1 //===-- StraightLineStrengthReduce.cpp - ------------------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 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. 15 // 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: 19 // 20 // Form 1: B + i * S 21 // Form 2: (B + i) * S 22 // Form 3: &B[i * S] 23 // 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, 27 // 28 // S1: X = B + i * S 29 // S2: Y = B + i' * S => X + (i' - i) * S 30 // 31 // S1: X = (B + i) * S 32 // S2: Y = (B + i') * S => X + (i' - i) * S 33 // 34 // S1: X = &B[i * S] 35 // S2: Y = &B[i' * S] => &X[(i' - i) * S] 36 // 37 // Note: (i' - i) * S is folded to the extent possible. 38 // 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. 43 // 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. 47 // 48 // TODO: 49 // 50 // - Floating point arithmetics when fast math is enabled. 51 // 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. 55 // 56 // - When (i' - i) is constant but i and i' are not, we could still perform 57 // SLSR. 58 #include <vector> 59 60 #include "llvm/ADT/DenseSet.h" 61 #include "llvm/ADT/FoldingSet.h" 62 #include "llvm/Analysis/ScalarEvolution.h" 63 #include "llvm/Analysis/TargetTransformInfo.h" 64 #include "llvm/Analysis/ValueTracking.h" 65 #include "llvm/IR/DataLayout.h" 66 #include "llvm/IR/Dominators.h" 67 #include "llvm/IR/IRBuilder.h" 68 #include "llvm/IR/Module.h" 69 #include "llvm/IR/PatternMatch.h" 70 #include "llvm/Support/raw_ostream.h" 71 #include "llvm/Transforms/Scalar.h" 72 #include "llvm/Transforms/Utils/Local.h" 73 74 using namespace llvm; 75 using namespace PatternMatch; 76 77 namespace { 78 79 class StraightLineStrengthReduce : public FunctionPass { 80 public: 81 // SLSR candidate. Such a candidate must be in one of the forms described in 82 // the header comments. 83 struct Candidate : public ilist_node<Candidate> { 84 enum Kind { 85 Invalid, // reserved for the default constructor 86 Add, // B + i * S 87 Mul, // (B + i) * S 88 GEP, // &B[..][i * S][..] 89 }; 90 91 Candidate() 92 : CandidateKind(Invalid), Base(nullptr), Index(nullptr), 93 Stride(nullptr), Ins(nullptr), Basis(nullptr) {} 94 Candidate(Kind CT, const SCEV *B, ConstantInt *Idx, Value *S, 95 Instruction *I) 96 : CandidateKind(CT), Base(B), Index(Idx), Stride(S), Ins(I), 97 Basis(nullptr) {} 98 Kind CandidateKind; 99 const SCEV *Base; 100 // Note that Index and Stride of a GEP candidate do not necessarily have the 101 // same integer type. In that case, during rewriting, Stride will be 102 // sign-extended or truncated to Index's type. 103 ConstantInt *Index; 104 Value *Stride; 105 // The instruction this candidate corresponds to. It helps us to rewrite a 106 // candidate with respect to its immediate basis. Note that one instruction 107 // can correspond to multiple candidates depending on how you associate the 108 // expression. For instance, 109 // 110 // (a + 1) * (b + 2) 111 // 112 // can be treated as 113 // 114 // <Base: a, Index: 1, Stride: b + 2> 115 // 116 // or 117 // 118 // <Base: b, Index: 2, Stride: a + 1> 119 Instruction *Ins; 120 // Points to the immediate basis of this candidate, or nullptr if we cannot 121 // find any basis for this candidate. 