670 lines · cpp
1//===- NaryReassociate.cpp - Reassociate n-ary expressions ----------------===//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-exception6//7//===----------------------------------------------------------------------===//8//9// This pass reassociates n-ary add expressions and eliminates the redundancy10// exposed by the reassociation.11//12// A motivating example:13//14// void foo(int a, int b) {15// bar(a + b);16// bar((a + 2) + b);17// }18//19// An ideal compiler should reassociate (a + 2) + b to (a + b) + 2 and simplify20// the above code to21//22// int t = a + b;23// bar(t);24// bar(t + 2);25//26// However, the Reassociate pass is unable to do that because it processes each27// instruction individually and believes (a + 2) + b is the best form according28// to its rank system.29//30// To address this limitation, NaryReassociate reassociates an expression in a31// form that reuses existing instructions. As a result, NaryReassociate can32// reassociate (a + 2) + b in the example to (a + b) + 2 because it detects that33// (a + b) is computed before.34//35// NaryReassociate works as follows. For every instruction in the form of (a +36// b) + c, it checks whether a + c or b + c is already computed by a dominating37// instruction. If so, it then reassociates (a + b) + c into (a + c) + b or (b +38// c) + a and removes the redundancy accordingly. To efficiently look up whether39// an expression is computed before, we store each instruction seen and its SCEV40// into an SCEV-to-instruction map.41//42// Although the algorithm pattern-matches only ternary additions, it43// automatically handles many >3-ary expressions by walking through the function44// in the depth-first order. For example, given45//46// (a + c) + d47// ((a + b) + c) + d48//49// NaryReassociate first rewrites (a + b) + c to (a + c) + b, and then rewrites50// ((a + c) + b) + d into ((a + c) + d) + b.51//52// Finally, the above dominator-based algorithm may need to be run multiple53// iterations before emitting optimal code. One source of this need is that we54// only split an operand when it is used only once. The above algorithm can55// eliminate an instruction and decrease the usage count of its operands. As a56// result, an instruction that previously had multiple uses may become a57// single-use instruction and thus eligible for split consideration. For58// example,59//60// ac = a + c61// ab = a + b62// abc = ab + c63// ab2 = ab + b64// ab2c = ab2 + c65//66// In the first iteration, we cannot reassociate abc to ac+b because ab is used67// twice. However, we can reassociate ab2c to abc+b in the first iteration. As a68// result, ab2 becomes dead and ab will be used only once in the second69// iteration.70//71// Limitations and TODO items:72//73// 1) We only considers n-ary adds and muls for now. This should be extended74// and generalized.75//76//===----------------------------------------------------------------------===//77 78#include "llvm/Transforms/Scalar/NaryReassociate.h"79#include "llvm/ADT/DepthFirstIterator.h"80#include "llvm/ADT/SmallVector.h"81#include "llvm/Analysis/AssumptionCache.h"82#include "llvm/Analysis/ScalarEvolution.h"83#include "llvm/Analysis/ScalarEvolutionExpressions.h"84#include "llvm/Analysis/TargetLibraryInfo.h"85#include "llvm/Analysis/TargetTransformInfo.h"86#include "llvm/Analysis/ValueTracking.h"87#include "llvm/IR/BasicBlock.h"88#include "llvm/IR/Constants.h"89#include "llvm/IR/DataLayout.h"90#include "llvm/IR/DerivedTypes.h"91#include "llvm/IR/Dominators.h"92#include "llvm/IR/Function.h"93#include "llvm/IR/GetElementPtrTypeIterator.h"94#include "llvm/IR/IRBuilder.h"95#include "llvm/IR/InstrTypes.h"96#include "llvm/IR/Instruction.h"97#include "llvm/IR/Instructions.h"98#include "llvm/IR/Module.h"99#include "llvm/IR/Operator.h"100#include "llvm/IR/PatternMatch.h"101#include "llvm/IR/Type.h"102#include "llvm/IR/Value.h"103#include "llvm/IR/ValueHandle.h"104#include "llvm/InitializePasses.h"105#include "llvm/Pass.h"106#include "llvm/Support/Casting.h"107#include "llvm/Support/ErrorHandling.h"108#include "llvm/Transforms/Scalar.h"109#include "llvm/Transforms/Utils/Local.h"110#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"111#include <cassert>112#include <cstdint>113 114using namespace llvm;115using namespace PatternMatch;116 117#define DEBUG_TYPE "nary-reassociate"118 119namespace {120 121class NaryReassociateLegacyPass : public FunctionPass {122public:123 static char ID;124 125 NaryReassociateLegacyPass() : FunctionPass(ID) {126 initializeNaryReassociateLegacyPassPass(*PassRegistry::getPassRegistry());127 }128 129 bool doInitialization(Module &M) override {130 return false;131 }132 133 bool runOnFunction(Function &F) override;134 135 void getAnalysisUsage(AnalysisUsage &AU) const override {136 AU.addPreserved<DominatorTreeWrapperPass>();137 AU.addPreserved<ScalarEvolutionWrapperPass>();138 AU.addPreserved<TargetLibraryInfoWrapperPass>();139 AU.addRequired<AssumptionCacheTracker>();140 AU.addRequired<DominatorTreeWrapperPass>();141 AU.addRequired<ScalarEvolutionWrapperPass>();142 AU.addRequired<TargetLibraryInfoWrapperPass>();143 AU.addRequired<TargetTransformInfoWrapperPass>();144 AU.setPreservesCFG();145 }146 147private:148 NaryReassociatePass Impl;149};150 151} // end anonymous namespace152 153char NaryReassociateLegacyPass::ID = 0;154 155INITIALIZE_PASS_BEGIN(NaryReassociateLegacyPass, "nary-reassociate",156 "Nary reassociation", false, false)157INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)158INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)159INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)160INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)161INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)162INITIALIZE_PASS_END(NaryReassociateLegacyPass, "nary-reassociate",163 "Nary reassociation", false, false)164 165FunctionPass *llvm::createNaryReassociatePass() {166 return new NaryReassociateLegacyPass();167}168 169bool NaryReassociateLegacyPass::runOnFunction(Function &F) {170 if (skipFunction(F))171 return false;172 173 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);174 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();175 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();176 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);177 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);178 179 return Impl.runImpl(F, AC, DT, SE, TLI, TTI);180}181 182PreservedAnalyses NaryReassociatePass::run(Function &F,183 FunctionAnalysisManager &AM) {184 auto *AC = &AM.getResult<AssumptionAnalysis>(F);185 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);186 auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);187 