LLVM 20.0.0git
StraightLineStrengthReduce.cpp
Go to the documentation of this file.
1//===- StraightLineStrengthReduce.cpp - -----------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements straight-line strength reduction (SLSR). Unlike loop
10// strength reduction, this algorithm is designed to reduce arithmetic
11// redundancy in straight-line code instead of loops. It has proven to be
12// effective in simplifying arithmetic statements derived from an unrolled loop.
13// It can also simplify the logic of SeparateConstOffsetFromGEP.
14//
15// There are many optimizations we can perform in the domain of SLSR. This file
16// for now contains only an initial step. Specifically, we look for strength
17// reduction candidates in the following forms:
18//
19// Form 1: B + i * S
20// Form 2: (B + i) * S
21// Form 3: &B[i * S]
22//
23// where S is an integer variable, and i is a constant integer. If we found two
24// candidates S1 and S2 in the same form and S1 dominates S2, we may rewrite S2
25// in a simpler way with respect to S1. For example,
26//
27// S1: X = B + i * S
28// S2: Y = B + i' * S => X + (i' - i) * S
29//
30// S1: X = (B + i) * S
31// S2: Y = (B + i') * S => X + (i' - i) * S
32//
33// S1: X = &B[i * S]
34// S2: Y = &B[i' * S] => &X[(i' - i) * S]
35//
36// Note: (i' - i) * S is folded to the extent possible.
37//
38// This rewriting is in general a good idea. The code patterns we focus on
39// usually come from loop unrolling, so (i' - i) * S is likely the same
40// across iterations and can be reused. When that happens, the optimized form
41// takes only one add starting from the second iteration.
42//
43// When such rewriting is possible, we call S1 a "basis" of S2. When S2 has
44// multiple bases, we choose to rewrite S2 with respect to its "immediate"
45// basis, the basis that is the closest ancestor in the dominator tree.
46//
47// TODO:
48//
49// - Floating point arithmetics when fast math is enabled.
50//
51// - SLSR may decrease ILP at the architecture level. Targets that are very
52// sensitive to ILP may want to disable it. Having SLSR to consider ILP is
53// left as future work.
54//
55// - When (i' - i) is constant but i and i' are not, we could still perform
56// SLSR.
57
58
65#include "llvm/IR/Constants.h"
66#include "llvm/IR/DataLayout.h"
68#include "llvm/IR/Dominators.h"
70#include "llvm/IR/IRBuilder.h"
71#include "llvm/IR/Instruction.h"
73#include "llvm/IR/Module.h"
74#include "llvm/IR/Operator.h"
76#include "llvm/IR/Type.h"
77#include "llvm/IR/Value.h"
79#include "llvm/Pass.h"
84#include <cassert>
85#include <cstdint>
86#include <limits>
87#include <list>
88#include <vector>
89
90using namespace llvm;
91using namespace PatternMatch;
92
94 std::numeric_limits<unsigned>::max();
95
96namespace {
97
98class StraightLineStrengthReduceLegacyPass : public FunctionPass {
99 const DataLayout *DL = nullptr;
100
101public:
102 static char ID;
103
104 StraightLineStrengthReduceLegacyPass() : FunctionPass(ID) {
107 }
108
109 void getAnalysisUsage(AnalysisUsage &AU) const override {
113 // We do not modify the shape of the CFG.
114 AU.setPreservesCFG();
115 }
116
117 bool doInitialization(Module &M) override {
118 DL = &M.getDataLayout();
119 return false;
120 }
121
122 bool runOnFunction(Function &F) override;
123};
124
125class StraightLineStrengthReduce {
126public:
127 StraightLineStrengthReduce(const DataLayout *DL, DominatorTree *DT,
129 : DL(DL), DT(DT), SE(SE), TTI(TTI) {}
130
131 // SLSR candidate. Such a candidate must be in one of the forms described in
133 struct Candidate {
134 enum Kind {
135 Invalid, // reserved for the default constructor
136 Add, // B + i * S
137 Mul, // (B + i) * S
138 GEP, // &B[..][i * S][..]
139 };
140
141 Candidate() = default;
142 Candidate(Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
143 Instruction *I)
144 : CandidateKind(CT), Base(B), Index(Idx), Stride(S), Ins(I) {}
145
146 Kind CandidateKind = Invalid;
147
148 const SCEV *Base = nullptr;
149
150 // Note that Index and Stride of a GEP candidate do not necessarily have the
151 // same integer type. In that case, during rewriting, Stride will be
152 // sign-extended or truncated to Index's type.