122 Candidate *Basis; 123 }; 124 125 static char ID; 126 127 StraightLineStrengthReduce() 128 : FunctionPass(ID), DL(nullptr), DT(nullptr), TTI(nullptr) { 129 initializeStraightLineStrengthReducePass(*PassRegistry::getPassRegistry()); 130 } 131 132 void getAnalysisUsage(AnalysisUsage &AU) const override { 133 AU.addRequired<DominatorTreeWrapperPass>(); 134 AU.addRequired<ScalarEvolutionWrapperPass>(); 135 AU.addRequired<TargetTransformInfoWrapperPass>(); 136 // We do not modify the shape of the CFG. 137 AU.setPreservesCFG(); 138 } 139 140 bool doInitialization(Module &M) override { 141 DL = &M.getDataLayout(); 142 return false; 143 } 144 145 bool runOnFunction(Function &F) override; 146 147 private: 148 // Returns true if Basis is a basis for C, i.e., Basis dominates C and they 149 // share the same base and stride. 150 bool isBasisFor(const Candidate &Basis, const Candidate &C); 151 // Returns whether the candidate can be folded into an addressing mode. 152 bool isFoldable(const Candidate &C, TargetTransformInfo *TTI, 153 const DataLayout *DL); 154 // Returns true if C is already in a simplest form and not worth being 155 // rewritten. 156 bool isSimplestForm(const Candidate &C); 157 // Checks whether I is in a candidate form. If so, adds all the matching forms 158 // to Candidates, and tries to find the immediate basis for each of them. 159 void allocateCandidatesAndFindBasis(Instruction *I); 160 // Allocate candidates and find bases for Add instructions. 161 void allocateCandidatesAndFindBasisForAdd(Instruction *I); 162 // Given I = LHS + RHS, factors RHS into i * S and makes (LHS + i * S) a 163 // candidate. 164 void allocateCandidatesAndFindBasisForAdd(Value *LHS, Value *RHS, 165 Instruction *I); 166 // Allocate candidates and find bases for Mul instructions. 167 void allocateCandidatesAndFindBasisForMul(Instruction *I); 168 // Splits LHS into Base + Index and, if succeeds, calls 169 // allocateCandidatesAndFindBasis. 170 void allocateCandidatesAndFindBasisForMul(Value *LHS, Value *RHS, 171 Instruction *I); 172 // Allocate candidates and find bases for GetElementPtr instructions. 173 void allocateCandidatesAndFindBasisForGEP(GetElementPtrInst *GEP); 174 // A helper function that scales Idx with ElementSize before invoking 175 // allocateCandidatesAndFindBasis. 176 void allocateCandidatesAndFindBasisForGEP(const SCEV *B, ConstantInt *Idx, 177 Value *S, uint64_t ElementSize, 178 Instruction *I); 179 // Adds the given form <CT, B, Idx, S> to Candidates, and finds its immediate 180 // basis. 181 void allocateCandidatesAndFindBasis(Candidate::Kind CT, const SCEV *B, 182 ConstantInt *Idx, Value *S, 183 Instruction *I); 184 // Rewrites candidate C with respect to Basis. 185 void rewriteCandidateWithBasis(const Candidate &C, const Candidate &Basis); 186 // A helper function that factors ArrayIdx to a product of a stride and a 187 // constant index, and invokes allocateCandidatesAndFindBasis with the 188 // factorings. 189 void factorArrayIndex(Value *ArrayIdx, const SCEV *Base, uint64_t ElementSize, 190 GetElementPtrInst *GEP); 191 // Emit code that computes the "bump" from Basis to C. If the candidate is a 192 // GEP and the bump is not divisible by the element size of the GEP, this 193 // function sets the BumpWithUglyGEP flag to notify its caller to bump the 194 // basis using an ugly GEP. 195 static Value *emitBump(const Candidate &Basis, const Candidate &C, 196 IRBuilder<> &Builder, const DataLayout *DL, 197 bool &BumpWithUglyGEP); 198 199 const DataLayout *DL; 200 DominatorTree *DT; 201 ScalarEvolution *SE; 202 TargetTransformInfo *TTI; 203 ilist<Candidate> Candidates; 204 // Temporarily holds all instructions that are unlinked (but not deleted) by 205 // rewriteCandidateWithBasis. These instructions will be actually removed 206 // after all rewriting finishes. 