InstructionCombining.cpp revision 221345
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"); 61STATISTIC(NumExpand, "Number of expansions"); 62STATISTIC(NumFactor , "Number of factorizations"); 63STATISTIC(NumReassoc , "Number of reassociations"); 64 65// Initialization Routines 66void llvm::initializeInstCombine(PassRegistry &Registry) { 67 initializeInstCombinerPass(Registry); 68} 69 70void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 71 initializeInstCombine(*unwrap(R)); 72} 73 74char InstCombiner::ID = 0; 75INITIALIZE_PASS(InstCombiner, "instcombine", 76 "Combine redundant instructions", false, false) 77 78void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 79 AU.setPreservesCFG(); 80} 81 82 83/// ShouldChangeType - Return true if it is desirable to convert a computation 84/// from 'From' to 'To'. We don't want to convert from a legal to an illegal 85/// type for example, or from a smaller to a larger illegal type. 86bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const { 87 assert(From->isIntegerTy() && To->isIntegerTy()); 88 89 // If we don't have TD, we don't know if the source/dest are legal. 90 if (!TD) return false; 91 92 unsigned FromWidth = From->getPrimitiveSizeInBits(); 93 unsigned ToWidth = To->getPrimitiveSizeInBits(); 94 bool FromLegal = TD->isLegalInteger(FromWidth); 95 bool ToLegal = TD->isLegalInteger(ToWidth); 96 97 // If this is a legal integer from type, and the result would be an illegal 98 // type, don't do the transformation. 99 if (FromLegal && !ToLegal) 100 return false; 101 102 // Otherwise, if both are illegal, do not increase the size of the result. We 103 // do allow things like i160 -> i64, but not i64 -> i160. 104 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 105 return false; 106 107 return true; 108} 109 110 111/// SimplifyAssociativeOrCommutative - This performs a few simplifications for 112/// operators which are associative or commutative: 113// 114// Commutative operators: 115// 116// 1. Order operands such that they are listed from right (least complex) to 117// left (most complex). This puts constants before unary operators before 118// binary operators. 119// 120// Associative operators: 121// 122// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 123// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 124// 125// Associative and commutative operators: 126// 127// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 128// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 129// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 130// if C1 and C2 are constants. 131// 132bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 133 Instruction::BinaryOps Opcode = I.getOpcode(); 134 bool Changed = false; 135 136 do { 137 // Order operands such that they are listed from right (least complex) to 138 // left (most complex). This puts constants before unary operators before 139 // binary operators. 140 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 141 getComplexity(I.getOperand(1))) 142 Changed = !I.swapOperands(); 143 144 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 145 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 146 147 if (I.isAssociative()) { 148 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 149 if (Op0 && Op0->getOpcode() == Opcode) { 150 Value *A = Op0->getOperand(0); 151 Value *B = Op0->getOperand(1); 152 Value *C = I.getOperand(1); 153 154 // Does "B op C" simplify? 155 if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) { 156 // It simplifies to V. Form "A op V". 157 I.setOperand(0, A); 158 I.setOperand(1, V); 159 // Conservatively clear the optional flags, since they may not be 160 // preserved by the reassociation. 161 I.clearSubclassOptionalData(); 162 Changed = true; 163 ++NumReassoc; 164 continue; 165 } 166 } 167 168 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 169 if (Op1 && Op1->getOpcode() == Opcode) { 170 Value *A = I.getOperand(0); 171 Value *B = Op1->getOperand(0); 172 Value *C = Op1->getOperand(1); 173 174 // Does "A op B" simplify? 175 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) { 176 // It simplifies to V. Form "V op C". 177 I.setOperand(0, V); 178 I.setOperand(1, C); 179 // Conservatively clear the optional flags, since they may not be 180 // preserved by the reassociation. 181 I.clearSubclassOptionalData(); 182 Changed = true; 183 ++NumReassoc; 184 continue; 185 } 186 } 187 } 188 189 if (I.isAssociative() && I.isCommutative()) { 190 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 191 if (Op0 && Op0->getOpcode() == Opcode) { 192 Value *A = Op0->getOperand(0); 193 Value *B = Op0->getOperand(1); 194 Value *C = I.getOperand(1); 195 196 // Does "C op A" simplify? 197 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) { 198 // It simplifies to V. Form "V op B". 199 I.setOperand(0, V); 200 I.setOperand(1, B); 201 // Conservatively clear the optional flags, since they may not be 202 // preserved by the reassociation. 203 I.clearSubclassOptionalData(); 204 Changed = true; 205 ++NumReassoc; 206 continue; 207 } 208 } 209 210 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 211 if (Op1 && Op1->getOpcode() == Opcode) { 212 Value *A = I.getOperand(0); 213 Value *B = Op1->getOperand(0); 214 Value *C = Op1->getOperand(1); 215 216 // Does "C op A" simplify? 217 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) { 218 // It simplifies to V. Form "B op V". 219 I.setOperand(0, B); 220 I.setOperand(1, V); 221 // Conservatively clear the optional flags, since they may not be 222 // preserved by the reassociation. 223 I.clearSubclassOptionalData(); 224 Changed = true; 225 ++NumReassoc; 226 continue; 227 } 228 } 229 230 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 231 // if C1 and C2 are constants. 232 if (Op0 && Op1 && 233 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 234 isa<Constant>(Op0->getOperand(1)) && 235 isa<Constant>(Op1->getOperand(1)) && 236 Op0->hasOneUse() && Op1->hasOneUse()) { 237 Value *A = Op0->getOperand(0); 238 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 239 Value *B = Op1->getOperand(0); 240 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 241 242 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 243 Instruction *New = BinaryOperator::Create(Opcode, A, B, Op1->getName(), 244 &I); 245 Worklist.Add(New); 246 I.setOperand(0, New); 247 I.setOperand(1, Folded); 248 // Conservatively clear the optional flags, since they may not be 249 // preserved by the reassociation. 250 I.clearSubclassOptionalData(); 251 Changed = true; 252 continue; 253 } 254 } 255 256 // No further simplifications. 257 return Changed; 258 } while (1); 259} 260 261/// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to 262/// "(X LOp Y) ROp (X LOp Z)". 263static bool LeftDistributesOverRight(Instruction::BinaryOps LOp, 264 Instruction::BinaryOps ROp) { 265 switch (LOp) { 266 default: 267 return false; 268 269 case Instruction::And: 270 // And distributes over Or and Xor. 271 switch (ROp) { 272 default: 273 return false; 274 case Instruction::Or: 275 case Instruction::Xor: 276 return true; 277 } 278 279 case Instruction::Mul: 280 // Multiplication distributes over addition and subtraction. 281 switch (ROp) { 282 default: 283 return false; 284 case Instruction::Add: 285 case Instruction::Sub: 286 return true; 287 } 288 289 case Instruction::Or: 290 // Or distributes over And. 291 switch (ROp) { 292 default: 293 return false; 294 case Instruction::And: 295 return true; 296 } 297 } 298} 299 300/// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to 301/// "(X ROp Z) LOp (Y ROp Z)". 