aboutsummaryrefslogtreecommitdiffstats
path: root/src/mesa/drivers/dri/i965/brw_schedule_instructions.cpp
blob: 56f69ea508b2f17375c1064e69f04889ff7f1805 (plain)
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
/*
 * Copyright © 2010 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 * Authors:
 *    Eric Anholt <eric@anholt.net>
 *
 */

#include "brw_fs.h"
#include "brw_vec4.h"
#include "brw_cfg.h"
#include "brw_shader.h"
#include "glsl/glsl_types.h"
#include "glsl/ir_optimization.h"

using namespace brw;

/** @file brw_fs_schedule_instructions.cpp
 *
 * List scheduling of FS instructions.
 *
 * The basic model of the list scheduler is to take a basic block,
 * compute a DAG of the dependencies (RAW ordering with latency, WAW
 * ordering with latency, WAR ordering), and make a list of the DAG heads.
 * Heuristically pick a DAG head, then put all the children that are
 * now DAG heads into the list of things to schedule.
 *
 * The heuristic is the important part.  We're trying to be cheap,
 * since actually computing the optimal scheduling is NP complete.
 * What we do is track a "current clock".  When we schedule a node, we
 * update the earliest-unblocked clock time of its children, and
 * increment the clock.  Then, when trying to schedule, we just pick
 * the earliest-unblocked instruction to schedule.
 *
 * Note that often there will be many things which could execute
 * immediately, and there are a range of heuristic options to choose
 * from in picking among those.
 */

static bool debug = false;

class instruction_scheduler;

class schedule_node : public exec_node
{
public:
   schedule_node(backend_instruction *inst, instruction_scheduler *sched);
   void set_latency_gen4();
   void set_latency_gen7(bool is_haswell);

   backend_instruction *inst;
   schedule_node **children;
   int *child_latency;
   int child_count;
   int parent_count;
   int child_array_size;
   int unblocked_time;
   int latency;

   /**
    * Which iteration of pushing groups of children onto the candidates list
    * this node was a part of.
    */
   unsigned cand_generation;

   /**
    * This is the sum of the instruction's latency plus the maximum delay of
    * its children, or just the issue_time if it's a leaf node.
    */
   int delay;
};

void
schedule_node::set_latency_gen4()
{
   int chans = 8;
   int math_latency = 22;

   switch (inst->opcode) {
   case SHADER_OPCODE_RCP:
      this->latency = 1 * chans * math_latency;
      break;
   case SHADER_OPCODE_RSQ:
      this->latency = 2 * chans * math_latency;
      break;
   case SHADER_OPCODE_INT_QUOTIENT:
   case SHADER_OPCODE_SQRT:
   case SHADER_OPCODE_LOG2:
      /* full precision log.  partial is 2. */
      this->latency = 3 * chans * math_latency;
      break;
   case SHADER_OPCODE_INT_REMAINDER:
   case SHADER_OPCODE_EXP2:
      /* full precision.  partial is 3, same throughput. */
      this->latency = 4 * chans * math_latency;
      break;
   case SHADER_OPCODE_POW:
      this->latency = 8 * chans * math_latency;
      break;
   case SHADER_OPCODE_SIN:
   case SHADER_OPCODE_COS:
      /* minimum latency, max is 12 rounds. */
      this->latency = 5 * chans * math_latency;
      break;
   default:
      this->latency = 2;
      break;
   }
}

void
schedule_node::set_latency_gen7(bool is_haswell)
{
   switch (inst->opcode) {
   case BRW_OPCODE_MAD:
      /* 2 cycles
       *  (since the last two src operands are in different register banks):
       * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
       *
       * 3 cycles on IVB, 4 on HSW
       *  (since the last two src operands are in the same register bank):
       * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
       *
       * 18 cycles on IVB, 16 on HSW
       *  (since the last two src operands are in different register banks):
       * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
       * mov(8) null   g4<4,5,1>F                     { align16 WE_normal 1Q };
       *
       * 20 cycles on IVB, 18 on HSW
       *  (since the last two src operands are in the same register bank):
       * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
       * mov(8) null   g4<4,4,1>F                     { align16 WE_normal 1Q };
       */

      /* Our register allocator doesn't know about register banks, so use the
       * higher latency.
       */
      latency = is_haswell ? 16 : 18;
      break;

   case BRW_OPCODE_LRP:
      /* 2 cycles
       *  (since the last two src operands are in different register banks):
       * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
       *
       * 3 cycles on IVB, 4 on HSW
       *  (since the last two src operands are in the same register bank):
       * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
       *
       * 16 cycles on IVB, 14 on HSW
       *  (since the last two src operands are in different register banks):
       * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
       * mov(8) null   g4<4,4,1>F                     { align16 WE_normal 1Q };
       *
       * 16 cycles
       *  (since the last two src operands are in the same register bank):
       * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
       * mov(8) null   g4<4,4,1>F                     { align16 WE_normal 1Q };
       */

      /* Our register allocator doesn't know about register banks, so use the
       * higher latency.
       */
      latency = 14;
      break;

   case SHADER_OPCODE_RCP:
   case SHADER_OPCODE_RSQ:
   case SHADER_OPCODE_SQRT:
   case SHADER_OPCODE_LOG2:
   case SHADER_OPCODE_EXP2:
   case SHADER_OPCODE_SIN:
   case SHADER_OPCODE_COS:
      /* 2 cycles:
       * math inv(8) g4<1>F g2<0,1,0>F      null       { align1 WE_normal 1Q };
       *
       * 18 cycles:
       * math inv(8) g4<1>F g2<0,1,0>F      null       { align1 WE_normal 1Q };
       * mov(8)      null   g4<8,8,1>F                 { align1 WE_normal 1Q };
       *
       * Same for exp2, log2, rsq, sqrt, sin, cos.
       */
      latency = is_haswell ? 14 : 16;
      break;

   case SHADER_OPCODE_POW:
      /* 2 cycles:
       * math pow(8) g4<1>F g2<0,1,0>F   g2.1<0,1,0>F  { align1 WE_normal 1Q };
       *
       * 26 cycles:
       * math pow(8) g4<1>F g2<0,1,0>F   g2.1<0,1,0>F  { align1 WE_normal 1Q };
       * mov(8)      null   g4<8,8,1>F                 { align1 WE_normal 1Q };
       */
      latency = is_haswell ? 22 : 24;
      break;

