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/IR/DataLayout.h"
65 #include "llvm/IR/Dominators.h"
66 #include "llvm/IR/IRBuilder.h"
67 #include "llvm/IR/Module.h"
68 #include "llvm/IR/PatternMatch.h"
69 #include "llvm/Support/raw_ostream.h"
70 #include "llvm/Transforms/Scalar.h"
71 #include "llvm/Transforms/Utils/Local.h"
74 using namespace PatternMatch;
78 class StraightLineStrengthReduce : public FunctionPass {
80 // SLSR candidate. Such a candidate must be in one of the forms described in
81 // the header comments.
82 struct Candidate : public ilist_node<Candidate> {
84 Invalid, // reserved for the default constructor
87 GEP, // &B[..][i * S][..]
91 : CandidateKind(Invalid), Base(nullptr), Index(nullptr),
92 Stride(nullptr), Ins(nullptr), Basis(nullptr) {}
93 Candidate(Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
95 : CandidateKind(CT), Base(B), Index(Idx), Stride(S), Ins(I),
99 // Note that Index and Stride of a GEP candidate do not necessarily have the
100 // same integer type. In that case, during rewriting, Stride will be
101 // sign-extended or truncated to Index's type.
104 // The instruction this candidate corresponds to. It helps us to rewrite a
105 // candidate with respect to its immediate basis. Note that one instruction
106 // can correspond to multiple candidates depending on how you associate the
107 // expression. For instance,
113 // <Base: a, Index: 1, Stride: b + 2>
117 // <Base: b, Index: 2, Stride: a + 1>
119 // Points to the immediate basis of this candidate, or nullptr if we cannot
120 // find any basis for this candidate.
126 StraightLineStrengthReduce()
127 : FunctionPass(ID), DL(nullptr), DT(nullptr), TTI(nullptr) {
128 initializeStraightLineStrengthReducePass(*PassRegistry::getPassRegistry());
131 void getAnalysisUsage(AnalysisUsage &AU) const override {
132 AU.addRequired<DominatorTreeWrapperPass>();
133 AU.addRequired<ScalarEvolution>();
134 AU.addRequired<TargetTransformInfoWrapperPass>();
135 // We do not modify the shape of the CFG.
136 AU.setPreservesCFG();
139 bool doInitialization(Module &M) override {
140 DL = &M.getDataLayout();
144 bool runOnFunction(Function &F) override;
147 // Returns true if Basis is a basis for C, i.e., Basis dominates C and they
148 // share the same base and stride.
149 bool isBasisFor(const Candidate &Basis, const Candidate &C);
150 // Returns whether the candidate can be folded into an addressing mode.
151 bool isFoldable(const Candidate &C, TargetTransformInfo *TTI,
152 const DataLayout *DL);
153 // Returns true if C is already in a simplest form and not worth being
155 bool isSimplestForm(const Candidate &C);
156 // Checks whether I is in a candidate form. If so, adds all the matching forms
157 // to Candidates, and tries to find the immediate basis for each of them.
158 void allocateCandidatesAndFindBasis(Instruction *I);
159 // Allocate candidates and find bases for Add instructions.
160 void allocateCandidatesAndFindBasisForAdd(Instruction *I);
161 // Given I = LHS + RHS, factors RHS into i * S and makes (LHS + i * S) a
163 void allocateCandidatesAndFindBasisForAdd(Value *LHS, Value *RHS,
165 // Allocate candidates and find bases for Mul instructions.
166 void allocateCandidatesAndFindBasisForMul(Instruction *I);
167 // Splits LHS into Base + Index and, if succeeds, calls
168 // allocateCandidatesAndFindBasis.
169 void allocateCandidatesAndFindBasisForMul(Value *LHS, Value *RHS,
171 // Allocate candidates and find bases for GetElementPtr instructions.
172 void allocateCandidatesAndFindBasisForGEP(GetElementPtrInst *GEP);
173 // A helper function that scales Idx with ElementSize before invoking
174 // allocateCandidatesAndFindBasis.
175 void allocateCandidatesAndFindBasisForGEP(const SCEV *B, ConstantInt *Idx,
176 Value *S, uint64_t ElementSize,
178 // Adds the given form <CT, B, Idx, S> to Candidates, and finds its immediate
180 void allocateCandidatesAndFindBasis(Candidate::Kind CT, const SCEV *B,
181 ConstantInt *Idx, Value *S,
183 // Rewrites candidate C with respect to Basis.
