinearScan* lsra)
{
return lsra->GetRegMaskForIndex(internalCandsIndex);
}
void TreeNodeInfo::setInternalCandidates(LinearScan* lsra, regMaskTP mask)
{
LinearScan::RegMaskIndex i = lsra->GetIndexForRegMask(mask);
assert(FitsIn<unsigned char>(i));
internalCandsIndex = (unsigned char)i;
}
void TreeNodeInfo::addInternalCandidates(LinearScan* lsra, regMaskTP mask)
{
LinearScan::RegMaskIndex i = lsra->GetIndexForRegMask(lsra->GetRegMaskForIndex(internalCandsIndex) | mask);
assert(FitsIn<unsigned char>(i));
internalCandsIndex = (unsigned char)i;
}
#if TRACK_LSRA_STATS
// ----------------------------------------------------------
// updateLsraStat: Increment LSRA stat counter.
//
// Arguments:
// stat - LSRA stat enum
// bbNum - Basic block to which LSRA stat needs to be
// associated with.
//
void LinearScan::updateLsraStat(LsraStat stat, unsigned bbNum)
{
if (bbNum > bbNumMaxBeforeResolution)
{
// This is a newly created basic block as part of resolution.
// These blocks contain resolution moves that are already accounted.
return;
}
switch (stat)
{
case LSRA_STAT_SPILL:
++(blockInfo[bbNum].spillCount);
break;
case LSRA_STAT_COPY_REG:
++(blockInfo[bbNum].copyRegCount);
break;
case LSRA_STAT_RESOLUTION_MOV:
++(blockInfo[bbNum].resolutionMovCount);
break;
case LSRA_STAT_SPLIT_EDGE:
++(blockInfo[bbNum].splitEdgeCount);
break;
default:
break;
}
}
// -----------------------------------------------------------
// dumpLsraStats - dumps Lsra stats to given file.
//
// Arguments:
// file - file to which stats are to be written.
//
void LinearScan::dumpLsraStats(FILE* file)
{
unsigned sumSpillCount = 0;
unsigned sumCopyRegCount = 0;
unsigned sumResolutionMovCount = 0;
unsigned sumSplitEdgeCount = 0;
UINT64 wtdSpillCount = 0;
UINT64 wtdCopyRegCount = 0;
UINT64 wtdResolutionMovCount = 0;
fprintf(file, "----------\n");
fprintf(file, "LSRA Stats");
#ifdef DEBUG
if (!VERBOSE)
{
fprintf(file, " : %s\n", compiler->info.compFullName);
}
else
{
// In verbose mode no need to print full name
// while printing lsra stats.
fprintf(file, "\n");
}
#else
fprintf(file, " : %s\n", compiler->eeGetMethodFullName(compiler->info.compCompHnd));
#endif
fprintf(file, "----------\n");
for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = block->bbNext)
{
if (block->bbNum > bbNumMaxBeforeResolution)
{
continue;
}
unsigned spillCount = blockInfo[block->bbNum].spillCount;
unsigned copyRegCount = blockInfo[block->bbNum].copyRegCount;
unsigned resolutionMovCount = blockInfo[block->bbNum].resolutionMovCount;
unsigned splitEdgeCount = blockInfo[block->bbNum].splitEdgeCount;
if (spillCount != 0 || copyRegCount != 0 || resolutionMovCount != 0 || splitEdgeCount != 0)
{
fprintf(file, "BB%02u [%8d]: ", block->bbNum, block->bbWeight);
fprintf(file, "SpillCount = %d, ResolutionMovs = %d, SplitEdges = %d, CopyReg = %d\n", spillCount,
resolutionMovCount, splitEdgeCount, copyRegCount);
}
sumSpillCount += spillCount;
sumCopyRegCount += copyRegCount;
sumResolutionMovCount += resolutionMovCount;
sumSplitEdgeCount += splitEdgeCount;
wtdSpillCount += (UINT64)spillCount * block->bbWeight;
wtdCopyRegCount += (UINT64)copyRegCount * block->bbWeight;
wtdResolutionMovCount += (UINT64)resolutionMovCount * block->bbWeight;
}
fprintf(file, "Total Tracked Vars: %d\n", compiler->lvaTrackedCount);
fprintf(file, "Total Reg Cand Vars: %d\n", regCandidateVarCount);
fprintf(file, "Total number of Intervals: %d\n", static_cast<unsigned>(intervals.size() - 1));
fprintf(file, "Total number of RefPositions: %d\n", static_cast<unsigned>(refPositions.size() - 1));
fprintf(file, "Total Spill Count: %d Weighted: %I64u\n", sumSpillCount, wtdSpillCount);
fprintf(file, "Total CopyReg Count: %d Weighted: %I64u\n", sumCopyRegCount, wtdCopyRegCount);
fprintf(file, "Total ResolutionMov Count: %d Weighted: %I64u\n", sumResolutionMovCount, wtdResolutionMovCount);
fprintf(file, "Total number of split edges: %d\n", sumSplitEdgeCount);
// compute total number of spill temps created
unsigned numSpillTemps = 0;
for (int i = 0; i < TYP_COUNT; i++)
{
numSpillTemps += maxSpill[i];
}
fprintf(file, "Total Number of spill temps created: %d\n\n", numSpillTemps);
}
#endif // TRACK_LSRA_STATS
#ifdef DEBUG
void dumpRegMask(regMaskTP regs)
{
if (regs == RBM_ALLINT)
{
printf("[allInt]");
}
else if (regs == (RBM_ALLINT & ~RBM_FPBASE))
{
printf("[allIntButFP]");
}
else if (regs == RBM_ALLFLOAT)
{
printf("[allFloat]");
}
else if (regs == RBM_ALLDOUBLE)
{
printf("[allDouble]");
}
else
{
dspRegMask(regs);
}
}
static const char* getRefTypeName(RefType refType)
{
switch (refType)
{
#define DEF_REFTYPE(memberName, memberValue, shortName) \
case memberName: \
return #memberName;
#include "lsra_reftypes.h"
#undef DEF_REFTYPE
default:
return nullptr;
}
}
static const char* getRefTypeShortName(RefType refType)
{
switch (refType)
{
#define DEF_REFTYPE(memberName, memberValue, shortName) \
case memberName: \
return shortName;
#include "lsra_reftypes.h"
#undef DEF_REFTYPE
default:
return nullptr;
}
}
void RefPosition::dump()
{
printf("<RefPosition #%-3u @%-3u", rpNum, nodeLocation);
if (nextRefPosition)
{
printf(" ->#%-3u", nextRefPosition->rpNum);
}
printf(" %s ", getRefTypeName(refType));
if (this->isPhysRegRef)
{
this->getReg()->tinyDump();
}
else if (getInterval())
{
this->getInterval()->tinyDump();
}
if (this->treeNode)
{
printf("%s ", treeNode->OpName(treeNode->OperGet()));
}
printf("BB%02u ", this->bbNum);
printf("regmask=");
dumpRegMask(registerAssignment);
if (this->lastUse)
{
printf(" last");
}
if (this->reload)
{
printf(" reload");
}
if (this->spillAfter)
{
printf(" spillAfter");
}
if (this->moveReg)
{
printf(" move");
}
if (this->copyReg)
{
printf(" copy");
}
if (this->isFixedRegRef)
{
printf(" fixed");
}
if (this->isLocalDefUse)
{
printf(" local");
}
if (this->delayRegFree)
{
printf(" delay");
}
if (this->outOfOrder)
{
printf(" outOfOrder");
}
if (this->AllocateIfProfitable())
{
printf(" regOptional");
}
printf(">\n");
}
void RegRecord::dump()
{
tinyDump();
}
void Interval::dump()
{
printf("Interval %2u:", intervalIndex);
if (isLocalVar)
{
printf(" (V%02u)", varNum);
}
if (isInternal)
{
printf(" (INTERNAL)");
}
if (isSpilled)
{
printf(" (SPILLED)");
}
if (isSplit)
{
printf(" (SPLIT)");
}
if (isStructField)
{
printf(" (struct)");
}
if (isSpecialPutArg)
{
printf(" (specialPutArg)");
}
if (isConstant)
{
printf(" (constant)");
}
printf(" RefPositions {");
for (RefPosition* refPosition = this->firstRefPosition; refPosition != nullptr;
refPosition = refPosition->nextRefPosition)
{
printf("#%u@%u", refPosition->rpNum, refPosition->nodeLocation);
if (refPosition->nextRefPosition)
{
printf(" ");
}
}
printf("}");
// this is not used (yet?)