207 std::vector<Instruction *> UnlinkedInstructions; 208 }; 209 } // anonymous namespace 210 211 char StraightLineStrengthReduce::ID = 0; 212 INITIALIZE_PASS_BEGIN(StraightLineStrengthReduce, "slsr", 213 "Straight line strength reduction", false, false) 214 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 215 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 216 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 217 INITIALIZE_PASS_END(StraightLineStrengthReduce, "slsr", 218 "Straight line strength reduction", false, false) 219 220 FunctionPass *llvm::createStraightLineStrengthReducePass() { 221 return new StraightLineStrengthReduce(); 222 } 223 224 bool StraightLineStrengthReduce::isBasisFor(const Candidate &Basis, 225 const Candidate &C) { 226 return (Basis.Ins != C.Ins && // skip the same instruction 227 // They must have the same type too. Basis.Base == C.Base doesn't 228 // guarantee their types are the same (PR23975). 229 Basis.Ins->getType() == C.Ins->getType() && 230 // Basis must dominate C in order to rewrite C with respect to Basis. 231 DT->dominates(Basis.Ins->getParent(), C.Ins->getParent()) && 232 // They share the same base, stride, and candidate kind. 233 Basis.Base == C.Base && Basis.Stride == C.Stride && 234 Basis.CandidateKind == C.CandidateKind); 235 } 236 237 // TODO: use TTI->getGEPCost. 238 static bool isGEPFoldable(GetElementPtrInst *GEP, 239 const TargetTransformInfo *TTI, 240 const DataLayout *DL) { 241 GlobalVariable *BaseGV = nullptr; 242 int64_t BaseOffset = 0; 243 bool HasBaseReg = false; 244 int64_t Scale = 0; 245 246 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand())) 247 BaseGV = GV; 248 else 249 HasBaseReg = true; 250 251 gep_type_iterator GTI = gep_type_begin(GEP); 252 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I, ++GTI) { 253 if (isa<SequentialType>(*GTI)) { 254 int64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType()); 255 if (ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) { 256 BaseOffset += ConstIdx->getSExtValue() * ElementSize; 257 } else { 258 // Needs scale register. 259 if (Scale != 0) { 260 // No addressing mode takes two scale registers. 261 return false; 262 } 263 Scale = ElementSize; 264 } 265 } else { 266 StructType *STy = cast<StructType>(*GTI); 267 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue(); 268 BaseOffset += DL->getStructLayout(STy)->getElementOffset(Field); 269 } 270 } 271 272 unsigned AddrSpace = GEP->getPointerAddressSpace(); 273 return TTI->isLegalAddressingMode(GEP->getType()->getElementType(), BaseGV, 274 BaseOffset, HasBaseReg, Scale, AddrSpace); 275 } 276 277 // Returns whether (Base + Index * Stride) can be folded to an addressing mode. 278 static bool isAddFoldable(const SCEV *Base, ConstantInt *Index, Value *Stride, 279 TargetTransformInfo *TTI) { 280 return TTI->isLegalAddressingMode(Base->getType(), nullptr, 0, true, 281 Index->getSExtValue()); 282 } 283 284 bool StraightLineStrengthReduce::isFoldable(const Candidate &C, 285 TargetTransformInfo *TTI, 286 const DataLayout *DL) { 287 if (C.CandidateKind == Candidate::Add) 288 return isAddFoldable(C.Base, C.Index, C.Stride, TTI); 289 if (C.CandidateKind == Candidate::GEP) 290 return isGEPFoldable(cast<GetElementPtrInst>(C.Ins), TTI, DL); 291 return false; 292 } 293 294 // Returns true if GEP has zero or one non-zero index. 