[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / Transforms / Vectorize / VPlanTransforms.cpp
blob8e6b48cdb2c8af3d9aa48242089d7337745121a2
1 //===-- VPlanTransforms.cpp - Utility VPlan to VPlan transforms -----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 ///
9 /// \file
10 /// This file implements a set of utility VPlan to VPlan transformations.
11 ///
12 //===----------------------------------------------------------------------===//
14 #include "VPlanTransforms.h"
15 #include "VPRecipeBuilder.h"
16 #include "VPlanAnalysis.h"
17 #include "VPlanCFG.h"
18 #include "VPlanDominatorTree.h"
19 #include "llvm/ADT/PostOrderIterator.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/Analysis/IVDescriptors.h"
23 #include "llvm/Analysis/VectorUtils.h"
24 #include "llvm/IR/Intrinsics.h"
25 #include "llvm/IR/PatternMatch.h"
27 using namespace llvm;
29 using namespace llvm::PatternMatch;
31 void VPlanTransforms::VPInstructionsToVPRecipes(
32 VPlanPtr &Plan,
33 function_ref<const InductionDescriptor *(PHINode *)>
34 GetIntOrFpInductionDescriptor,
35 ScalarEvolution &SE, const TargetLibraryInfo &TLI) {
37 ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
38 Plan->getEntry());
39 for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
40 VPRecipeBase *Term = VPBB->getTerminator();
41 auto EndIter = Term ? Term->getIterator() : VPBB->end();
42 // Introduce each ingredient into VPlan.
43 for (VPRecipeBase &Ingredient :
44 make_early_inc_range(make_range(VPBB->begin(), EndIter))) {
46 VPValue *VPV = Ingredient.getVPSingleValue();
47 Instruction *Inst = cast<Instruction>(VPV->getUnderlyingValue());
49 VPRecipeBase *NewRecipe = nullptr;
50 if (auto *VPPhi = dyn_cast<VPWidenPHIRecipe>(&Ingredient)) {
51 auto *Phi = cast<PHINode>(VPPhi->getUnderlyingValue());
52 if (const auto *II = GetIntOrFpInductionDescriptor(Phi)) {
53 VPValue *Start = Plan->getVPValueOrAddLiveIn(II->getStartValue());
54 VPValue *Step =
55 vputils::getOrCreateVPValueForSCEVExpr(*Plan, II->getStep(), SE);
56 NewRecipe = new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, *II);
57 } else {
58 Plan->addVPValue(Phi, VPPhi);
59 continue;
61 } else {
62 assert(isa<VPInstruction>(&Ingredient) &&
63 "only VPInstructions expected here");
64 assert(!isa<PHINode>(Inst) && "phis should be handled above");
65 // Create VPWidenMemoryInstructionRecipe for loads and stores.
66 if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
67 NewRecipe = new VPWidenMemoryInstructionRecipe(
68 *Load, Ingredient.getOperand(0), nullptr /*Mask*/,
69 false /*Consecutive*/, false /*Reverse*/);
70 } else if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) {
71 NewRecipe = new VPWidenMemoryInstructionRecipe(
72 *Store, Ingredient.getOperand(1), Ingredient.getOperand(0),
73 nullptr /*Mask*/, false /*Consecutive*/, false /*Reverse*/);
74 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
75 NewRecipe = new VPWidenGEPRecipe(GEP, Ingredient.operands());
76 } else if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
77 NewRecipe = new VPWidenCallRecipe(
78 *CI, drop_end(Ingredient.operands()),
79 getVectorIntrinsicIDForCall(CI, &TLI), CI->getDebugLoc());
80 } else if (SelectInst *SI = dyn_cast<SelectInst>(Inst)) {
81 NewRecipe = new VPWidenSelectRecipe(*SI, Ingredient.operands());
82 } else if (auto *CI = dyn_cast<CastInst>(Inst)) {
83 NewRecipe = new VPWidenCastRecipe(
84 CI->getOpcode(), Ingredient.getOperand(0), CI->getType(), *CI);
85 } else {
86 NewRecipe = new VPWidenRecipe(*Inst, Ingredient.operands());
90 NewRecipe->insertBefore(&Ingredient);
91 if (NewRecipe->getNumDefinedValues() == 1)
92 VPV->replaceAllUsesWith(NewRecipe->getVPSingleValue());
93 else
94 assert(NewRecipe->getNumDefinedValues() == 0 &&
95 "Only recpies with zero or one defined values expected");
96 Ingredient.eraseFromParent();
101 static bool sinkScalarOperands(VPlan &Plan) {
102 auto Iter = vp_depth_first_deep(Plan.getEntry());
103 bool Changed = false;
104 // First, collect the operands of all recipes in replicate blocks as seeds for
105 // sinking.
106 SetVector<std::pair<VPBasicBlock *, VPSingleDefRecipe *>> WorkList;
107 for (VPRegionBlock *VPR : VPBlockUtils::blocksOnly<VPRegionBlock>(Iter)) {
108 VPBasicBlock *EntryVPBB = VPR->getEntryBasicBlock();
109 if (!VPR->isReplicator() || EntryVPBB->getSuccessors().size() != 2)
110 continue;
111 VPBasicBlock *VPBB = dyn_cast<VPBasicBlock>(EntryVPBB->getSuccessors()[0]);
112 if (!VPBB || VPBB->getSingleSuccessor() != VPR->getExitingBasicBlock())
113 continue;
114 for (auto &Recipe : *VPBB) {
115 for (VPValue *Op : Recipe.operands())
116 if (auto *Def =
117 dyn_cast_or_null<VPSingleDefRecipe>(Op->getDefiningRecipe()))
118 WorkList.insert(std::make_pair(VPBB, Def));
122 bool ScalarVFOnly = Plan.hasScalarVFOnly();
123 // Try to sink each replicate or scalar IV steps recipe in the worklist.
124 for (unsigned I = 0; I != WorkList.size(); ++I) {
125 VPBasicBlock *SinkTo;
126 VPSingleDefRecipe *SinkCandidate;
127 std::tie(SinkTo, SinkCandidate) = WorkList[I];
128 if (SinkCandidate->getParent() == SinkTo ||
129 SinkCandidate->mayHaveSideEffects() ||
130 SinkCandidate->mayReadOrWriteMemory())
131 continue;
132 if (auto *RepR = dyn_cast<VPReplicateRecipe>(SinkCandidate)) {
133 if (!ScalarVFOnly && RepR->isUniform())
134 continue;
135 } else if (!isa<VPScalarIVStepsRecipe>(SinkCandidate))
136 continue;
138 bool NeedsDuplicating = false;
139 // All recipe users of the sink candidate must be in the same block SinkTo
140 // or all users outside of SinkTo must be uniform-after-vectorization (
141 // i.e., only first lane is used) . In the latter case, we need to duplicate
142 // SinkCandidate.
