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 for now. This should be extended and
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 Add for better CSE.
149 Instruction *tryReassociateAdd(BinaryOperator *I);
150 // A helper function for tryReassociateAdd. LHS and RHS are explicitly passed.
151 Instruction *tryReassociateAdd(Value *LHS, Value *RHS, Instruction *I);
152 // Rewrites I to LHS + RHS if LHS is computed already.
153 Instruction *tryReassociatedAdd(const SCEV *LHS, Value *RHS, Instruction *I);
155 // Returns the closest dominator of \c Dominatee that computes
156 // \c CandidateExpr. Returns null if not found.
157 Instruction *findClosestMatchingDominator(const SCEV *CandidateExpr,
158 Instruction *Dominatee);
159 // GetElementPtrInst implicitly sign-extends an index if the index is shorter
160 // than the pointer size. This function returns whether Index is shorter than
161 // GEP's pointer size, i.e., whether Index needs to be sign-extended in order
162 // to be an index of GEP.
163 bool requiresSignExtension(Value *Index, GetElementPtrInst *GEP);
164 // Returns whether V is known to be non-negative at context \c Ctxt.
165 bool isKnownNonNegative(Value *V, Instruction *Ctxt);
166 // Returns whether AO may sign overflow at context \c Ctxt. It computes a
167 // conservative result -- it answers true when not sure.
168 bool maySignOverflow(AddOperator *AO, Instruction *Ctxt);
171 const DataLayout *DL;
174 TargetLibraryInfo *TLI;
175 TargetTransformInfo *TTI;
176 // A lookup table quickly telling which instructions compute the given SCEV.
177 // Note that there can be multiple instructions at different locations
178 // computing to the same SCEV, so we map a SCEV to an instruction list. For
185 DenseMap<const SCEV *, SmallVector<Instruction *, 2>> SeenExprs;
187 } // anonymous namespace
189 char NaryReassociate::ID = 0;
190 INITIALIZE_PASS_BEGIN(NaryReassociate, "nary-reassociate", "Nary reassociation",
192 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
193 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
194 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
195 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
196 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
197 INITIALIZE_PASS_END(NaryReassociate, "nary-reassociate", "Nary reassociation",
200 FunctionPass *llvm::createNaryReassociatePass() {
201 return new NaryReassociate();
204 bool NaryReassociate::runOnFunction(Function &F) {
205 if (skipOptnoneFunction(F))
208 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
209 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
210 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
211 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
212 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
214 bool Changed = false, ChangedInThisIteration;
216 ChangedInThisIteration = doOneIteration(F);
217 Changed |= ChangedInThisIteration;
218 } while (ChangedInThisIteration);
222 // Whitelist the instruction types NaryReassociate handles for now.
223 static bool isPotentiallyNaryReassociable(Instruction *I) {
224 switch (I->getOpcode()) {
225 case Instruction::Add:
226 case Instruction::GetElementPtr:
233 bool NaryReassociate::doOneIteration(Function &F) {
234 bool Changed = false;
236 // Process the basic blocks in pre-order of the dominator tree. This order
237 // ensures that all bases of a candidate are in Candidates when we process it.
238 for (auto Node = GraphTraits<DominatorTree *>::nodes_begin(DT);
239 Node != GraphTraits<DominatorTree *>::nodes_end(DT); ++Node) {
240 BasicBlock *BB = Node->getBlock();
241 for (auto I = BB->begin(); I != BB->end(); ++I) {
242 if (SE->isSCEVable(I->getType()) && isPotentiallyNaryReassociable(I)) {
243 const SCEV *OldSCEV = SE->getSCEV(I);
244 if (Instruction *NewI = tryReassociate(I)) {
247 I->replaceAllUsesWith(NewI);
248 RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
251 // Add the rewritten instruction to SeenExprs; the original instruction
253 const SCEV *NewSCEV = SE->getSCEV(I);
254 SeenExprs[NewSCEV].push_back(I);
255 // Ideally, NewSCEV should equal OldSCEV because tryReassociate(I)
256 // is equivalent to I. However, ScalarEvolution::getSCEV may
257 // weaken nsw causing NewSCEV not to equal OldSCEV. For example, suppose
259 // I = &a[sext(i +nsw j)] // assuming sizeof(a[0]) = 4
261 // NewI = &a[sext(i)] + sext(j).
263 // ScalarEvolution computes
264 // getSCEV(I) = a + 4 * sext(i + j)
265 // getSCEV(newI) = a + 4 * sext(i) + 4 * sext(j)
266 // which are different SCEVs.
268 // To alleviate this issue of ScalarEvolution not always capturing
269 // equivalence, we add I to SeenExprs[OldSCEV] as well so that we can
270 // map both SCEV before and after tryReassociate(I) to I.
272 // This improvement is exercised in @reassociate_gep_nsw in nary-gep.ll.
