forked from llvm/llvm-project
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathVPlanTransforms.cpp
2649 lines (2376 loc) · 108 KB
/
VPlanTransforms.cpp
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
//===-- VPlanTransforms.cpp - Utility VPlan to VPlan transforms -----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This file implements a set of utility VPlan to VPlan transformations.
///
//===----------------------------------------------------------------------===//
#include "VPlanTransforms.h"
#include "VPRecipeBuilder.h"
#include "VPlan.h"
#include "VPlanAnalysis.h"
#include "VPlanCFG.h"
#include "VPlanDominatorTree.h"
#include "VPlanPatternMatch.h"
#include "VPlanUtils.h"
#include "VPlanVerifier.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/TypeSwitch.h"
#include "llvm/Analysis/IVDescriptors.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/TypeSize.h"
using namespace llvm;
bool VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(
VPlanPtr &Plan,
function_ref<const InductionDescriptor *(PHINode *)>
GetIntOrFpInductionDescriptor,
ScalarEvolution &SE, const TargetLibraryInfo &TLI) {
ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
Plan->getVectorLoopRegion());
for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
// Skip blocks outside region
if (!VPBB->getParent())
break;
VPRecipeBase *Term = VPBB->getTerminator();
auto EndIter = Term ? Term->getIterator() : VPBB->end();
// Introduce each ingredient into VPlan.
for (VPRecipeBase &Ingredient :
make_early_inc_range(make_range(VPBB->begin(), EndIter))) {
VPValue *VPV = Ingredient.getVPSingleValue();
Instruction *Inst = cast<Instruction>(VPV->getUnderlyingValue());
VPRecipeBase *NewRecipe = nullptr;
if (auto *VPPhi = dyn_cast<VPWidenPHIRecipe>(&Ingredient)) {
auto *Phi = cast<PHINode>(VPPhi->getUnderlyingValue());
const auto *II = GetIntOrFpInductionDescriptor(Phi);
if (!II)
continue;
VPValue *Start = Plan->getOrAddLiveIn(II->getStartValue());
VPValue *Step =
vputils::getOrCreateVPValueForSCEVExpr(*Plan, II->getStep(), SE);
NewRecipe = new VPWidenIntOrFpInductionRecipe(
Phi, Start, Step, &Plan->getVF(), *II, Ingredient.getDebugLoc());
} else {
assert(isa<VPInstruction>(&Ingredient) &&
"only VPInstructions expected here");
assert(!isa<PHINode>(Inst) && "phis should be handled above");
// Create VPWidenMemoryRecipe for loads and stores.
if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
NewRecipe = new VPWidenLoadRecipe(
*Load, Ingredient.getOperand(0), nullptr /*Mask*/,
false /*Consecutive*/, false /*Reverse*/,
Ingredient.getDebugLoc());
} else if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) {
NewRecipe = new VPWidenStoreRecipe(
*Store, Ingredient.getOperand(1), Ingredient.getOperand(0),
nullptr /*Mask*/, false /*Consecutive*/, false /*Reverse*/,
Ingredient.getDebugLoc());
} else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
NewRecipe = new VPWidenGEPRecipe(GEP, Ingredient.operands());
} else if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
Intrinsic::ID VectorID = getVectorIntrinsicIDForCall(CI, &TLI);
if (VectorID == Intrinsic::not_intrinsic)
return false;
NewRecipe = new VPWidenIntrinsicRecipe(
*CI, getVectorIntrinsicIDForCall(CI, &TLI),
{Ingredient.op_begin(), Ingredient.op_end() - 1}, CI->getType(),
CI->getDebugLoc());
} else if (SelectInst *SI = dyn_cast<SelectInst>(Inst)) {
NewRecipe = new VPWidenSelectRecipe(*SI, Ingredient.operands());
} else if (auto *CI = dyn_cast<CastInst>(Inst)) {
NewRecipe = new VPWidenCastRecipe(
CI->getOpcode(), Ingredient.getOperand(0), CI->getType(), *CI);
} else {
NewRecipe = new VPWidenRecipe(*Inst, Ingredient.operands());
}
}
NewRecipe->insertBefore(&Ingredient);
if (NewRecipe->getNumDefinedValues() == 1)
VPV->replaceAllUsesWith(NewRecipe->getVPSingleValue());
else
assert(NewRecipe->getNumDefinedValues() == 0 &&
"Only recpies with zero or one defined values expected");
Ingredient.eraseFromParent();
}
}
return true;
}
static bool sinkScalarOperands(VPlan &Plan) {
auto Iter = vp_depth_first_deep(Plan.getEntry());
bool Changed = false;
// First, collect the operands of all recipes in replicate blocks as seeds for
// sinking.
