1 //===- StraightLineStrengthReduce.cpp - -----------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
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 #include "llvm/Transforms/Scalar/StraightLineStrengthReduce.h"
59 #include "llvm/ADT/APInt.h"
60 #include "llvm/ADT/DepthFirstIterator.h"
61 #include "llvm/ADT/SmallVector.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/TargetTransformInfo.h"
64 #include "llvm/Analysis/ValueTracking.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/GetElementPtrTypeIterator.h"
70 #include "llvm/IR/IRBuilder.h"
71 #include "llvm/IR/Instruction.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/Module.h"
74 #include "llvm/IR/Operator.h"
75 #include "llvm/IR/PatternMatch.h"
76 #include "llvm/IR/Type.h"
77 #include "llvm/IR/Value.h"
78 #include "llvm/InitializePasses.h"
79 #include "llvm/Pass.h"
80 #include "llvm/Support/Casting.h"
81 #include "llvm/Support/ErrorHandling.h"
82 #include "llvm/Transforms/Scalar.h"
83 #include "llvm/Transforms/Utils/Local.h"
84 #include <cassert>
85 #include <cstdint>
86 #include <limits>
87 #include <list>
88 #include <vector>
89 
90 using namespace llvm;
91 using namespace PatternMatch;
92 
93 static const unsigned UnknownAddressSpace =
94     std::numeric_limits<unsigned>::max();
95 
96 namespace {
97 
98 class StraightLineStrengthReduceLegacyPass : public FunctionPass {
99   const DataLayout *DL = nullptr;
100 
101 public:
102   static char ID;
103 
104   StraightLineStrengthReduceLegacyPass() : FunctionPass(ID) {
105     initializeStraightLineStrengthReduceLegacyPassPass(
106         *PassRegistry::getPassRegistry());
107   }
108 
109   void getAnalysisUsage(AnalysisUsage &AU) const override {
110     AU.addRequired<DominatorTreeWrapperPass>();
111     AU.addRequired<ScalarEvolutionWrapperPass>();
112     AU.addRequired<TargetTransformInfoWrapperPass>();
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 
125 class StraightLineStrengthReduce {
126 public:
127   StraightLineStrengthReduce(const DataLayout *DL, DominatorTree *DT,
128                              ScalarEvolution *SE, TargetTransformInfo *TTI)
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
132   // the header comments.
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 
180 private:
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.
198   void allocateCandidatesAndFindBasisForAdd(Instruction *I);
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,
234                         GetElementPtrInst *GEP);
235 
236   // Emit code that computes the "bump" from Basis to C. If the candidate is a
237   // GEP and the bump is not divisible by the element size of the GEP, this
238   // function sets the BumpWithUglyGEP flag to notify its caller to bump the
239   // basis using an ugly GEP.
240   static Value *emitBump(const Candidate &Basis, const Candidate &C,
241                          IRBuilder<> &Builder, const DataLayout *DL,
242                          bool &BumpWithUglyGEP);
243 
244   const DataLayout *DL = nullptr;
245   DominatorTree *DT = nullptr;
246   ScalarEvolution *SE;
247   TargetTransformInfo *TTI = nullptr;
248   std::list<Candidate> Candidates;
249 
250   // Temporarily holds all instructions that are unlinked (but not deleted) by
251   // rewriteCandidateWithBasis. These instructions will be actually removed
252   // after all rewriting finishes.
253   std::vector<Instruction *> UnlinkedInstructions;
254 };
255 
256 } // end anonymous namespace
257 
258 char StraightLineStrengthReduceLegacyPass::ID = 0;
259 
260 INITIALIZE_PASS_BEGIN(StraightLineStrengthReduceLegacyPass, "slsr",
261                       "Straight line strength reduction", false, false)
262 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
263 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
264 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
265 INITIALIZE_PASS_END(StraightLineStrengthReduceLegacyPass, "slsr",
266                     "Straight line strength reduction", false, false)
267 
268 FunctionPass *llvm::createStraightLineStrengthReducePass() {
269   return new StraightLineStrengthReduceLegacyPass();
270 }
271 
272 bool StraightLineStrengthReduce::isBasisFor(const Candidate &Basis,
273                                             const Candidate &C) {
274   return (Basis.Ins != C.Ins && // skip the same instruction
275           // They must have the same type too. Basis.Base == C.Base doesn't
276           // guarantee their types are the same (PR23975).
