1 //===- NaryReassociate.cpp - Reassociate n-ary expressions ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass reassociates n-ary add expressions and eliminates the redundancy
11 // exposed by the reassociation.
13 // A motivating example:
15 // void foo(int a, int b) {
20 // An ideal compiler should reassociate (a + 2) + b to (a + b) + 2 and simplify
27 // However, the Reassociate pass is unable to do that because it processes each
28 // instruction individually and believes (a + 2) + b is the best form according
29 // to its rank system.
31 // To address this limitation, NaryReassociate reassociates an expression in a
32 // form that reuses existing instructions. As a result, NaryReassociate can
33 // reassociate (a + 2) + b in the example to (a + b) + 2 because it detects that
34 // (a + b) is computed before.
36 // NaryReassociate works as follows. For every instruction in the form of (a +
37 // b) + c, it checks whether a + c or b + c is already computed by a dominating
38 // instruction. If so, it then reassociates (a + b) + c into (a + c) + b or (b +
39 // c) + a and removes the redundancy accordingly. To efficiently look up whether
40 // an expression is computed before, we store each instruction seen and its SCEV
41 // into an SCEV-to-instruction map.
43 // Although the algorithm pattern-matches only ternary additions, it
44 // automatically handles many >3-ary expressions by walking through the function
45 // in the depth-first order. For example, given
50 // NaryReassociate first rewrites (a + b) + c to (a + c) + b, and then rewrites
51 // ((a + c) + b) + d into ((a + c) + d) + b.
53 // Finally, the above dominator-based algorithm may need to be run multiple
54 // iterations before emitting optimal code. One source of this need is that we
55 // only split an operand when it is used only once. The above algorithm can
56 // eliminate an instruction and decrease the usage count of its operands. As a
57 // result, an instruction that previously had multiple uses may become a
58 // single-use instruction and thus eligible for split consideration. For
67 // In the first iteration, we cannot reassociate abc to ac+b because ab is used
68 // twice. However, we can reassociate ab2c to abc+b in the first iteration. As a
69 // result, ab2 becomes dead and ab will be used only once in the second
72 // Limitations and TODO items:
74 // 1) We only considers n-ary adds and muls for now. This should be extended
77 //===----------------------------------------------------------------------===//
79 #include "llvm/Transforms/Scalar/NaryReassociate.h"
80 #include "llvm/Analysis/ValueTracking.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/Support/Debug.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/Transforms/Scalar.h"
86 #include "llvm/Transforms/Utils/Local.h"
88 using namespace PatternMatch;
90 #define DEBUG_TYPE "nary-reassociate"
93 class NaryReassociateLegacyPass : public FunctionPass {
97 NaryReassociateLegacyPass() : FunctionPass(ID) {
98 initializeNaryReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
101 bool doInitialization(Module &M) override {
104 bool runOnFunction(Function &F) override;
106 void getAnalysisUsage(AnalysisUsage &AU) const override {
107 AU.addPreserved<DominatorTreeWrapperPass>();
108 AU.addPreserved<ScalarEvolutionWrapperPass>();
109 AU.addPreserved<TargetLibraryInfoWrapperPass>();
110 AU.addRequired<AssumptionCacheTracker>();
111 AU.addRequired<DominatorTreeWrapperPass>();
112 AU.addRequired<ScalarEvolutionWrapperPass>();
113 AU.addRequired<TargetLibraryInfoWrapperPass>();
114 AU.addRequired<TargetTransformInfoWrapperPass>();
115 AU.setPreservesCFG();
119 NaryReassociatePass Impl;
121 } // anonymous namespace
123 char NaryReassociateLegacyPass::ID = 0;
124 INITIALIZE_PASS_BEGIN(NaryReassociateLegacyPass, "nary-reassociate",
125 "Nary reassociation", false, false)
126 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
127 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
128 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
129 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
130 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
131 INITIALIZE_PASS_END(NaryReassociateLegacyPass, "nary-reassociate",
132 "Nary reassociation", false, false)
134 FunctionPass *llvm::createNaryReassociatePass() {
135 return new NaryReassociateLegacyPass();
138 bool NaryReassociateLegacyPass::runOnFunction(Function &F) {
142 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
143 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
144 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
145 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
146 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
148 return Impl.runImpl(F, AC, DT, SE, TLI, TTI);
151 PreservedAnalyses NaryReassociatePass::run(Function &F,
152 FunctionAnalysisManager &AM) {
153 auto *AC = &AM.getResult<AssumptionAnalysis>(F);
154 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
155 auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
156 auto *TLI = &AM.getResult<TargetLibraryAnalysis>(F);
157 auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
159 if (!runImpl(F, AC, DT, SE, TLI, TTI))
160 return PreservedAnalyses::all();
162 PreservedAnalyses PA;
163 PA.preserveSet<CFGAnalyses>();
164 PA.preserve<ScalarEvolutionAnalysis>();
168 bool NaryReassociatePass::runImpl(Function &F, AssumptionCache *AC_,
169 DominatorTree *DT_, ScalarEvolution *SE_,
170 TargetLibraryInfo *TLI_,
171 TargetTransformInfo *TTI_) {
177 DL = &F.getParent()->getDataLayout();
179 bool Changed = false, ChangedInThisIteration;
181 ChangedInThisIteration = doOneIteration(F);
182 Changed |= ChangedInThisIteration;
183 } while (ChangedInThisIteration);
187 // Whitelist the instruction types NaryReassociate handles for now.