SetVector<std::pair<VPBasicBlock *, VPSingleDefRecipe *>> WorkList;
for (VPRegionBlock *VPR : VPBlockUtils::blocksOnly<VPRegionBlock>(Iter)) {
VPBasicBlock *EntryVPBB = VPR->getEntryBasicBlock();
if (!VPR->isReplicator() || EntryVPBB->getSuccessors().size() != 2)
continue;
VPBasicBlock *VPBB = dyn_cast<VPBasicBlock>(EntryVPBB->getSuccessors()[0]);
if (!VPBB || VPBB->getSingleSuccessor() != VPR->getExitingBasicBlock())
continue;
for (auto &Recipe : *VPBB) {
for (VPValue *Op : Recipe.operands())
if (auto *Def =
dyn_cast_or_null<VPSingleDefRecipe>(Op->getDefiningRecipe()))
WorkList.insert(std::make_pair(VPBB, Def));
}
}
bool ScalarVFOnly = Plan.hasScalarVFOnly();
// Try to sink each replicate or scalar IV steps recipe in the worklist.
for (unsigned I = 0; I != WorkList.size(); ++I) {
VPBasicBlock *SinkTo;
VPSingleDefRecipe *SinkCandidate;
std::tie(SinkTo, SinkCandidate) = WorkList[I];
if (SinkCandidate->getParent() == SinkTo ||
SinkCandidate->mayHaveSideEffects() ||
SinkCandidate->mayReadOrWriteMemory())
continue;
if (auto *RepR = dyn_cast<VPReplicateRecipe>(SinkCandidate)) {
if (!ScalarVFOnly && RepR->isUniform())
continue;
} else if (!isa<VPScalarIVStepsRecipe>(SinkCandidate))
continue;
bool NeedsDuplicating = false;
// All recipe users of the sink candidate must be in the same block SinkTo
// or all users outside of SinkTo must be uniform-after-vectorization (
// i.e., only first lane is used) . In the latter case, we need to duplicate
// SinkCandidate.
auto CanSinkWithUser = [SinkTo, &NeedsDuplicating,
SinkCandidate](VPUser *U) {
auto *UI = cast<VPRecipeBase>(U);
if (UI->getParent() == SinkTo)
return true;
NeedsDuplicating = UI->onlyFirstLaneUsed(SinkCandidate);
// We only know how to duplicate VPRecipeRecipes for now.
return NeedsDuplicating && isa<VPReplicateRecipe>(SinkCandidate);
};
if (!all_of(SinkCandidate->users(), CanSinkWithUser))
continue;
if (NeedsDuplicating) {
if (ScalarVFOnly)
continue;
Instruction *I = SinkCandidate->getUnderlyingInstr();
auto *Clone = new VPReplicateRecipe(I, SinkCandidate->operands(), true);
// TODO: add ".cloned" suffix to name of Clone's VPValue.
Clone->insertBefore(SinkCandidate);
SinkCandidate->replaceUsesWithIf(Clone, [SinkTo](VPUser &U, unsigned) {
return cast<VPRecipeBase>(&U)->getParent() != SinkTo;
});
}
SinkCandidate->moveBefore(*SinkTo, SinkTo->getFirstNonPhi());
for (VPValue *Op : SinkCandidate->operands())
if (auto *Def =
dyn_cast_or_null<VPSingleDefRecipe>(Op->getDefiningRecipe()))
WorkList.insert(std::make_pair(SinkTo, Def));
Changed = true;
}
return Changed;
}
/// If \p R is a region with a VPBranchOnMaskRecipe in the entry block, return
/// the mask.
VPValue *getPredicatedMask(VPRegionBlock *R) {
auto *EntryBB = dyn_cast<VPBasicBlock>(R->getEntry());
if (!EntryBB || EntryBB->size() != 1 ||
!isa<VPBranchOnMaskRecipe>(EntryBB->begin()))
return nullptr;
return cast<VPBranchOnMaskRecipe>(&*EntryBB->begin())->getOperand(0);
}
/// If \p R is a triangle region, return the 'then' block of the triangle.
static VPBasicBlock *getPredicatedThenBlock(VPRegionBlock *R) {
auto *EntryBB = cast<VPBasicBlock>(R->getEntry());
if (EntryBB->getNumSuccessors() != 2)
return nullptr;
auto *Succ0 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[0]);
auto *Succ1 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[1]);
if (!Succ0 || !Succ1)
return nullptr;
if (Succ0->getNumSuccessors() + Succ1->getNumSuccessors() != 1)
return nullptr;
if (Succ0->getSingleSuccessor() == Succ1)
return Succ0;
if (Succ1->getSingleSuccessor() == Succ0)
return Succ1;
return nullptr;
}
// Merge replicate regions in their successor region, if a replicate region
// is connected to a successor replicate region with the same predicate by a
// single, empty VPBasicBlock.
static bool mergeReplicateRegionsIntoSuccessors(VPlan &Plan) {
SmallPtrSet<VPRegionBlock *, 4> TransformedRegions;
// Collect replicate regions followed by an empty block, followed by another
// replicate region with matching masks to process front. This is to avoid
// iterator invalidation issues while merging regions.
SmallVector<VPRegionBlock *, 8> WorkList;
for (VPRegionBlock *Region1 : VPBlockUtils::blocksOnly<VPRegionBlock>(
vp_depth_first_deep(Plan.getEntry()))) {
if (!Region1->isReplicator())
continue;
auto *MiddleBasicBlock =
dyn_cast_or_null<VPBasicBlock>(Region1->getSingleSuccessor());
if (!MiddleBasicBlock || !MiddleBasicBlock->empty())
continue;
auto *Region2 =
dyn_cast_or_null<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor());
if (!Region2 || !Region2->isReplicator())
continue;
VPValue *Mask1 = getPredicatedMask(Region1);
VPValue *Mask2 = getPredicatedMask(Region2);
if (!Mask1 || Mask1 != Mask2)
continue;
assert(Mask1 && Mask2 && "both region must have conditions");
WorkList.push_back(Region1);
}
// Move recipes from Region1 to its successor region, if both are triangles.
for (VPRegionBlock *Region1 : WorkList) {
if (TransformedRegions.contains(Region1))
continue;
auto *MiddleBasicBlock = cast<VPBasicBlock>(Region1->getSingleSuccessor());
auto *Region2 = cast<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor());
VPBasicBlock *Then1 = getPredicatedThenBlock(Region1);
VPBasicBlock *Then2 = getPredicatedThenBlock(Region2);
if (!Then1 || !Then2)
continue;
// Note: No fusion-preventing memory dependencies are expected in either
// region. Such dependencies should be rejected during earlier dependence
// checks, which guarantee accesses can be re-ordered for vectorization.