143 auto CanSinkWithUser = [SinkTo, &NeedsDuplicating,
144 SinkCandidate](VPUser *U) {
145 auto *UI = dyn_cast<VPRecipeBase>(U);
146 if (!UI)
147 return false;
148 if (UI->getParent() == SinkTo)
149 return true;
150 NeedsDuplicating = UI->onlyFirstLaneUsed(SinkCandidate);
151 // We only know how to duplicate VPRecipeRecipes for now.
152 return NeedsDuplicating && isa<VPReplicateRecipe>(SinkCandidate);
154 if (!all_of(SinkCandidate->users(), CanSinkWithUser))
155 continue;
157 if (NeedsDuplicating) {
158 if (ScalarVFOnly)
159 continue;
160 Instruction *I = cast<Instruction>(
161 cast<VPReplicateRecipe>(SinkCandidate)->getUnderlyingValue());
162 auto *Clone = new VPReplicateRecipe(I, SinkCandidate->operands(), true);
163 // TODO: add ".cloned" suffix to name of Clone's VPValue.
165 Clone->insertBefore(SinkCandidate);
166 SinkCandidate->replaceUsesWithIf(Clone, [SinkTo](VPUser &U, unsigned) {
167 return cast<VPRecipeBase>(&U)->getParent() != SinkTo;
170 SinkCandidate->moveBefore(*SinkTo, SinkTo->getFirstNonPhi());
171 for (VPValue *Op : SinkCandidate->operands())
172 if (auto *Def =
173 dyn_cast_or_null<VPSingleDefRecipe>(Op->getDefiningRecipe()))
174 WorkList.insert(std::make_pair(SinkTo, Def));
175 Changed = true;
177 return Changed;
180 /// If \p R is a region with a VPBranchOnMaskRecipe in the entry block, return
181 /// the mask.
182 VPValue *getPredicatedMask(VPRegionBlock *R) {
183 auto *EntryBB = dyn_cast<VPBasicBlock>(R->getEntry());
184 if (!EntryBB || EntryBB->size() != 1 ||
185 !isa<VPBranchOnMaskRecipe>(EntryBB->begin()))
186 return nullptr;
188 return cast<VPBranchOnMaskRecipe>(&*EntryBB->begin())->getOperand(0);
191 /// If \p R is a triangle region, return the 'then' block of the triangle.
192 static VPBasicBlock *getPredicatedThenBlock(VPRegionBlock *R) {
193 auto *EntryBB = cast<VPBasicBlock>(R->getEntry());
194 if (EntryBB->getNumSuccessors() != 2)
195 return nullptr;
197 auto *Succ0 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[0]);
198 auto *Succ1 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[1]);
199 if (!Succ0 || !Succ1)
200 return nullptr;
202 if (Succ0->getNumSuccessors() + Succ1->getNumSuccessors() != 1)
203 return nullptr;
204 if (Succ0->getSingleSuccessor() == Succ1)
205 return Succ0;
206 if (Succ1->getSingleSuccessor() == Succ0)
207 return Succ1;
208 return nullptr;
211 // Merge replicate regions in their successor region, if a replicate region
212 // is connected to a successor replicate region with the same predicate by a
213 // single, empty VPBasicBlock.
214 static bool mergeReplicateRegionsIntoSuccessors(VPlan &Plan) {
215 SetVector<VPRegionBlock *> DeletedRegions;
217 // Collect replicate regions followed by an empty block, followed by another
218 // replicate region with matching masks to process front. This is to avoid
219 // iterator invalidation issues while merging regions.
220 SmallVector<VPRegionBlock *, 8> WorkList;
221 for (VPRegionBlock *Region1 : VPBlockUtils::blocksOnly<VPRegionBlock>(
222 vp_depth_first_deep(Plan.getEntry()))) {
223 if (!Region1->isReplicator())
224 continue;
225 auto *MiddleBasicBlock =
226 dyn_cast_or_null<VPBasicBlock>(Region1->getSingleSuccessor());
227 if (!MiddleBasicBlock || !MiddleBasicBlock->empty())
228 continue;
230 auto *Region2 =
231 dyn_cast_or_null<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor());
232 if (!Region2 || !Region2->isReplicator())
233 continue;
235 VPValue *Mask1 = getPredicatedMask(Region1);
236 VPValue *Mask2 = getPredicatedMask(Region2);
237 if (!Mask1 || Mask1 != Mask2)
238 continue;
240 assert(Mask1 && Mask2 && "both region must have conditions");
241 WorkList.push_back(Region1);
244 // Move recipes from Region1 to its successor region, if both are triangles.
245 for (VPRegionBlock *Region1 : WorkList) {
246 if (DeletedRegions.contains(Region1))
247 continue;
248 auto *MiddleBasicBlock = cast<VPBasicBlock>(Region1->getSingleSuccessor());
249 auto *Region2 = cast<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor());
251 VPBasicBlock *Then1 = getPredicatedThenBlock(Region1);
252 VPBasicBlock *Then2 = getPredicatedThenBlock(Region2);
253 if (!Then1 || !Then2)
254 continue;
256 // Note: No fusion-preventing memory dependencies are expected in either
257 // region. Such dependencies should be rejected during earlier dependence
258 // checks, which guarantee accesses can be re-ordered for vectorization.
260 // Move recipes to the successor region.
261 for (VPRecipeBase &ToMove : make_early_inc_range(reverse(*Then1)))
262 ToMove.moveBefore(*Then2, Then2->getFirstNonPhi());
264 auto *Merge1 = cast<VPBasicBlock>(Then1->getSingleSuccessor());
265 auto *Merge2 = cast<VPBasicBlock>(Then2->getSingleSuccessor());
267 // Move VPPredInstPHIRecipes from the merge block to the successor region's
268 // merge block. Update all users inside the successor region to use the
269 // original values.
270 for (VPRecipeBase &Phi1ToMove : make_early_inc_range(reverse(*Merge1))) {
271 VPValue *PredInst1 =
272 cast<VPPredInstPHIRecipe>(&Phi1ToMove)->getOperand(0);
273 VPValue *Phi1ToMoveV = Phi1ToMove.getVPSingleValue();
274 Phi1ToMoveV->replaceUsesWithIf(PredInst1, [Then2](VPUser &U, unsigned) {
275 auto *UI = dyn_cast<VPRecipeBase>(&U);
276 return UI && UI->getParent() == Then2;
279 Phi1ToMove.moveBefore(*Merge2, Merge2->begin());
282 // Finally, remove the first region.