   case SHADER_OPCODE_TEX:
   case SHADER_OPCODE_TXD:
   case SHADER_OPCODE_TXF:
   case SHADER_OPCODE_TXL:
      /* 18 cycles:
       * mov(8)  g115<1>F   0F                         { align1 WE_normal 1Q };
       * mov(8)  g114<1>F   0F                         { align1 WE_normal 1Q };
       * send(8) g4<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 0, 1) mlen 2 rlen 4         { align1 WE_normal 1Q };
       *
       * 697 +/-49 cycles (min 610, n=26):
       * mov(8)  g115<1>F   0F                         { align1 WE_normal 1Q };
       * mov(8)  g114<1>F   0F                         { align1 WE_normal 1Q };
       * send(8) g4<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 0, 1) mlen 2 rlen 4         { align1 WE_normal 1Q };
       * mov(8)  null       g4<8,8,1>F                 { align1 WE_normal 1Q };
       *
       * So the latency on our first texture load of the batchbuffer takes
       * ~700 cycles, since the caches are cold at that point.
       *
       * 840 +/- 92 cycles (min 720, n=25):
       * mov(8)  g115<1>F   0F                         { align1 WE_normal 1Q };
       * mov(8)  g114<1>F   0F                         { align1 WE_normal 1Q };
       * send(8) g4<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 0, 1) mlen 2 rlen 4         { align1 WE_normal 1Q };
       * mov(8)  null       g4<8,8,1>F                 { align1 WE_normal 1Q };
       * send(8) g4<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 0, 1) mlen 2 rlen 4         { align1 WE_normal 1Q };
       * mov(8)  null       g4<8,8,1>F                 { align1 WE_normal 1Q };
       *
       * On the second load, it takes just an extra ~140 cycles, and after
       * accounting for the 14 cycles of the MOV's latency, that makes ~130.
       *
       * 683 +/- 49 cycles (min = 602, n=47):
       * mov(8)  g115<1>F   0F                         { align1 WE_normal 1Q };
       * mov(8)  g114<1>F   0F                         { align1 WE_normal 1Q };
       * send(8) g4<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 0, 1) mlen 2 rlen 4         { align1 WE_normal 1Q };
       * send(8) g50<1>UW   g114<8,8,1>F
       *   sampler (10, 0, 0, 1) mlen 2 rlen 4         { align1 WE_normal 1Q };
       * mov(8)  null       g4<8,8,1>F                 { align1 WE_normal 1Q };
       *
       * The unit appears to be pipelined, since this matches up with the
       * cache-cold case, despite there being two loads here.  If you replace
       * the g4 in the MOV to null with g50, it's still 693 +/- 52 (n=39).
       *
       * So, take some number between the cache-hot 140 cycles and the
       * cache-cold 700 cycles.  No particular tuning was done on this.
       *
       * I haven't done significant testing of the non-TEX opcodes.  TXL at
       * least looked about the same as TEX.
       */
      latency = 200;
      break;

   case SHADER_OPCODE_TXS:
      /* Testing textureSize(sampler2D, 0), one load was 420 +/- 41
       * cycles (n=15):
       * mov(8)   g114<1>UD  0D                        { align1 WE_normal 1Q };
       * send(8)  g6<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 10, 1) mlen 1 rlen 4        { align1 WE_normal 1Q };
       * mov(16)  g6<1>F     g6<8,8,1>D                { align1 WE_normal 1Q };
       *
       *
       * Two loads was 535 +/- 30 cycles (n=19):
       * mov(16)   g114<1>UD  0D                       { align1 WE_normal 1H };
       * send(16)  g6<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 10, 2) mlen 2 rlen 8        { align1 WE_normal 1H };
       * mov(16)   g114<1>UD  0D                       { align1 WE_normal 1H };
       * mov(16)   g6<1>F     g6<8,8,1>D               { align1 WE_normal 1H };
       * send(16)  g8<1>UW    g114<8,8,1>F
       *   sampler (10, 0, 10, 2) mlen 2 rlen 8        { align1 WE_normal 1H };
       * mov(16)   g8<1>F     g8<8,8,1>D               { align1 WE_normal 1H };
       * add(16)   g6<1>F     g6<8,8,1>F   g8<8,8,1>F  { align1 WE_normal 1H };
       *
       * Since the only caches that should matter are just the
       * instruction/state cache containing the surface state, assume that we
       * always have hot caches.
       */
      latency = 100;
      break;

   case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
   case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
   case VS_OPCODE_PULL_CONSTANT_LOAD:
      /* testing using varying-index pull constants:
       *
       * 16 cycles:
       * mov(8)  g4<1>D  g2.1<0,1,0>F                  { align1 WE_normal 1Q };
       * send(8) g4<1>F  g4<8,8,1>D
       *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
       *
       * ~480 cycles:
       * mov(8)  g4<1>D  g2.1<0,1,0>F                  { align1 WE_normal 1Q };
       * send(8) g4<1>F  g4<8,8,1>D
       *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
       * mov(8)  null    g4<8,8,1>F                    { align1 WE_normal 1Q };
       *
       * ~620 cycles:
       * mov(8)  g4<1>D  g2.1<0,1,0>F                  { align1 WE_normal 1Q };
       * send(8) g4<1>F  g4<8,8,1>D
       *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
       * mov(8)  null    g4<8,8,1>F                    { align1 WE_normal 1Q };
       * send(8) g4<1>F  g4<8,8,1>D
       *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
       * mov(8)  null    g4<8,8,1>F                    { align1 WE_normal 1Q };
       *
       * So, if it's cache-hot, it's about 140.  If it's cache cold, it's
       * about 460.  We expect to mostly be cache hot, so pick something more
       * in that direction.
       */
      latency = 200;
      break;

   case SHADER_OPCODE_GEN7_SCRATCH_READ:
      /* Testing a load from offset 0, that had been previously written:
       *
       * send(8) g114<1>UW g0<8,8,1>F data (0, 0, 0) mlen 1 rlen 1 { align1 WE_normal 1Q };
       * mov(8)  null      g114<8,8,1>F { align1 WE_normal 1Q };
       *
       * The cycles spent seemed to be grouped around 40-50 (as low as 38),
       * then around 140.  Presumably this is cache hit vs miss.
       */
      latency = 50;
      break;

   case SHADER_OPCODE_UNTYPED_ATOMIC:
      /* Test code:
       *   mov(8)    g112<1>ud       0x00000000ud       { align1 WE_all 1Q };
       *   mov(1)    g112.7<1>ud     g1.7<0,1,0>ud      { align1 WE_all };
       *   mov(8)    g113<1>ud       0x00000000ud       { align1 WE_normal 1Q };
       *   send(8)   g4<1>ud         g112<8,8,1>ud
       *             data (38, 5, 6) mlen 2 rlen 1      { align1 WE_normal 1Q };
       *
       * Running it 100 times as fragment shader on a 128x128 quad
       * gives an average latency of 13867 cycles per atomic op,
       * standard deviation 3%.  Note that this is a rather
       * pessimistic estimate, the actual latency in cases with few
       * collisions between threads and favorable pipelining has been
       * seen to be reduced by a factor of 100.
       */
      latency = 14000;
      break;