277           Basis.Ins->getType() == C.Ins->getType() &&
278           // Basis must dominate C in order to rewrite C with respect to Basis.
279           DT->dominates(Basis.Ins->getParent(), C.Ins->getParent()) &&
280           // They share the same base, stride, and candidate kind.
281           Basis.Base == C.Base && Basis.Stride == C.Stride &&
282           Basis.CandidateKind == C.CandidateKind);
283 }
284 
285 static bool isGEPFoldable(GetElementPtrInst *GEP,
286                           const TargetTransformInfo *TTI) {
287   SmallVector<const Value *, 4> Indices(GEP->indices());
288   return TTI->getGEPCost(GEP->getSourceElementType(), GEP->getPointerOperand(),
289                          Indices) == TargetTransformInfo::TCC_Free;
290 }
291 
292 // Returns whether (Base + Index * Stride) can be folded to an addressing mode.
293 static bool isAddFoldable(const SCEV *Base, ConstantInt *Index, Value *Stride,
294                           TargetTransformInfo *TTI) {
295   // Index->getSExtValue() may crash if Index is wider than 64-bit.
296   return Index->getBitWidth() <= 64 &&
297          TTI->isLegalAddressingMode(Base->getType(), nullptr, 0, true,
298                                     Index->getSExtValue(), UnknownAddressSpace);
299 }
300 
301 bool StraightLineStrengthReduce::isFoldable(const Candidate &C,
302                                             TargetTransformInfo *TTI,
303                                             const DataLayout *DL) {
304   if (C.CandidateKind == Candidate::Add)
305     return isAddFoldable(C.Base, C.Index, C.Stride, TTI);
306   if (C.CandidateKind == Candidate::GEP)
307     return isGEPFoldable(cast<GetElementPtrInst>(C.Ins), TTI);
308   return false;
309 }
310 
311 // Returns true if GEP has zero or one non-zero index.
312 static bool hasOnlyOneNonZeroIndex(GetElementPtrInst *GEP) {
313   unsigned NumNonZeroIndices = 0;
314   for (Use &Idx : GEP->indices()) {
315     ConstantInt *ConstIdx = dyn_cast<ConstantInt>(Idx);
316     if (ConstIdx == nullptr || !ConstIdx->isZero())
317       ++NumNonZeroIndices;
318   }
319   return NumNonZeroIndices <= 1;
320 }
321 
322 bool StraightLineStrengthReduce::isSimplestForm(const Candidate &C) {
323   if (C.CandidateKind == Candidate::Add) {
324     // B + 1 * S or B + (-1) * S
325     return C.Index->isOne() || C.Index->isMinusOne();
326   }
327   if (C.CandidateKind == Candidate::Mul) {
328     // (B + 0) * S
329     return C.Index->isZero();
330   }
331   if (C.CandidateKind == Candidate::GEP) {
332     // (char*)B + S or (char*)B - S
333     return ((C.Index->isOne() || C.Index->isMinusOne()) &&
334             hasOnlyOneNonZeroIndex(cast<GetElementPtrInst>(C.Ins)));
335   }
336   return false;
337 }
338 
339 // TODO: We currently implement an algorithm whose time complexity is linear in
340 // the number of existing candidates. However, we could do better by using
341 // ScopedHashTable. Specifically, while traversing the dominator tree, we could
342 // maintain all the candidates that dominate the basic block being traversed in
343 // a ScopedHashTable. This hash table is indexed by the base and the stride of
344 // a candidate. Therefore, finding the immediate basis of a candidate boils down
345 // to one hash-table look up.