295 static bool hasOnlyOneNonZeroIndex(GetElementPtrInst *GEP) { 296 unsigned NumNonZeroIndices = 0; 297 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I) { 298 ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I); 299 if (ConstIdx == nullptr || !ConstIdx->isZero()) 300 ++NumNonZeroIndices; 301 } 302 return NumNonZeroIndices <= 1; 303 } 304 305 bool StraightLineStrengthReduce::isSimplestForm(const Candidate &C) { 306 if (C.CandidateKind == Candidate::Add) { 307 // B + 1 * S or B + (-1) * S 308 return C.Index->isOne() || C.Index->isMinusOne(); 309 } 310 if (C.CandidateKind == Candidate::Mul) { 311 // (B + 0) * S 312 return C.Index->isZero(); 313 } 314 if (C.CandidateKind == Candidate::GEP) { 315 // (char*)B + S or (char*)B - S 316 return ((C.Index->isOne() || C.Index->isMinusOne()) && 317 hasOnlyOneNonZeroIndex(cast<GetElementPtrInst>(C.Ins))); 318 } 319 return false; 320 } 321 322 // TODO: We currently implement an algorithm whose time complexity is linear in 323 // the number of existing candidates. However, we could do better by using 324 // ScopedHashTable. Specifically, while traversing the dominator tree, we could 325 // maintain all the candidates that dominate the basic block being traversed in 326 // a ScopedHashTable. This hash table is indexed by the base and the stride of 327 // a candidate. Therefore, finding the immediate basis of a candidate boils down 328 // to one hash-table look up. 329 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis( 330 Candidate::Kind CT, const SCEV *B, ConstantInt *Idx, Value *S, 331 Instruction *I) { 332 Candidate C(CT, B, Idx, S, I); 333 // SLSR can complicate an instruction in two cases: 334 // 335 // 1. If we can fold I into an addressing mode, computing I is likely free or 336 // takes only one instruction. 337 // 338 // 2. I is already in a simplest form. For example, when 339 // X = B + 8 * S 340 // Y = B + S, 341 // rewriting Y to X - 7 * S is probably a bad idea. 342 // 343 // In the above cases, we still add I to the candidate list so that I can be 344 // the basis of other candidates, but we leave I's basis blank so that I 345 // won't be rewritten. 346 if (!isFoldable(C, TTI, DL) && !isSimplestForm(C)) { 347 // Try to compute the immediate basis of C. 348 unsigned NumIterations = 0; 349 // Limit the scan radius to avoid running in quadratice time. 350 static const unsigned MaxNumIterations = 50; 351 for (auto Basis = Candidates.rbegin(); 352 Basis != Candidates.rend() && NumIterations < MaxNumIterations; 353 ++Basis, ++NumIterations) { 354 if (isBasisFor(*Basis, C)) { 355 C.Basis = &(*Basis); 356 break; 357 } 358 } 359 } 360 // Regardless of whether we find a basis for C, we need to push C to the 361 // candidate list so that it can be the basis of other candidates. 362 Candidates.push_back(C); 363 } 364 365 void StraightLineStrengthReduce::allocateCandidatesAndFindBasis( 366 Instruction *I) { 367 switch (I->getOpcode()) { 368 case Instruction::Add: 369 allocateCandidatesAndFindBasisForAdd(I); 370 break; 371 case Instruction::Mul: 372 allocateCandidatesAndFindBasisForMul(I); 373 break; 374 case Instruction::GetElementPtr: 375 allocateCandidatesAndFindBasisForGEP(cast<GetElementPtrInst>(I)); 376 break; 377 } 378 } 379 380 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd( 381 Instruction *I) { 382 // Try matching B + i * S. 383 if (!isa<IntegerType>(I->getType())) 384 