   case SHADER_OPCODE_UNTYPED_SURFACE_READ:
      /* Test code:
       *   mov(8)    g112<1>UD       0x00000000UD       { align1 WE_all 1Q };
       *   mov(1)    g112.7<1>UD     g1.7<0,1,0>UD      { align1 WE_all };
       *   mov(8)    g113<1>UD       0x00000000UD       { align1 WE_normal 1Q };
       *   send(8)   g4<1>UD         g112<8,8,1>UD
       *             data (38, 6, 5) mlen 2 rlen 1      { align1 WE_normal 1Q };
       *   .
       *   . [repeats 8 times]
       *   .
       *   mov(8)    g112<1>UD       0x00000000UD       { align1 WE_all 1Q };
       *   mov(1)    g112.7<1>UD     g1.7<0,1,0>UD      { align1 WE_all };
       *   mov(8)    g113<1>UD       0x00000000UD       { align1 WE_normal 1Q };
       *   send(8)   g4<1>UD         g112<8,8,1>UD
       *             data (38, 6, 5) mlen 2 rlen 1      { align1 WE_normal 1Q };
       *
       * Running it 100 times as fragment shader on a 128x128 quad
       * gives an average latency of 583 cycles per surface read,
       * standard deviation 0.9%.
       */
      latency = is_haswell ? 300 : 600;
      break;

   default:
      /* 2 cycles:
       * mul(8) g4<1>F g2<0,1,0>F      0.5F            { align1 WE_normal 1Q };
       *
       * 16 cycles:
       * mul(8) g4<1>F g2<0,1,0>F      0.5F            { align1 WE_normal 1Q };
       * mov(8) null   g4<8,8,1>F                      { align1 WE_normal 1Q };
       */
      latency = 14;
      break;
   }
}

class instruction_scheduler {
public:
   instruction_scheduler(backend_visitor *v, int grf_count,
                         instruction_scheduler_mode mode)
   {
      this->bv = v;
      this->mem_ctx = ralloc_context(NULL);
      this->grf_count = grf_count;
      this->instructions.make_empty();
      this->instructions_to_schedule = 0;
      this->post_reg_alloc = (mode == SCHEDULE_POST);
      this->mode = mode;
      this->time = 0;
      if (!post_reg_alloc) {
         this->remaining_grf_uses = rzalloc_array(mem_ctx, int, grf_count);
         this->grf_active = rzalloc_array(mem_ctx, bool, grf_count);
      } else {
         this->remaining_grf_uses = NULL;
         this->grf_active = NULL;
      }
   }

   ~instruction_scheduler()
   {
      ralloc_free(this->mem_ctx);
   }
   void add_barrier_deps(schedule_node *n);
   void add_dep(schedule_node *before, schedule_node *after, int latency);
   void add_dep(schedule_node *before, schedule_node *after);

   void run(cfg_t *cfg);
   void add_insts_from_block(bblock_t *block);
   void compute_delay(schedule_node *node);
   virtual void calculate_deps() = 0;
   virtual schedule_node *choose_instruction_to_schedule() = 0;

   /**
    * Returns how many cycles it takes the instruction to issue.
    *
    * Instructions in gen hardware are handled one simd4 vector at a time,
    * with 1 cycle per vector dispatched.  Thus SIMD8 pixel shaders take 2
    * cycles to dispatch and SIMD16 (compressed) instructions take 4.
    */
   virtual int issue_time(backend_instruction *inst) = 0;

   virtual void count_remaining_grf_uses(backend_instruction *inst) = 0;
   virtual void update_register_pressure(backend_instruction *inst) = 0;
   virtual int get_register_pressure_benefit(backend_instruction *inst) = 0;

   void schedule_instructions(bblock_t *block);

   void *mem_ctx;

   bool post_reg_alloc;
   int instructions_to_schedule;
   int grf_count;
   int time;
   exec_list instructions;
   backend_visitor *bv;

   instruction_scheduler_mode mode;

   /**
    * Number of instructions left to schedule that reference each vgrf.
    *
    * Used so that we can prefer scheduling instructions that will end the
    * live intervals of multiple variables, to reduce register pressure.
    */
   int *remaining_grf_uses;

   /**
    * Tracks whether each VGRF has had an instruction scheduled that uses it.
    *
    * This is used to estimate whether scheduling a new instruction will
    * increase register pressure.
    */
   bool *grf_active;
};

class fs_instruction_scheduler : public instruction_scheduler
{
public:
   fs_instruction_scheduler(fs_visitor *v, int grf_count,
                            instruction_scheduler_mode mode);
   void calculate_deps();
   bool is_compressed(fs_inst *inst);
   schedule_node *choose_instruction_to_schedule();
   int issue_time(backend_instruction *inst);
   fs_visitor *v;

   void count_remaining_grf_uses(backend_instruction *inst);
   void update_register_pressure(backend_instruction *inst);
   int get_register_pressure_benefit(backend_instruction *inst);
};

fs_instruction_scheduler::fs_instruction_scheduler(fs_visitor *v,
                                                   int grf_count,
                                                   instruction_scheduler_mode mode)
   : instruction_scheduler(v, grf_count, mode),
     v(v)
{
}

void
fs_instruction_scheduler::count_remaining_grf_uses(backend_instruction *be)
{
   fs_inst *inst = (fs_inst *)be;

   if (!remaining_grf_uses)
      return;

   if (inst->dst.file == GRF)
      remaining_grf_uses[inst->dst.reg]++;

   for (int i = 0; i < inst->sources; i++) {
      if (inst->src[i].file != GRF)
         continue;

      remaining_grf_uses[inst->src[i].reg]++;
   }
}

void
fs_instruction_scheduler::update_register_pressure(backend_instruction *be)
{
   fs_inst *inst = (fs_inst *)be;

   if (!remaining_grf_uses)
      return;

   if (inst->dst.file == GRF) {
      remaining_grf_uses[inst->dst.reg]--;
      grf_active[inst->dst.reg] = true;
   }

   for (int i = 0; i < inst->sources; i++) {
      if (inst->src[i].file == GRF) {
         remaining_grf_uses[inst->src[i].reg]--;
         grf_active[inst->src[i].reg] = true;
      }
   }
}