184 void rewriteCandidateWithBasis(const Candidate &C, const Candidate &Basis);
185 // A helper function that factors ArrayIdx to a product of a stride and a
186 // constant index, and invokes allocateCandidatesAndFindBasis with the
188 void factorArrayIndex(Value *ArrayIdx, const SCEV *Base, uint64_t ElementSize,
189 GetElementPtrInst *GEP);
190 // Emit code that computes the "bump" from Basis to C. If the candidate is a
191 // GEP and the bump is not divisible by the element size of the GEP, this
192 // function sets the BumpWithUglyGEP flag to notify its caller to bump the
193 // basis using an ugly GEP.
194 static Value *emitBump(const Candidate &Basis, const Candidate &C,
195 IRBuilder<> &Builder, const DataLayout *DL,
196 bool &BumpWithUglyGEP);
198 const DataLayout *DL;
201 TargetTransformInfo *TTI;
202 ilist<Candidate> Candidates;
203 // Temporarily holds all instructions that are unlinked (but not deleted) by
204 // rewriteCandidateWithBasis. These instructions will be actually removed
205 // after all rewriting finishes.
206 std::vector<Instruction *> UnlinkedInstructions;
208 } // anonymous namespace
210 char StraightLineStrengthReduce::ID = 0;
211 INITIALIZE_PASS_BEGIN(StraightLineStrengthReduce, "slsr",
212 "Straight line strength reduction", false, false)
213 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
214 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
215 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
216 INITIALIZE_PASS_END(StraightLineStrengthReduce, "slsr",
217 "Straight line strength reduction", false, false)
219 FunctionPass *llvm::createStraightLineStrengthReducePass() {
220 return new StraightLineStrengthReduce();
223 bool StraightLineStrengthReduce::isBasisFor(const Candidate &Basis,
224 const Candidate &C) {
225 return (Basis.Ins != C.Ins && // skip the same instruction
226 // Basis must dominate C in order to rewrite C with respect to Basis.
227 DT->dominates(Basis.Ins->getParent(), C.Ins->getParent()) &&
228 // They share the same base, stride, and candidate kind.
229 Basis.Base == C.Base &&
230 Basis.Stride == C.Stride &&
231 Basis.CandidateKind == C.CandidateKind);
234 static bool isGEPFoldable(GetElementPtrInst *GEP,
235 const TargetTransformInfo *TTI,
236 const DataLayout *DL) {
237 GlobalVariable *BaseGV = nullptr;
238 int64_t BaseOffset = 0;
239 bool HasBaseReg = false;
242 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand()))
247 gep_type_iterator GTI = gep_type_begin(GEP);
248 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I, ++GTI) {
249 if (isa<SequentialType>(*GTI)) {
250 int64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
251 if (ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
252 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
254 // Needs scale register.
256 // No addressing mode takes two scale registers.
262 StructType *STy = cast<StructType>(*GTI);
263 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
264 BaseOffset += DL->getStructLayout(STy)->getElementOffset(Field);
267 return TTI->isLegalAddressingMode(GEP->getType()->getElementType(), BaseGV,
268 BaseOffset, HasBaseReg, Scale);
271 // Returns whether (Base + Index * Stride) can be folded to an addressing mode.
272 static bool isAddFoldable(const SCEV *Base, ConstantInt *Index, Value *Stride,
273 TargetTransformInfo *TTI) {
274 return TTI->isLegalAddressingMode(Base->getType(), nullptr, 0, true,
275 Index->getSExtValue());
278 bool StraightLineStrengthReduce::isFoldable(const Candidate &C,
279 TargetTransformInfo *TTI,
280 const DataLayout *DL) {
281 if (C.CandidateKind == Candidate::Add)
282 return isAddFoldable(C.Base, C.Index, C.Stride, TTI);
283 if (C.CandidateKind == Candidate::GEP)
284 return isGEPFoldable(cast<GetElementPtrInst>(C.Ins), TTI, DL);
288 // Returns true if GEP has zero or one non-zero index.
