InstructionCombining.cpp revision 096aa79276b8527a3cbbb3691e40e729dea09523
1//===- InstructionCombining.cpp - Combine multiple instructions -----------===// 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// InstructionCombining - Combine instructions to form fewer, simple 11// instructions. This pass does not modify the CFG. This pass is where 12// algebraic simplification happens. 13// 14// This pass combines things like: 15// %Y = add i32 %X, 1 16// %Z = add i32 %Y, 1 17// into: 18// %Z = add i32 %X, 2 19// 20// This is a simple worklist driven algorithm. 21// 22// This pass guarantees that the following canonicalizations are performed on 23// the program: 24// 1. If a binary operator has a constant operand, it is moved to the RHS 25// 2. Bitwise operators with constant operands are always grouped so that 26// shifts are performed first, then or's, then and's, then xor's. 27// 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible 28// 4. All cmp instructions on boolean values are replaced with logical ops 29// 5. add X, X is represented as (X*2) => (X << 1) 30// 6. Multiplies with a power-of-two constant argument are transformed into 31// shifts. 32// ... etc. 33// 34//===----------------------------------------------------------------------===// 35 36#define DEBUG_TYPE "instcombine" 37#include "llvm/Transforms/Scalar.h" 38#include "InstCombine.h" 39#include "llvm/IntrinsicInst.h" 40#include "llvm/Analysis/ConstantFolding.h" 41#include "llvm/Analysis/InstructionSimplify.h" 42#include "llvm/Analysis/MemoryBuiltins.h" 43#include "llvm/Target/TargetData.h" 44#include "llvm/Transforms/Utils/Local.h" 45#include "llvm/Support/CFG.h" 46#include "llvm/Support/Debug.h" 47#include "llvm/Support/GetElementPtrTypeIterator.h" 48#include "llvm/Support/PatternMatch.h" 49#include "llvm/ADT/SmallPtrSet.h" 50#include "llvm/ADT/Statistic.h" 51#include "llvm-c/Initialization.h" 52#include <algorithm> 53#include <climits> 54using namespace llvm; 55using namespace llvm::PatternMatch; 56 57STATISTIC(NumCombined , "Number of insts combined"); 58STATISTIC(NumConstProp, "Number of constant folds"); 59STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 60STATISTIC(NumSunkInst , "Number of instructions sunk"); 61 62// Initialization Routines 63void llvm::initializeInstCombine(PassRegistry &Registry) { 64 initializeInstCombinerPass(Registry); 65} 66 67void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 68 initializeInstCombine(*unwrap(R)); 69} 70 71char InstCombiner::ID = 0; 72INITIALIZE_PASS(InstCombiner, "instcombine", 73 "Combine redundant instructions", false, false) 74 75void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 76 AU.addPreservedID(LCSSAID); 77 AU.setPreservesCFG(); 78} 79 80 81/// ShouldChangeType - Return true if it is desirable to convert a computation 82/// from 'From' to 'To'. We don't want to convert from a legal to an illegal 83/// type for example, or from a smaller to a larger illegal type. 84bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const { 85 assert(From->isIntegerTy() && To->isIntegerTy()); 86 87 // If we don't have TD, we don't know if the source/dest are legal. 88 if (!TD) return false; 89 90 unsigned FromWidth = From->getPrimitiveSizeInBits(); 91 unsigned ToWidth = To->getPrimitiveSizeInBits(); 92 bool FromLegal = TD->isLegalInteger(FromWidth); 93 bool ToLegal = TD->isLegalInteger(ToWidth); 94 95 // If this is a legal integer from type, and the result would be an illegal 96 // type, don't do the transformation. 97 if (FromLegal && !ToLegal) 98 return false; 99 100 // Otherwise, if both are illegal, do not increase the size of the result. We 101 // do allow things like i160 -> i64, but not i64 -> i160. 102 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 103 return false; 104 105 return true; 106} 107 108 109/// SimplifyAssociativeOrCommutative - This performs a few simplifications for 110/// operators which are associative or commutative: 111// 112// Commutative operators: 113// 114// 1. Order operands such that they are listed from right (least complex) to 115// left (most complex). This puts constants before unary operators before 116// binary operators. 117// 118// Associative operators: 119// 120// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 121// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 122// 123// Associative and commutative operators: 124// 125// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 126// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 127// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 128// if C1 and C2 are constants. 129// 130bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 131 Instruction::BinaryOps Opcode = I.getOpcode(); 132 bool Changed = false; 133 134 do { 135 // Order operands such that they are listed from right (least complex) to 136 // left (most complex). This puts constants before unary operators before 137 // binary operators. 138 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 139 getComplexity(I.getOperand(1))) 140 Changed = !I.swapOperands(); 141 142 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 143 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 144 145 if (I.isAssociative()) { 146 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 147 if (Op0 && Op0->getOpcode() == Opcode) { 148 Value *A = Op0->getOperand(0); 149 Value *B = Op0->getOperand(1); 150 Value *C = I.getOperand(1); 151 152 // Does "B op C" simplify? 153 if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) { 154 // It simplifies to V. Form "A op V". 155 I.setOperand(0, A); 156 I.setOperand(1, V); 157 Changed = true; 158 continue; 159 } 160 } 161 162 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 163 if (Op1 && Op1->getOpcode() == Opcode) { 164 Value *A = I.getOperand(0); 165 Value *B = Op1->getOperand(0); 166 Value *C = Op1->getOperand(1); 167 168 // Does "A op B" simplify? 169 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) { 170 // It simplifies to V. Form "V op C". 171 I.setOperand(0, V); 172 I.setOperand(1, C); 173 Changed = true; 174 continue; 175 } 176 } 177 } 178 179 if (I.isAssociative() && I.isCommutative()) { 180 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 181 if (Op0 && Op0->getOpcode() == Opcode) { 182 Value *A = Op0->getOperand(0); 183 Value *B = Op0->getOperand(1); 184 Value *C = I.getOperand(1); 185 186 // Does "C op A" simplify? 