InstructionCombining.cpp revision 916973706d9b2559c7f88cef8b0f4e40a6450c06
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 <algorithm> 52#include <climits> 53using namespace llvm; 54using namespace llvm::PatternMatch; 55 56STATISTIC(NumCombined , "Number of insts combined"); 57STATISTIC(NumConstProp, "Number of constant folds"); 58STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 59STATISTIC(NumSunkInst , "Number of instructions sunk"); 60 61 62char InstCombiner::ID = 0; 63static RegisterPass<InstCombiner> 64X("instcombine", "Combine redundant instructions"); 65 66void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 67 AU.addPreservedID(LCSSAID); 68 AU.setPreservesCFG(); 69} 70 71 72/// ShouldChangeType - Return true if it is desirable to convert a computation 73/// from 'From' to 'To'. We don't want to convert from a legal to an illegal 74/// type for example, or from a smaller to a larger illegal type. 75bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const { 76 assert(From->isIntegerTy() && To->isIntegerTy()); 77 78 // If we don't have TD, we don't know if the source/dest are legal. 79 if (!TD) return false; 80 81 unsigned FromWidth = From->getPrimitiveSizeInBits(); 82 unsigned ToWidth = To->getPrimitiveSizeInBits(); 83 bool FromLegal = TD->isLegalInteger(FromWidth); 84 bool ToLegal = TD->isLegalInteger(ToWidth); 85 86 // If this is a legal integer from type, and the result would be an illegal 87 // type, don't do the transformation. 88 if (FromLegal && !ToLegal) 89 return false; 90 91 // Otherwise, if both are illegal, do not increase the size of the result. We 92 // do allow things like i160 -> i64, but not i64 -> i160. 93 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 94 return false; 95 96 return true; 97} 98 99 100// SimplifyCommutative - This performs a few simplifications for commutative 101// operators: 102// 103// 1. Order operands such that they are listed from right (least complex) to 104// left (most complex). This puts constants before unary operators before 105// binary operators. 106// 107// 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2)) 108// 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2)) 109// 110bool InstCombiner::SimplifyCommutative(BinaryOperator &I) { 111 bool Changed = false; 112 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) 113 Changed = !I.swapOperands(); 114 115 if (!I.isAssociative()) return Changed; 116 117 Instruction::BinaryOps Opcode = I.getOpcode(); 118 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0))) 119 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) { 120 if (isa<Constant>(I.getOperand(1))) { 121 Constant *Folded = ConstantExpr::get(I.getOpcode(), 122 cast<Constant>(I.getOperand(1)), 123 cast<Constant>(Op->getOperand(1))); 124 I.setOperand(0, Op->getOperand(0)); 125 I.setOperand(1, Folded); 126 return true; 127 } 128 129 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1))) 130 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) && 131 Op->hasOneUse() && Op1->hasOneUse()) { 132 Constant *C1 = cast<Constant>(Op->getOperand(1)); 133 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 134 135 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2)) 136 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2); 137 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0), 138 Op1->getOperand(0), 139 Op1->getName(), &I); 140 Worklist.Add(New); 141 I.setOperand(0, New); 142 I.setOperand(1, Folded); 143 return true; 144 } 145 } 146 return Changed; 147} 148 149// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 150// if the LHS is a constant zero (which is the 'negate' form). 151// 152Value *InstCombiner::dyn_castNegVal(Value *V) const { 153 if (BinaryOperator::isNeg(V)) 154 return BinaryOperator::getNegArgument(V); 155 156 // Constants can be considered to be negated values if they can be folded. 157 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 158 return ConstantExpr::getNeg(C); 159 160 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) 161 if (C->getType()->getElementType()->isIntegerTy()) 162 return ConstantExpr::getNeg(C); 163 164 return 0; 165} 166 167// dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 168// instruction if the LHS is a constant negative zero (which is the 'negate' 169// form). 170// 171Value *InstCombiner::dyn_castFNegVal(Value *V) const { 172 if (BinaryOperator::isFNeg(V)) 173 return BinaryOperator::getFNegArgument(V); 174 175 // Constants can be considered to be negated values if they can be folded. 176 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 177 return ConstantExpr::getFNeg(C); 178 179 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) 180 if (C->getType()->getElementType()->isFloatingPointTy()) 181 return ConstantExpr::getFNeg(C); 182 183 return 0; 184} 185 186static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 187 InstCombiner *IC) { 188 if (CastInst *CI = dyn_cast<CastInst>(&I)) 189 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 190 191 // Figure out if the constant is the left or the right argument. 192 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 193 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 194 195 if (Constant *SOC = dyn_cast<Constant>(SO)) { 196 if (ConstIsRHS) 197 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 198 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 199 } 200 201 Value *Op0 = SO, *Op1 = ConstOperand; 202 if (!ConstIsRHS) 203 std::swap(Op0, Op1); 204 205 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) 206 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 207 SO->getName()+".op"); 208 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 209 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 210 SO->getName()+".cmp"); 211 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 212 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 213 SO->getName()+".cmp"); 214 llvm_unreachable("Unknown binary instruction type!"); 215} 216 217// FoldOpIntoSelect - Given an instruction with a select as one operand and a 218// constant as the other operand, try to fold the binary operator into the 219// select arguments. This also works for Cast instructions, which obviously do 220// not have a second operand. 221Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 222 // Don't modify shared select instructions 223 if (!SI->hasOneUse()) return 0; 224 Value *TV = SI->getOperand(1); 225 Value *FV = SI->getOperand(2); 226 227 if (isa<Constant>(TV) || isa<Constant>(FV)) { 228 // Bool selects with constant operands can be folded to logical ops. 229 if (SI->getType()->isIntegerTy(1)) return 0; 230 231 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 232 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 233 234 return SelectInst::Create(SI->getCondition(), SelectTrueVal, 235 SelectFalseVal); 236 } 237 return 0; 238} 239 240 241/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 242/// has a PHI node as operand #0, see if we can fold the instruction into the 243/// PHI (which is only possible if all operands to the PHI are constants). 244/// 245/// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms 246/// that would normally be unprofitable because they strongly encourage jump 247/// threading. 248Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I, 249 bool AllowAggressive) { 250 AllowAggressive = false; 251 PHINode *PN = cast<PHINode>(I.getOperand(0)); 252 unsigned NumPHIValues = PN->getNumIncomingValues(); 253 if (NumPHIValues == 0 || 254 // We normally only transform phis with a single use, unless we're trying 255 // hard to make jump threading happen. 256 (!PN->hasOneUse() && !AllowAggressive)) 257 return 0; 258 259 260 // Check to see if all of the operands of the PHI are simple constants 261 // (constantint/constantfp/undef). If there is one non-constant value, 262 // remember the BB it is in. If there is more than one or if *it* is a PHI, 263 // bail out. We don't do arbitrary constant expressions here because moving 264 // their computation can be expensive without a cost model. 265 BasicBlock *NonConstBB = 0; 266 for (unsigned i = 0; i != NumPHIValues; ++i) 267 if (!isa<Constant>(PN->getIncomingValue(i)) || 268 isa<ConstantExpr>(PN->getIncomingValue(i))) { 269 if (NonConstBB) return 0; // More than one non-const value. 270 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi. 271 NonConstBB = PN->getIncomingBlock(i); 272 273 // If the incoming non-constant value is in I's block, we have an infinite 274 // loop. 275 if (NonConstBB == I.getParent()) 276 return 0; 277 } 278 279 // If there is exactly one non-constant value, we can insert a copy of the 280 // operation in that block. However, if this is a critical edge, we would be 281 // inserting the computation one some other paths (e.g. inside a loop). Only 282 // do this if the pred block is unconditionally branching into the phi block. 283 if (NonConstBB != 0 && !AllowAggressive) { 284 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 285 if (!BI || !BI->isUnconditional()) return 0; 286 } 287 288 // Okay, we can do the transformation: create the new PHI node. 289 PHINode *NewPN = PHINode::Create(I.getType(), ""); 290 NewPN->reserveOperandSpace(PN->getNumOperands()/2); 291 InsertNewInstBefore(NewPN, *PN); 292 NewPN->takeName(PN); 293 294 // Next, add all of the operands to the PHI. 295 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 296 // We only currently try to fold the condition of a select when it is a phi, 297 // not the true/false values. 298 Value *TrueV = SI->getTrueValue(); 299 Value *FalseV = SI->getFalseValue(); 300 BasicBlock *PhiTransBB = PN->getParent(); 301 for (unsigned i = 0; i != NumPHIValues; ++i) { 302 BasicBlock *ThisBB = PN->getIncomingBlock(i); 303 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 304 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 305 Value *InV = 0; 306 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) { 307 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 308 } else { 309 assert(PN->getIncomingBlock(i) == NonConstBB); 310 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred, 311 FalseVInPred, 312 "phitmp", NonConstBB->getTerminator()); 313 Worklist.Add(cast<Instruction>(InV)); 314 } 315 NewPN->addIncoming(InV, ThisBB); 316 } 317 } else if (I.getNumOperands() == 2) { 318 Constant *C = cast<Constant>(I.getOperand(1)); 319 for (unsigned i = 0; i != NumPHIValues; ++i) { 320 Value *InV = 0; 321 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) { 322 if (CmpInst *CI = dyn_cast<CmpInst>(&I)) 323 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 324 else 325 InV = ConstantExpr::get(I.getOpcode(), InC, C); 326 } else { 327 assert(PN->getIncomingBlock(i) == NonConstBB); 328 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) 329 InV = BinaryOperator::Create(BO->getOpcode(), 330 PN->getIncomingValue(i), C, "phitmp", 331 NonConstBB->getTerminator()); 332 else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) 333 InV = CmpInst::Create(CI->getOpcode(), 334 CI->getPredicate(), 335 PN->getIncomingValue(i), C, "phitmp", 336 NonConstBB->getTerminator()); 337 else 338 llvm_unreachable("Unknown binop!"); 339 340 Worklist.Add(cast<Instruction>(InV)); 341 } 342 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 343 } 344 } else { 345 CastInst *CI = cast<CastInst>(&I); 346 const Type *RetTy = CI->getType(); 347 for (unsigned i = 0; i != NumPHIValues; ++i) { 348 Value *InV; 349 