// printf(" SpillOffset %d", this->spillOffset);
printf(" physReg:%s", getRegName(physReg));
printf(" Preferences=");
dumpRegMask(this->registerPreferences);
if (relatedInterval)
{
printf(" RelatedInterval ");
relatedInterval->microDump();
printf("[%p]", dspPtr(relatedInterval));
}
printf("\n");
}
// print out very concise representation
void Interval::tinyDump()
{
printf("<Ivl:%u", intervalIndex);
if (isLocalVar)
{
printf(" V%02u", varNum);
}
if (isInternal)
{
printf(" internal");
}
printf("> ");
}
// print out extremely concise representation
void Interval::microDump()
{
char intervalTypeChar = 'I';
if (isInternal)
{
intervalTypeChar = 'T';
}
else if (isLocalVar)
{
intervalTypeChar = 'L';
}
printf("<%c%u>", intervalTypeChar, intervalIndex);
}
void RegRecord::tinyDump()
{
printf("<Reg:%-3s> ", getRegName(regNum));
}
void TreeNodeInfo::dump(LinearScan* lsra)
{
printf("<TreeNodeInfo @ %2u %d=%d %di %df", loc, dstCount, srcCount, internalIntCount, internalFloatCount);
printf(" src=");
dumpRegMask(getSrcCandidates(lsra));
printf(" int=");
dumpRegMask(getInternalCandidates(lsra));
printf(" dst=");
dumpRegMask(getDstCandidates(lsra));
if (isLocalDefUse)
{
printf(" L");
}
if (isInitialized)
{
printf(" I");
}
if (isLsraAdded)
{
printf(" A");
}
if (isDelayFree)
{
printf(" D");
}
if (isTgtPref)
{
printf(" P");
}
if (regOptional)
{
printf(" O");
}
if (isInternalRegDelayFree)
{
printf(" ID");
}
printf(">\n");
}
void LinearScan::lsraDumpIntervals(const char* msg)
{
Interval* interval;
printf("\nLinear scan intervals %s:\n", msg);
for (auto& interval : intervals)
{
// only dump something if it has references
// if (interval->firstRefPosition)
interval.dump();
}
printf("\n");
}
// Dumps a tree node as a destination or source operand, with the style
// of dump dependent on the mode
void LinearScan::lsraGetOperandString(GenTreePtr tree,
LsraTupleDumpMode mode,
char* operandString,
unsigned operandStringLength)
{
const char* lastUseChar = "";
if ((tree->gtFlags & GTF_VAR_DEATH) != 0)
{
lastUseChar = "*";
}
switch (mode)
{
case LinearScan::LSRA_DUMP_PRE:
_snprintf_s(operandString, operandStringLength, operandStringLength, "t%d%s", tree->gtTreeID, lastUseChar);
break;
case LinearScan::LSRA_DUMP_REFPOS:
_snprintf_s(operandString, operandStringLength, operandStringLength, "t%d%s", tree->gtTreeID, lastUseChar);
break;
case LinearScan::LSRA_DUMP_POST:
{
Compiler* compiler = JitTls::GetCompiler();
if (!tree->gtHasReg())
{
_snprintf_s(operandString, operandStringLength, operandStringLength, "STK%s", lastUseChar);
}
else
{
_snprintf_s(operandString, operandStringLength, operandStringLength, "%s%s",
getRegName(tree->gtRegNum, useFloatReg(tree->TypeGet())), lastUseChar);
}
}
break;
default:
printf("ERROR: INVALID TUPLE DUMP MODE\n");
break;
}
}
void LinearScan::lsraDispNode(GenTreePtr tree, LsraTupleDumpMode mode, bool hasDest)
{
Compiler* compiler = JitTls::GetCompiler();
const unsigned operandStringLength = 16;
char operandString[operandStringLength];
const char* emptyDestOperand = " ";
char spillChar = ' ';
if (mode == LinearScan::LSRA_DUMP_POST)
{
if ((tree->gtFlags & GTF_SPILL) != 0)
{
spillChar = 'S';
}
if (!hasDest && tree->gtHasReg())
{
// A node can define a register, but not produce a value for a parent to consume,
// i.e. in the "localDefUse" case.
// There used to be an assert here that we wouldn't spill such a node.
// However, we can have unused lclVars that wind up being the node at which
// it is spilled. This probably indicates a bug, but we don't realy want to
// assert during a dump.
if (spillChar == 'S')
{
spillChar = '$';
}
else
{
spillChar = '*';
}
hasDest = true;
}
}
printf("%c N%03u. ", spillChar, tree->gtSeqNum);
LclVarDsc* varDsc = nullptr;
unsigned varNum = UINT_MAX;
if (tree->IsLocal())
{
varNum = tree->gtLclVarCommon.gtLclNum;
varDsc = &(compiler->lvaTable[varNum]);
if (varDsc->lvLRACandidate)
{
hasDest = false;
}
}
if (hasDest)
{
if (mode == LinearScan::LSRA_DUMP_POST && tree->gtFlags & GTF_SPILLED)
{
assert(tree->gtHasReg());
}
lsraGetOperandString(tree, mode, operandString, operandStringLength);
printf("%-15s =", operandString);
}
else
{
printf("%-15s ", emptyDestOperand);
}
if (varDsc != nullptr)
{
if (varDsc->lvLRACandidate)
{
if (mode == LSRA_DUMP_REFPOS)
{
printf(" V%02u(L%d)", varNum, getIntervalForLocalVar(varDsc->lvVarIndex)->intervalIndex);
}
else
{
lsraGetOperandString(tree, mode, operandString, operandStringLength);
printf(" V%02u(%s)", varNum, operandString);
if (mode == LinearScan::LSRA_DUMP_POST && tree->gtFlags & GTF_SPILLED)
{
printf("R");
}
}
}
else
{
printf(" V%02u MEM", varNum);
}
}
else if (tree->OperIsAssignment())
{
assert(!tree->gtHasReg());
printf(" asg%s ", GenTree::OpName(tree->OperGet()));
}
else
{
compiler->gtDispNodeName(tree);
if (tree->OperKind() & GTK_LEAF)
{
compiler->gtDispLeaf(tree, nullptr);
}
}
}
//------------------------------------------------------------------------
// ComputeOperandDstCount: computes the number of registers defined by a
// node.
//
// For most nodes, this is simple:
// - Nodes that do not produce values (e.g. stores and other void-typed
// nodes) and nodes that immediately use the registers they define
// produce no registers
// - Nodes that are marked as defining N registers define N registers.
//
// For contained nodes, however, things are more complicated: for purposes
// of bookkeeping, a contained node is treated as producing the transitive
// closure of the registers produced by its sources.
//
// Arguments:
// operand - The operand for which to compute a register count.
//
// Returns:
// The number of registers defined by `operand`.