187 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) { 188 // It simplifies to V. Form "V op B". 189 I.setOperand(0, V); 190 I.setOperand(1, B); 191 Changed = true; 192 continue; 193 } 194 } 195 196 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 197 if (Op1 && Op1->getOpcode() == Opcode) { 198 Value *A = I.getOperand(0); 199 Value *B = Op1->getOperand(0); 200 Value *C = Op1->getOperand(1); 201 202 // Does "C op A" simplify? 203 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) { 204 // It simplifies to V. Form "B op V". 205 I.setOperand(0, B); 206 I.setOperand(1, V); 207 Changed = true; 208 continue; 209 } 210 } 211 212 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 213 // if C1 and C2 are constants. 214 if (Op0 && Op1 && 215 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 216 isa<Constant>(Op0->getOperand(1)) && 217 isa<Constant>(Op1->getOperand(1)) && 218 Op0->hasOneUse() && Op1->hasOneUse()) { 219 Value *A = Op0->getOperand(0); 220 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 221 Value *B = Op1->getOperand(0); 222 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 223 224 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 225 Instruction *New = BinaryOperator::Create(Opcode, A, B, Op1->getName(), 226 &I); 227 Worklist.Add(New); 228 I.setOperand(0, New); 229 I.setOperand(1, Folded); 230 Changed = true; 231 continue; 232 } 233 } 234 235 // No further simplifications. 236 return Changed; 237 } while (1); 238} 239 240// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 241// if the LHS is a constant zero (which is the 'negate' form). 242// 243Value *InstCombiner::dyn_castNegVal(Value *V) const { 244 if (BinaryOperator::isNeg(V)) 245 return BinaryOperator::getNegArgument(V); 246 247 // Constants can be considered to be negated values if they can be folded. 248 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 249 return ConstantExpr::getNeg(C); 250 251 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) 252 if (C->getType()->getElementType()->isIntegerTy()) 253 return ConstantExpr::getNeg(C); 254 255 return 0; 256} 257 258// dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 259// instruction if the LHS is a constant negative zero (which is the 'negate' 260// form). 261// 262Value *InstCombiner::dyn_castFNegVal(Value *V) const { 263 if (BinaryOperator::isFNeg(V)) 264 return BinaryOperator::getFNegArgument(V); 265 266 // Constants can be considered to be negated values if they can be folded. 267 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 268 return ConstantExpr::getFNeg(C); 269 270 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) 271 if (C->getType()->getElementType()->isFloatingPointTy()) 272 return ConstantExpr::getFNeg(C); 273 274 return 0; 275} 276 277static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 278 InstCombiner *IC) { 279 if (CastInst *CI = dyn_cast<CastInst>(&I)) 280 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 281 282 // Figure out if the constant is the left or the right argument. 283 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 284 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 285 286 if (Constant *SOC = dyn_cast<Constant>(SO)) { 287 if (ConstIsRHS) 288 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 289 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 290 } 291 292 Value *Op0 = SO, *Op1 = ConstOperand; 293 if (!ConstIsRHS) 294 std::swap(Op0, Op1); 295 296 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) 297 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 298 SO->getName()+".op"); 299 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 300 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 301 SO->getName()+".cmp"); 302 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 303 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 304 SO->getName()+".cmp"); 305 llvm_unreachable("Unknown binary instruction type!"); 306} 307 308// FoldOpIntoSelect - Given an instruction with a select as one operand and a 309// constant as the other operand, try to fold the binary operator into the 310// select arguments. This also works for Cast instructions, which obviously do 311// not have a second operand. 312Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 313 // Don't modify shared select instructions 314 if (!SI->hasOneUse()) return 0; 315 Value *TV = SI->getOperand(1); 316 Value *FV = SI->getOperand(2); 317 318 if (isa<Constant>(TV) || isa<Constant>(FV)) { 319 // Bool selects with constant operands can be folded to logical ops. 320 if (SI->getType()->isIntegerTy(1)) return 0; 321 322 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 323 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 324 325 return SelectInst::Create(SI->getCondition(), SelectTrueVal, 326 SelectFalseVal); 327 } 328 return 0; 329} 330 331 332/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 333/// has a PHI node as operand #0, see if we can fold the instruction into the 334/// PHI (which is only possible if all operands to the PHI are constants). 335/// 336/// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms 337/// that would normally be unprofitable because they strongly encourage jump 338/// threading. 339Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I, 340 bool AllowAggressive) { 341 AllowAggressive = false; 342 PHINode *PN = cast<PHINode>(I.getOperand(0)); 343 unsigned NumPHIValues = PN->getNumIncomingValues(); 344 if (NumPHIValues == 0 || 345 // We normally only transform phis with a single use, unless we're trying 346 // hard to make jump threading happen. 347 (!PN->hasOneUse() && !AllowAggressive)) 348 return 0; 349 350 351 // Check to see if all of the operands of the PHI are simple constants 352 // (constantint/constantfp/undef). If there is one non-constant value, 353 // remember the BB it is in. If there is more than one or if *it* is a PHI, 354 // bail out. We don't do arbitrary constant expressions here because moving 355 // their computation can be expensive without a cost model. 356 BasicBlock *NonConstBB = 0; 357 for (unsigned i = 0; i != NumPHIValues; ++i) 358 if (!isa<Constant>(PN->getIncomingValue(i)) || 359 isa<ConstantExpr>(PN->getIncomingValue(i))) { 360 if (NonConstBB) return 0; // More than one non-const value. 