1 //===-- StraightLineStrengthReduce.cpp - ------------------------*- C++ -*-===//
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 file implements straight-line strength reduction (SLSR). Unlike loop
11 // strength reduction, this algorithm is designed to reduce arithmetic
12 // redundancy in straight-line code instead of loops. It has proven to be
13 // effective in simplifying arithmetic statements derived from an unrolled loop.
14 // It can also simplify the logic of SeparateConstOffsetFromGEP.
16 // There are many optimizations we can perform in the domain of SLSR. This file
17 // for now contains only an initial step. Specifically, we look for strength
18 // reduction candidates in the following forms:
21 // Form 2: (B + i) * S
24 // where S is an integer variable, and i is a constant integer. If we found two
25 // candidates S1 and S2 in the same form and S1 dominates S2, we may rewrite S2
26 // in a simpler way with respect to S1. For example,
29 // S2: Y = B + i' * S => X + (i' - i) * S
31 // S1: X = (B + i) * S
32 // S2: Y = (B + i') * S => X + (i' - i) * S
35 // S2: Y = &B[i' * S] => &X[(i' - i) * S]
37 // Note: (i' - i) * S is folded to the extent possible.
39 // This rewriting is in general a good idea. The code patterns we focus on
40 // usually come from loop unrolling, so (i' - i) * S is likely the same
41 // across iterations and can be reused. When that happens, the optimized form
42 // takes only one add starting from the second iteration.
44 // When such rewriting is possible, we call S1 a "basis" of S2. When S2 has
45 // multiple bases, we choose to rewrite S2 with respect to its "immediate"
46 // basis, the basis that is the closest ancestor in the dominator tree.
50 // - Floating point arithmetics when fast math is enabled.
52 // - SLSR may decrease ILP at the architecture level. Targets that are very
53 // sensitive to ILP may want to disable it. Having SLSR to consider ILP is
54 // left as future work.
56 // - When (i' - i) is constant but i and i' are not, we could still perform
60 #include "llvm/ADT/DenseSet.h"
61 #include "llvm/ADT/FoldingSet.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/TargetTransformInfo.h"
64 #include "llvm/Analysis/ValueTracking.h"
65 #include "llvm/IR/DataLayout.h"
66 #include "llvm/IR/Dominators.h"
67 #include "llvm/IR/IRBuilder.h"
68 #include "llvm/IR/Module.h"
69 #include "llvm/IR/PatternMatch.h"
70 #include "llvm/Support/raw_ostream.h"
71 #include "llvm/Transforms/Scalar.h"
72 #include "llvm/Transforms/Utils/Local.h"
75 using namespace PatternMatch;
79 class StraightLineStrengthReduce : public FunctionPass {
81 // SLSR candidate. Such a candidate must be in one of the forms described in
82 // the header comments.
83 struct Candidate : public ilist_node<Candidate> {
85 Invalid, // reserved for the default constructor
88 GEP, // &B[..][i * S][..]
92 : CandidateKind(Invalid), Base(nullptr), Index(nullptr),
93 Stride(nullptr), Ins(nullptr), Basis(nullptr) {}
94 Candidate(Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
96 : CandidateKind(CT), Base(B), Index(Idx), Stride(S), Ins(I),
100 // Note that Index and Stride of a GEP candidate do not necessarily have the
101 // same integer type. In that case, during rewriting, Stride will be
102 // sign-extended or truncated to Index's type.
105 // The instruction this candidate corresponds to. It helps us to rewrite a
106 // candidate with respect to its immediate basis. Note that one instruction
107 // can correspond to multiple candidates depending on how you associate the
108 // expression. For instance,
114 // <Base: a, Index: 1, Stride: b + 2>
118 // <Base: b, Index: 2, Stride: a + 1>
120 // Points to the immediate basis of this candidate, or nullptr if we cannot
121 // find any basis for this candidate.
127 StraightLineStrengthReduce()
128 : FunctionPass(ID), DL(nullptr), DT(nullptr), TTI(nullptr) {
129 initializeStraightLineStrengthReducePass(*PassRegistry::getPassRegistry());
132 void getAnalysisUsage(AnalysisUsage &AU) const override {
133 AU.addRequired<DominatorTreeWrapperPass>();
134 AU.addRequired<ScalarEvolution>();
135 AU.addRequired<TargetTransformInfoWrapperPass>();
136 // We do not modify the shape of the CFG.
