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);
166 const DataLayout *DL;
169 TargetLibraryInfo *TLI;
170 TargetTransformInfo *TTI;
171 // A lookup table quickly telling which instructions compute the given SCEV.
172 // Note that there can be multiple instructions at different locations
173 // computing to the same SCEV, so we map a SCEV to an instruction list. For
180 DenseMap<const SCEV *, SmallVector<Instruction *, 2>> SeenExprs;
182 } // anonymous namespace
184 char NaryReassociate::ID = 0;
185 INITIALIZE_PASS_BEGIN(NaryReassociate, "nary-reassociate", "Nary reassociation",
187 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
188 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
189 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
190 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
191 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
192 INITIALIZE_PASS_END(NaryReassociate, "nary-reassociate", "Nary reassociation",
195 FunctionPass *llvm::createNaryReassociatePass() {
196 return new NaryReassociate();
199 bool NaryReassociate::runOnFunction(Function &F) {
200 if (skipOptnoneFunction(F))
203 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
204 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
205 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
206 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
207 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
209 bool Changed = false, ChangedInThisIteration;
211 ChangedInThisIteration = doOneIteration(F);
212 Changed |= ChangedInThisIteration;
213 } while (ChangedInThisIteration);
217 // Whitelist the instruction types NaryReassociate handles for now.
218 static bool isPotentiallyNaryReassociable(Instruction *I) {
219 switch (I->getOpcode()) {
220 case Instruction::Add:
221 case Instruction::GetElementPtr:
228 bool NaryReassociate::doOneIteration(Function &F) {
229 bool Changed = false;
231 // Process the basic blocks in pre-order of the dominator tree. This order
232 // ensures that all bases of a candidate are in Candidates when we process it.
233 for (auto Node = GraphTraits<DominatorTree *>::nodes_begin(DT);
234 Node != GraphTraits<DominatorTree *>::nodes_end(DT); ++Node) {
235 BasicBlock *BB = Node->getBlock();
236 for (auto I = BB->begin(); I != BB->end(); ++I) {
237 if (SE->isSCEVable(I->getType()) && isPotentiallyNaryReassociable(I)) {
238 const SCEV *OldSCEV = SE->getSCEV(I);
239 if (Instruction *NewI = tryReassociate(I)) {
242 I->replaceAllUsesWith(NewI);
243 RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
246 // Add the rewritten instruction to SeenExprs; the original instruction
248 const SCEV *NewSCEV = SE->getSCEV(I);
249 SeenExprs[NewSCEV].push_back(I);
250 // Ideally, NewSCEV should equal OldSCEV because tryReassociate(I)
251 // is equivalent to I. However, ScalarEvolution::getSCEV may
252 // weaken nsw causing NewSCEV not to equal OldSCEV. For example, suppose
254 // I = &a[sext(i +nsw j)] // assuming sizeof(a[0]) = 4
256 // NewI = &a[sext(i)] + sext(j).
258 // ScalarEvolution computes
259 // getSCEV(I) = a + 4 * sext(i + j)
260 // getSCEV(newI) = a + 4 * sext(i) + 4 * sext(j)
261 // which are different SCEVs.
263 // To alleviate this issue of ScalarEvolution not always capturing
264 // equivalence, we add I to SeenExprs[OldSCEV] as well so that we can
265 // map both SCEV before and after tryReassociate(I) to I.
267 // This improvement is exercised in @reassociate_gep_nsw in nary-gep.ll.
268 if (NewSCEV != OldSCEV)
269 SeenExprs[OldSCEV].push_back(I);
276 Instruction *NaryReassociate::tryReassociate(Instruction *I) {
277 switch (I->getOpcode()) {
278 case Instruction::Add:
279 return tryReassociateAdd(cast<BinaryOperator>(I));
280 case Instruction::GetElementPtr:
281 return tryReassociateGEP(cast<GetElementPtrInst>(I));
283 llvm_unreachable("should be filtered out by isPotentiallyNaryReassociable");
287 // FIXME: extract this method into TTI->getGEPCost.