361 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi. 362 NonConstBB = PN->getIncomingBlock(i); 363 364 // If the incoming non-constant value is in I's block, we have an infinite 365 // loop. 366 if (NonConstBB == I.getParent()) 367 return 0; 368 } 369 370 // If there is exactly one non-constant value, we can insert a copy of the 371 // operation in that block. However, if this is a critical edge, we would be 372 // inserting the computation one some other paths (e.g. inside a loop). Only 373 // do this if the pred block is unconditionally branching into the phi block. 374 if (NonConstBB != 0 && !AllowAggressive) { 375 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 376 if (!BI || !BI->isUnconditional()) return 0; 377 } 378 379 // Okay, we can do the transformation: create the new PHI node. 380 PHINode *NewPN = PHINode::Create(I.getType(), ""); 381 NewPN->reserveOperandSpace(PN->getNumOperands()/2); 382 InsertNewInstBefore(NewPN, *PN); 383 NewPN->takeName(PN); 384 385 // Next, add all of the operands to the PHI. 386 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 387 // We only currently try to fold the condition of a select when it is a phi, 388 // not the true/false values. 389 Value *TrueV = SI->getTrueValue(); 390 Value *FalseV = SI->getFalseValue(); 391 BasicBlock *PhiTransBB = PN->getParent(); 392 for (unsigned i = 0; i != NumPHIValues; ++i) { 393 BasicBlock *ThisBB = PN->getIncomingBlock(i); 394 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 395 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 396 Value *InV = 0; 397 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) { 398 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 399 } else { 400 assert(PN->getIncomingBlock(i) == NonConstBB); 401 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred, 402 FalseVInPred, 403 "phitmp", NonConstBB->getTerminator()); 404 Worklist.Add(cast<Instruction>(InV)); 405 } 406 NewPN->addIncoming(InV, ThisBB); 407 } 408 } else if (I.getNumOperands() == 2) { 409 Constant *C = cast<Constant>(I.getOperand(1)); 410 for (unsigned i = 0; i != NumPHIValues; ++i) { 411 Value *InV = 0; 412 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) { 413 if (CmpInst *CI = dyn_cast<CmpInst>(&I)) 414 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 415 else 416 InV = ConstantExpr::get(I.getOpcode(), InC, C); 417 } else { 418 assert(PN->getIncomingBlock(i) == NonConstBB); 419 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) 420 InV = BinaryOperator::Create(BO->getOpcode(), 421 PN->getIncomingValue(i), C, "phitmp", 422 NonConstBB->getTerminator()); 423 else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) 424 InV = CmpInst::Create(CI->getOpcode(), 425 CI->getPredicate(), 426 PN->getIncomingValue(i), C, "phitmp", 427 NonConstBB->getTerminator()); 428 else 429 llvm_unreachable("Unknown binop!"); 430 431 Worklist.Add(cast<Instruction>(InV)); 432 } 433 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 434 } 435 } else { 436 CastInst *CI = cast<CastInst>(&I); 437 const Type *RetTy = CI->getType(); 438 for (unsigned i = 0; i != NumPHIValues; ++i) { 439 Value *InV; 440 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) { 441 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 442 } else { 443 assert(PN->getIncomingBlock(i) == NonConstBB); 444 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i), 445 I.getType(), "phitmp", 446 NonConstBB->getTerminator()); 447 Worklist.Add(cast<Instruction>(InV)); 448 } 449 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 450 } 451 } 452 return ReplaceInstUsesWith(I, NewPN); 453} 454 455/// FindElementAtOffset - Given a type and a constant offset, determine whether 456/// or not there is a sequence of GEP indices into the type that will land us at 457/// the specified offset. If so, fill them into NewIndices and return the 458/// resultant element type, otherwise return null. 459const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset, 460 SmallVectorImpl<Value*> &NewIndices) { 461 if (!TD) return 0; 462 if (!Ty->isSized()) return 0; 463 464 // Start with the index over the outer type. Note that the type size 465 // might be zero (even if the offset isn't zero) if the indexed type 466 // is something like [0 x {int, int}] 467 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 468 int64_t FirstIdx = 0; 469 if (int64_t TySize = TD->getTypeAllocSize(Ty)) { 470 FirstIdx = Offset/TySize; 471 Offset -= FirstIdx*TySize; 472 473 // Handle hosts where % returns negative instead of values [0..TySize). 474 if (Offset < 0) { 475 --FirstIdx; 476 Offset += TySize; 477 assert(Offset >= 0); 478 } 479 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 480 } 481 482 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 483 484 // Index into the types. If we fail, set OrigBase to null. 485 while (Offset) { 486 // Indexing into tail padding between struct/array elements. 487 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty)) 488 return 0; 489 490 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 491 const StructLayout *SL = TD->getStructLayout(STy); 492 assert(Offset < (int64_t)SL->getSizeInBytes() && 493 "Offset must stay within the indexed type"); 494 495 unsigned Elt = SL->getElementContainingOffset(Offset); 496 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 497 Elt)); 498 499 Offset -= SL->getElementOffset(Elt); 500 Ty = STy->getElementType(Elt); 501 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 502 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType()); 503 assert(EltSize && "Cannot index into a zero-sized array"); 504 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 505 Offset %= EltSize; 506 Ty = AT->getElementType(); 507 } else { 508 // Otherwise, we can't index into the middle of this atomic type, bail. 509 return 0; 510 } 511 } 512 513 return Ty; 514} 515 516 517 518Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 519 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 520 521 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD)) 522 return ReplaceInstUsesWith(GEP, V); 523 524 Value *PtrOp = GEP.getOperand(0); 525 526 if (isa<UndefValue>(GEP.getOperand(0))) 527 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType())); 528 529 // Eliminate unneeded casts for indices. 