//
void LinearScan::DumpOperandDefs(
GenTree* operand, bool& first, LsraTupleDumpMode mode, char* operandString, const unsigned operandStringLength)
{
assert(operand != nullptr);
assert(operandString != nullptr);
if (ComputeOperandDstCount(operand) == 0)
{
return;
}
if (operand->gtLsraInfo.dstCount != 0)
{
// This operand directly produces registers; print it.
for (int i = 0; i < operand->gtLsraInfo.dstCount; i++)
{
if (!first)
{
printf(",");
}
lsraGetOperandString(operand, mode, operandString, operandStringLength);
printf("%s", operandString);
first = false;
}
}
else
{
// This is a contained node. Dump the defs produced by its operands.
for (GenTree* op : operand->Operands())
{
DumpOperandDefs(op, first, mode, operandString, operandStringLength);
}
}
}
void LinearScan::TupleStyleDump(LsraTupleDumpMode mode)
{
BasicBlock* block;
LsraLocation currentLoc = 1; // 0 is the entry
const unsigned operandStringLength = 16;
char operandString[operandStringLength];
// currentRefPosition is not used for LSRA_DUMP_PRE
// We keep separate iterators for defs, so that we can print them
// on the lhs of the dump
auto currentRefPosition = refPositions.begin();
switch (mode)
{
case LSRA_DUMP_PRE:
printf("TUPLE STYLE DUMP BEFORE LSRA\n");
break;
case LSRA_DUMP_REFPOS:
printf("TUPLE STYLE DUMP WITH REF POSITIONS\n");
break;
case LSRA_DUMP_POST:
printf("TUPLE STYLE DUMP WITH REGISTER ASSIGNMENTS\n");
break;
default:
printf("ERROR: INVALID TUPLE DUMP MODE\n");
return;
}
if (mode != LSRA_DUMP_PRE)
{
printf("Incoming Parameters: ");
for (; currentRefPosition != refPositions.end() && currentRefPosition->refType != RefTypeBB;
++currentRefPosition)
{
Interval* interval = currentRefPosition->getInterval();
assert(interval != nullptr && interval->isLocalVar);
printf(" V%02d", interval->varNum);
if (mode == LSRA_DUMP_POST)
{
regNumber reg;
if (currentRefPosition->registerAssignment == RBM_NONE)
{
reg = REG_STK;
}
else
{
reg = currentRefPosition->assignedReg();
}
LclVarDsc* varDsc = &(compiler->lvaTable[interval->varNum]);
printf("(");
regNumber assignedReg = varDsc->lvRegNum;
regNumber argReg = (varDsc->lvIsRegArg) ? varDsc->lvArgReg : REG_STK;
assert(reg == assignedReg || varDsc->lvRegister == false);
if (reg != argReg)
{
printf(getRegName(argReg, isFloatRegType(interval->registerType)));
printf("=>");
}
printf("%s)", getRegName(reg, isFloatRegType(interval->registerType)));
}
}
printf("\n");
}
for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock())
{
currentLoc += 2;
if (mode == LSRA_DUMP_REFPOS)
{
bool printedBlockHeader = false;
// We should find the boundary RefPositions in the order of exposed uses, dummy defs, and the blocks
for (; currentRefPosition != refPositions.end() &&
(currentRefPosition->refType == RefTypeExpUse || currentRefPosition->refType == RefTypeDummyDef ||
(currentRefPosition->refType == RefTypeBB && !printedBlockHeader));
++currentRefPosition)
{
Interval* interval = nullptr;
if (currentRefPosition->isIntervalRef())
{
interval = currentRefPosition->getInterval();
}
switch (currentRefPosition->refType)
{
case RefTypeExpUse:
assert(interval != nullptr);
assert(interval->isLocalVar);
printf(" Exposed use of V%02u at #%d\n", interval->varNum, currentRefPosition->rpNum);
break;
case RefTypeDummyDef:
assert(interval != nullptr);
assert(interval->isLocalVar);
printf(" Dummy def of V%02u at #%d\n", interval->varNum, currentRefPosition->rpNum);
break;
case RefTypeBB:
block->dspBlockHeader(compiler);
printedBlockHeader = true;
printf("=====\n");
break;
default:
printf("Unexpected RefPosition type at #%d\n", currentRefPosition->rpNum);
break;
}
}
}
else
{
block->dspBlockHeader(compiler);
printf("=====\n");
}
if (enregisterLocalVars && mode == LSRA_DUMP_POST && block != compiler->fgFirstBB &&
block->bbNum <= bbNumMaxBeforeResolution)
{
printf("Predecessor for variable locations: BB%02u\n", blockInfo[block->bbNum].predBBNum);
dumpInVarToRegMap(block);
}
if (block->bbNum > bbNumMaxBeforeResolution)
{
SplitEdgeInfo splitEdgeInfo;
splitBBNumToTargetBBNumMap->Lookup(block->bbNum, &splitEdgeInfo);
assert(splitEdgeInfo.toBBNum <= bbNumMaxBeforeResolution);
assert(splitEdgeInfo.fromBBNum <= bbNumMaxBeforeResolution);
printf("New block introduced for resolution from BB%02u to BB%02u\n", splitEdgeInfo.fromBBNum,
splitEdgeInfo.toBBNum);
}
for (GenTree* node : LIR::AsRange(block).NonPhiNodes())
{
GenTree* tree = node;
genTreeOps oper = tree->OperGet();
TreeNodeInfo& info = tree->gtLsraInfo;
if (tree->gtLsraInfo.isLsraAdded)
{
// This must be one of the nodes that we add during LSRA
if (oper == GT_LCL_VAR)
{
info.srcCount = 0;
info.dstCount = 1;
}
else if (oper == GT_RELOAD || oper == GT_COPY)
{
info.srcCount = 1;
info.dstCount = 1;
}
#ifdef FEATURE_SIMD
else if (oper == GT_SIMD)
{
if (tree->gtSIMD.gtSIMDIntrinsicID == SIMDIntrinsicUpperSave)
{
info.srcCount = 1;
info.dstCount = 1;
}
else
{
assert(tree->gtSIMD.gtSIMDIntrinsicID == SIMDIntrinsicUpperRestore);
info.srcCount = 2;
info.dstCount = 0;
}
}
#endif // FEATURE_SIMD
else
{
assert(oper == GT_SWAP);
info.srcCount = 2;
info.dstCount = 0;
}
info.internalIntCount = 0;
info.internalFloatCount = 0;
}
int consume = info.srcCount;
int produce = info.dstCount;
regMaskTP killMask = RBM_NONE;
regMaskTP fixedMask = RBM_NONE;
lsraDispNode(tree, mode, produce != 0 && mode != LSRA_DUMP_REFPOS);
if (mode != LSRA_DUMP_REFPOS)
{
if (consume > 0)
{
printf("; ");
bool first = true;
for (GenTree* operand : tree->Operands())
{
DumpOperandDefs(operand, first, mode, operandString, operandStringLength);
}
}
}
else
{
// Print each RefPosition on a new line, but
// printing all the kills for each node on a single line
// and combining the fixed regs with their associated def or use
bool killPrinted = false;
RefPosition* lastFixedRegRefPos = nullptr;
for (; currentRefPosition != refPositions.end() &&
(currentRefPosition->refType == RefTypeUse || currentRefPosition->refType == RefTypeFixedReg ||
currentRefPosition->refType == RefTypeKill || currentRefPosition->refType == RefTypeDef) &&
(currentRefPosition->nodeLocation == tree->gtSeqNum ||
currentRefPosition->nodeLocation == tree->gtSeqNum + 1);
++currentRefPosition)
{
Interval* interval = nullptr;
if (currentRefPosition->isIntervalRef())
{
interval = currentRefPosition->getInterval();
}
switch (currentRefPosition->refType)
{
case RefTypeUse:
if (currentRefPosition->isPhysRegRef)
{
printf("\n Use:R%d(#%d)",