346 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
347     Candidate::Kind CT, const SCEV *B, ConstantInt *Idx, Value *S,
348     Instruction *I) {
349   Candidate C(CT, B, Idx, S, I);
350   // SLSR can complicate an instruction in two cases:
351   //
352   // 1. If we can fold I into an addressing mode, computing I is likely free or
353   // takes only one instruction.
354   //
355   // 2. I is already in a simplest form. For example, when
356   //      X = B + 8 * S
357   //      Y = B + S,
358   //    rewriting Y to X - 7 * S is probably a bad idea.
359   //
360   // In the above cases, we still add I to the candidate list so that I can be
361   // the basis of other candidates, but we leave I's basis blank so that I
362   // won't be rewritten.
363   if (!isFoldable(C, TTI, DL) && !isSimplestForm(C)) {
364     // Try to compute the immediate basis of C.
365     unsigned NumIterations = 0;
366     // Limit the scan radius to avoid running in quadratice time.
367     static const unsigned MaxNumIterations = 50;
368     for (auto Basis = Candidates.rbegin();
369          Basis != Candidates.rend() && NumIterations < MaxNumIterations;
370          ++Basis, ++NumIterations) {
371       if (isBasisFor(*Basis, C)) {
372         C.Basis = &(*Basis);
373         break;
374       }
375     }
376   }
377   // Regardless of whether we find a basis for C, we need to push C to the
378   // candidate list so that it can be the basis of other candidates.
379   Candidates.push_back(C);
380 }
381 
382 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis(
383     Instruction *I) {
384   switch (I->getOpcode()) {
385   case Instruction::Add:
386     allocateCandidatesAndFindBasisForAdd(I);
387     break;
388   case Instruction::Mul:
389     allocateCandidatesAndFindBasisForMul(I);
390     break;
391   case Instruction::GetElementPtr:
392     allocateCandidatesAndFindBasisForGEP(cast<GetElementPtrInst>(I));
393     break;
394   }
395 }
396 
397 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd(
398     Instruction *I) {
399   // Try matching B + i * S.
400   if (!isa<IntegerType>(I->getType()))
401     return;
402 
403   assert(I->getNumOperands() == 2 && "isn't I an add?");
404   Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
405   allocateCandidatesAndFindBasisForAdd(LHS, RHS, I);
406   if (LHS != RHS)
407     allocateCandidatesAndFindBasisForAdd(RHS, LHS, I);
408 }
409 
410 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd(
411     Value *LHS, Value *RHS, Instruction *I) {
412   Value *S = nullptr;
413   ConstantInt *Idx = nullptr;
414   if (match(RHS, m_Mul(m_Value(S), m_ConstantInt(Idx)))) {
415     // I = LHS + RHS = LHS + Idx * S
416     allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
417   } else if (match(RHS, m_Shl(m_Value(S), m_ConstantInt(Idx)))) {
418     // I = LHS + RHS = LHS + (S << Idx) = LHS + S * (1 << Idx)
419     APInt One(Idx->getBitWidth(), 1);
420     Idx = ConstantInt::get(Idx->getContext(), One << Idx->getValue());
421     allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I);
422   } else {
423     // At least, I = LHS + 1 * RHS
424     ConstantInt *One = ConstantInt::get(cast<IntegerType>(I->getType()), 1);
425     allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), One, RHS,
426                                    I);
427   }
428 }
429 
430 // Returns true if A matches B + C where C is constant.
431 static bool matchesAdd(Value *A, Value *&B, ConstantInt *&C) {
432   return (match(A, m_Add(m_Value(B), m_ConstantInt(C))) ||
433           match(A, m_Add(m_ConstantInt(C), m_Value(B))));
434 }
435 
436 // Returns true if A matches B | C where C is constant.
437 static bool matchesOr(Value *A, Value *&B, ConstantInt *&C) {
438   return (match(A, m_Or(m_Value(B), m_ConstantInt(C))) ||
439           match(A, m_Or(m_ConstantInt(C), m_Value(B))));
440 }
441 
442 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
443     Value *LHS, Value *RHS, Instruction *I) {
444   Value *B = nullptr;
445   ConstantInt *Idx = nullptr;
446   if (matchesAdd(LHS, B, Idx)) {
447     // If LHS is in the form of "Base + Index", then I is in the form of
448     // "(Base + Index) * RHS".
