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/Analysis/AssumptionCache.h"
80 #include "llvm/Analysis/ScalarEvolution.h"
81 #include "llvm/Analysis/TargetLibraryInfo.h"
82 #include "llvm/Analysis/TargetTransformInfo.h"
83 #include "llvm/Analysis/ValueTracking.h"
84 #include "llvm/IR/Dominators.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/Local.h"
92 using namespace PatternMatch;
94 #define DEBUG_TYPE "nary-reassociate"
97 class NaryReassociate : public FunctionPass {
101 NaryReassociate(): FunctionPass(ID) {
102 initializeNaryReassociatePass(*PassRegistry::getPassRegistry());
105 bool doInitialization(Module &M) override {
106 DL = &M.getDataLayout();
109 bool runOnFunction(Function &F) override;
111 void getAnalysisUsage(AnalysisUsage &AU) const override {
112 AU.addPreserved<DominatorTreeWrapperPass>();
113 AU.addPreserved<ScalarEvolutionWrapperPass>();
114 AU.addPreserved<TargetLibraryInfoWrapperPass>();
115 AU.addRequired<AssumptionCacheTracker>();
116 AU.addRequired<DominatorTreeWrapperPass>();
117 AU.addRequired<ScalarEvolutionWrapperPass>();
118 AU.addRequired<TargetLibraryInfoWrapperPass>();
119 AU.addRequired<TargetTransformInfoWrapperPass>();
120 AU.setPreservesCFG();
124 // Runs only one iteration of the dominator-based algorithm. See the header
125 // comments for why we need multiple iterations.
126 bool doOneIteration(Function &F);
128 // Reassociates I for better CSE.
129 Instruction *tryReassociate(Instruction *I);
131 // Reassociate GEP for better CSE.
132 Instruction *tryReassociateGEP(GetElementPtrInst *GEP);
133 // Try splitting GEP at the I-th index and see whether either part can be
134 // CSE'ed. This is a helper function for tryReassociateGEP.
136 // \p IndexedType The element type indexed by GEP's I-th index. This is
138 // GEP->getIndexedType(GEP->getPointerOperand(), 0-th index,
140 GetElementPtrInst *tryReassociateGEPAtIndex(GetElementPtrInst *GEP,
141 unsigned I, Type *IndexedType);
142 // Given GEP's I-th index = LHS + RHS, see whether &Base[..][LHS][..] or
143 // &Base[..][RHS][..] can be CSE'ed and rewrite GEP accordingly.
144 GetElementPtrInst *tryReassociateGEPAtIndex(GetElementPtrInst *GEP,
145 unsigned I, Value *LHS,
146 Value *RHS, Type *IndexedType);
148 // Reassociate binary operators for better CSE.
149 Instruction *tryReassociateBinaryOp(BinaryOperator *I);
151 // A helper function for tryReassociateBinaryOp. LHS and RHS are explicitly
153 Instruction *tryReassociateBinaryOp(Value *LHS, Value *RHS,
155 // Rewrites I to (LHS op RHS) if LHS is computed already.
156 Instruction *tryReassociatedBinaryOp(const SCEV *LHS, Value *RHS,
159 // Tries to match Op1 and Op2 by using V.
160 bool matchTernaryOp(BinaryOperator *I, Value *V, Value *&Op1, Value *&Op2);
162 // Gets SCEV for (LHS op RHS).
163 const SCEV *getBinarySCEV(BinaryOperator *I, const SCEV *LHS,
166 // Returns the closest dominator of \c Dominatee that computes
167 // \c CandidateExpr. Returns null if not found.
168 Instruction *findClosestMatchingDominator(const SCEV *CandidateExpr,
169 Instruction *Dominatee);
170 // GetElementPtrInst implicitly sign-extends an index if the index is shorter
171 // than the pointer size. This function returns whether Index is shorter than
172 // GEP's pointer size, i.e., whether Index needs to be sign-extended in order
173 // to be an index of GEP.
174 bool requiresSignExtension(Value *Index, GetElementPtrInst *GEP);
177 const DataLayout *DL;
180 TargetLibraryInfo *TLI;
181 TargetTransformInfo *TTI;
182 // A lookup table quickly telling which instructions compute the given SCEV.