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) { 350 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 351 } else { 352 assert(PN->getIncomingBlock(i) == NonConstBB); 353 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i), 354 I.getType(), "phitmp", 355 NonConstBB->getTerminator()); 356 Worklist.Add(cast<Instruction>(InV)); 357 } 358 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 359 } 360 } 361 return ReplaceInstUsesWith(I, NewPN); 362} 363 364/// FindElementAtOffset - Given a type and a constant offset, determine whether 365/// or not there is a sequence of GEP indices into the type that will land us at 366/// the specified offset. If so, fill them into NewIndices and return the 367/// resultant element type, otherwise return null. 368const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset, 369 SmallVectorImpl<Value*> &NewIndices) { 370 if (!TD) return 0; 371 if (!Ty->isSized()) return 0; 372 373 // Start with the index over the outer type. Note that the type size 374 // might be zero (even if the offset isn't zero) if the indexed type 375 // is something like [0 x {int, int}] 376 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 377 int64_t FirstIdx = 0; 378 if (int64_t TySize = TD->getTypeAllocSize(Ty)) { 379 FirstIdx = Offset/TySize; 380 Offset -= FirstIdx*TySize; 381 382 // Handle hosts where % returns negative instead of values [0..TySize). 383 if (Offset < 0) { 384 --FirstIdx; 385 Offset += TySize; 386 assert(Offset >= 0); 387 } 388 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 389 } 390 391 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 392 393 // Index into the types. If we fail, set OrigBase to null. 394 while (Offset) { 395 // Indexing into tail padding between struct/array elements. 396 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty)) 397 return 0; 398 399 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 400 const StructLayout *SL = TD->getStructLayout(STy); 401 assert(Offset < (int64_t)SL->getSizeInBytes() && 402 "Offset must stay within the indexed type"); 403 404 unsigned Elt = SL->getElementContainingOffset(Offset); 405 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 406 Elt)); 407 408 Offset -= SL->getElementOffset(Elt); 409 Ty = STy->getElementType(Elt); 410 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 411 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType()); 412 assert(EltSize && "Cannot index into a zero-sized array"); 413 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 414 Offset %= EltSize; 415 Ty = AT->getElementType(); 416 } else { 417 // Otherwise, we can't index into the middle of this atomic type, bail. 418 return 0; 419 } 420 } 421 422 return Ty; 423} 424 425 426 427Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 428 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 429 430 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD)) 431 return ReplaceInstUsesWith(GEP, V); 432 433 Value *PtrOp = GEP.getOperand(0); 434 435 if (isa<UndefValue>(GEP.getOperand(0))) 436 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType())); 437 438 // Eliminate unneeded casts for indices. 439 if (TD) { 440 bool MadeChange = false; 441 unsigned PtrSize = TD->getPointerSizeInBits(); 442 443 gep_type_iterator GTI = gep_type_begin(GEP); 444 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 445 I != E; ++I, ++GTI) { 446 if (!isa<SequentialType>(*GTI)) continue; 447 448 // If we are using a wider index than needed for this platform, shrink it 449 // to what we need. If narrower, sign-extend it to what we need. This 450 // explicit cast can make subsequent optimizations more obvious. 451 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth(); 452 if (OpBits == PtrSize) 453 continue; 454 455 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true); 456 MadeChange = true; 457 } 458 if (MadeChange) return &GEP; 459 } 460 461 // Combine Indices - If the source pointer to this getelementptr instruction 462 // is a getelementptr instruction, combine the indices of the two 463 // getelementptr instructions into a single instruction. 464 // 465 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 466 // Note that if our source is a gep chain itself that we wait for that 467 // chain to be resolved before we perform this transformation. This 468 // avoids us creating a TON of code in some cases. 469 // 470 if (GetElementPtrInst *SrcGEP = 471 dyn_cast<GetElementPtrInst>(Src->getOperand(0))) 472 if (SrcGEP->getNumOperands() == 2) 473 return 0; // Wait until our source is folded to completion. 474 475 SmallVector<Value*, 8> Indices; 476 477 // Find out whether the last index in the source GEP is a sequential idx. 478 bool EndsWithSequential = false; 479 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 480 I != E; ++I) 481 EndsWithSequential = !(*I)->isStructTy(); 482 483 // Can we combine the two pointer arithmetics offsets? 484 if (EndsWithSequential) { 485 // Replace: gep (gep %P, long B), long A, ... 486 // With: T = long A+B; gep %P, T, ... 487 // 488 Value *Sum; 489 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 490 Value *GO1 = GEP.getOperand(1); 491 if (SO1 == Constant::getNullValue(SO1->getType())) { 492 Sum = GO1; 493 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 494 Sum = SO1; 495 } else { 496 // If they aren't the same type, then the input hasn't been processed 497 // by the loop above yet (which canonicalizes sequential index types to 498 // intptr_t). Just avoid transforming this until the input has been 499 // normalized. 500 if (SO1->getType() != GO1->getType()) 501 return 0; 502 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 503 } 504 505 // Update the GEP in place if possible. 