//
// Move recipes to the successor region.
for (VPRecipeBase &ToMove : make_early_inc_range(reverse(*Then1)))
ToMove.moveBefore(*Then2, Then2->getFirstNonPhi());
auto *Merge1 = cast<VPBasicBlock>(Then1->getSingleSuccessor());
auto *Merge2 = cast<VPBasicBlock>(Then2->getSingleSuccessor());
// Move VPPredInstPHIRecipes from the merge block to the successor region's
// merge block. Update all users inside the successor region to use the
// original values.
for (VPRecipeBase &Phi1ToMove : make_early_inc_range(reverse(*Merge1))) {
VPValue *PredInst1 =
cast<VPPredInstPHIRecipe>(&Phi1ToMove)->getOperand(0);
VPValue *Phi1ToMoveV = Phi1ToMove.getVPSingleValue();
Phi1ToMoveV->replaceUsesWithIf(PredInst1, [Then2](VPUser &U, unsigned) {
return cast<VPRecipeBase>(&U)->getParent() == Then2;
});
// Remove phi recipes that are unused after merging the regions.
if (Phi1ToMove.getVPSingleValue()->getNumUsers() == 0) {
Phi1ToMove.eraseFromParent();
continue;
}
Phi1ToMove.moveBefore(*Merge2, Merge2->begin());
}
// Remove the dead recipes in Region1's entry block.
for (VPRecipeBase &R :
make_early_inc_range(reverse(*Region1->getEntryBasicBlock())))
R.eraseFromParent();
// Finally, remove the first region.
for (VPBlockBase *Pred : make_early_inc_range(Region1->getPredecessors())) {
VPBlockUtils::disconnectBlocks(Pred, Region1);
VPBlockUtils::connectBlocks(Pred, MiddleBasicBlock);
}
VPBlockUtils::disconnectBlocks(Region1, MiddleBasicBlock);
TransformedRegions.insert(Region1);
}
return !TransformedRegions.empty();
}
static VPRegionBlock *createReplicateRegion(VPReplicateRecipe *PredRecipe,
VPlan &Plan) {
Instruction *Instr = PredRecipe->getUnderlyingInstr();
// Build the triangular if-then region.
std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str();
assert(Instr->getParent() && "Predicated instruction not in any basic block");
auto *BlockInMask = PredRecipe->getMask();
auto *MaskDef = BlockInMask->getDefiningRecipe();
auto *BOMRecipe = new VPBranchOnMaskRecipe(
BlockInMask, MaskDef ? MaskDef->getDebugLoc() : DebugLoc());
auto *Entry =
Plan.createVPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe);
// Replace predicated replicate recipe with a replicate recipe without a
// mask but in the replicate region.
auto *RecipeWithoutMask = new VPReplicateRecipe(
PredRecipe->getUnderlyingInstr(),
make_range(PredRecipe->op_begin(), std::prev(PredRecipe->op_end())),
PredRecipe->isUniform());
auto *Pred =
Plan.createVPBasicBlock(Twine(RegionName) + ".if", RecipeWithoutMask);
VPPredInstPHIRecipe *PHIRecipe = nullptr;
if (PredRecipe->getNumUsers() != 0) {
PHIRecipe = new VPPredInstPHIRecipe(RecipeWithoutMask,
RecipeWithoutMask->getDebugLoc());
PredRecipe->replaceAllUsesWith(PHIRecipe);
PHIRecipe->setOperand(0, RecipeWithoutMask);
}
PredRecipe->eraseFromParent();
auto *Exiting =
Plan.createVPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe);
VPRegionBlock *Region =
Plan.createVPRegionBlock(Entry, Exiting, RegionName, true);
// Note: first set Entry as region entry and then connect successors starting
// from it in order, to propagate the "parent" of each VPBasicBlock.
VPBlockUtils::insertTwoBlocksAfter(Pred, Exiting, Entry);
VPBlockUtils::connectBlocks(Pred, Exiting);
return Region;
}
static void addReplicateRegions(VPlan &Plan) {
SmallVector<VPReplicateRecipe *> WorkList;
for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
vp_depth_first_deep(Plan.getEntry()))) {
for (VPRecipeBase &R : *VPBB)
if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) {
if (RepR->isPredicated())
WorkList.push_back(RepR);
}
}
unsigned BBNum = 0;
for (VPReplicateRecipe *RepR : WorkList) {
VPBasicBlock *CurrentBlock = RepR->getParent();
VPBasicBlock *SplitBlock = CurrentBlock->splitAt(RepR->getIterator());
BasicBlock *OrigBB = RepR->getUnderlyingInstr()->getParent();
SplitBlock->setName(
OrigBB->hasName() ? OrigBB->getName() + "." + Twine(BBNum++) : "");
// Record predicated instructions for above packing optimizations.