449     allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
450   } else if (matchesOr(LHS, B, Idx) && haveNoCommonBitsSet(B, Idx, *DL)) {
451     // If LHS is in the form of "Base | Index" and Base and Index have no common
452     // bits set, then
453     //   Base | Index = Base + Index
454     // and I is thus in the form of "(Base + Index) * RHS".
455     allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I);
456   } else {
457     // Otherwise, at least try the form (LHS + 0) * RHS.
458     ConstantInt *Zero = ConstantInt::get(cast<IntegerType>(I->getType()), 0);
459     allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(LHS), Zero, RHS,
460                                    I);
461   }
462 }
463 
464 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul(
465     Instruction *I) {
466   // Try matching (B + i) * S.
467   // TODO: we could extend SLSR to float and vector types.
468   if (!isa<IntegerType>(I->getType()))
469     return;
470 
471   assert(I->getNumOperands() == 2 && "isn't I a mul?");
472   Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
473   allocateCandidatesAndFindBasisForMul(LHS, RHS, I);
474   if (LHS != RHS) {
475     // Symmetrically, try to split RHS to Base + Index.
476     allocateCandidatesAndFindBasisForMul(RHS, LHS, I);
477   }
478 }
479 
480 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
481     const SCEV *B, ConstantInt *Idx, Value *S, uint64_t ElementSize,
482     Instruction *I) {
483   // I = B + sext(Idx *nsw S) * ElementSize
484   //   = B + (sext(Idx) * sext(S)) * ElementSize
485   //   = B + (sext(Idx) * ElementSize) * sext(S)
486   // Casting to IntegerType is safe because we skipped vector GEPs.
487   IntegerType *PtrIdxTy = cast<IntegerType>(DL->getIndexType(I->getType()));
488   ConstantInt *ScaledIdx = ConstantInt::get(
489       PtrIdxTy, Idx->getSExtValue() * (int64_t)ElementSize, true);
490   allocateCandidatesAndFindBasis(Candidate::GEP, B, ScaledIdx, S, I);
491 }
492 
493 void StraightLineStrengthReduce::factorArrayIndex(Value *ArrayIdx,
494                                                   const SCEV *Base,
495                                                   uint64_t ElementSize,
496                                                   GetElementPtrInst *GEP) {
497   // At least, ArrayIdx = ArrayIdx *nsw 1.
498   allocateCandidatesAndFindBasisForGEP(
499       Base, ConstantInt::get(cast<IntegerType>(ArrayIdx->getType()), 1),
500       ArrayIdx, ElementSize, GEP);
501   Value *LHS = nullptr;
502   ConstantInt *RHS = nullptr;
503   // One alternative is matching the SCEV of ArrayIdx instead of ArrayIdx
504   // itself. This would allow us to handle the shl case for free. However,
505   // matching SCEVs has two issues:
506   //
507   // 1. this would complicate rewriting because the rewriting procedure
508   // would have to translate SCEVs back to IR instructions. This translation
509   // is difficult when LHS is further evaluated to a composite SCEV.
510   //
511   // 2. ScalarEvolution is designed to be control-flow oblivious. It tends
512   // to strip nsw/nuw flags which are critical for SLSR to trace into
513   // sext'ed multiplication.
514   if (match(ArrayIdx, m_NSWMul(m_Value(LHS), m_ConstantInt(RHS)))) {
515     // SLSR is currently unsafe if i * S may overflow.
