1e60b36cfSFlorian Hahn //===-- VPlanTransforms.cpp - Utility VPlan to VPlan transforms -----------===// 2e60b36cfSFlorian Hahn // 3e60b36cfSFlorian Hahn // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4e60b36cfSFlorian Hahn // See https://llvm.org/LICENSE.txt for license information. 5e60b36cfSFlorian Hahn // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6e60b36cfSFlorian Hahn // 7e60b36cfSFlorian Hahn //===----------------------------------------------------------------------===// 8e60b36cfSFlorian Hahn /// 9e60b36cfSFlorian Hahn /// \file 10e60b36cfSFlorian Hahn /// This file implements a set of utility VPlan to VPlan transformations. 11e60b36cfSFlorian Hahn /// 12e60b36cfSFlorian Hahn //===----------------------------------------------------------------------===// 13e60b36cfSFlorian Hahn 14e60b36cfSFlorian Hahn #include "VPlanTransforms.h" 15e60b36cfSFlorian Hahn #include "llvm/ADT/PostOrderIterator.h" 16e60b36cfSFlorian Hahn 17e60b36cfSFlorian Hahn using namespace llvm; 18e60b36cfSFlorian Hahn 19e60b36cfSFlorian Hahn void VPlanTransforms::VPInstructionsToVPRecipes( 20e60b36cfSFlorian Hahn Loop *OrigLoop, VPlanPtr &Plan, 21d0d38df0SDavid Green LoopVectorizationLegality::InductionList &Inductions, 220de8aeaeSMauri Mustonen SmallPtrSetImpl<Instruction *> &DeadInstructions, ScalarEvolution &SE) { 23e60b36cfSFlorian Hahn 24e60b36cfSFlorian Hahn auto *TopRegion = cast<VPRegionBlock>(Plan->getEntry()); 25e60b36cfSFlorian Hahn ReversePostOrderTraversal<VPBlockBase *> RPOT(TopRegion->getEntry()); 26e60b36cfSFlorian Hahn 27e60b36cfSFlorian Hahn for (VPBlockBase *Base : RPOT) { 28e60b36cfSFlorian Hahn // Do not widen instructions in pre-header and exit blocks. 29e60b36cfSFlorian Hahn if (Base->getNumPredecessors() == 0 || Base->getNumSuccessors() == 0) 30e60b36cfSFlorian Hahn continue; 31e60b36cfSFlorian Hahn 32e60b36cfSFlorian Hahn VPBasicBlock *VPBB = Base->getEntryBasicBlock(); 33e60b36cfSFlorian Hahn // Introduce each ingredient into VPlan. 34e60b36cfSFlorian Hahn for (auto I = VPBB->begin(), E = VPBB->end(); I != E;) { 35e60b36cfSFlorian Hahn VPRecipeBase *Ingredient = &*I++; 36a0e1313cSFlorian Hahn VPValue *VPV = Ingredient->getVPSingleValue(); 3715a74b64SFlorian Hahn Instruction *Inst = cast<Instruction>(VPV->getUnderlyingValue()); 38e60b36cfSFlorian Hahn if (DeadInstructions.count(Inst)) { 3976afbf60SFlorian Hahn VPValue DummyValue; 4015a74b64SFlorian Hahn VPV->replaceAllUsesWith(&DummyValue); 41e60b36cfSFlorian Hahn Ingredient->eraseFromParent(); 42e60b36cfSFlorian Hahn continue; 43e60b36cfSFlorian Hahn } 44e60b36cfSFlorian Hahn 45e60b36cfSFlorian Hahn VPRecipeBase *NewRecipe = nullptr; 4615a74b64SFlorian Hahn if (auto *VPPhi = dyn_cast<VPWidenPHIRecipe>(Ingredient)) { 4715a74b64SFlorian Hahn auto *Phi = cast<PHINode>(VPPhi->getUnderlyingValue()); 48d0d38df0SDavid Green InductionDescriptor II = Inductions.lookup(Phi); 49e60b36cfSFlorian Hahn if (II.getKind() == InductionDescriptor::IK_IntInduction || 50e60b36cfSFlorian Hahn II.getKind() == InductionDescriptor::IK_FpInduction) { 51816dba48SFlorian Hahn VPValue *Start = Plan->getOrAddVPValue(II.getStartValue()); 52daaa0e35SFlorian Hahn NewRecipe = new VPWidenIntOrFpInductionRecipe(Phi, Start, nullptr); 5315a74b64SFlorian Hahn } else { 5415a74b64SFlorian Hahn Plan->addVPValue(Phi, VPPhi); 5515a74b64SFlorian Hahn continue; 5615a74b64SFlorian Hahn } 5715a74b64SFlorian Hahn } else { 5815a74b64SFlorian Hahn assert(isa<VPInstruction>(Ingredient) && 5915a74b64SFlorian Hahn "only VPInstructions expected here"); 6015a74b64SFlorian Hahn assert(!isa<PHINode>(Inst) && "phis should be handled above"); 6115a74b64SFlorian Hahn // Create VPWidenMemoryInstructionRecipe for loads and stores. 6215a74b64SFlorian Hahn if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) { 6315a74b64SFlorian Hahn NewRecipe = new VPWidenMemoryInstructionRecipe( 6415a74b64SFlorian Hahn *Load, Plan->getOrAddVPValue(getLoadStorePointerOperand(Inst)), 6515a74b64SFlorian Hahn nullptr /*Mask*/); 6615a74b64SFlorian Hahn } else if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) { 6715a74b64SFlorian Hahn NewRecipe = new VPWidenMemoryInstructionRecipe( 6815a74b64SFlorian Hahn *Store, Plan->getOrAddVPValue(getLoadStorePointerOperand(Inst)), 6915a74b64SFlorian Hahn Plan->getOrAddVPValue(Store->getValueOperand()), 7015a74b64SFlorian Hahn nullptr /*Mask*/); 71e60b36cfSFlorian Hahn } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 72c0cdba72SFlorian Hahn NewRecipe = new VPWidenGEPRecipe( 73c0cdba72SFlorian Hahn GEP, Plan->mapToVPValues(GEP->operands()), OrigLoop); 74494b5ba3SMauri Mustonen } else if (CallInst *CI = dyn_cast<CallInst>(Inst)) { 75494b5ba3SMauri Mustonen NewRecipe = new VPWidenCallRecipe( 76494b5ba3SMauri Mustonen *CI, Plan->mapToVPValues(CI->arg_operands())); 770de8aeaeSMauri Mustonen } else if (SelectInst *SI = dyn_cast<SelectInst>(Inst)) { 780de8aeaeSMauri Mustonen bool InvariantCond = 790de8aeaeSMauri Mustonen SE.isLoopInvariant(SE.getSCEV(SI->getOperand(0)), OrigLoop); 800de8aeaeSMauri Mustonen NewRecipe = new VPWidenSelectRecipe( 810de8aeaeSMauri Mustonen *SI, Plan->mapToVPValues(SI->operands()), InvariantCond); 8215a74b64SFlorian Hahn } else { 83e8937985SFlorian Hahn NewRecipe = 84e8937985SFlorian Hahn new VPWidenRecipe(*Inst, Plan->mapToVPValues(Inst->operands())); 8515a74b64SFlorian Hahn } 8615a74b64SFlorian Hahn } 87e60b36cfSFlorian Hahn 88e60b36cfSFlorian Hahn NewRecipe->insertBefore(Ingredient); 8976afbf60SFlorian Hahn if (NewRecipe->getNumDefinedValues() == 1) 90a0e1313cSFlorian Hahn VPV->replaceAllUsesWith(NewRecipe->getVPSingleValue()); 9176afbf60SFlorian Hahn else 9276afbf60SFlorian Hahn assert(NewRecipe->getNumDefinedValues() == 0 && 9376afbf60SFlorian Hahn "Only recpies with zero or one defined values expected"); 94e60b36cfSFlorian Hahn Ingredient->eraseFromParent(); 9554a14c26SFlorian Hahn Plan->removeVPValueFor(Inst); 9654a14c26SFlorian Hahn for (auto *Def : NewRecipe->definedValues()) { 9754a14c26SFlorian Hahn Plan->addVPValue(Inst, Def); 9854a14c26SFlorian Hahn } 99e60b36cfSFlorian Hahn } 100e60b36cfSFlorian Hahn } 101e60b36cfSFlorian Hahn } 10265d3dd7cSFlorian Hahn 10365d3dd7cSFlorian Hahn bool VPlanTransforms::sinkScalarOperands(VPlan &Plan) { 10465d3dd7cSFlorian Hahn auto Iter = depth_first( 10565d3dd7cSFlorian Hahn VPBlockRecursiveTraversalWrapper<VPBlockBase *>(Plan.getEntry())); 10665d3dd7cSFlorian Hahn bool Changed = false; 10765d3dd7cSFlorian Hahn // First, collect the operands of all predicated replicate recipes as seeds 10865d3dd7cSFlorian Hahn // for sinking. 