506 if (Src->getNumOperands() == 2) { 507 GEP.setOperand(0, Src->getOperand(0)); 508 GEP.setOperand(1, Sum); 509 return &GEP; 510 } 511 Indices.append(Src->op_begin()+1, Src->op_end()-1); 512 Indices.push_back(Sum); 513 Indices.append(GEP.op_begin()+2, GEP.op_end()); 514 } else if (isa<Constant>(*GEP.idx_begin()) && 515 cast<Constant>(*GEP.idx_begin())->isNullValue() && 516 Src->getNumOperands() != 1) { 517 // Otherwise we can do the fold if the first index of the GEP is a zero 518 Indices.append(Src->op_begin()+1, Src->op_end()); 519 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 520 } 521 522 if (!Indices.empty()) 523 return (GEP.isInBounds() && Src->isInBounds()) ? 524 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(), 525 Indices.end(), GEP.getName()) : 526 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(), 527 Indices.end(), GEP.getName()); 528 } 529 530 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 531 Value *StrippedPtr = PtrOp->stripPointerCasts(); 532 if (StrippedPtr != PtrOp) { 533 const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType()); 534 535 bool HasZeroPointerIndex = false; 536 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 537 HasZeroPointerIndex = C->isZero(); 538 539 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 540 // into : GEP [10 x i8]* X, i32 0, ... 541 // 542 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 543 // into : GEP i8* X, ... 544 // 545 // This occurs when the program declares an array extern like "int X[];" 546 if (HasZeroPointerIndex) { 547 const PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 548 if (const ArrayType *CATy = 549 dyn_cast<ArrayType>(CPTy->getElementType())) { 550 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 551 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 552 // -> GEP i8* X, ... 553 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 554 GetElementPtrInst *Res = 555 GetElementPtrInst::Create(StrippedPtr, Idx.begin(), 556 Idx.end(), GEP.getName()); 557 Res->setIsInBounds(GEP.isInBounds()); 558 return Res; 559 } 560 561 if (const ArrayType *XATy = 562 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 563 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 564 if (CATy->getElementType() == XATy->getElementType()) { 565 // -> GEP [10 x i8]* X, i32 0, ... 566 // At this point, we know that the cast source type is a pointer 567 // to an array of the same type as the destination pointer 568 // array. Because the array type is never stepped over (there 569 // is a leading zero) we can fold the cast into this GEP. 570 GEP.setOperand(0, StrippedPtr); 571 return &GEP; 572 } 573 } 574 } 575 } else if (GEP.getNumOperands() == 2) { 576 // Transform things like: 577 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 578 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 579 const Type *SrcElTy = StrippedPtrTy->getElementType(); 580 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType(); 581 if (TD && SrcElTy->isArrayTy() && 582 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) == 583 TD->getTypeAllocSize(ResElTy)) { 584 Value *Idx[2]; 585 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); 586 Idx[1] = GEP.getOperand(1); 587 Value *NewGEP = GEP.isInBounds() ? 588 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) : 589 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()); 590 // V and GEP are both pointer types --> BitCast 591 return new BitCastInst(NewGEP, GEP.getType()); 592 } 593 594 // Transform things like: 595 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 596 // (where tmp = 8*tmp2) into: 597 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 598 599 if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) { 600 uint64_t ArrayEltSize = 601 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()); 602 603 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We 604 // allow either a mul, shift, or constant here. 605 Value *NewIdx = 0; 606 ConstantInt *Scale = 0; 607 if (ArrayEltSize == 1) { 608 NewIdx = GEP.getOperand(1); 609 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1); 610 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) { 611 NewIdx = ConstantInt::get(CI->getType(), 1); 612 Scale = CI; 613 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){ 614 if (Inst->getOpcode() == Instruction::Shl && 615 isa<ConstantInt>(Inst->getOperand(1))) { 616 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1)); 617 uint32_t ShAmtVal = ShAmt->getLimitedValue(64); 618 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()), 619 1ULL << ShAmtVal); 620 NewIdx = Inst->getOperand(0); 621 } else if (Inst->getOpcode() == Instruction::Mul && 622 isa<ConstantInt>(Inst->getOperand(1))) { 623 Scale = cast<ConstantInt>(Inst->getOperand(1)); 624 NewIdx = Inst->getOperand(0); 625 } 626 } 627 628 // If the index will be to exactly the right offset with the scale taken 629 // out, perform the transformation. Note, we don't know whether Scale is 630 // signed or not. We'll use unsigned version of division/modulo 631 // operation after making sure Scale doesn't have the sign bit set. 632 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL && 633 Scale->getZExtValue() % ArrayEltSize == 0) { 634 Scale = ConstantInt::get(Scale->getType(), 635 Scale->getZExtValue() / ArrayEltSize); 636 if (Scale->getZExtValue() != 1) { 637 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(), 638 false /*ZExt*/); 639 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale"); 640 } 641 642 // Insert the new GEP instruction. 643 Value *Idx[2]; 644 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); 645 Idx[1] = NewIdx; 646 Value *NewGEP = GEP.isInBounds() ? 