183 // Note that there can be multiple instructions at different locations
184 // computing to the same SCEV, so we map a SCEV to an instruction list. For
191 DenseMap<const SCEV *, SmallVector<WeakVH, 2>> SeenExprs;
193 } // anonymous namespace
195 char NaryReassociate::ID = 0;
196 INITIALIZE_PASS_BEGIN(NaryReassociate, "nary-reassociate", "Nary reassociation",
198 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
199 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
200 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
201 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
202 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
203 INITIALIZE_PASS_END(NaryReassociate, "nary-reassociate", "Nary reassociation",
206 FunctionPass *llvm::createNaryReassociatePass() {
207 return new NaryReassociate();
210 bool NaryReassociate::runOnFunction(Function &F) {
211 if (skipOptnoneFunction(F))
214 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
215 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
216 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
217 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
218 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
220 bool Changed = false, ChangedInThisIteration;
222 ChangedInThisIteration = doOneIteration(F);
223 Changed |= ChangedInThisIteration;
224 } while (ChangedInThisIteration);
228 // Whitelist the instruction types NaryReassociate handles for now.
229 static bool isPotentiallyNaryReassociable(Instruction *I) {
230 switch (I->getOpcode()) {
231 case Instruction::Add:
232 case Instruction::GetElementPtr:
233 case Instruction::Mul:
240 bool NaryReassociate::doOneIteration(Function &F) {
241 bool Changed = false;
243 // Process the basic blocks in pre-order of the dominator tree. This order
244 // ensures that all bases of a candidate are in Candidates when we process it.
245 for (auto Node = GraphTraits<DominatorTree *>::nodes_begin(DT);
246 Node != GraphTraits<DominatorTree *>::nodes_end(DT); ++Node) {
247 BasicBlock *BB = Node->getBlock();
248 for (auto I = BB->begin(); I != BB->end(); ++I) {
249 if (SE->isSCEVable(I->getType()) && isPotentiallyNaryReassociable(&*I)) {
250 const SCEV *OldSCEV = SE->getSCEV(&*I);
251 if (Instruction *NewI = tryReassociate(&*I)) {
253 SE->forgetValue(&*I);
254 I->replaceAllUsesWith(NewI);
255 // If SeenExprs constains I's WeakVH, that entry will be replaced with
257 RecursivelyDeleteTriviallyDeadInstructions(&*I, TLI);
258 I = NewI->getIterator();
260 // Add the rewritten instruction to SeenExprs; the original instruction
262 const SCEV *NewSCEV = SE->getSCEV(&*I);
263 SeenExprs[NewSCEV].push_back(WeakVH(&*I));
264 // Ideally, NewSCEV should equal OldSCEV because tryReassociate(I)
265 // is equivalent to I. However, ScalarEvolution::getSCEV may
266 // weaken nsw causing NewSCEV not to equal OldSCEV. For example, suppose
268 // I = &a[sext(i +nsw j)] // assuming sizeof(a[0]) = 4
270 // NewI = &a[sext(i)] + sext(j).
272 // ScalarEvolution computes
273 // getSCEV(I) = a + 4 * sext(i + j)
274 // getSCEV(newI) = a + 4 * sext(i) + 4 * sext(j)
275 // which are different SCEVs.
277 // To alleviate this issue of ScalarEvolution not always capturing
278 // equivalence, we add I to SeenExprs[OldSCEV] as well so that we can
279 // map both SCEV before and after tryReassociate(I) to I.
281 // This improvement is exercised in @reassociate_gep_nsw in nary-gep.ll.
282 if (NewSCEV != OldSCEV)
283 SeenExprs[OldSCEV].push_back(WeakVH(&*I));
290 Instruction *NaryReassociate::tryReassociate(Instruction *I) {
291 switch (I->getOpcode()) {
292 case Instruction::Add:
293 case Instruction::Mul:
294 return tryReassociateBinaryOp(cast<BinaryOperator>(I));
295 case Instruction::GetElementPtr:
296 return tryReassociateGEP(cast<GetElementPtrInst>(I));
298 llvm_unreachable("should be filtered out by isPotentiallyNaryReassociable");
302 // FIXME: extract this method into TTI->getGEPCost.
303 static bool isGEPFoldable(GetElementPtrInst *GEP,
304 const TargetTransformInfo *TTI,
305 const DataLayout *DL) {
306 GlobalVariable *BaseGV = nullptr;
307 int64_t BaseOffset = 0;
308 bool HasBaseReg = false;
311 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand()))
316 gep_type_iterator GTI = gep_type_begin(GEP);
317 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I, ++GTI) {
318 if (isa<SequentialType>(*GTI)) {
319 int64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
320 if (ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
321 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
323 // Needs scale register.
325 // No addressing mode takes two scale registers.
