InstructionCombining.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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-c/Initialization.h" 40#include "llvm/ADT/SmallPtrSet.h" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/ADT/StringSwitch.h" 43#include "llvm/Analysis/ConstantFolding.h" 44#include "llvm/Analysis/InstructionSimplify.h" 45#include "llvm/Analysis/MemoryBuiltins.h" 46#include "llvm/IR/CFG.h" 47#include "llvm/IR/DataLayout.h" 48#include "llvm/IR/GetElementPtrTypeIterator.h" 49#include "llvm/IR/IntrinsicInst.h" 50#include "llvm/IR/PatternMatch.h" 51#include "llvm/IR/ValueHandle.h" 52#include "llvm/Support/CommandLine.h" 53#include "llvm/Support/Debug.h" 54#include "llvm/Target/TargetLibraryInfo.h" 55#include "llvm/Transforms/Utils/Local.h" 56#include <algorithm> 57#include <climits> 58using namespace llvm; 59using namespace llvm::PatternMatch; 60 61STATISTIC(NumCombined , "Number of insts combined"); 62STATISTIC(NumConstProp, "Number of constant folds"); 63STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 64STATISTIC(NumSunkInst , "Number of instructions sunk"); 65STATISTIC(NumExpand, "Number of expansions"); 66STATISTIC(NumFactor , "Number of factorizations"); 67STATISTIC(NumReassoc , "Number of reassociations"); 68 69static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden, 70 cl::init(false), 71 cl::desc("Enable unsafe double to float " 72 "shrinking for math lib calls")); 73 74// Initialization Routines 75void llvm::initializeInstCombine(PassRegistry &Registry) { 76 initializeInstCombinerPass(Registry); 77} 78 79void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 80 initializeInstCombine(*unwrap(R)); 81} 82 83char InstCombiner::ID = 0; 84INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine", 85 "Combine redundant instructions", false, false) 86INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 87INITIALIZE_PASS_END(InstCombiner, "instcombine", 88 "Combine redundant instructions", false, false) 89 90void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 91 AU.setPreservesCFG(); 92 AU.addRequired<TargetLibraryInfo>(); 93} 94 95 96Value *InstCombiner::EmitGEPOffset(User *GEP) { 97 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP); 98} 99 100/// ShouldChangeType - Return true if it is desirable to convert a computation 101/// from 'From' to 'To'. We don't want to convert from a legal to an illegal 102/// type for example, or from a smaller to a larger illegal type. 103bool InstCombiner::ShouldChangeType(Type *From, Type *To) const { 104 assert(From->isIntegerTy() && To->isIntegerTy()); 105 106 // If we don't have DL, we don't know if the source/dest are legal. 107 if (!DL) return false; 108 109 unsigned FromWidth = From->getPrimitiveSizeInBits(); 110 unsigned ToWidth = To->getPrimitiveSizeInBits(); 111 bool FromLegal = DL->isLegalInteger(FromWidth); 112 bool ToLegal = DL->isLegalInteger(ToWidth); 113 114 // If this is a legal integer from type, and the result would be an illegal 115 // type, don't do the transformation. 116 if (FromLegal && !ToLegal) 117 return false; 118 119 // Otherwise, if both are illegal, do not increase the size of the result. We 120 // do allow things like i160 -> i64, but not i64 -> i160. 121 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 122 return false; 123 124 return true; 125} 126 127// Return true, if No Signed Wrap should be maintained for I. 128// The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C", 129// where both B and C should be ConstantInts, results in a constant that does 130// not overflow. This function only handles the Add and Sub opcodes. For 131// all other opcodes, the function conservatively returns false. 132static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { 133 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I); 134 if (!OBO || !OBO->hasNoSignedWrap()) { 135 return false; 136 } 137 138 // We reason about Add and Sub Only. 139 Instruction::BinaryOps Opcode = I.getOpcode(); 140 if (Opcode != Instruction::Add && 141 Opcode != Instruction::Sub) { 142 return false; 143 } 144 145 ConstantInt *CB = dyn_cast<ConstantInt>(B); 146 ConstantInt *CC = dyn_cast<ConstantInt>(C); 147 148 if (!CB || !CC) { 149 return false; 150 } 151 152 const APInt &BVal = CB->getValue(); 153 const APInt &CVal = CC->getValue(); 154 bool Overflow = false; 155 156 if (Opcode == Instruction::Add) { 157 BVal.sadd_ov(CVal, Overflow); 158 } else { 159 BVal.ssub_ov(CVal, Overflow); 160 } 161 162 return !Overflow; 163} 164 165/// Conservatively clears subclassOptionalData after a reassociation or 166/// commutation. We preserve fast-math flags when applicable as they can be 167/// preserved. 168static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { 169 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I); 170 if (!FPMO) { 171 I.clearSubclassOptionalData(); 172 return; 173 } 174 175 FastMathFlags FMF = I.getFastMathFlags(); 176 I.clearSubclassOptionalData(); 177 I.setFastMathFlags(FMF); 178} 179 180/// SimplifyAssociativeOrCommutative - This performs a few simplifications for 181/// operators which are associative or commutative: 182// 183// Commutative operators: 184// 185// 1. Order operands such that they are listed from right (least complex) to 186// left (most complex). This puts constants before unary operators before 187// binary operators. 188// 189// Associative operators: 190// 191// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 192// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 193// 194// Associative and commutative operators: 195// 196// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 197// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 198// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 199// if C1 and C2 are constants. 200// 201bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 202 Instruction::BinaryOps Opcode = I.getOpcode(); 203 bool Changed = false; 204 205 do { 206 // Order operands such that they are listed from right (least complex) to 207 // left (most complex). This puts constants before unary operators before 208 // binary operators. 209 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 210 getComplexity(I.getOperand(1))) 211 Changed = !I.swapOperands(); 212 213 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 214 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 215 216 if (I.isAssociative()) { 217 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 218 if (Op0 && Op0->getOpcode() == Opcode) { 219 Value *A = Op0->getOperand(0); 220 Value *B = Op0->getOperand(1); 221 Value *C = I.getOperand(1); 222 223 // Does "B op C" simplify? 224 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) { 225 // It simplifies to V. Form "A op V". 226 I.setOperand(0, A); 227 I.setOperand(1, V); 228 // Conservatively clear the optional flags, since they may not be 229 // preserved by the reassociation. 230 if (MaintainNoSignedWrap(I, B, C) && 231 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) { 232 // Note: this is only valid because SimplifyBinOp doesn't look at 233 // the operands to Op0. 234 I.clearSubclassOptionalData(); 235 I.setHasNoSignedWrap(true); 236 } else { 237 ClearSubclassDataAfterReassociation(I); 238 } 239 240 Changed = true; 241 ++NumReassoc; 242 continue; 243 } 244 } 245 246 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 247 if (Op1 && Op1->getOpcode() == Opcode) { 248 Value *A = I.getOperand(0); 249 Value *B = Op1->getOperand(0); 250 Value *C = Op1->getOperand(1); 251 252 // Does "A op B" simplify? 253 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) { 254 // It simplifies to V. Form "V op C". 255 I.setOperand(0, V); 256 I.setOperand(1, C); 257 // Conservatively clear the optional flags, since they may not be 258 // preserved by the reassociation. 259 ClearSubclassDataAfterReassociation(I); 260 Changed = true; 261 ++NumReassoc; 262 continue; 263 } 264 } 265 } 266 267 if (I.isAssociative() && I.isCommutative()) { 268 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 269 if (Op0 && Op0->getOpcode() == Opcode) { 270 Value *A = Op0->getOperand(0); 271 Value *B = Op0->getOperand(1); 272 Value *C = I.getOperand(1); 273 274 // Does "C op A" simplify? 275 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 276 // It simplifies to V. Form "V op B". 277 I.setOperand(0, V); 278 I.setOperand(1, B); 279 // Conservatively clear the optional flags, since they may not be 280 // preserved by the reassociation. 281 ClearSubclassDataAfterReassociation(I); 282 Changed = true; 283 ++NumReassoc; 284 continue; 285 } 286 } 287 288 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 289 if (Op1 && Op1->getOpcode() == Opcode) { 290 Value *A = I.getOperand(0); 291 Value *B = Op1->getOperand(0); 292 Value *C = Op1->getOperand(1); 293 294 // Does "C op A" simplify? 295 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 296 // It simplifies to V. Form "B op V". 297 I.setOperand(0, B); 298 I.setOperand(1, V); 299 // Conservatively clear the optional flags, since they may not be 300 // preserved by the reassociation. 301 ClearSubclassDataAfterReassociation(I); 302 Changed = true; 303 ++NumReassoc; 304 continue; 305 } 306 } 307 308 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 309 // if C1 and C2 are constants. 310 if (Op0 && Op1 && 311 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 312 isa<Constant>(Op0->getOperand(1)) && 313 isa<Constant>(Op1->getOperand(1)) && 314 Op0->hasOneUse() && Op1->hasOneUse()) { 315 Value *A = Op0->getOperand(0); 316 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 317 Value *B = Op1->getOperand(0); 318 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 319 320 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 321 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B); 322 if (isa<FPMathOperator>(New)) { 323 FastMathFlags Flags = I.getFastMathFlags(); 324 Flags &= Op0->getFastMathFlags(); 325 Flags &= Op1->getFastMathFlags(); 326 New->setFastMathFlags(Flags); 327 } 328 InsertNewInstWith(New, I); 329 New->takeName(Op1); 330 I.setOperand(0, New); 331 I.setOperand(1, Folded); 332 // Conservatively clear the optional flags, since they may not be 333 // preserved by the reassociation. 334 ClearSubclassDataAfterReassociation(I); 335 336 Changed = true; 337 continue; 338 } 339 } 340 341 // No further simplifications. 342 return Changed; 343 } while (1); 344} 345 346/// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to 347/// "(X LOp Y) ROp (X LOp Z)". 348static bool LeftDistributesOverRight(Instruction::BinaryOps LOp, 349 Instruction::BinaryOps ROp) { 350 switch (LOp) { 351 default: 352 return false; 353 354 case Instruction::And: 355 // And distributes over Or and Xor. 356 switch (ROp) { 357 default: 358 return false; 359 case Instruction::Or: 360 case Instruction::Xor: 361 return true; 362 } 363 364 case Instruction::Mul: 365 // Multiplication distributes over addition and subtraction. 366 switch (ROp) { 367 default: 368 return false; 369 case Instruction::Add: 370 case Instruction::Sub: 371 return true; 372 } 373 374 case Instruction::Or: 375 // Or distributes over And. 376 switch (ROp) { 377 default: 378 return false; 379 case Instruction::And: 380 return true; 381 } 382 } 383} 384 385/// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to 386/// "(X ROp Z) LOp (Y ROp Z)". 387static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, 388 Instruction::BinaryOps ROp) { 389 if (Instruction::isCommutative(ROp)) 390 return LeftDistributesOverRight(ROp, LOp); 391 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", 392 // but this requires knowing that the addition does not overflow and other 393 // such subtleties. 