647 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()): 648 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()); 649 // The NewGEP must be pointer typed, so must the old one -> BitCast 650 return new BitCastInst(NewGEP, GEP.getType()); 651 } 652 } 653 } 654 } 655 656 /// See if we can simplify: 657 /// X = bitcast A* to B* 658 /// Y = gep X, <...constant indices...> 659 /// into a gep of the original struct. This is important for SROA and alias 660 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 661 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 662 if (TD && 663 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) { 664 // Determine how much the GEP moves the pointer. We are guaranteed to get 665 // a constant back from EmitGEPOffset. 666 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP)); 667 int64_t Offset = OffsetV->getSExtValue(); 668 669 // If this GEP instruction doesn't move the pointer, just replace the GEP 670 // with a bitcast of the real input to the dest type. 671 if (Offset == 0) { 672 // If the bitcast is of an allocation, and the allocation will be 673 // converted to match the type of the cast, don't touch this. 674 if (isa<AllocaInst>(BCI->getOperand(0)) || 675 isMalloc(BCI->getOperand(0))) { 676 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 677 if (Instruction *I = visitBitCast(*BCI)) { 678 if (I != BCI) { 679 I->takeName(BCI); 680 BCI->getParent()->getInstList().insert(BCI, I); 681 ReplaceInstUsesWith(*BCI, I); 682 } 683 return &GEP; 684 } 685 } 686 return new BitCastInst(BCI->getOperand(0), GEP.getType()); 687 } 688 689 // Otherwise, if the offset is non-zero, we need to find out if there is a 690 // field at Offset in 'A's type. If so, we can pull the cast through the 691 // GEP. 692 SmallVector<Value*, 8> NewIndices; 693 const Type *InTy = 694 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType(); 695 if (FindElementAtOffset(InTy, Offset, NewIndices)) { 696 Value *NGEP = GEP.isInBounds() ? 697 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(), 698 NewIndices.end()) : 699 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(), 700 NewIndices.end()); 701 702 if (NGEP->getType() == GEP.getType()) 703 return ReplaceInstUsesWith(GEP, NGEP); 704 NGEP->takeName(&GEP); 705 return new BitCastInst(NGEP, GEP.getType()); 706 } 707 } 708 } 709 710 return 0; 711} 712 713 714 715static bool IsOnlyNullComparedAndFreed(const Value &V) { 716 for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end(); 717 UI != UE; ++UI) { 718 if (isFreeCall(*UI)) 719 continue; 720 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) 721 if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1))) 722 continue; 723 return false; 724 } 725 return true; 726} 727 728Instruction *InstCombiner::visitMalloc(Instruction &MI) { 729 // If we have a malloc call which is only used in any amount of comparisons 730 // to null and free calls, delete the calls and replace the comparisons with 731 // true or false as appropriate. 732 if (IsOnlyNullComparedAndFreed(MI)) { 733 for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end(); 734 UI != UE;) { 735 // We can assume that every remaining use is a free call or an icmp eq/ne 736 // to null, so the cast is safe. 737 Instruction *I = cast<Instruction>(*UI); 738 739 // Early increment here, as we're about to get rid of the user. 740 ++UI; 741 742 if (isFreeCall(I)) { 743 EraseInstFromFunction(*cast<CallInst>(I)); 744 continue; 745 } 746 // Again, the cast is safe. 747 ICmpInst *C = cast<ICmpInst>(I); 748 ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()), 749 C->isFalseWhenEqual())); 750 EraseInstFromFunction(*C); 751 } 752 return EraseInstFromFunction(MI); 753 } 754 return 0; 755} 756 757 758 759Instruction *InstCombiner::visitFree(CallInst &FI) { 760 Value *Op = FI.getArgOperand(0); 761 762 // free undef -> unreachable. 763 if (isa<UndefValue>(Op)) { 764 // Insert a new store to null because we cannot modify the CFG here. 765 new StoreInst(ConstantInt::getTrue(FI.getContext()), 766 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI); 767 return EraseInstFromFunction(FI); 768 } 769 770 // If we have 'free null' delete the instruction. This can happen in stl code 771 // when lots of inlining happens. 772 if (isa<ConstantPointerNull>(Op)) 773 return EraseInstFromFunction(FI); 774 775 return 0; 776} 777 778 779 780Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 781 // Change br (not X), label True, label False to: br X, label False, True 782 Value *X = 0; 783 BasicBlock *TrueDest; 784 BasicBlock *FalseDest; 785 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 786 !isa<Constant>(X)) { 787 // Swap Destinations and condition... 788 BI.setCondition(X); 789 BI.setSuccessor(0, FalseDest); 790 BI.setSuccessor(1, TrueDest); 791 return &BI; 792 } 793 794 // Cannonicalize fcmp_one -> fcmp_oeq 795 FCmpInst::Predicate FPred; Value *Y; 796 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 797 TrueDest, FalseDest)) && 798 BI.getCondition()->hasOneUse()) 799 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 800 FPred == FCmpInst::FCMP_OGE) { 801 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 802 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 803 804 // Swap Destinations and condition. 805 BI.setSuccessor(0, FalseDest); 806 BI.setSuccessor(1, TrueDest); 807 Worklist.Add(Cond); 808 return &BI; 809 } 810 811 // Cannonicalize icmp_ne -> icmp_eq 812 ICmpInst::Predicate IPred; 813 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 814 TrueDest, FalseDest)) && 815 BI.getCondition()->hasOneUse()) 816 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 817 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 818 IPred == ICmpInst::ICMP_SGE) { 819 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 820 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 821 // Swap Destinations and condition. 