516     // GEP = Base + sext(LHS *nsw RHS) * ElementSize
517     allocateCandidatesAndFindBasisForGEP(Base, RHS, LHS, ElementSize, GEP);
518   } else if (match(ArrayIdx, m_NSWShl(m_Value(LHS), m_ConstantInt(RHS)))) {
519     // GEP = Base + sext(LHS <<nsw RHS) * ElementSize
520     //     = Base + sext(LHS *nsw (1 << RHS)) * ElementSize
521     APInt One(RHS->getBitWidth(), 1);
522     ConstantInt *PowerOf2 =
523         ConstantInt::get(RHS->getContext(), One << RHS->getValue());
524     allocateCandidatesAndFindBasisForGEP(Base, PowerOf2, LHS, ElementSize, GEP);
525   }
526 }
527 
528 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP(
529     GetElementPtrInst *GEP) {
530   // TODO: handle vector GEPs
531   if (GEP->getType()->isVectorTy())
532     return;
533 
534   SmallVector<const SCEV *, 4> IndexExprs;
535   for (Use &Idx : GEP->indices())
536     IndexExprs.push_back(SE->getSCEV(Idx));
537 
538   gep_type_iterator GTI = gep_type_begin(GEP);
539   for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
540     if (GTI.isStruct())
541       continue;
542 
543     const SCEV *OrigIndexExpr = IndexExprs[I - 1];
544     IndexExprs[I - 1] = SE->getZero(OrigIndexExpr->getType());
545 
546     // The base of this candidate is GEP's base plus the offsets of all
547     // indices except this current one.
548     const SCEV *BaseExpr = SE->getGEPExpr(cast<GEPOperator>(GEP), IndexExprs);
549     Value *ArrayIdx = GEP->getOperand(I);
550     uint64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
551     if (ArrayIdx->getType()->getIntegerBitWidth() <=
552         DL->getIndexSizeInBits(GEP->getAddressSpace())) {
553       // Skip factoring if ArrayIdx is wider than the index size, because
554       // ArrayIdx is implicitly truncated to the index size.
555       factorArrayIndex(ArrayIdx, BaseExpr, ElementSize, GEP);
556     }
557     // When ArrayIdx is the sext of a value, we try to factor that value as
558     // well.  Handling this case is important because array indices are
559     // typically sign-extended to the pointer index size.
560     Value *TruncatedArrayIdx = nullptr;
561     if (match(ArrayIdx, m_SExt(m_Value(TruncatedArrayIdx))) &&
562         TruncatedArrayIdx->getType()->getIntegerBitWidth() <=
563             DL->getIndexSizeInBits(GEP->getAddressSpace())) {
564       // Skip factoring if TruncatedArrayIdx is wider than the pointer size,
565       // because TruncatedArrayIdx is implicitly truncated to the pointer size.
566       factorArrayIndex(TruncatedArrayIdx, BaseExpr, ElementSize, GEP);
567     }
568 
569     IndexExprs[I - 1] = OrigIndexExpr;
570   }
571 }
572 
573 // A helper function that unifies the bitwidth of A and B.
574 static void unifyBitWidth(APInt &A, APInt &B) {
575   if (A.getBitWidth() < B.getBitWidth())
576     A = A.sext(B.getBitWidth());
577   else if (A.getBitWidth() > B.getBitWidth())
578     B = B.sext(A.getBitWidth());
579 }
580 
581 Value *StraightLineStrengthReduce::emitBump(const Candidate &Basis,
582                                             const Candidate &C,
583                                             IRBuilder<> &Builder,
584                                             const DataLayout *DL,
585                                             bool &BumpWithUglyGEP) {
586   APInt Idx = C.Index->getValue(), BasisIdx = Basis.Index->getValue();
587   unifyBitWidth(Idx, BasisIdx);
588   APInt IndexOffset = Idx - BasisIdx;
589 
590   BumpWithUglyGEP = false;
591   if (Basis.CandidateKind == Candidate::GEP) {
592     APInt ElementSize(
593         IndexOffset.getBitWidth(),
594         DL->getTypeAllocSize(
595             cast<GetElementPtrInst>(Basis.Ins)->getResultElementType()));
596     APInt Q, R;
597     APInt::sdivrem(IndexOffset, ElementSize, Q, R);
598     if (R == 0)
599       IndexOffset = Q;
600     else
601       BumpWithUglyGEP = true;
602   }
603 
604   // Compute Bump = C - Basis = (i' - i) * S.