394 return false; 395} 396 397/// SimplifyUsingDistributiveLaws - This tries to simplify binary operations 398/// which some other binary operation distributes over either by factorizing 399/// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this 400/// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is 401/// a win). Returns the simplified value, or null if it didn't simplify. 402Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { 403 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 404 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 405 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 406 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op 407 408 // Factorization. 409 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) { 410 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize 411 // a common term. 412 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1); 413 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1); 414 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 415 416 // Does "X op' Y" always equal "Y op' X"? 417 bool InnerCommutative = Instruction::isCommutative(InnerOpcode); 418 419 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? 420 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) 421 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the 422 // commutative case, "(A op' B) op (C op' A)"? 423 if (A == C || (InnerCommutative && A == D)) { 424 if (A != C) 425 std::swap(C, D); 426 // Consider forming "A op' (B op D)". 427 // If "B op D" simplifies then it can be formed with no cost. 428 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL); 429 // If "B op D" doesn't simplify then only go on if both of the existing 430 // operations "A op' B" and "C op' D" will be zapped as no longer used. 431 if (!V && Op0->hasOneUse() && Op1->hasOneUse()) 432 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName()); 433 if (V) { 434 ++NumFactor; 435 V = Builder->CreateBinOp(InnerOpcode, A, V); 436 V->takeName(&I); 437 return V; 438 } 439 } 440 441 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? 442 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) 443 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the 444 // commutative case, "(A op' B) op (B op' D)"? 445 if (B == D || (InnerCommutative && B == C)) { 446 if (B != D) 447 std::swap(C, D); 448 // Consider forming "(A op C) op' B". 449 // If "A op C" simplifies then it can be formed with no cost. 450 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL); 451 // If "A op C" doesn't simplify then only go on if both of the existing 452 // operations "A op' B" and "C op' D" will be zapped as no longer used. 453 if (!V && Op0->hasOneUse() && Op1->hasOneUse()) 454 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName()); 455 if (V) { 456 ++NumFactor; 457 V = Builder->CreateBinOp(InnerOpcode, V, B); 458 V->takeName(&I); 459 return V; 460 } 461 } 462 } 463 464 // Expansion. 465 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { 466 // The instruction has the form "(A op' B) op C". See if expanding it out 467 // to "(A op C) op' (B op C)" results in simplifications. 468 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 469 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 470 471 // Do "A op C" and "B op C" both simplify? 472 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL)) 473 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) { 474 // They do! Return "L op' R". 475 ++NumExpand; 476 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 477 if ((L == A && R == B) || 478 (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) 479 return Op0; 480 // Otherwise return "L op' R" if it simplifies. 481 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 482 return V; 483 // Otherwise, create a new instruction. 484 C = Builder->CreateBinOp(InnerOpcode, L, R); 485 C->takeName(&I); 486 return C; 487 } 488 } 489 490 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) { 491 // The instruction has the form "A op (B op' C)". See if expanding it out 492 // to "(A op B) op' (A op C)" results in simplifications. 493 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 494 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' 495 496 // Do "A op B" and "A op C" both simplify? 497 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL)) 498 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) { 499 // They do! Return "L op' R". 500 ++NumExpand; 501 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 502 if ((L == B && R == C) || 503 (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) 504 return Op1; 505 // Otherwise return "L op' R" if it simplifies. 506 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 507 return V; 508 // Otherwise, create a new instruction. 509 A = Builder->CreateBinOp(InnerOpcode, L, R); 510 A->takeName(&I); 511 return A; 512 } 513 } 514 515 return 0; 516} 517 518// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 519// if the LHS is a constant zero (which is the 'negate' form). 520// 521Value *InstCombiner::dyn_castNegVal(Value *V) const { 522 if (BinaryOperator::isNeg(V)) 523 return BinaryOperator::getNegArgument(V); 524 525 // Constants can be considered to be negated values if they can be folded. 526 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 527 return ConstantExpr::getNeg(C); 528 529 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 530 if (C->getType()->getElementType()->isIntegerTy()) 531 return ConstantExpr::getNeg(C); 532 533 return 0; 534} 535 536// dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 537// instruction if the LHS is a constant negative zero (which is the 'negate' 538// form). 539// 540Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const { 541 if (BinaryOperator::isFNeg(V, IgnoreZeroSign)) 542 return BinaryOperator::getFNegArgument(V); 543 544 // Constants can be considered to be negated values if they can be folded. 545 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 546 return ConstantExpr::getFNeg(C); 547 548 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 549 if (C->getType()->getElementType()->isFloatingPointTy()) 550 return ConstantExpr::getFNeg(C); 551 552 return 0; 553} 554 555static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 556 InstCombiner *IC) { 557 if (CastInst *CI = dyn_cast<CastInst>(&I)) { 558 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 559 } 560 561 // Figure out if the constant is the left or the right argument. 562 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 563 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 564 565 if (Constant *SOC = dyn_cast<Constant>(SO)) { 566 if (ConstIsRHS) 567 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 568 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 569 } 570 571 Value *Op0 = SO, *Op1 = ConstOperand; 572 if (!ConstIsRHS) 573 std::swap(Op0, Op1); 574 575 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) { 576 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 577 SO->getName()+".op"); 578 Instruction *FPInst = dyn_cast<Instruction>(RI); 579 if (FPInst && isa<FPMathOperator>(FPInst)) 580 FPInst->copyFastMathFlags(BO); 581 return RI; 582 } 583 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 584 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 585 SO->getName()+".cmp"); 586 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 587 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 588 SO->getName()+".cmp"); 589 llvm_unreachable("Unknown binary instruction type!"); 590} 591 592// FoldOpIntoSelect - Given an instruction with a select as one operand and a 593// constant as the other operand, try to fold the binary operator into the 594// select arguments. This also works for Cast instructions, which obviously do 595// not have a second operand. 596Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 597 // Don't modify shared select instructions 598 if (!SI->hasOneUse()) return 0; 599 Value *TV = SI->getOperand(1); 600 Value *FV = SI->getOperand(2); 601 602 if (isa<Constant>(TV) || isa<Constant>(FV)) { 603 // Bool selects with constant operands can be folded to logical ops. 604 if (SI->getType()->isIntegerTy(1)) return 0; 605 606 // If it's a bitcast involving vectors, make sure it has the same number of 607 // elements on both sides. 608 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) { 609 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy()); 610 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy()); 611 612 // Verify that either both or neither are vectors. 613 if ((SrcTy == NULL) != (DestTy == NULL)) return 0; 614 // If vectors, verify that they have the same number of elements. 615 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) 616 return 0; 617 } 618 619 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 620 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 621 622 return SelectInst::Create(SI->getCondition(), 623 SelectTrueVal, SelectFalseVal); 624 } 625 return 0; 626} 627 628 629/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 630/// has a PHI node as operand #0, see if we can fold the instruction into the 631/// PHI (which is only possible if all operands to the PHI are constants). 632/// 633Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { 634 PHINode *PN = cast<PHINode>(I.getOperand(0)); 635 unsigned NumPHIValues = PN->getNumIncomingValues(); 636 if (NumPHIValues == 0) 637 return 0; 638 639 // We normally only transform phis with a single use. However, if a PHI has 640 // multiple uses and they are all the same operation, we can fold *all* of the 641 // uses into the PHI. 642 if (!PN->hasOneUse()) { 643 // Walk the use list for the instruction, comparing them to I. 644 for (User *U : PN->users()) { 645 Instruction *UI = cast<Instruction>(U); 646 if (UI != &I && !I.isIdenticalTo(UI)) 647 return 0; 648 } 649 // Otherwise, we can replace *all* users with the new PHI we form. 650 } 651 652 // Check to see if all of the operands of the PHI are simple constants 653 // (constantint/constantfp/undef). If there is one non-constant value, 654 // remember the BB it is in. If there is more than one or if *it* is a PHI, 655 // bail out. We don't do arbitrary constant expressions here because moving 656 // their computation can be expensive without a cost model. 657 BasicBlock *NonConstBB = 0; 658 for (unsigned i = 0; i != NumPHIValues; ++i) { 659 Value *InVal = PN->getIncomingValue(i); 660 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal)) 661 continue; 662 663 if (isa<PHINode>(InVal)) return 0; // Itself a phi. 664 if (NonConstBB) return 0; // More than one non-const value. 665 666 NonConstBB = PN->getIncomingBlock(i); 667 668 // If the InVal is an invoke at the end of the pred block, then we can't 669 // insert a computation after it without breaking the edge. 670 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 671 if (II->getParent() == NonConstBB) 672 return 0; 673 674 // If the incoming non-constant value is in I's block, we will remove one 675 // instruction, but insert another equivalent one, leading to infinite 676 // instcombine. 677 if (NonConstBB == I.getParent()) 678 return 0; 679 } 680 681 // If there is exactly one non-constant value, we can insert a copy of the 682 // operation in that block. However, if this is a critical edge, we would be 683 // inserting the computation one some other paths (e.g. inside a loop). Only 684 // do this if the pred block is unconditionally branching into the phi block. 