int
fs_instruction_scheduler::get_register_pressure_benefit(backend_instruction *be)
{
   fs_inst *inst = (fs_inst *)be;
   int benefit = 0;

   if (inst->dst.file == GRF) {
      if (remaining_grf_uses[inst->dst.reg] == 1)
         benefit += v->alloc.sizes[inst->dst.reg];
      if (!grf_active[inst->dst.reg])
         benefit -= v->alloc.sizes[inst->dst.reg];
   }

   for (int i = 0; i < inst->sources; i++) {
      if (inst->src[i].file != GRF)
         continue;

      if (remaining_grf_uses[inst->src[i].reg] == 1)
         benefit += v->alloc.sizes[inst->src[i].reg];
      if (!grf_active[inst->src[i].reg])
         benefit -= v->alloc.sizes[inst->src[i].reg];
   }

   return benefit;
}

class vec4_instruction_scheduler : public instruction_scheduler
{
public:
   vec4_instruction_scheduler(vec4_visitor *v, int grf_count);
   void calculate_deps();
   schedule_node *choose_instruction_to_schedule();
   int issue_time(backend_instruction *inst);
   vec4_visitor *v;

   void count_remaining_grf_uses(backend_instruction *inst);
   void update_register_pressure(backend_instruction *inst);
   int get_register_pressure_benefit(backend_instruction *inst);
};

vec4_instruction_scheduler::vec4_instruction_scheduler(vec4_visitor *v,
                                                       int grf_count)
   : instruction_scheduler(v, grf_count, SCHEDULE_POST),
     v(v)
{
}

void
vec4_instruction_scheduler::count_remaining_grf_uses(backend_instruction *be)
{
}

void
vec4_instruction_scheduler::update_register_pressure(backend_instruction *be)
{
}

int
vec4_instruction_scheduler::get_register_pressure_benefit(backend_instruction *be)
{
   return 0;
}

schedule_node::schedule_node(backend_instruction *inst,
                             instruction_scheduler *sched)
{
   struct brw_context *brw = sched->bv->brw;

   this->inst = inst;
   this->child_array_size = 0;
   this->children = NULL;
   this->child_latency = NULL;
   this->child_count = 0;
   this->parent_count = 0;
   this->unblocked_time = 0;
   this->cand_generation = 0;
   this->delay = 0;

   /* We can't measure Gen6 timings directly but expect them to be much
    * closer to Gen7 than Gen4.
    */
   if (!sched->post_reg_alloc)
      this->latency = 1;
   else if (brw->gen >= 6)
      set_latency_gen7(brw->is_haswell);
   else
      set_latency_gen4();
}

void
instruction_scheduler::add_insts_from_block(bblock_t *block)
{
   /* Removing the last instruction from a basic block removes the block as
    * well, so put a NOP at the end to keep it alive.
    */
   if (!block->end()->is_control_flow()) {
      backend_instruction *nop = new(mem_ctx) backend_instruction();
      nop->opcode = BRW_OPCODE_NOP;
      block->end()->insert_after(block, nop);
   }

   foreach_inst_in_block_safe(backend_instruction, inst, block) {
      if (inst->opcode == BRW_OPCODE_NOP || inst->is_control_flow())
         continue;

      schedule_node *n = new(mem_ctx) schedule_node(inst, this);

      this->instructions_to_schedule++;

      inst->remove(block);
      instructions.push_tail(n);
   }
}

/** Recursive computation of the delay member of a node. */
void
instruction_scheduler::compute_delay(schedule_node *n)
{
   if (!n->child_count) {
      n->delay = issue_time(n->inst);
   } else {
      for (int i = 0; i < n->child_count; i++) {
         if (!n->children[i]->delay)
            compute_delay(n->children[i]);
         n->delay = MAX2(n->delay, n->latency + n->children[i]->delay);
      }
   }
}

/**
 * Add a dependency between two instruction nodes.
 *
 * The @after node will be scheduled after @before.  We will try to
 * schedule it @latency cycles after @before, but no guarantees there.
 */
void
instruction_scheduler::add_dep(schedule_node *before, schedule_node *after,
                               int latency)
{
   if (!before || !after)
      return;

   assert(before != after);

   for (int i = 0; i < before->child_count; i++) {
      if (before->children[i] == after) {
         before->child_latency[i] = MAX2(before->child_latency[i], latency);
         return;
      }
   }

   if (before->child_array_size <= before->child_count) {
      if (before->child_array_size < 16)
         before->child_array_size = 16;
      else
         before->child_array_size *= 2;

      before->children = reralloc(mem_ctx, before->children,
                                  schedule_node *,
                                  before->child_array_size);
      before->child_latency = reralloc(mem_ctx, before->child_latency,
                                       int, before->child_array_size);
   }

   before->children[before->child_count] = after;
   before->child_latency[before->child_count] = latency;
   before->child_count++;
   after->parent_count++;
}

void
instruction_scheduler::add_dep(schedule_node *before, schedule_node *after)
{
   if (!before)
      return;

   add_dep(before, after, before->latency);
}

/**
 * Sometimes we really want this node to execute after everything that
 * was before it and before everything that followed it.  This adds
 * the deps to do so.
 */
void
instruction_scheduler::add_barrier_deps(schedule_node *n)
{
   schedule_node *prev = (schedule_node *)n->prev;
   schedule_node *next = (schedule_node *)n->next;

   if (prev) {
      while (!prev->is_head_sentinel()) {
         add_dep(prev, n, 0);
         prev = (schedule_node *)prev->prev;
      }
   }

   if (next) {
      while (!next->is_tail_sentinel()) {
         add_dep(n, next, 0);
         next = (schedule_node *)next->next;
      }
   }
}

/* instruction scheduling needs to be aware of when an MRF write
 * actually writes 2 MRFs.
 */
bool
fs_instruction_scheduler::is_compressed(fs_inst *inst)
{
   return inst->exec_size == 16;
}

void
fs_instruction_scheduler::calculate_deps()
{
   /* Pre-register-allocation, this tracks the last write per VGRF offset.
    * After register allocation, reg_offsets are gone and we track individual
    * GRF registers.
    */
   schedule_node *last_grf_write[grf_count * 16];
   schedule_node *last_mrf_write[BRW_MAX_MRF];
   schedule_node *last_conditional_mod[2] = { NULL, NULL };
   schedule_node *last_accumulator_write = NULL;
   /* Fixed HW registers are assumed to be separate from the virtual
    * GRFs, so they can be tracked separately.  We don't really write
    * to fixed GRFs much, so don't bother tracking them on a more
    * granular level.
    */
   schedule_node *last_fixed_grf_write = NULL;
   int reg_width = v->dispatch_width / 8;