auto *TLI = &AM.getResult<TargetLibraryAnalysis>(F);188 auto *TTI = &AM.getResult<TargetIRAnalysis>(F);189 190 if (!runImpl(F, AC, DT, SE, TLI, TTI))191 return PreservedAnalyses::all();192 193 PreservedAnalyses PA;194 PA.preserveSet<CFGAnalyses>();195 PA.preserve<ScalarEvolutionAnalysis>();196 return PA;197}198 199bool NaryReassociatePass::runImpl(Function &F, AssumptionCache *AC_,200 DominatorTree *DT_, ScalarEvolution *SE_,201 TargetLibraryInfo *TLI_,202 TargetTransformInfo *TTI_) {203 AC = AC_;204 DT = DT_;205 SE = SE_;206 TLI = TLI_;207 TTI = TTI_;208 DL = &F.getDataLayout();209 210 bool Changed = false, ChangedInThisIteration;211 do {212 ChangedInThisIteration = doOneIteration(F);213 Changed |= ChangedInThisIteration;214 } while (ChangedInThisIteration);215 return Changed;216}217 218bool NaryReassociatePass::doOneIteration(Function &F) {219 bool Changed = false;220 SeenExprs.clear();221 // Process the basic blocks in a depth first traversal of the dominator222 // tree. This order ensures that all bases of a candidate are in Candidates223 // when we process it.224 SmallVector<WeakTrackingVH, 16> DeadInsts;225 for (const auto Node : depth_first(DT)) {226 BasicBlock *BB = Node->getBlock();227 for (Instruction &OrigI : *BB) {228 const SCEV *OrigSCEV = nullptr;229 if (Instruction *NewI = tryReassociate(&OrigI, OrigSCEV)) {230 Changed = true;231 OrigI.replaceAllUsesWith(NewI);232 233 // Add 'OrigI' to the list of dead instructions.234 DeadInsts.push_back(WeakTrackingVH(&OrigI));235 // Add the rewritten instruction to SeenExprs; the original236 // instruction is deleted.237 const SCEV *NewSCEV = SE->getSCEV(NewI);238 SeenExprs[NewSCEV].push_back(WeakTrackingVH(NewI));239 240 // Ideally, NewSCEV should equal OldSCEV because tryReassociate(I)241 // is equivalent to I. However, ScalarEvolution::getSCEV may242 // weaken nsw causing NewSCEV not to equal OldSCEV. For example,243 // suppose we reassociate244 // I = &a[sext(i +nsw j)] // assuming sizeof(a[0]) = 4245 // to246 // NewI = &a[sext(i)] + sext(j).247 //248 // ScalarEvolution computes249 // getSCEV(I) = a + 4 * sext(i + j)250 // getSCEV(newI) = a + 4 * sext(i) + 4 * sext(j)251 // which are different SCEVs.252 //253 // To alleviate this issue of ScalarEvolution not always capturing254 // equivalence, we add I to SeenExprs[OldSCEV] as well so that we can255 // map both SCEV before and after tryReassociate(I) to I.256 //257 // This improvement is exercised in @reassociate_gep_nsw in258 // nary-gep.ll.259 if (NewSCEV != OrigSCEV)260 SeenExprs[OrigSCEV].push_back(WeakTrackingVH(NewI));261 } else if (OrigSCEV)262 SeenExprs[OrigSCEV].push_back(WeakTrackingVH(&OrigI));263 }264 }265 // Delete all dead instructions from 'DeadInsts'.266 // Please note ScalarEvolution is updated along the way.267 RecursivelyDeleteTriviallyDeadInstructionsPermissive(268 DeadInsts, TLI, nullptr, [this](Value *V) { SE->forgetValue(V); });269 270 return Changed;271}272 273template <typename PredT>274Instruction *275NaryReassociatePass::matchAndReassociateMinOrMax(Instruction *I,276 const SCEV *&OrigSCEV) {277 Value *LHS = nullptr;278 Value *RHS = nullptr;279 280 auto MinMaxMatcher =281 MaxMin_match<ICmpInst, bind_ty<Value>, bind_ty<Value>, PredT>(282 m_Value(LHS), m_Value(RHS));283 if (match(I, MinMaxMatcher)) {284 OrigSCEV = SE->getSCEV(I);285 if (auto *NewMinMax = dyn_cast_or_null<Instruction>(286 tryReassociateMinOrMax(I, MinMaxMatcher, LHS, RHS)))287 return NewMinMax;288 if (auto *NewMinMax = dyn_cast_or_null<Instruction>(289 tryReassociateMinOrMax(I, MinMaxMatcher, RHS, LHS)))290 return NewMinMax;291 }292 return nullptr;293}294 295Instruction *NaryReassociatePass::tryReassociate(Instruction * I,296 const SCEV *&OrigSCEV) {297 298 if (!SE->isSCEVable(I->getType()))299 return nullptr;300 301 switch (I->getOpcode()) {302 case Instruction::Add:303 case Instruction::Mul:304 OrigSCEV = SE->getSCEV(I);305 return tryReassociateBinaryOp(cast<BinaryOperator>(I));306 case Instruction::GetElementPtr:307 OrigSCEV = SE->getSCEV(I);308 return tryReassociateGEP(cast<GetElementPtrInst>(I));309 default:310 break;311 }312 313 // Try to match signed/unsigned Min/Max.314 Instruction *ResI = nullptr;315 // TODO: Currently min/max reassociation is restricted to integer types only316 // due to use of SCEVExpander which my introduce incompatible forms of min/max317 // for pointer types.318 if (I->getType()->isIntegerTy())319 if ((ResI = matchAndReassociateMinOrMax<umin_pred_ty>(I, OrigSCEV)) ||320 (ResI = matchAndReassociateMinOrMax<smin_pred_ty>(I, OrigSCEV)) ||321 (ResI = matchAndReassociateMinOrMax<umax_pred_ty>(I, OrigSCEV)) ||322 (ResI = matchAndReassociateMinOrMax<smax_pred_ty>(I, OrigSCEV)))323 return ResI;324 325 return nullptr;326}327 328static bool isGEPFoldable(GetElementPtrInst *GEP,329 const TargetTransformInfo *TTI) {330 SmallVector<const Value *, 4> Indices(GEP->indices());331 return TTI->getGEPCost(GEP->getSourceElementType(), GEP->getPointerOperand(),332 Indices) == TargetTransformInfo::TCC_Free;333}334 335Instruction *NaryReassociatePass::tryReassociateGEP(GetElementPtrInst *GEP) {336 // Not worth reassociating GEP if it is foldable.337 if (isGEPFoldable(GEP, TTI))338 return nullptr;339 340 gep_type_iterator GTI = gep_type_begin(*GEP);341 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {342 if (GTI.isSequential()) {343 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I - 1,344 GTI.getIndexedType())) {345 return NewGEP;346 }347 }348 }349 return nullptr;350}351 352bool NaryReassociatePass::requiresSignExtension(Value *Index,353 GetElementPtrInst *GEP) {354 unsigned IndexSizeInBits =355 DL->getIndexSizeInBits(GEP->getType()->getPointerAddressSpace());356 return cast<IntegerType>(Index->getType())->getBitWidth() < IndexSizeInBits;357}358 359GetElementPtrInst *360NaryReassociatePass::tryReassociateGEPAtIndex(GetElementPtrInst *GEP,361 unsigned I, Type *IndexedType) {362 SimplifyQuery SQ(*DL, DT, AC, GEP);363 Value *IndexToSplit = GEP->getOperand(I + 1);364 if (SExtInst *SExt = dyn_cast<SExtInst>(IndexToSplit)) {365 IndexToSplit = SExt->getOperand(0);366 } else if (ZExtInst *ZExt = dyn_cast<ZExtInst>(IndexToSplit)) {367 // zext can be treated as sext if the source is non-negative.368 if (isKnownNonNegative(ZExt->getOperand(0), SQ))369 IndexToSplit = ZExt->getOperand(0);370 }371 372 if (AddOperator *AO = dyn_cast<AddOperator>(IndexToSplit)) {373 // If the I-th index needs sext and the underlying add is not equipped with374 // nsw, we cannot split the add because375 // sext(LHS + RHS) != sext(LHS) + sext(RHS).376 if (requiresSignExtension(IndexToSplit, GEP) &&377 computeOverflowForSignedAdd(AO, SQ) != OverflowResult::NeverOverflows)378 return nullptr;379 