currentRefPosition->getReg()->regNum, currentRefPosition->rpNum);
}
else
{
assert(interval != nullptr);
printf("\n Use:");
interval->microDump();
printf("(#%d)", currentRefPosition->rpNum);
if (currentRefPosition->isFixedRegRef)
{
assert(genMaxOneBit(currentRefPosition->registerAssignment));
assert(lastFixedRegRefPos != nullptr);
printf(" Fixed:%s(#%d)", getRegName(currentRefPosition->assignedReg(),
isFloatRegType(interval->registerType)),
lastFixedRegRefPos->rpNum);
lastFixedRegRefPos = nullptr;
}
if (currentRefPosition->isLocalDefUse)
{
printf(" LocalDefUse");
}
if (currentRefPosition->lastUse)
{
printf(" *");
}
}
break;
case RefTypeDef:
{
// Print each def on a new line
assert(interval != nullptr);
printf("\n Def:");
interval->microDump();
printf("(#%d)", currentRefPosition->rpNum);
if (currentRefPosition->isFixedRegRef)
{
assert(genMaxOneBit(currentRefPosition->registerAssignment));
printf(" %s", getRegName(currentRefPosition->assignedReg(),
isFloatRegType(interval->registerType)));
}
if (currentRefPosition->isLocalDefUse)
{
printf(" LocalDefUse");
}
if (currentRefPosition->lastUse)
{
printf(" *");
}
if (interval->relatedInterval != nullptr)
{
printf(" Pref:");
interval->relatedInterval->microDump();
}
}
break;
case RefTypeKill:
if (!killPrinted)
{
printf("\n Kill: ");
killPrinted = true;
}
printf(getRegName(currentRefPosition->assignedReg(),
isFloatRegType(currentRefPosition->getReg()->registerType)));
printf(" ");
break;
case RefTypeFixedReg:
lastFixedRegRefPos = currentRefPosition;
break;
default:
printf("Unexpected RefPosition type at #%d\n", currentRefPosition->rpNum);
break;
}
}
}
printf("\n");
if (info.internalIntCount != 0 && mode != LSRA_DUMP_REFPOS)
{
printf("\tinternal (%d):\t", info.internalIntCount);
if (mode == LSRA_DUMP_POST)
{
dumpRegMask(tree->gtRsvdRegs);
}
else if ((info.getInternalCandidates(this) & allRegs(TYP_INT)) != allRegs(TYP_INT))
{
dumpRegMask(info.getInternalCandidates(this) & allRegs(TYP_INT));
}
printf("\n");
}
if (info.internalFloatCount != 0 && mode != LSRA_DUMP_REFPOS)
{
printf("\tinternal (%d):\t", info.internalFloatCount);
if (mode == LSRA_DUMP_POST)
{
dumpRegMask(tree->gtRsvdRegs);
}
else if ((info.getInternalCandidates(this) & allRegs(TYP_INT)) != allRegs(TYP_INT))
{
dumpRegMask(info.getInternalCandidates(this) & allRegs(TYP_INT));
}
printf("\n");
}
}
if (enregisterLocalVars && mode == LSRA_DUMP_POST)
{
dumpOutVarToRegMap(block);
}
printf("\n");
}
printf("\n\n");
}
void LinearScan::dumpLsraAllocationEvent(LsraDumpEvent event,
Interval* interval,
regNumber reg,
BasicBlock* currentBlock)
{
if (!(VERBOSE))
{
return;
}
switch (event)
{
// Conflicting def/use
case LSRA_EVENT_DEFUSE_CONFLICT:
if (!dumpTerse)
{
printf(" Def and Use have conflicting register requirements:");
}
else
{
printf("DUconflict ");
dumpRegRecords();
}
break;
case LSRA_EVENT_DEFUSE_FIXED_DELAY_USE:
if (!dumpTerse)
{
printf(" Can't change useAssignment ");
}
break;
case LSRA_EVENT_DEFUSE_CASE1:
if (!dumpTerse)
{
printf(" case #1, use the defRegAssignment\n");
}
else
{
printf(indentFormat, " case #1 use defRegAssignment");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_DEFUSE_CASE2:
if (!dumpTerse)
{
printf(" case #2, use the useRegAssignment\n");
}
else
{
printf(indentFormat, " case #2 use useRegAssignment");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_DEFUSE_CASE3:
if (!dumpTerse)
{
printf(" case #3, change the defRegAssignment to the use regs\n");
}
else
{
printf(indentFormat, " case #3 use useRegAssignment");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_DEFUSE_CASE4:
if (!dumpTerse)
{
printf(" case #4, change the useRegAssignment to the def regs\n");
}
else
{
printf(indentFormat, " case #4 use defRegAssignment");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_DEFUSE_CASE5:
if (!dumpTerse)
{
printf(" case #5, Conflicting Def and Use single-register requirements require copies - set def to all "
"regs of the appropriate type\n");
}
else
{
printf(indentFormat, " case #5 set def to all regs");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_DEFUSE_CASE6:
if (!dumpTerse)
{
printf(" case #6, Conflicting Def and Use register requirements require a copy\n");
}
else
{
printf(indentFormat, " case #6 need a copy");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_SPILL:
if (!dumpTerse)
{
printf("Spilled:\n");
interval->dump();
}
else
{
assert(interval != nullptr && interval->assignedReg != nullptr);
printf("Spill %-4s ", getRegName(interval->assignedReg->regNum));
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_SPILL_EXTENDED_LIFETIME:
if (!dumpTerse)
{
printf(" Spilled extended lifetime var V%02u at last use; not marked for actual spill.",
interval->intervalIndex);
}
break;
// Restoring the previous register
case LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL_AFTER_SPILL:
assert(interval != nullptr);
if (!dumpTerse)
{
printf(" Assign register %s to previous interval Ivl:%d after spill\n", getRegName(reg),
interval->intervalIndex);
}
else
{
// If we spilled, then the dump is already pre-indented, but we need to pre-indent for the subsequent
// allocation
// with a dumpEmptyRefPosition().
printf("SRstr %-4s ", getRegName(reg));
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL:
assert(interval != nullptr);
if (!dumpTerse)
{
printf(" Assign register %s to previous interval Ivl:%d\n", getRegName(reg), interval->intervalIndex);
}
else
{
if (activeRefPosition == nullptr)
{
printf(emptyRefPositionFormat, "");
}
printf("Restr %-4s ", getRegName(reg));
dumpRegRecords();
if (activeRefPosition != nullptr)
{
printf(emptyRefPositionFormat, "");
}
}
break;
// Done with GC Kills
case LSRA_EVENT_DONE_KILL_GC_REFS:
printf("DoneKillGC ");
break;
// Block boundaries
case LSRA_EVENT_START_BB:
assert(currentBlock != nullptr);
if (!dumpTerse)
{
printf("\n\n Live Vars(Regs) at start of BB%02u (from pred BB%02u):", currentBlock->bbNum,
blockInfo[currentBlock->bbNum].predBBNum);
dumpVarToRegMap(inVarToRegMaps[currentBlock->bbNum]);
}
break;
case LSRA_EVENT_END_BB:
if (!dumpTerse)
{
printf("\n\n Live Vars(Regs) after BB%02u:", currentBlock->bbNum);
dumpVarToRegMap(outVarToRegMaps[currentBlock->bbNum]);
}
break;
case LSRA_EVENT_FREE_REGS:
if (!dumpTerse)
{
printf("Freeing registers:\n");
}
break;
// Characteristics of the current RefPosition
case LSRA_EVENT_INCREMENT_RANGE_END:
if (!dumpTerse)
{
printf(" Incrementing nextPhysRegLocation for %s\n", getRegName(reg));
}
// else ???