153 ConstantInt *Index = nullptr;
154
155 Value *Stride = nullptr;
156
157 // The instruction this candidate corresponds to. It helps us to rewrite a
158 // candidate with respect to its immediate basis. Note that one instruction
159 // can correspond to multiple candidates depending on how you associate the
160 // expression. For instance,
161 //
162 // (a + 1) * (b + 2)
163 //
164 // can be treated as
165 //
166 // <Base: a, Index: 1, Stride: b + 2>
167 //
168 // or
169 //
170 // <Base: b, Index: 2, Stride: a + 1>
171 Instruction *Ins = nullptr;
172
173 // Points to the immediate basis of this candidate, or nullptr if we cannot
174 // find any basis for this candidate.
175 Candidate *Basis = nullptr;
176 };
177
178 bool runOnFunction(Function &F);
179
180private:
181 // Returns true if Basis is a basis for C, i.e., Basis dominates C and they
182 // share the same base and stride.
183 bool isBasisFor(const Candidate &Basis, const Candidate &C);
184
185 // Returns whether the candidate can be folded into an addressing mode.
186 bool isFoldable(const Candidate &C, TargetTransformInfo *TTI,
187 const DataLayout *DL);
188
189 // Returns true if C is already in a simplest form and not worth being
190 // rewritten.
191 bool isSimplestForm(const Candidate &C);
192
193 // Checks whether I is in a candidate form. If so, adds all the matching forms
194 // to Candidates, and tries to find the immediate basis for each of them.
195 void allocateCandidatesAndFindBasis(Instruction *I);
196
197 // Allocate candidates and find bases for Add instructions.
199
200 // Given I = LHS + RHS, factors RHS into i * S and makes (LHS + i * S) a
201 // candidate.
202 void allocateCandidatesAndFindBasisForAdd(Value *LHS, Value *RHS,
203 Instruction *I);
204 // Allocate candidates and find bases for Mul instructions.
205 void allocateCandidatesAndFindBasisForMul(Instruction *I);
206
207 // Splits LHS into Base + Index and, if succeeds, calls
208 // allocateCandidatesAndFindBasis.
209 void allocateCandidatesAndFindBasisForMul(Value *LHS, Value *RHS,
210 Instruction *I);
211
212 // Allocate candidates and find bases for GetElementPtr instructions.
213 void allocateCandidatesAndFindBasisForGEP(GetElementPtrInst *GEP);
214
215 // A helper function that scales Idx with ElementSize before invoking
216 // allocateCandidatesAndFindBasis.
217 void allocateCandidatesAndFindBasisForGEP(const SCEV *B, ConstantInt *Idx,
218 Value *S, uint64_t ElementSize,
219 Instruction *I);
220
221 // Adds the given form <CT, B, Idx, S> to Candidates, and finds its immediate
222 // basis.
223 void allocateCandidatesAndFindBasis(Candidate::Kind CT, const SCEV *B,
224 ConstantInt *Idx, Value *S,
225 Instruction *I);
226
227 // Rewrites candidate C with respect to Basis.
228 void rewriteCandidateWithBasis(const Candidate &C, const Candidate &Basis);
229
230 // A helper function that factors ArrayIdx to a product of a stride and a
231 // constant index, and invokes allocateCandidatesAndFindBasis with the
232 // factorings.
233 void factorArrayIndex(Value *ArrayIdx, const SCEV *Base, uint64_t ElementSize,
235
236 // Emit code that computes the "bump" from Basis to C.
237 static Value *emitBump(const Candidate &Basis, const Candidate &C,
238 IRBuilder<> &Builder, const DataLayout *DL);
239
240 const DataLayout *DL = nullptr;
241 DominatorTree *DT = nullptr;
242 ScalarEvolution *SE;
243 TargetTransformInfo *TTI = nullptr;
244 std::list<Candidate> Candidates;
245
246 // Temporarily holds all instructions that are unlinked (but not deleted) by
247 // rewriteCandidateWithBasis. These instructions will be actually removed
248 // after all rewriting finishes.