VPBlockBase *Region = createReplicateRegion(RepR, Plan);
Region->setParent(CurrentBlock->getParent());
VPBlockUtils::insertOnEdge(CurrentBlock, SplitBlock, Region);
}
}
/// Remove redundant VPBasicBlocks by merging them into their predecessor if
/// the predecessor has a single successor.
static bool mergeBlocksIntoPredecessors(VPlan &Plan) {
SmallVector<VPBasicBlock *> WorkList;
for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
vp_depth_first_deep(Plan.getEntry()))) {
// Don't fold the blocks in the skeleton of the Plan into their single
// predecessors for now.
// TODO: Remove restriction once more of the skeleton is modeled in VPlan.
if (!VPBB->getParent())
continue;
auto *PredVPBB =
dyn_cast_or_null<VPBasicBlock>(VPBB->getSinglePredecessor());
if (!PredVPBB || PredVPBB->getNumSuccessors() != 1 ||
isa<VPIRBasicBlock>(PredVPBB))
continue;
WorkList.push_back(VPBB);
}
for (VPBasicBlock *VPBB : WorkList) {
VPBasicBlock *PredVPBB = cast<VPBasicBlock>(VPBB->getSinglePredecessor());
for (VPRecipeBase &R : make_early_inc_range(*VPBB))
R.moveBefore(*PredVPBB, PredVPBB->end());
VPBlockUtils::disconnectBlocks(PredVPBB, VPBB);
auto *ParentRegion = VPBB->getParent();
if (ParentRegion && ParentRegion->getExiting() == VPBB)
ParentRegion->setExiting(PredVPBB);
for (auto *Succ : to_vector(VPBB->successors())) {
VPBlockUtils::disconnectBlocks(VPBB, Succ);
VPBlockUtils::connectBlocks(PredVPBB, Succ);
}
// VPBB is now dead and will be cleaned up when the plan gets destroyed.
}
return !WorkList.empty();
}
void VPlanTransforms::createAndOptimizeReplicateRegions(VPlan &Plan) {
// Convert masked VPReplicateRecipes to if-then region blocks.
addReplicateRegions(Plan);
bool ShouldSimplify = true;
while (ShouldSimplify) {
ShouldSimplify = sinkScalarOperands(Plan);
ShouldSimplify |= mergeReplicateRegionsIntoSuccessors(Plan);
ShouldSimplify |= mergeBlocksIntoPredecessors(Plan);
}
}
/// Remove redundant casts of inductions.
///
/// Such redundant casts are casts of induction variables that can be ignored,
/// because we already proved that the casted phi is equal to the uncasted phi
/// in the vectorized loop. There is no need to vectorize the cast - the same
/// value can be used for both the phi and casts in the vector loop.
static void removeRedundantInductionCasts(VPlan &Plan) {
for (auto &Phi : Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) {
auto *IV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
if (!IV || IV->getTruncInst())
continue;
// A sequence of IR Casts has potentially been recorded for IV, which
// *must be bypassed* when the IV is vectorized, because the vectorized IV
// will produce the desired casted value. This sequence forms a def-use
// chain and is provided in reverse order, ending with the cast that uses
// the IV phi. Search for the recipe of the last cast in the chain and
// replace it with the original IV. Note that only the final cast is
// expected to have users outside the cast-chain and the dead casts left
// over will be cleaned up later.
auto &Casts = IV->getInductionDescriptor().getCastInsts();
VPValue *FindMyCast = IV;
for (Instruction *IRCast : reverse(Casts)) {
VPSingleDefRecipe *FoundUserCast = nullptr;
for (auto *U : FindMyCast->users()) {
auto *UserCast = dyn_cast<VPSingleDefRecipe>(U);
if (UserCast && UserCast->getUnderlyingValue() == IRCast) {
FoundUserCast = UserCast;
break;
}
}
FindMyCast = FoundUserCast;
}
FindMyCast->replaceAllUsesWith(IV);
}
}
/// Try to replace VPWidenCanonicalIVRecipes with a widened canonical IV
/// recipe, if it exists.
static void removeRedundantCanonicalIVs(VPlan &Plan) {
VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
VPWidenCanonicalIVRecipe *WidenNewIV = nullptr;
for (VPUser *U : CanonicalIV->users()) {
WidenNewIV = dyn_cast<VPWidenCanonicalIVRecipe>(U);
if (WidenNewIV)
break;
}
if (!WidenNewIV)
return;
VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
if (!WidenOriginalIV || !WidenOriginalIV->isCanonical())
continue;
// Replace WidenNewIV with WidenOriginalIV if WidenOriginalIV provides
// everything WidenNewIV's users need. That is, WidenOriginalIV will
// generate a vector phi or all users of WidenNewIV demand the first lane
// only.
if (any_of(WidenOriginalIV->users(),
[WidenOriginalIV](VPUser *U) {
return !U->usesScalars(WidenOriginalIV);
}) ||
vputils::onlyFirstLaneUsed(WidenNewIV)) {
WidenNewIV->replaceAllUsesWith(WidenOriginalIV);
WidenNewIV->eraseFromParent();
return;
}
}
}
/// Returns true if \p R is dead and can be removed.