530 if (TD) { 531 bool MadeChange = false; 532 unsigned PtrSize = TD->getPointerSizeInBits(); 533 534 gep_type_iterator GTI = gep_type_begin(GEP); 535 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 536 I != E; ++I, ++GTI) { 537 if (!isa<SequentialType>(*GTI)) continue; 538 539 // If we are using a wider index than needed for this platform, shrink it 540 // to what we need. If narrower, sign-extend it to what we need. This 541 // explicit cast can make subsequent optimizations more obvious. 542 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth(); 543 if (OpBits == PtrSize) 544 continue; 545 546 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true); 547 MadeChange = true; 548 } 549 if (MadeChange) return &GEP; 550 } 551 552 // Combine Indices - If the source pointer to this getelementptr instruction 553 // is a getelementptr instruction, combine the indices of the two 554 // getelementptr instructions into a single instruction. 555 // 556 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 557 // Note that if our source is a gep chain itself that we wait for that 558 // chain to be resolved before we perform this transformation. This 559 // avoids us creating a TON of code in some cases. 560 // 561 if (GetElementPtrInst *SrcGEP = 562 dyn_cast<GetElementPtrInst>(Src->getOperand(0))) 563 if (SrcGEP->getNumOperands() == 2) 564 return 0; // Wait until our source is folded to completion. 565 566 SmallVector<Value*, 8> Indices; 567 568 // Find out whether the last index in the source GEP is a sequential idx. 569 bool EndsWithSequential = false; 570 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 571 I != E; ++I) 572 EndsWithSequential = !(*I)->isStructTy(); 573 574 // Can we combine the two pointer arithmetics offsets? 575 if (EndsWithSequential) { 576 // Replace: gep (gep %P, long B), long A, ... 577 // With: T = long A+B; gep %P, T, ... 578 // 579 Value *Sum; 580 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 581 Value *GO1 = GEP.getOperand(1); 582 if (SO1 == Constant::getNullValue(SO1->getType())) { 583 Sum = GO1; 584 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 585 Sum = SO1; 586 } else { 587 // If they aren't the same type, then the input hasn't been processed 588 // by the loop above yet (which canonicalizes sequential index types to 589 // intptr_t). Just avoid transforming this until the input has been 590 // normalized. 591 if (SO1->getType() != GO1->getType()) 592 return 0; 593 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 594 } 595 596 // Update the GEP in place if possible. 597 if (Src->getNumOperands() == 2) { 598 GEP.setOperand(0, Src->getOperand(0)); 599 GEP.setOperand(1, Sum); 600 return &GEP; 601 } 602 Indices.append(Src->op_begin()+1, Src->op_end()-1); 603 Indices.push_back(Sum); 604 Indices.append(GEP.op_begin()+2, GEP.op_end()); 605 } else if (isa<Constant>(*GEP.idx_begin()) && 606 cast<Constant>(*GEP.idx_begin())->isNullValue() && 607 Src->getNumOperands() != 1) { 608 // Otherwise we can do the fold if the first index of the GEP is a zero 609 Indices.append(Src->op_begin()+1, Src->op_end()); 610 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 611 } 612 613 if (!Indices.empty()) 614 return (GEP.isInBounds() && Src->isInBounds()) ? 615 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(), 616 Indices.end(), GEP.getName()) : 617 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(), 618 Indices.end(), GEP.getName()); 619 } 620 621 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 622 Value *StrippedPtr = PtrOp->stripPointerCasts(); 623 if (StrippedPtr != PtrOp) { 624 const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType()); 625 626 bool HasZeroPointerIndex = false; 627 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 628 HasZeroPointerIndex = C->isZero(); 629 630 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 631 // into : GEP [10 x i8]* X, i32 0, ... 632 // 633 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 634 // into : GEP i8* X, ... 635 // 636 // This occurs when the program declares an array extern like "int X[];" 637 if (HasZeroPointerIndex) { 638 const PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 639 if (const ArrayType *CATy = 640 dyn_cast<ArrayType>(CPTy->getElementType())) { 641 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 642 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 643 // -> GEP i8* X, ... 644 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 645 GetElementPtrInst *Res = 646 GetElementPtrInst::Create(StrippedPtr, Idx.begin(), 647 Idx.end(), GEP.getName()); 648 Res->setIsInBounds(GEP.isInBounds()); 649 return Res; 650 } 651 652 if (const ArrayType *XATy = 653 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 654 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 655 if (CATy->getElementType() == XATy->getElementType()) { 656 // -> GEP [10 x i8]* X, i32 0, ... 657 // At this point, we know that the cast source type is a pointer 658 // to an array of the same type as the destination pointer 659 // array. Because the array type is never stepped over (there 660 // is a leading zero) we can fold the cast into this GEP. 661 GEP.setOperand(0, StrippedPtr); 662 return &GEP; 663 } 664 } 665 } 666 } else if (GEP.getNumOperands() == 2) { 667 // Transform things like: 668 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 669 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 670 const Type *SrcElTy = StrippedPtrTy->getElementType(); 671 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType(); 672 if (TD && SrcElTy->isArrayTy() && 673 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) == 674 TD->getTypeAllocSize(ResElTy)) { 675 Value *Idx[2]; 676 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); 677 Idx[1] = GEP.getOperand(1); 678 Value *NewGEP = GEP.isInBounds() ? 