250};
251
252} // end anonymous namespace
253
254char StraightLineStrengthReduceLegacyPass::ID = 0;
255
256INITIALIZE_PASS_BEGIN(StraightLineStrengthReduceLegacyPass, "slsr",
257 "Straight line strength reduction", false, false)
261INITIALIZE_PASS_END(StraightLineStrengthReduceLegacyPass, "slsr",
262 "Straight line strength reduction", false, false)
263
265 return new StraightLineStrengthReduceLegacyPass();
266}
267
268bool StraightLineStrengthReduce::isBasisFor(const Candidate &Basis,
269 const Candidate &C) {
270 return (Basis.Ins != C.Ins && // skip the same instruction
271 // They must have the same type too. Basis.Base == C.Base doesn't
272 // guarantee their types are the same (PR23975).
273 Basis.Ins->getType() == C.Ins->getType() &&
274 // Basis must dominate C in order to rewrite C with respect to Basis.
275 DT->dominates(Basis.Ins->getParent(), C.Ins->getParent()) &&
276 // They share the same base, stride, and candidate kind.
277 Basis.Base == C.Base && Basis.Stride == C.Stride &&
278 Basis.CandidateKind == C.CandidateKind);
279}
280
282 const TargetTransformInfo *TTI) {
283 SmallVector<const Value *, 4> Indices(GEP->indices());
284 return TTI->getGEPCost(GEP->getSourceElementType(), GEP->getPointerOperand(),
286}
287
288// Returns whether (Base + Index * Stride) can be folded to an addressing mode.
289static bool isAddFoldable(const SCEV *Base, ConstantInt *Index, Value *Stride,
291 // Index->getSExtValue() may crash if Index is wider than 64-bit.
292 return Index->getBitWidth() <= 64 &&
295}
296
297bool StraightLineStrengthReduce::isFoldable(const Candidate &C,
299 const DataLayout *DL) {
301 return isAddFoldable(C.Base, C.Index, C.Stride, TTI);
302 if (C.CandidateKind == Candidate::GEP)
303 return isGEPFoldable(cast<GetElementPtrInst>(C.Ins), TTI);
304 return false;
305}
306
307// Returns true if GEP has zero or one non-zero index.
309 unsigned NumNonZeroIndices = 0;
310 for (Use &Idx : GEP->indices()) {
311 ConstantInt *ConstIdx = dyn_cast<ConstantInt>(Idx);
312 if (ConstIdx == nullptr || !ConstIdx->isZero())
313 ++NumNonZeroIndices;
314 }
315 return NumNonZeroIndices <= 1;
316}
317
318bool StraightLineStrengthReduce::isSimplestForm(const Candidate &C) {
319 if (C.CandidateKind == Candidate::Add) {
320 // B + 1 * S or B + (-1) * S
321 return C.Index->isOne() || C.Index->isMinusOne();
322 }
323 if (C.CandidateKind == Candidate::Mul) {
324 // (B + 0) * S
325 return C.Index->isZero();
326 }
327 if (C.CandidateKind == Candidate::GEP) {
328 // (char*)B + S or (char*)B - S
329 return ((C.Index->isOne() || C.Index->isMinusOne()) &&
330 hasOnlyOneNonZeroIndex(cast<GetElementPtrInst>(C.Ins)));
331 }
332 return false;
333}
334
335// TODO: We currently implement an algorithm whose time complexity is linear in
336// the number of existing candidates. However, we could do better by using
337// ScopedHashTable. Specifically, while traversing the dominator tree, we could
338// maintain all the candidates that dominate the basic block being traversed in
339// a ScopedHashTable. This hash table is indexed by the base and the stride of
340// a candidate. Therefore, finding the immediate basis of a candidate boils down
341// to one hash-table look up.
342void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
343 Candidate::Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
344 Instruction *I) {
345 Candidate C(CT, B, Idx, S, I);
346 // SLSR can complicate an instruction in two cases:
347 //
348 // 1. If we can fold I into an addressing mode, computing I is likely free or
349 // takes only one instruction.