331 StructType *STy = cast<StructType>(*GTI);
332 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
333 BaseOffset += DL->getStructLayout(STy)->getElementOffset(Field);
337 unsigned AddrSpace = GEP->getPointerAddressSpace();
338 return TTI->isLegalAddressingMode(GEP->getType()->getElementType(), BaseGV,
339 BaseOffset, HasBaseReg, Scale, AddrSpace);
342 Instruction *NaryReassociate::tryReassociateGEP(GetElementPtrInst *GEP) {
343 // Not worth reassociating GEP if it is foldable.
344 if (isGEPFoldable(GEP, TTI, DL))
347 gep_type_iterator GTI = gep_type_begin(*GEP);
348 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I) {
349 if (isa<SequentialType>(*GTI++)) {
350 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I - 1, *GTI)) {
358 bool NaryReassociate::requiresSignExtension(Value *Index,
359 GetElementPtrInst *GEP) {
360 unsigned PointerSizeInBits =
361 DL->getPointerSizeInBits(GEP->getType()->getPointerAddressSpace());
362 return cast<IntegerType>(Index->getType())->getBitWidth() < PointerSizeInBits;
366 NaryReassociate::tryReassociateGEPAtIndex(GetElementPtrInst *GEP, unsigned I,
368 Value *IndexToSplit = GEP->getOperand(I + 1);
369 if (SExtInst *SExt = dyn_cast<SExtInst>(IndexToSplit)) {
370 IndexToSplit = SExt->getOperand(0);
371 } else if (ZExtInst *ZExt = dyn_cast<ZExtInst>(IndexToSplit)) {
372 // zext can be treated as sext if the source is non-negative.
373 if (isKnownNonNegative(ZExt->getOperand(0), *DL, 0, AC, GEP, DT))
374 IndexToSplit = ZExt->getOperand(0);
377 if (AddOperator *AO = dyn_cast<AddOperator>(IndexToSplit)) {
378 // If the I-th index needs sext and the underlying add is not equipped with
379 // nsw, we cannot split the add because
380 // sext(LHS + RHS) != sext(LHS) + sext(RHS).
381 if (requiresSignExtension(IndexToSplit, GEP) &&
382 computeOverflowForSignedAdd(AO, *DL, AC, GEP, DT) !=
383 OverflowResult::NeverOverflows)
386 Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);
387 // IndexToSplit = LHS + RHS.
388 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType))
390 // Symmetrically, try IndexToSplit = RHS + LHS.
393 tryReassociateGEPAtIndex(GEP, I, RHS, LHS, IndexedType))
400 GetElementPtrInst *NaryReassociate::tryReassociateGEPAtIndex(
401 GetElementPtrInst *GEP, unsigned I, Value *LHS, Value *RHS,
403 // Look for GEP's closest dominator that has the same SCEV as GEP except that
404 // the I-th index is replaced with LHS.
405 SmallVector<const SCEV *, 4> IndexExprs;
406 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
407 IndexExprs.push_back(SE->getSCEV(*Index));
408 // Replace the I-th index with LHS.
409 IndexExprs[I] = SE->getSCEV(LHS);
410 if (isKnownNonNegative(LHS, *DL, 0, AC, GEP, DT) &&
411 DL->getTypeSizeInBits(LHS->getType()) <
412 DL->getTypeSizeInBits(GEP->getOperand(I)->getType())) {
413 // Zero-extend LHS if it is non-negative. InstCombine canonicalizes sext to
414 // zext if the source operand is proved non-negative. We should do that
415 // consistently so that CandidateExpr more likely appears before. See
416 // @reassociate_gep_assume for an example of this canonicalization.
418 SE->getZeroExtendExpr(IndexExprs[I], GEP->getOperand(I)->getType());
420 const SCEV *CandidateExpr = SE->getGEPExpr(
421 GEP->getSourceElementType(), SE->getSCEV(GEP->getPointerOperand()),
422 IndexExprs, GEP->isInBounds());
424 auto *Candidate = findClosestMatchingDominator(CandidateExpr, GEP);
425 if (Candidate == nullptr)
428 PointerType *TypeOfCandidate = dyn_cast<PointerType>(Candidate->getType());
429 // Pretty rare but theoretically possible when a numeric value happens to
430 // share CandidateExpr.
431 if (TypeOfCandidate == nullptr)
434 // NewGEP = (char *)Candidate + RHS * sizeof(IndexedType)
435 uint64_t IndexedSize = DL->getTypeAllocSize(IndexedType);
436 Type *ElementType = TypeOfCandidate->getElementType();
437 uint64_t ElementSize = DL->getTypeAllocSize(ElementType);
438 // Another less rare case: because I is not necessarily the last index of the
439 // GEP, the size of the type at the I-th index (IndexedSize) is not
440 // necessarily divisible by ElementSize. For example,
449 // sizeof(S) = 100 is indivisible by sizeof(int64) = 8.