605   // Common case 1: if (i' - i) is 1, Bump = S.
606   if (IndexOffset == 1)
607     return C.Stride;
608   // Common case 2: if (i' - i) is -1, Bump = -S.
609   if (IndexOffset.isAllOnes())
610     return Builder.CreateNeg(C.Stride);
611 
612   // Otherwise, Bump = (i' - i) * sext/trunc(S). Note that (i' - i) and S may
613   // have different bit widths.
614   IntegerType *DeltaType =
615       IntegerType::get(Basis.Ins->getContext(), IndexOffset.getBitWidth());
616   Value *ExtendedStride = Builder.CreateSExtOrTrunc(C.Stride, DeltaType);
617   if (IndexOffset.isPowerOf2()) {
618     // If (i' - i) is a power of 2, Bump = sext/trunc(S) << log(i' - i).
619     ConstantInt *Exponent = ConstantInt::get(DeltaType, IndexOffset.logBase2());
620     return Builder.CreateShl(ExtendedStride, Exponent);
621   }
622   if (IndexOffset.isNegatedPowerOf2()) {
623     // If (i - i') is a power of 2, Bump = -sext/trunc(S) << log(i' - i).
624     ConstantInt *Exponent =
625         ConstantInt::get(DeltaType, (-IndexOffset).logBase2());
626     return Builder.CreateNeg(Builder.CreateShl(ExtendedStride, Exponent));
627   }
628   Constant *Delta = ConstantInt::get(DeltaType, IndexOffset);
629   return Builder.CreateMul(ExtendedStride, Delta);
630 }
631 
632 void StraightLineStrengthReduce::rewriteCandidateWithBasis(
633     const Candidate &C, const Candidate &Basis) {
634   assert(C.CandidateKind == Basis.CandidateKind && C.Base == Basis.Base &&
635          C.Stride == Basis.Stride);
636   // We run rewriteCandidateWithBasis on all candidates in a post-order, so the
637   // basis of a candidate cannot be unlinked before the candidate.
638   assert(Basis.Ins->getParent() != nullptr && "the basis is unlinked");
639 
640   // An instruction can correspond to multiple candidates. Therefore, instead of
641   // simply deleting an instruction when we rewrite it, we mark its parent as
642   // nullptr (i.e. unlink it) so that we can skip the candidates whose
643   // instruction is already rewritten.
644   if (!C.Ins->getParent())
645     return;
646 
647   IRBuilder<> Builder(C.Ins);
648   bool BumpWithUglyGEP;
649   Value *Bump = emitBump(Basis, C, Builder, DL, BumpWithUglyGEP);
650   Value *Reduced = nullptr; // equivalent to but weaker than C.Ins
651   switch (C.CandidateKind) {
652   case Candidate::Add:
653   case Candidate::Mul: {
654     // C = Basis + Bump
655     Value *NegBump;
656     if (match(Bump, m_Neg(m_Value(NegBump)))) {
657       // If Bump is a neg instruction, emit C = Basis - (-Bump).
658       Reduced = Builder.CreateSub(Basis.Ins, NegBump);
659       // We only use the negative argument of Bump, and Bump itself may be
660       // trivially dead.
661       RecursivelyDeleteTriviallyDeadInstructions(Bump);
662     } else {
663       // It's tempting to preserve nsw on Bump and/or Reduced. However, it's
664       // usually unsound, e.g.,
665       //
666       // X = (-2 +nsw 1) *nsw INT_MAX
667       // Y = (-2 +nsw 3) *nsw INT_MAX
668       //   =>
669       // Y = X + 2 * INT_MAX
670       //
671       // Neither + and * in the resultant expression are nsw.