   /* The last instruction always needs to still be the last
    * instruction.  Either it's flow control (IF, ELSE, ENDIF, DO,
    * WHILE) and scheduling other things after it would disturb the
    * basic block, or it's FB_WRITE and we should do a better job at
    * dead code elimination anyway.
    */
   schedule_node *last = (schedule_node *)instructions.get_tail();
   add_barrier_deps(last);

   memset(last_grf_write, 0, sizeof(last_grf_write));
   memset(last_mrf_write, 0, sizeof(last_mrf_write));

   /* top-to-bottom dependencies: RAW and WAW. */
   foreach_in_list(schedule_node, n, &instructions) {
      fs_inst *inst = (fs_inst *)n->inst;

      if (inst->opcode == FS_OPCODE_PLACEHOLDER_HALT ||
         inst->has_side_effects())
         add_barrier_deps(n);

      /* read-after-write deps. */
      for (int i = 0; i < inst->sources; i++) {
         if (inst->src[i].file == GRF) {
            if (post_reg_alloc) {
               for (int r = 0; r < inst->regs_read(i); r++)
                  add_dep(last_grf_write[inst->src[i].reg + r], n);
            } else {
               for (int r = 0; r < inst->regs_read(i); r++) {
                  add_dep(last_grf_write[inst->src[i].reg * 16 + inst->src[i].reg_offset + r], n);
               }
            }
         } else if (inst->src[i].file == HW_REG &&
                    (inst->src[i].fixed_hw_reg.file ==
                     BRW_GENERAL_REGISTER_FILE)) {
            if (post_reg_alloc) {
               int size = reg_width;
               if (inst->src[i].fixed_hw_reg.vstride == BRW_VERTICAL_STRIDE_0)
                  size = 1;
               for (int r = 0; r < size; r++)
                  add_dep(last_grf_write[inst->src[i].fixed_hw_reg.nr + r], n);
            } else {
               add_dep(last_fixed_grf_write, n);
            }
         } else if (inst->src[i].is_accumulator()) {
            add_dep(last_accumulator_write, n);
         } else if (inst->src[i].file != BAD_FILE &&
                    inst->src[i].file != IMM &&
                    inst->src[i].file != UNIFORM &&
                    (inst->src[i].file != HW_REG ||
                     inst->src[i].fixed_hw_reg.file != IMM)) {
            assert(inst->src[i].file != MRF);
            add_barrier_deps(n);
         }
      }

      if (inst->base_mrf != -1) {
         for (int i = 0; i < inst->mlen; i++) {
            /* It looks like the MRF regs are released in the send
             * instruction once it's sent, not when the result comes
             * back.
             */
            add_dep(last_mrf_write[inst->base_mrf + i], n);
         }
      }

      if (inst->reads_flag()) {
         add_dep(last_conditional_mod[inst->flag_subreg], n);
      }

      if (inst->reads_accumulator_implicitly()) {
         add_dep(last_accumulator_write, n);
      }

      /* write-after-write deps. */
      if (inst->dst.file == GRF) {
         if (post_reg_alloc) {
            for (int r = 0; r < inst->regs_written; r++) {
               add_dep(last_grf_write[inst->dst.reg + r], n);
               last_grf_write[inst->dst.reg + r] = n;
            }
         } else {
            for (int r = 0; r < inst->regs_written; r++) {
               add_dep(last_grf_write[inst->dst.reg * 16 + inst->dst.reg_offset + r], n);
               last_grf_write[inst->dst.reg * 16 + inst->dst.reg_offset + r] = n;
            }
         }
      } else if (inst->dst.file == MRF) {
         int reg = inst->dst.reg & ~BRW_MRF_COMPR4;

         add_dep(last_mrf_write[reg], n);
         last_mrf_write[reg] = n;
         if (is_compressed(inst)) {
            if (inst->dst.reg & BRW_MRF_COMPR4)
               reg += 4;
            else
               reg++;
            add_dep(last_mrf_write[reg], n);
            last_mrf_write[reg] = n;
         }
      } else if (inst->dst.file == HW_REG &&
                 inst->dst.fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
         if (post_reg_alloc) {
            for (int r = 0; r < reg_width; r++)
               last_grf_write[inst->dst.fixed_hw_reg.nr + r] = n;
         } else {
            last_fixed_grf_write = n;
         }
      } else if (inst->dst.is_accumulator()) {
         add_dep(last_accumulator_write, n);
         last_accumulator_write = n;
      } else if (inst->dst.file != BAD_FILE &&
                 !inst->dst.is_null()) {
         add_barrier_deps(n);
      }

      if (inst->mlen > 0 && inst->base_mrf != -1) {
         for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
            add_dep(last_mrf_write[inst->base_mrf + i], n);
            last_mrf_write[inst->base_mrf + i] = n;
         }
      }

      if (inst->writes_flag()) {
         add_dep(last_conditional_mod[inst->flag_subreg], n, 0);
         last_conditional_mod[inst->flag_subreg] = n;
      }

      if (inst->writes_accumulator_implicitly(v->brw) &&
          !inst->dst.is_accumulator()) {
         add_dep(last_accumulator_write, n);
         last_accumulator_write = n;
      }
   }

   /* bottom-to-top dependencies: WAR */
   memset(last_grf_write, 0, sizeof(last_grf_write));
   memset(last_mrf_write, 0, sizeof(last_mrf_write));
   memset(last_conditional_mod, 0, sizeof(last_conditional_mod));
   last_accumulator_write = NULL;
   last_fixed_grf_write = NULL;

   exec_node *node;
   exec_node *prev;
   for (node = instructions.get_tail(), prev = node->prev;
        !node->is_head_sentinel();
        node = prev, prev = node->prev) {
      schedule_node *n = (schedule_node *)node;
      fs_inst *inst = (fs_inst *)n->inst;