685 if (NonConstBB != 0) { 686 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 687 if (!BI || !BI->isUnconditional()) return 0; 688 } 689 690 // Okay, we can do the transformation: create the new PHI node. 691 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues()); 692 InsertNewInstBefore(NewPN, *PN); 693 NewPN->takeName(PN); 694 695 // If we are going to have to insert a new computation, do so right before the 696 // predecessors terminator. 697 if (NonConstBB) 698 Builder->SetInsertPoint(NonConstBB->getTerminator()); 699 700 // Next, add all of the operands to the PHI. 701 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 702 // We only currently try to fold the condition of a select when it is a phi, 703 // not the true/false values. 704 Value *TrueV = SI->getTrueValue(); 705 Value *FalseV = SI->getFalseValue(); 706 BasicBlock *PhiTransBB = PN->getParent(); 707 for (unsigned i = 0; i != NumPHIValues; ++i) { 708 BasicBlock *ThisBB = PN->getIncomingBlock(i); 709 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 710 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 711 Value *InV = 0; 712 // Beware of ConstantExpr: it may eventually evaluate to getNullValue, 713 // even if currently isNullValue gives false. 714 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)); 715 if (InC && !isa<ConstantExpr>(InC)) 716 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 717 else 718 InV = Builder->CreateSelect(PN->getIncomingValue(i), 719 TrueVInPred, FalseVInPred, "phitmp"); 720 NewPN->addIncoming(InV, ThisBB); 721 } 722 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 723 Constant *C = cast<Constant>(I.getOperand(1)); 724 for (unsigned i = 0; i != NumPHIValues; ++i) { 725 Value *InV = 0; 726 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 727 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 728 else if (isa<ICmpInst>(CI)) 729 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 730 C, "phitmp"); 731 else 732 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 733 C, "phitmp"); 734 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 735 } 736 } else if (I.getNumOperands() == 2) { 737 Constant *C = cast<Constant>(I.getOperand(1)); 738 for (unsigned i = 0; i != NumPHIValues; ++i) { 739 Value *InV = 0; 740 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 741 InV = ConstantExpr::get(I.getOpcode(), InC, C); 742 else 743 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 744 PN->getIncomingValue(i), C, "phitmp"); 745 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 746 } 747 } else { 748 CastInst *CI = cast<CastInst>(&I); 749 Type *RetTy = CI->getType(); 750 for (unsigned i = 0; i != NumPHIValues; ++i) { 751 Value *InV; 752 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 753 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 754 else 755 InV = Builder->CreateCast(CI->getOpcode(), 756 PN->getIncomingValue(i), I.getType(), "phitmp"); 757 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 758 } 759 } 760 761 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { 762 Instruction *User = cast<Instruction>(*UI++); 763 if (User == &I) continue; 764 ReplaceInstUsesWith(*User, NewPN); 765 EraseInstFromFunction(*User); 766 } 767 return ReplaceInstUsesWith(I, NewPN); 768} 769 770/// FindElementAtOffset - Given a pointer type and a constant offset, determine 771/// whether or not there is a sequence of GEP indices into the pointed type that 772/// will land us at the specified offset. If so, fill them into NewIndices and 773/// return the resultant element type, otherwise return null. 774Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset, 775 SmallVectorImpl<Value*> &NewIndices) { 776 assert(PtrTy->isPtrOrPtrVectorTy()); 777 778 if (!DL) 779 return 0; 780 781 Type *Ty = PtrTy->getPointerElementType(); 782 if (!Ty->isSized()) 783 return 0; 784 785 // Start with the index over the outer type. Note that the type size 786 // might be zero (even if the offset isn't zero) if the indexed type 787 // is something like [0 x {int, int}] 788 Type *IntPtrTy = DL->getIntPtrType(PtrTy); 789 int64_t FirstIdx = 0; 790 if (int64_t TySize = DL->getTypeAllocSize(Ty)) { 791 FirstIdx = Offset/TySize; 792 Offset -= FirstIdx*TySize; 793 794 // Handle hosts where % returns negative instead of values [0..TySize). 795 if (Offset < 0) { 796 --FirstIdx; 797 Offset += TySize; 798 assert(Offset >= 0); 799 } 800 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 801 } 802 803 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 804 805 // Index into the types. If we fail, set OrigBase to null. 806 while (Offset) { 807 // Indexing into tail padding between struct/array elements. 808 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty)) 809 return 0; 810 811 if (StructType *STy = dyn_cast<StructType>(Ty)) { 812 const StructLayout *SL = DL->getStructLayout(STy); 813 assert(Offset < (int64_t)SL->getSizeInBytes() && 814 "Offset must stay within the indexed type"); 815 816 unsigned Elt = SL->getElementContainingOffset(Offset); 817 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 818 Elt)); 819 820 Offset -= SL->getElementOffset(Elt); 821 Ty = STy->getElementType(Elt); 822 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 823 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType()); 824 assert(EltSize && "Cannot index into a zero-sized array"); 825 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 826 Offset %= EltSize; 827 Ty = AT->getElementType(); 828 } else { 829 // Otherwise, we can't index into the middle of this atomic type, bail. 830 return 0; 831 } 832 } 833 834 return Ty; 835} 836 837static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { 838 // If this GEP has only 0 indices, it is the same pointer as 839 // Src. If Src is not a trivial GEP too, don't combine 840 // the indices. 841 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() && 842 !Src.hasOneUse()) 843 return false; 844 return true; 845} 846 847/// Descale - Return a value X such that Val = X * Scale, or null if none. If 848/// the multiplication is known not to overflow then NoSignedWrap is set. 849Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { 850 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); 851 assert(cast<IntegerType>(Val->getType())->getBitWidth() == 852 Scale.getBitWidth() && "Scale not compatible with value!"); 853 854 // If Val is zero or Scale is one then Val = Val * Scale. 855 if (match(Val, m_Zero()) || Scale == 1) { 856 NoSignedWrap = true; 857 return Val; 858 } 859 860 // If Scale is zero then it does not divide Val. 861 if (Scale.isMinValue()) 862 return 0; 863 864 // Look through chains of multiplications, searching for a constant that is 865 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 866 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by 867 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore 868 // down from Val: 869 // 870 // Val = M1 * X || Analysis starts here and works down 871 // M1 = M2 * Y || Doesn't descend into terms with more 872 // M2 = Z * 4 \/ than one use 873 // 874 // Then to modify a term at the bottom: 875 // 876 // Val = M1 * X 877 // M1 = Z * Y || Replaced M2 with Z 878 // 879 // Then to work back up correcting nsw flags. 880 881 // Op - the term we are currently analyzing. Starts at Val then drills down. 882 // Replaced with its descaled value before exiting from the drill down loop. 883 Value *Op = Val; 884 885 // Parent - initially null, but after drilling down notes where Op came from. 886 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the 887 // 0'th operand of Val. 888 std::pair<Instruction*, unsigned> Parent; 889 890 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper 891 // levels that doesn't overflow. 892 bool RequireNoSignedWrap = false; 893 894 // logScale - log base 2 of the scale. Negative if not a power of 2. 895 int32_t logScale = Scale.exactLogBase2(); 896 897 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down 898 899 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 900 // If Op is a constant divisible by Scale then descale to the quotient. 901 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. 902 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); 903 if (!Remainder.isMinValue()) 904 // Not divisible by Scale. 905 return 0; 906 // Replace with the quotient in the parent. 907 Op = ConstantInt::get(CI->getType(), Quotient); 908 NoSignedWrap = true; 909 break; 910 } 911 912 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { 913 914 if (BO->getOpcode() == Instruction::Mul) { 915 // Multiplication. 916 NoSignedWrap = BO->hasNoSignedWrap(); 917 if (RequireNoSignedWrap && !NoSignedWrap) 918 return 0; 919 920 // There are three cases for multiplication: multiplication by exactly 921 // the scale, multiplication by a constant different to the scale, and 922 // multiplication by something else. 923 Value *LHS = BO->getOperand(0); 924 Value *RHS = BO->getOperand(1); 925 926 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 927 // Multiplication by a constant. 928 if (CI->getValue() == Scale) { 929 // Multiplication by exactly the scale, replace the multiplication 930 // by its left-hand side in the parent. 931 Op = LHS; 932 break; 933 } 934 935 // Otherwise drill down into the constant. 936 if (!Op->hasOneUse()) 937 return 0; 938 939 Parent = std::make_pair(BO, 1); 940 continue; 941 } 942 943 // Multiplication by something else. Drill down into the left-hand side 944 // since that's where the reassociate pass puts the good stuff. 945 if (!Op->hasOneUse()) 946 return 0; 947 948 Parent = std::make_pair(BO, 0); 949 continue; 950 } 951 952 if (logScale > 0 && BO->getOpcode() == Instruction::Shl && 953 isa<ConstantInt>(BO->getOperand(1))) { 954 // Multiplication by a power of 2. 955 NoSignedWrap = BO->hasNoSignedWrap(); 956 if (RequireNoSignedWrap && !NoSignedWrap) 957 return 0; 958 959 Value *LHS = BO->getOperand(0); 960 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> 961 getLimitedValue(Scale.getBitWidth()); 962 // Op = LHS << Amt. 963 964 if (Amt == logScale) { 965 // Multiplication by exactly the scale, replace the multiplication 966 // by its left-hand side in the parent. 967 Op = LHS; 968 break; 969 } 970 if (Amt < logScale || !Op->hasOneUse()) 971 return 0; 972 973 // Multiplication by more than the scale. Reduce the multiplying amount 974 // by the scale in the parent. 975 Parent = std::make_pair(BO, 1); 976 Op = ConstantInt::get(BO->getType(), Amt - logScale); 977 break; 978 } 979 } 980 981 if (!Op->hasOneUse()) 982 return 0; 983 984 if (CastInst *Cast = dyn_cast<CastInst>(Op)) { 985 if (Cast->getOpcode() == Instruction::SExt) { 986 // Op is sign-extended from a smaller type, descale in the smaller type. 987 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 988 APInt SmallScale = Scale.trunc(SmallSize); 989 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to 990 // descale Op as (sext Y) * Scale. In order to have 991 // sext (Y * SmallScale) = (sext Y) * Scale 992 // some conditions need to hold however: SmallScale must sign-extend to 993 // Scale and the multiplication Y * SmallScale should not overflow. 994 if (SmallScale.sext(Scale.getBitWidth()) != Scale) 995 // SmallScale does not sign-extend to Scale. 996 return 0; 997 assert(SmallScale.exactLogBase2() == logScale); 998 // Require that Y * SmallScale must not overflow. 