289 static bool hasOnlyOneNonZeroIndex(GetElementPtrInst *GEP) {
290 unsigned NumNonZeroIndices = 0;
291 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I) {
292 ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
293 if (ConstIdx == nullptr || !ConstIdx->isZero())
296 return NumNonZeroIndices <= 1;
299 bool StraightLineStrengthReduce::isSimplestForm(const Candidate &C) {
300 if (C.CandidateKind == Candidate::Add) {
301 // B + 1 * S or B + (-1) * S
302 return C.Index->isOne() || C.Index->isMinusOne();
304 if (C.CandidateKind == Candidate::Mul) {
306 return C.Index->isZero();
308 if (C.CandidateKind == Candidate::GEP) {
309 // (char*)B + S or (char*)B - S
310 return ((C.Index->isOne() || C.Index->isMinusOne()) &&
311 hasOnlyOneNonZeroIndex(cast<GetElementPtrInst>(C.Ins)));
316 // TODO: We currently implement an algorithm whose time complexity is linear in
317 // the number of existing candidates. However, we could do better by using
318 // ScopedHashTable. Specifically, while traversing the dominator tree, we could
319 // maintain all the candidates that dominate the basic block being traversed in
320 // a ScopedHashTable. This hash table is indexed by the base and the stride of
321 // a candidate. Therefore, finding the immediate basis of a candidate boils down
322 // to one hash-table look up.
323 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
324 Candidate::Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
326 Candidate C(CT, B, Idx, S, I);
327 // SLSR can complicate an instruction in two cases:
329 // 1. If we can fold I into an addressing mode, computing I is likely free or
330 // takes only one instruction.
332 // 2. I is already in a simplest form. For example, when
335 // rewriting Y to X - 7 * S is probably a bad idea.
337 // In the above cases, we still add I to the candidate list so that I can be
338 // the basis of other candidates, but we leave I's basis blank so that I
339 // won't be rewritten.
340 if (!isFoldable(C, TTI, DL) && !isSimplestForm(C)) {
341 // Try to compute the immediate basis of C.
342 unsigned NumIterations = 0;
343 // Limit the scan radius to avoid running in quadratice time.
344 static const unsigned MaxNumIterations = 50;
345 for (auto Basis = Candidates.rbegin();
346 Basis != Candidates.rend() && NumIterations < MaxNumIterations;
347 ++Basis, ++NumIterations) {
348 if (isBasisFor(*Basis, C)) {
354 // Regardless of whether we find a basis for C, we need to push C to the
355 // candidate list so that it can be the basis of other candidates.
356 Candidates.push_back(C);
359 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
361 switch (I->getOpcode()) {
362 case Instruction::Add:
363 allocateCandidatesAndFindBasisForAdd(I);
365 case Instruction::Mul:
366 allocateCandidatesAndFindBasisForMul(I);
368 case Instruction::GetElementPtr:
369 allocateCandidatesAndFindBasisForGEP(cast<GetElementPtrInst>(I));
374 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd(
376 // Try matching B + i * S.
377 if (!isa<IntegerType>(I->getType()))
380 assert(I->getNumOperands() == 2 && "isn't I an add?");
381 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
382 allocateCandidatesAndFindBasisForAdd(LHS, RHS, I);
384 allocateCandidatesAndFindBasisForAdd(RHS, LHS, I);
387 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd(
388 Value *LHS, Value *RHS, Instruction *I) {
390 ConstantInt *Idx = nullptr;
391 if (match(RHS, m_Mul(m_Value(S), m_ConstantInt(Idx)))) {
392 // I = LHS + RHS = LHS + Idx * S
393 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
394 } else if (match(RHS, m_Shl(m_Value(S), m_ConstantInt(Idx)))) {
395 // I = LHS + RHS = LHS + (S << Idx) = LHS + S * (1 << Idx)
396 APInt One(Idx->getBitWidth(), 1);
397 Idx = ConstantInt::get(Idx->getContext(), One << Idx->getValue());
398 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
400 // At least, I = LHS + 1 * RHS
401 ConstantInt *One = ConstantInt::get(cast<IntegerType>(I->getType()), 1);
402 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), One, RHS,
407 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
408 Value *LHS, Value *RHS, Instruction *I) {
410 ConstantInt *Idx = nullptr;
411 // Only handle the canonical operand ordering.