672       Reduced = Builder.CreateAdd(Basis.Ins, Bump);
673     }
674     break;
675   }
676   case Candidate::GEP:
677     {
678     Type *OffsetTy = DL->getIndexType(C.Ins->getType());
679     bool InBounds = cast<GetElementPtrInst>(C.Ins)->isInBounds();
680     if (BumpWithUglyGEP) {
681       // C = (char *)Basis + Bump
682       unsigned AS = Basis.Ins->getType()->getPointerAddressSpace();
683       Type *CharTy = Type::getInt8PtrTy(Basis.Ins->getContext(), AS);
684       Reduced = Builder.CreateBitCast(Basis.Ins, CharTy);
685       Reduced =
686           Builder.CreateGEP(Builder.getInt8Ty(), Reduced, Bump, "", InBounds);
687       Reduced = Builder.CreateBitCast(Reduced, C.Ins->getType());
688     } else {
689       // C = gep Basis, Bump
690       // Canonicalize bump to pointer size.
691       Bump = Builder.CreateSExtOrTrunc(Bump, OffsetTy);
692       Reduced = Builder.CreateGEP(
693           cast<GetElementPtrInst>(Basis.Ins)->getResultElementType(), Basis.Ins,
694           Bump, "", InBounds);
695     }
696       break;
697     }
698   default:
699     llvm_unreachable("C.CandidateKind is invalid");
700   };
701   Reduced->takeName(C.Ins);
702   C.Ins->replaceAllUsesWith(Reduced);
703   // Unlink C.Ins so that we can skip other candidates also corresponding to
704   // C.Ins. The actual deletion is postponed to the end of runOnFunction.
705   C.Ins->removeFromParent();
706   UnlinkedInstructions.push_back(C.Ins);
707 }
708 
709 bool StraightLineStrengthReduceLegacyPass::runOnFunction(Function &F) {
710   if (skipFunction(F))
711     return false;
712 
713   auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
714   auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
715   auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
716   return StraightLineStrengthReduce(DL, DT, SE, TTI).runOnFunction(F);
717 }
718 
719 bool StraightLineStrengthReduce::runOnFunction(Function &F) {
720   // Traverse the dominator tree in the depth-first order. This order makes sure
721   // all bases of a candidate are in Candidates when we process it.
722   for (const auto Node : depth_first(DT))
723     for (auto &I : *(Node->getBlock()))
724       allocateCandidatesAndFindBasis(&I);
725 
726   // Rewrite candidates in the reverse depth-first order. This order makes sure
727   // a candidate being rewritten is not a basis for any other candidate.
728   while (!Candidates.empty()) {
729     const Candidate &C = Candidates.back();
730     if (C.Basis != nullptr) {
731       rewriteCandidateWithBasis(C, *C.Basis);
732     }
733     Candidates.pop_back();
734   }
735 
736   // Delete all unlink instructions.
737   for (auto *UnlinkedInst : UnlinkedInstructions) {
738     for (unsigned I = 0, E = UnlinkedInst->getNumOperands(); I != E; ++I) {
739       Value *Op = UnlinkedInst->getOperand(I);
740       UnlinkedInst->setOperand(I, nullptr);
741       RecursivelyDeleteTriviallyDeadInstructions(Op);
742     }
743     UnlinkedInst->deleteValue();
744   }
745   bool Ret = !UnlinkedInstructions.empty();
746   UnlinkedInstructions.clear();
747   return Ret;
748 }
749 
750 namespace llvm {
751 
752 PreservedAnalyses
753 StraightLineStrengthReducePass::run(Function &F, FunctionAnalysisManager &AM) {
754   const DataLayout *DL = &F.getParent()->getDataLayout();
755   auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
756   auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
757   auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
758 
759   if (!StraightLineStrengthReduce(DL, DT, SE, TTI).runOnFunction(F))
760     return PreservedAnalyses::all();
761 
762   PreservedAnalyses PA;
763   PA.preserveSet<CFGAnalyses>();
764   PA.preserve<DominatorTreeAnalysis>();
765   PA.preserve<ScalarEvolutionAnalysis>();
766   PA.preserve<TargetIRAnalysis>();
767   return PA;
768 }
769 
770 } // namespace llvm
771