      /* write-after-read deps. */
      for (int i = 0; i < inst->sources; i++) {
         if (inst->src[i].file == GRF) {
            if (post_reg_alloc) {
               for (int r = 0; r < inst->regs_read(i); r++)
                  add_dep(n, last_grf_write[inst->src[i].reg + r]);
            } else {
               for (int r = 0; r < inst->regs_read(i); r++) {
                  add_dep(n, last_grf_write[inst->src[i].reg * 16 + inst->src[i].reg_offset + r]);
               }
            }
         } else if (inst->src[i].file == HW_REG &&
                    (inst->src[i].fixed_hw_reg.file ==
                     BRW_GENERAL_REGISTER_FILE)) {
            if (post_reg_alloc) {
               int size = reg_width;
               if (inst->src[i].fixed_hw_reg.vstride == BRW_VERTICAL_STRIDE_0)
                  size = 1;
               for (int r = 0; r < size; r++)
                  add_dep(n, last_grf_write[inst->src[i].fixed_hw_reg.nr + r]);
            } else {
               add_dep(n, last_fixed_grf_write);
            }
         } else if (inst->src[i].is_accumulator()) {
            add_dep(n, last_accumulator_write);
         } else if (inst->src[i].file != BAD_FILE &&
                    inst->src[i].file != IMM &&
                    inst->src[i].file != UNIFORM &&
                    (inst->src[i].file != HW_REG ||
                     inst->src[i].fixed_hw_reg.file != IMM)) {
            assert(inst->src[i].file != MRF);
            add_barrier_deps(n);
         }
      }

      if (inst->base_mrf != -1) {
         for (int i = 0; i < inst->mlen; i++) {
            /* It looks like the MRF regs are released in the send
             * instruction once it's sent, not when the result comes
             * back.
             */
            add_dep(n, last_mrf_write[inst->base_mrf + i], 2);
         }
      }

      if (inst->reads_flag()) {
         add_dep(n, last_conditional_mod[inst->flag_subreg]);
      }

      if (inst->reads_accumulator_implicitly()) {
         add_dep(n, last_accumulator_write);
      }

      /* Update the things this instruction wrote, so earlier reads
       * can mark this as WAR dependency.
       */
      if (inst->dst.file == GRF) {
         if (post_reg_alloc) {
            for (int r = 0; r < inst->regs_written; r++)
               last_grf_write[inst->dst.reg + r] = n;
         } else {
            for (int r = 0; r < inst->regs_written; r++) {
               last_grf_write[inst->dst.reg * 16 + inst->dst.reg_offset + r] = n;
            }
         }
      } else if (inst->dst.file == MRF) {
         int reg = inst->dst.reg & ~BRW_MRF_COMPR4;

         last_mrf_write[reg] = n;

         if (is_compressed(inst)) {
            if (inst->dst.reg & BRW_MRF_COMPR4)
               reg += 4;
            else
               reg++;

            last_mrf_write[reg] = n;
         }
      } else if (inst->dst.file == HW_REG &&
                 inst->dst.fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
         if (post_reg_alloc) {
            for (int r = 0; r < reg_width; r++)
               last_grf_write[inst->dst.fixed_hw_reg.nr + r] = n;
         } else {
            last_fixed_grf_write = n;
         }
      } else if (inst->dst.is_accumulator()) {
         last_accumulator_write = n;
      } else if (inst->dst.file != BAD_FILE &&
                 !inst->dst.is_null()) {
         add_barrier_deps(n);
      }

      if (inst->mlen > 0 && inst->base_mrf != -1) {
         for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
            last_mrf_write[inst->base_mrf + i] = n;
         }
      }

      if (inst->writes_flag()) {
         last_conditional_mod[inst->flag_subreg] = n;
      }

      if (inst->writes_accumulator_implicitly(v->brw)) {
         last_accumulator_write = n;
      }
   }
}

void
vec4_instruction_scheduler::calculate_deps()
{
   schedule_node *last_grf_write[grf_count];
   schedule_node *last_mrf_write[BRW_MAX_MRF];
   schedule_node *last_conditional_mod = NULL;
   schedule_node *last_accumulator_write = NULL;
   /* Fixed HW registers are assumed to be separate from the virtual
    * GRFs, so they can be tracked separately.  We don't really write
    * to fixed GRFs much, so don't bother tracking them on a more
    * granular level.
    */
   schedule_node *last_fixed_grf_write = NULL;

   /* The last instruction always needs to still be the last instruction.
    * Either it's flow control (IF, ELSE, ENDIF, DO, WHILE) and scheduling
    * other things after it would disturb the basic block, or it's the EOT
    * URB_WRITE and we should do a better job at dead code eliminating
    * anything that could have been scheduled after it.
    */
   schedule_node *last = (schedule_node *)instructions.get_tail();
   add_barrier_deps(last);

   memset(last_grf_write, 0, sizeof(last_grf_write));
   memset(last_mrf_write, 0, sizeof(last_mrf_write));

   /* top-to-bottom dependencies: RAW and WAW. */
   foreach_in_list(schedule_node, n, &instructions) {
      vec4_instruction *inst = (vec4_instruction *)n->inst;

      if (inst->has_side_effects())
         add_barrier_deps(n);

      /* read-after-write deps. */
      for (int i = 0; i < 3; i++) {
         if (inst->src[i].file == GRF) {
            for (unsigned j = 0; j < inst->regs_read(i); ++j)
               add_dep(last_grf_write[inst->src[i].reg + j], n);
         } else if (inst->src[i].file == HW_REG &&
                    (inst->src[i].fixed_hw_reg.file ==
                     BRW_GENERAL_REGISTER_FILE)) {
            add_dep(last_fixed_grf_write, n);
         } else if (inst->src[i].is_accumulator()) {
            assert(last_accumulator_write);
            add_dep(last_accumulator_write, n);
         } else if (inst->src[i].file != BAD_FILE &&
                    inst->src[i].file != IMM &&
                    inst->src[i].file != UNIFORM &&
                    (inst->src[i].file != HW_REG ||
                     inst->src[i].fixed_hw_reg.file != IMM)) {
            /* No reads from MRF, and ATTR is already translated away */
            assert(inst->src[i].file != MRF &&
                   inst->src[i].file != ATTR);
            add_barrier_deps(n);
         }
      }

      if (!inst->is_send_from_grf()) {
         for (int i = 0; i < inst->mlen; i++) {
            /* It looks like the MRF regs are released in the send
             * instruction once it's sent, not when the result comes
             * back.
             */
            add_dep(last_mrf_write[inst->base_mrf + i], n);
         }
      }

      if (inst->reads_flag()) {
         assert(last_conditional_mod);
         add_dep(last_conditional_mod, n);
      }

      if (inst->reads_accumulator_implicitly()) {
         assert(last_accumulator_write);
         add_dep(last_accumulator_write, n);
      }