188 static bool isPotentiallyNaryReassociable(Instruction *I) {
189 switch (I->getOpcode()) {
190 case Instruction::Add:
191 case Instruction::GetElementPtr:
192 case Instruction::Mul:
199 bool NaryReassociatePass::doOneIteration(Function &F) {
200 bool Changed = false;
202 // Process the basic blocks in a depth first traversal of the dominator
203 // tree. This order ensures that all bases of a candidate are in Candidates
204 // when we process it.
205 for (const auto Node : depth_first(DT)) {
206 BasicBlock *BB = Node->getBlock();
207 for (auto I = BB->begin(); I != BB->end(); ++I) {
208 if (SE->isSCEVable(I->getType()) && isPotentiallyNaryReassociable(&*I)) {
209 const SCEV *OldSCEV = SE->getSCEV(&*I);
210 if (Instruction *NewI = tryReassociate(&*I)) {
212 SE->forgetValue(&*I);
213 I->replaceAllUsesWith(NewI);
214 // If SeenExprs constains I's WeakVH, that entry will be replaced with
216 RecursivelyDeleteTriviallyDeadInstructions(&*I, TLI);
217 I = NewI->getIterator();
219 // Add the rewritten instruction to SeenExprs; the original instruction
221 const SCEV *NewSCEV = SE->getSCEV(&*I);
222 SeenExprs[NewSCEV].push_back(WeakVH(&*I));
223 // Ideally, NewSCEV should equal OldSCEV because tryReassociate(I)
224 // is equivalent to I. However, ScalarEvolution::getSCEV may
225 // weaken nsw causing NewSCEV not to equal OldSCEV. For example, suppose
227 // I = &a[sext(i +nsw j)] // assuming sizeof(a[0]) = 4
229 // NewI = &a[sext(i)] + sext(j).
231 // ScalarEvolution computes
232 // getSCEV(I) = a + 4 * sext(i + j)
233 // getSCEV(newI) = a + 4 * sext(i) + 4 * sext(j)
234 // which are different SCEVs.
236 // To alleviate this issue of ScalarEvolution not always capturing
237 // equivalence, we add I to SeenExprs[OldSCEV] as well so that we can
238 // map both SCEV before and after tryReassociate(I) to I.
240 // This improvement is exercised in @reassociate_gep_nsw in nary-gep.ll.
241 if (NewSCEV != OldSCEV)
242 SeenExprs[OldSCEV].push_back(WeakVH(&*I));
249 Instruction *NaryReassociatePass::tryReassociate(Instruction *I) {
250 switch (I->getOpcode()) {
251 case Instruction::Add:
252 case Instruction::Mul:
253 return tryReassociateBinaryOp(cast<BinaryOperator>(I));
254 case Instruction::GetElementPtr:
255 return tryReassociateGEP(cast<GetElementPtrInst>(I));
257 llvm_unreachable("should be filtered out by isPotentiallyNaryReassociable");
261 static bool isGEPFoldable(GetElementPtrInst *GEP,
262 const TargetTransformInfo *TTI) {
263 SmallVector<const Value*, 4> Indices;
264 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I)
265 Indices.push_back(*I);
266 return TTI->getGEPCost(GEP->getSourceElementType(), GEP->getPointerOperand(),
267 Indices) == TargetTransformInfo::TCC_Free;
270 Instruction *NaryReassociatePass::tryReassociateGEP(GetElementPtrInst *GEP) {
271 // Not worth reassociating GEP if it is foldable.