static bool isDeadRecipe(VPRecipeBase &R) {
using namespace llvm::PatternMatch;
// Do remove conditional assume instructions as their conditions may be
// flattened.
auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
bool IsConditionalAssume =
RepR && RepR->isPredicated() &&
match(RepR->getUnderlyingInstr(), m_Intrinsic<Intrinsic::assume>());
if (IsConditionalAssume)
return true;
if (R.mayHaveSideEffects())
return false;
// Recipe is dead if no user keeps the recipe alive.
return all_of(R.definedValues(),
[](VPValue *V) { return V->getNumUsers() == 0; });
}
void VPlanTransforms::removeDeadRecipes(VPlan &Plan) {
ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
Plan.getEntry());
for (VPBasicBlock *VPBB : reverse(VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT))) {
// The recipes in the block are processed in reverse order, to catch chains
// of dead recipes.
for (VPRecipeBase &R : make_early_inc_range(reverse(*VPBB))) {
if (isDeadRecipe(R))
R.eraseFromParent();
}
}
}
static VPScalarIVStepsRecipe *
createScalarIVSteps(VPlan &Plan, InductionDescriptor::InductionKind Kind,
Instruction::BinaryOps InductionOpcode,
FPMathOperator *FPBinOp, Instruction *TruncI,
VPValue *StartV, VPValue *Step, DebugLoc DL,
VPBuilder &Builder) {
VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
VPSingleDefRecipe *BaseIV = Builder.createDerivedIV(
Kind, FPBinOp, StartV, CanonicalIV, Step, "offset.idx");
// Truncate base induction if needed.
Type *CanonicalIVType = CanonicalIV->getScalarType();
VPTypeAnalysis TypeInfo(CanonicalIVType);
Type *ResultTy = TypeInfo.inferScalarType(BaseIV);
if (TruncI) {
Type *TruncTy = TruncI->getType();
assert(ResultTy->getScalarSizeInBits() > TruncTy->getScalarSizeInBits() &&
"Not truncating.");
assert(ResultTy->isIntegerTy() && "Truncation requires an integer type");
BaseIV = Builder.createScalarCast(Instruction::Trunc, BaseIV, TruncTy, DL);
ResultTy = TruncTy;
}
// Truncate step if needed.
Type *StepTy = TypeInfo.inferScalarType(Step);
if (ResultTy != StepTy) {
assert(StepTy->getScalarSizeInBits() > ResultTy->getScalarSizeInBits() &&
"Not truncating.");
assert(StepTy->isIntegerTy() && "Truncation requires an integer type");
auto *VecPreheader =
cast<VPBasicBlock>(HeaderVPBB->getSingleHierarchicalPredecessor());
VPBuilder::InsertPointGuard Guard(Builder);
Builder.setInsertPoint(VecPreheader);
Step = Builder.createScalarCast(Instruction::Trunc, Step, ResultTy, DL);
}
return Builder.createScalarIVSteps(InductionOpcode, FPBinOp, BaseIV, Step,
&Plan.getVF());
}
static SmallVector<VPUser *> collectUsersRecursively(VPValue *V) {
SetVector<VPUser *> Users(V->user_begin(), V->user_end());
for (unsigned I = 0; I != Users.size(); ++I) {
VPRecipeBase *Cur = cast<VPRecipeBase>(Users[I]);
if (isa<VPHeaderPHIRecipe>(Cur))
continue;
for (VPValue *V : Cur->definedValues())
Users.insert_range(V->users());
}
return Users.takeVector();
}
/// Legalize VPWidenPointerInductionRecipe, by replacing it with a PtrAdd
/// (IndStart, ScalarIVSteps (0, Step)) if only its scalar values are used, as
/// VPWidenPointerInductionRecipe will generate vectors only. If some users
/// require vectors while other require scalars, the scalar uses need to extract
/// the scalars from the generated vectors (Note that this is different to how
/// int/fp inductions are handled). Legalize extract-from-ends using uniform
/// VPReplicateRecipe of wide inductions to use regular VPReplicateRecipe, so
/// the correct end value is available. Also optimize
/// VPWidenIntOrFpInductionRecipe, if any of its users needs scalar values, by
/// providing them scalar steps built on the canonical scalar IV and update the
/// original IV's users. This is an optional optimization to reduce the needs of
/// vector extracts.
static void legalizeAndOptimizeInductions(VPlan &Plan) {
using namespace llvm::VPlanPatternMatch;
VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
bool HasOnlyVectorVFs = !Plan.hasScalarVFOnly();
VPBuilder Builder(HeaderVPBB, HeaderVPBB->getFirstNonPhi());
for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
auto *PhiR = dyn_cast<VPWidenInductionRecipe>(&Phi);
if (!PhiR)
continue;
// Try to narrow wide and replicating recipes to uniform recipes, based on
// VPlan analysis.