137 AU.setPreservesCFG();
140 bool doInitialization(Module &M) override {
141 DL = &M.getDataLayout();
145 bool runOnFunction(Function &F) override;
148 // Returns true if Basis is a basis for C, i.e., Basis dominates C and they
149 // share the same base and stride.
150 bool isBasisFor(const Candidate &Basis, const Candidate &C);
151 // Returns whether the candidate can be folded into an addressing mode.
152 bool isFoldable(const Candidate &C, TargetTransformInfo *TTI,
153 const DataLayout *DL);
154 // Returns true if C is already in a simplest form and not worth being
156 bool isSimplestForm(const Candidate &C);
157 // Checks whether I is in a candidate form. If so, adds all the matching forms
158 // to Candidates, and tries to find the immediate basis for each of them.
159 void allocateCandidatesAndFindBasis(Instruction *I);
160 // Allocate candidates and find bases for Add instructions.
161 void allocateCandidatesAndFindBasisForAdd(Instruction *I);
162 // Given I = LHS + RHS, factors RHS into i * S and makes (LHS + i * S) a
164 void allocateCandidatesAndFindBasisForAdd(Value *LHS, Value *RHS,
166 // Allocate candidates and find bases for Mul instructions.
167 void allocateCandidatesAndFindBasisForMul(Instruction *I);
168 // Splits LHS into Base + Index and, if succeeds, calls
169 // allocateCandidatesAndFindBasis.
170 void allocateCandidatesAndFindBasisForMul(Value *LHS, Value *RHS,
172 // Allocate candidates and find bases for GetElementPtr instructions.
173 void allocateCandidatesAndFindBasisForGEP(GetElementPtrInst *GEP);
174 // A helper function that scales Idx with ElementSize before invoking
175 // allocateCandidatesAndFindBasis.
176 void allocateCandidatesAndFindBasisForGEP(const SCEV *B, ConstantInt *Idx,
177 Value *S, uint64_t ElementSize,
179 // Adds the given form <CT, B, Idx, S> to Candidates, and finds its immediate
181 void allocateCandidatesAndFindBasis(Candidate::Kind CT, const SCEV *B,
182 ConstantInt *Idx, Value *S,
184 // Rewrites candidate C with respect to Basis.
185 void rewriteCandidateWithBasis(const Candidate &C, const Candidate &Basis);
186 // A helper function that factors ArrayIdx to a product of a stride and a
187 // constant index, and invokes allocateCandidatesAndFindBasis with the
189 void factorArrayIndex(Value *ArrayIdx, const SCEV *Base, uint64_t ElementSize,
190 GetElementPtrInst *GEP);
191 // Emit code that computes the "bump" from Basis to C. If the candidate is a
192 // GEP and the bump is not divisible by the element size of the GEP, this
193 // function sets the BumpWithUglyGEP flag to notify its caller to bump the
194 // basis using an ugly GEP.
195 static Value *emitBump(const Candidate &Basis, const Candidate &C,
196 IRBuilder<> &Builder, const DataLayout *DL,
197 bool &BumpWithUglyGEP);
199 const DataLayout *DL;
202 TargetTransformInfo *TTI;
203 ilist<Candidate> Candidates;
204 // Temporarily holds all instructions that are unlinked (but not deleted) by
205 // rewriteCandidateWithBasis. These instructions will be actually removed
206 // after all rewriting finishes.
207 std::vector<Instruction *> UnlinkedInstructions;
209 } // anonymous namespace
211 char StraightLineStrengthReduce::ID = 0;
212 INITIALIZE_PASS_BEGIN(StraightLineStrengthReduce, "slsr",
213 "Straight line strength reduction", false, false)
214 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
215 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
216 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
217 INITIALIZE_PASS_END(StraightLineStrengthReduce, "slsr",
218 "Straight line strength reduction", false, false)
220 FunctionPass *llvm::createStraightLineStrengthReducePass() {
221 return new StraightLineStrengthReduce();
224 bool StraightLineStrengthReduce::isBasisFor(const Candidate &Basis,
225 const Candidate &C) {
226 return (Basis.Ins != C.Ins && // skip the same instruction
227 // They must have the same type too. Basis.Base == C.Base doesn't
228 // guarantee their types are the same (PR23975).