679 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) : 680 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()); 681 // V and GEP are both pointer types --> BitCast 682 return new BitCastInst(NewGEP, GEP.getType()); 683 } 684 685 // Transform things like: 686 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 687 // (where tmp = 8*tmp2) into: 688 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 689 690 if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) { 691 uint64_t ArrayEltSize = 692 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()); 693 694 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We 695 // allow either a mul, shift, or constant here. 696 Value *NewIdx = 0; 697 ConstantInt *Scale = 0; 698 if (ArrayEltSize == 1) { 699 NewIdx = GEP.getOperand(1); 700 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1); 701 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) { 702 NewIdx = ConstantInt::get(CI->getType(), 1); 703 Scale = CI; 704 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){ 705 if (Inst->getOpcode() == Instruction::Shl && 706 isa<ConstantInt>(Inst->getOperand(1))) { 707 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1)); 708 uint32_t ShAmtVal = ShAmt->getLimitedValue(64); 709 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()), 710 1ULL << ShAmtVal); 711 NewIdx = Inst->getOperand(0); 712 } else if (Inst->getOpcode() == Instruction::Mul && 713 isa<ConstantInt>(Inst->getOperand(1))) { 714 Scale = cast<ConstantInt>(Inst->getOperand(1)); 715 NewIdx = Inst->getOperand(0); 716 } 717 } 718 719 // If the index will be to exactly the right offset with the scale taken 720 // out, perform the transformation. Note, we don't know whether Scale is 721 // signed or not. We'll use unsigned version of division/modulo 722 // operation after making sure Scale doesn't have the sign bit set. 723 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL && 724 Scale->getZExtValue() % ArrayEltSize == 0) { 725 Scale = ConstantInt::get(Scale->getType(), 726 Scale->getZExtValue() / ArrayEltSize); 727 if (Scale->getZExtValue() != 1) { 728 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(), 729 false /*ZExt*/); 730 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale"); 731 } 732 733 // Insert the new GEP instruction. 734 Value *Idx[2]; 735 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); 736 Idx[1] = NewIdx; 737 Value *NewGEP = GEP.isInBounds() ? 738 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()): 739 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()); 740 // The NewGEP must be pointer typed, so must the old one -> BitCast 741 return new BitCastInst(NewGEP, GEP.getType()); 742 } 743 } 744 } 745 } 746 747 /// See if we can simplify: 748 /// X = bitcast A* to B* 749 /// Y = gep X, <...constant indices...> 750 /// into a gep of the original struct. This is important for SROA and alias 751 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 752 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 753 if (TD && 754 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) { 755 // Determine how much the GEP moves the pointer. We are guaranteed to get 756 // a constant back from EmitGEPOffset. 757 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP)); 758 int64_t Offset = OffsetV->getSExtValue(); 759 760 // If this GEP instruction doesn't move the pointer, just replace the GEP 761 // with a bitcast of the real input to the dest type. 762 if (Offset == 0) { 763 // If the bitcast is of an allocation, and the allocation will be 764 // converted to match the type of the cast, don't touch this. 765 if (isa<AllocaInst>(BCI->getOperand(0)) || 766 isMalloc(BCI->getOperand(0))) { 767 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 768 if (Instruction *I = visitBitCast(*BCI)) { 769 if (I != BCI) { 770 I->takeName(BCI); 771 BCI->getParent()->getInstList().insert(BCI, I); 772 ReplaceInstUsesWith(*BCI, I); 773 } 774 return &GEP; 775 } 776 } 777 return new BitCastInst(BCI->getOperand(0), GEP.getType()); 778 } 779 780 // Otherwise, if the offset is non-zero, we need to find out if there is a 781 // field at Offset in 'A's type. If so, we can pull the cast through the 782 // GEP. 783 SmallVector<Value*, 8> NewIndices; 784 const Type *InTy = 785 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType(); 786 if (FindElementAtOffset(InTy, Offset, NewIndices)) { 787 Value *NGEP = GEP.isInBounds() ? 788 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(), 789 NewIndices.end()) : 790 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(), 791 NewIndices.end()); 792 793 if (NGEP->getType() == GEP.getType()) 794 return ReplaceInstUsesWith(GEP, NGEP); 795 NGEP->takeName(&GEP); 796 return new BitCastInst(NGEP, GEP.getType()); 797 } 798 } 799 } 800 801 return 0; 802} 803 804 805 806static bool IsOnlyNullComparedAndFreed(const Value &V) { 807 for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end(); 808 UI != UE; ++UI) { 809 const User *U = *UI; 810 if (isFreeCall(U)) 811 continue; 812 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(U)) 813 if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1))) 814 continue; 815 return false; 816 } 817 return true; 818} 819 820Instruction *InstCombiner::visitMalloc(Instruction &MI) { 821 // If we have a malloc call which is only used in any amount of comparisons 822 // to null and free calls, delete the calls and replace the comparisons with 823 // true or false as appropriate. 824 if (IsOnlyNullComparedAndFreed(MI)) { 825 for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end(); 826 UI != UE;) { 827 // We can assume that every remaining use is a free call or an icmp eq/ne 828 // to null, so the cast is safe. 829 Instruction *I = cast<Instruction>(*UI); 830 831 // Early increment here, as we're about to get rid of the user. 832 ++UI; 833 834 if (isFreeCall(I)) { 835 EraseInstFromFunction(*cast<CallInst>(I)); 836 continue; 837 } 838 // Again, the cast is safe. 839 ICmpInst *C = cast<ICmpInst>(I); 840 ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()), 841 C->isFalseWhenEqual())); 842 EraseInstFromFunction(*C); 843 } 844 return EraseInstFromFunction(MI); 845 } 846 return 0; 847} 848 849 850 851Instruction *InstCombiner::visitFree(CallInst &FI) { 852 Value *Op = FI.getArgOperand(0); 853 854 // free undef -> unreachable. 855 if (isa<UndefValue>(Op)) { 856 // Insert a new store to null because we cannot modify the CFG here. 857 new StoreInst(ConstantInt::getTrue(FI.getContext()), 858 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI); 859 return EraseInstFromFunction(FI); 860 } 861 862 // If we have 'free null' delete the instruction. This can happen in stl code 863 // when lots of inlining happens. 