break;
case LSRA_EVENT_LAST_USE:
if (!dumpTerse)
{
printf(" Last use, marked to be freed\n");
}
break;
case LSRA_EVENT_LAST_USE_DELAYED:
if (!dumpTerse)
{
printf(" Last use, marked to be freed (delayed)\n");
}
break;
case LSRA_EVENT_NEEDS_NEW_REG:
if (!dumpTerse)
{
printf(" Needs new register; mark %s to be freed\n", getRegName(reg));
}
else
{
printf("Free %-4s ", getRegName(reg));
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
// Allocation decisions
case LSRA_EVENT_FIXED_REG:
case LSRA_EVENT_EXP_USE:
if (!dumpTerse)
{
printf("No allocation\n");
}
else
{
printf("Keep %-4s ", getRegName(reg));
}
break;
case LSRA_EVENT_ZERO_REF:
assert(interval != nullptr && interval->isLocalVar);
if (!dumpTerse)
{
printf("Marking V%02u as last use there are no actual references\n", interval->varNum);
}
else
{
printf("NoRef ");
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_KEPT_ALLOCATION:
if (!dumpTerse)
{
printf("already allocated %4s\n", getRegName(reg));
}
else
{
printf("Keep %-4s ", getRegName(reg));
}
break;
case LSRA_EVENT_COPY_REG:
assert(interval != nullptr && interval->recentRefPosition != nullptr);
if (!dumpTerse)
{
printf("allocated %s as copyReg\n\n", getRegName(reg));
}
else
{
printf("Copy %-4s ", getRegName(reg));
}
break;
case LSRA_EVENT_MOVE_REG:
assert(interval != nullptr && interval->recentRefPosition != nullptr);
if (!dumpTerse)
{
printf(" needs a new register; marked as moveReg\n");
}
else
{
printf("Move %-4s ", getRegName(reg));
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_ALLOC_REG:
if (!dumpTerse)
{
printf("allocated %s\n", getRegName(reg));
}
else
{
printf("Alloc %-4s ", getRegName(reg));
}
break;
case LSRA_EVENT_REUSE_REG:
if (!dumpTerse)
{
printf("reused constant in %s\n", getRegName(reg));
}
else
{
printf("Reuse %-4s ", getRegName(reg));
}
break;
case LSRA_EVENT_ALLOC_SPILLED_REG:
if (!dumpTerse)
{
printf("allocated spilled register %s\n", getRegName(reg));
}
else
{
printf("Steal %-4s ", getRegName(reg));
}
break;
case LSRA_EVENT_NO_ENTRY_REG_ALLOCATED:
assert(interval != nullptr && interval->isLocalVar);
if (!dumpTerse)
{
printf("Not allocating an entry register for V%02u due to low ref count\n", interval->varNum);
}
else
{
printf("LoRef ");
}
break;
case LSRA_EVENT_NO_REG_ALLOCATED:
if (!dumpTerse)
{
printf("no register allocated\n");
}
else
{
printf("NoReg ");
}
break;
case LSRA_EVENT_RELOAD:
if (!dumpTerse)
{
printf(" Marked for reload\n");
}
else
{
printf("ReLod %-4s ", getRegName(reg));
dumpRegRecords();
dumpEmptyRefPosition();
}
break;
case LSRA_EVENT_SPECIAL_PUTARG:
if (!dumpTerse)
{
printf(" Special case of putArg - using lclVar that's in the expected reg\n");
}
else
{
printf("PtArg %-4s ", getRegName(reg));
}
break;
default:
break;
}
}
//------------------------------------------------------------------------
// dumpRegRecordHeader: Dump the header for a column-based dump of the register state.
//
// Arguments:
// None.
//
// Return Value:
// None.
//
// Assumptions:
// Reg names fit in 4 characters (minimum width of the columns)
//
// Notes:
// In order to make the table as dense as possible (for ease of reading the dumps),
// we determine the minimum regColumnWidth width required to represent:
// regs, by name (e.g. eax or xmm0) - this is fixed at 4 characters.
// intervals, as Vnn for lclVar intervals, or as I<num> for other intervals.
// The table is indented by the amount needed for dumpRefPositionShort, which is
// captured in shortRefPositionDumpWidth.
//
void LinearScan::dumpRegRecordHeader()
{
printf("The following table has one or more rows for each RefPosition that is handled during allocation.\n"
"The first column provides the basic information about the RefPosition, with its type (e.g. Def,\n"
"Use, Fixd) followed by a '*' if it is a last use, and a 'D' if it is delayRegFree, and then the\n"
"action taken during allocation (e.g. Alloc a new register, or Keep an existing one).\n"
"The subsequent columns show the Interval occupying each register, if any, followed by 'a' if it is\n"
"active, and 'i'if it is inactive. Columns are only printed up to the last modifed register, which\n"
"may increase during allocation, in which case additional columns will appear. Registers which are\n"
"not marked modified have ---- in their column.\n\n");
// First, determine the width of each register column (which holds a reg name in the
// header, and an interval name in each subsequent row).
int intervalNumberWidth = (int)log10((double)intervals.size()) + 1;
// The regColumnWidth includes the identifying character (I or V) and an 'i' or 'a' (inactive or active)
regColumnWidth = intervalNumberWidth + 2;
if (regColumnWidth < 4)
{
regColumnWidth = 4;
}
sprintf_s(intervalNameFormat, MAX_FORMAT_CHARS, "%%c%%-%dd", regColumnWidth - 2);
sprintf_s(regNameFormat, MAX_FORMAT_CHARS, "%%-%ds", regColumnWidth);
// Next, determine the width of the short RefPosition (see dumpRefPositionShort()).
// This is in the form:
// nnn.#mmm NAME TYPEld
// Where:
// nnn is the Location, right-justified to the width needed for the highest location.
// mmm is the RefPosition rpNum, left-justified to the width needed for the highest rpNum.
// NAME is dumped by dumpReferentName(), and is "regColumnWidth".
// TYPE is RefTypeNameShort, and is 4 characters
// l is either '*' (if a last use) or ' ' (otherwise)
// d is either 'D' (if a delayed use) or ' ' (otherwise)
maxNodeLocation = (maxNodeLocation == 0)
? 1
: maxNodeLocation; // corner case of a method with an infinite loop without any gentree nodes
assert(maxNodeLocation >= 1);
assert(refPositions.size() >= 1);
int nodeLocationWidth = (int)log10((double)maxNodeLocation) + 1;
int refPositionWidth = (int)log10((double)refPositions.size()) + 1;
int refTypeInfoWidth = 4 /*TYPE*/ + 2 /* last-use and delayed */ + 1 /* space */;
int locationAndRPNumWidth = nodeLocationWidth + 2 /* .# */ + refPositionWidth + 1 /* space */;
int shortRefPositionDumpWidth = locationAndRPNumWidth + regColumnWidth + 1 /* space */ + refTypeInfoWidth;
sprintf_s(shortRefPositionFormat, MAX_FORMAT_CHARS, "%%%dd.#%%-%dd ", nodeLocationWidth, refPositionWidth);
sprintf_s(emptyRefPositionFormat, MAX_FORMAT_CHARS, "%%-%ds", shortRefPositionDumpWidth);
// The width of the "allocation info"
// - a 5-character allocation decision
// - a space
// - a 4-character register
// - a space
int allocationInfoWidth = 5 + 1 + 4 + 1;
// Next, determine the width of the legend for each row. This includes:
// - a short RefPosition dump (shortRefPositionDumpWidth), which includes a space
// - the allocation info (allocationInfoWidth), which also includes a space
regTableIndent = shortRefPositionDumpWidth + allocationInfoWidth;
// BBnn printed left-justified in the NAME Typeld and allocationInfo space.