999 RequireNoSignedWrap = true; 1000 1001 // Drill down through the cast. 1002 Parent = std::make_pair(Cast, 0); 1003 Scale = SmallScale; 1004 continue; 1005 } 1006 1007 if (Cast->getOpcode() == Instruction::Trunc) { 1008 // Op is truncated from a larger type, descale in the larger type. 1009 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then 1010 // trunc (Y * sext Scale) = (trunc Y) * Scale 1011 // always holds. However (trunc Y) * Scale may overflow even if 1012 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared 1013 // from this point up in the expression (see later). 1014 if (RequireNoSignedWrap) 1015 return 0; 1016 1017 // Drill down through the cast. 1018 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1019 Parent = std::make_pair(Cast, 0); 1020 Scale = Scale.sext(LargeSize); 1021 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) 1022 logScale = -1; 1023 assert(Scale.exactLogBase2() == logScale); 1024 continue; 1025 } 1026 } 1027 1028 // Unsupported expression, bail out. 1029 return 0; 1030 } 1031 1032 // We know that we can successfully descale, so from here on we can safely 1033 // modify the IR. Op holds the descaled version of the deepest term in the 1034 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known 1035 // not to overflow. 1036 1037 if (!Parent.first) 1038 // The expression only had one term. 1039 return Op; 1040 1041 // Rewrite the parent using the descaled version of its operand. 1042 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); 1043 assert(Op != Parent.first->getOperand(Parent.second) && 1044 "Descaling was a no-op?"); 1045 Parent.first->setOperand(Parent.second, Op); 1046 Worklist.Add(Parent.first); 1047 1048 // Now work back up the expression correcting nsw flags. The logic is based 1049 // on the following observation: if X * Y is known not to overflow as a signed 1050 // multiplication, and Y is replaced by a value Z with smaller absolute value, 1051 // then X * Z will not overflow as a signed multiplication either. As we work 1052 // our way up, having NoSignedWrap 'true' means that the descaled value at the 1053 // current level has strictly smaller absolute value than the original. 1054 Instruction *Ancestor = Parent.first; 1055 do { 1056 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { 1057 // If the multiplication wasn't nsw then we can't say anything about the 1058 // value of the descaled multiplication, and we have to clear nsw flags 1059 // from this point on up. 1060 bool OpNoSignedWrap = BO->hasNoSignedWrap(); 1061 NoSignedWrap &= OpNoSignedWrap; 1062 if (NoSignedWrap != OpNoSignedWrap) { 1063 BO->setHasNoSignedWrap(NoSignedWrap); 1064 Worklist.Add(Ancestor); 1065 } 1066 } else if (Ancestor->getOpcode() == Instruction::Trunc) { 1067 // The fact that the descaled input to the trunc has smaller absolute 1068 // value than the original input doesn't tell us anything useful about 1069 // the absolute values of the truncations. 1070 NoSignedWrap = false; 1071 } 1072 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && 1073 "Failed to keep proper track of nsw flags while drilling down?"); 1074 1075 if (Ancestor == Val) 1076 // Got to the top, all done! 1077 return Val; 1078 1079 // Move up one level in the expression. 1080 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); 1081 Ancestor = Ancestor->user_back(); 1082 } while (1); 1083} 1084 1085Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 1086 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 1087 1088 if (Value *V = SimplifyGEPInst(Ops, DL)) 1089 return ReplaceInstUsesWith(GEP, V); 1090 1091 Value *PtrOp = GEP.getOperand(0); 1092 1093 // Eliminate unneeded casts for indices, and replace indices which displace 1094 // by multiples of a zero size type with zero. 1095 if (DL) { 1096 bool MadeChange = false; 1097 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType()); 1098 1099 gep_type_iterator GTI = gep_type_begin(GEP); 1100 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 1101 I != E; ++I, ++GTI) { 1102 // Skip indices into struct types. 1103 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 1104 if (!SeqTy) continue; 1105 1106 // If the element type has zero size then any index over it is equivalent 1107 // to an index of zero, so replace it with zero if it is not zero already. 1108 if (SeqTy->getElementType()->isSized() && 1109 DL->getTypeAllocSize(SeqTy->getElementType()) == 0) 1110 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 1111 *I = Constant::getNullValue(IntPtrTy); 1112 MadeChange = true; 1113 } 1114 1115 Type *IndexTy = (*I)->getType(); 1116 if (IndexTy != IntPtrTy) { 1117 // If we are using a wider index than needed for this platform, shrink 1118 // it to what we need. If narrower, sign-extend it to what we need. 1119 // This explicit cast can make subsequent optimizations more obvious. 1120 *I = Builder->CreateIntCast(*I, IntPtrTy, true); 1121 MadeChange = true; 1122 } 1123 } 1124 if (MadeChange) return &GEP; 1125 } 1126 1127 // Combine Indices - If the source pointer to this getelementptr instruction 1128 // is a getelementptr instruction, combine the indices of the two 1129 // getelementptr instructions into a single instruction. 1130 // 1131 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 1132 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) 1133 return 0; 1134 1135 // Note that if our source is a gep chain itself then we wait for that 1136 // chain to be resolved before we perform this transformation. This 1137 // avoids us creating a TON of code in some cases. 1138 if (GEPOperator *SrcGEP = 1139 dyn_cast<GEPOperator>(Src->getOperand(0))) 1140 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) 1141 return 0; // Wait until our source is folded to completion. 1142 1143 SmallVector<Value*, 8> Indices; 1144 1145 // Find out whether the last index in the source GEP is a sequential idx. 1146 bool EndsWithSequential = false; 1147 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 1148 I != E; ++I) 1149 EndsWithSequential = !(*I)->isStructTy(); 1150 1151 // Can we combine the two pointer arithmetics offsets? 1152 if (EndsWithSequential) { 1153 // Replace: gep (gep %P, long B), long A, ... 1154 // With: T = long A+B; gep %P, T, ... 1155 // 1156 Value *Sum; 1157 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 1158 Value *GO1 = GEP.getOperand(1); 1159 if (SO1 == Constant::getNullValue(SO1->getType())) { 1160 Sum = GO1; 1161 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 1162 Sum = SO1; 1163 } else { 1164 // If they aren't the same type, then the input hasn't been processed 1165 // by the loop above yet (which canonicalizes sequential index types to 1166 // intptr_t). Just avoid transforming this until the input has been 1167 // normalized. 1168 if (SO1->getType() != GO1->getType()) 1169 return 0; 1170 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 1171 } 1172 1173 // Update the GEP in place if possible. 1174 if (Src->getNumOperands() == 2) { 1175 GEP.setOperand(0, Src->getOperand(0)); 1176 GEP.setOperand(1, Sum); 1177 return &GEP; 1178 } 1179 Indices.append(Src->op_begin()+1, Src->op_end()-1); 1180 Indices.push_back(Sum); 1181 Indices.append(GEP.op_begin()+2, GEP.op_end()); 1182 } else if (isa<Constant>(*GEP.idx_begin()) && 1183 cast<Constant>(*GEP.idx_begin())->isNullValue() && 1184 Src->getNumOperands() != 1) { 1185 // Otherwise we can do the fold if the first index of the GEP is a zero 1186 Indices.append(Src->op_begin()+1, Src->op_end()); 1187 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 1188 } 1189 1190 if (!Indices.empty()) 1191 return (GEP.isInBounds() && Src->isInBounds()) ? 1192 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices, 1193 GEP.getName()) : 1194 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName()); 1195 } 1196 1197 // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y)) 1198 // The GEP pattern is emitted by the SCEV expander for certain kinds of 1199 // pointer arithmetic. 1200 if (DL && GEP.getNumIndices() == 1 && 1201 match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) { 1202 unsigned AS = GEP.getPointerAddressSpace(); 1203 if (GEP.getType() == Builder->getInt8PtrTy(AS) && 1204 GEP.getOperand(1)->getType()->getScalarSizeInBits() == 1205 DL->getPointerSizeInBits(AS)) { 1206 Operator *Index = cast<Operator>(GEP.getOperand(1)); 1207 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); 1208 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); 1209 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); 1210 } 1211 } 1212 1213 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 1214 Value *StrippedPtr = PtrOp->stripPointerCasts(); 1215 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType()); 1216 1217 // We do not handle pointer-vector geps here. 1218 if (!StrippedPtrTy) 1219 return 0; 1220 1221 if (StrippedPtr != PtrOp) { 1222 bool HasZeroPointerIndex = false; 1223 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 1224 HasZeroPointerIndex = C->isZero(); 1225 1226 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 1227 // into : GEP [10 x i8]* X, i32 0, ... 1228 // 1229 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 1230 // into : GEP i8* X, ... 1231 // 1232 // This occurs when the program declares an array extern like "int X[];" 1233 if (HasZeroPointerIndex) { 1234 PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 1235 if (ArrayType *CATy = 1236 dyn_cast<ArrayType>(CPTy->getElementType())) { 1237 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 1238 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 1239 // -> GEP i8* X, ... 1240 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 1241 GetElementPtrInst *Res = 1242 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName()); 1243 Res->setIsInBounds(GEP.isInBounds()); 1244 return Res; 1245 } 1246 1247 if (ArrayType *XATy = 1248 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 1249 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 1250 if (CATy->getElementType() == XATy->getElementType()) { 1251 // -> GEP [10 x i8]* X, i32 0, ... 1252 // At this point, we know that the cast source type is a pointer 1253 // to an array of the same type as the destination pointer 1254 // array. Because the array type is never stepped over (there 1255 // is a leading zero) we can fold the cast into this GEP. 1256 GEP.setOperand(0, StrippedPtr); 1257 return &GEP; 1258 } 1259 } 1260 } 1261 } else if (GEP.getNumOperands() == 2) { 1262 // Transform things like: 1263 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 1264 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 1265 Type *SrcElTy = StrippedPtrTy->getElementType(); 1266 Type *ResElTy = PtrOp->getType()->getPointerElementType(); 1267 if (DL && SrcElTy->isArrayTy() && 1268 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) == 1269 DL->getTypeAllocSize(ResElTy)) { 1270 Type *IdxType = DL->getIntPtrType(GEP.getType()); 1271 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; 1272 Value *NewGEP = GEP.isInBounds() ? 1273 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : 1274 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); 1275 1276 // V and GEP are both pointer types --> BitCast 1277 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1278 return new BitCastInst(NewGEP, GEP.getType()); 1279 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1280 } 1281 1282 // Transform things like: 1283 // %V = mul i64 %N, 4 1284 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V 1285 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast 1286 if (DL && ResElTy->isSized() && SrcElTy->isSized()) { 1287 // Check that changing the type amounts to dividing the index by a scale 1288 // factor. 