302static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, 303 Instruction::BinaryOps ROp) { 304 if (Instruction::isCommutative(ROp)) 305 return LeftDistributesOverRight(ROp, LOp); 306 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", 307 // but this requires knowing that the addition does not overflow and other 308 // such subtleties. 309 return false; 310} 311 312/// SimplifyUsingDistributiveLaws - This tries to simplify binary operations 313/// which some other binary operation distributes over either by factorizing 314/// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this 315/// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is 316/// a win). Returns the simplified value, or null if it didn't simplify. 317Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { 318 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 319 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 320 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 321 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op 322 323 // Factorization. 324 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) { 325 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize 326 // a common term. 327 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1); 328 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1); 329 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 330 331 // Does "X op' Y" always equal "Y op' X"? 332 bool InnerCommutative = Instruction::isCommutative(InnerOpcode); 333 334 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? 335 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) 336 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the 337 // commutative case, "(A op' B) op (C op' A)"? 338 if (A == C || (InnerCommutative && A == D)) { 339 if (A != C) 340 std::swap(C, D); 341 // Consider forming "A op' (B op D)". 342 // If "B op D" simplifies then it can be formed with no cost. 343 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD); 344 // If "B op D" doesn't simplify then only go on if both of the existing 345 // operations "A op' B" and "C op' D" will be zapped as no longer used. 346 if (!V && Op0->hasOneUse() && Op1->hasOneUse()) 347 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName()); 348 if (V) { 349 ++NumFactor; 350 V = Builder->CreateBinOp(InnerOpcode, A, V); 351 V->takeName(&I); 352 return V; 353 } 354 } 355 356 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? 357 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) 358 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the 359 // commutative case, "(A op' B) op (B op' D)"? 360 if (B == D || (InnerCommutative && B == C)) { 361 if (B != D) 362 std::swap(C, D); 363 // Consider forming "(A op C) op' B". 364 // If "A op C" simplifies then it can be formed with no cost. 365 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD); 366 // If "A op C" doesn't simplify then only go on if both of the existing 367 // operations "A op' B" and "C op' D" will be zapped as no longer used. 368 if (!V && Op0->hasOneUse() && Op1->hasOneUse()) 369 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName()); 370 if (V) { 371 ++NumFactor; 372 V = Builder->CreateBinOp(InnerOpcode, V, B); 373 V->takeName(&I); 374 return V; 375 } 376 } 377 } 378 379 // Expansion. 380 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { 381 // The instruction has the form "(A op' B) op C". See if expanding it out 382 // to "(A op C) op' (B op C)" results in simplifications. 383 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 384 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 385 386 // Do "A op C" and "B op C" both simplify? 387 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD)) 388 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) { 389 // They do! Return "L op' R". 390 ++NumExpand; 391 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 392 if ((L == A && R == B) || 393 (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) 394 return Op0; 395 // Otherwise return "L op' R" if it simplifies. 396 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD)) 397 return V; 398 // Otherwise, create a new instruction. 399 C = Builder->CreateBinOp(InnerOpcode, L, R); 400 C->takeName(&I); 401 return C; 402 } 403 } 404 405 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) { 406 // The instruction has the form "A op (B op' C)". See if expanding it out 407 // to "(A op B) op' (A op C)" results in simplifications. 408 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 409 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' 410 411 // Do "A op B" and "A op C" both simplify? 412 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD)) 413 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) { 414 // They do! Return "L op' R". 415 ++NumExpand; 416 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 417 if ((L == B && R == C) || 418 (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) 419 return Op1; 420 // Otherwise return "L op' R" if it simplifies. 421 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD)) 422 return V; 423 // Otherwise, create a new instruction. 424 A = Builder->CreateBinOp(InnerOpcode, L, R); 425 A->takeName(&I); 426 return A; 427 } 428 } 429 430 return 0; 431} 432 433// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 434// if the LHS is a constant zero (which is the 'negate' form). 435// 436Value *InstCombiner::dyn_castNegVal(Value *V) const { 437 if (BinaryOperator::isNeg(V)) 438 return BinaryOperator::getNegArgument(V); 439 440 // Constants can be considered to be negated values if they can be folded. 441 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 442 return ConstantExpr::getNeg(C); 443 444 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) 445 if (C->getType()->getElementType()->isIntegerTy()) 446 return ConstantExpr::getNeg(C); 447 448 return 0; 449} 450 451// dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 452// instruction if the LHS is a constant negative zero (which is the 'negate' 453// form). 454// 455Value *InstCombiner::dyn_castFNegVal(Value *V) const { 456 if (BinaryOperator::isFNeg(V)) 457 return BinaryOperator::getFNegArgument(V); 458 459 // Constants can be considered to be negated values if they can be folded. 460 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 461 return ConstantExpr::getFNeg(C); 462 463 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) 464 if (C->getType()->getElementType()->isFloatingPointTy()) 465 return ConstantExpr::getFNeg(C); 466 467 return 0; 468} 469 470static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 471 InstCombiner *IC) { 472 if (CastInst *CI = dyn_cast<CastInst>(&I)) { 473 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 474 } 475 476 // Figure out if the constant is the left or the right argument. 