272 if (isGEPFoldable(GEP, TTI))
275 gep_type_iterator GTI = gep_type_begin(*GEP);
276 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
277 if (GTI.isSequential()) {
278 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I - 1,
279 GTI.getIndexedType())) {
287 bool NaryReassociatePass::requiresSignExtension(Value *Index,
288 GetElementPtrInst *GEP) {
289 unsigned PointerSizeInBits =
290 DL->getPointerSizeInBits(GEP->getType()->getPointerAddressSpace());
291 return cast<IntegerType>(Index->getType())->getBitWidth() < PointerSizeInBits;
295 NaryReassociatePass::tryReassociateGEPAtIndex(GetElementPtrInst *GEP,
296 unsigned I, Type *IndexedType) {
297 Value *IndexToSplit = GEP->getOperand(I + 1);
298 if (SExtInst *SExt = dyn_cast<SExtInst>(IndexToSplit)) {
299 IndexToSplit = SExt->getOperand(0);
300 } else if (ZExtInst *ZExt = dyn_cast<ZExtInst>(IndexToSplit)) {
301 // zext can be treated as sext if the source is non-negative.
302 if (isKnownNonNegative(ZExt->getOperand(0), *DL, 0, AC, GEP, DT))
303 IndexToSplit = ZExt->getOperand(0);
306 if (AddOperator *AO = dyn_cast<AddOperator>(IndexToSplit)) {
307 // If the I-th index needs sext and the underlying add is not equipped with
308 // nsw, we cannot split the add because
309 // sext(LHS + RHS) != sext(LHS) + sext(RHS).
310 if (requiresSignExtension(IndexToSplit, GEP) &&
311 computeOverflowForSignedAdd(AO, *DL, AC, GEP, DT) !=
312 OverflowResult::NeverOverflows)
315 Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);
316 // IndexToSplit = LHS + RHS.
317 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType))
319 // Symmetrically, try IndexToSplit = RHS + LHS.
322 tryReassociateGEPAtIndex(GEP, I, RHS, LHS, IndexedType))
330 NaryReassociatePass::tryReassociateGEPAtIndex(GetElementPtrInst *GEP,
331 unsigned I, Value *LHS,
332 Value *RHS, Type *IndexedType) {
333 // Look for GEP's closest dominator that has the same SCEV as GEP except that
334 // the I-th index is replaced with LHS.
335 SmallVector<const SCEV *, 4> IndexExprs;
336 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
337 IndexExprs.push_back(SE->getSCEV(*Index));
338 // Replace the I-th index with LHS.
339 IndexExprs[I] = SE->getSCEV(LHS);
340 if (isKnownNonNegative(LHS, *DL, 0, AC, GEP, DT) &&
341 DL->getTypeSizeInBits(LHS->getType()) <
342 DL->getTypeSizeInBits(GEP->getOperand(I)->getType())) {
343 // Zero-extend LHS if it is non-negative. InstCombine canonicalizes sext to
344 // zext if the source operand is proved non-negative. We should do that
345 // consistently so that CandidateExpr more likely appears before. See
346 // @reassociate_gep_assume for an example of this canonicalization.
348 SE->getZeroExtendExpr(IndexExprs[I], GEP->getOperand(I)->getType());
350 const SCEV *CandidateExpr = SE->getGEPExpr(cast<GEPOperator>(GEP),
353 Value *Candidate = findClosestMatchingDominator(CandidateExpr, GEP);
354 if (Candidate == nullptr)
357 IRBuilder<> Builder(GEP);
358 // Candidate does not necessarily have the same pointer type as GEP. Use
359 // bitcast or pointer cast to make sure they have the same type, so that the
360 // later RAUW doesn't complain.
361 Candidate = Builder.CreateBitOrPointerCast(Candidate, GEP->getType());
362 assert(Candidate->getType() == GEP->getType());
364 // NewGEP = (char *)Candidate + RHS * sizeof(IndexedType)
365 uint64_t IndexedSize = DL->getTypeAllocSize(IndexedType);
366 Type *ElementType = GEP->getResultElementType();
367 uint64_t ElementSize = DL->getTypeAllocSize(ElementType);
368 // Another less rare case: because I is not necessarily the last index of the
369 // GEP, the size of the type at the I-th index (IndexedSize) is not
370 // necessarily divisible by ElementSize. For example,
379 // sizeof(S) = 100 is indivisible by sizeof(int64) = 8.