273 if (NewSCEV != OldSCEV)
274 SeenExprs[OldSCEV].push_back(I);
281 Instruction *NaryReassociate::tryReassociate(Instruction *I) {
282 switch (I->getOpcode()) {
283 case Instruction::Add:
284 return tryReassociateAdd(cast<BinaryOperator>(I));
285 case Instruction::GetElementPtr:
286 return tryReassociateGEP(cast<GetElementPtrInst>(I));
288 llvm_unreachable("should be filtered out by isPotentiallyNaryReassociable");
292 // FIXME: extract this method into TTI->getGEPCost.
293 static bool isGEPFoldable(GetElementPtrInst *GEP,
294 const TargetTransformInfo *TTI,
295 const DataLayout *DL) {
296 GlobalVariable *BaseGV = nullptr;
297 int64_t BaseOffset = 0;
298 bool HasBaseReg = false;
301 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand()))
306 gep_type_iterator GTI = gep_type_begin(GEP);
307 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I, ++GTI) {
308 if (isa<SequentialType>(*GTI)) {
309 int64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
310 if (ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
311 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
313 // Needs scale register.
315 // No addressing mode takes two scale registers.
321 StructType *STy = cast<StructType>(*GTI);
322 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
323 BaseOffset += DL->getStructLayout(STy)->getElementOffset(Field);
327 unsigned AddrSpace = GEP->getPointerAddressSpace();
328 return TTI->isLegalAddressingMode(GEP->getType()->getElementType(), BaseGV,
329 BaseOffset, HasBaseReg, Scale, AddrSpace);
332 Instruction *NaryReassociate::tryReassociateGEP(GetElementPtrInst *GEP) {
333 // Not worth reassociating GEP if it is foldable.
334 if (isGEPFoldable(GEP, TTI, DL))
337 gep_type_iterator GTI = gep_type_begin(*GEP);
338 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I) {
339 if (isa<SequentialType>(*GTI++)) {
340 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I - 1, *GTI)) {
348 bool NaryReassociate::requiresSignExtension(Value *Index,
349 GetElementPtrInst *GEP) {
350 unsigned PointerSizeInBits =
351 DL->getPointerSizeInBits(GEP->getType()->getPointerAddressSpace());
352 return cast<IntegerType>(Index->getType())->getBitWidth() < PointerSizeInBits;
355 bool NaryReassociate::isKnownNonNegative(Value *V, Instruction *Ctxt) {
356 bool NonNegative, Negative;
357 // TODO: ComputeSignBits is expensive. Consider caching the results.
358 ComputeSignBit(V, NonNegative, Negative, *DL, 0, AC, Ctxt, DT);
362 bool NaryReassociate::maySignOverflow(AddOperator *AO, Instruction *Ctxt) {
363 if (AO->hasNoSignedWrap())
366 Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);
367 // If LHS or RHS has the same sign as the sum, AO doesn't sign overflow.
368 // TODO: handle the negative case as well.
369 if (isKnownNonNegative(AO, Ctxt) &&
370 (isKnownNonNegative(LHS, Ctxt) || isKnownNonNegative(RHS, Ctxt)))
377 NaryReassociate::tryReassociateGEPAtIndex(GetElementPtrInst *GEP, unsigned I,
379 Value *IndexToSplit = GEP->getOperand(I + 1);
380 if (SExtInst *SExt = dyn_cast<SExtInst>(IndexToSplit)) {
381 IndexToSplit = SExt->getOperand(0);
382 } else if (ZExtInst *ZExt = dyn_cast<ZExtInst>(IndexToSplit)) {
383 // zext can be treated as sext if the source is non-negative.
384 if (isKnownNonNegative(ZExt->getOperand(0), GEP))
385 IndexToSplit = ZExt->getOperand(0);
388 if (AddOperator *AO = dyn_cast<AddOperator>(IndexToSplit)) {
389 // If the I-th index needs sext and the underlying add is not equipped with
390 // nsw, we cannot split the add because
391 // sext(LHS + RHS) != sext(LHS) + sext(RHS).
392 if (requiresSignExtension(IndexToSplit, GEP) && maySignOverflow(AO, GEP))
394 Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);
395 // IndexToSplit = LHS + RHS.
396 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType))
398 // Symmetrically, try IndexToSplit = RHS + LHS.
401 tryReassociateGEPAtIndex(GEP, I, RHS, LHS, IndexedType))
408 GetElementPtrInst *NaryReassociate::tryReassociateGEPAtIndex(
409 GetElementPtrInst *GEP, unsigned I, Value *LHS, Value *RHS,
411 // Look for GEP's closest dominator that has the same SCEV as GEP except that
412 // the I-th index is replaced with LHS.
413 SmallVector<const SCEV *, 4> IndexExprs;
414 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
415 IndexExprs.push_back(SE->getSCEV(*Index));
416 // Replace the I-th index with LHS.