return; 385 386 assert(I->getNumOperands() == 2 && "isn't I an add?"); 387 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 388 allocateCandidatesAndFindBasisForAdd(LHS, RHS, I); 389 if (LHS != RHS) 390 allocateCandidatesAndFindBasisForAdd(RHS, LHS, I); 391 } 392 393 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForAdd( 394 Value *LHS, Value *RHS, Instruction *I) { 395 Value *S = nullptr; 396 ConstantInt *Idx = nullptr; 397 if (match(RHS, m_Mul(m_Value(S), m_ConstantInt(Idx)))) { 398 // I = LHS + RHS = LHS + Idx * S 399 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I); 400 } else if (match(RHS, m_Shl(m_Value(S), m_ConstantInt(Idx)))) { 401 // I = LHS + RHS = LHS + (S << Idx) = LHS + S * (1 << Idx) 402 APInt One(Idx->getBitWidth(), 1); 403 Idx = ConstantInt::get(Idx->getContext(), One << Idx->getValue()); 404 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), Idx, S, I); 405 } else { 406 // At least, I = LHS + 1 * RHS 407 ConstantInt *One = ConstantInt::get(cast<IntegerType>(I->getType()), 1); 408 allocateCandidatesAndFindBasis(Candidate::Add, SE->getSCEV(LHS), One, RHS, 409 I); 410 } 411 } 412 413 // Returns true if A matches B + C where C is constant. 414 static bool matchesAdd(Value *A, Value *&B, ConstantInt *&C) { 415 return (match(A, m_Add(m_Value(B), m_ConstantInt(C))) || 416 match(A, m_Add(m_ConstantInt(C), m_Value(B)))); 417 } 418 419 // Returns true if A matches B | C where C is constant. 420 static bool matchesOr(Value *A, Value *&B, ConstantInt *&C) { 421 return (match(A, m_Or(m_Value(B), m_ConstantInt(C))) || 422 match(A, m_Or(m_ConstantInt(C), m_Value(B)))); 423 } 424 425 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul( 426 Value *LHS, Value *RHS, Instruction *I) { 427 Value *B = nullptr; 428 ConstantInt *Idx = nullptr; 429 if (matchesAdd(LHS, B, Idx)) { 430 // If LHS is in the form of "Base + Index", then I is in the form of 431 // "(Base + Index) * RHS". 432 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I); 433 } else if (matchesOr(LHS, B, Idx) && haveNoCommonBitsSet(B, Idx, *DL)) { 434 // If LHS is in the form of "Base | Index" and Base and Index have no common 435 // bits set, then 436 // Base | Index = Base + Index 437 // and I is thus in the form of "(Base + Index) * RHS". 438 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(B), Idx, RHS, I); 439 } else { 440 // Otherwise, at least try the form (LHS + 0) * RHS. 441 ConstantInt *Zero = ConstantInt::get(cast<IntegerType>(I->getType()), 0); 442 allocateCandidatesAndFindBasis(Candidate::Mul, SE->getSCEV(LHS), Zero, RHS, 443 I); 444 } 445 } 446 447 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForMul( 448 Instruction *I) { 449 // Try matching (B + i) * S. 450 // TODO: we could extend SLSR to float and vector types. 451 if (!isa<IntegerType>(I->getType())) 452 return; 453 454 assert(I->getNumOperands() == 2 && "isn't I a mul?"); 455 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 456 allocateCandidatesAndFindBasisForMul(LHS, RHS, I); 457 if (LHS != RHS) { 458 // Symmetrically, try to split RHS to Base + Index. 