350 //
351 // 2. I is already in a simplest form. For example, when
352 // X = B + 8 * S
353 // Y = B + S,
354 // rewriting Y to X - 7 * S is probably a bad idea.
355 //
356 // In the above cases, we still add I to the candidate list so that I can be
357 // the basis of other candidates, but we leave I's basis blank so that I
358 // won't be rewritten.
359 if (!isFoldable(C, TTI, DL) && !isSimplestForm(C)) {
360 // Try to compute the immediate basis of C.
361 unsigned NumIterations = 0;
362 // Limit the scan radius to avoid running in quadratice time.
363 static const unsigned MaxNumIterations = 50;
364 for (auto Basis = Candidates.rbegin();
365 Basis != Candidates.rend() && NumIterations < MaxNumIterations;
366 ++Basis, ++NumIterations) {
367 if (isBasisFor(*Basis, C)) {
368 C.Basis = &(*Basis);
369 break;
370 }
371 }
372 }
373 // Regardless of whether we find a basis for C, we need to push C to the
374 // candidate list so that it can be the basis of other candidates.
375 Candidates.push_back(C);
376}
377
378void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
379 Instruction *I) {
380 switch (I->getOpcode()) {
383 break;
384 case Instruction::Mul:
385 allocateCandidatesAndFindBasisForMul(I);
386 break;
387 case Instruction::GetElementPtr:
388 allocateCandidatesAndFindBasisForGEP(cast<GetElementPtrInst>(I));
389 break;
390 }
391}
392
394 Instruction *I) {
395 // Try matching B + i * S.
396 if (!isa<IntegerType>(I->getType()))
397 return;
398
399 assert(I->getNumOperands() == 2 && "isn't I an add?");
400 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
402 if (LHS != RHS)
404}
405
407 Value *LHS, Value *RHS, Instruction *I) {
408 Value *S = nullptr;
409 ConstantInt *Idx = nullptr;
410 if (match(RHS, m_Mul(m_Value(S), m_ConstantInt(Idx)))) {
411 // I = LHS + RHS = LHS + Idx * S
412 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
413 } else if (match(RHS, m_Shl(m_Value(S), m_ConstantInt(Idx)))) {
414 // I = LHS + RHS = LHS + (S << Idx) = LHS + S * (1 << Idx)
415 APInt One(Idx->getBitWidth(), 1);
416 Idx = ConstantInt::get(Idx->getContext(), One << Idx->getValue());
417 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
418 } else {
419 // At least, I = LHS + 1 * RHS
420 ConstantInt *One = ConstantInt::get(cast<IntegerType>(I->getType()), 1);
422 I);
423 }
424}
425
426// Returns true if A matches B + C where C is constant.
427static bool matchesAdd(Value *A, Value *&B, ConstantInt *&C) {
429}
430
431// Returns true if A matches B | C where C is constant.
432static bool matchesOr(Value *A, Value *&B, ConstantInt *&C) {
433 return match(A, m_c_Or(m_Value(B), m_ConstantInt(C)));
434}
435
436void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
437 Value *LHS, Value *RHS, Instruction *I) {
438 Value *B = nullptr;
439 ConstantInt *Idx = nullptr;
440 if (matchesAdd(LHS, B, Idx)) {
441 // If LHS is in the form of "Base + Index", then I is in the form of
442 // "(Base + Index) * RHS".
443 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
444 } else if (matchesOr(LHS, B, Idx) && haveNoCommonBitsSet(B, Idx, *DL)) {
445 // If LHS is in the form of "Base | Index" and Base and Index have no common
446 // bits set, then
447 // Base | Index = Base + Index
448 // and I is thus in the form of "(Base + Index) * RHS".
449 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
450 } else {
451 // Otherwise, at least try the form (LHS + 0) * RHS.
452 ConstantInt *Zero = ConstantInt::get(cast<IntegerType>(I->getType()), 0);
453 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(LHS), Zero, RHS,
454 I);
455 }
456}
457
458void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
459 Instruction *I) {
460 // Try matching (B + i) * S.
461 // TODO: we could extend SLSR to float and vector types.
462 if (!isa<IntegerType>(I->getType()))
463 return;
464
465 assert(I->getNumOperands() == 2 && "isn't I a mul?");
466 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
467 allocateCandidatesAndFindBasisForMul(LHS, RHS, I);
468 if (LHS != RHS) {
469 // Symmetrically, try to split RHS to Base + Index.