822 BI.setSuccessor(0, FalseDest); 823 BI.setSuccessor(1, TrueDest); 824 Worklist.Add(Cond); 825 return &BI; 826 } 827 828 return 0; 829} 830 831Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 832 Value *Cond = SI.getCondition(); 833 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 834 if (I->getOpcode() == Instruction::Add) 835 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 836 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 837 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) 838 SI.setOperand(i, 839 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)), 840 AddRHS)); 841 SI.setOperand(0, I->getOperand(0)); 842 Worklist.Add(I); 843 return &SI; 844 } 845 } 846 return 0; 847} 848 849Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 850 Value *Agg = EV.getAggregateOperand(); 851 852 if (!EV.hasIndices()) 853 return ReplaceInstUsesWith(EV, Agg); 854 855 if (Constant *C = dyn_cast<Constant>(Agg)) { 856 if (isa<UndefValue>(C)) 857 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType())); 858 859 if (isa<ConstantAggregateZero>(C)) 860 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType())); 861 862 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) { 863 // Extract the element indexed by the first index out of the constant 864 Value *V = C->getOperand(*EV.idx_begin()); 865 if (EV.getNumIndices() > 1) 866 // Extract the remaining indices out of the constant indexed by the 867 // first index 868 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end()); 869 else 870 return ReplaceInstUsesWith(EV, V); 871 } 872 return 0; // Can't handle other constants 873 } 874 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 875 // We're extracting from an insertvalue instruction, compare the indices 876 const unsigned *exti, *exte, *insi, *inse; 877 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 878 exte = EV.idx_end(), inse = IV->idx_end(); 879 exti != exte && insi != inse; 880 ++exti, ++insi) { 881 if (*insi != *exti) 882 // The insert and extract both reference distinctly different elements. 883 // This means the extract is not influenced by the insert, and we can 884 // replace the aggregate operand of the extract with the aggregate 885 // operand of the insert. i.e., replace 886 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 887 // %E = extractvalue { i32, { i32 } } %I, 0 888 // with 889 // %E = extractvalue { i32, { i32 } } %A, 0 890 return ExtractValueInst::Create(IV->getAggregateOperand(), 891 EV.idx_begin(), EV.idx_end()); 892 } 893 if (exti == exte && insi == inse) 894 // Both iterators are at the end: Index lists are identical. Replace 895 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 896 // %C = extractvalue { i32, { i32 } } %B, 1, 0 897 // with "i32 42" 898 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 899 if (exti == exte) { 900 // The extract list is a prefix of the insert list. i.e. replace 901 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 902 // %E = extractvalue { i32, { i32 } } %I, 1 903 // with 904 // %X = extractvalue { i32, { i32 } } %A, 1 905 // %E = insertvalue { i32 } %X, i32 42, 0 906 // by switching the order of the insert and extract (though the 907 // insertvalue should be left in, since it may have other uses). 908 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 909 EV.idx_begin(), EV.idx_end()); 910 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 911 insi, inse); 912 } 913 if (insi == inse) 914 // The insert list is a prefix of the extract list 915 // We can simply remove the common indices from the extract and make it 916 // operate on the inserted value instead of the insertvalue result. 917 // i.e., replace 918 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 919 // %E = extractvalue { i32, { i32 } } %I, 1, 0 920 // with 921 // %E extractvalue { i32 } { i32 42 }, 0 922 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 923 exti, exte); 924 } 925 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 926 // We're extracting from an intrinsic, see if we're the only user, which 927 // allows us to simplify multiple result intrinsics to simpler things that 928 // just get one value. 929 if (II->hasOneUse()) { 930 // Check if we're grabbing the overflow bit or the result of a 'with 931 // overflow' intrinsic. If it's the latter we can remove the intrinsic 932 // and replace it with a traditional binary instruction. 933 switch (II->getIntrinsicID()) { 934 case Intrinsic::uadd_with_overflow: 935 case Intrinsic::sadd_with_overflow: 936 if (*EV.idx_begin() == 0) { // Normal result. 937 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 938 II->replaceAllUsesWith(UndefValue::get(II->getType())); 939 EraseInstFromFunction(*II); 940 return BinaryOperator::CreateAdd(LHS, RHS); 941 } 942 break; 943 case Intrinsic::usub_with_overflow: 944 case Intrinsic::ssub_with_overflow: 945 if (*EV.idx_begin() == 0) { // Normal result. 946 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 947 II->replaceAllUsesWith(UndefValue::get(II->getType())); 948 EraseInstFromFunction(*II); 949 return BinaryOperator::CreateSub(LHS, RHS); 950 } 951 break; 952 case Intrinsic::umul_with_overflow: 953 case Intrinsic::smul_with_overflow: 954 if (*EV.idx_begin() == 0) { // Normal result. 955 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 956 II->replaceAllUsesWith(UndefValue::get(II->getType())); 957 EraseInstFromFunction(*II); 958 return BinaryOperator::CreateMul(LHS, RHS); 959 } 960 break; 961 default: 962 break; 963 } 964 } 965 } 966 // Can't simplify extracts from other values. Note that nested extracts are 967 // already simplified implicitely by the above (extract ( extract (insert) ) 968 // will be translated into extract ( insert ( extract ) ) first and then just 969 // the value inserted, if appropriate). 