381 // TODO: bail out on this case for now. We could emit uglygep.
382 if (IndexedSize % ElementSize != 0)
385 // NewGEP = &Candidate[RHS * (sizeof(IndexedType) / sizeof(Candidate[0])));
386 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
387 if (RHS->getType() != IntPtrTy)
388 RHS = Builder.CreateSExtOrTrunc(RHS, IntPtrTy);
389 if (IndexedSize != ElementSize) {
390 RHS = Builder.CreateMul(
391 RHS, ConstantInt::get(IntPtrTy, IndexedSize / ElementSize));
393 GetElementPtrInst *NewGEP =
394 cast<GetElementPtrInst>(Builder.CreateGEP(Candidate, RHS));
395 NewGEP->setIsInBounds(GEP->isInBounds());
396 NewGEP->takeName(GEP);
400 Instruction *NaryReassociatePass::tryReassociateBinaryOp(BinaryOperator *I) {
401 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
402 if (auto *NewI = tryReassociateBinaryOp(LHS, RHS, I))
404 if (auto *NewI = tryReassociateBinaryOp(RHS, LHS, I))
409 Instruction *NaryReassociatePass::tryReassociateBinaryOp(Value *LHS, Value *RHS,
411 Value *A = nullptr, *B = nullptr;
412 // To be conservative, we reassociate I only when it is the only user of (A op
414 if (LHS->hasOneUse() && matchTernaryOp(I, LHS, A, B)) {
415 // I = (A op B) op RHS
416 // = (A op RHS) op B or (B op RHS) op A
417 const SCEV *AExpr = SE->getSCEV(A), *BExpr = SE->getSCEV(B);
418 const SCEV *RHSExpr = SE->getSCEV(RHS);
419 if (BExpr != RHSExpr) {
421 tryReassociatedBinaryOp(getBinarySCEV(I, AExpr, RHSExpr), B, I))
424 if (AExpr != RHSExpr) {
426 tryReassociatedBinaryOp(getBinarySCEV(I, BExpr, RHSExpr), A, I))
433 Instruction *NaryReassociatePass::tryReassociatedBinaryOp(const SCEV *LHSExpr,
436 // Look for the closest dominator LHS of I that computes LHSExpr, and replace
437 // I with LHS op RHS.
438 auto *LHS = findClosestMatchingDominator(LHSExpr, I);
442 Instruction *NewI = nullptr;
443 switch (I->getOpcode()) {
444 case Instruction::Add:
445 NewI = BinaryOperator::CreateAdd(LHS, RHS, "", I);
447 case Instruction::Mul:
448 NewI = BinaryOperator::CreateMul(LHS, RHS, "", I);
451 llvm_unreachable("Unexpected instruction.");
457 bool NaryReassociatePass::matchTernaryOp(BinaryOperator *I, Value *V,
458 Value *&Op1, Value *&Op2) {
459 switch (I->getOpcode()) {
460 case Instruction::Add:
461 return match(V, m_Add(m_Value(Op1), m_Value(Op2)));
462 case Instruction::Mul:
463 return match(V, m_Mul(m_Value(Op1), m_Value(Op2)));
465 llvm_unreachable("Unexpected instruction.");
470 const SCEV *NaryReassociatePass::getBinarySCEV(BinaryOperator *I,
473 switch (I->getOpcode()) {
474 case Instruction::Add:
475 return SE->getAddExpr(LHS, RHS);
476 case Instruction::Mul:
477 return SE->getMulExpr(LHS, RHS);
479 llvm_unreachable("Unexpected instruction.");
485 NaryReassociatePass::findClosestMatchingDominator(const SCEV *CandidateExpr,
486 Instruction *Dominatee) {
487 auto Pos = SeenExprs.find(CandidateExpr);
488 if (Pos == SeenExprs.end())
491 auto &Candidates = Pos->second;
492 // Because we process the basic blocks in pre-order of the dominator tree, a
493 // candidate that doesn't dominate the current instruction won't dominate any
494 // future instruction either. Therefore, we pop it out of the stack. This
495 // optimization makes the algorithm O(n).
496 while (!Candidates.empty()) {
497 // Candidates stores WeakVHs, so a candidate can be nullptr if it's removed
499 if (Value *Candidate = Candidates.back()) {
500 Instruction *CandidateInstruction = cast<Instruction>(Candidate);
501 if (DT->dominates(CandidateInstruction, Dominatee))
502 return CandidateInstruction;
504 Candidates.pop_back();