int bbDumpWidth = regColumnWidth + 1 + refTypeInfoWidth + allocationInfoWidth;
int bbNumWidth = (int)log10((double)compiler->fgBBNumMax) + 1;
// In the unlikely event that BB numbers overflow the space, we'll simply omit the predBB
int predBBNumDumpSpace = regTableIndent - locationAndRPNumWidth - bbNumWidth - 9; // 'BB' + ' PredBB'
if (predBBNumDumpSpace < bbNumWidth)
{
sprintf_s(bbRefPosFormat, MAX_LEGEND_FORMAT_CHARS, "BB%%-%dd", shortRefPositionDumpWidth - 2);
}
else
{
sprintf_s(bbRefPosFormat, MAX_LEGEND_FORMAT_CHARS, "BB%%-%dd PredBB%%-%dd", bbNumWidth, predBBNumDumpSpace);
}
if (compiler->shouldDumpASCIITrees())
{
columnSeparator = "|";
line = "-";
leftBox = "+";
middleBox = "+";
rightBox = "+";
}
else
{
columnSeparator = "\xe2\x94\x82";
line = "\xe2\x94\x80";
leftBox = "\xe2\x94\x9c";
middleBox = "\xe2\x94\xbc";
rightBox = "\xe2\x94\xa4";
}
sprintf_s(indentFormat, MAX_FORMAT_CHARS, "%%-%ds", regTableIndent);
// Now, set up the legend format for the RefPosition info
sprintf_s(legendFormat, MAX_LEGEND_FORMAT_CHARS, "%%-%d.%ds%%-%d.%ds%%-%ds%%s", nodeLocationWidth + 1,
nodeLocationWidth + 1, refPositionWidth + 2, refPositionWidth + 2, regColumnWidth + 1);
// Finally, print a "title row" including the legend and the reg names
dumpRegRecordTitle();
}
int LinearScan::getLastUsedRegNumIndex()
{
int lastUsedRegNumIndex = 0;
regMaskTP usedRegsMask = compiler->codeGen->regSet.rsGetModifiedRegsMask();
int lastRegNumIndex = compiler->compFloatingPointUsed ? REG_FP_LAST : REG_INT_LAST;
for (int regNumIndex = 0; regNumIndex <= lastRegNumIndex; regNumIndex++)
{
if ((usedRegsMask & genRegMask((regNumber)regNumIndex)) != 0)
{
lastUsedRegNumIndex = regNumIndex;
}
}
return lastUsedRegNumIndex;
}
void LinearScan::dumpRegRecordTitleLines()
{
for (int i = 0; i < regTableIndent; i++)
{
printf("%s", line);
}
int lastUsedRegNumIndex = getLastUsedRegNumIndex();
for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++)
{
printf("%s", middleBox);
for (int i = 0; i < regColumnWidth; i++)
{
printf("%s", line);
}
}
printf("%s\n", rightBox);
}
void LinearScan::dumpRegRecordTitle()
{
dumpRegRecordTitleLines();
// Print out the legend for the RefPosition info
printf(legendFormat, "Loc ", "RP# ", "Name ", "Type Action Reg ");
// Print out the register name column headers
char columnFormatArray[MAX_FORMAT_CHARS];
sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%s%%-%d.%ds", columnSeparator, regColumnWidth, regColumnWidth);
int lastUsedRegNumIndex = getLastUsedRegNumIndex();
for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++)
{
regNumber regNum = (regNumber)regNumIndex;
const char* regName = getRegName(regNum);
printf(columnFormatArray, regName);
}
printf("%s\n", columnSeparator);
rowCountSinceLastTitle = 0;
dumpRegRecordTitleLines();
}
void LinearScan::dumpRegRecords()
{
static char columnFormatArray[18];
int lastUsedRegNumIndex = getLastUsedRegNumIndex();
regMaskTP usedRegsMask = compiler->codeGen->regSet.rsGetModifiedRegsMask();
for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++)
{
printf("%s", columnSeparator);
RegRecord& regRecord = physRegs[regNumIndex];
Interval* interval = regRecord.assignedInterval;
if (interval != nullptr)
{
dumpIntervalName(interval);
char activeChar = interval->isActive ? 'a' : 'i';
printf("%c", activeChar);
}
else if (regRecord.isBusyUntilNextKill)
{
printf(columnFormatArray, "Busy");
}
else if ((usedRegsMask & genRegMask((regNumber)regNumIndex)) == 0)
{
sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%%-%ds", regColumnWidth);
printf(columnFormatArray, "----");
}
else
{
sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%%-%ds", regColumnWidth);
printf(columnFormatArray, "");
}
}
printf("%s\n", columnSeparator);
if (rowCountSinceLastTitle > MAX_ROWS_BETWEEN_TITLES)
{
dumpRegRecordTitle();
}
rowCountSinceLastTitle++;
}
void LinearScan::dumpIntervalName(Interval* interval)
{
if (interval->isLocalVar)
{
printf(intervalNameFormat, 'V', interval->varNum);
}
else if (interval->isConstant)
{
printf(intervalNameFormat, 'C', interval->intervalIndex);
}
else
{
printf(intervalNameFormat, 'I', interval->intervalIndex);
}
}
void LinearScan::dumpEmptyRefPosition()
{
printf(emptyRefPositionFormat, "");
}
// Note that the size of this dump is computed in dumpRegRecordHeader().
//
void LinearScan::dumpRefPositionShort(RefPosition* refPosition, BasicBlock* currentBlock)
{
BasicBlock* block = currentBlock;
if (refPosition->refType == RefTypeBB)
{
// Always print a title row before a RefTypeBB (except for the first, because we
// will already have printed it before the parameters)
if (refPosition->refType == RefTypeBB && block != compiler->fgFirstBB && block != nullptr)
{
dumpRegRecordTitle();
}
}
printf(shortRefPositionFormat, refPosition->nodeLocation, refPosition->rpNum);
if (refPosition->refType == RefTypeBB)
{
if (block == nullptr)
{
printf(regNameFormat, "END");
printf(" ");
printf(regNameFormat, "");
}
else
{
printf(bbRefPosFormat, block->bbNum, block == compiler->fgFirstBB ? 0 : blockInfo[block->bbNum].predBBNum);
}
}
else if (refPosition->isIntervalRef())
{
Interval* interval = refPosition->getInterval();
dumpIntervalName(interval);
char lastUseChar = ' ';
char delayChar = ' ';
if (refPosition->lastUse)
{
lastUseChar = '*';
if (refPosition->delayRegFree)
{
delayChar = 'D';
}
}
printf(" %s%c%c ", getRefTypeShortName(refPosition->refType), lastUseChar, delayChar);
}
else if (refPosition->isPhysRegRef)
{
RegRecord* regRecord = refPosition->getReg();
printf(regNameFormat, getRegName(regRecord->regNum));
printf(" %s ", getRefTypeShortName(refPosition->refType));
}
else
{
assert(refPosition->refType == RefTypeKillGCRefs);
// There's no interval or reg name associated with this.
printf(regNameFormat, " ");
printf(" %s ", getRefTypeShortName(refPosition->refType));
}
}
//------------------------------------------------------------------------
// LinearScan::IsResolutionMove:
// Returns true if the given node is a move inserted by LSRA
// resolution.
//
// Arguments:
// node - the node to check.
//
bool LinearScan::IsResolutionMove(GenTree* node)
{
if (!node->gtLsraInfo.isLsraAdded)
{
return false;
}
switch (node->OperGet())
{
case GT_LCL_VAR:
case GT_COPY:
return node->gtLsraInfo.isLocalDefUse;
case GT_SWAP:
return true;
default:
return false;
}
}
//------------------------------------------------------------------------
// LinearScan::IsResolutionNode:
// Returns true if the given node is either a move inserted by LSRA
// resolution or an operand to such a move.
//
// Arguments:
// containingRange - the range that contains the node to check.
// node - the node to check.
//
bool LinearScan::IsResolutionNode(LIR::Range& containingRange, GenTree* node)
{
for (;;)
{
if (IsResolutionMove(node))
{
return true;
}
if (!node->gtLsraInfo.isLsraAdded || (node->OperGet() != GT_LCL_VAR))
{
return false;
}
LIR::Use use;
bool foundUse = containingRange.TryGetUse(node, &use);
assert(foundUse);
node = use.User();
}
}
//------------------------------------------------------------------------
// verifyFinalAllocation: Traverse the RefPositions and verify various invariants.
//
// Arguments:
// None.
//
// Return Value:
// None.
//
// Notes:
// If verbose is set, this will also dump a table of the final allocations.
void LinearScan::verifyFinalAllocation()
{
if (VERBOSE)
{
printf("\nFinal allocation\n");
}
// Clear register assignments.
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
RegRecord* physRegRecord = getRegisterRecord(reg);
physRegRecord->assignedInterval = nullptr;
}
for (auto& interval : intervals)
{
interval.assignedReg = nullptr;
interval.physReg = REG_NA;
}
DBEXEC(VERBOSE, dumpRegRecordTitle());
BasicBlock* currentBlock = nullptr;
GenTree* firstBlockEndResolutionNode = nullptr;
regMaskTP regsToFree = RBM_NONE;
regMaskTP delayRegsToFree = RBM_NONE;
LsraLocation currentLocation = MinLocation;
for (auto& refPosition : refPositions)
{
RefPosition* currentRefPosition = &refPosition;
Interval* interval = nullptr;
RegRecord* regRecord = nullptr;
regNumber regNum = REG_NA;
if (currentRefPosition->refType == RefTypeBB)
{
regsToFree |= delayRegsToFree;
delayRegsToFree = RBM_NONE;
// For BB RefPositions, wait until we dump the "end of block" info before dumping the basic RefPosition
// info.