229 Basis.Ins->getType() == C.Ins->getType() &&
230 // Basis must dominate C in order to rewrite C with respect to Basis.
231 DT->dominates(Basis.Ins->getParent(), C.Ins->getParent()) &&
232 // They share the same base, stride, and candidate kind.
233 Basis.Base == C.Base && Basis.Stride == C.Stride &&
234 Basis.CandidateKind == C.CandidateKind);
237 static bool isGEPFoldable(GetElementPtrInst *GEP,
238 const TargetTransformInfo *TTI,
239 const DataLayout *DL) {
240 GlobalVariable *BaseGV = nullptr;
241 int64_t BaseOffset = 0;
242 bool HasBaseReg = false;
245 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand()))
250 gep_type_iterator GTI = gep_type_begin(GEP);
251 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I, ++GTI) {
252 if (isa<SequentialType>(*GTI)) {
253 int64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
254 if (ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
255 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
257 // Needs scale register.
259 // No addressing mode takes two scale registers.
265 StructType *STy = cast<StructType>(*GTI);
266 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
267 BaseOffset += DL->getStructLayout(STy)->getElementOffset(Field);
271 unsigned AddrSpace = GEP->getPointerAddressSpace();
272 return TTI->isLegalAddressingMode(GEP->getType()->getElementType(), BaseGV,
273 BaseOffset, HasBaseReg, Scale, AddrSpace);
276 // Returns whether (Base + Index * Stride) can be folded to an addressing mode.
277 static bool isAddFoldable(const SCEV *Base, ConstantInt *Index, Value *Stride,
278 TargetTransformInfo *TTI) {
279 return TTI->isLegalAddressingMode(Base->getType(), nullptr, 0, true,
280 Index->getSExtValue());
283 bool StraightLineStrengthReduce::isFoldable(const Candidate &C,
284 TargetTransformInfo *TTI,
285 const DataLayout *DL) {
286 if (C.CandidateKind == Candidate::Add)
287 return isAddFoldable(C.Base, C.Index, C.Stride, TTI);
288 if (C.CandidateKind == Candidate::GEP)
289 return isGEPFoldable(cast<GetElementPtrInst>(C.Ins), TTI, DL);
293 // Returns true if GEP has zero or one non-zero index.
294 static bool hasOnlyOneNonZeroIndex(GetElementPtrInst *GEP) {
295 unsigned NumNonZeroIndices = 0;
296 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I) {
297 ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
298 if (ConstIdx == nullptr || !ConstIdx->isZero())
301 return NumNonZeroIndices <= 1;
304 bool StraightLineStrengthReduce::isSimplestForm(const Candidate &C) {
305 if (C.CandidateKind == Candidate::Add) {
306 // B + 1 * S or B + (-1) * S
307 return C.Index->isOne() || C.Index->isMinusOne();
309 if (C.CandidateKind == Candidate::Mul) {
311 return C.Index->isZero();
313 if (C.CandidateKind == Candidate::GEP) {
314 // (char*)B + S or (char*)B - S
315 return ((C.Index->isOne() || C.Index->isMinusOne()) &&
316 hasOnlyOneNonZeroIndex(cast<GetElementPtrInst>(C.Ins)));
321 // TODO: We currently implement an algorithm whose time complexity is linear in
322 // the number of existing candidates. However, we could do better by using
323 // ScopedHashTable. Specifically, while traversing the dominator tree, we could
324 // maintain all the candidates that dominate the basic block being traversed in
325 // a ScopedHashTable. This hash table is indexed by the base and the stride of
326 // a candidate. Therefore, finding the immediate basis of a candidate boils down
327 // to one hash-table look up.
328 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
329 Candidate::Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
331 Candidate C(CT, B, Idx, S, I);
332 // SLSR can complicate an instruction in two cases:
334 // 1. If we can fold I into an addressing mode, computing I is likely free or
335 // takes only one instruction.