412 if (match(LHS, m_Add(m_Value(B), m_ConstantInt(Idx)))) {
413 // If LHS is in the form of "Base + Index", then I is in the form of
414 // "(Base + Index) * RHS".
415 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
417 // Otherwise, at least try the form (LHS + 0) * RHS.
418 ConstantInt *Zero = ConstantInt::get(cast<IntegerType>(I->getType()), 0);
419 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(LHS), Zero, RHS,
424 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
426 // Try matching (B + i) * S.
427 // TODO: we could extend SLSR to float and vector types.
428 if (!isa<IntegerType>(I->getType()))
431 assert(I->getNumOperands() == 2 && "isn't I a mul?");
432 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
433 allocateCandidatesAndFindBasisForMul(LHS, RHS, I);
435 // Symmetrically, try to split RHS to Base + Index.
436 allocateCandidatesAndFindBasisForMul(RHS, LHS, I);
440 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
441 const SCEV *B, ConstantInt *Idx, Value *S, uint64_t ElementSize,
443 // I = B + sext(Idx *nsw S) * ElementSize
444 // = B + (sext(Idx) * sext(S)) * ElementSize
445 // = B + (sext(Idx) * ElementSize) * sext(S)
446 // Casting to IntegerType is safe because we skipped vector GEPs.
447 IntegerType *IntPtrTy = cast<IntegerType>(DL->getIntPtrType(I->getType()));
448 ConstantInt *ScaledIdx = ConstantInt::get(
449 IntPtrTy, Idx->getSExtValue() * (int64_t)ElementSize, true);
450 allocateCandidatesAndFindBasis(Candidate::GEP, B, ScaledIdx, S, I);
453 void StraightLineStrengthReduce::factorArrayIndex(Value *ArrayIdx,
455 uint64_t ElementSize,
456 GetElementPtrInst *GEP) {
457 // At least, ArrayIdx = ArrayIdx *nsw 1.
458 allocateCandidatesAndFindBasisForGEP(
459 Base, ConstantInt::get(cast<IntegerType>(ArrayIdx->getType()), 1),
460 ArrayIdx, ElementSize, GEP);
461 Value *LHS = nullptr;
462 ConstantInt *RHS = nullptr;
463 // One alternative is matching the SCEV of ArrayIdx instead of ArrayIdx
464 // itself. This would allow us to handle the shl case for free. However,
465 // matching SCEVs has two issues:
467 // 1. this would complicate rewriting because the rewriting procedure
468 // would have to translate SCEVs back to IR instructions. This translation
469 // is difficult when LHS is further evaluated to a composite SCEV.
471 // 2. ScalarEvolution is designed to be control-flow oblivious. It tends
472 // to strip nsw/nuw flags which are critical for SLSR to trace into
473 // sext'ed multiplication.
474 if (match(ArrayIdx, m_NSWMul(m_Value(LHS), m_ConstantInt(RHS)))) {
475 // SLSR is currently unsafe if i * S may overflow.
476 // GEP = Base + sext(LHS *nsw RHS) * ElementSize
477 allocateCandidatesAndFindBasisForGEP(Base, RHS, LHS, ElementSize, GEP);
478 } else if (match(ArrayIdx, m_NSWShl(m_Value(LHS), m_ConstantInt(RHS)))) {
479 // GEP = Base + sext(LHS <<nsw RHS) * ElementSize
480 // = Base + sext(LHS *nsw (1 << RHS)) * ElementSize
481 APInt One(RHS->getBitWidth(), 1);
482 ConstantInt *PowerOf2 =
483 ConstantInt::get(RHS->getContext(), One << RHS->getValue());
484 allocateCandidatesAndFindBasisForGEP(Base, PowerOf2, LHS, ElementSize, GEP);
488 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
489 GetElementPtrInst *GEP) {
490 // TODO: handle vector GEPs
491 if (GEP->getType()->isVectorTy())
494 const SCEV *GEPExpr = SE->getSCEV(GEP);
495 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
497 gep_type_iterator GTI = gep_type_begin(GEP);
498 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I) {
499 if (!isa<SequentialType>(*GTI++))
501 Value *ArrayIdx = *I;
502 // Compute the byte offset of this index.