970 return 0; 971} 972 973 974 975 976/// TryToSinkInstruction - Try to move the specified instruction from its 977/// current block into the beginning of DestBlock, which can only happen if it's 978/// safe to move the instruction past all of the instructions between it and the 979/// end of its block. 980static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 981 assert(I->hasOneUse() && "Invariants didn't hold!"); 982 983 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 984 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I)) 985 return false; 986 987 // Do not sink alloca instructions out of the entry block. 988 if (isa<AllocaInst>(I) && I->getParent() == 989 &DestBlock->getParent()->getEntryBlock()) 990 return false; 991 992 // We can only sink load instructions if there is nothing between the load and 993 // the end of block that could change the value. 994 if (I->mayReadFromMemory()) { 995 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 996 Scan != E; ++Scan) 997 if (Scan->mayWriteToMemory()) 998 return false; 999 } 1000 1001 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI(); 1002 1003 I->moveBefore(InsertPos); 1004 ++NumSunkInst; 1005 return true; 1006} 1007 1008 1009/// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 1010/// all reachable code to the worklist. 1011/// 1012/// This has a couple of tricks to make the code faster and more powerful. In 1013/// particular, we constant fold and DCE instructions as we go, to avoid adding 1014/// them to the worklist (this significantly speeds up instcombine on code where 1015/// many instructions are dead or constant). Additionally, if we find a branch 1016/// whose condition is a known constant, we only visit the reachable successors. 1017/// 1018static bool AddReachableCodeToWorklist(BasicBlock *BB, 1019 SmallPtrSet<BasicBlock*, 64> &Visited, 1020 InstCombiner &IC, 1021 const TargetData *TD) { 1022 bool MadeIRChange = false; 1023 SmallVector<BasicBlock*, 256> Worklist; 1024 Worklist.push_back(BB); 1025 1026 std::vector<Instruction*> InstrsForInstCombineWorklist; 1027 InstrsForInstCombineWorklist.reserve(128); 1028 1029 SmallPtrSet<ConstantExpr*, 64> FoldedConstants; 1030 1031 do { 1032 BB = Worklist.pop_back_val(); 1033 1034 // We have now visited this block! If we've already been here, ignore it. 1035 if (!Visited.insert(BB)) continue; 1036 1037 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 1038 Instruction *Inst = BBI++; 1039 1040 // DCE instruction if trivially dead. 1041 if (isInstructionTriviallyDead(Inst)) { 1042 ++NumDeadInst; 1043 DEBUG(errs() << "IC: DCE: " << *Inst << '\n'); 1044 Inst->eraseFromParent(); 1045 continue; 1046 } 1047 1048 // ConstantProp instruction if trivially constant. 1049 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 1050 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1051 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " 1052 << *Inst << '\n'); 1053 Inst->replaceAllUsesWith(C); 1054 ++NumConstProp; 1055 Inst->eraseFromParent(); 1056 continue; 1057 } 1058 1059 if (TD) { 1060 // See if we can constant fold its operands. 1061 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 1062 i != e; ++i) { 1063 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 1064 if (CE == 0) continue; 1065 1066 // If we already folded this constant, don't try again. 1067 if (!FoldedConstants.insert(CE)) 1068 continue; 1069 1070 Constant *NewC = ConstantFoldConstantExpression(CE, TD); 1071 if (NewC && NewC != CE) { 1072 *i = NewC; 1073 MadeIRChange = true; 1074 } 1075 } 1076 } 1077 1078 InstrsForInstCombineWorklist.push_back(Inst); 1079 } 1080 1081 // Recursively visit successors. If this is a branch or switch on a 1082 // constant, only visit the reachable successor. 1083 TerminatorInst *TI = BB->getTerminator(); 1084 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1085 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 1086 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 1087 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 1088 Worklist.push_back(ReachableBB); 1089 continue; 1090 } 1091 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1092 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 1093 // See if this is an explicit destination. 1094 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 1095 if (SI->getCaseValue(i) == Cond) { 1096 BasicBlock *ReachableBB = SI->getSuccessor(i); 1097 Worklist.push_back(ReachableBB); 1098 continue; 1099 } 1100 1101 // Otherwise it is the default destination. 1102 Worklist.push_back(SI->getSuccessor(0)); 1103 continue; 1104 } 1105 } 1106 1107 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 1108 Worklist.push_back(TI->getSuccessor(i)); 1109 } while (!Worklist.empty()); 1110 1111 // Once we've found all of the instructions to add to instcombine's worklist, 1112 // add them in reverse order. This way instcombine will visit from the top 1113 // of the function down. This jives well with the way that it adds all uses 1114 // of instructions to the worklist after doing a transformation, thus avoiding 1115 // some N^2 behavior in pathological cases. 1116 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 1117 InstrsForInstCombineWorklist.size()); 1118 1119 return MadeIRChange; 1120} 1121 1122bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 1123 MadeIRChange = false; 1124 1125 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 1126 << F.getNameStr() << "\n"); 1127 1128 { 1129 // Do a depth-first traversal of the function, populate the worklist with 1130 // the reachable instructions. Ignore blocks that are not reachable. Keep 1131 // track of which blocks we visit. 