1289 uint64_t ResSize = DL->getTypeAllocSize(ResElTy); 1290 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy); 1291 if (ResSize && SrcSize % ResSize == 0) { 1292 Value *Idx = GEP.getOperand(1); 1293 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1294 uint64_t Scale = SrcSize / ResSize; 1295 1296 // Earlier transforms ensure that the index has type IntPtrType, which 1297 // considerably simplifies the logic by eliminating implicit casts. 1298 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && 1299 "Index not cast to pointer width?"); 1300 1301 bool NSW; 1302 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1303 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1304 // If the multiplication NewIdx * Scale may overflow then the new 1305 // GEP may not be "inbounds". 1306 Value *NewGEP = GEP.isInBounds() && NSW ? 1307 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) : 1308 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName()); 1309 1310 // The NewGEP must be pointer typed, so must the old one -> BitCast 1311 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1312 return new BitCastInst(NewGEP, GEP.getType()); 1313 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1314 } 1315 } 1316 } 1317 1318 // Similarly, transform things like: 1319 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 1320 // (where tmp = 8*tmp2) into: 1321 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 1322 if (DL && ResElTy->isSized() && SrcElTy->isSized() && 1323 SrcElTy->isArrayTy()) { 1324 // Check that changing to the array element type amounts to dividing the 1325 // index by a scale factor. 1326 uint64_t ResSize = DL->getTypeAllocSize(ResElTy); 1327 uint64_t ArrayEltSize 1328 = DL->getTypeAllocSize(SrcElTy->getArrayElementType()); 1329 if (ResSize && ArrayEltSize % ResSize == 0) { 1330 Value *Idx = GEP.getOperand(1); 1331 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1332 uint64_t Scale = ArrayEltSize / ResSize; 1333 1334 // Earlier transforms ensure that the index has type IntPtrType, which 1335 // considerably simplifies the logic by eliminating implicit casts. 1336 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && 1337 "Index not cast to pointer width?"); 1338 1339 bool NSW; 1340 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1341 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1342 // If the multiplication NewIdx * Scale may overflow then the new 1343 // GEP may not be "inbounds". 1344 Value *Off[2] = { 1345 Constant::getNullValue(DL->getIntPtrType(GEP.getType())), 1346 NewIdx 1347 }; 1348 1349 Value *NewGEP = GEP.isInBounds() && NSW ? 1350 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) : 1351 Builder->CreateGEP(StrippedPtr, Off, GEP.getName()); 1352 // The NewGEP must be pointer typed, so must the old one -> BitCast 1353 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1354 return new BitCastInst(NewGEP, GEP.getType()); 1355 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1356 } 1357 } 1358 } 1359 } 1360 } 1361 1362 if (!DL) 1363 return 0; 1364 1365 /// See if we can simplify: 1366 /// X = bitcast A* to B* 1367 /// Y = gep X, <...constant indices...> 1368 /// into a gep of the original struct. This is important for SROA and alias 1369 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 1370 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 1371 Value *Operand = BCI->getOperand(0); 1372 PointerType *OpType = cast<PointerType>(Operand->getType()); 1373 unsigned OffsetBits = DL->getPointerTypeSizeInBits(OpType); 1374 APInt Offset(OffsetBits, 0); 1375 if (!isa<BitCastInst>(Operand) && 1376 GEP.accumulateConstantOffset(*DL, Offset) && 1377 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 1378 1379 // If this GEP instruction doesn't move the pointer, just replace the GEP 1380 // with a bitcast of the real input to the dest type. 1381 if (!Offset) { 1382 // If the bitcast is of an allocation, and the allocation will be 1383 // converted to match the type of the cast, don't touch this. 1384 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) { 1385 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1386 if (Instruction *I = visitBitCast(*BCI)) { 1387 if (I != BCI) { 1388 I->takeName(BCI); 1389 BCI->getParent()->getInstList().insert(BCI, I); 1390 ReplaceInstUsesWith(*BCI, I); 1391 } 1392 return &GEP; 1393 } 1394 } 1395 return new BitCastInst(Operand, GEP.getType()); 1396 } 1397 1398 // Otherwise, if the offset is non-zero, we need to find out if there is a 1399 // field at Offset in 'A's type. If so, we can pull the cast through the 1400 // GEP. 1401 SmallVector<Value*, 8> NewIndices; 1402 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { 1403 Value *NGEP = GEP.isInBounds() ? 1404 Builder->CreateInBoundsGEP(Operand, NewIndices) : 1405 Builder->CreateGEP(Operand, NewIndices); 1406 1407 if (NGEP->getType() == GEP.getType()) 1408 return ReplaceInstUsesWith(GEP, NGEP); 1409 NGEP->takeName(&GEP); 1410 return new BitCastInst(NGEP, GEP.getType()); 1411 } 1412 } 1413 } 1414 1415 return 0; 1416} 1417 1418static bool 1419isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, 1420 const TargetLibraryInfo *TLI) { 1421 SmallVector<Instruction*, 4> Worklist; 1422 Worklist.push_back(AI); 1423 1424 do { 1425 Instruction *PI = Worklist.pop_back_val(); 1426 for (User *U : PI->users()) { 1427 Instruction *I = cast<Instruction>(U); 1428 switch (I->getOpcode()) { 1429 default: 1430 // Give up the moment we see something we can't handle. 1431 return false; 1432 1433 case Instruction::BitCast: 1434 case Instruction::GetElementPtr: 1435 Users.push_back(I); 1436 Worklist.push_back(I); 1437 continue; 1438 1439 case Instruction::ICmp: { 1440 ICmpInst *ICI = cast<ICmpInst>(I); 1441 // We can fold eq/ne comparisons with null to false/true, respectively. 1442 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1))) 1443 return false; 1444 Users.push_back(I); 1445 continue; 1446 } 1447 1448 case Instruction::Call: 1449 // Ignore no-op and store intrinsics. 1450 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1451 switch (II->getIntrinsicID()) { 1452 default: 1453 return false; 1454 1455 case Intrinsic::memmove: 1456 case Intrinsic::memcpy: 1457 case Intrinsic::memset: { 1458 MemIntrinsic *MI = cast<MemIntrinsic>(II); 1459 if (MI->isVolatile() || MI->getRawDest() != PI) 1460 return false; 1461 } 1462 // fall through 1463 case Intrinsic::dbg_declare: 1464 case Intrinsic::dbg_value: 1465 case Intrinsic::invariant_start: 1466 case Intrinsic::invariant_end: 1467 case Intrinsic::lifetime_start: 1468 case Intrinsic::lifetime_end: 1469 case Intrinsic::objectsize: 1470 Users.push_back(I); 1471 continue; 1472 } 1473 } 1474 1475 if (isFreeCall(I, TLI)) { 1476 Users.push_back(I); 1477 continue; 1478 } 1479 return false; 1480 1481 case Instruction::Store: { 1482 StoreInst *SI = cast<StoreInst>(I); 1483 if (SI->isVolatile() || SI->getPointerOperand() != PI) 1484 return false; 1485 Users.push_back(I); 1486 continue; 1487 } 1488 } 1489 llvm_unreachable("missing a return?"); 1490 } 1491 } while (!Worklist.empty()); 1492 return true; 1493} 1494 1495Instruction *InstCombiner::visitAllocSite(Instruction &MI) { 1496 // If we have a malloc call which is only used in any amount of comparisons 1497 // to null and free calls, delete the calls and replace the comparisons with 1498 // true or false as appropriate. 1499 SmallVector<WeakVH, 64> Users; 1500 if (isAllocSiteRemovable(&MI, Users, TLI)) { 1501 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 1502 Instruction *I = cast_or_null<Instruction>(&*Users[i]); 1503 if (!I) continue; 1504 1505 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) { 1506 ReplaceInstUsesWith(*C, 1507 ConstantInt::get(Type::getInt1Ty(C->getContext()), 1508 C->isFalseWhenEqual())); 1509 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 1510 ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); 1511 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1512 if (II->getIntrinsicID() == Intrinsic::objectsize) { 1513 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1)); 1514 uint64_t DontKnow = CI->isZero() ? -1ULL : 0; 1515 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); 1516 } 1517 } 1518 EraseInstFromFunction(*I); 1519 } 1520 1521 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) { 1522 // Replace invoke with a NOP intrinsic to maintain the original CFG 1523 Module *M = II->getParent()->getParent()->getParent(); 1524 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); 1525 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), 1526 None, "", II->getParent()); 1527 } 1528 return EraseInstFromFunction(MI); 1529 } 1530 return 0; 1531} 1532 1533/// \brief Move the call to free before a NULL test. 1534/// 1535/// Check if this free is accessed after its argument has been test 1536/// against NULL (property 0). 1537/// If yes, it is legal to move this call in its predecessor block. 1538/// 1539/// The move is performed only if the block containing the call to free 1540/// will be removed, i.e.: 1541/// 1. it has only one predecessor P, and P has two successors 1542/// 2. it contains the call and an unconditional branch 1543/// 3. its successor is the same as its predecessor's successor 1544/// 1545/// The profitability is out-of concern here and this function should 1546/// be called only if the caller knows this transformation would be 1547/// profitable (e.g., for code size). 1548static Instruction * 1549tryToMoveFreeBeforeNullTest(CallInst &FI) { 1550 Value *Op = FI.getArgOperand(0); 1551 BasicBlock *FreeInstrBB = FI.getParent(); 1552 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); 1553 1554 // Validate part of constraint #1: Only one predecessor 1555 // FIXME: We can extend the number of predecessor, but in that case, we 1556 // would duplicate the call to free in each predecessor and it may 1557 // not be profitable even for code size. 1558 if (!PredBB) 1559 return 0; 1560 1561 // Validate constraint #2: Does this block contains only the call to 1562 // free and an unconditional branch? 1563 // FIXME: We could check if we can speculate everything in the 1564 // predecessor block 1565 if (FreeInstrBB->size() != 2) 1566 return 0; 1567 BasicBlock *SuccBB; 1568 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) 1569 return 0; 1570 1571 // Validate the rest of constraint #1 by matching on the pred branch. 1572 TerminatorInst *TI = PredBB->getTerminator(); 1573 BasicBlock *TrueBB, *FalseBB; 1574 ICmpInst::Predicate Pred; 1575 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) 1576 return 0; 1577 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) 1578 return 0; 1579 1580 // Validate constraint #3: Ensure the null case just falls through. 1581 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) 1582 return 0; 1583 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && 1584 "Broken CFG: missing edge from predecessor to successor"); 1585 1586 FI.moveBefore(TI); 1587 return &FI; 1588} 1589 1590 1591Instruction *InstCombiner::visitFree(CallInst &FI) { 1592 Value *Op = FI.getArgOperand(0); 1593 1594 // free undef -> unreachable. 1595 if (isa<UndefValue>(Op)) { 1596 // Insert a new store to null because we cannot modify the CFG here. 1597 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()), 1598 UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); 1599 return EraseInstFromFunction(FI); 1600 } 1601 1602 // If we have 'free null' delete the instruction. This can happen in stl code 1603 // when lots of inlining happens. 