283 for (VPBlockBase *Pred : make_early_inc_range(Region1->getPredecessors())) {
284 VPBlockUtils::disconnectBlocks(Pred, Region1);
285 VPBlockUtils::connectBlocks(Pred, MiddleBasicBlock);
287 VPBlockUtils::disconnectBlocks(Region1, MiddleBasicBlock);
288 DeletedRegions.insert(Region1);
291 for (VPRegionBlock *ToDelete : DeletedRegions)
292 delete ToDelete;
293 return !DeletedRegions.empty();
296 static VPRegionBlock *createReplicateRegion(VPReplicateRecipe *PredRecipe,
297 VPlan &Plan) {
298 Instruction *Instr = PredRecipe->getUnderlyingInstr();
299 // Build the triangular if-then region.
300 std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str();
301 assert(Instr->getParent() && "Predicated instruction not in any basic block");
302 auto *BlockInMask = PredRecipe->getMask();
303 auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask);
304 auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe);
306 // Replace predicated replicate recipe with a replicate recipe without a
307 // mask but in the replicate region.
308 auto *RecipeWithoutMask = new VPReplicateRecipe(
309 PredRecipe->getUnderlyingInstr(),
310 make_range(PredRecipe->op_begin(), std::prev(PredRecipe->op_end())),
311 PredRecipe->isUniform());
312 auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", RecipeWithoutMask);
314 VPPredInstPHIRecipe *PHIRecipe = nullptr;
315 if (PredRecipe->getNumUsers() != 0) {
316 PHIRecipe = new VPPredInstPHIRecipe(RecipeWithoutMask);
317 PredRecipe->replaceAllUsesWith(PHIRecipe);
318 PHIRecipe->setOperand(0, RecipeWithoutMask);
320 PredRecipe->eraseFromParent();
321 auto *Exiting = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe);
322 VPRegionBlock *Region = new VPRegionBlock(Entry, Exiting, RegionName, true);
324 // Note: first set Entry as region entry and then connect successors starting
325 // from it in order, to propagate the "parent" of each VPBasicBlock.
326 VPBlockUtils::insertTwoBlocksAfter(Pred, Exiting, Entry);
327 VPBlockUtils::connectBlocks(Pred, Exiting);
329 return Region;
332 static void addReplicateRegions(VPlan &Plan) {
333 SmallVector<VPReplicateRecipe *> WorkList;
334 for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
335 vp_depth_first_deep(Plan.getEntry()))) {
336 for (VPRecipeBase &R : *VPBB)
337 if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) {
338 if (RepR->isPredicated())
339 WorkList.push_back(RepR);
343 unsigned BBNum = 0;
344 for (VPReplicateRecipe *RepR : WorkList) {
345 VPBasicBlock *CurrentBlock = RepR->getParent();
346 VPBasicBlock *SplitBlock = CurrentBlock->splitAt(RepR->getIterator());
348 BasicBlock *OrigBB = RepR->getUnderlyingInstr()->getParent();
349 SplitBlock->setName(
350 OrigBB->hasName() ? OrigBB->getName() + "." + Twine(BBNum++) : "");
351 // Record predicated instructions for above packing optimizations.
352 VPBlockBase *Region = createReplicateRegion(RepR, Plan);
353 Region->setParent(CurrentBlock->getParent());
354 VPBlockUtils::disconnectBlocks(CurrentBlock, SplitBlock);
355 VPBlockUtils::connectBlocks(CurrentBlock, Region);
356 VPBlockUtils::connectBlocks(Region, SplitBlock);
360 void VPlanTransforms::createAndOptimizeReplicateRegions(VPlan &Plan) {
361 // Convert masked VPReplicateRecipes to if-then region blocks.
362 addReplicateRegions(Plan);
364 bool ShouldSimplify = true;
365 while (ShouldSimplify) {
366 ShouldSimplify = sinkScalarOperands(Plan);
367 ShouldSimplify |= mergeReplicateRegionsIntoSuccessors(Plan);
368 ShouldSimplify |= VPlanTransforms::mergeBlocksIntoPredecessors(Plan);
371 bool VPlanTransforms::mergeBlocksIntoPredecessors(VPlan &Plan) {
372 SmallVector<VPBasicBlock *> WorkList;
373 for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
374 vp_depth_first_deep(Plan.getEntry()))) {
375 auto *PredVPBB =
376 dyn_cast_or_null<VPBasicBlock>(VPBB->getSinglePredecessor());
377 if (PredVPBB && PredVPBB->getNumSuccessors() == 1)
378 WorkList.push_back(VPBB);
381 for (VPBasicBlock *VPBB : WorkList) {
382 VPBasicBlock *PredVPBB = cast<VPBasicBlock>(VPBB->getSinglePredecessor());
383 for (VPRecipeBase &R : make_early_inc_range(*VPBB))
384 R.moveBefore(*PredVPBB, PredVPBB->end());
385 VPBlockUtils::disconnectBlocks(PredVPBB, VPBB);
386 auto *ParentRegion = cast_or_null<VPRegionBlock>(VPBB->getParent());
387 if (ParentRegion && ParentRegion->getExiting() == VPBB)
388 ParentRegion->setExiting(PredVPBB);
389 for (auto *Succ : to_vector(VPBB->successors())) {
390 VPBlockUtils::disconnectBlocks(VPBB, Succ);
391 VPBlockUtils::connectBlocks(PredVPBB, Succ);
393 delete VPBB;
395 return !WorkList.empty();
398 void VPlanTransforms::removeRedundantInductionCasts(VPlan &Plan) {
399 for (auto &Phi : Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) {
400 auto *IV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
401 if (!IV || IV->getTruncInst())
402 continue;
404 // A sequence of IR Casts has potentially been recorded for IV, which
405 // *must be bypassed* when the IV is vectorized, because the vectorized IV
406 // will produce the desired casted value. This sequence forms a def-use
407 // chain and is provided in reverse order, ending with the cast that uses
408 // the IV phi. Search for the recipe of the last cast in the chain and
409 // replace it with the original IV. Note that only the final cast is
410 // expected to have users outside the cast-chain and the dead casts left
411 // over will be cleaned up later.