503 uint64_t ElementSize = DL->getTypeAllocSize(*GTI);
504 const SCEV *ElementSizeExpr = SE->getSizeOfExpr(IntPtrTy, *GTI);
505 const SCEV *ArrayIdxExpr = SE->getSCEV(ArrayIdx);
506 ArrayIdxExpr = SE->getTruncateOrSignExtend(ArrayIdxExpr, IntPtrTy);
507 const SCEV *LocalOffset =
508 SE->getMulExpr(ArrayIdxExpr, ElementSizeExpr, SCEV::FlagNSW);
509 // The base of this candidate equals GEPExpr less the byte offset of this
511 const SCEV *Base = SE->getMinusSCEV(GEPExpr, LocalOffset);
512 factorArrayIndex(ArrayIdx, Base, ElementSize, GEP);
513 // When ArrayIdx is the sext of a value, we try to factor that value as
514 // well. Handling this case is important because array indices are
515 // typically sign-extended to the pointer size.
516 Value *TruncatedArrayIdx = nullptr;
517 if (match(ArrayIdx, m_SExt(m_Value(TruncatedArrayIdx))))
518 factorArrayIndex(TruncatedArrayIdx, Base, ElementSize, GEP);
522 // A helper function that unifies the bitwidth of A and B.
523 static void unifyBitWidth(APInt &A, APInt &B) {
524 if (A.getBitWidth() < B.getBitWidth())
525 A = A.sext(B.getBitWidth());
526 else if (A.getBitWidth() > B.getBitWidth())
527 B = B.sext(A.getBitWidth());
530 Value *StraightLineStrengthReduce::emitBump(const Candidate &Basis,
532 IRBuilder<> &Builder,
533 const DataLayout *DL,
534 bool &BumpWithUglyGEP) {
535 APInt Idx = C.Index->getValue(), BasisIdx = Basis.Index->getValue();
536 unifyBitWidth(Idx, BasisIdx);
537 APInt IndexOffset = Idx - BasisIdx;
539 BumpWithUglyGEP = false;
540 if (Basis.CandidateKind == Candidate::GEP) {
542 IndexOffset.getBitWidth(),
543 DL->getTypeAllocSize(
544 cast<GetElementPtrInst>(Basis.Ins)->getType()->getElementType()));
546 APInt::sdivrem(IndexOffset, ElementSize, Q, R);
547 if (R.getSExtValue() == 0)
550 BumpWithUglyGEP = true;
553 // Compute Bump = C - Basis = (i' - i) * S.
554 // Common case 1: if (i' - i) is 1, Bump = S.
555 if (IndexOffset.getSExtValue() == 1)
557 // Common case 2: if (i' - i) is -1, Bump = -S.
558 if (IndexOffset.getSExtValue() == -1)
559 return Builder.CreateNeg(C.Stride);
561 // Otherwise, Bump = (i' - i) * sext/trunc(S). Note that (i' - i) and S may
562 // have different bit widths.
563 IntegerType *DeltaType =
564 IntegerType::get(Basis.Ins->getContext(), IndexOffset.getBitWidth());
565 Value *ExtendedStride = Builder.CreateSExtOrTrunc(C.Stride, DeltaType);
566 if (IndexOffset.isPowerOf2()) {
567 // If (i' - i) is a power of 2, Bump = sext/trunc(S) << log(i' - i).
568 ConstantInt *Exponent = ConstantInt::get(DeltaType, IndexOffset.logBase2());
569 return Builder.CreateShl(ExtendedStride, Exponent);
571 if ((-IndexOffset).isPowerOf2()) {
572 // If (i - i') is a power of 2, Bump = -sext/trunc(S) << log(i' - i).
573 ConstantInt *Exponent =
574 ConstantInt::get(DeltaType, (-IndexOffset).logBase2());
575 return Builder.CreateNeg(Builder.CreateShl(ExtendedStride, Exponent));
577 Constant *Delta = ConstantInt::get(DeltaType, IndexOffset);
578 return Builder.CreateMul(ExtendedStride, Delta);
581 void StraightLineStrengthReduce::rewriteCandidateWithBasis(
582 const Candidate &C, const Candidate &Basis) {
583 assert(C.CandidateKind == Basis.CandidateKind && C.Base == Basis.Base &&
584 C.Stride == Basis.Stride);
585 // We run rewriteCandidateWithBasis on all candidates in a post-order, so the
586 // basis of a candidate cannot be unlinked before the candidate.