470 allocateCandidatesAndFindBasisForMul(RHS, LHS, I);
471 }
472}
473
474void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
475 const SCEV *B, ConstantInt *Idx, Value *S, uint64_t ElementSize,
476 Instruction *I) {
477 // I = B + sext(Idx *nsw S) * ElementSize
478 // = B + (sext(Idx) * sext(S)) * ElementSize
479 // = B + (sext(Idx) * ElementSize) * sext(S)
480 // Casting to IntegerType is safe because we skipped vector GEPs.
481 IntegerType *PtrIdxTy = cast<IntegerType>(DL->getIndexType(I->getType()));
482 ConstantInt *ScaledIdx = ConstantInt::get(
483 PtrIdxTy, Idx->getSExtValue() * (int64_t)ElementSize, true);
484 allocateCandidatesAndFindBasis(Candidate::GEP, B, ScaledIdx, S, I);
485}
486
487void StraightLineStrengthReduce::factorArrayIndex(Value *ArrayIdx,
488 const SCEV *Base,
489 uint64_t ElementSize,
491 // At least, ArrayIdx = ArrayIdx *nsw 1.
492 allocateCandidatesAndFindBasisForGEP(
493 Base, ConstantInt::get(cast<IntegerType>(ArrayIdx->getType()), 1),
494 ArrayIdx, ElementSize, GEP);
495 Value *LHS = nullptr;
496 ConstantInt *RHS = nullptr;
497 // One alternative is matching the SCEV of ArrayIdx instead of ArrayIdx
498 // itself. This would allow us to handle the shl case for free. However,
499 // matching SCEVs has two issues:
500 //
501 // 1. this would complicate rewriting because the rewriting procedure
502 // would have to translate SCEVs back to IR instructions. This translation
503 // is difficult when LHS is further evaluated to a composite SCEV.
504 //
505 // 2. ScalarEvolution is designed to be control-flow oblivious. It tends
506 // to strip nsw/nuw flags which are critical for SLSR to trace into
507 // sext'ed multiplication.
508 if (match(ArrayIdx, m_NSWMul(m_Value(LHS), m_ConstantInt(RHS)))) {
509 // SLSR is currently unsafe if i * S may overflow.
510 // GEP = Base + sext(LHS *nsw RHS) * ElementSize
511 allocateCandidatesAndFindBasisForGEP(Base, RHS, LHS, ElementSize, GEP);
512 } else if (match(ArrayIdx, m_NSWShl(m_Value(LHS), m_ConstantInt(RHS)))) {
513 // GEP = Base + sext(LHS <<nsw RHS) * ElementSize
514 // = Base + sext(LHS *nsw (1 << RHS)) * ElementSize
515 APInt One(RHS->getBitWidth(), 1);
516 ConstantInt *PowerOf2 =
517 ConstantInt::get(RHS->getContext(), One << RHS->getValue());
518 allocateCandidatesAndFindBasisForGEP(Base, PowerOf2, LHS, ElementSize, GEP);
519 }
520}
521
522void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
524 // TODO: handle vector GEPs
525 if (GEP->getType()->isVectorTy())
526 return;
527
529 for (Use &Idx : GEP->indices())
530 IndexExprs.push_back(SE->getSCEV(Idx));
531
533 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
534 if (GTI.isStruct())
535 continue;
536
537 const SCEV *OrigIndexExpr = IndexExprs[I - 1];
538 IndexExprs[I - 1] = SE->getZero(OrigIndexExpr->getType());
539
540 // The base of this candidate is GEP's base plus the offsets of all
541 // indices except this current one.
542 const SCEV *BaseExpr = SE->getGEPExpr(cast<GEPOperator>(GEP), IndexExprs);
543 Value *ArrayIdx = GEP->getOperand(I);
544 uint64_t ElementSize = GTI.getSequentialElementStride(*DL);
545 if (ArrayIdx->getType()->getIntegerBitWidth() <=
547 // Skip factoring if ArrayIdx is wider than the index size, because
548 // ArrayIdx is implicitly truncated to the index size.