412 auto &Casts = IV->getInductionDescriptor().getCastInsts();
413 VPValue *FindMyCast = IV;
414 for (Instruction *IRCast : reverse(Casts)) {
415 VPSingleDefRecipe *FoundUserCast = nullptr;
416 for (auto *U : FindMyCast->users()) {
417 auto *UserCast = dyn_cast<VPSingleDefRecipe>(U);
418 if (UserCast && UserCast->getUnderlyingValue() == IRCast) {
419 FoundUserCast = UserCast;
420 break;
423 FindMyCast = FoundUserCast;
425 FindMyCast->replaceAllUsesWith(IV);
429 void VPlanTransforms::removeRedundantCanonicalIVs(VPlan &Plan) {
430 VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
431 VPWidenCanonicalIVRecipe *WidenNewIV = nullptr;
432 for (VPUser *U : CanonicalIV->users()) {
433 WidenNewIV = dyn_cast<VPWidenCanonicalIVRecipe>(U);
434 if (WidenNewIV)
435 break;
438 if (!WidenNewIV)
439 return;
441 VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
442 for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
443 auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
445 if (!WidenOriginalIV || !WidenOriginalIV->isCanonical() ||
446 WidenOriginalIV->getScalarType() != WidenNewIV->getScalarType())
447 continue;
449 // Replace WidenNewIV with WidenOriginalIV if WidenOriginalIV provides
450 // everything WidenNewIV's users need. That is, WidenOriginalIV will
451 // generate a vector phi or all users of WidenNewIV demand the first lane
452 // only.
453 if (any_of(WidenOriginalIV->users(),
454 [WidenOriginalIV](VPUser *U) {
455 return !U->usesScalars(WidenOriginalIV);
456 }) ||
457 vputils::onlyFirstLaneUsed(WidenNewIV)) {
458 WidenNewIV->replaceAllUsesWith(WidenOriginalIV);
459 WidenNewIV->eraseFromParent();
460 return;
465 void VPlanTransforms::removeDeadRecipes(VPlan &Plan) {
466 ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
467 Plan.getEntry());
469 for (VPBasicBlock *VPBB : reverse(VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT))) {
470 // The recipes in the block are processed in reverse order, to catch chains
471 // of dead recipes.
472 for (VPRecipeBase &R : make_early_inc_range(reverse(*VPBB))) {
473 // A user keeps R alive:
474 if (any_of(R.definedValues(),
475 [](VPValue *V) { return V->getNumUsers(); }))
476 continue;
478 // Having side effects keeps R alive, but do remove conditional assume
479 // instructions as their conditions may be flattened.
480 auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
481 bool IsConditionalAssume =
482 RepR && RepR->isPredicated() &&
483 match(RepR->getUnderlyingInstr(), m_Intrinsic<Intrinsic::assume>());
484 if (R.mayHaveSideEffects() && !IsConditionalAssume)
485 continue;
487 R.eraseFromParent();
492 static VPValue *createScalarIVSteps(VPlan &Plan, const InductionDescriptor &ID,
493 ScalarEvolution &SE, Instruction *TruncI,
494 Type *IVTy, VPValue *StartV,
495 VPValue *Step) {
496 VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
497 auto IP = HeaderVPBB->getFirstNonPhi();
498 VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
499 Type *TruncTy = TruncI ? TruncI->getType() : IVTy;
500 VPValue *BaseIV = CanonicalIV;
501 if (!CanonicalIV->isCanonical(ID.getKind(), StartV, Step, TruncTy)) {
502 BaseIV = new VPDerivedIVRecipe(ID, StartV, CanonicalIV, Step,
503 TruncI ? TruncI->getType() : nullptr);
504 HeaderVPBB->insert(BaseIV->getDefiningRecipe(), IP);
507 VPScalarIVStepsRecipe *Steps = new VPScalarIVStepsRecipe(ID, BaseIV, Step);
508 HeaderVPBB->insert(Steps, IP);
509 return Steps;
512 void VPlanTransforms::optimizeInductions(VPlan &Plan, ScalarEvolution &SE) {
513 SmallVector<VPRecipeBase *> ToRemove;
514 VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
515 bool HasOnlyVectorVFs = !Plan.hasVF(ElementCount::getFixed(1));
516 for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
517 auto *WideIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
518 if (!WideIV)
519 continue;
520 if (HasOnlyVectorVFs && none_of(WideIV->users(), [WideIV](VPUser *U) {
521 return U->usesScalars(WideIV);
523 continue;
525 const InductionDescriptor &ID = WideIV->getInductionDescriptor();
526 VPValue *Steps = createScalarIVSteps(
527 Plan, ID, SE, WideIV->getTruncInst(), WideIV->getPHINode()->getType(),
528 WideIV->getStartValue(), WideIV->getStepValue());
530 // Update scalar users of IV to use Step instead.
531 if (!HasOnlyVectorVFs)
532 WideIV->replaceAllUsesWith(Steps);
533 else
534 WideIV->replaceUsesWithIf(Steps, [WideIV](VPUser &U, unsigned) {
535 return U.usesScalars(WideIV);
540 void VPlanTransforms::removeRedundantExpandSCEVRecipes(VPlan &Plan) {
541 DenseMap<const SCEV *, VPValue *> SCEV2VPV;
543 for (VPRecipeBase &R :
544 make_early_inc_range(*Plan.getEntry()->getEntryBasicBlock())) {
545 auto *ExpR = dyn_cast<VPExpandSCEVRecipe>(&R);
546 if (!ExpR)
547 continue;
549 auto I = SCEV2VPV.insert({ExpR->getSCEV(), ExpR});
550 if (I.second)
551 continue;
552 ExpR->replaceAllUsesWith(I.first->second);
553 ExpR->eraseFromParent();
557 static bool canSimplifyBranchOnCond(VPInstruction *Term) {
558 VPInstruction *Not = dyn_cast<VPInstruction>(Term->getOperand(0));
559 if (!Not || Not->getOpcode() != VPInstruction::Not)
560 return false;
562 VPInstruction *ALM = dyn_cast<VPInstruction>(Not->getOperand(0));
563 return ALM && ALM->getOpcode() == VPInstruction::ActiveLaneMask;
566 void VPlanTransforms::optimizeForVFAndUF(VPlan &Plan, ElementCount BestVF,
567 unsigned BestUF,
568 PredicatedScalarEvolution &PSE) {
569 assert(Plan.hasVF(BestVF) && "BestVF is not available in Plan");
570 assert(Plan.hasUF(BestUF) && "BestUF is not available in Plan");
571 VPBasicBlock *ExitingVPBB =
572 Plan.getVectorLoopRegion()->getExitingBasicBlock();
573 auto *Term = dyn_cast<VPInstruction>(&ExitingVPBB->back());
574 // Try to simplify the branch condition if TC <= VF * UF when preparing to
575 // execute the plan for the main vector loop. We only do this if the
576 // terminator is:
577 // 1. BranchOnCount, or
578 // 2. BranchOnCond where the input is Not(ActiveLaneMask).