477 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 478 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 479 480 if (Constant *SOC = dyn_cast<Constant>(SO)) { 481 if (ConstIsRHS) 482 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 483 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 484 } 485 486 Value *Op0 = SO, *Op1 = ConstOperand; 487 if (!ConstIsRHS) 488 std::swap(Op0, Op1); 489 490 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) 491 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 492 SO->getName()+".op"); 493 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 494 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 495 SO->getName()+".cmp"); 496 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 497 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 498 SO->getName()+".cmp"); 499 llvm_unreachable("Unknown binary instruction type!"); 500} 501 502// FoldOpIntoSelect - Given an instruction with a select as one operand and a 503// constant as the other operand, try to fold the binary operator into the 504// select arguments. This also works for Cast instructions, which obviously do 505// not have a second operand. 506Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 507 // Don't modify shared select instructions 508 if (!SI->hasOneUse()) return 0; 509 Value *TV = SI->getOperand(1); 510 Value *FV = SI->getOperand(2); 511 512 if (isa<Constant>(TV) || isa<Constant>(FV)) { 513 // Bool selects with constant operands can be folded to logical ops. 514 if (SI->getType()->isIntegerTy(1)) return 0; 515 516 // If it's a bitcast involving vectors, make sure it has the same number of 517 // elements on both sides. 518 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) { 519 const VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy()); 520 const VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy()); 521 522 // Verify that either both or neither are vectors. 523 if ((SrcTy == NULL) != (DestTy == NULL)) return 0; 524 // If vectors, verify that they have the same number of elements. 525 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) 526 return 0; 527 } 528 529 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 530 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 531 532 return SelectInst::Create(SI->getCondition(), 533 SelectTrueVal, SelectFalseVal); 534 } 535 return 0; 536} 537 538 539/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 540/// has a PHI node as operand #0, see if we can fold the instruction into the 541/// PHI (which is only possible if all operands to the PHI are constants). 542/// 543Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { 544 PHINode *PN = cast<PHINode>(I.getOperand(0)); 545 unsigned NumPHIValues = PN->getNumIncomingValues(); 546 if (NumPHIValues == 0) 547 return 0; 548 549 // We normally only transform phis with a single use. However, if a PHI has 550 // multiple uses and they are all the same operation, we can fold *all* of the 551 // uses into the PHI. 552 if (!PN->hasOneUse()) { 553 // Walk the use list for the instruction, comparing them to I. 554 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); 555 UI != E; ++UI) { 556 Instruction *User = cast<Instruction>(*UI); 557 if (User != &I && !I.isIdenticalTo(User)) 558 return 0; 559 } 560 // Otherwise, we can replace *all* users with the new PHI we form. 561 } 562 563 // Check to see if all of the operands of the PHI are simple constants 564 // (constantint/constantfp/undef). If there is one non-constant value, 565 // remember the BB it is in. If there is more than one or if *it* is a PHI, 566 // bail out. We don't do arbitrary constant expressions here because moving 567 // their computation can be expensive without a cost model. 568 BasicBlock *NonConstBB = 0; 569 for (unsigned i = 0; i != NumPHIValues; ++i) { 570 Value *InVal = PN->getIncomingValue(i); 571 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal)) 572 continue; 573 574 if (isa<PHINode>(InVal)) return 0; // Itself a phi. 575 if (NonConstBB) return 0; // More than one non-const value. 576 577 NonConstBB = PN->getIncomingBlock(i); 578 579 // If the InVal is an invoke at the end of the pred block, then we can't 580 // insert a computation after it without breaking the edge. 581 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 582 if (II->getParent() == NonConstBB) 583 return 0; 584 585 // If the incoming non-constant value is in I's block, we will remove one 586 // instruction, but insert another equivalent one, leading to infinite 587 // instcombine. 588 if (NonConstBB == I.getParent()) 589 return 0; 590 } 591 592 // If there is exactly one non-constant value, we can insert a copy of the 593 // operation in that block. However, if this is a critical edge, we would be 594 // inserting the computation one some other paths (e.g. inside a loop). Only 595 // do this if the pred block is unconditionally branching into the phi block. 596 if (NonConstBB != 0) { 597 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 598 if (!BI || !BI->isUnconditional()) return 0; 599 } 600 601 // Okay, we can do the transformation: create the new PHI node. 602 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues(), ""); 603 InsertNewInstBefore(NewPN, *PN); 604 NewPN->takeName(PN); 605 606 // If we are going to have to insert a new computation, do so right before the 607 // predecessors terminator. 608 if (NonConstBB) 609 Builder->SetInsertPoint(NonConstBB->getTerminator()); 610 611 // Next, add all of the operands to the PHI. 612 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 613 // We only currently try to fold the condition of a select when it is a phi, 614 // not the true/false values. 615 Value *TrueV = SI->getTrueValue(); 616 Value *FalseV = SI->getFalseValue(); 617 BasicBlock *PhiTransBB = PN->getParent(); 618 for (unsigned i = 0; i != NumPHIValues; ++i) { 619 BasicBlock *ThisBB = PN->getIncomingBlock(i); 620 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 621 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 622 Value *InV = 0; 623 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 624 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 625 else 626 InV = Builder->CreateSelect(PN->getIncomingValue(i), 627 TrueVInPred, FalseVInPred, "phitmp"); 628 NewPN->addIncoming(InV, ThisBB); 629 } 630 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 631 Constant *C = cast<Constant>(I.getOperand(1)); 632 for (unsigned i = 0; i != NumPHIValues; ++i) { 633 Value *InV = 0; 634 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 635 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 636 else if (isa<ICmpInst>(CI)) 637 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 638 C, "phitmp"); 639 else 640 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 641 C, "phitmp"); 642 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 643 } 644 } else if (I.getNumOperands() == 2) { 645 Constant *C = cast<Constant>(I.getOperand(1)); 646 for (unsigned i = 0; i != NumPHIValues; ++i) { 647 Value *InV = 0; 648 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 649 InV = ConstantExpr::get(I.getOpcode(), InC, C); 650 else 651 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 652 PN->getIncomingValue(i), C, "phitmp"); 653 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 654 } 655 } else { 656 CastInst *CI = cast<CastInst>(&I); 657 const Type *RetTy = CI->getType(); 658 for (unsigned i = 0; i != NumPHIValues; ++i) { 659 Value *InV; 660 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 661 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 662 else 663 InV = Builder->CreateCast(CI->getOpcode(), 664 PN->getIncomingValue(i), I.getType(), "phitmp"); 665 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 666 } 667 } 668 669 