1132 SmallPtrSet<BasicBlock*, 64> Visited; 1133 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD); 1134 1135 // Do a quick scan over the function. If we find any blocks that are 1136 // unreachable, remove any instructions inside of them. This prevents 1137 // the instcombine code from having to deal with some bad special cases. 1138 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 1139 if (!Visited.count(BB)) { 1140 Instruction *Term = BB->getTerminator(); 1141 while (Term != BB->begin()) { // Remove instrs bottom-up 1142 BasicBlock::iterator I = Term; --I; 1143 1144 DEBUG(errs() << "IC: DCE: " << *I << '\n'); 1145 // A debug intrinsic shouldn't force another iteration if we weren't 1146 // going to do one without it. 1147 if (!isa<DbgInfoIntrinsic>(I)) { 1148 ++NumDeadInst; 1149 MadeIRChange = true; 1150 } 1151 1152 // If I is not void type then replaceAllUsesWith undef. 1153 // This allows ValueHandlers and custom metadata to adjust itself. 1154 if (!I->getType()->isVoidTy()) 1155 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1156 I->eraseFromParent(); 1157 } 1158 } 1159 } 1160 1161 while (!Worklist.isEmpty()) { 1162 Instruction *I = Worklist.RemoveOne(); 1163 if (I == 0) continue; // skip null values. 1164 1165 // Check to see if we can DCE the instruction. 1166 if (isInstructionTriviallyDead(I)) { 1167 DEBUG(errs() << "IC: DCE: " << *I << '\n'); 1168 EraseInstFromFunction(*I); 1169 ++NumDeadInst; 1170 MadeIRChange = true; 1171 continue; 1172 } 1173 1174 // Instruction isn't dead, see if we can constant propagate it. 1175 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 1176 if (Constant *C = ConstantFoldInstruction(I, TD)) { 1177 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 1178 1179 // Add operands to the worklist. 1180 ReplaceInstUsesWith(*I, C); 1181 ++NumConstProp; 1182 EraseInstFromFunction(*I); 1183 MadeIRChange = true; 1184 continue; 1185 } 1186 1187 // See if we can trivially sink this instruction to a successor basic block. 1188 if (I->hasOneUse()) { 1189 BasicBlock *BB = I->getParent(); 1190 Instruction *UserInst = cast<Instruction>(I->use_back()); 1191 BasicBlock *UserParent; 1192 1193 // Get the block the use occurs in. 1194 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 1195 UserParent = PN->getIncomingBlock(I->use_begin().getUse()); 1196 else 1197 UserParent = UserInst->getParent(); 1198 1199 if (UserParent != BB) { 1200 bool UserIsSuccessor = false; 1201 // See if the user is one of our successors. 1202 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 1203 if (*SI == UserParent) { 1204 UserIsSuccessor = true; 1205 break; 1206 } 1207 1208 // If the user is one of our immediate successors, and if that successor 1209 // only has us as a predecessors (we'd have to split the critical edge 1210 // otherwise), we can keep going. 1211 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 1212 // Okay, the CFG is simple enough, try to sink this instruction. 1213 MadeIRChange |= TryToSinkInstruction(I, UserParent); 1214 } 1215 } 1216 1217 // Now that we have an instruction, try combining it to simplify it. 1218 Builder->SetInsertPoint(I->getParent(), I); 1219 1220#ifndef NDEBUG 1221 std::string OrigI; 1222#endif 1223 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 1224 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n'); 1225 1226 if (Instruction *Result = visit(*I)) { 1227 ++NumCombined; 1228 // Should we replace the old instruction with a new one? 1229 if (Result != I) { 1230 DEBUG(errs() << "IC: Old = " << *I << '\n' 1231 << " New = " << *Result << '\n'); 1232 1233 // Everything uses the new instruction now. 1234 I->replaceAllUsesWith(Result); 1235 1236 // Push the new instruction and any users onto the worklist. 1237 Worklist.Add(Result); 1238 Worklist.AddUsersToWorkList(*Result); 1239 1240 // Move the name to the new instruction first. 1241 Result->takeName(I); 1242 1243 // Insert the new instruction into the basic block... 1244 BasicBlock *InstParent = I->getParent(); 1245 BasicBlock::iterator InsertPos = I; 1246 1247 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert 1248 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs. 1249 ++InsertPos; 1250 1251 InstParent->getInstList().insert(InsertPos, Result); 1252 1253 EraseInstFromFunction(*I); 1254 } else { 1255#ifndef NDEBUG 1256 DEBUG(errs() << "IC: Mod = " << OrigI << '\n' 1257 << " New = " << *I << '\n'); 1258#endif 1259 1260 // If the instruction was modified, it's possible that it is now dead. 1261 // if so, remove it. 1262 if (isInstructionTriviallyDead(I)) { 1263 EraseInstFromFunction(*I); 1264 } else { 1265 Worklist.Add(I); 1266 Worklist.AddUsersToWorkList(*I); 1267 } 1268 } 1269 MadeIRChange = true; 1270 } 1271 } 1272 1273 Worklist.Zap(); 1274 return MadeIRChange; 1275} 1276 1277 1278bool InstCombiner::runOnFunction(Function &F) { 1279 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID); 1280 TD = getAnalysisIfAvailable<TargetData>(); 1281 1282 1283 /// Builder - This is an IRBuilder that automatically inserts new 1284 /// instructions into the worklist when they are created. 1285 IRBuilder<true, TargetFolder, InstCombineIRInserter> 1286 TheBuilder(F.getContext(), TargetFolder(TD), 1287 InstCombineIRInserter(Worklist)); 1288 Builder = &TheBuilder; 1289 1290 bool EverMadeChange = false; 1291 1292 // Iterate while there is work to do. 1293 unsigned Iteration = 0; 1294 while (DoOneIteration(F, Iteration++)) 1295 EverMadeChange = true; 1296 1297 Builder = 0; 1298 return EverMadeChange; 1299} 1300 1301FunctionPass *llvm::createInstructionCombiningPass() { 1302 return new InstCombiner(); 1303} 1304