579 if (!Term || (Term->getOpcode() != VPInstruction::BranchOnCount &&
580 (Term->getOpcode() != VPInstruction::BranchOnCond ||
581 !canSimplifyBranchOnCond(Term))))
582 return;
584 Type *IdxTy =
585 Plan.getCanonicalIV()->getStartValue()->getLiveInIRValue()->getType();
586 const SCEV *TripCount = createTripCountSCEV(IdxTy, PSE);
587 ScalarEvolution &SE = *PSE.getSE();
588 const SCEV *C =
589 SE.getConstant(TripCount->getType(), BestVF.getKnownMinValue() * BestUF);
590 if (TripCount->isZero() ||
591 !SE.isKnownPredicate(CmpInst::ICMP_ULE, TripCount, C))
592 return;
594 LLVMContext &Ctx = SE.getContext();
595 auto *BOC = new VPInstruction(
596 VPInstruction::BranchOnCond,
597 {Plan.getVPValueOrAddLiveIn(ConstantInt::getTrue(Ctx))});
598 Term->eraseFromParent();
599 ExitingVPBB->appendRecipe(BOC);
600 Plan.setVF(BestVF);
601 Plan.setUF(BestUF);
602 // TODO: Further simplifications are possible
603 // 1. Replace inductions with constants.
604 // 2. Replace vector loop region with VPBasicBlock.
607 #ifndef NDEBUG
608 static VPRegionBlock *GetReplicateRegion(VPRecipeBase *R) {
609 auto *Region = dyn_cast_or_null<VPRegionBlock>(R->getParent()->getParent());
610 if (Region && Region->isReplicator()) {
611 assert(Region->getNumSuccessors() == 1 &&
612 Region->getNumPredecessors() == 1 && "Expected SESE region!");
613 assert(R->getParent()->size() == 1 &&
614 "A recipe in an original replicator region must be the only "
615 "recipe in its block");
616 return Region;
618 return nullptr;
620 #endif
622 static bool properlyDominates(const VPRecipeBase *A, const VPRecipeBase *B,
623 VPDominatorTree &VPDT) {
624 if (A == B)
625 return false;
627 auto LocalComesBefore = [](const VPRecipeBase *A, const VPRecipeBase *B) {
628 for (auto &R : *A->getParent()) {
629 if (&R == A)
630 return true;
631 if (&R == B)
632 return false;
634 llvm_unreachable("recipe not found");
636 const VPBlockBase *ParentA = A->getParent();
637 const VPBlockBase *ParentB = B->getParent();
638 if (ParentA == ParentB)
639 return LocalComesBefore(A, B);
641 assert(!GetReplicateRegion(const_cast<VPRecipeBase *>(A)) &&
642 "No replicate regions expected at this point");
643 assert(!GetReplicateRegion(const_cast<VPRecipeBase *>(B)) &&
644 "No replicate regions expected at this point");
645 return VPDT.properlyDominates(ParentA, ParentB);
648 /// Sink users of \p FOR after the recipe defining the previous value \p
649 /// Previous of the recurrence. \returns true if all users of \p FOR could be
650 /// re-arranged as needed or false if it is not possible.
651 static bool
652 sinkRecurrenceUsersAfterPrevious(VPFirstOrderRecurrencePHIRecipe *FOR,
653 VPRecipeBase *Previous,
654 VPDominatorTree &VPDT) {
655 // Collect recipes that need sinking.
656 SmallVector<VPRecipeBase *> WorkList;
657 SmallPtrSet<VPRecipeBase *, 8> Seen;
658 Seen.insert(Previous);
659 auto TryToPushSinkCandidate = [&](VPRecipeBase *SinkCandidate) {
660 // The previous value must not depend on the users of the recurrence phi. In
661 // that case, FOR is not a fixed order recurrence.
662 if (SinkCandidate == Previous)
663 return false;
665 if (isa<VPHeaderPHIRecipe>(SinkCandidate) ||
666 !Seen.insert(SinkCandidate).second ||
667 properlyDominates(Previous, SinkCandidate, VPDT))
668 return true;
670 if (SinkCandidate->mayHaveSideEffects())
671 return false;
673 WorkList.push_back(SinkCandidate);
674 return true;
677 // Recursively sink users of FOR after Previous.
678 WorkList.push_back(FOR);
679 for (unsigned I = 0; I != WorkList.size(); ++I) {
680 VPRecipeBase *Current = WorkList[I];
681 assert(Current->getNumDefinedValues() == 1 &&
682 "only recipes with a single defined value expected");
684 for (VPUser *User : Current->getVPSingleValue()->users()) {
685 if (auto *R = dyn_cast<VPRecipeBase>(User))
686 if (!TryToPushSinkCandidate(R))
687 return false;
691 // Keep recipes to sink ordered by dominance so earlier instructions are
692 // processed first.
693 sort(WorkList, [&VPDT](const VPRecipeBase *A, const VPRecipeBase *B) {
694 return properlyDominates(A, B, VPDT);
697 for (VPRecipeBase *SinkCandidate : WorkList) {
698 if (SinkCandidate == FOR)
699 continue;
701 SinkCandidate->moveAfter(Previous);
702 Previous = SinkCandidate;
704 return true;
707 bool VPlanTransforms::adjustFixedOrderRecurrences(VPlan &Plan,
708 VPBuilder &Builder) {
709 VPDominatorTree VPDT;
710 VPDT.recalculate(Plan);
712 SmallVector<VPFirstOrderRecurrencePHIRecipe *> RecurrencePhis;
713 for (VPRecipeBase &R :
714 Plan.getVectorLoopRegion()->getEntry()->getEntryBasicBlock()->phis())
715 if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R))
716 RecurrencePhis.push_back(FOR);
718 for (VPFirstOrderRecurrencePHIRecipe *FOR : RecurrencePhis) {
719 SmallPtrSet<VPFirstOrderRecurrencePHIRecipe *, 4> SeenPhis;
720 VPRecipeBase *Previous = FOR->getBackedgeValue()->getDefiningRecipe();
721 // Fixed-order recurrences do not contain cycles, so this loop is guaranteed
722 // to terminate.
723 while (auto *PrevPhi =
724 dyn_cast_or_null<VPFirstOrderRecurrencePHIRecipe>(Previous)) {
725 assert(PrevPhi->getParent() == FOR->getParent());
726 assert(SeenPhis.insert(PrevPhi).second);
727 Previous = PrevPhi->getBackedgeValue()->getDefiningRecipe();
730 if (!sinkRecurrenceUsersAfterPrevious(FOR, Previous, VPDT))
731 return false;
733 // Introduce a recipe to combine the incoming and previous values of a
734 // fixed-order recurrence.
735 VPBasicBlock *InsertBlock = Previous->getParent();
736 if (isa<VPHeaderPHIRecipe>(Previous))
737 Builder.setInsertPoint(InsertBlock, InsertBlock->getFirstNonPhi());
738 else
739 Builder.setInsertPoint(InsertBlock, std::next(Previous->getIterator()));
741 auto *RecurSplice = cast<VPInstruction>(
742 Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice,
743 {FOR, FOR->getBackedgeValue()}));
745 FOR->replaceAllUsesWith(RecurSplice);
746 // Set the first operand of RecurSplice to FOR again, after replacing
747 // all users.