864 if (isa<ConstantPointerNull>(Op)) 865 return EraseInstFromFunction(FI); 866 867 return 0; 868} 869 870 871 872Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 873 // Change br (not X), label True, label False to: br X, label False, True 874 Value *X = 0; 875 BasicBlock *TrueDest; 876 BasicBlock *FalseDest; 877 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 878 !isa<Constant>(X)) { 879 // Swap Destinations and condition... 880 BI.setCondition(X); 881 BI.setSuccessor(0, FalseDest); 882 BI.setSuccessor(1, TrueDest); 883 return &BI; 884 } 885 886 // Cannonicalize fcmp_one -> fcmp_oeq 887 FCmpInst::Predicate FPred; Value *Y; 888 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 889 TrueDest, FalseDest)) && 890 BI.getCondition()->hasOneUse()) 891 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 892 FPred == FCmpInst::FCMP_OGE) { 893 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 894 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 895 896 // Swap Destinations and condition. 897 BI.setSuccessor(0, FalseDest); 898 BI.setSuccessor(1, TrueDest); 899 Worklist.Add(Cond); 900 return &BI; 901 } 902 903 // Cannonicalize icmp_ne -> icmp_eq 904 ICmpInst::Predicate IPred; 905 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 906 TrueDest, FalseDest)) && 907 BI.getCondition()->hasOneUse()) 908 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 909 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 910 IPred == ICmpInst::ICMP_SGE) { 911 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 912 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 913 // Swap Destinations and condition. 914 BI.setSuccessor(0, FalseDest); 915 BI.setSuccessor(1, TrueDest); 916 Worklist.Add(Cond); 917 return &BI; 918 } 919 920 return 0; 921} 922 923Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 924 Value *Cond = SI.getCondition(); 925 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 926 if (I->getOpcode() == Instruction::Add) 927 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 928 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 929 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) 930 SI.setOperand(i, 931 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)), 932 AddRHS)); 933 SI.setOperand(0, I->getOperand(0)); 934 Worklist.Add(I); 935 return &SI; 936 } 937 } 938 return 0; 939} 940 941Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 942 Value *Agg = EV.getAggregateOperand(); 943 944 if (!EV.hasIndices()) 945 return ReplaceInstUsesWith(EV, Agg); 946 947 if (Constant *C = dyn_cast<Constant>(Agg)) { 948 if (isa<UndefValue>(C)) 949 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType())); 950 951 if (isa<ConstantAggregateZero>(C)) 952 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType())); 953 954 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) { 955 // Extract the element indexed by the first index out of the constant 956 Value *V = C->getOperand(*EV.idx_begin()); 957 if (EV.getNumIndices() > 1) 958 // Extract the remaining indices out of the constant indexed by the 959 // first index 960 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end()); 961 else 962 return ReplaceInstUsesWith(EV, V); 963 } 964 return 0; // Can't handle other constants 965 } 966 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 967 // We're extracting from an insertvalue instruction, compare the indices 968 const unsigned *exti, *exte, *insi, *inse; 969 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 970 exte = EV.idx_end(), inse = IV->idx_end(); 971 exti != exte && insi != inse; 972 ++exti, ++insi) { 973 if (*insi != *exti) 974 // The insert and extract both reference distinctly different elements. 975 // This means the extract is not influenced by the insert, and we can 976 // replace the aggregate operand of the extract with the aggregate 977 // operand of the insert. i.e., replace 978 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 979 // %E = extractvalue { i32, { i32 } } %I, 0 980 // with 981 // %E = extractvalue { i32, { i32 } } %A, 0 982 return ExtractValueInst::Create(IV->getAggregateOperand(), 983 EV.idx_begin(), EV.idx_end()); 984 } 985 if (exti == exte && insi == inse) 986 // Both iterators are at the end: Index lists are identical. Replace 987 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 988 // %C = extractvalue { i32, { i32 } } %B, 1, 0 989 // with "i32 42" 990 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 991 if (exti == exte) { 992 // The extract list is a prefix of the insert list. i.e. replace 993 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 994 // %E = extractvalue { i32, { i32 } } %I, 1 995 // with 996 // %X = extractvalue { i32, { i32 } } %A, 1 997 // %E = insertvalue { i32 } %X, i32 42, 0 998 // by switching the order of the insert and extract (though the 999 // insertvalue should be left in, since it may have other uses). 1000 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 1001 EV.idx_begin(), EV.idx_end()); 1002 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 1003 insi, inse); 1004 } 1005 if (insi == inse) 1006 // The insert list is a prefix of the extract list 1007 // We can simply remove the common indices from the extract and make it 1008 // operate on the inserted value instead of the insertvalue result. 1009 // i.e., replace 1010 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1011 // %E = extractvalue { i32, { i32 } } %I, 1, 0 1012 // with 1013 // %E extractvalue { i32 } { i32 42 }, 0 1014 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 1015 exti, exte); 1016 } 1017 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 1018 // We're extracting from an intrinsic, see if we're the only user, which 1019 // allows us to simplify multiple result intrinsics to simpler things that 1020 // just get one value. 1021 if (II->hasOneUse()) { 1022 // Check if we're grabbing the overflow bit or the result of a 'with 1023 // overflow' intrinsic. If it's the latter we can remove the intrinsic 1024 // and replace it with a traditional binary instruction. 1025 switch (II->getIntrinsicID()) { 1026 case Intrinsic::uadd_with_overflow: 1027 case Intrinsic::sadd_with_overflow: 1028 if (*EV.idx_begin() == 0) { // Normal result. 