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); 670 UI != E; ) { 671 Instruction *User = cast<Instruction>(*UI++); 672 if (User == &I) continue; 673 ReplaceInstUsesWith(*User, NewPN); 674 EraseInstFromFunction(*User); 675 } 676 return ReplaceInstUsesWith(I, NewPN); 677} 678 679/// FindElementAtOffset - Given a type and a constant offset, determine whether 680/// or not there is a sequence of GEP indices into the type that will land us at 681/// the specified offset. If so, fill them into NewIndices and return the 682/// resultant element type, otherwise return null. 683const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset, 684 SmallVectorImpl<Value*> &NewIndices) { 685 if (!TD) return 0; 686 if (!Ty->isSized()) return 0; 687 688 // Start with the index over the outer type. Note that the type size 689 // might be zero (even if the offset isn't zero) if the indexed type 690 // is something like [0 x {int, int}] 691 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 692 int64_t FirstIdx = 0; 693 if (int64_t TySize = TD->getTypeAllocSize(Ty)) { 694 FirstIdx = Offset/TySize; 695 Offset -= FirstIdx*TySize; 696 697 // Handle hosts where % returns negative instead of values [0..TySize). 698 if (Offset < 0) { 699 --FirstIdx; 700 Offset += TySize; 701 assert(Offset >= 0); 702 } 703 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 704 } 705 706 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 707 708 // Index into the types. If we fail, set OrigBase to null. 709 while (Offset) { 710 // Indexing into tail padding between struct/array elements. 711 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty)) 712 return 0; 713 714 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 715 const StructLayout *SL = TD->getStructLayout(STy); 716 assert(Offset < (int64_t)SL->getSizeInBytes() && 717 "Offset must stay within the indexed type"); 718 719 unsigned Elt = SL->getElementContainingOffset(Offset); 720 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 721 Elt)); 722 723 Offset -= SL->getElementOffset(Elt); 724 Ty = STy->getElementType(Elt); 725 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 726 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType()); 727 assert(EltSize && "Cannot index into a zero-sized array"); 728 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 729 Offset %= EltSize; 730 Ty = AT->getElementType(); 731 } else { 732 // Otherwise, we can't index into the middle of this atomic type, bail. 733 return 0; 734 } 735 } 736 737 return Ty; 738} 739 740 741 742Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 743 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 744 745 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD)) 746 return ReplaceInstUsesWith(GEP, V); 747 748 Value *PtrOp = GEP.getOperand(0); 749 750 // Eliminate unneeded casts for indices, and replace indices which displace 751 // by multiples of a zero size type with zero. 752 if (TD) { 753 bool MadeChange = false; 754 const Type *IntPtrTy = TD->getIntPtrType(GEP.getContext()); 755 756 gep_type_iterator GTI = gep_type_begin(GEP); 757 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 758 I != E; ++I, ++GTI) { 759 // Skip indices into struct types. 760 const SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 761 if (!SeqTy) continue; 762 763 // If the element type has zero size then any index over it is equivalent 764 // to an index of zero, so replace it with zero if it is not zero already. 765 if (SeqTy->getElementType()->isSized() && 766 TD->getTypeAllocSize(SeqTy->getElementType()) == 0) 767 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 768 *I = Constant::getNullValue(IntPtrTy); 769 MadeChange = true; 770 } 771 772 if ((*I)->getType() != IntPtrTy) { 773 // If we are using a wider index than needed for this platform, shrink 774 // it to what we need. If narrower, sign-extend it to what we need. 775 // This explicit cast can make subsequent optimizations more obvious. 776 *I = Builder->CreateIntCast(*I, IntPtrTy, true); 777 MadeChange = true; 778 } 779 } 780 if (MadeChange) return &GEP; 781 } 782 783 // Combine Indices - If the source pointer to this getelementptr instruction 784 // is a getelementptr instruction, combine the indices of the two 785 // getelementptr instructions into a single instruction. 786 // 787 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 788 // Note that if our source is a gep chain itself that we wait for that 789 // chain to be resolved before we perform this transformation. This 790 // avoids us creating a TON of code in some cases. 791 // 792 if (GetElementPtrInst *SrcGEP = 793 dyn_cast<GetElementPtrInst>(Src->getOperand(0))) 794 if (SrcGEP->getNumOperands() == 2) 795 return 0; // Wait until our source is folded to completion. 796 797 SmallVector<Value*, 8> Indices; 798 799 // Find out whether the last index in the source GEP is a sequential idx. 800 bool EndsWithSequential = false; 801 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 802 I != E; ++I) 803 EndsWithSequential = !(*I)->isStructTy(); 804 805 // Can we combine the two pointer arithmetics offsets? 806 if (EndsWithSequential) { 807 // Replace: gep (gep %P, long B), long A, ... 808 // With: T = long A+B; gep %P, T, ... 809 // 810 Value *Sum; 811 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 812 Value *GO1 = GEP.getOperand(1); 813 if (SO1 == Constant::getNullValue(SO1->getType())) { 814 Sum = GO1; 815 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 816 Sum = SO1; 817 } else { 818 // If they aren't the same type, then the input hasn't been processed 819 // by the loop above yet (which canonicalizes sequential index types to 820 // intptr_t). Just avoid transforming this until the input has been 821 // normalized. 822 if (SO1->getType() != GO1->getType()) 823 return 0; 824 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 825 } 826 827 // Update the GEP in place if possible. 828 if (Src->getNumOperands() == 2) { 829 GEP.setOperand(0, Src->getOperand(0)); 830 GEP.setOperand(1, Sum); 831 return &GEP; 832 } 833 Indices.append(Src->op_begin()+1, Src->op_end()-1); 834 Indices.push_back(Sum); 835 Indices.append(GEP.op_begin()+2, GEP.op_end()); 836 } else if (isa<Constant>(*GEP.idx_begin()) && 837 cast<Constant>(*GEP.idx_begin())->isNullValue() && 838 Src->getNumOperands() != 1) { 839 // Otherwise we can do the fold if the first index of the GEP is a zero 840 Indices.append(Src->op_begin()+1, Src->op_end()); 841 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 842 } 843 844 if (!Indices.empty()) 845 return (GEP.isInBounds() && Src->isInBounds()) ? 846 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(), 847 Indices.end(), GEP.getName()) : 848 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(), 849 Indices.end(), GEP.getName()); 850 } 851 852 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 853 Value *StrippedPtr = PtrOp->stripPointerCasts(); 854 const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType()); 855 if (StrippedPtr != PtrOp && 856 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 857 858 bool HasZeroPointerIndex = false; 859 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 860 HasZeroPointerIndex = C->isZero(); 861 862 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 863 // into : GEP [10 x i8]* X, i32 0, ... 864 // 865 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 866 // into : GEP i8* X, ... 