549 factorArrayIndex(ArrayIdx, BaseExpr, ElementSize, GEP);
550 }
551 // When ArrayIdx is the sext of a value, we try to factor that value as
552 // well. Handling this case is important because array indices are
553 // typically sign-extended to the pointer index size.
554 Value *TruncatedArrayIdx = nullptr;
555 if (match(ArrayIdx, m_SExt(m_Value(TruncatedArrayIdx))) &&
556 TruncatedArrayIdx->getType()->getIntegerBitWidth() <=
558 // Skip factoring if TruncatedArrayIdx is wider than the pointer size,
559 // because TruncatedArrayIdx is implicitly truncated to the pointer size.
560 factorArrayIndex(TruncatedArrayIdx, BaseExpr, ElementSize, GEP);
561 }
562
563 IndexExprs[I - 1] = OrigIndexExpr;
564 }
565}
566
567// A helper function that unifies the bitwidth of A and B.
568static void unifyBitWidth(APInt &A, APInt &B) {
569 if (A.getBitWidth() < B.getBitWidth())
570 A = A.sext(B.getBitWidth());
571 else if (A.getBitWidth() > B.getBitWidth())
572 B = B.sext(A.getBitWidth());
573}
574
575Value *StraightLineStrengthReduce::emitBump(const Candidate &Basis,
576 const Candidate &C,
577 IRBuilder<> &Builder,
578 const DataLayout *DL) {
579 APInt Idx = C.Index->getValue(), BasisIdx = Basis.Index->getValue();
580 unifyBitWidth(Idx, BasisIdx);
581 APInt IndexOffset = Idx - BasisIdx;
582
583 // Compute Bump = C - Basis = (i' - i) * S.
584 // Common case 1: if (i' - i) is 1, Bump = S.
585 if (IndexOffset == 1)
586 return C.Stride;
587 // Common case 2: if (i' - i) is -1, Bump = -S.
588 if (IndexOffset.isAllOnes())
589 return Builder.CreateNeg(C.Stride);
590
591 // Otherwise, Bump = (i' - i) * sext/trunc(S). Note that (i' - i) and S may
592 // have different bit widths.
593 IntegerType *DeltaType =
594 IntegerType::get(Basis.Ins->getContext(), IndexOffset.getBitWidth());
595 Value *ExtendedStride = Builder.CreateSExtOrTrunc(C.Stride, DeltaType);
596 if (IndexOffset.isPowerOf2()) {
597 // If (i' - i) is a power of 2, Bump = sext/trunc(S) << log(i' - i).
598 ConstantInt *Exponent = ConstantInt::get(DeltaType, IndexOffset.logBase2());
599 return Builder.CreateShl(ExtendedStride, Exponent);
600 }
601 if (IndexOffset.isNegatedPowerOf2()) {
602 // If (i - i') is a power of 2, Bump = -sext/trunc(S) << log(i' - i).
604 ConstantInt::get(DeltaType, (-IndexOffset).logBase2());
605 return Builder.CreateNeg(Builder.CreateShl(ExtendedStride, Exponent));
606 }
607 Constant *Delta = ConstantInt::get(DeltaType, IndexOffset);
608 return Builder.CreateMul(ExtendedStride, Delta);
609}
610
611void StraightLineStrengthReduce::rewriteCandidateWithBasis(
612 const Candidate &C, const Candidate &Basis) {
613 assert(C.CandidateKind == Basis.CandidateKind && C.Base == Basis.Base &&
614 C.Stride == Basis.Stride);
615 // We run rewriteCandidateWithBasis on all candidates in a post-order, so the
616 // basis of a candidate cannot be unlinked before the candidate.
617 assert(Basis.Ins->getParent() != nullptr && "the basis is unlinked");
618
619 // An instruction can correspond to multiple candidates. Therefore, instead of
620 // simply deleting an instruction when we rewrite it, we mark its parent as
621 // nullptr (i.e. unlink it) so that we can skip the candidates whose
622 // instruction is already rewritten.
623 if (!C.Ins->getParent())
624 return;
625
626 IRBuilder<> Builder(C.Ins);
627 Value *Bump = emitBump(Basis, C, Builder, DL);
628 Value *Reduced = nullptr; // equivalent to but weaker than C.Ins
629 switch (C.CandidateKind) {
631 case Candidate::Mul: {
632 // C = Basis + Bump
633 Value *NegBump;
634 if (match(Bump, m_Neg(m_Value(NegBump)))) {
635 // If Bump is a neg instruction, emit C = Basis - (-Bump).