337 // 2. I is already in a simplest form. For example, when
340 // rewriting Y to X - 7 * S is probably a bad idea.
342 // In the above cases, we still add I to the candidate list so that I can be
343 // the basis of other candidates, but we leave I's basis blank so that I
344 // won't be rewritten.
345 if (!isFoldable(C, TTI, DL) && !isSimplestForm(C)) {
346 // Try to compute the immediate basis of C.
347 unsigned NumIterations = 0;
348 // Limit the scan radius to avoid running in quadratice time.
349 static const unsigned MaxNumIterations = 50;
350 for (auto Basis = Candidates.rbegin();
351 Basis != Candidates.rend() && NumIterations < MaxNumIterations;
352 ++Basis, ++NumIterations) {
353 if (isBasisFor(*Basis, C)) {
359 // Regardless of whether we find a basis for C, we need to push C to the
360 // candidate list so that it can be the basis of other candidates.
361 Candidates.push_back(C);
364 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
366 switch (I->getOpcode()) {
367 case Instruction::Add:
368 allocateCandidatesAndFindBasisForAdd(I);
370 case Instruction::Mul:
371 allocateCandidatesAndFindBasisForMul(I);
373 case Instruction::GetElementPtr:
374 allocateCandidatesAndFindBasisForGEP(cast<GetElementPtrInst>(I));
379 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd(
381 // Try matching B + i * S.
382 if (!isa<IntegerType>(I->getType()))
385 assert(I->getNumOperands() == 2 && "isn't I an add?");
386 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
387 allocateCandidatesAndFindBasisForAdd(LHS, RHS, I);
389 allocateCandidatesAndFindBasisForAdd(RHS, LHS, I);
392 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd(
393 Value *LHS, Value *RHS, Instruction *I) {
395 ConstantInt *Idx = nullptr;
396 if (match(RHS, m_Mul(m_Value(S), m_ConstantInt(Idx)))) {
397 // I = LHS + RHS = LHS + Idx * S
398 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
399 } else if (match(RHS, m_Shl(m_Value(S), m_ConstantInt(Idx)))) {
400 // I = LHS + RHS = LHS + (S << Idx) = LHS + S * (1 << Idx)
401 APInt One(Idx->getBitWidth(), 1);
402 Idx = ConstantInt::get(Idx->getContext(), One << Idx->getValue());
403 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
405 // At least, I = LHS + 1 * RHS
406 ConstantInt *One = ConstantInt::get(cast<IntegerType>(I->getType()), 1);
407 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), One, RHS,
412 // Returns true if A matches B + C where C is constant.
413 static bool matchesAdd(Value *A, Value *&B, ConstantInt *&C) {
414 return (match(A, m_Add(m_Value(B), m_ConstantInt(C))) ||
415 match(A, m_Add(m_ConstantInt(C), m_Value(B))));
418 // Returns true if A matches B | C where C is constant.
419 static bool matchesOr(Value *A, Value *&B, ConstantInt *&C) {
420 return (match(A, m_Or(m_Value(B), m_ConstantInt(C))) ||
421 match(A, m_Or(m_ConstantInt(C), m_Value(B))));
424 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
425 Value *LHS, Value *RHS, Instruction *I) {
427 ConstantInt *Idx = nullptr;
428 if (matchesAdd(LHS, B, Idx)) {
429 // If LHS is in the form of "Base + Index", then I is in the form of
430 // "(Base + Index) * RHS".
431 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
432 } else if (matchesOr(LHS, B, Idx) && haveNoCommonBitsSet(B, Idx, *DL)) {
433 // If LHS is in the form of "Base | Index" and Base and Index have no common
435 // Base | Index = Base + Index
436 // and I is thus in the form of "(Base + Index) * RHS".
437 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
439 // Otherwise, at least try the form (LHS + 0) * RHS.
440 ConstantInt *Zero = ConstantInt::get(cast<IntegerType>(I->getType()), 0);
441 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(LHS), Zero, RHS,
446 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
448 // Try matching (B + i) * S.
449 // TODO: we could extend SLSR to float and vector types.