459 allocateCandidatesAndFindBasisForMul(RHS, LHS, I); 460 } 461 } 462 463 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP( 464 const SCEV *B, ConstantInt *Idx, Value *S, uint64_t ElementSize, 465 Instruction *I) { 466 // I = B + sext(Idx *nsw S) * ElementSize 467 // = B + (sext(Idx) * sext(S)) * ElementSize 468 // = B + (sext(Idx) * ElementSize) * sext(S) 469 // Casting to IntegerType is safe because we skipped vector GEPs. 470 IntegerType *IntPtrTy = cast<IntegerType>(DL->getIntPtrType(I->getType())); 471 ConstantInt *ScaledIdx = ConstantInt::get( 472 IntPtrTy, Idx->getSExtValue() * (int64_t)ElementSize, true); 473 allocateCandidatesAndFindBasis(Candidate::GEP, B, ScaledIdx, S, I); 474 } 475 476 void StraightLineStrengthReduce::factorArrayIndex(Value *ArrayIdx, 477 const SCEV *Base, 478 uint64_t ElementSize, 479 GetElementPtrInst *GEP) { 480 // At least, ArrayIdx = ArrayIdx *nsw 1. 481 allocateCandidatesAndFindBasisForGEP( 482 Base, ConstantInt::get(cast<IntegerType>(ArrayIdx->getType()), 1), 483 ArrayIdx, ElementSize, GEP); 484 Value *LHS = nullptr; 485 ConstantInt *RHS = nullptr; 486 // One alternative is matching the SCEV of ArrayIdx instead of ArrayIdx 487 // itself. This would allow us to handle the shl case for free. However, 488 // matching SCEVs has two issues: 489 // 490 // 1. this would complicate rewriting because the rewriting procedure 491 // would have to translate SCEVs back to IR instructions. This translation 492 // is difficult when LHS is further evaluated to a composite SCEV. 493 // 494 // 2. ScalarEvolution is designed to be control-flow oblivious. It tends 495 // to strip nsw/nuw flags which are critical for SLSR to trace into 496 // sext'ed multiplication. 497 if (match(ArrayIdx, m_NSWMul(m_Value(LHS), m_ConstantInt(RHS)))) { 498 // SLSR is currently unsafe if i * S may overflow. 499 // GEP = Base + sext(LHS *nsw RHS) * ElementSize 500 allocateCandidatesAndFindBasisForGEP(Base, RHS, LHS, ElementSize, GEP); 501 } else if (match(ArrayIdx, m_NSWShl(m_Value(LHS), m_ConstantInt(RHS)))) { 502 // GEP = Base + sext(LHS <<nsw RHS) * ElementSize 503 // = Base + sext(LHS *nsw (1 << RHS)) * ElementSize 504 APInt One(RHS->getBitWidth(), 1); 505 ConstantInt *PowerOf2 = 506 ConstantInt::get(RHS->getContext(), One << RHS->getValue()); 507 allocateCandidatesAndFindBasisForGEP(Base, PowerOf2, LHS, ElementSize, GEP); 508 } 509 } 510 511 void StraightLineStrengthReduce::allocateCandidatesAndFindBasisForGEP( 512 GetElementPtrInst *GEP) { 513 // TODO: handle vector GEPs 514 if (GEP->getType()->isVectorTy()) 515 return; 516 517 SmallVector<const SCEV *, 4> IndexExprs; 518 for (auto I = GEP->idx_begin(); I != GEP->idx_end(); ++I) 519 IndexExprs.push_back(SE->getSCEV(*I)); 520 521 gep_type_iterator GTI = gep_type_begin(GEP); 522 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I) { 523 if (!isa<SequentialType>(*GTI++)) 524 continue; 525 526 const SCEV *OrigIndexExpr = IndexExprs[I - 1]; 527 IndexExprs[I - 1] = SE->getZero(OrigIndexExpr->getType()); 528 529 // The base of this candidate is GEP's base plus the offsets of all 530 // indices except this current one. 531 const SCEV *BaseExpr = SE->getGEPExpr(GEP->getSourceElementType(), 532 SE->getSCEV(GEP->getPointerOperand()), 533 IndexExprs, GEP->isInBounds()); 534 Value *ArrayIdx = GEP->getOperand(I); 535 uint64_t ElementSize = DL->getTypeAllocSize(*GTI); 536 factorArrayIndex(ArrayIdx, BaseExpr, ElementSize, GEP); 537 // When ArrayIdx is the sext of a value, we try to factor that value as 538 // well. Handling this case is important because array indices are 539 // typically sign-extended to the pointer size. 540 Value *TruncatedArrayIdx = nullptr; 541 if (match(ArrayIdx, m_SExt(m_Value(TruncatedArrayIdx)))) 542 factorArrayIndex(TruncatedArrayIdx, BaseExpr, ElementSize, GEP); 543 544 