1029 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1030 II->replaceAllUsesWith(UndefValue::get(II->getType())); 1031 EraseInstFromFunction(*II); 1032 return BinaryOperator::CreateAdd(LHS, RHS); 1033 } 1034 break; 1035 case Intrinsic::usub_with_overflow: 1036 case Intrinsic::ssub_with_overflow: 1037 if (*EV.idx_begin() == 0) { // Normal result. 1038 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1039 II->replaceAllUsesWith(UndefValue::get(II->getType())); 1040 EraseInstFromFunction(*II); 1041 return BinaryOperator::CreateSub(LHS, RHS); 1042 } 1043 break; 1044 case Intrinsic::umul_with_overflow: 1045 case Intrinsic::smul_with_overflow: 1046 if (*EV.idx_begin() == 0) { // Normal result. 1047 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1048 II->replaceAllUsesWith(UndefValue::get(II->getType())); 1049 EraseInstFromFunction(*II); 1050 return BinaryOperator::CreateMul(LHS, RHS); 1051 } 1052 break; 1053 default: 1054 break; 1055 } 1056 } 1057 } 1058 // Can't simplify extracts from other values. Note that nested extracts are 1059 // already simplified implicitely by the above (extract ( extract (insert) ) 1060 // will be translated into extract ( insert ( extract ) ) first and then just 1061 // the value inserted, if appropriate). 1062 return 0; 1063} 1064 1065 1066 1067 1068/// TryToSinkInstruction - Try to move the specified instruction from its 1069/// current block into the beginning of DestBlock, which can only happen if it's 1070/// safe to move the instruction past all of the instructions between it and the 1071/// end of its block. 1072static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 1073 assert(I->hasOneUse() && "Invariants didn't hold!"); 1074 1075 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1076 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I)) 1077 return false; 1078 1079 // Do not sink alloca instructions out of the entry block. 1080 if (isa<AllocaInst>(I) && I->getParent() == 1081 &DestBlock->getParent()->getEntryBlock()) 1082 return false; 1083 1084 // We can only sink load instructions if there is nothing between the load and 1085 // the end of block that could change the value. 1086 if (I->mayReadFromMemory()) { 1087 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 1088 Scan != E; ++Scan) 1089 if (Scan->mayWriteToMemory()) 1090 return false; 1091 } 1092 1093 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI(); 1094 1095 I->moveBefore(InsertPos); 1096 ++NumSunkInst; 1097 return true; 1098} 1099 1100 1101/// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 1102/// all reachable code to the worklist. 1103/// 1104/// This has a couple of tricks to make the code faster and more powerful. In 1105/// particular, we constant fold and DCE instructions as we go, to avoid adding 1106/// them to the worklist (this significantly speeds up instcombine on code where 1107/// many instructions are dead or constant). Additionally, if we find a branch 1108/// whose condition is a known constant, we only visit the reachable successors. 1109/// 1110static bool AddReachableCodeToWorklist(BasicBlock *BB, 1111 SmallPtrSet<BasicBlock*, 64> &Visited, 1112 InstCombiner &IC, 1113 const TargetData *TD) { 1114 bool MadeIRChange = false; 1115 SmallVector<BasicBlock*, 256> Worklist; 1116 Worklist.push_back(BB); 1117 1118 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 1119 SmallPtrSet<ConstantExpr*, 64> FoldedConstants; 1120 1121 do { 1122 BB = Worklist.pop_back_val(); 1123 1124 // We have now visited this block! If we've already been here, ignore it. 1125 if (!Visited.insert(BB)) continue; 1126 1127 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 1128 Instruction *Inst = BBI++; 1129 1130 // DCE instruction if trivially dead. 1131 if (isInstructionTriviallyDead(Inst)) { 1132 ++NumDeadInst; 1133 DEBUG(errs() << "IC: DCE: " << *Inst << '\n'); 1134 Inst->eraseFromParent(); 1135 continue; 1136 } 1137 1138 // ConstantProp instruction if trivially constant. 1139 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 1140 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1141 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " 1142 << *Inst << '\n'); 1143 Inst->replaceAllUsesWith(C); 1144 ++NumConstProp; 1145 Inst->eraseFromParent(); 1146 continue; 1147 } 1148 1149 if (TD) { 1150 // See if we can constant fold its operands. 1151 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 1152 i != e; ++i) { 1153 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 1154 if (CE == 0) continue; 1155 1156 // If we already folded this constant, don't try again. 1157 if (!FoldedConstants.insert(CE)) 1158 continue; 1159 1160 Constant *NewC = ConstantFoldConstantExpression(CE, TD); 1161 if (NewC && NewC != CE) { 1162 *i = NewC; 1163 MadeIRChange = true; 1164 } 1165 } 1166 } 1167 1168 InstrsForInstCombineWorklist.push_back(Inst); 1169 } 1170 1171 // Recursively visit successors. If this is a branch or switch on a 1172 // constant, only visit the reachable successor. 1173 TerminatorInst *TI = BB->getTerminator(); 1174 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1175 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 1176 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 1177 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 1178 Worklist.push_back(ReachableBB); 1179 continue; 1180 } 1181 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1182 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 1183 // See if this is an explicit destination. 1184 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 1185 if (SI->getCaseValue(i) == Cond) { 1186 BasicBlock *ReachableBB = SI->getSuccessor(i); 1187 Worklist.push_back(ReachableBB); 1188 continue; 1189 } 1190 1191 // Otherwise it is the default destination. 1192 Worklist.push_back(SI->getSuccessor(0)); 1193 continue; 1194 } 1195 } 1196 1197 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 1198 Worklist.push_back(TI->getSuccessor(i)); 1199 } while (!Worklist.empty()); 1200 1201 // Once we've found all of the instructions to add to instcombine's worklist, 1202 // add them in reverse order. This way instcombine will visit from the top 1203 // of the function down. This jives well with the way that it adds all uses 1204 // of instructions to the worklist after doing a transformation, thus avoiding 1205 // some N^2 behavior in pathological cases. 