867 // 868 // This occurs when the program declares an array extern like "int X[];" 869 if (HasZeroPointerIndex) { 870 const PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 871 if (const ArrayType *CATy = 872 dyn_cast<ArrayType>(CPTy->getElementType())) { 873 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 874 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 875 // -> GEP i8* X, ... 876 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 877 GetElementPtrInst *Res = 878 GetElementPtrInst::Create(StrippedPtr, Idx.begin(), 879 Idx.end(), GEP.getName()); 880 Res->setIsInBounds(GEP.isInBounds()); 881 return Res; 882 } 883 884 if (const ArrayType *XATy = 885 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 886 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 887 if (CATy->getElementType() == XATy->getElementType()) { 888 // -> GEP [10 x i8]* X, i32 0, ... 889 // At this point, we know that the cast source type is a pointer 890 // to an array of the same type as the destination pointer 891 // array. Because the array type is never stepped over (there 892 // is a leading zero) we can fold the cast into this GEP. 893 GEP.setOperand(0, StrippedPtr); 894 return &GEP; 895 } 896 } 897 } 898 } else if (GEP.getNumOperands() == 2) { 899 // Transform things like: 900 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 901 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 902 const Type *SrcElTy = StrippedPtrTy->getElementType(); 903 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType(); 904 if (TD && SrcElTy->isArrayTy() && 905 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) == 906 TD->getTypeAllocSize(ResElTy)) { 907 Value *Idx[2]; 908 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); 909 Idx[1] = GEP.getOperand(1); 910 Value *NewGEP = GEP.isInBounds() ? 911 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) : 912 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()); 913 // V and GEP are both pointer types --> BitCast 914 return new BitCastInst(NewGEP, GEP.getType()); 915 } 916 917 // Transform things like: 918 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 919 // (where tmp = 8*tmp2) into: 920 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 921 922 if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) { 923 uint64_t ArrayEltSize = 924 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()); 925 926 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We 927 // allow either a mul, shift, or constant here. 928 Value *NewIdx = 0; 929 ConstantInt *Scale = 0; 930 if (ArrayEltSize == 1) { 931 NewIdx = GEP.getOperand(1); 932 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1); 933 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) { 934 NewIdx = ConstantInt::get(CI->getType(), 1); 935 Scale = CI; 936 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){ 937 if (Inst->getOpcode() == Instruction::Shl && 938 isa<ConstantInt>(Inst->getOperand(1))) { 939 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1)); 940 uint32_t ShAmtVal = ShAmt->getLimitedValue(64); 941 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()), 942 1ULL << ShAmtVal); 943 NewIdx = Inst->getOperand(0); 944 } else if (Inst->getOpcode() == Instruction::Mul && 945 isa<ConstantInt>(Inst->getOperand(1))) { 946 Scale = cast<ConstantInt>(Inst->getOperand(1)); 947 NewIdx = Inst->getOperand(0); 948 } 949 } 950 951 // If the index will be to exactly the right offset with the scale taken 952 // out, perform the transformation. Note, we don't know whether Scale is 953 // signed or not. We'll use unsigned version of division/modulo 954 // operation after making sure Scale doesn't have the sign bit set. 955 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL && 956 Scale->getZExtValue() % ArrayEltSize == 0) { 957 Scale = ConstantInt::get(Scale->getType(), 958 Scale->getZExtValue() / ArrayEltSize); 959 if (Scale->getZExtValue() != 1) { 960 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(), 961 false /*ZExt*/); 962 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale"); 963 } 964 965 // Insert the new GEP instruction. 966 Value *Idx[2]; 967 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); 968 Idx[1] = NewIdx; 969 Value *NewGEP = GEP.isInBounds() ? 970 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()): 971 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()); 972 // The NewGEP must be pointer typed, so must the old one -> BitCast 973 return new BitCastInst(NewGEP, GEP.getType()); 974 } 975 } 976 } 977 } 978 979 /// See if we can simplify: 980 /// X = bitcast A* to B* 981 /// Y = gep X, <...constant indices...> 982 /// into a gep of the original struct. This is important for SROA and alias 983 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 984 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 985 if (TD && 986 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices() && 987 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 988 989 // Determine how much the GEP moves the pointer. We are guaranteed to get 990 // a constant back from EmitGEPOffset. 991 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP)); 992 int64_t Offset = OffsetV->getSExtValue(); 993 994 // If this GEP instruction doesn't move the pointer, just replace the GEP 995 // with a bitcast of the real input to the dest type. 996 if (Offset == 0) { 997 // If the bitcast is of an allocation, and the allocation will be 998 // converted to match the type of the cast, don't touch this. 999 if (isa<AllocaInst>(BCI->getOperand(0)) || 1000 isMalloc(BCI->getOperand(0))) { 1001 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1002 if (Instruction *I = visitBitCast(*BCI)) { 1003 if (I != BCI) { 1004 I->takeName(BCI); 1005 BCI->getParent()->getInstList().insert(BCI, I); 1006 ReplaceInstUsesWith(*BCI, I); 1007 } 1008 return &GEP; 1009 } 1010 } 1011 return new BitCastInst(BCI->getOperand(0), GEP.getType()); 1012 } 1013 1014 // Otherwise, if the offset is non-zero, we need to find out if there is a 1015 // field at Offset in 'A's type. If so, we can pull the cast through the 1016 // GEP. 1017 SmallVector<Value*, 8> NewIndices; 1018 const Type *InTy = 1019 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType(); 1020 if (FindElementAtOffset(InTy, Offset, NewIndices)) { 1021 Value *NGEP = GEP.isInBounds() ? 1022 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(), 1023 NewIndices.end()) : 1024 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(), 1025 NewIndices.end()); 1026 1027 if (NGEP->getType() == GEP.getType()) 1028 return ReplaceInstUsesWith(GEP, NGEP); 1029 NGEP->takeName(&GEP); 1030 return new BitCastInst(NGEP, GEP.getType()); 1031 } 1032 } 1033 } 1034 1035 return 0; 1036} 1037 1038 1039 1040static bool IsOnlyNullComparedAndFreed(const Value &V) { 1041 for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end(); 1042 UI != UE; ++UI) { 1043 const User *U = *UI; 1044 if (isFreeCall(U)) 1045 continue; 1046 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(U)) 1047 if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1))) 1048 continue; 1049 return false; 1050 } 1051 return true; 1052} 1053 1054Instruction *InstCombiner::visitMalloc(Instruction &MI) { 1055 // If we have a malloc call which is only used in any amount of comparisons 1056 // to null and free calls, delete the calls and replace the comparisons with 1057 // true or false as appropriate. 