      /* write-after-write deps. */
      if (inst->dst.file == GRF) {
         for (unsigned j = 0; j < inst->regs_written; ++j) {
            add_dep(last_grf_write[inst->dst.reg + j], n);
            last_grf_write[inst->dst.reg + j] = n;
         }
      } else if (inst->dst.file == MRF) {
         add_dep(last_mrf_write[inst->dst.reg], n);
         last_mrf_write[inst->dst.reg] = n;
     } else if (inst->dst.file == HW_REG &&
                 inst->dst.fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
         last_fixed_grf_write = n;
      } else if (inst->dst.is_accumulator()) {
         add_dep(last_accumulator_write, n);
         last_accumulator_write = n;
      } else if (inst->dst.file != BAD_FILE &&
                 !inst->dst.is_null()) {
         add_barrier_deps(n);
      }

      if (inst->mlen > 0 && !inst->is_send_from_grf()) {
         for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
            add_dep(last_mrf_write[inst->base_mrf + i], n);
            last_mrf_write[inst->base_mrf + i] = n;
         }
      }

      if (inst->writes_flag()) {
         add_dep(last_conditional_mod, n, 0);
         last_conditional_mod = n;
      }

      if (inst->writes_accumulator_implicitly(v->brw) &&
          !inst->dst.is_accumulator()) {
         add_dep(last_accumulator_write, n);
         last_accumulator_write = n;
      }
   }

   /* bottom-to-top dependencies: WAR */
   memset(last_grf_write, 0, sizeof(last_grf_write));
   memset(last_mrf_write, 0, sizeof(last_mrf_write));
   last_conditional_mod = NULL;
   last_accumulator_write = NULL;
   last_fixed_grf_write = NULL;

   exec_node *node;
   exec_node *prev;
   for (node = instructions.get_tail(), prev = node->prev;
        !node->is_head_sentinel();
        node = prev, prev = node->prev) {
      schedule_node *n = (schedule_node *)node;
      vec4_instruction *inst = (vec4_instruction *)n->inst;

      /* write-after-read deps. */
      for (int i = 0; i < 3; i++) {
         if (inst->src[i].file == GRF) {
            for (unsigned j = 0; j < inst->regs_read(i); ++j)
               add_dep(n, last_grf_write[inst->src[i].reg + j]);
         } else if (inst->src[i].file == HW_REG &&
                    (inst->src[i].fixed_hw_reg.file ==
                     BRW_GENERAL_REGISTER_FILE)) {
            add_dep(n, last_fixed_grf_write);
         } else if (inst->src[i].is_accumulator()) {
            add_dep(n, last_accumulator_write);
         } else if (inst->src[i].file != BAD_FILE &&
                    inst->src[i].file != IMM &&
                    inst->src[i].file != UNIFORM &&
                    (inst->src[i].file != HW_REG ||
                     inst->src[i].fixed_hw_reg.file != IMM)) {
            assert(inst->src[i].file != MRF &&
                   inst->src[i].file != ATTR);
            add_barrier_deps(n);
         }
      }

      if (!inst->is_send_from_grf()) {
         for (int i = 0; i < inst->mlen; i++) {
            /* It looks like the MRF regs are released in the send
             * instruction once it's sent, not when the result comes
             * back.
             */
            add_dep(n, last_mrf_write[inst->base_mrf + i], 2);
         }
      }

      if (inst->reads_flag()) {
         add_dep(n, last_conditional_mod);
      }

      if (inst->reads_accumulator_implicitly()) {
         add_dep(n, last_accumulator_write);
      }

      /* Update the things this instruction wrote, so earlier reads
       * can mark this as WAR dependency.
       */
      if (inst->dst.file == GRF) {
         for (unsigned j = 0; j < inst->regs_written; ++j)
            last_grf_write[inst->dst.reg + j] = n;
      } else if (inst->dst.file == MRF) {
         last_mrf_write[inst->dst.reg] = n;
      } else if (inst->dst.file == HW_REG &&
                 inst->dst.fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
         last_fixed_grf_write = n;
      } else if (inst->dst.is_accumulator()) {
         last_accumulator_write = n;
      } else if (inst->dst.file != BAD_FILE &&
                 !inst->dst.is_null()) {
         add_barrier_deps(n);
      }

      if (inst->mlen > 0 && !inst->is_send_from_grf()) {
         for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
            last_mrf_write[inst->base_mrf + i] = n;
         }
      }

      if (inst->writes_flag()) {
         last_conditional_mod = n;
      }

      if (inst->writes_accumulator_implicitly(v->brw)) {
         last_accumulator_write = n;
      }
   }
}

schedule_node *
fs_instruction_scheduler::choose_instruction_to_schedule()
{
   struct brw_context *brw = v->brw;
   schedule_node *chosen = NULL;

   if (mode == SCHEDULE_PRE || mode == SCHEDULE_POST) {
      int chosen_time = 0;

      /* Of the instructions ready to execute or the closest to
       * being ready, choose the oldest one.
       */
      foreach_in_list(schedule_node, n, &instructions) {
         if (!chosen || n->unblocked_time < chosen_time) {
            chosen = n;
            chosen_time = n->unblocked_time;
         }
      }
   } else {
      /* Before register allocation, we don't care about the latencies of
       * instructions.  All we care about is reducing live intervals of
       * variables so that we can avoid register spilling, or get SIMD16
       * shaders which naturally do a better job of hiding instruction
       * latency.
       */
      foreach_in_list(schedule_node, n, &instructions) {
         fs_inst *inst = (fs_inst *)n->inst;

         if (!chosen) {
            chosen = n;
            continue;
         }

         /* Most important: If we can definitely reduce register pressure, do
          * so immediately.
          */
         int register_pressure_benefit = get_register_pressure_benefit(n->inst);
         int chosen_register_pressure_benefit =
            get_register_pressure_benefit(chosen->inst);

         if (register_pressure_benefit > 0 &&
             register_pressure_benefit > chosen_register_pressure_benefit) {
            chosen = n;
            continue;
         } else if (chosen_register_pressure_benefit > 0 &&
                    (register_pressure_benefit <
                     chosen_register_pressure_benefit)) {
            continue;
         }

         if (mode == SCHEDULE_PRE_LIFO) {
            /* Prefer instructions that recently became available for
             * scheduling.  These are the things that are most likely to
             * (eventually) make a variable dead and reduce register pressure.
             * Typical register pressure estimates don't work for us because
             * most of our pressure comes from texturing, where no single
             * instruction to schedule will make a vec4 value dead.
             */
            if (n->cand_generation > chosen->cand_generation) {
               chosen = n;
               continue;
            } else if (n->cand_generation < chosen->cand_generation) {
               continue;
            }

            /* On MRF-using chips, prefer non-SEND instructions.  If we don't
             * do this, then because we prefer instructions that just became
             * candidates, we'll end up in a pattern of scheduling a SEND,
             * then the MRFs for the next SEND, then the next SEND, then the
             * MRFs, etc., without ever consuming the results of a send.
             */
            if (brw->gen < 7) {
               fs_inst *chosen_inst = (fs_inst *)chosen->inst;