288 static bool isGEPFoldable(GetElementPtrInst *GEP,
289 const TargetTransformInfo *TTI,
290 const DataLayout *DL) {
291 GlobalVariable *BaseGV = nullptr;
292 int64_t BaseOffset = 0;
293 bool HasBaseReg = false;
296 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand()))
301 gep_type_iterator GTI = gep_type_begin(GEP);
302 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I, ++GTI) {
303 if (isa<SequentialType>(*GTI)) {
304 int64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
305 if (ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
306 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
308 // Needs scale register.
310 // No addressing mode takes two scale registers.
316 StructType *STy = cast<StructType>(*GTI);
317 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
318 BaseOffset += DL->getStructLayout(STy)->getElementOffset(Field);
322 unsigned AddrSpace = GEP->getPointerAddressSpace();
323 return TTI->isLegalAddressingMode(GEP->getType()->getElementType(), BaseGV,
324 BaseOffset, HasBaseReg, Scale, AddrSpace);
327 Instruction *NaryReassociate::tryReassociateGEP(GetElementPtrInst *GEP) {
328 // Not worth reassociating GEP if it is foldable.
329 if (isGEPFoldable(GEP, TTI, DL))
332 gep_type_iterator GTI = gep_type_begin(*GEP);
333 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I) {
334 if (isa<SequentialType>(*GTI++)) {
335 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I - 1, *GTI)) {
343 bool NaryReassociate::requiresSignExtension(Value *Index,
344 GetElementPtrInst *GEP) {
345 unsigned PointerSizeInBits =
346 DL->getPointerSizeInBits(GEP->getType()->getPointerAddressSpace());
347 return cast<IntegerType>(Index->getType())->getBitWidth() < PointerSizeInBits;
351 NaryReassociate::tryReassociateGEPAtIndex(GetElementPtrInst *GEP, unsigned I,
353 Value *IndexToSplit = GEP->getOperand(I + 1);
354 if (SExtInst *SExt = dyn_cast<SExtInst>(IndexToSplit)) {
355 IndexToSplit = SExt->getOperand(0);
356 } else if (ZExtInst *ZExt = dyn_cast<ZExtInst>(IndexToSplit)) {
357 // zext can be treated as sext if the source is non-negative.
358 if (isKnownNonNegative(ZExt->getOperand(0), *DL, 0, AC, GEP, DT))
359 IndexToSplit = ZExt->getOperand(0);
362 if (AddOperator *AO = dyn_cast<AddOperator>(IndexToSplit)) {
363 // If the I-th index needs sext and the underlying add is not equipped with
364 // nsw, we cannot split the add because
365 // sext(LHS + RHS) != sext(LHS) + sext(RHS).
366 if (requiresSignExtension(IndexToSplit, GEP) &&
367 computeOverflowForSignedAdd(AO, *DL, AC, GEP, DT) !=
368 OverflowResult::NeverOverflows)
371 Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);
372 // IndexToSplit = LHS + RHS.
373 if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType))
375 // Symmetrically, try IndexToSplit = RHS + LHS.
378 tryReassociateGEPAtIndex(GEP, I, RHS, LHS, IndexedType))
385 GetElementPtrInst *NaryReassociate::tryReassociateGEPAtIndex(
386 GetElementPtrInst *GEP, unsigned I, Value *LHS, Value *RHS,
388 // Look for GEP's closest dominator that has the same SCEV as GEP except that
389 // the I-th index is replaced with LHS.
390 SmallVector<const SCEV *, 4> IndexExprs;
391 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
392 IndexExprs.push_back(SE->getSCEV(*Index));
393 // Replace the I-th index with LHS.
394 IndexExprs[I] = SE->getSCEV(LHS);
395 if (isKnownNonNegative(LHS, *DL, 0, AC, GEP, DT) &&
396 DL->getTypeSizeInBits(LHS->getType()) <
397 DL->getTypeSizeInBits(GEP->getOperand(I)->getType())) {
398 // Zero-extend LHS if it is non-negative. InstCombine canonicalizes sext to
399 // zext if the source operand is proved non-negative. We should do that
400 // consistently so that CandidateExpr more likely appears before. See
401 // @reassociate_gep_assume for an example of this canonicalization.