636 Reduced = Builder.CreateSub(Basis.Ins, NegBump);
637 // We only use the negative argument of Bump, and Bump itself may be
640 } else {
641 // It's tempting to preserve nsw on Bump and/or Reduced. However, it's
642 // usually unsound, e.g.,
643 //
644 // X = (-2 +nsw 1) *nsw INT_MAX
645 // Y = (-2 +nsw 3) *nsw INT_MAX
646 // =>
647 // Y = X + 2 * INT_MAX
648 //
649 // Neither + and * in the resultant expression are nsw.
651 }
652 break;
653 }
654 case Candidate::GEP: {
655 bool InBounds = cast<GetElementPtrInst>(C.Ins)->isInBounds();
656 // C = (char *)Basis + Bump
657 Reduced = Builder.CreatePtrAdd(Basis.Ins, Bump, "", InBounds);
658 break;
659 }
660 default:
661 llvm_unreachable("C.CandidateKind is invalid");
662 };
663 Reduced->takeName(C.Ins);
664 C.Ins->replaceAllUsesWith(Reduced);
665 // Unlink C.Ins so that we can skip other candidates also corresponding to
666 // C.Ins. The actual deletion is postponed to the end of runOnFunction.
667 C.Ins->removeFromParent();
669}
670
671bool StraightLineStrengthReduceLegacyPass::runOnFunction(Function &F) {
672 if (skipFunction(F))
673 return false;
674
675 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
676 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
677 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
678 return StraightLineStrengthReduce(DL, DT, SE, TTI).runOnFunction(F);
679}
680
681bool StraightLineStrengthReduce::runOnFunction(Function &F) {
682 // Traverse the dominator tree in the depth-first order. This order makes sure
683 // all bases of a candidate are in Candidates when we process it.
684 for (const auto Node : depth_first(DT))
685 for (auto &I : *(Node->getBlock()))
686 allocateCandidatesAndFindBasis(&I);
687
688 // Rewrite candidates in the reverse depth-first order. This order makes sure
689 // a candidate being rewritten is not a basis for any other candidate.
690 while (!Candidates.empty()) {
691 const Candidate &C = Candidates.back();
692 if (C.Basis != nullptr) {
693 rewriteCandidateWithBasis(C, *C.Basis);
694 }
695 Candidates.pop_back();
696 }
697
698 // Delete all unlink instructions.
700 for (unsigned I = 0, E = UnlinkedInst->getNumOperands(); I != E; ++I) {
704 }
706 }
709 return Ret;
710}
711
712namespace llvm {
713
716 const DataLayout *DL = &F.getDataLayout();
717 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
718 auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
719 auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
720
721 if (!StraightLineStrengthReduce(DL, DT, SE, TTI).runOnFunction(F))
722 return PreservedAnalyses::all();
723
729 return PA;
730}
731
732} // namespace llvm
This file implements a class to represent arbitrary precision integral constant values and operations...
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
This file builds on the ADT/GraphTraits.h file to build generic depth first graph iterator.
static bool runOnFunction(Function &F, bool PostInlining)
Hexagon Common GEP
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
Module.h This file contains the declarations for the Module class.
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:55
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:57
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:52
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file defines the SmallVector class.
static bool matchesOr(Value *A, Value *&B, ConstantInt *&C)
static bool isAddFoldable(const SCEV *Base, ConstantInt *Index, Value *Stride, TargetTransformInfo *TTI)
static bool hasOnlyOneNonZeroIndex(GetElementPtrInst *GEP)
static void unifyBitWidth(APInt &A, APInt &B)
static bool matchesAdd(Value *A, Value *&B, ConstantInt *&C)
static bool isGEPFoldable(GetElementPtrInst *GEP, const TargetTransformInfo *TTI)
Straight line strength reduction
This pass exposes codegen information to IR-level passes.
Value * RHS
Value * LHS
BinaryOperator * Mul
Class for arbitrary precision integers.