// TODO: Apply to all recipes in the future, to replace legacy uniformity
// analysis.
auto Users = collectUsersRecursively(PhiR);
for (VPUser *U : reverse(Users)) {
auto *Def = dyn_cast<VPSingleDefRecipe>(U);
auto *RepR = dyn_cast<VPReplicateRecipe>(U);
// Skip recipes that shouldn't be narrowed.
if (!Def || !isa<VPReplicateRecipe, VPWidenRecipe>(Def) ||
Def->getNumUsers() == 0 || !Def->getUnderlyingValue() ||
(RepR && (RepR->isUniform() || RepR->isPredicated())))
continue;
// Skip recipes that may have other lanes than their first used.
if (!vputils::isUniformAfterVectorization(Def) &&
!vputils::onlyFirstLaneUsed(Def))
continue;
auto *Clone = new VPReplicateRecipe(Def->getUnderlyingInstr(),
Def->operands(), /*IsUniform*/ true);
Clone->insertAfter(Def);
Def->replaceAllUsesWith(Clone);
}
// Replace wide pointer inductions which have only their scalars used by
// PtrAdd(IndStart, ScalarIVSteps (0, Step)).
if (auto *PtrIV = dyn_cast<VPWidenPointerInductionRecipe>(&Phi)) {
if (!PtrIV->onlyScalarsGenerated(Plan.hasScalableVF()))
continue;
const InductionDescriptor &ID = PtrIV->getInductionDescriptor();
VPValue *StartV =
Plan.getOrAddLiveIn(ConstantInt::get(ID.getStep()->getType(), 0));
VPValue *StepV = PtrIV->getOperand(1);
VPScalarIVStepsRecipe *Steps = createScalarIVSteps(
Plan, InductionDescriptor::IK_IntInduction, Instruction::Add, nullptr,
nullptr, StartV, StepV, PtrIV->getDebugLoc(), Builder);
VPValue *PtrAdd = Builder.createPtrAdd(PtrIV->getStartValue(), Steps,
PtrIV->getDebugLoc(), "next.gep");
PtrIV->replaceAllUsesWith(PtrAdd);
continue;
}
// Replace widened induction with scalar steps for users that only use
// scalars.
auto *WideIV = cast<VPWidenIntOrFpInductionRecipe>(&Phi);
if (HasOnlyVectorVFs && none_of(WideIV->users(), [WideIV](VPUser *U) {
return U->usesScalars(WideIV);
}))
continue;
const InductionDescriptor &ID = WideIV->getInductionDescriptor();
VPScalarIVStepsRecipe *Steps = createScalarIVSteps(
Plan, ID.getKind(), ID.getInductionOpcode(),
dyn_cast_or_null<FPMathOperator>(ID.getInductionBinOp()),
WideIV->getTruncInst(), WideIV->getStartValue(), WideIV->getStepValue(),
WideIV->getDebugLoc(), Builder);
// Update scalar users of IV to use Step instead.
if (!HasOnlyVectorVFs)
WideIV->replaceAllUsesWith(Steps);
else
WideIV->replaceUsesWithIf(Steps, [WideIV](VPUser &U, unsigned) {
return U.usesScalars(WideIV);
});
}
}
/// Check if \p VPV is an untruncated wide induction, either before or after the
/// increment. If so return the header IV (before the increment), otherwise
/// return null.
static VPWidenInductionRecipe *getOptimizableIVOf(VPValue *VPV) {
auto *WideIV = dyn_cast<VPWidenInductionRecipe>(VPV);
if (WideIV) {
// VPV itself is a wide induction, separately compute the end value for exit
// users if it is not a truncated IV.
auto *IntOrFpIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(WideIV);
return (IntOrFpIV && IntOrFpIV->getTruncInst()) ? nullptr : WideIV;
}
// Check if VPV is an optimizable induction increment.
VPRecipeBase *Def = VPV->getDefiningRecipe();
if (!Def || Def->getNumOperands() != 2)
return nullptr;
WideIV = dyn_cast<VPWidenInductionRecipe>(Def->getOperand(0));
if (!WideIV)
WideIV = dyn_cast<VPWidenInductionRecipe>(Def->getOperand(1));
if (!WideIV)
return nullptr;
auto IsWideIVInc = [&]() {
using namespace VPlanPatternMatch;
auto &ID = WideIV->getInductionDescriptor();
// Check if VPV increments the induction by the induction step.
VPValue *IVStep = WideIV->getStepValue();
switch (ID.getInductionOpcode()) {
case Instruction::Add:
return match(VPV, m_c_Binary<Instruction::Add>(m_Specific(WideIV),
m_Specific(IVStep)));
case Instruction::FAdd:
return match(VPV, m_c_Binary<Instruction::FAdd>(m_Specific(WideIV),
m_Specific(IVStep)));
case Instruction::FSub:
return match(VPV, m_Binary<Instruction::FSub>(m_Specific(WideIV),
m_Specific(IVStep)));
case Instruction::Sub: {
// IVStep will be the negated step of the subtraction. Check if Step == -1
// * IVStep.