IndexExprs[I - 1] = OrigIndexExpr; 545 } 546 } 547 548 // A helper function that unifies the bitwidth of A and B. 549 static void unifyBitWidth(APInt &A, APInt &B) { 550 if (A.getBitWidth() < B.getBitWidth()) 551 A = A.sext(B.getBitWidth()); 552 else if (A.getBitWidth() > B.getBitWidth()) 553 B = B.sext(A.getBitWidth()); 554 } 555 556 Value *StraightLineStrengthReduce::emitBump(const Candidate &Basis, 557 const Candidate &C, 558 IRBuilder<> &Builder, 559 const DataLayout *DL, 560 bool &BumpWithUglyGEP) { 561 APInt Idx = C.Index->getValue(), BasisIdx = Basis.Index->getValue(); 562 unifyBitWidth(Idx, BasisIdx); 563 APInt IndexOffset = Idx - BasisIdx; 564 565 BumpWithUglyGEP = false; 566 if (Basis.CandidateKind == Candidate::GEP) { 567 APInt ElementSize( 568 IndexOffset.getBitWidth(), 569 DL->getTypeAllocSize( 570 cast<GetElementPtrInst>(Basis.Ins)->getType()->getElementType())); 571 APInt Q, R; 572 APInt::sdivrem(IndexOffset, ElementSize, Q, R); 573 if (R.getSExtValue() == 0) 574 IndexOffset = Q; 575 else 576 BumpWithUglyGEP = true; 577 } 578 579 // Compute Bump = C - Basis = (i' - i) * S. 580 // Common case 1: if (i' - i) is 1, Bump = S. 581 if (IndexOffset.getSExtValue() == 1) 582 return C.Stride; 583 // Common case 2: if (i' - i) is -1, Bump = -S. 584 if (IndexOffset.getSExtValue() == -1) 585 return Builder.CreateNeg(C.Stride); 586 587 // Otherwise, Bump = (i' - i) * sext/trunc(S). Note that (i' - i) and S may 588 // have different bit widths. 589 IntegerType *DeltaType = 590 IntegerType::get(Basis.Ins->getContext(), IndexOffset.getBitWidth()); 591 Value *ExtendedStride = Builder.CreateSExtOrTrunc(C.Stride, DeltaType); 592 if (IndexOffset.isPowerOf2()) { 593 // If (i' - i) is a power of 2, Bump = sext/trunc(S) << log(i' - i). 594 ConstantInt *Exponent = ConstantInt::get(DeltaType, IndexOffset.logBase2()); 595 return Builder.CreateShl(ExtendedStride, Exponent); 596 } 597 if ((-IndexOffset).isPowerOf2()) { 598 // If (i - i') is a power of 2, Bump = -sext/trunc(S) << log(i' - i). 599 ConstantInt *Exponent = 600 ConstantInt::get(DeltaType, (-IndexOffset).logBase2()); 601 return Builder.CreateNeg(Builder.CreateShl(ExtendedStride, Exponent)); 602 } 603 Constant *Delta = ConstantInt::get(DeltaType, IndexOffset); 604 return Builder.CreateMul(ExtendedStride, Delta); 605 } 606 607 void StraightLineStrengthReduce::rewriteCandidateWithBasis( 608 const Candidate &C, const Candidate &Basis) { 609 assert(C.CandidateKind == Basis.CandidateKind && C.Base == Basis.Base && 610 C.Stride == Basis.Stride); 611 // We run rewriteCandidateWithBasis on all candidates in a post-order, so the 612 // basis of a candidate cannot be unlinked before the candidate. 613 assert(Basis.Ins->getParent() != nullptr && "the basis is unlinked"); 614 615 // An instruction can correspond to multiple candidates. Therefore, instead of 616 // simply deleting an instruction when we rewrite it, we mark its parent as 617 // nullptr (i.e. unlink it) so that we can skip the candidates whose 618 // instruction is already rewritten. 619 if (!C.Ins->getParent()) 620 return; 621 622 IRBuilder<> Builder(C.Ins); 623 bool BumpWithUglyGEP; 624 Value *Bump = emitBump(Basis, C, Builder, DL, BumpWithUglyGEP); 625 Value *Reduced = nullptr; // equivalent to but weaker than C.Ins 626 switch (C.CandidateKind) { 627 case Candidate::Add: 628 case Candidate::Mul: 629 // C = Basis + Bump 630 if (BinaryOperator::isNeg(Bump)) { 631 // If Bump is a neg instruction, emit C = Basis - (-Bump). 