1604 if (isa<ConstantPointerNull>(Op)) 1605 return EraseInstFromFunction(FI); 1606 1607 // If we optimize for code size, try to move the call to free before the null 1608 // test so that simplify cfg can remove the empty block and dead code 1609 // elimination the branch. I.e., helps to turn something like: 1610 // if (foo) free(foo); 1611 // into 1612 // free(foo); 1613 if (MinimizeSize) 1614 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) 1615 return I; 1616 1617 return 0; 1618} 1619 1620 1621 1622Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 1623 // Change br (not X), label True, label False to: br X, label False, True 1624 Value *X = 0; 1625 BasicBlock *TrueDest; 1626 BasicBlock *FalseDest; 1627 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 1628 !isa<Constant>(X)) { 1629 // Swap Destinations and condition... 1630 BI.setCondition(X); 1631 BI.swapSuccessors(); 1632 return &BI; 1633 } 1634 1635 // Canonicalize fcmp_one -> fcmp_oeq 1636 FCmpInst::Predicate FPred; Value *Y; 1637 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 1638 TrueDest, FalseDest)) && 1639 BI.getCondition()->hasOneUse()) 1640 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 1641 FPred == FCmpInst::FCMP_OGE) { 1642 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 1643 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 1644 1645 // Swap Destinations and condition. 1646 BI.swapSuccessors(); 1647 Worklist.Add(Cond); 1648 return &BI; 1649 } 1650 1651 // Canonicalize icmp_ne -> icmp_eq 1652 ICmpInst::Predicate IPred; 1653 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 1654 TrueDest, FalseDest)) && 1655 BI.getCondition()->hasOneUse()) 1656 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 1657 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 1658 IPred == ICmpInst::ICMP_SGE) { 1659 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 1660 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 1661 // Swap Destinations and condition. 1662 BI.swapSuccessors(); 1663 Worklist.Add(Cond); 1664 return &BI; 1665 } 1666 1667 return 0; 1668} 1669 1670Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 1671 Value *Cond = SI.getCondition(); 1672 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 1673 if (I->getOpcode() == Instruction::Add) 1674 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 1675 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 1676 // Skip the first item since that's the default case. 1677 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); 1678 i != e; ++i) { 1679 ConstantInt* CaseVal = i.getCaseValue(); 1680 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal), 1681 AddRHS); 1682 assert(isa<ConstantInt>(NewCaseVal) && 1683 "Result of expression should be constant"); 1684 i.setValue(cast<ConstantInt>(NewCaseVal)); 1685 } 1686 SI.setCondition(I->getOperand(0)); 1687 Worklist.Add(I); 1688 return &SI; 1689 } 1690 } 1691 return 0; 1692} 1693 1694Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 1695 Value *Agg = EV.getAggregateOperand(); 1696 1697 if (!EV.hasIndices()) 1698 return ReplaceInstUsesWith(EV, Agg); 1699 1700 if (Constant *C = dyn_cast<Constant>(Agg)) { 1701 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) { 1702 if (EV.getNumIndices() == 0) 1703 return ReplaceInstUsesWith(EV, C2); 1704 // Extract the remaining indices out of the constant indexed by the 1705 // first index 1706 return ExtractValueInst::Create(C2, EV.getIndices().slice(1)); 1707 } 1708 return 0; // Can't handle other constants 1709 } 1710 1711 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 1712 // We're extracting from an insertvalue instruction, compare the indices 1713 const unsigned *exti, *exte, *insi, *inse; 1714 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 1715 exte = EV.idx_end(), inse = IV->idx_end(); 1716 exti != exte && insi != inse; 1717 ++exti, ++insi) { 1718 if (*insi != *exti) 1719 // The insert and extract both reference distinctly different elements. 1720 // This means the extract is not influenced by the insert, and we can 1721 // replace the aggregate operand of the extract with the aggregate 1722 // operand of the insert. i.e., replace 1723 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1724 // %E = extractvalue { i32, { i32 } } %I, 0 1725 // with 1726 // %E = extractvalue { i32, { i32 } } %A, 0 1727 return ExtractValueInst::Create(IV->getAggregateOperand(), 1728 EV.getIndices()); 1729 } 1730 if (exti == exte && insi == inse) 1731 // Both iterators are at the end: Index lists are identical. Replace 1732 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1733 // %C = extractvalue { i32, { i32 } } %B, 1, 0 1734 // with "i32 42" 1735 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 1736 if (exti == exte) { 1737 // The extract list is a prefix of the insert list. i.e. replace 1738 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1739 // %E = extractvalue { i32, { i32 } } %I, 1 1740 // with 1741 // %X = extractvalue { i32, { i32 } } %A, 1 1742 // %E = insertvalue { i32 } %X, i32 42, 0 1743 // by switching the order of the insert and extract (though the 1744 // insertvalue should be left in, since it may have other uses). 1745 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 1746 EV.getIndices()); 1747 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 1748 makeArrayRef(insi, inse)); 1749 } 1750 if (insi == inse) 1751 // The insert list is a prefix of the extract list 1752 // We can simply remove the common indices from the extract and make it 1753 // operate on the inserted value instead of the insertvalue result. 1754 // i.e., replace 1755 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1756 // %E = extractvalue { i32, { i32 } } %I, 1, 0 1757 // with 1758 // %E extractvalue { i32 } { i32 42 }, 0 1759 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 1760 makeArrayRef(exti, exte)); 1761 } 1762 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 1763 // We're extracting from an intrinsic, see if we're the only user, which 1764 // allows us to simplify multiple result intrinsics to simpler things that 1765 // just get one value. 1766 if (II->hasOneUse()) { 1767 // Check if we're grabbing the overflow bit or the result of a 'with 1768 // overflow' intrinsic. If it's the latter we can remove the intrinsic 1769 // and replace it with a traditional binary instruction. 1770 switch (II->getIntrinsicID()) { 1771 case Intrinsic::uadd_with_overflow: 1772 case Intrinsic::sadd_with_overflow: 1773 if (*EV.idx_begin() == 0) { // Normal result. 1774 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1775 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1776 EraseInstFromFunction(*II); 1777 return BinaryOperator::CreateAdd(LHS, RHS); 1778 } 1779 1780 // If the normal result of the add is dead, and the RHS is a constant, 1781 // we can transform this into a range comparison. 1782 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 1783 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 1784 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 1785 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 1786 ConstantExpr::getNot(CI)); 1787 break; 1788 case Intrinsic::usub_with_overflow: 1789 case Intrinsic::ssub_with_overflow: 1790 if (*EV.idx_begin() == 0) { // Normal result. 1791 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1792 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1793 EraseInstFromFunction(*II); 1794 return BinaryOperator::CreateSub(LHS, RHS); 1795 } 1796 break; 1797 case Intrinsic::umul_with_overflow: 1798 case Intrinsic::smul_with_overflow: 1799 if (*EV.idx_begin() == 0) { // Normal result. 1800 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1801 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1802 EraseInstFromFunction(*II); 1803 return BinaryOperator::CreateMul(LHS, RHS); 1804 } 1805 break; 1806 default: 1807 break; 1808 } 1809 } 1810 } 1811 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 1812 // If the (non-volatile) load only has one use, we can rewrite this to a 1813 // load from a GEP. This reduces the size of the load. 1814 // FIXME: If a load is used only by extractvalue instructions then this 1815 // could be done regardless of having multiple uses. 1816 if (L->isSimple() && L->hasOneUse()) { 1817 // extractvalue has integer indices, getelementptr has Value*s. Convert. 1818 SmallVector<Value*, 4> Indices; 1819 // Prefix an i32 0 since we need the first element. 1820 Indices.push_back(Builder->getInt32(0)); 1821 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 1822 I != E; ++I) 1823 Indices.push_back(Builder->getInt32(*I)); 1824 1825 // We need to insert these at the location of the old load, not at that of 1826 // the extractvalue. 1827 Builder->SetInsertPoint(L->getParent(), L); 1828 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices); 1829 // Returning the load directly will cause the main loop to insert it in 1830 // the wrong spot, so use ReplaceInstUsesWith(). 1831 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 1832 } 1833 // We could simplify extracts from other values. Note that nested extracts may 1834 // already be simplified implicitly by the above: extract (extract (insert) ) 1835 // will be translated into extract ( insert ( extract ) ) first and then just 1836 // the value inserted, if appropriate. Similarly for extracts from single-use 1837 // loads: extract (extract (load)) will be translated to extract (load (gep)) 1838 // and if again single-use then via load (gep (gep)) to load (gep). 1839 // However, double extracts from e.g. function arguments or return values 1840 // aren't handled yet. 1841 return 0; 1842} 1843 1844enum Personality_Type { 1845 Unknown_Personality, 1846 GNU_Ada_Personality, 1847 GNU_CXX_Personality, 1848 GNU_ObjC_Personality 1849}; 1850 1851/// RecognizePersonality - See if the given exception handling personality 1852/// function is one that we understand. If so, return a description of it; 1853/// otherwise return Unknown_Personality. 1854static Personality_Type RecognizePersonality(Value *Pers) { 1855 Function *F = dyn_cast<Function>(Pers->stripPointerCasts()); 1856 if (!F) 1857 return Unknown_Personality; 1858 return StringSwitch<Personality_Type>(F->getName()) 1859 .Case("__gnat_eh_personality", GNU_Ada_Personality) 1860 .Case("__gxx_personality_v0", GNU_CXX_Personality) 1861 .Case("__objc_personality_v0", GNU_ObjC_Personality) 1862 .Default(Unknown_Personality); 1863} 1864 1865/// isCatchAll - Return 'true' if the given typeinfo will match anything. 1866static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) { 1867 switch (Personality) { 1868 case Unknown_Personality: 1869 return false; 1870 case GNU_Ada_Personality: 1871 // While __gnat_all_others_value will match any Ada exception, it doesn't 1872 // match foreign exceptions (or didn't, before gcc-4.7). 1873 return false; 1874 case GNU_CXX_Personality: 1875 case GNU_ObjC_Personality: 1876 return TypeInfo->isNullValue(); 1877 } 1878 llvm_unreachable("Unknown personality!"); 1879} 1880 1881static bool shorter_filter(const Value *LHS, const Value *RHS) { 1882 return 1883 cast<ArrayType>(LHS->getType())->getNumElements() 1884 < 1885 cast<ArrayType>(RHS->getType())->getNumElements(); 1886} 1887 1888Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { 1889 // The logic here should be correct for any real-world personality function. 1890 // However if that turns out not to be true, the offending logic can always 1891 // be conditioned on the personality function, like the catch-all logic is. 1892 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn()); 1893 1894 // Simplify the list of clauses, eg by removing repeated catch clauses 1895 // (these are often created by inlining). 