748 RecurSplice->setOperand(0, FOR);
750 return true;
753 void VPlanTransforms::clearReductionWrapFlags(VPlan &Plan) {
754 for (VPRecipeBase &R :
755 Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) {
756 auto *PhiR = dyn_cast<VPReductionPHIRecipe>(&R);
757 if (!PhiR)
758 continue;
759 const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
760 RecurKind RK = RdxDesc.getRecurrenceKind();
761 if (RK != RecurKind::Add && RK != RecurKind::Mul)
762 continue;
764 SmallSetVector<VPValue *, 8> Worklist;
765 Worklist.insert(PhiR);
767 for (unsigned I = 0; I != Worklist.size(); ++I) {
768 VPValue *Cur = Worklist[I];
769 if (auto *RecWithFlags =
770 dyn_cast<VPRecipeWithIRFlags>(Cur->getDefiningRecipe())) {
771 RecWithFlags->dropPoisonGeneratingFlags();
774 for (VPUser *U : Cur->users()) {
775 auto *UserRecipe = dyn_cast<VPRecipeBase>(U);
776 if (!UserRecipe)
777 continue;
778 for (VPValue *V : UserRecipe->definedValues())
779 Worklist.insert(V);
785 /// Returns true is \p V is constant one.
786 static bool isConstantOne(VPValue *V) {
787 if (!V->isLiveIn())
788 return false;
789 auto *C = dyn_cast<ConstantInt>(V->getLiveInIRValue());
790 return C && C->isOne();
793 /// Returns the llvm::Instruction opcode for \p R.
794 static unsigned getOpcodeForRecipe(VPRecipeBase &R) {
795 if (auto *WidenR = dyn_cast<VPWidenRecipe>(&R))
796 return WidenR->getUnderlyingInstr()->getOpcode();
797 if (auto *WidenC = dyn_cast<VPWidenCastRecipe>(&R))
798 return WidenC->getOpcode();
799 if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R))
800 return RepR->getUnderlyingInstr()->getOpcode();
801 if (auto *VPI = dyn_cast<VPInstruction>(&R))
802 return VPI->getOpcode();
803 return 0;
806 /// Try to simplify recipe \p R.
807 static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
808 switch (getOpcodeForRecipe(R)) {
809 case Instruction::Mul: {
810 VPValue *A = R.getOperand(0);
811 VPValue *B = R.getOperand(1);
812 if (isConstantOne(A))
813 return R.getVPSingleValue()->replaceAllUsesWith(B);
814 if (isConstantOne(B))
815 return R.getVPSingleValue()->replaceAllUsesWith(A);
816 break;
818 case Instruction::Trunc: {
819 VPRecipeBase *Ext = R.getOperand(0)->getDefiningRecipe();
820 if (!Ext)
821 break;
822 unsigned ExtOpcode = getOpcodeForRecipe(*Ext);
823 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
824 break;
825 VPValue *A = Ext->getOperand(0);
826 VPValue *Trunc = R.getVPSingleValue();
827 Type *TruncTy = TypeInfo.inferScalarType(Trunc);
828 Type *ATy = TypeInfo.inferScalarType(A);
829 if (TruncTy == ATy) {
830 Trunc->replaceAllUsesWith(A);
831 } else {
832 // Don't replace a scalarizing recipe with a widened cast.
833 if (isa<VPReplicateRecipe>(&R))
834 break;
835 if (ATy->getScalarSizeInBits() < TruncTy->getScalarSizeInBits()) {
836 auto *VPC =
837 new VPWidenCastRecipe(Instruction::CastOps(ExtOpcode), A, TruncTy);
838 VPC->insertBefore(&R);
839 Trunc->replaceAllUsesWith(VPC);
840 } else if (ATy->getScalarSizeInBits() > TruncTy->getScalarSizeInBits()) {
841 auto *VPC = new VPWidenCastRecipe(Instruction::Trunc, A, TruncTy);
842 VPC->insertBefore(&R);
843 Trunc->replaceAllUsesWith(VPC);
846 #ifndef NDEBUG
847 // Verify that the cached type info is for both A and its users is still
848 // accurate by comparing it to freshly computed types.
849 VPTypeAnalysis TypeInfo2(TypeInfo.getContext());
850 assert(TypeInfo.inferScalarType(A) == TypeInfo2.inferScalarType(A));
851 for (VPUser *U : A->users()) {
852 auto *R = dyn_cast<VPRecipeBase>(U);
853 if (!R)
854 continue;
855 for (VPValue *VPV : R->definedValues())
856 assert(TypeInfo.inferScalarType(VPV) == TypeInfo2.inferScalarType(VPV));
858 #endif
859 break;
861 default:
862 break;
866 /// Try to simplify the recipes in \p Plan.
867 static void simplifyRecipes(VPlan &Plan, LLVMContext &Ctx) {
868 ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
869 Plan.getEntry());
870 VPTypeAnalysis TypeInfo(Ctx);
871 for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
872 for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
873 simplifyRecipe(R, TypeInfo);
878 void VPlanTransforms::truncateToMinimalBitwidths(
879 VPlan &Plan, const MapVector<Instruction *, uint64_t> &MinBWs,
880 LLVMContext &Ctx) {
881 #ifndef NDEBUG
882 // Count the processed recipes and cross check the count later with MinBWs
883 // size, to make sure all entries in MinBWs have been handled.
884 unsigned NumProcessedRecipes = 0;
885 #endif
886 // Keep track of created truncates, so they can be re-used. Note that we
887 // cannot use RAUW after creating a new truncate, as this would could make
888 // other uses have different types for their operands, making them invalidly
889 // typed.
890 DenseMap<VPValue *, VPWidenCastRecipe *> ProcessedTruncs;
891 VPTypeAnalysis TypeInfo(Ctx);
892 VPBasicBlock *PH = Plan.getEntry();
893 for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
894 vp_depth_first_deep(Plan.getVectorLoopRegion()))) {
895 for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
896 if (!isa<VPWidenRecipe, VPWidenCastRecipe, VPReplicateRecipe,
897 VPWidenSelectRecipe, VPWidenMemoryInstructionRecipe>(&R))
898 continue;
899 if (isa<VPWidenMemoryInstructionRecipe>(&R) &&
900 cast<VPWidenMemoryInstructionRecipe>(&R)->isStore())
901 continue;
903 VPValue *ResultVPV = R.getVPSingleValue();
904 auto *UI = cast_or_null<Instruction>(ResultVPV->getUnderlyingValue());
905 unsigned NewResSizeInBits = MinBWs.lookup(UI);
906 if (!NewResSizeInBits)
907 continue;
909 #ifndef NDEBUG
910 NumProcessedRecipes++;
911 #endif
912 // If the value wasn't vectorized, we must maintain the original scalar
913 // type. Skip those here, after incrementing NumProcessedRecipes. Also
914 // skip casts which do not need to be handled explicitly here, as
915 // redundant casts will be removed during recipe simplification.