Definition: APInt.h:78
bool isNegatedPowerOf2() const
Check if this APInt's negated value is a power of two greater than zero.
Definition: APInt.h:429
bool isAllOnes() const
Determine if all bits are set. This is true for zero-width values.
Definition: APInt.h:351
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1448
unsigned logBase2() const
Definition: APInt.h:1719
bool isPowerOf2() const
Check if this APInt's value is a power of two greater than zero.
Definition: APInt.h:420
A container for analyses that lazily runs them and caches their results.
Definition: PassManager.h:253
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:405
Represent the analysis usage information of a pass.
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:256
Represents analyses that only rely on functions' control flow.
Definition: Analysis.h:72
This is the shared class of boolean and integer constants.
Definition: Constants.h:81
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:206
This is an important base class in LLVM.
Definition: Constant.h:42
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:109
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:279
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:317
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Definition: Dominators.h:162
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:310
virtual bool runOnFunction(Function &F)=0
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass.
an instruction for type-safe pointer arithmetic to access elements of arrays and structs
Definition: Instructions.h:915
Value * CreatePtrAdd(Value *Ptr, Value *Offset, const Twine &Name="", GEPNoWrapFlags NW=GEPNoWrapFlags::none())
Definition: IRBuilder.h:1984
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNSW=false)
Definition: IRBuilder.h:1726
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1349
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1421
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1332
Value * CreateSExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a SExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:2062
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1366
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:2674
Class to represent integer types.
Definition: DerivedTypes.h:40
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:266
A Module instance is used to store all the information related to an LLVM module.
Definition: Module.h:65
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
virtual void getAnalysisUsage(AnalysisUsage &) const
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: Pass.cpp:98
virtual bool doInitialization(Module &)
doInitialization - Virtual method overridden by subclasses to do any necessary initialization before ...
Definition: Pass.h:119
A set of analyses that are preserved following a run of a transformation pass.
Definition: Analysis.h:111
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: Analysis.h:117
void preserveSet()
Mark an analysis set as preserved.
Definition: Analysis.h:146
void preserve()
Mark an analysis as preserved.
Definition: Analysis.h:131
This class represents an analyzed expression in the program.
Type * getType() const
Return the LLVM type of this SCEV expression.
Analysis pass that exposes the ScalarEvolution for a function.
The main scalar evolution driver.
void push_back(const T &Elt)
Definition: SmallVector.h:427
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1210
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Analysis pass providing the TargetTransformInfo.
Wrapper pass for TargetTransformInfo.
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
InstructionCost getGEPCost(Type *PointeeType, const Value *Ptr, ArrayRef< const Value * > Operands, Type *AccessType=nullptr, TargetCostKind CostKind=TCK_SizeAndLatency) const
Estimate the cost of a GEP operation when lowered.
bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace=0, Instruction *I=nullptr, int64_t ScalableOffset=0) const
Return true if the addressing mode represented by AM is legal for this target, for a load/store of th...
@ TCC_Free
Expected to fold away in lowering.
unsigned getIntegerBitWidth() const
A Use represents the edge between a Value definition and its users.
Definition: Use.h:43
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:1075
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:383
TypeSize getSequentialElementStride(const DataLayout &DL) const
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:168
BinaryOp_match< LHS, RHS, Instruction::Mul > m_Mul(const LHS &L, const RHS &R)
BinaryOp_match< cst_pred_ty< is_zero_int >, ValTy, Instruction::Sub > m_Neg(const ValTy &V)
Matches a 'Neg' as 'sub 0, V'.
OverflowingBinaryOp_match< LHS, RHS, Instruction::Shl, OverflowingBinaryOperator::NoSignedWrap > m_NSWShl(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)
Matches a Add with LHS and RHS in either order.
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:92
BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)
CastInst_match< OpTy, SExtInst > m_SExt(const OpTy &Op)
Matches SExt.
BinaryOp_match< LHS, RHS, Instruction::Or, true > m_c_Or(const LHS &L, const RHS &R)
Matches an Or with LHS and RHS in either order.
OverflowingBinaryOp_match< LHS, RHS, Instruction::Mul, OverflowingBinaryOperator::NoSignedWrap > m_NSWMul(const LHS &L, const RHS &R)
This is an optimization pass for GlobalISel generic memory operations.