VPValue *Step;
if (!match(VPV,
m_Binary<Instruction::Sub>(m_VPValue(), m_VPValue(Step))) ||
!Step->isLiveIn() || !IVStep->isLiveIn())
return false;
auto *StepCI = dyn_cast<ConstantInt>(Step->getLiveInIRValue());
auto *IVStepCI = dyn_cast<ConstantInt>(IVStep->getLiveInIRValue());
return StepCI && IVStepCI &&
StepCI->getValue() == (-1 * IVStepCI->getValue());
}
default:
return ID.getKind() == InductionDescriptor::IK_PtrInduction &&
match(VPV, m_GetElementPtr(m_Specific(WideIV),
m_Specific(WideIV->getStepValue())));
}
llvm_unreachable("should have been covered by switch above");
};
return IsWideIVInc() ? WideIV : nullptr;
}
/// Attempts to optimize the induction variable exit values for users in the
/// early exit block.
static VPValue *optimizeEarlyExitInductionUser(VPlan &Plan,
VPTypeAnalysis &TypeInfo,
VPBlockBase *PredVPBB,
VPValue *Op) {
using namespace VPlanPatternMatch;
VPValue *Incoming, *Mask;
if (!match(Op, m_VPInstruction<Instruction::ExtractElement>(
m_VPValue(Incoming),
m_VPInstruction<VPInstruction::FirstActiveLane>(
m_VPValue(Mask)))))
return nullptr;
auto *WideIV = getOptimizableIVOf(Incoming);
if (!WideIV)
return nullptr;
auto *WideIntOrFp = dyn_cast<VPWidenIntOrFpInductionRecipe>(WideIV);
if (WideIntOrFp && WideIntOrFp->getTruncInst())
return nullptr;
// Calculate the final index.
VPValue *EndValue = Plan.getCanonicalIV();
auto CanonicalIVType = Plan.getCanonicalIV()->getScalarType();
VPBuilder B(cast<VPBasicBlock>(PredVPBB));
DebugLoc DL = cast<VPInstruction>(Op)->getDebugLoc();
VPValue *FirstActiveLane =
B.createNaryOp(VPInstruction::FirstActiveLane, Mask, DL);
Type *FirstActiveLaneType = TypeInfo.inferScalarType(FirstActiveLane);
if (CanonicalIVType != FirstActiveLaneType) {
Instruction::CastOps CastOp =
CanonicalIVType->getScalarSizeInBits() <
FirstActiveLaneType->getScalarSizeInBits()
? Instruction::Trunc
: Instruction::ZExt;
FirstActiveLane =
B.createScalarCast(CastOp, FirstActiveLane, CanonicalIVType, DL);
}
EndValue = B.createNaryOp(Instruction::Add, {EndValue, FirstActiveLane}, DL);
// `getOptimizableIVOf()` always returns the pre-incremented IV, so if it
// changed it means the exit is using the incremented value, so we need to
// add the step.
if (Incoming != WideIV) {
VPValue *One = Plan.getOrAddLiveIn(ConstantInt::get(CanonicalIVType, 1));
EndValue = B.createNaryOp(Instruction::Add, {EndValue, One}, DL);
}
if (!WideIntOrFp || !WideIntOrFp->isCanonical()) {
const InductionDescriptor &ID = WideIV->getInductionDescriptor();
VPValue *Start = WideIV->getStartValue();
VPValue *Step = WideIV->getStepValue();
EndValue = B.createDerivedIV(
ID.getKind(), dyn_cast_or_null<FPMathOperator>(ID.getInductionBinOp()),
Start, EndValue, Step);
}
return EndValue;
}
/// Attempts to optimize the induction variable exit values for users in the
/// exit block coming from the latch in the original scalar loop.
static VPValue *
optimizeLatchExitInductionUser(VPlan &Plan, VPTypeAnalysis &TypeInfo,
VPBlockBase *PredVPBB, VPValue *Op,
DenseMap<VPValue *, VPValue *> &EndValues) {
using namespace VPlanPatternMatch;
VPValue *Incoming;
if (!match(Op, m_VPInstruction<VPInstruction::ExtractFromEnd>(
m_VPValue(Incoming), m_SpecificInt(1))))
return nullptr;
auto *WideIV = getOptimizableIVOf(Incoming);
if (!WideIV)
return nullptr;
VPValue *EndValue = EndValues.lookup(WideIV);
assert(EndValue && "end value must have been pre-computed");
// `getOptimizableIVOf()` always returns the pre-incremented IV, so if it
// changed it means the exit is using the incremented value, so we don't
// need to subtract the step.
if (Incoming != WideIV)
return EndValue;
// Otherwise, subtract the step from the EndValue.
VPBuilder B(cast<VPBasicBlock>(PredVPBB)->getTerminator());
VPValue *Step = WideIV->getStepValue();
Type *ScalarTy = TypeInfo.inferScalarType(WideIV);
if (ScalarTy->isIntegerTy())
return B.createNaryOp(Instruction::Sub, {EndValue, Step}, {}, "ind.escape");
if (ScalarTy->isPointerTy()) {
auto *Zero = Plan.getOrAddLiveIn(
ConstantInt::get(Step->getLiveInIRValue()->getType(), 0));
return B.createPtrAdd(EndValue,
B.createNaryOp(Instruction::Sub, {Zero, Step}), {},
"ind.escape");
}
if (ScalarTy->isFloatingPointTy()) {
const auto &ID = WideIV->getInductionDescriptor();
return B.createNaryOp(
ID.getInductionBinOp()->getOpcode() == Instruction::FAdd
? Instruction::FSub
: Instruction::FAdd,
{EndValue, Step}, {ID.getInductionBinOp()->getFastMathFlags()});
}
llvm_unreachable("all possible induction types must be handled");
return nullptr;
}
void VPlanTransforms::optimizeInductionExitUsers(
VPlan &Plan, DenseMap<VPValue *, VPValue *> &EndValues) {
VPBlockBase *MiddleVPBB = Plan.getMiddleBlock();
VPTypeAnalysis TypeInfo(Plan.getCanonicalIV()->getScalarType());
for (VPIRBasicBlock *ExitVPBB : Plan.getExitBlocks()) {
for (VPRecipeBase &R : *ExitVPBB) {
auto *ExitIRI = dyn_cast<VPIRPhi>(&R);
if (!ExitIRI)
break;
for (auto [Idx, PredVPBB] : enumerate(ExitVPBB->getPredecessors())) {
VPValue *Escape = nullptr;
if (PredVPBB == MiddleVPBB)
Escape = optimizeLatchExitInductionUser(
Plan, TypeInfo, PredVPBB, ExitIRI->getOperand(Idx), EndValues);
else
Escape = optimizeEarlyExitInductionUser(Plan, TypeInfo, PredVPBB,
ExitIRI->getOperand(Idx));
if (Escape)
ExitIRI->setOperand(Idx, Escape);
}
}
}
}
/// Remove redundant EpxandSCEVRecipes in \p Plan's entry block by replacing
/// them with already existing recipes expanding the same SCEV expression.