916 if (isa<VPReplicateRecipe, VPWidenCastRecipe>(&R)) {
917 #ifndef NDEBUG
918 // If any of the operands is a live-in and not used by VPWidenRecipe or
919 // VPWidenSelectRecipe, but in MinBWs, make sure it is counted as
920 // processed as well. When MinBWs is currently constructed, there is no
921 // information about whether recipes are widened or replicated and in
922 // case they are reciplicated the operands are not truncated. Counting
923 // them them here ensures we do not miss any recipes in MinBWs.
924 // TODO: Remove once the analysis is done on VPlan.
925 for (VPValue *Op : R.operands()) {
926 if (!Op->isLiveIn())
927 continue;
928 auto *UV = dyn_cast_or_null<Instruction>(Op->getUnderlyingValue());
929 if (UV && MinBWs.contains(UV) && !ProcessedTruncs.contains(Op) &&
930 all_of(Op->users(), [](VPUser *U) {
931 return !isa<VPWidenRecipe, VPWidenSelectRecipe>(U);
932 })) {
933 // Add an entry to ProcessedTruncs to avoid counting the same
934 // operand multiple times.
935 ProcessedTruncs[Op] = nullptr;
936 NumProcessedRecipes += 1;
939 #endif
940 continue;
943 Type *OldResTy = TypeInfo.inferScalarType(ResultVPV);
944 unsigned OldResSizeInBits = OldResTy->getScalarSizeInBits();
945 assert(OldResTy->isIntegerTy() && "only integer types supported");
946 if (OldResSizeInBits == NewResSizeInBits)
947 continue;
948 assert(OldResSizeInBits > NewResSizeInBits && "Nothing to shrink?");
949 (void)OldResSizeInBits;
951 auto *NewResTy = IntegerType::get(Ctx, NewResSizeInBits);
953 // Any wrapping introduced by shrinking this operation shouldn't be
954 // considered undefined behavior. So, we can't unconditionally copy
955 // arithmetic wrapping flags to VPW.
956 if (auto *VPW = dyn_cast<VPRecipeWithIRFlags>(&R))
957 VPW->dropPoisonGeneratingFlags();
959 // Extend result to original width.
960 auto *Ext = new VPWidenCastRecipe(Instruction::ZExt, ResultVPV, OldResTy);
961 Ext->insertAfter(&R);
962 ResultVPV->replaceAllUsesWith(Ext);
963 Ext->setOperand(0, ResultVPV);
965 if (isa<VPWidenMemoryInstructionRecipe>(&R)) {
966 assert(!cast<VPWidenMemoryInstructionRecipe>(&R)->isStore() && "stores cannot be narrowed");
967 continue;
970 // Shrink operands by introducing truncates as needed.
971 unsigned StartIdx = isa<VPWidenSelectRecipe>(&R) ? 1 : 0;
972 for (unsigned Idx = StartIdx; Idx != R.getNumOperands(); ++Idx) {
973 auto *Op = R.getOperand(Idx);
974 unsigned OpSizeInBits =
975 TypeInfo.inferScalarType(Op)->getScalarSizeInBits();
976 if (OpSizeInBits == NewResSizeInBits)
977 continue;
978 assert(OpSizeInBits > NewResSizeInBits && "nothing to truncate");
979 auto [ProcessedIter, IterIsEmpty] =
980 ProcessedTruncs.insert({Op, nullptr});
981 VPWidenCastRecipe *NewOp =
982 IterIsEmpty
983 ? new VPWidenCastRecipe(Instruction::Trunc, Op, NewResTy)
984 : ProcessedIter->second;
985 R.setOperand(Idx, NewOp);
986 if (!IterIsEmpty)
987 continue;
988 ProcessedIter->second = NewOp;
989 if (!Op->isLiveIn()) {
990 NewOp->insertBefore(&R);
991 } else {
992 PH->appendRecipe(NewOp);
993 #ifndef NDEBUG
994 auto *OpInst = dyn_cast<Instruction>(Op->getLiveInIRValue());
995 bool IsContained = MinBWs.contains(OpInst);
996 NumProcessedRecipes += IsContained;
997 #endif
1004 assert(MinBWs.size() == NumProcessedRecipes &&
1005 "some entries in MinBWs haven't been processed");
1008 void VPlanTransforms::optimize(VPlan &Plan, ScalarEvolution &SE) {
1009 removeRedundantCanonicalIVs(Plan);
1010 removeRedundantInductionCasts(Plan);
1012 optimizeInductions(Plan, SE);
1013 simplifyRecipes(Plan, SE.getContext());
1014 removeDeadRecipes(Plan);
1016 createAndOptimizeReplicateRegions(Plan);
1018 removeRedundantExpandSCEVRecipes(Plan);
1019 mergeBlocksIntoPredecessors(Plan);
1022 // Add a VPActiveLaneMaskPHIRecipe and related recipes to \p Plan and replace
1023 // the loop terminator with a branch-on-cond recipe with the negated
1024 // active-lane-mask as operand. Note that this turns the loop into an
1025 // uncountable one. Only the existing terminator is replaced, all other existing
1026 // recipes/users remain unchanged, except for poison-generating flags being
1027 // dropped from the canonical IV increment. Return the created
1028 // VPActiveLaneMaskPHIRecipe.
1030 // The function uses the following definitions:
1032 // %TripCount = DataWithControlFlowWithoutRuntimeCheck ?
1033 // calculate-trip-count-minus-VF (original TC) : original TC
1034 // %IncrementValue = DataWithControlFlowWithoutRuntimeCheck ?
1035 // CanonicalIVPhi : CanonicalIVIncrement
1036 // %StartV is the canonical induction start value.
1038 // The function adds the following recipes:
1040 // vector.ph:
1041 // %TripCount = calculate-trip-count-minus-VF (original TC)
1042 // [if DataWithControlFlowWithoutRuntimeCheck]
1043 // %EntryInc = canonical-iv-increment-for-part %StartV
1044 // %EntryALM = active-lane-mask %EntryInc, %TripCount
1046 // vector.body:
1047 // ...
1048 // %P = active-lane-mask-phi [ %EntryALM, %vector.ph ], [ %ALM, %vector.body ]
1049 // ...