451 // TODO: bail out on this case for now. We could emit uglygep.
452 if (IndexedSize % ElementSize != 0)
455 // NewGEP = &Candidate[RHS * (sizeof(IndexedType) / sizeof(Candidate[0])));
456 IRBuilder<> Builder(GEP);
457 Type *IntPtrTy = DL->getIntPtrType(TypeOfCandidate);
458 if (RHS->getType() != IntPtrTy)
459 RHS = Builder.CreateSExtOrTrunc(RHS, IntPtrTy);
460 if (IndexedSize != ElementSize) {
461 RHS = Builder.CreateMul(
462 RHS, ConstantInt::get(IntPtrTy, IndexedSize / ElementSize));
464 GetElementPtrInst *NewGEP =
465 cast<GetElementPtrInst>(Builder.CreateGEP(Candidate, RHS));
466 NewGEP->setIsInBounds(GEP->isInBounds());
467 NewGEP->takeName(GEP);
471 Instruction *NaryReassociate::tryReassociateBinaryOp(BinaryOperator *I) {
472 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
473 if (auto *NewI = tryReassociateBinaryOp(LHS, RHS, I))
475 if (auto *NewI = tryReassociateBinaryOp(RHS, LHS, I))
480 Instruction *NaryReassociate::tryReassociateBinaryOp(Value *LHS, Value *RHS,
482 Value *A = nullptr, *B = nullptr;
483 // To be conservative, we reassociate I only when it is the only user of (A op
485 if (LHS->hasOneUse() && matchTernaryOp(I, LHS, A, B)) {
486 // I = (A op B) op RHS
487 // = (A op RHS) op B or (B op RHS) op A
488 const SCEV *AExpr = SE->getSCEV(A), *BExpr = SE->getSCEV(B);
489 const SCEV *RHSExpr = SE->getSCEV(RHS);
490 if (BExpr != RHSExpr) {
492 tryReassociatedBinaryOp(getBinarySCEV(I, AExpr, RHSExpr), B, I))
495 if (AExpr != RHSExpr) {
497 tryReassociatedBinaryOp(getBinarySCEV(I, BExpr, RHSExpr), A, I))
504 Instruction *NaryReassociate::tryReassociatedBinaryOp(const SCEV *LHSExpr,
507 // Look for the closest dominator LHS of I that computes LHSExpr, and replace
508 // I with LHS op RHS.
509 auto *LHS = findClosestMatchingDominator(LHSExpr, I);
513 Instruction *NewI = nullptr;
514 switch (I->getOpcode()) {
515 case Instruction::Add:
516 NewI = BinaryOperator::CreateAdd(LHS, RHS, "", I);
518 case Instruction::Mul:
519 NewI = BinaryOperator::CreateMul(LHS, RHS, "", I);
522 llvm_unreachable("Unexpected instruction.");
528 bool NaryReassociate::matchTernaryOp(BinaryOperator *I, Value *V, Value *&Op1,
530 switch (I->getOpcode()) {
531 case Instruction::Add:
532 return match(V, m_Add(m_Value(Op1), m_Value(Op2)));
533 case Instruction::Mul:
534 return match(V, m_Mul(m_Value(Op1), m_Value(Op2)));
536 llvm_unreachable("Unexpected instruction.");
541 const SCEV *NaryReassociate::getBinarySCEV(BinaryOperator *I, const SCEV *LHS,
543 switch (I->getOpcode()) {
544 case Instruction::Add:
545 return SE->getAddExpr(LHS, RHS);
546 case Instruction::Mul:
547 return SE->getMulExpr(LHS, RHS);
549 llvm_unreachable("Unexpected instruction.");
555 NaryReassociate::findClosestMatchingDominator(const SCEV *CandidateExpr,
556 Instruction *Dominatee) {
557 auto Pos = SeenExprs.find(CandidateExpr);
558 if (Pos == SeenExprs.end())
561 auto &Candidates = Pos->second;
562 // Because we process the basic blocks in pre-order of the dominator tree, a
563 // candidate that doesn't dominate the current instruction won't dominate any
564 // future instruction either. Therefore, we pop it out of the stack. This
565 // optimization makes the algorithm O(n).
566 while (!Candidates.empty()) {
567 // Candidates stores WeakVHs, so a candidate can be nullptr if it's removed
569 if (Value *Candidate = Candidates.back()) {
570 Instruction *CandidateInstruction = cast<Instruction>(Candidate);
571 if (DT->dominates(CandidateInstruction, Dominatee))
572 return CandidateInstruction;
574 Candidates.pop_back();