               /* We use regs_written > 1 as our test for the kind of send
                * instruction to avoid -- only sends generate many regs, and a
                * single-result send is probably actually reducing register
                * pressure.
                */
               if (inst->regs_written <= inst->dst.width / 8 &&
                   chosen_inst->regs_written > chosen_inst->dst.width / 8) {
                  chosen = n;
                  continue;
               } else if (inst->regs_written > chosen_inst->regs_written) {
                  continue;
               }
            }
         }

         /* For instructions pushed on the cands list at the same time, prefer
          * the one with the highest delay to the end of the program.  This is
          * most likely to have its values able to be consumed first (such as
          * for a large tree of lowered ubo loads, which appear reversed in
          * the instruction stream with respect to when they can be consumed).
          */
         if (n->delay > chosen->delay) {
            chosen = n;
            continue;
         } else if (n->delay < chosen->delay) {
            continue;
         }

         /* If all other metrics are equal, we prefer the first instruction in
          * the list (program execution).
          */
      }
   }

   return chosen;
}

schedule_node *
vec4_instruction_scheduler::choose_instruction_to_schedule()
{
   schedule_node *chosen = NULL;
   int chosen_time = 0;

   /* Of the instructions ready to execute or the closest to being ready,
    * choose the oldest one.
    */
   foreach_in_list(schedule_node, n, &instructions) {
      if (!chosen || n->unblocked_time < chosen_time) {
         chosen = n;
         chosen_time = n->unblocked_time;
      }
   }

   return chosen;
}

int
fs_instruction_scheduler::issue_time(backend_instruction *inst)
{
   if (is_compressed((fs_inst *)inst))
      return 4;
   else
      return 2;
}

int
vec4_instruction_scheduler::issue_time(backend_instruction *inst)
{
   /* We always execute as two vec4s in parallel. */
   return 2;
}

void
instruction_scheduler::schedule_instructions(bblock_t *block)
{
   struct brw_context *brw = bv->brw;
   backend_instruction *inst = block->end();
   time = 0;

   /* Remove non-DAG heads from the list. */
   foreach_in_list_safe(schedule_node, n, &instructions) {
      if (n->parent_count != 0)
         n->remove();
   }

   unsigned cand_generation = 1;
   while (!instructions.is_empty()) {
      schedule_node *chosen = choose_instruction_to_schedule();

      /* Schedule this instruction. */
      assert(chosen);
      chosen->remove();
      inst->insert_before(block, chosen->inst);
      instructions_to_schedule--;
      update_register_pressure(chosen->inst);

      /* Update the clock for how soon an instruction could start after the
       * chosen one.
       */
      time += issue_time(chosen->inst);

      /* If we expected a delay for scheduling, then bump the clock to reflect
       * that as well.  In reality, the hardware will switch to another
       * hyperthread and may not return to dispatching our thread for a while
       * even after we're unblocked.
       */
      time = MAX2(time, chosen->unblocked_time);

      if (debug) {
         fprintf(stderr, "clock %4d, scheduled: ", time);
         bv->dump_instruction(chosen->inst);
      }

      /* Now that we've scheduled a new instruction, some of its
       * children can be promoted to the list of instructions ready to
       * be scheduled.  Update the children's unblocked time for this
       * DAG edge as we do so.
       */
      for (int i = chosen->child_count - 1; i >= 0; i--) {
         schedule_node *child = chosen->children[i];

         child->unblocked_time = MAX2(child->unblocked_time,
                                      time + chosen->child_latency[i]);

         if (debug) {
            fprintf(stderr, "\tchild %d, %d parents: ", i, child->parent_count);
            bv->dump_instruction(child->inst);
         }

         child->cand_generation = cand_generation;
         child->parent_count--;
         if (child->parent_count == 0) {
            if (debug) {
               fprintf(stderr, "\t\tnow available\n");
            }
            instructions.push_head(child);
         }
      }
      cand_generation++;

      /* Shared resource: the mathbox.  There's one mathbox per EU on Gen6+
       * but it's more limited pre-gen6, so if we send something off to it then
       * the next math instruction isn't going to make progress until the first
       * is done.
       */
      if (brw->gen < 6 && chosen->inst->is_math()) {
         foreach_in_list(schedule_node, n, &instructions) {
            if (n->inst->is_math())
               n->unblocked_time = MAX2(n->unblocked_time,
                                        time + chosen->latency);
         }
      }
   }

   if (block->end()->opcode == BRW_OPCODE_NOP)
      block->end()->remove(block);
   assert(instructions_to_schedule == 0);
}

void
instruction_scheduler::run(cfg_t *cfg)
{
   if (debug) {
      fprintf(stderr, "\nInstructions before scheduling (reg_alloc %d)\n",
              post_reg_alloc);
      bv->dump_instructions();
   }

   /* Populate the remaining GRF uses array to improve the pre-regalloc
    * scheduling.
    */
   if (remaining_grf_uses) {
      foreach_block_and_inst(block, backend_instruction, inst, cfg) {
         count_remaining_grf_uses(inst);
      }
   }

   foreach_block(block, cfg) {
      if (block->end_ip - block->start_ip <= 1)
         continue;

      add_insts_from_block(block);

      calculate_deps();

      foreach_in_list(schedule_node, n, &instructions) {
         compute_delay(n);
      }

      schedule_instructions(block);
   }

   if (debug) {
      fprintf(stderr, "\nInstructions after scheduling (reg_alloc %d)\n",
              post_reg_alloc);
      bv->dump_instructions();
   }
}

void
fs_visitor::schedule_instructions(instruction_scheduler_mode mode)
{
   int grf_count;
   if (mode == SCHEDULE_POST)
      grf_count = grf_used;
   else
      grf_count = alloc.count;

   fs_instruction_scheduler sched(this, grf_count, mode);
   sched.run(cfg);

   if (unlikely(debug_enabled) && mode == SCHEDULE_POST) {
      fprintf(stderr, "%s%d estimated execution time: %d cycles\n",
              stage_abbrev, dispatch_width, sched.time);
   }

   invalidate_live_intervals();
}

void
vec4_visitor::opt_schedule_instructions()
{
   vec4_instruction_scheduler sched(this, prog_data->total_grf);
   sched.run(cfg);

   if (unlikely(debug_enabled)) {
      fprintf(stderr, "%s estimated execution time: %d cycles\n",
              stage_abbrev, sched.time);
   }

   invalidate_live_intervals();
}