1058 if (IsOnlyNullComparedAndFreed(MI)) { 1059 for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end(); 1060 UI != UE;) { 1061 // We can assume that every remaining use is a free call or an icmp eq/ne 1062 // to null, so the cast is safe. 1063 Instruction *I = cast<Instruction>(*UI); 1064 1065 // Early increment here, as we're about to get rid of the user. 1066 ++UI; 1067 1068 if (isFreeCall(I)) { 1069 EraseInstFromFunction(*cast<CallInst>(I)); 1070 continue; 1071 } 1072 // Again, the cast is safe. 1073 ICmpInst *C = cast<ICmpInst>(I); 1074 ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()), 1075 C->isFalseWhenEqual())); 1076 EraseInstFromFunction(*C); 1077 } 1078 return EraseInstFromFunction(MI); 1079 } 1080 return 0; 1081} 1082 1083 1084 1085Instruction *InstCombiner::visitFree(CallInst &FI) { 1086 Value *Op = FI.getArgOperand(0); 1087 1088 // free undef -> unreachable. 1089 if (isa<UndefValue>(Op)) { 1090 // Insert a new store to null because we cannot modify the CFG here. 1091 new StoreInst(ConstantInt::getTrue(FI.getContext()), 1092 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI); 1093 return EraseInstFromFunction(FI); 1094 } 1095 1096 // If we have 'free null' delete the instruction. This can happen in stl code 1097 // when lots of inlining happens. 1098 if (isa<ConstantPointerNull>(Op)) 1099 return EraseInstFromFunction(FI); 1100 1101 return 0; 1102} 1103 1104 1105 1106Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 1107 // Change br (not X), label True, label False to: br X, label False, True 1108 Value *X = 0; 1109 BasicBlock *TrueDest; 1110 BasicBlock *FalseDest; 1111 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 1112 !isa<Constant>(X)) { 1113 // Swap Destinations and condition... 1114 BI.setCondition(X); 1115 BI.setSuccessor(0, FalseDest); 1116 BI.setSuccessor(1, TrueDest); 1117 return &BI; 1118 } 1119 1120 // Cannonicalize fcmp_one -> fcmp_oeq 1121 FCmpInst::Predicate FPred; Value *Y; 1122 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 1123 TrueDest, FalseDest)) && 1124 BI.getCondition()->hasOneUse()) 1125 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 1126 FPred == FCmpInst::FCMP_OGE) { 1127 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 1128 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 1129 1130 // Swap Destinations and condition. 1131 BI.setSuccessor(0, FalseDest); 1132 BI.setSuccessor(1, TrueDest); 1133 Worklist.Add(Cond); 1134 return &BI; 1135 } 1136 1137 // Cannonicalize icmp_ne -> icmp_eq 1138 ICmpInst::Predicate IPred; 1139 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 1140 TrueDest, FalseDest)) && 1141 BI.getCondition()->hasOneUse()) 1142 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 1143 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 1144 IPred == ICmpInst::ICMP_SGE) { 1145 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 1146 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 1147 // Swap Destinations and condition. 1148 BI.setSuccessor(0, FalseDest); 1149 BI.setSuccessor(1, TrueDest); 1150 Worklist.Add(Cond); 1151 return &BI; 1152 } 1153 1154 return 0; 1155} 1156 1157Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 1158 Value *Cond = SI.getCondition(); 1159 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 1160 if (I->getOpcode() == Instruction::Add) 1161 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 1162 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 1163 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2) 1164 SI.setOperand(i, 1165 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)), 1166 AddRHS)); 1167 SI.setOperand(0, I->getOperand(0)); 1168 Worklist.Add(I); 1169 return &SI; 1170 } 1171 } 1172 return 0; 1173} 1174 1175Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 1176 Value *Agg = EV.getAggregateOperand(); 1177 1178 if (!EV.hasIndices()) 1179 return ReplaceInstUsesWith(EV, Agg); 1180 1181 if (Constant *C = dyn_cast<Constant>(Agg)) { 1182 if (isa<UndefValue>(C)) 1183 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType())); 1184 1185 if (isa<ConstantAggregateZero>(C)) 1186 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType())); 1187 1188 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) { 1189 // Extract the element indexed by the first index out of the constant 1190 Value *V = C->getOperand(*EV.idx_begin()); 1191 if (EV.getNumIndices() > 1) 1192 // Extract the remaining indices out of the constant indexed by the 1193 // first index 1194 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end()); 1195 else 1196 return ReplaceInstUsesWith(EV, V); 1197 } 1198 return 0; // Can't handle other constants 1199 } 1200 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 1201 // We're extracting from an insertvalue instruction, compare the indices 1202 const unsigned *exti, *exte, *insi, *inse; 1203 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 1204 exte = EV.idx_end(), inse = IV->idx_end(); 1205 exti != exte && insi != inse; 1206 ++exti, ++insi) { 1207 if (*insi != *exti) 1208 // The insert and extract both reference distinctly different elements. 1209 // This means the extract is not influenced by the insert, and we can 1210 // replace the aggregate operand of the extract with the aggregate 1211 // operand of the insert. i.e., replace 1212 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1213 // %E = extractvalue { i32, { i32 } } %I, 0 1214 // with 1215 // %E = extractvalue { i32, { i32 } } %A, 0 1216 return ExtractValueInst::Create(IV->getAggregateOperand(), 1217 EV.idx_begin(), EV.idx_end()); 1218 } 1219 if (exti == exte && insi == inse) 1220 // Both iterators are at the end: Index lists are identical. Replace 1221 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1222 // %C = extractvalue { i32, { i32 } } %B, 1, 0 1223 // with "i32 42" 1224 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 1225 if (exti == exte) { 1226 // The extract list is a prefix of the insert list. i.e. replace 1227 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1228 // %E = extractvalue { i32, { i32 } } %I, 1 1229 // with 1230 // %X = extractvalue { i32, { i32 } } %A, 1 1231 // %E = insertvalue { i32 } %X, i32 42, 0 1232 // by switching the order of the insert and extract (though the 1233 // insertvalue should be left in, since it may have other uses). 1234 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 1235 EV.idx_begin(), EV.idx_end()); 1236 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 1237 insi, inse); 1238 } 1239 if (insi == inse) 1240 // The insert list is a prefix of the extract list 1241 // We can simply remove the common indices from the extract and make it 1242 // operate on the inserted value instead of the insertvalue result. 1243 // i.e., replace 1244 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1245 // %E = extractvalue { i32, { i32 } } %I, 1, 0 1246 // with 1247 // %E extractvalue { i32 } { i32 42 }, 0 1248 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 1249 exti, exte); 1250 } 1251 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 1252 // We're extracting from an intrinsic, see if we're the only user, which 1253 // allows us to simplify multiple result intrinsics to simpler things that 1254 // just get one value. 1255 if (II->hasOneUse()) { 1256 // Check if we're grabbing the overflow bit or the result of a 'with 1257 // overflow' intrinsic. If it's the latter we can remove the intrinsic 1258 // and replace it with a traditional binary instruction. 