632 Reduced = 633 Builder.CreateSub(Basis.Ins, BinaryOperator::getNegArgument(Bump)); 634 // We only use the negative argument of Bump, and Bump itself may be 635 // trivially dead. 636 RecursivelyDeleteTriviallyDeadInstructions(Bump); 637 } else { 638 // It's tempting to preserve nsw on Bump and/or Reduced. However, it's 639 // usually unsound, e.g., 640 // 641 // X = (-2 +nsw 1) *nsw INT_MAX 642 // Y = (-2 +nsw 3) *nsw INT_MAX 643 // => 644 // Y = X + 2 * INT_MAX 645 // 646 // Neither + and * in the resultant expression are nsw. 647 Reduced = Builder.CreateAdd(Basis.Ins, Bump); 648 } 649 break; 650 case Candidate::GEP: 651 { 652 Type *IntPtrTy = DL->getIntPtrType(C.Ins->getType()); 653 bool InBounds = cast<GetElementPtrInst>(C.Ins)->isInBounds(); 654 if (BumpWithUglyGEP) { 655 // C = (char *)Basis + Bump 656 unsigned AS = Basis.Ins->getType()->getPointerAddressSpace(); 657 Type *CharTy = Type::getInt8PtrTy(Basis.Ins->getContext(), AS); 658 Reduced = Builder.CreateBitCast(Basis.Ins, CharTy); 659 if (InBounds) 660 Reduced = 661 Builder.CreateInBoundsGEP(Builder.getInt8Ty(), Reduced, Bump); 662 else 663 Reduced = Builder.CreateGEP(Builder.getInt8Ty(), Reduced, Bump); 664 Reduced = Builder.CreateBitCast(Reduced, C.Ins->getType()); 665 } else { 666 // C = gep Basis, Bump 667 // Canonicalize bump to pointer size. 668 Bump = Builder.CreateSExtOrTrunc(Bump, IntPtrTy); 669 if (InBounds) 670 Reduced = Builder.CreateInBoundsGEP(nullptr, Basis.Ins, Bump); 671 else 672 Reduced = Builder.CreateGEP(nullptr, Basis.Ins, Bump); 673 } 674 } 675 break; 676 default: 677 llvm_unreachable("C.CandidateKind is invalid"); 678 }; 679 Reduced->takeName(C.Ins); 680 C.Ins->replaceAllUsesWith(Reduced); 681 // Unlink C.Ins so that we can skip other candidates also corresponding to 682 // C.Ins. The actual deletion is postponed to the end of runOnFunction. 683 C.Ins->removeFromParent(); 684 UnlinkedInstructions.push_back(C.Ins); 685 } 686 687 bool StraightLineStrengthReduce::runOnFunction(Function &F) { 688 if (skipOptnoneFunction(F)) 689 return false; 690 691 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 692 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 693 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 694 // Traverse the dominator tree in the depth-first order. This order makes sure 695 // all bases of a candidate are in Candidates when we process it. 696 for (auto node = GraphTraits<DominatorTree *>::nodes_begin(DT); 697 node != GraphTraits<DominatorTree *>::nodes_end(DT); ++node) { 698 for (auto &I : *node->getBlock()) 699 allocateCandidatesAndFindBasis(&I); 700 } 701 702 // Rewrite candidates in the reverse depth-first order. This order makes sure 703 // a candidate being rewritten is not a basis for any other candidate. 704 while (!Candidates.empty()) { 705 const Candidate &C = Candidates.back(); 706 if (C.Basis != nullptr) { 707 rewriteCandidateWithBasis(C, *C.Basis); 708 } 709 Candidates.pop_back(); 710 } 711 712 // Delete all unlink instructions. 713 for (auto *UnlinkedInst : UnlinkedInstructions) { 714 for (unsigned I = 0, E = UnlinkedInst->getNumOperands(); I != E; ++I) { 715 Value *Op = UnlinkedInst->getOperand(I); 716 UnlinkedInst->setOperand(I, nullptr); 717 RecursivelyDeleteTriviallyDeadInstructions(Op); 718 } 719 delete UnlinkedInst; 720 } 721 bool Ret = !UnlinkedInstructions.empty(); 722 UnlinkedInstructions.clear(); 723 return Ret; 724 } 725