1896 bool MakeNewInstruction = false; // If true, recreate using the following: 1897 SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction; 1898 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. 1899 1900 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already. 1901 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) { 1902 bool isLastClause = i + 1 == e; 1903 if (LI.isCatch(i)) { 1904 // A catch clause. 1905 Value *CatchClause = LI.getClause(i); 1906 Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts()); 1907 1908 // If we already saw this clause, there is no point in having a second 1909 // copy of it. 1910 if (AlreadyCaught.insert(TypeInfo)) { 1911 // This catch clause was not already seen. 1912 NewClauses.push_back(CatchClause); 1913 } else { 1914 // Repeated catch clause - drop the redundant copy. 1915 MakeNewInstruction = true; 1916 } 1917 1918 // If this is a catch-all then there is no point in keeping any following 1919 // clauses or marking the landingpad as having a cleanup. 1920 if (isCatchAll(Personality, TypeInfo)) { 1921 if (!isLastClause) 1922 MakeNewInstruction = true; 1923 CleanupFlag = false; 1924 break; 1925 } 1926 } else { 1927 // A filter clause. If any of the filter elements were already caught 1928 // then they can be dropped from the filter. It is tempting to try to 1929 // exploit the filter further by saying that any typeinfo that does not 1930 // occur in the filter can't be caught later (and thus can be dropped). 1931 // However this would be wrong, since typeinfos can match without being 1932 // equal (for example if one represents a C++ class, and the other some 1933 // class derived from it). 1934 assert(LI.isFilter(i) && "Unsupported landingpad clause!"); 1935 Value *FilterClause = LI.getClause(i); 1936 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType()); 1937 unsigned NumTypeInfos = FilterType->getNumElements(); 1938 1939 // An empty filter catches everything, so there is no point in keeping any 1940 // following clauses or marking the landingpad as having a cleanup. By 1941 // dealing with this case here the following code is made a bit simpler. 1942 if (!NumTypeInfos) { 1943 NewClauses.push_back(FilterClause); 1944 if (!isLastClause) 1945 MakeNewInstruction = true; 1946 CleanupFlag = false; 1947 break; 1948 } 1949 1950 bool MakeNewFilter = false; // If true, make a new filter. 1951 SmallVector<Constant *, 16> NewFilterElts; // New elements. 1952 if (isa<ConstantAggregateZero>(FilterClause)) { 1953 // Not an empty filter - it contains at least one null typeinfo. 1954 assert(NumTypeInfos > 0 && "Should have handled empty filter already!"); 1955 Constant *TypeInfo = 1956 Constant::getNullValue(FilterType->getElementType()); 1957 // If this typeinfo is a catch-all then the filter can never match. 1958 if (isCatchAll(Personality, TypeInfo)) { 1959 // Throw the filter away. 1960 MakeNewInstruction = true; 1961 continue; 1962 } 1963 1964 // There is no point in having multiple copies of this typeinfo, so 1965 // discard all but the first copy if there is more than one. 1966 NewFilterElts.push_back(TypeInfo); 1967 if (NumTypeInfos > 1) 1968 MakeNewFilter = true; 1969 } else { 1970 ConstantArray *Filter = cast<ConstantArray>(FilterClause); 1971 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements. 1972 NewFilterElts.reserve(NumTypeInfos); 1973 1974 // Remove any filter elements that were already caught or that already 1975 // occurred in the filter. While there, see if any of the elements are 1976 // catch-alls. If so, the filter can be discarded. 1977 bool SawCatchAll = false; 1978 for (unsigned j = 0; j != NumTypeInfos; ++j) { 1979 Value *Elt = Filter->getOperand(j); 1980 Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts()); 1981 if (isCatchAll(Personality, TypeInfo)) { 1982 // This element is a catch-all. Bail out, noting this fact. 1983 SawCatchAll = true; 1984 break; 1985 } 1986 if (AlreadyCaught.count(TypeInfo)) 1987 // Already caught by an earlier clause, so having it in the filter 1988 // is pointless. 1989 continue; 1990 // There is no point in having multiple copies of the same typeinfo in 1991 // a filter, so only add it if we didn't already. 1992 if (SeenInFilter.insert(TypeInfo)) 1993 NewFilterElts.push_back(cast<Constant>(Elt)); 1994 } 1995 // A filter containing a catch-all cannot match anything by definition. 1996 if (SawCatchAll) { 1997 // Throw the filter away. 1998 MakeNewInstruction = true; 1999 continue; 2000 } 2001 2002 // If we dropped something from the filter, make a new one. 2003 if (NewFilterElts.size() < NumTypeInfos) 2004 MakeNewFilter = true; 2005 } 2006 if (MakeNewFilter) { 2007 FilterType = ArrayType::get(FilterType->getElementType(), 2008 NewFilterElts.size()); 2009 FilterClause = ConstantArray::get(FilterType, NewFilterElts); 2010 MakeNewInstruction = true; 2011 } 2012 2013 NewClauses.push_back(FilterClause); 2014 2015 // If the new filter is empty then it will catch everything so there is 2016 // no point in keeping any following clauses or marking the landingpad 2017 // as having a cleanup. The case of the original filter being empty was 2018 // already handled above. 2019 if (MakeNewFilter && !NewFilterElts.size()) { 2020 assert(MakeNewInstruction && "New filter but not a new instruction!"); 2021 CleanupFlag = false; 2022 break; 2023 } 2024 } 2025 } 2026 2027 // If several filters occur in a row then reorder them so that the shortest 2028 // filters come first (those with the smallest number of elements). This is 2029 // advantageous because shorter filters are more likely to match, speeding up 2030 // unwinding, but mostly because it increases the effectiveness of the other 2031 // filter optimizations below. 2032 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) { 2033 unsigned j; 2034 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters. 2035 for (j = i; j != e; ++j) 2036 if (!isa<ArrayType>(NewClauses[j]->getType())) 2037 break; 2038 2039 // Check whether the filters are already sorted by length. We need to know 2040 // if sorting them is actually going to do anything so that we only make a 2041 // new landingpad instruction if it does. 2042 for (unsigned k = i; k + 1 < j; ++k) 2043 if (shorter_filter(NewClauses[k+1], NewClauses[k])) { 2044 // Not sorted, so sort the filters now. Doing an unstable sort would be 2045 // correct too but reordering filters pointlessly might confuse users. 2046 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j, 2047 shorter_filter); 2048 MakeNewInstruction = true; 2049 break; 2050 } 2051 2052 // Look for the next batch of filters. 2053 i = j + 1; 2054 } 2055 2056 // If typeinfos matched if and only if equal, then the elements of a filter L 2057 // that occurs later than a filter F could be replaced by the intersection of 2058 // the elements of F and L. In reality two typeinfos can match without being 2059 // equal (for example if one represents a C++ class, and the other some class 2060 // derived from it) so it would be wrong to perform this transform in general. 2061 // However the transform is correct and useful if F is a subset of L. In that 2062 // case L can be replaced by F, and thus removed altogether since repeating a 2063 // filter is pointless. So here we look at all pairs of filters F and L where 2064 // L follows F in the list of clauses, and remove L if every element of F is 2065 // an element of L. This can occur when inlining C++ functions with exception 2066 // specifications. 2067 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) { 2068 // Examine each filter in turn. 2069 Value *Filter = NewClauses[i]; 2070 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType()); 2071 if (!FTy) 2072 // Not a filter - skip it. 2073 continue; 2074 unsigned FElts = FTy->getNumElements(); 2075 // Examine each filter following this one. Doing this backwards means that 2076 // we don't have to worry about filters disappearing under us when removed. 2077 for (unsigned j = NewClauses.size() - 1; j != i; --j) { 2078 Value *LFilter = NewClauses[j]; 2079 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType()); 2080 if (!LTy) 2081 // Not a filter - skip it. 2082 continue; 2083 // If Filter is a subset of LFilter, i.e. every element of Filter is also 2084 // an element of LFilter, then discard LFilter. 2085 SmallVectorImpl<Value *>::iterator J = NewClauses.begin() + j; 2086 // If Filter is empty then it is a subset of LFilter. 2087 if (!FElts) { 2088 // Discard LFilter. 2089 NewClauses.erase(J); 2090 MakeNewInstruction = true; 2091 // Move on to the next filter. 2092 continue; 2093 } 2094 unsigned LElts = LTy->getNumElements(); 2095 // If Filter is longer than LFilter then it cannot be a subset of it. 2096 if (FElts > LElts) 2097 // Move on to the next filter. 2098 continue; 2099 // At this point we know that LFilter has at least one element. 2100 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros. 2101 // Filter is a subset of LFilter iff Filter contains only zeros (as we 2102 // already know that Filter is not longer than LFilter). 2103 if (isa<ConstantAggregateZero>(Filter)) { 2104 assert(FElts <= LElts && "Should have handled this case earlier!"); 2105 // Discard LFilter. 2106 NewClauses.erase(J); 2107 MakeNewInstruction = true; 2108 } 2109 // Move on to the next filter. 2110 continue; 2111 } 2112 ConstantArray *LArray = cast<ConstantArray>(LFilter); 2113 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros. 2114 // Since Filter is non-empty and contains only zeros, it is a subset of 2115 // LFilter iff LFilter contains a zero. 2116 assert(FElts > 0 && "Should have eliminated the empty filter earlier!"); 2117 for (unsigned l = 0; l != LElts; ++l) 2118 if (LArray->getOperand(l)->isNullValue()) { 2119 // LFilter contains a zero - discard it. 2120 NewClauses.erase(J); 2121 MakeNewInstruction = true; 2122 break; 2123 } 2124 // Move on to the next filter. 2125 continue; 2126 } 2127 // At this point we know that both filters are ConstantArrays. Loop over 2128 // operands to see whether every element of Filter is also an element of 2129 // LFilter. Since filters tend to be short this is probably faster than 2130 // using a method that scales nicely. 2131 ConstantArray *FArray = cast<ConstantArray>(Filter); 2132 bool AllFound = true; 2133 for (unsigned f = 0; f != FElts; ++f) { 2134 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts(); 2135 AllFound = false; 2136 for (unsigned l = 0; l != LElts; ++l) { 2137 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts(); 2138 if (LTypeInfo == FTypeInfo) { 2139 AllFound = true; 2140 break; 2141 } 2142 } 2143 if (!AllFound) 2144 break; 2145 } 2146 if (AllFound) { 2147 // Discard LFilter. 2148 NewClauses.erase(J); 2149 MakeNewInstruction = true; 2150 } 2151 // Move on to the next filter. 2152 } 2153 } 2154 2155 // If we changed any of the clauses, replace the old landingpad instruction 2156 // with a new one. 2157 if (MakeNewInstruction) { 2158 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(), 2159 LI.getPersonalityFn(), 2160 NewClauses.size()); 2161 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i) 2162 NLI->addClause(NewClauses[i]); 2163 // A landing pad with no clauses must have the cleanup flag set. It is 2164 // theoretically possible, though highly unlikely, that we eliminated all 2165 // clauses. If so, force the cleanup flag to true. 2166 if (NewClauses.empty()) 2167 CleanupFlag = true; 2168 NLI->setCleanup(CleanupFlag); 2169 return NLI; 2170 } 2171 2172 // Even if none of the clauses changed, we may nonetheless have understood 2173 // that the cleanup flag is pointless. Clear it if so. 2174 if (LI.isCleanup() != CleanupFlag) { 2175 assert(!CleanupFlag && "Adding a cleanup, not removing one?!"); 2176 LI.setCleanup(CleanupFlag); 2177 return &LI; 2178 } 2179 2180 return 0; 2181} 2182 2183 2184 2185 2186/// TryToSinkInstruction - Try to move the specified instruction from its 2187/// current block into the beginning of DestBlock, which can only happen if it's 2188/// safe to move the instruction past all of the instructions between it and the 2189/// end of its block. 