450 if (!isa<IntegerType>(I->getType()))
453 assert(I->getNumOperands() == 2 && "isn't I a mul?");
454 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
455 allocateCandidatesAndFindBasisForMul(LHS, RHS, I);
457 // Symmetrically, try to split RHS to Base + Index.
458 allocateCandidatesAndFindBasisForMul(RHS, LHS, I);
462 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
463 const SCEV *B, ConstantInt *Idx, Value *S, uint64_t ElementSize,
465 // I = B + sext(Idx *nsw S) * ElementSize
466 // = B + (sext(Idx) * sext(S)) * ElementSize
467 // = B + (sext(Idx) * ElementSize) * sext(S)
468 // Casting to IntegerType is safe because we skipped vector GEPs.
469 IntegerType *IntPtrTy = cast<IntegerType>(DL->getIntPtrType(I->getType()));
470 ConstantInt *ScaledIdx = ConstantInt::get(
471 IntPtrTy, Idx->getSExtValue() * (int64_t)ElementSize, true);
472 allocateCandidatesAndFindBasis(Candidate::GEP, B, ScaledIdx, S, I);
475 void StraightLineStrengthReduce::factorArrayIndex(Value *ArrayIdx,
477 uint64_t ElementSize,
478 GetElementPtrInst *GEP) {
479 // At least, ArrayIdx = ArrayIdx *nsw 1.
480 allocateCandidatesAndFindBasisForGEP(
481 Base, ConstantInt::get(cast<IntegerType>(ArrayIdx->getType()), 1),
482 ArrayIdx, ElementSize, GEP);
483 Value *LHS = nullptr;
484 ConstantInt *RHS = nullptr;
485 // One alternative is matching the SCEV of ArrayIdx instead of ArrayIdx
486 // itself. This would allow us to handle the shl case for free. However,
487 // matching SCEVs has two issues:
489 // 1. this would complicate rewriting because the rewriting procedure
490 // would have to translate SCEVs back to IR instructions. This translation
491 // is difficult when LHS is further evaluated to a composite SCEV.
493 // 2. ScalarEvolution is designed to be control-flow oblivious. It tends
494 // to strip nsw/nuw flags which are critical for SLSR to trace into
495 // sext'ed multiplication.
496 if (match(ArrayIdx, m_NSWMul(m_Value(LHS), m_ConstantInt(RHS)))) {
497 // SLSR is currently unsafe if i * S may overflow.
498 // GEP = Base + sext(LHS *nsw RHS) * ElementSize
499 allocateCandidatesAndFindBasisForGEP(Base, RHS, LHS, ElementSize, GEP);
500 } else if (match(ArrayIdx, m_NSWShl(m_Value(LHS), m_ConstantInt(RHS)))) {
501 // GEP = Base + sext(LHS <<nsw RHS) * ElementSize
502 // = Base + sext(LHS *nsw (1 << RHS)) * ElementSize
503 APInt One(RHS->getBitWidth(), 1);
504 ConstantInt *PowerOf2 =
505 ConstantInt::get(RHS->getContext(), One << RHS->getValue());
506 allocateCandidatesAndFindBasisForGEP(Base, PowerOf2, LHS, ElementSize, GEP);
510 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
511 GetElementPtrInst *GEP) {
512 // TODO: handle vector GEPs
513 if (GEP->getType()->isVectorTy())
516 SmallVector<const SCEV *, 4> IndexExprs;
517 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I)
518 IndexExprs.push_back(SE->getSCEV(*I));
520 gep_type_iterator GTI = gep_type_begin(GEP);
521 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I) {
522 if (!isa<SequentialType>(*GTI++))
525 const SCEV *OrigIndexExpr = IndexExprs[I - 1];
526 IndexExprs[I - 1] = SE->getConstant(OrigIndexExpr->getType(), 0);
528 // The base of this candidate is GEP's base plus the offsets of all
529 // indices except this current one.
530 const SCEV *BaseExpr = SE->getGEPExpr(GEP->getSourceElementType(),
531 SE->getSCEV(GEP->getPointerOperand()),
532 IndexExprs, GEP->isInBounds());
533 Value *ArrayIdx = GEP->getOperand(I);
534 uint64_t ElementSize = DL->getTypeAllocSize(*GTI);
535 factorArrayIndex(ArrayIdx, BaseExpr, ElementSize, GEP);
536 // When ArrayIdx is the sext of a value, we try to factor that value as
537 // well. Handling this case is important because array indices are
538 // typically sign-extended to the pointer size.