380 Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);381 // IndexToSplit = LHS + RHS.382 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType))383 return NewGEP;384 // Symmetrically, try IndexToSplit = RHS + LHS.385 if (LHS != RHS) {386 if (auto *NewGEP =387 tryReassociateGEPAtIndex(GEP, I, RHS, LHS, IndexedType))388 return NewGEP;389 }390 }391 return nullptr;392}393 394GetElementPtrInst *395NaryReassociatePass::tryReassociateGEPAtIndex(GetElementPtrInst *GEP,396 unsigned I, Value *LHS,397 Value *RHS, Type *IndexedType) {398 // Look for GEP's closest dominator that has the same SCEV as GEP except that399 // the I-th index is replaced with LHS.400 SmallVector<const SCEV *, 4> IndexExprs;401 for (Use &Index : GEP->indices())402 IndexExprs.push_back(SE->getSCEV(Index));403 // Replace the I-th index with LHS.404 IndexExprs[I] = SE->getSCEV(LHS);405 Type *GEPArgType = SE->getEffectiveSCEVType(GEP->getOperand(I)->getType());406 Type *LHSType = SE->getEffectiveSCEVType(LHS->getType());407 size_t LHSSize = DL->getTypeSizeInBits(LHSType).getFixedValue();408 size_t GEPArgSize = DL->getTypeSizeInBits(GEPArgType).getFixedValue();409 if (isKnownNonNegative(LHS, SimplifyQuery(*DL, DT, AC, GEP)) &&410 LHSSize < GEPArgSize) {411 // Zero-extend LHS if it is non-negative. InstCombine canonicalizes sext to412 // zext if the source operand is proved non-negative. We should do that413 // consistently so that CandidateExpr more likely appears before. See414 // @reassociate_gep_assume for an example of this canonicalization.415 IndexExprs[I] = SE->getZeroExtendExpr(IndexExprs[I], GEPArgType);416 }417 const SCEV *CandidateExpr = SE->getGEPExpr(cast<GEPOperator>(GEP),418 IndexExprs);419 420 Value *Candidate = findClosestMatchingDominator(CandidateExpr, GEP);421 if (Candidate == nullptr)422 return nullptr;423 424 IRBuilder<> Builder(GEP);425 // Candidate should have the same pointer type as GEP.426 assert(Candidate->getType() == GEP->getType());427 428 // NewGEP = (char *)Candidate + RHS * sizeof(IndexedType)429 uint64_t IndexedSize = DL->getTypeAllocSize(IndexedType);430 Type *ElementType = GEP->getResultElementType();431 uint64_t ElementSize = DL->getTypeAllocSize(ElementType);432 // Another less rare case: because I is not necessarily the last index of the433 // GEP, the size of the type at the I-th index (IndexedSize) is not434 // necessarily divisible by ElementSize. For example,435 //436 // #pragma pack(1)437 // struct S {438 // int a[3];439 // int64 b[8];440 // };441 // #pragma pack()442 //443 // sizeof(S) = 100 is indivisible by sizeof(int64) = 8.444 //445 // TODO: bail out on this case for now. We could emit uglygep.446 if (IndexedSize % ElementSize != 0)447 return nullptr;448 449 // NewGEP = &Candidate[RHS * (sizeof(IndexedType) / sizeof(Candidate[0])));450 Type *PtrIdxTy = DL->getIndexType(GEP->getType());451 if (RHS->getType() != PtrIdxTy)452 RHS = Builder.CreateSExtOrTrunc(RHS, PtrIdxTy);453 if (IndexedSize != ElementSize) {454 RHS = Builder.CreateMul(455 RHS, ConstantInt::get(PtrIdxTy, IndexedSize / ElementSize));456 }457 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(458 Builder.CreateGEP(GEP->getResultElementType(), Candidate, RHS));459 NewGEP->setIsInBounds(GEP->isInBounds());460 NewGEP->takeName(GEP);461 return NewGEP;462}463 464Instruction *NaryReassociatePass::tryReassociateBinaryOp(BinaryOperator *I) {465 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);466 // There is no need to reassociate 0.467 if (SE->getSCEV(I)->isZero())468 return nullptr;469 if (auto *NewI = tryReassociateBinaryOp(LHS, RHS, I))470 return NewI;471 if (auto *NewI = tryReassociateBinaryOp(RHS, LHS, I))472 return NewI;473 return nullptr;474}475 476Instruction *NaryReassociatePass::tryReassociateBinaryOp(Value *LHS, Value *RHS,477 BinaryOperator *I) {478 Value *A = nullptr, *B = nullptr;479 // To be conservative, we reassociate I only when it is the only user of (A op480 // B).481 if (LHS->hasOneUse() && matchTernaryOp(I, LHS, A, B)) {482 // I = (A op B) op RHS483 // = (A op RHS) op B or (B op RHS) op A484 const SCEV *AExpr = SE->getSCEV(A), *BExpr = SE->getSCEV(B);485 const SCEV *RHSExpr = SE->getSCEV(RHS);486 if (BExpr != RHSExpr) {487 if (auto *NewI =488 tryReassociatedBinaryOp(getBinarySCEV(I, AExpr, RHSExpr), B, I))489 return NewI;490 }491 if (AExpr != RHSExpr) {492 if (auto *NewI =493 tryReassociatedBinaryOp(getBinarySCEV(I, BExpr, RHSExpr), A, I))494 return NewI;495 }496 }497 return nullptr;498}499 500Instruction *NaryReassociatePass::tryReassociatedBinaryOp(const SCEV *LHSExpr,501 Value *RHS,502 BinaryOperator *I) {503 // Look for the closest dominator LHS of I that computes LHSExpr, and replace504 // I with LHS op RHS.505 auto *LHS = findClosestMatchingDominator(LHSExpr, I);506 if (LHS == nullptr)507 return nullptr;508 509 Instruction *NewI = nullptr;510 switch (I->getOpcode()) {511 case Instruction::Add:512 NewI = BinaryOperator::CreateAdd(LHS, RHS, "", I->getIterator());513 break;514 case Instruction::Mul:515 NewI = BinaryOperator::CreateMul(LHS, RHS, "", I->getIterator());516 break;517 default:518 llvm_unreachable("Unexpected instruction.");519 }520 NewI->setDebugLoc(I->getDebugLoc());521 NewI->takeName(I);522 return NewI;523}524 525bool NaryReassociatePass::matchTernaryOp(BinaryOperator *I, Value *V,526 Value *&Op1, Value *&Op2) {527 switch (I->getOpcode()) {528 case Instruction::Add:529 return match(V, m_Add(m_Value(Op1), m_Value(Op2)));530 case Instruction::Mul:531 return match(V, m_Mul(m_Value(Op1), m_Value(Op2)));532 default:533 llvm_unreachable("Unexpected instruction.");534 }535 return false;536}537 538const SCEV *NaryReassociatePass::getBinarySCEV(BinaryOperator *I,539 const SCEV *LHS,540 const SCEV *RHS) {541 switch (I->getOpcode()) {542 case Instruction::Add:543 return SE->getAddExpr(LHS, RHS);544 case Instruction::Mul:545 return SE->getMulExpr(LHS, RHS);546 default:547 llvm_unreachable("Unexpected instruction.");548 }549 return nullptr;550}551 552Instruction *553NaryReassociatePass::findClosestMatchingDominator(const SCEV *CandidateExpr,554 Instruction *Dominatee) {555 auto Pos = SeenExprs.find(CandidateExpr);556 if (Pos == SeenExprs.end())557 return nullptr;558 559 auto &Candidates = Pos->second;560 // Because we process the basic blocks in pre-order of the dominator tree, a561 // candidate that doesn't dominate the current instruction won't dominate any562 // future instruction either. Therefore, we pop it out of the stack. This563 // optimization makes the algorithm O(n).564 while (!Candidates.empty()) {565 // Candidates stores WeakTrackingVHs, so a candidate can be nullptr if it's566 // removed during rewriting.567 if (Value *Candidate = Candidates.pop_back_val()) {568 