}
else
{
// For other RefPosition types, we can dump the basic RefPosition info now.
DBEXEC(VERBOSE, dumpRefPositionShort(currentRefPosition, currentBlock));
if (currentRefPosition->isPhysRegRef)
{
regRecord = currentRefPosition->getReg();
regRecord->recentRefPosition = currentRefPosition;
regNum = regRecord->regNum;
}
else if (currentRefPosition->isIntervalRef())
{
interval = currentRefPosition->getInterval();
interval->recentRefPosition = currentRefPosition;
if (currentRefPosition->registerAssignment != RBM_NONE)
{
if (!genMaxOneBit(currentRefPosition->registerAssignment))
{
assert(currentRefPosition->refType == RefTypeExpUse ||
currentRefPosition->refType == RefTypeDummyDef);
}
else
{
regNum = currentRefPosition->assignedReg();
regRecord = getRegisterRecord(regNum);
}
}
}
}
LsraLocation newLocation = currentRefPosition->nodeLocation;
if (newLocation > currentLocation)
{
// Free Registers.
// We could use the freeRegisters() method, but we'd have to carefully manage the active intervals.
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
regMaskTP regMask = genRegMask(reg);
if ((regsToFree & regMask) != RBM_NONE)
{
RegRecord* physRegRecord = getRegisterRecord(reg);
physRegRecord->assignedInterval = nullptr;
}
}
regsToFree = delayRegsToFree;
regsToFree = RBM_NONE;
}
currentLocation = newLocation;
switch (currentRefPosition->refType)
{
case RefTypeBB:
{
if (currentBlock == nullptr)
{
currentBlock = startBlockSequence();
}
else
{
// Verify the resolution moves at the end of the previous block.
for (GenTree* node = firstBlockEndResolutionNode; node != nullptr; node = node->gtNext)
{
assert(enregisterLocalVars);
// Only verify nodes that are actually moves; don't bother with the nodes that are
// operands to moves.
if (IsResolutionMove(node))
{
verifyResolutionMove(node, currentLocation);
}
}
// Validate the locations at the end of the previous block.
if (enregisterLocalVars)
{
VarToRegMap outVarToRegMap = outVarToRegMaps[currentBlock->bbNum];
VarSetOps::Iter iter(compiler, currentBlock->bbLiveOut);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
if (localVarIntervals[varIndex] == nullptr)
{
assert(!compiler->lvaTable[compiler->lvaTrackedToVarNum[varIndex]].lvLRACandidate);
continue;
}
regNumber regNum = getVarReg(outVarToRegMap, varIndex);
interval = getIntervalForLocalVar(varIndex);
assert(interval->physReg == regNum || (interval->physReg == REG_NA && regNum == REG_STK));
interval->physReg = REG_NA;
interval->assignedReg = nullptr;
interval->isActive = false;
}
}
// Clear register assignments.
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
RegRecord* physRegRecord = getRegisterRecord(reg);
physRegRecord->assignedInterval = nullptr;
}
// Now, record the locations at the beginning of this block.
currentBlock = moveToNextBlock();
}
if (currentBlock != nullptr)
{
if (enregisterLocalVars)
{
VarToRegMap inVarToRegMap = inVarToRegMaps[currentBlock->bbNum];
VarSetOps::Iter iter(compiler, currentBlock->bbLiveIn);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
if (localVarIntervals[varIndex] == nullptr)
{
assert(!compiler->lvaTable[compiler->lvaTrackedToVarNum[varIndex]].lvLRACandidate);
continue;
}
regNumber regNum = getVarReg(inVarToRegMap, varIndex);
interval = getIntervalForLocalVar(varIndex);
interval->physReg = regNum;
interval->assignedReg = &(physRegs[regNum]);
interval->isActive = true;
physRegs[regNum].assignedInterval = interval;
}
}
if (VERBOSE)
{
dumpRefPositionShort(currentRefPosition, currentBlock);
dumpRegRecords();
}
// Finally, handle the resolution moves, if any, at the beginning of the next block.
firstBlockEndResolutionNode = nullptr;
bool foundNonResolutionNode = false;
LIR::Range& currentBlockRange = LIR::AsRange(currentBlock);
for (GenTree* node : currentBlockRange.NonPhiNodes())
{
if (IsResolutionNode(currentBlockRange, node))
{
assert(enregisterLocalVars);
if (foundNonResolutionNode)
{
firstBlockEndResolutionNode = node;
break;
}
else if (IsResolutionMove(node))
{
// Only verify nodes that are actually moves; don't bother with the nodes that are
// operands to moves.
verifyResolutionMove(node, currentLocation);
}
}
else
{
foundNonResolutionNode = true;
}
}
}
}
break;
case RefTypeKill:
assert(regRecord != nullptr);
assert(regRecord->assignedInterval == nullptr);
dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock);
break;
case RefTypeFixedReg:
assert(regRecord != nullptr);
dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock);
break;
case RefTypeUpperVectorSaveDef:
case RefTypeUpperVectorSaveUse:
case RefTypeDef:
case RefTypeUse:
case RefTypeParamDef:
case RefTypeZeroInit:
assert(interval != nullptr);
if (interval->isSpecialPutArg)
{
dumpLsraAllocationEvent(LSRA_EVENT_SPECIAL_PUTARG, interval, regNum);
break;
}
if (currentRefPosition->reload)
{
interval->isActive = true;
assert(regNum != REG_NA);
interval->physReg = regNum;
interval->assignedReg = regRecord;
regRecord->assignedInterval = interval;
dumpLsraAllocationEvent(LSRA_EVENT_RELOAD, nullptr, regRecord->regNum, currentBlock);
}
if (regNum == REG_NA)
{
dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, interval);
}
else if (RefTypeIsDef(currentRefPosition->refType))
{
interval->isActive = true;
if (VERBOSE)
{
if (interval->isConstant && (currentRefPosition->treeNode != nullptr) &&
currentRefPosition->treeNode->IsReuseRegVal())
{
dumpLsraAllocationEvent(LSRA_EVENT_REUSE_REG, nullptr, regRecord->regNum, currentBlock);
}
else
{
dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, nullptr, regRecord->regNum, currentBlock);
}
}
}
else if (currentRefPosition->copyReg)
{
dumpLsraAllocationEvent(LSRA_EVENT_COPY_REG, interval, regRecord->regNum, currentBlock);
}
else if (currentRefPosition->moveReg)
{
assert(interval->assignedReg != nullptr);
interval->assignedReg->assignedInterval = nullptr;
interval->physReg = regNum;
interval->assignedReg = regRecord;
regRecord->assignedInterval = interval;
if (VERBOSE)
{
printf("Move %-4s ", getRegName(regRecord->regNum));
}
}
else
{
dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock);
}
if (currentRefPosition->lastUse || currentRefPosition->spillAfter)
{
interval->isActive = false;
}
if (regNum != REG_NA)
{
if (currentRefPosition->spillAfter)
{
if (VERBOSE)
{
// If refPos is marked as copyReg, then the reg that is spilled
// is the homeReg of the interval not the reg currently assigned
// to refPos.
regNumber spillReg = regNum;
if (currentRefPosition->copyReg)
{
assert(interval != nullptr);
spillReg = interval->physReg;
}
dumpRegRecords();
dumpEmptyRefPosition();
printf("Spill %-4s ", getRegName(spillReg));
}
}
else if (currentRefPosition->copyReg)
{
regRecord->assignedInterval = interval;
}
else
{
interval->physReg = regNum;
interval->assignedReg = regRecord;
regRecord->assignedInterval = interval;
}
}
break;
case RefTypeKillGCRefs:
// No action to take.
// However, we will assert that, at resolution time, no registers contain GC refs.
{
DBEXEC(VERBOSE, printf(" "));
regMaskTP candidateRegs = currentRefPosition->registerAssignment;
while (candidateRegs != RBM_NONE)
{
regMaskTP nextRegBit = genFindLowestBit(candidateRegs);
candidateRegs &= ~nextRegBit;
regNumber nextReg = genRegNumFromMask(nextRegBit);
RegRecord* regRecord = getRegisterRecord(nextReg);
Interval* assignedInterval = regRecord->assignedInterval;
assert(assignedInterval == nullptr || !varTypeIsGC(assignedInterval->registerType));
}
}
break;
case RefTypeExpUse:
case RefTypeDummyDef:
// Do nothing; these will be handled by the RefTypeBB.