539 Value *TruncatedArrayIdx = nullptr;
540 if (match(ArrayIdx, m_SExt(m_Value(TruncatedArrayIdx))))
541 factorArrayIndex(TruncatedArrayIdx, BaseExpr, ElementSize, GEP);
543 IndexExprs[I - 1] = OrigIndexExpr;
547 // A helper function that unifies the bitwidth of A and B.
548 static void unifyBitWidth(APInt &A, APInt &B) {
549 if (A.getBitWidth() < B.getBitWidth())
550 A = A.sext(B.getBitWidth());
551 else if (A.getBitWidth() > B.getBitWidth())
552 B = B.sext(A.getBitWidth());
555 Value *StraightLineStrengthReduce::emitBump(const Candidate &Basis,
557 IRBuilder<> &Builder,
558 const DataLayout *DL,
559 bool &BumpWithUglyGEP) {
560 APInt Idx = C.Index->getValue(), BasisIdx = Basis.Index->getValue();
561 unifyBitWidth(Idx, BasisIdx);
562 APInt IndexOffset = Idx - BasisIdx;
564 BumpWithUglyGEP = false;
565 if (Basis.CandidateKind == Candidate::GEP) {
567 IndexOffset.getBitWidth(),
568 DL->getTypeAllocSize(
569 cast<GetElementPtrInst>(Basis.Ins)->getType()->getElementType()));
571 APInt::sdivrem(IndexOffset, ElementSize, Q, R);
572 if (R.getSExtValue() == 0)
575 BumpWithUglyGEP = true;
578 // Compute Bump = C - Basis = (i' - i) * S.
579 // Common case 1: if (i' - i) is 1, Bump = S.
580 if (IndexOffset.getSExtValue() == 1)
582 // Common case 2: if (i' - i) is -1, Bump = -S.
583 if (IndexOffset.getSExtValue() == -1)
584 return Builder.CreateNeg(C.Stride);
586 // Otherwise, Bump = (i' - i) * sext/trunc(S). Note that (i' - i) and S may
587 // have different bit widths.
588 IntegerType *DeltaType =
589 IntegerType::get(Basis.Ins->getContext(), IndexOffset.getBitWidth());
590 Value *ExtendedStride = Builder.CreateSExtOrTrunc(C.Stride, DeltaType);
591 if (IndexOffset.isPowerOf2()) {
592 // If (i' - i) is a power of 2, Bump = sext/trunc(S) << log(i' - i).
593 ConstantInt *Exponent = ConstantInt::get(DeltaType, IndexOffset.logBase2());
594 return Builder.CreateShl(ExtendedStride, Exponent);
596 if ((-IndexOffset).isPowerOf2()) {
597 // If (i - i') is a power of 2, Bump = -sext/trunc(S) << log(i' - i).
598 ConstantInt *Exponent =
599 ConstantInt::get(DeltaType, (-IndexOffset).logBase2());
600 return Builder.CreateNeg(Builder.CreateShl(ExtendedStride, Exponent));
602 Constant *Delta = ConstantInt::get(DeltaType, IndexOffset);
603 return Builder.CreateMul(ExtendedStride, Delta);
606 void StraightLineStrengthReduce::rewriteCandidateWithBasis(
607 const Candidate &C, const Candidate &Basis) {
608 assert(C.CandidateKind == Basis.CandidateKind && C.Base == Basis.Base &&
609 C.Stride == Basis.Stride);
610 // We run rewriteCandidateWithBasis on all candidates in a post-order, so the
611 // basis of a candidate cannot be unlinked before the candidate.
612 assert(Basis.Ins->getParent() != nullptr && "the basis is unlinked");
614 // An instruction can correspond to multiple candidates. Therefore, instead of
615 // simply deleting an instruction when we rewrite it, we mark its parent as
616 // nullptr (i.e. unlink it) so that we can skip the candidates whose
617 // instruction is already rewritten.