417 IndexExprs[I] = SE->getSCEV(LHS);
418 if (isKnownNonNegative(LHS, GEP) &&
419 DL->getTypeSizeInBits(LHS->getType()) <
420 DL->getTypeSizeInBits(GEP->getOperand(I)->getType())) {
421 // Zero-extend LHS if it is non-negative. InstCombine canonicalizes sext to
422 // zext if the source operand is proved non-negative. We should do that
423 // consistently so that CandidateExpr more likely appears before. See
424 // @reassociate_gep_assume for an example of this canonicalization.
426 SE->getZeroExtendExpr(IndexExprs[I], GEP->getOperand(I)->getType());
428 const SCEV *CandidateExpr = SE->getGEPExpr(
429 GEP->getSourceElementType(), SE->getSCEV(GEP->getPointerOperand()),
430 IndexExprs, GEP->isInBounds());
432 auto *Candidate = findClosestMatchingDominator(CandidateExpr, GEP);
433 if (Candidate == nullptr)
436 PointerType *TypeOfCandidate = dyn_cast<PointerType>(Candidate->getType());
437 // Pretty rare but theoretically possible when a numeric value happens to
438 // share CandidateExpr.
439 if (TypeOfCandidate == nullptr)
442 // NewGEP = (char *)Candidate + RHS * sizeof(IndexedType)
443 uint64_t IndexedSize = DL->getTypeAllocSize(IndexedType);
444 Type *ElementType = TypeOfCandidate->getElementType();
445 uint64_t ElementSize = DL->getTypeAllocSize(ElementType);
446 // Another less rare case: because I is not necessarily the last index of the
447 // GEP, the size of the type at the I-th index (IndexedSize) is not
448 // necessarily divisible by ElementSize. For example,
457 // sizeof(S) = 100 is indivisible by sizeof(int64) = 8.
459 // TODO: bail out on this case for now. We could emit uglygep.
460 if (IndexedSize % ElementSize != 0)
463 // NewGEP = &Candidate[RHS * (sizeof(IndexedType) / sizeof(Candidate[0])));
464 IRBuilder<> Builder(GEP);
465 Type *IntPtrTy = DL->getIntPtrType(TypeOfCandidate);
466 if (RHS->getType() != IntPtrTy)
467 RHS = Builder.CreateSExtOrTrunc(RHS, IntPtrTy);
468 if (IndexedSize != ElementSize) {
469 RHS = Builder.CreateMul(
470 RHS, ConstantInt::get(IntPtrTy, IndexedSize / ElementSize));
472 GetElementPtrInst *NewGEP =
473 cast<GetElementPtrInst>(Builder.CreateGEP(Candidate, RHS));
474 NewGEP->setIsInBounds(GEP->isInBounds());
475 NewGEP->takeName(GEP);
479 Instruction *NaryReassociate::tryReassociateAdd(BinaryOperator *I) {
480 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
481 if (auto *NewI = tryReassociateAdd(LHS, RHS, I))
483 if (auto *NewI = tryReassociateAdd(RHS, LHS, I))
488 Instruction *NaryReassociate::tryReassociateAdd(Value *LHS, Value *RHS,
490 Value *A = nullptr, *B = nullptr;
491 // To be conservative, we reassociate I only when it is the only user of A+B.
492 if (LHS->hasOneUse() && match(LHS, m_Add(m_Value(A), m_Value(B)))) {
494 // = (A + RHS) + B or (B + RHS) + A
495 const SCEV *AExpr = SE->getSCEV(A), *BExpr = SE->getSCEV(B);
496 const SCEV *RHSExpr = SE->getSCEV(RHS);
497 if (BExpr != RHSExpr) {
498 if (auto *NewI = tryReassociatedAdd(SE->getAddExpr(AExpr, RHSExpr), B, I))
501 if (AExpr != RHSExpr) {
502 if (auto *NewI = tryReassociatedAdd(SE->getAddExpr(BExpr, RHSExpr), A, I))
509 Instruction *NaryReassociate::tryReassociatedAdd(const SCEV *LHSExpr,
510 Value *RHS, Instruction *I) {
511 // Look for the closest dominator LHS of I that computes LHSExpr, and replace
513 auto *LHS = findClosestMatchingDominator(LHSExpr, I);
517 Instruction *NewI = BinaryOperator::CreateAdd(LHS, RHS, "", I);
523 NaryReassociate::findClosestMatchingDominator(const SCEV *CandidateExpr,
524 Instruction *Dominatee) {
525 auto Pos = SeenExprs.find(CandidateExpr);
526 if (Pos == SeenExprs.end())
529 auto &Candidates = Pos->second;
530 // Because we process the basic blocks in pre-order of the dominator tree, a
531 // candidate that doesn't dominate the current instruction won't dominate any
532 // future instruction either. Therefore, we pop it out of the stack. This
533 // optimization makes the algorithm O(n).
534 while (!Candidates.empty()) {
535 Instruction *Candidate = Candidates.back();
536 if (DT->dominates(Candidate, Dominatee))
538 Candidates.pop_back();