1259 switch (II->getIntrinsicID()) { 1260 case Intrinsic::uadd_with_overflow: 1261 case Intrinsic::sadd_with_overflow: 1262 if (*EV.idx_begin() == 0) { // Normal result. 1263 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1264 II->replaceAllUsesWith(UndefValue::get(II->getType())); 1265 EraseInstFromFunction(*II); 1266 return BinaryOperator::CreateAdd(LHS, RHS); 1267 } 1268 1269 // If the normal result of the add is dead, and the RHS is a constant, 1270 // we can transform this into a range comparison. 1271 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 1272 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 1273 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 1274 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 1275 ConstantExpr::getNot(CI)); 1276 break; 1277 case Intrinsic::usub_with_overflow: 1278 case Intrinsic::ssub_with_overflow: 1279 if (*EV.idx_begin() == 0) { // Normal result. 1280 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1281 II->replaceAllUsesWith(UndefValue::get(II->getType())); 1282 EraseInstFromFunction(*II); 1283 return BinaryOperator::CreateSub(LHS, RHS); 1284 } 1285 break; 1286 case Intrinsic::umul_with_overflow: 1287 case Intrinsic::smul_with_overflow: 1288 if (*EV.idx_begin() == 0) { // Normal result. 1289 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1290 II->replaceAllUsesWith(UndefValue::get(II->getType())); 1291 EraseInstFromFunction(*II); 1292 return BinaryOperator::CreateMul(LHS, RHS); 1293 } 1294 break; 1295 default: 1296 break; 1297 } 1298 } 1299 } 1300 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 1301 // If the (non-volatile) load only has one use, we can rewrite this to a 1302 // load from a GEP. This reduces the size of the load. 1303 // FIXME: If a load is used only by extractvalue instructions then this 1304 // could be done regardless of having multiple uses. 1305 if (!L->isVolatile() && L->hasOneUse()) { 1306 // extractvalue has integer indices, getelementptr has Value*s. Convert. 1307 SmallVector<Value*, 4> Indices; 1308 // Prefix an i32 0 since we need the first element. 1309 Indices.push_back(Builder->getInt32(0)); 1310 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 1311 I != E; ++I) 1312 Indices.push_back(Builder->getInt32(*I)); 1313 1314 // We need to insert these at the location of the old load, not at that of 1315 // the extractvalue. 1316 Builder->SetInsertPoint(L->getParent(), L); 1317 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), 1318 Indices.begin(), Indices.end()); 1319 // Returning the load directly will cause the main loop to insert it in 1320 // the wrong spot, so use ReplaceInstUsesWith(). 1321 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 1322 } 1323 // We could simplify extracts from other values. Note that nested extracts may 1324 // already be simplified implicitly by the above: extract (extract (insert) ) 1325 // will be translated into extract ( insert ( extract ) ) first and then just 1326 // the value inserted, if appropriate. Similarly for extracts from single-use 1327 // loads: extract (extract (load)) will be translated to extract (load (gep)) 1328 // and if again single-use then via load (gep (gep)) to load (gep). 1329 // However, double extracts from e.g. function arguments or return values 1330 // aren't handled yet. 1331 return 0; 1332} 1333 1334 1335 1336 1337/// TryToSinkInstruction - Try to move the specified instruction from its 1338/// current block into the beginning of DestBlock, which can only happen if it's 1339/// safe to move the instruction past all of the instructions between it and the 1340/// end of its block. 1341static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 1342 assert(I->hasOneUse() && "Invariants didn't hold!"); 1343 1344 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1345 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I)) 1346 return false; 1347 1348 // Do not sink alloca instructions out of the entry block. 1349 if (isa<AllocaInst>(I) && I->getParent() == 1350 &DestBlock->getParent()->getEntryBlock()) 1351 return false; 1352 1353 // We can only sink load instructions if there is nothing between the load and 1354 // the end of block that could change the value. 1355 if (I->mayReadFromMemory()) { 1356 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 1357 Scan != E; ++Scan) 1358 if (Scan->mayWriteToMemory()) 1359 return false; 1360 } 1361 1362 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI(); 1363 1364 I->moveBefore(InsertPos); 1365 ++NumSunkInst; 1366 return true; 1367} 1368 1369 1370/// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 1371/// all reachable code to the worklist. 1372/// 1373/// This has a couple of tricks to make the code faster and more powerful. In 1374/// particular, we constant fold and DCE instructions as we go, to avoid adding 1375/// them to the worklist (this significantly speeds up instcombine on code where 1376/// many instructions are dead or constant). Additionally, if we find a branch 1377/// whose condition is a known constant, we only visit the reachable successors. 1378/// 1379static bool AddReachableCodeToWorklist(BasicBlock *BB, 1380 SmallPtrSet<BasicBlock*, 64> &Visited, 1381 InstCombiner &IC, 1382 const TargetData *TD) { 1383 bool MadeIRChange = false; 1384 SmallVector<BasicBlock*, 256> Worklist; 1385 Worklist.push_back(BB); 1386 1387 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 1388 SmallPtrSet<ConstantExpr*, 64> FoldedConstants; 1389 1390 do { 1391 BB = Worklist.pop_back_val(); 1392 1393 // We have now visited this block! If we've already been here, ignore it. 1394 if (!Visited.insert(BB)) continue; 1395 1396 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 1397 Instruction *Inst = BBI++; 1398 1399 // DCE instruction if trivially dead. 1400 if (isInstructionTriviallyDead(Inst)) { 1401 ++NumDeadInst; 1402 DEBUG(errs() << "IC: DCE: " << *Inst << '\n'); 1403 Inst->eraseFromParent(); 1404 continue; 1405 } 1406 1407 // ConstantProp instruction if trivially constant. 1408 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 1409 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1410 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " 1411 << *Inst << '\n'); 1412 Inst->replaceAllUsesWith(C); 1413 ++NumConstProp; 1414 Inst->eraseFromParent(); 1415 continue; 1416 } 1417 1418 if (TD) { 1419 // See if we can constant fold its operands. 1420 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 1421 i != e; ++i) { 1422 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 1423 if (CE == 0) continue; 1424 1425 // If we already folded this constant, don't try again. 1426 if (!FoldedConstants.insert(CE)) 1427 continue; 1428 1429 Constant *NewC = ConstantFoldConstantExpression(CE, TD); 1430 if (NewC && NewC != CE) { 1431 *i = NewC; 1432 MadeIRChange = true; 1433 } 1434 } 1435 } 1436 1437 InstrsForInstCombineWorklist.push_back(Inst); 1438 } 1439 1440 // Recursively visit successors. If this is a branch or switch on a 1441 // constant, only visit the reachable successor. 1442 TerminatorInst *TI = BB->getTerminator(); 1443 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1444 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 1445 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 1446 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 1447 Worklist.push_back(ReachableBB); 1448 continue; 1449 } 1450 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1451 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 1452 // See if this is an explicit destination. 