618 if (!C.Ins->getParent())
621 IRBuilder<> Builder(C.Ins);
622 bool BumpWithUglyGEP;
623 Value *Bump = emitBump(Basis, C, Builder, DL, BumpWithUglyGEP);
624 Value *Reduced = nullptr; // equivalent to but weaker than C.Ins
625 switch (C.CandidateKind) {
629 if (BinaryOperator::isNeg(Bump)) {
630 // If Bump is a neg instruction, emit C = Basis - (-Bump).
632 Builder.CreateSub(Basis.Ins, BinaryOperator::getNegArgument(Bump));
633 // We only use the negative argument of Bump, and Bump itself may be
635 RecursivelyDeleteTriviallyDeadInstructions(Bump);
637 // It's tempting to preserve nsw on Bump and/or Reduced. However, it's
638 // usually unsound, e.g.,
640 // X = (-2 +nsw 1) *nsw INT_MAX
641 // Y = (-2 +nsw 3) *nsw INT_MAX
643 // Y = X + 2 * INT_MAX
645 // Neither + and * in the resultant expression are nsw.
646 Reduced = Builder.CreateAdd(Basis.Ins, Bump);
651 Type *IntPtrTy = DL->getIntPtrType(C.Ins->getType());
652 bool InBounds = cast<GetElementPtrInst>(C.Ins)->isInBounds();
653 if (BumpWithUglyGEP) {
654 // C = (char *)Basis + Bump
655 unsigned AS = Basis.Ins->getType()->getPointerAddressSpace();
656 Type *CharTy = Type::getInt8PtrTy(Basis.Ins->getContext(), AS);
657 Reduced = Builder.CreateBitCast(Basis.Ins, CharTy);
660 Builder.CreateInBoundsGEP(Builder.getInt8Ty(), Reduced, Bump);
662 Reduced = Builder.CreateGEP(Builder.getInt8Ty(), Reduced, Bump);
663 Reduced = Builder.CreateBitCast(Reduced, C.Ins->getType());
665 // C = gep Basis, Bump
666 // Canonicalize bump to pointer size.
667 Bump = Builder.CreateSExtOrTrunc(Bump, IntPtrTy);
669 Reduced = Builder.CreateInBoundsGEP(nullptr, Basis.Ins, Bump);
671 Reduced = Builder.CreateGEP(nullptr, Basis.Ins, Bump);
676 llvm_unreachable("C.CandidateKind is invalid");
678 Reduced->takeName(C.Ins);
679 C.Ins->replaceAllUsesWith(Reduced);
680 // Unlink C.Ins so that we can skip other candidates also corresponding to
681 // C.Ins. The actual deletion is postponed to the end of runOnFunction.
682 C.Ins->removeFromParent();
683 UnlinkedInstructions.push_back(C.Ins);
686 bool StraightLineStrengthReduce::runOnFunction(Function &F) {
687 if (skipOptnoneFunction(F))
690 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
691 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
692 SE = &getAnalysis<ScalarEvolution>();
693 // Traverse the dominator tree in the depth-first order. This order makes sure
694 // all bases of a candidate are in Candidates when we process it.
695 for (auto node = GraphTraits<DominatorTree *>::nodes_begin(DT);
696 node != GraphTraits<DominatorTree *>::nodes_end(DT); ++node) {
697 for (auto &I : *node->getBlock())
698 allocateCandidatesAndFindBasis(&I);
701 // Rewrite candidates in the reverse depth-first order. This order makes sure
702 // a candidate being rewritten is not a basis for any other candidate.
703 while (!Candidates.empty()) {
704 const Candidate &C = Candidates.back();
705 if (C.Basis != nullptr) {
706 rewriteCandidateWithBasis(C, *C.Basis);
708 Candidates.pop_back();
711 // Delete all unlink instructions.
712 for (auto *UnlinkedInst : UnlinkedInstructions) {
713 for (unsigned I = 0, E = UnlinkedInst->getNumOperands(); I != E; ++I) {
714 Value *Op = UnlinkedInst->getOperand(I);
715 UnlinkedInst->setOperand(I, nullptr);
716 RecursivelyDeleteTriviallyDeadInstructions(Op);
720 bool Ret = !UnlinkedInstructions.empty();
721 UnlinkedInstructions.clear();