1206 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 1207 InstrsForInstCombineWorklist.size()); 1208 1209 return MadeIRChange; 1210} 1211 1212bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 1213 MadeIRChange = false; 1214 1215 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 1216 << F.getNameStr() << "\n"); 1217 1218 { 1219 // Do a depth-first traversal of the function, populate the worklist with 1220 // the reachable instructions. Ignore blocks that are not reachable. Keep 1221 // track of which blocks we visit. 1222 SmallPtrSet<BasicBlock*, 64> Visited; 1223 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD); 1224 1225 // Do a quick scan over the function. If we find any blocks that are 1226 // unreachable, remove any instructions inside of them. This prevents 1227 // the instcombine code from having to deal with some bad special cases. 1228 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 1229 if (!Visited.count(BB)) { 1230 Instruction *Term = BB->getTerminator(); 1231 while (Term != BB->begin()) { // Remove instrs bottom-up 1232 BasicBlock::iterator I = Term; --I; 1233 1234 DEBUG(errs() << "IC: DCE: " << *I << '\n'); 1235 // A debug intrinsic shouldn't force another iteration if we weren't 1236 // going to do one without it. 1237 if (!isa<DbgInfoIntrinsic>(I)) { 1238 ++NumDeadInst; 1239 MadeIRChange = true; 1240 } 1241 1242 // If I is not void type then replaceAllUsesWith undef. 1243 // This allows ValueHandlers and custom metadata to adjust itself. 1244 if (!I->getType()->isVoidTy()) 1245 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1246 I->eraseFromParent(); 1247 } 1248 } 1249 } 1250 1251 while (!Worklist.isEmpty()) { 1252 Instruction *I = Worklist.RemoveOne(); 1253 if (I == 0) continue; // skip null values. 1254 1255 // Check to see if we can DCE the instruction. 1256 if (isInstructionTriviallyDead(I)) { 1257 DEBUG(errs() << "IC: DCE: " << *I << '\n'); 1258 EraseInstFromFunction(*I); 1259 ++NumDeadInst; 1260 MadeIRChange = true; 1261 continue; 1262 } 1263 1264 // Instruction isn't dead, see if we can constant propagate it. 1265 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 1266 if (Constant *C = ConstantFoldInstruction(I, TD)) { 1267 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 1268 1269 // Add operands to the worklist. 1270 ReplaceInstUsesWith(*I, C); 1271 ++NumConstProp; 1272 EraseInstFromFunction(*I); 1273 MadeIRChange = true; 1274 continue; 1275 } 1276 1277 // See if we can trivially sink this instruction to a successor basic block. 1278 if (I->hasOneUse()) { 1279 BasicBlock *BB = I->getParent(); 1280 Instruction *UserInst = cast<Instruction>(I->use_back()); 1281 BasicBlock *UserParent; 1282 1283 // Get the block the use occurs in. 1284 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 1285 UserParent = PN->getIncomingBlock(I->use_begin().getUse()); 1286 else 1287 UserParent = UserInst->getParent(); 1288 1289 if (UserParent != BB) { 1290 bool UserIsSuccessor = false; 1291 // See if the user is one of our successors. 1292 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 1293 if (*SI == UserParent) { 1294 UserIsSuccessor = true; 1295 break; 1296 } 1297 1298 // If the user is one of our immediate successors, and if that successor 1299 // only has us as a predecessors (we'd have to split the critical edge 1300 // otherwise), we can keep going. 1301 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 1302 // Okay, the CFG is simple enough, try to sink this instruction. 1303 MadeIRChange |= TryToSinkInstruction(I, UserParent); 1304 } 1305 } 1306 1307 // Now that we have an instruction, try combining it to simplify it. 1308 Builder->SetInsertPoint(I->getParent(), I); 1309 1310#ifndef NDEBUG 1311 std::string OrigI; 1312#endif 1313 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 1314 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n'); 1315 1316 if (Instruction *Result = visit(*I)) { 1317 ++NumCombined; 1318 // Should we replace the old instruction with a new one? 1319 if (Result != I) { 1320 DEBUG(errs() << "IC: Old = " << *I << '\n' 1321 << " New = " << *Result << '\n'); 1322 1323 // Everything uses the new instruction now. 1324 I->replaceAllUsesWith(Result); 1325 1326 // Push the new instruction and any users onto the worklist. 1327 Worklist.Add(Result); 1328 Worklist.AddUsersToWorkList(*Result); 1329 1330 // Move the name to the new instruction first. 1331 Result->takeName(I); 1332 1333 // Insert the new instruction into the basic block... 1334 BasicBlock *InstParent = I->getParent(); 1335 BasicBlock::iterator InsertPos = I; 1336 1337 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert 1338 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs. 1339 ++InsertPos; 1340 1341 InstParent->getInstList().insert(InsertPos, Result); 1342 1343 EraseInstFromFunction(*I); 1344 } else { 1345#ifndef NDEBUG 1346 DEBUG(errs() << "IC: Mod = " << OrigI << '\n' 1347 << " New = " << *I << '\n'); 1348#endif 1349 1350 // If the instruction was modified, it's possible that it is now dead. 1351 // if so, remove it. 1352 if (isInstructionTriviallyDead(I)) { 1353 EraseInstFromFunction(*I); 1354 } else { 1355 Worklist.Add(I); 1356 Worklist.AddUsersToWorkList(*I); 1357 } 1358 } 1359 MadeIRChange = true; 1360 } 1361 } 1362 1363 Worklist.Zap(); 1364 return MadeIRChange; 1365} 1366 1367 1368bool InstCombiner::runOnFunction(Function &F) { 1369 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID); 1370 TD = getAnalysisIfAvailable<TargetData>(); 1371 1372 1373 /// Builder - This is an IRBuilder that automatically inserts new 1374 /// instructions into the worklist when they are created. 1375 IRBuilder<true, TargetFolder, InstCombineIRInserter> 1376 TheBuilder(F.getContext(), TargetFolder(TD), 1377 InstCombineIRInserter(Worklist)); 1378 Builder = &TheBuilder; 1379 1380 bool EverMadeChange = false; 1381 1382 // Iterate while there is work to do. 1383 unsigned Iteration = 0; 1384 while (DoOneIteration(F, Iteration++)) 1385 EverMadeChange = true; 1386 1387 Builder = 0; 1388 return EverMadeChange; 1389} 1390 1391FunctionPass *llvm::createInstructionCombiningPass() { 1392 return new InstCombiner(); 1393} 1394