DBEXEC(VERBOSE, printf(" "));
break;
case RefTypeInvalid:
// for these 'currentRefPosition->refType' values, No action to take
break;
}
if (currentRefPosition->refType != RefTypeBB)
{
DBEXEC(VERBOSE, dumpRegRecords());
if (interval != nullptr)
{
if (currentRefPosition->copyReg)
{
assert(interval->physReg != regNum);
regRecord->assignedInterval = nullptr;
assert(interval->assignedReg != nullptr);
regRecord = interval->assignedReg;
}
if (currentRefPosition->spillAfter || currentRefPosition->lastUse)
{
interval->physReg = REG_NA;
interval->assignedReg = nullptr;
// regRegcord could be null if the RefPosition does not require a register.
if (regRecord != nullptr)
{
regRecord->assignedInterval = nullptr;
}
else
{
assert(!currentRefPosition->RequiresRegister());
}
}
}
}
}
// Now, verify the resolution blocks.
// Currently these are nearly always at the end of the method, but that may not alwyas be the case.
// So, we'll go through all the BBs looking for blocks whose bbNum is greater than bbNumMaxBeforeResolution.
for (BasicBlock* currentBlock = compiler->fgFirstBB; currentBlock != nullptr; currentBlock = currentBlock->bbNext)
{
if (currentBlock->bbNum > bbNumMaxBeforeResolution)
{
// If we haven't enregistered an lclVars, we have no resolution blocks.
assert(enregisterLocalVars);
if (VERBOSE)
{
dumpRegRecordTitle();
printf(shortRefPositionFormat, 0, 0);
assert(currentBlock->bbPreds != nullptr && currentBlock->bbPreds->flBlock != nullptr);
printf(bbRefPosFormat, currentBlock->bbNum, currentBlock->bbPreds->flBlock->bbNum);
dumpRegRecords();
}
// Clear register assignments.
for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg))
{
RegRecord* physRegRecord = getRegisterRecord(reg);
physRegRecord->assignedInterval = nullptr;
}
// Set the incoming register assignments
VarToRegMap inVarToRegMap = getInVarToRegMap(currentBlock->bbNum);
VarSetOps::Iter iter(compiler, currentBlock->bbLiveIn);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
if (localVarIntervals[varIndex] == nullptr)
{
assert(!compiler->lvaTable[compiler->lvaTrackedToVarNum[varIndex]].lvLRACandidate);
continue;
}
regNumber regNum = getVarReg(inVarToRegMap, varIndex);
Interval* interval = getIntervalForLocalVar(varIndex);
interval->physReg = regNum;
interval->assignedReg = &(physRegs[regNum]);
interval->isActive = true;
physRegs[regNum].assignedInterval = interval;
}
// Verify the moves in this block
LIR::Range& currentBlockRange = LIR::AsRange(currentBlock);
for (GenTree* node : currentBlockRange.NonPhiNodes())
{
assert(IsResolutionNode(currentBlockRange, node));
if (IsResolutionMove(node))
{
// Only verify nodes that are actually moves; don't bother with the nodes that are
// operands to moves.
verifyResolutionMove(node, currentLocation);
}
}
// Verify the outgoing register assignments
{
VarToRegMap outVarToRegMap = getOutVarToRegMap(currentBlock->bbNum);
VarSetOps::Iter iter(compiler, currentBlock->bbLiveOut);
unsigned varIndex = 0;
while (iter.NextElem(&varIndex))
{
if (localVarIntervals[varIndex] == nullptr)
{
assert(!compiler->lvaTable[compiler->lvaTrackedToVarNum[varIndex]].lvLRACandidate);
continue;
}
regNumber regNum = getVarReg(outVarToRegMap, varIndex);
Interval* interval = getIntervalForLocalVar(varIndex);
assert(interval->physReg == regNum || (interval->physReg == REG_NA && regNum == REG_STK));
interval->physReg = REG_NA;
interval->assignedReg = nullptr;
interval->isActive = false;
}
}
}
}
DBEXEC(VERBOSE, printf("\n"));
}
//------------------------------------------------------------------------
// verifyResolutionMove: Verify a resolution statement. Called by verifyFinalAllocation()
//
// Arguments:
// resolutionMove - A GenTree* that must be a resolution move.
// currentLocation - The LsraLocation of the most recent RefPosition that has been verified.
//
// Return Value:
// None.
//
// Notes:
// If verbose is set, this will also dump the moves into the table of final allocations.
void LinearScan::verifyResolutionMove(GenTree* resolutionMove, LsraLocation currentLocation)
{
GenTree* dst = resolutionMove;
assert(IsResolutionMove(dst));
if (dst->OperGet() == GT_SWAP)
{
GenTreeLclVarCommon* left = dst->gtGetOp1()->AsLclVarCommon();
GenTreeLclVarCommon* right = dst->gtGetOp2()->AsLclVarCommon();
regNumber leftRegNum = left->gtRegNum;
regNumber rightRegNum = right->gtRegNum;
LclVarDsc* leftVarDsc = compiler->lvaTable + left->gtLclNum;
LclVarDsc* rightVarDsc = compiler->lvaTable + right->gtLclNum;
Interval* leftInterval = getIntervalForLocalVar(leftVarDsc->lvVarIndex);
Interval* rightInterval = getIntervalForLocalVar(rightVarDsc->lvVarIndex);
assert(leftInterval->physReg == leftRegNum && rightInterval->physReg == rightRegNum);
leftInterval->physReg = rightRegNum;
rightInterval->physReg = leftRegNum;
leftInterval->assignedReg = &physRegs[rightRegNum];
rightInterval->assignedReg = &physRegs[leftRegNum];
physRegs[rightRegNum].assignedInterval = leftInterval;
physRegs[leftRegNum].assignedInterval = rightInterval;
if (VERBOSE)
{
printf(shortRefPositionFormat, currentLocation, 0);
dumpIntervalName(leftInterval);
printf(" Swap ");
printf(" %-4s ", getRegName(rightRegNum));
dumpRegRecords();
printf(shortRefPositionFormat, currentLocation, 0);
dumpIntervalName(rightInterval);
printf(" \" ");
printf(" %-4s ", getRegName(leftRegNum));
dumpRegRecords();
}
return;
}
regNumber dstRegNum = dst->gtRegNum;
regNumber srcRegNum;
GenTreeLclVarCommon* lcl;
if (dst->OperGet() == GT_COPY)
{
lcl = dst->gtGetOp1()->AsLclVarCommon();
srcRegNum = lcl->gtRegNum;
}
else
{
lcl = dst->AsLclVarCommon();
if ((lcl->gtFlags & GTF_SPILLED) != 0)
{
srcRegNum = REG_STK;
}
else
{
assert((lcl->gtFlags & GTF_SPILL) != 0);
srcRegNum = dstRegNum;
dstRegNum = REG_STK;
}
}
Interval* interval = getIntervalForLocalVarNode(lcl);
assert(interval->physReg == srcRegNum || (srcRegNum == REG_STK && interval->physReg == REG_NA));
if (srcRegNum != REG_STK)
{
physRegs[srcRegNum].assignedInterval = nullptr;
}
if (dstRegNum != REG_STK)
{
interval->physReg = dstRegNum;
interval->assignedReg = &(physRegs[dstRegNum]);
physRegs[dstRegNum].assignedInterval = interval;
interval->isActive = true;
}
else
{
interval->physReg = REG_NA;
interval->assignedReg = nullptr;
interval->isActive = false;
}
if (VERBOSE)
{
printf(shortRefPositionFormat, currentLocation, 0);
dumpIntervalName(interval);
printf(" Move ");
printf(" %-4s ", getRegName(dstRegNum));
dumpRegRecords();
}
}
#endif // DEBUG
#endif // !LEGACY_BACKEND
|