1050 // %InLoopInc = canonical-iv-increment-for-part %IncrementValue
1051 // %ALM = active-lane-mask %InLoopInc, TripCount
1052 // %Negated = Not %ALM
1053 // branch-on-cond %Negated
1055 static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
1056 VPlan &Plan, bool DataAndControlFlowWithoutRuntimeCheck) {
1057 VPRegionBlock *TopRegion = Plan.getVectorLoopRegion();
1058 VPBasicBlock *EB = TopRegion->getExitingBasicBlock();
1059 auto *CanonicalIVPHI = Plan.getCanonicalIV();
1060 VPValue *StartV = CanonicalIVPHI->getStartValue();
1062 auto *CanonicalIVIncrement =
1063 cast<VPInstruction>(CanonicalIVPHI->getBackedgeValue());
1064 // TODO: Check if dropping the flags is needed if
1065 // !DataAndControlFlowWithoutRuntimeCheck.
1066 CanonicalIVIncrement->dropPoisonGeneratingFlags();
1067 DebugLoc DL = CanonicalIVIncrement->getDebugLoc();
1068 // We can't use StartV directly in the ActiveLaneMask VPInstruction, since
1069 // we have to take unrolling into account. Each part needs to start at
1070 // Part * VF
1071 auto *VecPreheader = cast<VPBasicBlock>(TopRegion->getSinglePredecessor());
1072 VPBuilder Builder(VecPreheader);
1074 // Create the ActiveLaneMask instruction using the correct start values.
1075 VPValue *TC = Plan.getTripCount();
1077 VPValue *TripCount, *IncrementValue;
1078 if (!DataAndControlFlowWithoutRuntimeCheck) {
1079 // When the loop is guarded by a runtime overflow check for the loop
1080 // induction variable increment by VF, we can increment the value before
1081 // the get.active.lane mask and use the unmodified tripcount.
1082 IncrementValue = CanonicalIVIncrement;
1083 TripCount = TC;
1084 } else {
1085 // When avoiding a runtime check, the active.lane.mask inside the loop
1086 // uses a modified trip count and the induction variable increment is
1087 // done after the active.lane.mask intrinsic is called.
1088 IncrementValue = CanonicalIVPHI;
1089 TripCount = Builder.createNaryOp(VPInstruction::CalculateTripCountMinusVF,
1090 {TC}, DL);
1092 auto *EntryIncrement = Builder.createOverflowingOp(
1093 VPInstruction::CanonicalIVIncrementForPart, {StartV}, {false, false}, DL,
1094 "index.part.next");
1096 // Create the active lane mask instruction in the VPlan preheader.
1097 auto *EntryALM =
1098 Builder.createNaryOp(VPInstruction::ActiveLaneMask, {EntryIncrement, TC},
1099 DL, "active.lane.mask.entry");
1101 // Now create the ActiveLaneMaskPhi recipe in the main loop using the
1102 // preheader ActiveLaneMask instruction.
1103 auto LaneMaskPhi = new VPActiveLaneMaskPHIRecipe(EntryALM, DebugLoc());
1104 LaneMaskPhi->insertAfter(CanonicalIVPHI);
1106 // Create the active lane mask for the next iteration of the loop before the
1107 // original terminator.
1108 VPRecipeBase *OriginalTerminator = EB->getTerminator();
1109 Builder.setInsertPoint(OriginalTerminator);
1110 auto *InLoopIncrement =
1111 Builder.createOverflowingOp(VPInstruction::CanonicalIVIncrementForPart,
1112 {IncrementValue}, {false, false}, DL);
1113 auto *ALM = Builder.createNaryOp(VPInstruction::ActiveLaneMask,
1114 {InLoopIncrement, TripCount}, DL,
1115 "active.lane.mask.next");
1116 LaneMaskPhi->addOperand(ALM);
1118 // Replace the original terminator with BranchOnCond. We have to invert the
1119 // mask here because a true condition means jumping to the exit block.
1120 auto *NotMask = Builder.createNot(ALM, DL);
1121 Builder.createNaryOp(VPInstruction::BranchOnCond, {NotMask}, DL);
1122 OriginalTerminator->eraseFromParent();
1123 return LaneMaskPhi;
1126 void VPlanTransforms::addActiveLaneMask(
1127 VPlan &Plan, bool UseActiveLaneMaskForControlFlow,
1128 bool DataAndControlFlowWithoutRuntimeCheck) {
1129 assert((!DataAndControlFlowWithoutRuntimeCheck ||
1130 UseActiveLaneMaskForControlFlow) &&
1131 "DataAndControlFlowWithoutRuntimeCheck implies "
1132 "UseActiveLaneMaskForControlFlow");
1134 auto FoundWidenCanonicalIVUser =
1135 find_if(Plan.getCanonicalIV()->users(),
1136 [](VPUser *U) { return isa<VPWidenCanonicalIVRecipe>(U); });
1137 assert(FoundWidenCanonicalIVUser &&
1138 "Must have widened canonical IV when tail folding!");
1139 auto *WideCanonicalIV =
1140 cast<VPWidenCanonicalIVRecipe>(*FoundWidenCanonicalIVUser);
1141 VPSingleDefRecipe *LaneMask;
1142 if (UseActiveLaneMaskForControlFlow) {
1143 LaneMask = addVPLaneMaskPhiAndUpdateExitBranch(
1144 Plan, DataAndControlFlowWithoutRuntimeCheck);
1145 } else {
1146 LaneMask = new VPInstruction(VPInstruction::ActiveLaneMask,
1147 {WideCanonicalIV, Plan.getTripCount()},
1148 nullptr, "active.lane.mask");
1149 LaneMask->insertAfter(WideCanonicalIV);
1152 // Walk users of WideCanonicalIV and replace all compares of the form
1153 // (ICMP_ULE, WideCanonicalIV, backedge-taken-count) with an
1154 // active-lane-mask.
1155 VPValue *BTC = Plan.getOrCreateBackedgeTakenCount();
1156 for (VPUser *U : SmallVector<VPUser *>(WideCanonicalIV->users())) {
1157 auto *CompareToReplace = dyn_cast<VPInstruction>(U);
1158 if (!CompareToReplace ||
1159 CompareToReplace->getOpcode() != Instruction::ICmp ||
1160 CompareToReplace->getPredicate() != CmpInst::ICMP_ULE ||
1161 CompareToReplace->getOperand(1) != BTC)
1162 continue;
1164 assert(CompareToReplace->getOperand(0) == WideCanonicalIV &&
1165 "WidenCanonicalIV must be the first operand of the compare");
1166 CompareToReplace->replaceAllUsesWith(LaneMask);
1167 CompareToReplace->eraseFromParent();