Instruction *CandidateInstruction = cast<Instruction>(Candidate);569 if (!DT->dominates(CandidateInstruction, Dominatee))570 continue;571 572 // Make sure that the instruction is safe to reuse without introducing573 // poison.574 SmallVector<Instruction *> DropPoisonGeneratingInsts;575 if (!SE->canReuseInstruction(CandidateExpr, CandidateInstruction,576 DropPoisonGeneratingInsts))577 continue;578 579 for (Instruction *I : DropPoisonGeneratingInsts)580 I->dropPoisonGeneratingAnnotations();581 582 return CandidateInstruction;583 }584 }585 return nullptr;586}587 588template <typename MaxMinT> static SCEVTypes convertToSCEVype(MaxMinT &MM) {589 if (std::is_same_v<smax_pred_ty, typename MaxMinT::PredType>)590 return scSMaxExpr;591 else if (std::is_same_v<umax_pred_ty, typename MaxMinT::PredType>)592 return scUMaxExpr;593 else if (std::is_same_v<smin_pred_ty, typename MaxMinT::PredType>)594 return scSMinExpr;595 else if (std::is_same_v<umin_pred_ty, typename MaxMinT::PredType>)596 return scUMinExpr;597 598 llvm_unreachable("Can't convert MinMax pattern to SCEV type");599 return scUnknown;600}601 602// Parameters:603// I - instruction matched by MaxMinMatch matcher604// MaxMinMatch - min/max idiom matcher605// LHS - first operand of I606// RHS - second operand of I607template <typename MaxMinT>608Value *NaryReassociatePass::tryReassociateMinOrMax(Instruction *I,609 MaxMinT MaxMinMatch,610 Value *LHS, Value *RHS) {611 Value *A = nullptr, *B = nullptr;612 MaxMinT m_MaxMin(m_Value(A), m_Value(B));613 614 if (!match(LHS, m_MaxMin))615 return nullptr;616 617 if (LHS->hasNUsesOrMore(3) ||618 // The optimization is profitable only if LHS can be removed in the end.619 // In other words LHS should be used (directly or indirectly) by I only.620 llvm::any_of(LHS->users(), [&](auto *U) {621 return U != I && !(U->hasOneUser() && *U->users().begin() == I);622 }))623 return nullptr;624 625 auto tryCombination = [&](Value *A, const SCEV *AExpr, Value *B,626 const SCEV *BExpr, Value *C,627 const SCEV *CExpr) -> Value * {628 SmallVector<const SCEV *, 2> Ops1{BExpr, AExpr};629 const SCEVTypes SCEVType = convertToSCEVype(m_MaxMin);630 const SCEV *R1Expr = SE->getMinMaxExpr(SCEVType, Ops1);631 632 Instruction *R1MinMax = findClosestMatchingDominator(R1Expr, I);633 634 if (!R1MinMax)635 return nullptr;636 637 LLVM_DEBUG(dbgs() << "NARY: Found common sub-expr: " << *R1MinMax << "\n");638 639 SmallVector<const SCEV *, 2> Ops2{SE->getUnknown(C),640 SE->getUnknown(R1MinMax)};641 const SCEV *R2Expr = SE->getMinMaxExpr(SCEVType, Ops2);642 643 SCEVExpander Expander(*SE, *DL, "nary-reassociate");644 Value *NewMinMax = Expander.expandCodeFor(R2Expr, I->getType(), I);645 NewMinMax->setName(Twine(I->getName()).concat(".nary"));646 647 LLVM_DEBUG(dbgs() << "NARY: Deleting: " << *I << "\n"648 << "NARY: Inserting: " << *NewMinMax << "\n");649 return NewMinMax;650 };651 652 const SCEV *AExpr = SE->getSCEV(A);653 const SCEV *BExpr = SE->getSCEV(B);654 const SCEV *RHSExpr = SE->getSCEV(RHS);655 656 if (BExpr != RHSExpr) {657 // Try (A op RHS) op B658 if (auto *NewMinMax = tryCombination(A, AExpr, RHS, RHSExpr, B, BExpr))659 return NewMinMax;660 }661 662 if (AExpr != RHSExpr) {663 // Try (RHS op B) op A664 if (auto *NewMinMax = tryCombination(RHS, RHSExpr, B, BExpr, A, AExpr))665 return NewMinMax;666 }667 668 return nullptr;669}670