403 SE->getZeroExtendExpr(IndexExprs[I], GEP->getOperand(I)->getType());
405 const SCEV *CandidateExpr = SE->getGEPExpr(
406 GEP->getSourceElementType(), SE->getSCEV(GEP->getPointerOperand()),
407 IndexExprs, GEP->isInBounds());
409 auto *Candidate = findClosestMatchingDominator(CandidateExpr, GEP);
410 if (Candidate == nullptr)
413 PointerType *TypeOfCandidate = dyn_cast<PointerType>(Candidate->getType());
414 // Pretty rare but theoretically possible when a numeric value happens to
415 // share CandidateExpr.
416 if (TypeOfCandidate == nullptr)
419 // NewGEP = (char *)Candidate + RHS * sizeof(IndexedType)
420 uint64_t IndexedSize = DL->getTypeAllocSize(IndexedType);
421 Type *ElementType = TypeOfCandidate->getElementType();
422 uint64_t ElementSize = DL->getTypeAllocSize(ElementType);
423 // Another less rare case: because I is not necessarily the last index of the
424 // GEP, the size of the type at the I-th index (IndexedSize) is not
425 // necessarily divisible by ElementSize. For example,
434 // sizeof(S) = 100 is indivisible by sizeof(int64) = 8.
436 // TODO: bail out on this case for now. We could emit uglygep.
437 if (IndexedSize % ElementSize != 0)
440 // NewGEP = &Candidate[RHS * (sizeof(IndexedType) / sizeof(Candidate[0])));
441 IRBuilder<> Builder(GEP);
442 Type *IntPtrTy = DL->getIntPtrType(TypeOfCandidate);
443 if (RHS->getType() != IntPtrTy)
444 RHS = Builder.CreateSExtOrTrunc(RHS, IntPtrTy);
445 if (IndexedSize != ElementSize) {
446 RHS = Builder.CreateMul(
447 RHS, ConstantInt::get(IntPtrTy, IndexedSize / ElementSize));
449 GetElementPtrInst *NewGEP =
450 cast<GetElementPtrInst>(Builder.CreateGEP(Candidate, RHS));
451 NewGEP->setIsInBounds(GEP->isInBounds());
452 NewGEP->takeName(GEP);
456 Instruction *NaryReassociate::tryReassociateAdd(BinaryOperator *I) {
457 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
458 if (auto *NewI = tryReassociateAdd(LHS, RHS, I))
460 if (auto *NewI = tryReassociateAdd(RHS, LHS, I))
465 Instruction *NaryReassociate::tryReassociateAdd(Value *LHS, Value *RHS,
467 Value *A = nullptr, *B = nullptr;
468 // To be conservative, we reassociate I only when it is the only user of A+B.
469 if (LHS->hasOneUse() && match(LHS, m_Add(m_Value(A), m_Value(B)))) {
471 // = (A + RHS) + B or (B + RHS) + A
472 const SCEV *AExpr = SE->getSCEV(A), *BExpr = SE->getSCEV(B);
473 const SCEV *RHSExpr = SE->getSCEV(RHS);
474 if (BExpr != RHSExpr) {
475 if (auto *NewI = tryReassociatedAdd(SE->getAddExpr(AExpr, RHSExpr), B, I))
478 if (AExpr != RHSExpr) {
479 if (auto *NewI = tryReassociatedAdd(SE->getAddExpr(BExpr, RHSExpr), A, I))
486 Instruction *NaryReassociate::tryReassociatedAdd(const SCEV *LHSExpr,
487 Value *RHS, Instruction *I) {
488 // Look for the closest dominator LHS of I that computes LHSExpr, and replace
490 auto *LHS = findClosestMatchingDominator(LHSExpr, I);
494 Instruction *NewI = BinaryOperator::CreateAdd(LHS, RHS, "", I);
500 NaryReassociate::findClosestMatchingDominator(const SCEV *CandidateExpr,
501 Instruction *Dominatee) {
502 auto Pos = SeenExprs.find(CandidateExpr);
503 if (Pos == SeenExprs.end())
506 auto &Candidates = Pos->second;
507 // Because we process the basic blocks in pre-order of the dominator tree, a
508 // candidate that doesn't dominate the current instruction won't dominate any
509 // future instruction either. Therefore, we pop it out of the stack. This
510 // optimization makes the algorithm O(n).
511 while (!Candidates.empty()) {
512 Instruction *Candidate = Candidates.back();
513 if (DT->dominates(Candidate, Dominatee))
515 Candidates.pop_back();