109*7359450eSFlorian Hahn SetVector<std::pair<VPBasicBlock *, VPValue *>> WorkList; 11065d3dd7cSFlorian Hahn for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { 11165d3dd7cSFlorian Hahn for (auto &Recipe : *VPBB) { 11265d3dd7cSFlorian Hahn auto *RepR = dyn_cast<VPReplicateRecipe>(&Recipe); 11365d3dd7cSFlorian Hahn if (!RepR || !RepR->isPredicated()) 11465d3dd7cSFlorian Hahn continue; 115*7359450eSFlorian Hahn for (VPValue *Op : RepR->operands()) 116*7359450eSFlorian Hahn WorkList.insert(std::make_pair(RepR->getParent(), Op)); 11765d3dd7cSFlorian Hahn } 11865d3dd7cSFlorian Hahn } 11965d3dd7cSFlorian Hahn 12065d3dd7cSFlorian Hahn // Try to sink each replicate recipe in the worklist. 12165d3dd7cSFlorian Hahn while (!WorkList.empty()) { 122*7359450eSFlorian Hahn VPBasicBlock *SinkTo; 123*7359450eSFlorian Hahn VPValue *C; 124*7359450eSFlorian Hahn std::tie(SinkTo, C) = WorkList.pop_back_val(); 12565d3dd7cSFlorian Hahn auto *SinkCandidate = dyn_cast_or_null<VPReplicateRecipe>(C->Def); 126*7359450eSFlorian Hahn if (!SinkCandidate || SinkCandidate->isUniform() || 127*7359450eSFlorian Hahn SinkCandidate->getParent() == SinkTo || 12865d3dd7cSFlorian Hahn SinkCandidate->mayHaveSideEffects() || 12965d3dd7cSFlorian Hahn SinkCandidate->mayReadOrWriteMemory()) 13065d3dd7cSFlorian Hahn continue; 13165d3dd7cSFlorian Hahn 13265d3dd7cSFlorian Hahn // All recipe users of the sink candidate must be in the same block SinkTo. 13365d3dd7cSFlorian Hahn if (any_of(SinkCandidate->users(), [SinkTo](VPUser *U) { 13465d3dd7cSFlorian Hahn auto *UI = dyn_cast<VPRecipeBase>(U); 13565d3dd7cSFlorian Hahn return !UI || UI->getParent() != SinkTo; 13665d3dd7cSFlorian Hahn })) 13765d3dd7cSFlorian Hahn continue; 13865d3dd7cSFlorian Hahn 13965d3dd7cSFlorian Hahn SinkCandidate->moveBefore(*SinkTo, SinkTo->getFirstNonPhi()); 140*7359450eSFlorian Hahn for (VPValue *Op : SinkCandidate->operands()) 141*7359450eSFlorian Hahn WorkList.insert(std::make_pair(SinkTo, Op)); 14265d3dd7cSFlorian Hahn Changed = true; 14365d3dd7cSFlorian Hahn } 14465d3dd7cSFlorian Hahn return Changed; 14565d3dd7cSFlorian Hahn } 14680aa7e14SFlorian Hahn 14780aa7e14SFlorian Hahn /// If \p R is a region with a VPBranchOnMaskRecipe in the entry block, return 14880aa7e14SFlorian Hahn /// the mask. 14980aa7e14SFlorian Hahn VPValue *getPredicatedMask(VPRegionBlock *R) { 15080aa7e14SFlorian Hahn auto *EntryBB = dyn_cast<VPBasicBlock>(R->getEntry()); 15180aa7e14SFlorian Hahn if (!EntryBB || EntryBB->size() != 1 || 15280aa7e14SFlorian Hahn !isa<VPBranchOnMaskRecipe>(EntryBB->begin())) 15380aa7e14SFlorian Hahn return nullptr; 15480aa7e14SFlorian Hahn 15580aa7e14SFlorian Hahn return cast<VPBranchOnMaskRecipe>(&*EntryBB->begin())->getOperand(0); 15680aa7e14SFlorian Hahn } 15780aa7e14SFlorian Hahn 15880aa7e14SFlorian Hahn /// If \p R is a triangle region, return the 'then' block of the triangle. 