static void removeRedundantExpandSCEVRecipes(VPlan &Plan) {
DenseMap<const SCEV *, VPValue *> SCEV2VPV;
for (VPRecipeBase &R :
make_early_inc_range(*Plan.getEntry()->getEntryBasicBlock())) {
auto *ExpR = dyn_cast<VPExpandSCEVRecipe>(&R);
if (!ExpR)
continue;
auto I = SCEV2VPV.insert({ExpR->getSCEV(), ExpR});
if (I.second)
continue;
ExpR->replaceAllUsesWith(I.first->second);
ExpR->eraseFromParent();
}
}
static void recursivelyDeleteDeadRecipes(VPValue *V) {
SmallVector<VPValue *> WorkList;
SmallPtrSet<VPValue *, 8> Seen;
WorkList.push_back(V);
while (!WorkList.empty()) {
VPValue *Cur = WorkList.pop_back_val();
if (!Seen.insert(Cur).second)
continue;
VPRecipeBase *R = Cur->getDefiningRecipe();
if (!R)
continue;
if (!isDeadRecipe(*R))
continue;
WorkList.append(R->op_begin(), R->op_end());
R->eraseFromParent();
}
}
/// Try to simplify recipe \p R.
static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
using namespace llvm::VPlanPatternMatch;
if (auto *Blend = dyn_cast<VPBlendRecipe>(&R)) {
// Try to remove redundant blend recipes.
SmallPtrSet<VPValue *, 4> UniqueValues;
if (Blend->isNormalized() || !match(Blend->getMask(0), m_False()))
UniqueValues.insert(Blend->getIncomingValue(0));
for (unsigned I = 1; I != Blend->getNumIncomingValues(); ++I)
if (!match(Blend->getMask(I), m_False()))
UniqueValues.insert(Blend->getIncomingValue(I));
if (UniqueValues.size() == 1) {
Blend->replaceAllUsesWith(*UniqueValues.begin());
Blend->eraseFromParent();
return;
}
if (Blend->isNormalized())
return;
// Normalize the blend so its first incoming value is used as the initial
// value with the others blended into it.
unsigned StartIndex = 0;
for (unsigned I = 0; I != Blend->getNumIncomingValues(); ++I) {
// If a value's mask is used only by the blend then is can be deadcoded.
// TODO: Find the most expensive mask that can be deadcoded, or a mask
// that's used by multiple blends where it can be removed from them all.
VPValue *Mask = Blend->getMask(I);
if (Mask->getNumUsers() == 1 && !match(Mask, m_False())) {
StartIndex = I;
break;
}
}
SmallVector<VPValue *, 4> OperandsWithMask;
OperandsWithMask.push_back(Blend->getIncomingValue(StartIndex));
for (unsigned I = 0; I != Blend->getNumIncomingValues(); ++I) {
if (I == StartIndex)
continue;
OperandsWithMask.push_back(Blend->getIncomingValue(I));
OperandsWithMask.push_back(Blend->getMask(I));
}
auto *NewBlend = new VPBlendRecipe(
cast<PHINode>(Blend->getUnderlyingValue()), OperandsWithMask);
NewBlend->insertBefore(&R);
VPValue *DeadMask = Blend->getMask(StartIndex);
Blend->replaceAllUsesWith(NewBlend);
Blend->eraseFromParent();
recursivelyDeleteDeadRecipes(DeadMask);
/// Simplify BLEND %a, %b, Not(%mask) -> BLEND %b, %a, %mask.
VPValue *NewMask;
if (NewBlend->getNumOperands() == 3 &&
match(NewBlend->getMask(1), m_Not(m_VPValue(NewMask)))) {
VPValue *Inc0 = NewBlend->getOperand(0);
VPValue *Inc1 = NewBlend->getOperand(1);
VPValue *OldMask = NewBlend->getOperand(2);
NewBlend->setOperand(0, Inc1);
NewBlend->setOperand(1, Inc0);
NewBlend->setOperand(2, NewMask);
if (OldMask->getNumUsers() == 0)
cast<VPInstruction>(OldMask)->eraseFromParent();
}
return;
}
// VPScalarIVSteps can only be simplified after unrolling. VPScalarIVSteps for
// part 0 can be replaced by their start value, if only the first lane is
// demanded.
if (auto *Steps = dyn_cast<VPScalarIVStepsRecipe>(&R)) {