587 assert(Basis.Ins->getParent() != nullptr && "the basis is unlinked");
589 // An instruction can correspond to multiple candidates. Therefore, instead of
590 // simply deleting an instruction when we rewrite it, we mark its parent as
591 // nullptr (i.e. unlink it) so that we can skip the candidates whose
592 // instruction is already rewritten.
593 if (!C.Ins->getParent())
596 IRBuilder<> Builder(C.Ins);
597 bool BumpWithUglyGEP;
598 Value *Bump = emitBump(Basis, C, Builder, DL, BumpWithUglyGEP);
599 Value *Reduced = nullptr; // equivalent to but weaker than C.Ins
600 switch (C.CandidateKind) {
604 if (BinaryOperator::isNeg(Bump)) {
605 // If Bump is a neg instruction, emit C = Basis - (-Bump).
607 Builder.CreateSub(Basis.Ins, BinaryOperator::getNegArgument(Bump));
608 // We only use the negative argument of Bump, and Bump itself may be
610 RecursivelyDeleteTriviallyDeadInstructions(Bump);
612 Reduced = Builder.CreateAdd(Basis.Ins, Bump);
617 Type *IntPtrTy = DL->getIntPtrType(C.Ins->getType());
618 bool InBounds = cast<GetElementPtrInst>(C.Ins)->isInBounds();
619 if (BumpWithUglyGEP) {
620 // C = (char *)Basis + Bump
621 unsigned AS = Basis.Ins->getType()->getPointerAddressSpace();
622 Type *CharTy = Type::getInt8PtrTy(Basis.Ins->getContext(), AS);
623 Reduced = Builder.CreateBitCast(Basis.Ins, CharTy);
626 Builder.CreateInBoundsGEP(Builder.getInt8Ty(), Reduced, Bump);
628 Reduced = Builder.CreateGEP(Builder.getInt8Ty(), Reduced, Bump);
629 Reduced = Builder.CreateBitCast(Reduced, C.Ins->getType());
631 // C = gep Basis, Bump
632 // Canonicalize bump to pointer size.
633 Bump = Builder.CreateSExtOrTrunc(Bump, IntPtrTy);
635 Reduced = Builder.CreateInBoundsGEP(nullptr, Basis.Ins, Bump);
637 Reduced = Builder.CreateGEP(nullptr, Basis.Ins, Bump);
642 llvm_unreachable("C.CandidateKind is invalid");
644 Reduced->takeName(C.Ins);
645 C.Ins->replaceAllUsesWith(Reduced);
646 // Unlink C.Ins so that we can skip other candidates also corresponding to
647 // C.Ins. The actual deletion is postponed to the end of runOnFunction.
648 C.Ins->removeFromParent();
649 UnlinkedInstructions.push_back(C.Ins);
652 bool StraightLineStrengthReduce::runOnFunction(Function &F) {
653 if (skipOptnoneFunction(F))
656 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
657 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
658 SE = &getAnalysis<ScalarEvolution>();
659 // Traverse the dominator tree in the depth-first order. This order makes sure
660 // all bases of a candidate are in Candidates when we process it.
661 for (auto node = GraphTraits<DominatorTree *>::nodes_begin(DT);
662 node != GraphTraits<DominatorTree *>::nodes_end(DT); ++node) {
663 for (auto &I : *node->getBlock())
664 allocateCandidatesAndFindBasis(&I);
667 // Rewrite candidates in the reverse depth-first order. This order makes sure
668 // a candidate being rewritten is not a basis for any other candidate.
669 while (!Candidates.empty()) {
670 const Candidate &C = Candidates.back();
671 if (C.Basis != nullptr) {
672 rewriteCandidateWithBasis(C, *C.Basis);
674 Candidates.pop_back();
677 // Delete all unlink instructions.
678 for (auto *UnlinkedInst : UnlinkedInstructions) {
679 for (unsigned I = 0, E = UnlinkedInst->getNumOperands(); I != E; ++I) {
680 Value *Op = UnlinkedInst->getOperand(I);
681 UnlinkedInst->setOperand(I, nullptr);
682 RecursivelyDeleteTriviallyDeadInstructions(Op);
686 bool Ret = !UnlinkedInstructions.empty();
687 UnlinkedInstructions.clear();