15980aa7e14SFlorian Hahn static VPBasicBlock *getPredicatedThenBlock(VPRegionBlock *R) { 16080aa7e14SFlorian Hahn auto *EntryBB = cast<VPBasicBlock>(R->getEntry()); 16180aa7e14SFlorian Hahn if (EntryBB->getNumSuccessors() != 2) 16280aa7e14SFlorian Hahn return nullptr; 16380aa7e14SFlorian Hahn 16480aa7e14SFlorian Hahn auto *Succ0 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[0]); 16580aa7e14SFlorian Hahn auto *Succ1 = dyn_cast<VPBasicBlock>(EntryBB->getSuccessors()[1]); 16680aa7e14SFlorian Hahn if (!Succ0 || !Succ1) 16780aa7e14SFlorian Hahn return nullptr; 16880aa7e14SFlorian Hahn 16980aa7e14SFlorian Hahn if (Succ0->getNumSuccessors() + Succ1->getNumSuccessors() != 1) 17080aa7e14SFlorian Hahn return nullptr; 17180aa7e14SFlorian Hahn if (Succ0->getSingleSuccessor() == Succ1) 17280aa7e14SFlorian Hahn return Succ0; 17380aa7e14SFlorian Hahn if (Succ1->getSingleSuccessor() == Succ0) 17480aa7e14SFlorian Hahn return Succ1; 17580aa7e14SFlorian Hahn return nullptr; 17680aa7e14SFlorian Hahn } 17780aa7e14SFlorian Hahn 17880aa7e14SFlorian Hahn bool VPlanTransforms::mergeReplicateRegions(VPlan &Plan) { 17980aa7e14SFlorian Hahn SetVector<VPRegionBlock *> DeletedRegions; 18080aa7e14SFlorian Hahn bool Changed = false; 18180aa7e14SFlorian Hahn 18280aa7e14SFlorian Hahn // Collect region blocks to process up-front, to avoid iterator invalidation 18380aa7e14SFlorian Hahn // issues while merging regions. 18480aa7e14SFlorian Hahn SmallVector<VPRegionBlock *, 8> CandidateRegions( 18580aa7e14SFlorian Hahn VPBlockUtils::blocksOnly<VPRegionBlock>(depth_first( 18680aa7e14SFlorian Hahn VPBlockRecursiveTraversalWrapper<VPBlockBase *>(Plan.getEntry())))); 18780aa7e14SFlorian Hahn 18880aa7e14SFlorian Hahn // Check if Base is a predicated triangle, followed by an empty block, 18980aa7e14SFlorian Hahn // followed by another predicate triangle. If that's the case, move the 19080aa7e14SFlorian Hahn // recipes from the first to the second triangle. 19180aa7e14SFlorian Hahn for (VPRegionBlock *Region1 : CandidateRegions) { 19280aa7e14SFlorian Hahn if (DeletedRegions.contains(Region1)) 19380aa7e14SFlorian Hahn continue; 19480aa7e14SFlorian Hahn auto *MiddleBasicBlock = 19580aa7e14SFlorian Hahn dyn_cast_or_null<VPBasicBlock>(Region1->getSingleSuccessor()); 19680aa7e14SFlorian Hahn if (!MiddleBasicBlock || !MiddleBasicBlock->empty()) 19780aa7e14SFlorian Hahn continue; 19880aa7e14SFlorian Hahn 19980aa7e14SFlorian Hahn auto *Region2 = 20080aa7e14SFlorian Hahn dyn_cast_or_null<VPRegionBlock>(MiddleBasicBlock->getSingleSuccessor()); 20180aa7e14SFlorian Hahn if (!Region2) 20280aa7e14SFlorian Hahn continue; 20380aa7e14SFlorian Hahn 20480aa7e14SFlorian Hahn VPValue *Mask1 = getPredicatedMask(Region1); 20580aa7e14SFlorian Hahn VPValue *Mask2 = getPredicatedMask(Region2); 20680aa7e14SFlorian Hahn if (!Mask1 || Mask1 != Mask2) 20780aa7e14SFlorian Hahn continue; 20880aa7e14SFlorian Hahn VPBasicBlock *Then1 = getPredicatedThenBlock(Region1); 20980aa7e14SFlorian Hahn VPBasicBlock *Then2 = getPredicatedThenBlock(Region2); 21080aa7e14SFlorian Hahn if (!Then1 || !Then2) 21180aa7e14SFlorian Hahn continue; 21280aa7e14SFlorian Hahn 21380aa7e14SFlorian Hahn assert(Mask1 && Mask2 && "both region must have conditions"); 21480aa7e14SFlorian Hahn 21580aa7e14SFlorian Hahn // Note: No fusion-preventing memory dependencies are expected in either 21680aa7e14SFlorian Hahn // region. Such dependencies should be rejected during earlier dependence 21780aa7e14SFlorian Hahn // checks, which guarantee accesses can be re-ordered for vectorization. 21880aa7e14SFlorian Hahn // 21980aa7e14SFlorian Hahn // Move recipes to the successor region. 22080aa7e14SFlorian Hahn for (VPRecipeBase &ToMove : make_early_inc_range(reverse(*Then1))) 22180aa7e14SFlorian Hahn ToMove.moveBefore(*Then2, Then2->getFirstNonPhi()); 22280aa7e14SFlorian Hahn 22380aa7e14SFlorian Hahn auto *Merge1 = cast<VPBasicBlock>(Then1->getSingleSuccessor()); 22480aa7e14SFlorian Hahn auto *Merge2 = cast<VPBasicBlock>(Then2->getSingleSuccessor()); 22580aa7e14SFlorian Hahn 22680aa7e14SFlorian Hahn // Move VPPredInstPHIRecipes from the merge block to the successor region's 22780aa7e14SFlorian Hahn // merge block. Update all users inside the successor region to use the 22880aa7e14SFlorian Hahn // original values. 22980aa7e14SFlorian Hahn for (VPRecipeBase &Phi1ToMove : make_early_inc_range(reverse(*Merge1))) { 23080aa7e14SFlorian Hahn VPValue *PredInst1 = 23180aa7e14SFlorian Hahn cast<VPPredInstPHIRecipe>(&Phi1ToMove)->getOperand(0); 232368af755SFlorian Hahn VPValue *Phi1ToMoveV = Phi1ToMove.getVPSingleValue(); 233368af755SFlorian Hahn SmallVector<VPUser *> Users(Phi1ToMoveV->user_begin(), 234368af755SFlorian Hahn Phi1ToMoveV->user_end()); 235368af755SFlorian Hahn for (VPUser *U : Users) { 23680aa7e14SFlorian Hahn auto *UI = dyn_cast<VPRecipeBase>(U); 23780aa7e14SFlorian Hahn if (!UI || UI->getParent() != Then2) 23880aa7e14SFlorian Hahn continue; 23980aa7e14SFlorian Hahn for (unsigned I = 0, E = U->getNumOperands(); I != E; ++I) { 240368af755SFlorian Hahn if (Phi1ToMoveV != U->getOperand(I)) 24180aa7e14SFlorian Hahn continue; 24280aa7e14SFlorian Hahn U->setOperand(I, PredInst1); 24380aa7e14SFlorian Hahn } 24480aa7e14SFlorian Hahn } 24580aa7e14SFlorian Hahn 24680aa7e14SFlorian Hahn Phi1ToMove.moveBefore(*Merge2, Merge2->begin()); 24780aa7e14SFlorian Hahn } 24880aa7e14SFlorian Hahn 24980aa7e14SFlorian Hahn // Finally, remove the first region. 25080aa7e14SFlorian Hahn for (VPBlockBase *Pred : make_early_inc_range(Region1->getPredecessors())) { 25180aa7e14SFlorian Hahn VPBlockUtils::disconnectBlocks(Pred, Region1); 25280aa7e14SFlorian Hahn VPBlockUtils::connectBlocks(Pred, MiddleBasicBlock); 25380aa7e14SFlorian Hahn } 25480aa7e14SFlorian Hahn VPBlockUtils::disconnectBlocks(Region1, MiddleBasicBlock); 25580aa7e14SFlorian Hahn DeletedRegions.insert(Region1); 25680aa7e14SFlorian Hahn } 25780aa7e14SFlorian Hahn 25880aa7e14SFlorian Hahn for (VPRegionBlock *ToDelete : DeletedRegions) 25980aa7e14SFlorian Hahn delete ToDelete; 26080aa7e14SFlorian Hahn return Changed; 26180aa7e14SFlorian Hahn } 262