1453 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 1454 if (SI->getCaseValue(i) == Cond) { 1455 BasicBlock *ReachableBB = SI->getSuccessor(i); 1456 Worklist.push_back(ReachableBB); 1457 continue; 1458 } 1459 1460 // Otherwise it is the default destination. 1461 Worklist.push_back(SI->getSuccessor(0)); 1462 continue; 1463 } 1464 } 1465 1466 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 1467 Worklist.push_back(TI->getSuccessor(i)); 1468 } while (!Worklist.empty()); 1469 1470 // Once we've found all of the instructions to add to instcombine's worklist, 1471 // add them in reverse order. This way instcombine will visit from the top 1472 // of the function down. This jives well with the way that it adds all uses 1473 // of instructions to the worklist after doing a transformation, thus avoiding 1474 // some N^2 behavior in pathological cases. 1475 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 1476 InstrsForInstCombineWorklist.size()); 1477 1478 return MadeIRChange; 1479} 1480 1481bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 1482 MadeIRChange = false; 1483 1484 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 1485 << F.getNameStr() << "\n"); 1486 1487 { 1488 // Do a depth-first traversal of the function, populate the worklist with 1489 // the reachable instructions. Ignore blocks that are not reachable. Keep 1490 // track of which blocks we visit. 1491 SmallPtrSet<BasicBlock*, 64> Visited; 1492 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD); 1493 1494 // Do a quick scan over the function. If we find any blocks that are 1495 // unreachable, remove any instructions inside of them. This prevents 1496 // the instcombine code from having to deal with some bad special cases. 1497 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 1498 if (!Visited.count(BB)) { 1499 Instruction *Term = BB->getTerminator(); 1500 while (Term != BB->begin()) { // Remove instrs bottom-up 1501 BasicBlock::iterator I = Term; --I; 1502 1503 DEBUG(errs() << "IC: DCE: " << *I << '\n'); 1504 // A debug intrinsic shouldn't force another iteration if we weren't 1505 // going to do one without it. 1506 if (!isa<DbgInfoIntrinsic>(I)) { 1507 ++NumDeadInst; 1508 MadeIRChange = true; 1509 } 1510 1511 // If I is not void type then replaceAllUsesWith undef. 1512 // This allows ValueHandlers and custom metadata to adjust itself. 1513 if (!I->getType()->isVoidTy()) 1514 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1515 I->eraseFromParent(); 1516 } 1517 } 1518 } 1519 1520 while (!Worklist.isEmpty()) { 1521 Instruction *I = Worklist.RemoveOne(); 1522 if (I == 0) continue; // skip null values. 1523 1524 // Check to see if we can DCE the instruction. 1525 if (isInstructionTriviallyDead(I)) { 1526 DEBUG(errs() << "IC: DCE: " << *I << '\n'); 1527 EraseInstFromFunction(*I); 1528 ++NumDeadInst; 1529 MadeIRChange = true; 1530 continue; 1531 } 1532 1533 // Instruction isn't dead, see if we can constant propagate it. 1534 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 1535 if (Constant *C = ConstantFoldInstruction(I, TD)) { 1536 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 1537 1538 // Add operands to the worklist. 1539 ReplaceInstUsesWith(*I, C); 1540 ++NumConstProp; 1541 EraseInstFromFunction(*I); 1542 MadeIRChange = true; 1543 continue; 1544 } 1545 1546 // See if we can trivially sink this instruction to a successor basic block. 1547 if (I->hasOneUse()) { 1548 BasicBlock *BB = I->getParent(); 1549 Instruction *UserInst = cast<Instruction>(I->use_back()); 1550 BasicBlock *UserParent; 1551 1552 // Get the block the use occurs in. 1553 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 1554 UserParent = PN->getIncomingBlock(I->use_begin().getUse()); 1555 else 1556 UserParent = UserInst->getParent(); 1557 1558 if (UserParent != BB) { 1559 bool UserIsSuccessor = false; 1560 // See if the user is one of our successors. 1561 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 1562 if (*SI == UserParent) { 1563 UserIsSuccessor = true; 1564 break; 1565 } 1566 1567 // If the user is one of our immediate successors, and if that successor 1568 // only has us as a predecessors (we'd have to split the critical edge 1569 // otherwise), we can keep going. 1570 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 1571 // Okay, the CFG is simple enough, try to sink this instruction. 1572 MadeIRChange |= TryToSinkInstruction(I, UserParent); 1573 } 1574 } 1575 1576 // Now that we have an instruction, try combining it to simplify it. 1577 Builder->SetInsertPoint(I->getParent(), I); 1578 1579#ifndef NDEBUG 1580 std::string OrigI; 1581#endif 1582 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 1583 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n'); 1584 1585 if (Instruction *Result = visit(*I)) { 1586 ++NumCombined; 1587 // Should we replace the old instruction with a new one? 1588 if (Result != I) { 1589 DEBUG(errs() << "IC: Old = " << *I << '\n' 1590 << " New = " << *Result << '\n'); 1591 1592 Result->setDebugLoc(I->getDebugLoc()); 1593 // Everything uses the new instruction now. 1594 I->replaceAllUsesWith(Result); 1595 1596 // Push the new instruction and any users onto the worklist. 1597 Worklist.Add(Result); 1598 Worklist.AddUsersToWorkList(*Result); 1599 1600 // Move the name to the new instruction first. 1601 Result->takeName(I); 1602 1603 // Insert the new instruction into the basic block... 1604 BasicBlock *InstParent = I->getParent(); 1605 BasicBlock::iterator InsertPos = I; 1606 1607 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert 1608 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs. 1609 ++InsertPos; 1610 1611 InstParent->getInstList().insert(InsertPos, Result); 1612 1613 EraseInstFromFunction(*I); 1614 } else { 1615#ifndef NDEBUG 1616 DEBUG(errs() << "IC: Mod = " << OrigI << '\n' 1617 << " New = " << *I << '\n'); 1618#endif 1619 1620 // If the instruction was modified, it's possible that it is now dead. 1621 // if so, remove it. 1622 if (isInstructionTriviallyDead(I)) { 1623 EraseInstFromFunction(*I); 1624 } else { 1625 Worklist.Add(I); 1626 Worklist.AddUsersToWorkList(*I); 1627 } 1628 } 1629 MadeIRChange = true; 1630 } 1631 } 1632 1633 Worklist.Zap(); 1634 return MadeIRChange; 1635} 1636 1637 1638bool InstCombiner::runOnFunction(Function &F) { 1639 TD = getAnalysisIfAvailable<TargetData>(); 1640 1641 1642 /// Builder - This is an IRBuilder that automatically inserts new 1643 /// instructions into the worklist when they are created. 1644 IRBuilder<true, TargetFolder, InstCombineIRInserter> 1645 TheBuilder(F.getContext(), TargetFolder(TD), 1646 InstCombineIRInserter(Worklist)); 1647 Builder = &TheBuilder; 1648 1649 bool EverMadeChange = false; 1650 1651 // Lower dbg.declare intrinsics otherwise their value may be clobbered 1652 // by instcombiner. 1653 EverMadeChange = LowerDbgDeclare(F); 1654 1655 // Iterate while there is work to do. 1656 unsigned Iteration = 0; 1657 while (DoOneIteration(F, Iteration++)) 1658 EverMadeChange = true; 1659 1660 Builder = 0; 1661 return EverMadeChange; 1662} 1663 1664FunctionPass *llvm::createInstructionCombiningPass() { 1665 return new InstCombiner(); 1666} 1667