2190static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 2191 assert(I->hasOneUse() && "Invariants didn't hold!"); 2192 2193 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 2194 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() || 2195 isa<TerminatorInst>(I)) 2196 return false; 2197 2198 // Do not sink alloca instructions out of the entry block. 2199 if (isa<AllocaInst>(I) && I->getParent() == 2200 &DestBlock->getParent()->getEntryBlock()) 2201 return false; 2202 2203 // We can only sink load instructions if there is nothing between the load and 2204 // the end of block that could change the value. 2205 if (I->mayReadFromMemory()) { 2206 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 2207 Scan != E; ++Scan) 2208 if (Scan->mayWriteToMemory()) 2209 return false; 2210 } 2211 2212 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt(); 2213 I->moveBefore(InsertPos); 2214 ++NumSunkInst; 2215 return true; 2216} 2217 2218 2219/// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 2220/// all reachable code to the worklist. 2221/// 2222/// This has a couple of tricks to make the code faster and more powerful. In 2223/// particular, we constant fold and DCE instructions as we go, to avoid adding 2224/// them to the worklist (this significantly speeds up instcombine on code where 2225/// many instructions are dead or constant). Additionally, if we find a branch 2226/// whose condition is a known constant, we only visit the reachable successors. 2227/// 2228static bool AddReachableCodeToWorklist(BasicBlock *BB, 2229 SmallPtrSet<BasicBlock*, 64> &Visited, 2230 InstCombiner &IC, 2231 const DataLayout *DL, 2232 const TargetLibraryInfo *TLI) { 2233 bool MadeIRChange = false; 2234 SmallVector<BasicBlock*, 256> Worklist; 2235 Worklist.push_back(BB); 2236 2237 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 2238 DenseMap<ConstantExpr*, Constant*> FoldedConstants; 2239 2240 do { 2241 BB = Worklist.pop_back_val(); 2242 2243 // We have now visited this block! If we've already been here, ignore it. 2244 if (!Visited.insert(BB)) continue; 2245 2246 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 2247 Instruction *Inst = BBI++; 2248 2249 // DCE instruction if trivially dead. 2250 if (isInstructionTriviallyDead(Inst, TLI)) { 2251 ++NumDeadInst; 2252 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); 2253 Inst->eraseFromParent(); 2254 continue; 2255 } 2256 2257 // ConstantProp instruction if trivially constant. 2258 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 2259 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) { 2260 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " 2261 << *Inst << '\n'); 2262 Inst->replaceAllUsesWith(C); 2263 ++NumConstProp; 2264 Inst->eraseFromParent(); 2265 continue; 2266 } 2267 2268 if (DL) { 2269 // See if we can constant fold its operands. 2270 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 2271 i != e; ++i) { 2272 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 2273 if (CE == 0) continue; 2274 2275 Constant*& FoldRes = FoldedConstants[CE]; 2276 if (!FoldRes) 2277 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); 2278 if (!FoldRes) 2279 FoldRes = CE; 2280 2281 if (FoldRes != CE) { 2282 *i = FoldRes; 2283 MadeIRChange = true; 2284 } 2285 } 2286 } 2287 2288 InstrsForInstCombineWorklist.push_back(Inst); 2289 } 2290 2291 // Recursively visit successors. If this is a branch or switch on a 2292 // constant, only visit the reachable successor. 2293 TerminatorInst *TI = BB->getTerminator(); 2294 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2295 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 2296 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 2297 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 2298 Worklist.push_back(ReachableBB); 2299 continue; 2300 } 2301 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2302 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 2303 // See if this is an explicit destination. 2304 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2305 i != e; ++i) 2306 if (i.getCaseValue() == Cond) { 2307 BasicBlock *ReachableBB = i.getCaseSuccessor(); 2308 Worklist.push_back(ReachableBB); 2309 continue; 2310 } 2311 2312 // Otherwise it is the default destination. 2313 Worklist.push_back(SI->getDefaultDest()); 2314 continue; 2315 } 2316 } 2317 2318 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 2319 Worklist.push_back(TI->getSuccessor(i)); 2320 } while (!Worklist.empty()); 2321 2322 // Once we've found all of the instructions to add to instcombine's worklist, 2323 // add them in reverse order. This way instcombine will visit from the top 2324 // of the function down. This jives well with the way that it adds all uses 2325 // of instructions to the worklist after doing a transformation, thus avoiding 2326 // some N^2 behavior in pathological cases. 2327 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 2328 InstrsForInstCombineWorklist.size()); 2329 2330 return MadeIRChange; 2331} 2332 2333bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 2334 MadeIRChange = false; 2335 2336 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 2337 << F.getName() << "\n"); 2338 2339 { 2340 // Do a depth-first traversal of the function, populate the worklist with 2341 // the reachable instructions. Ignore blocks that are not reachable. Keep 2342 // track of which blocks we visit. 2343 SmallPtrSet<BasicBlock*, 64> Visited; 2344 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL, 2345 TLI); 2346 2347 // Do a quick scan over the function. If we find any blocks that are 2348 // unreachable, remove any instructions inside of them. This prevents 2349 // the instcombine code from having to deal with some bad special cases. 2350 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 2351 if (Visited.count(BB)) continue; 2352 2353 // Delete the instructions backwards, as it has a reduced likelihood of 2354 // having to update as many def-use and use-def chains. 2355 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 2356 while (EndInst != BB->begin()) { 2357 // Delete the next to last instruction. 2358 BasicBlock::iterator I = EndInst; 2359 Instruction *Inst = --I; 2360 if (!Inst->use_empty()) 2361 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 2362 if (isa<LandingPadInst>(Inst)) { 2363 EndInst = Inst; 2364 continue; 2365 } 2366 if (!isa<DbgInfoIntrinsic>(Inst)) { 2367 ++NumDeadInst; 2368 MadeIRChange = true; 2369 } 2370 Inst->eraseFromParent(); 2371 } 2372 } 2373 } 2374 2375 while (!Worklist.isEmpty()) { 2376 Instruction *I = Worklist.RemoveOne(); 2377 if (I == 0) continue; // skip null values. 2378 2379 // Check to see if we can DCE the instruction. 2380 if (isInstructionTriviallyDead(I, TLI)) { 2381 DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); 2382 EraseInstFromFunction(*I); 2383 ++NumDeadInst; 2384 MadeIRChange = true; 2385 continue; 2386 } 2387 2388 // Instruction isn't dead, see if we can constant propagate it. 2389 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 2390 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) { 2391 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 2392 2393 // Add operands to the worklist. 2394 ReplaceInstUsesWith(*I, C); 2395 ++NumConstProp; 2396 EraseInstFromFunction(*I); 2397 MadeIRChange = true; 2398 continue; 2399 } 2400 2401 // See if we can trivially sink this instruction to a successor basic block. 2402 if (I->hasOneUse()) { 2403 BasicBlock *BB = I->getParent(); 2404 Instruction *UserInst = cast<Instruction>(*I->user_begin()); 2405 BasicBlock *UserParent; 2406 2407 // Get the block the use occurs in. 2408 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 2409 UserParent = PN->getIncomingBlock(*I->use_begin()); 2410 else 2411 UserParent = UserInst->getParent(); 2412 2413 if (UserParent != BB) { 2414 bool UserIsSuccessor = false; 2415 // See if the user is one of our successors. 2416 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 2417 if (*SI == UserParent) { 2418 UserIsSuccessor = true; 2419 break; 2420 } 2421 2422 // If the user is one of our immediate successors, and if that successor 2423 // only has us as a predecessors (we'd have to split the critical edge 2424 // otherwise), we can keep going. 2425 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 2426 // Okay, the CFG is simple enough, try to sink this instruction. 2427 MadeIRChange |= TryToSinkInstruction(I, UserParent); 2428 } 2429 } 2430 2431 // Now that we have an instruction, try combining it to simplify it. 2432 Builder->SetInsertPoint(I->getParent(), I); 2433 Builder->SetCurrentDebugLocation(I->getDebugLoc()); 2434 2435#ifndef NDEBUG 2436 std::string OrigI; 2437#endif 2438 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 2439 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); 2440 2441 if (Instruction *Result = visit(*I)) { 2442 ++NumCombined; 2443 // Should we replace the old instruction with a new one? 2444 if (Result != I) { 2445 DEBUG(dbgs() << "IC: Old = " << *I << '\n' 2446 << " New = " << *Result << '\n'); 2447 2448 if (!I->getDebugLoc().isUnknown()) 2449 Result->setDebugLoc(I->getDebugLoc()); 2450 // Everything uses the new instruction now. 2451 I->replaceAllUsesWith(Result); 2452 2453 // Move the name to the new instruction first. 2454 Result->takeName(I); 2455 2456 // Push the new instruction and any users onto the worklist. 2457 Worklist.Add(Result); 2458 Worklist.AddUsersToWorkList(*Result); 2459 2460 // Insert the new instruction into the basic block... 2461 BasicBlock *InstParent = I->getParent(); 2462 BasicBlock::iterator InsertPos = I; 2463 2464 // If we replace a PHI with something that isn't a PHI, fix up the 2465 // insertion point. 2466 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos)) 2467 InsertPos = InstParent->getFirstInsertionPt(); 2468 2469 InstParent->getInstList().insert(InsertPos, Result); 2470 2471 EraseInstFromFunction(*I); 2472 } else { 2473#ifndef NDEBUG 2474 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' 2475 << " New = " << *I << '\n'); 2476#endif 2477 2478 // If the instruction was modified, it's possible that it is now dead. 2479 // if so, remove it. 2480 if (isInstructionTriviallyDead(I, TLI)) { 2481 EraseInstFromFunction(*I); 2482 } else { 2483 Worklist.Add(I); 2484 Worklist.AddUsersToWorkList(*I); 2485 } 2486 } 2487 MadeIRChange = true; 2488 } 2489 } 2490 2491 Worklist.Zap(); 2492 return MadeIRChange; 2493} 2494 2495namespace { 2496class InstCombinerLibCallSimplifier : public LibCallSimplifier { 2497 InstCombiner *IC; 2498public: 2499 InstCombinerLibCallSimplifier(const DataLayout *DL, 2500 const TargetLibraryInfo *TLI, 2501 InstCombiner *IC) 2502 : LibCallSimplifier(DL, TLI, UnsafeFPShrink) { 2503 this->IC = IC; 2504 } 2505 2506 /// replaceAllUsesWith - override so that instruction replacement 2507 /// can be defined in terms of the instruction combiner framework. 2508 void replaceAllUsesWith(Instruction *I, Value *With) const override { 2509 IC->ReplaceInstUsesWith(*I, With); 2510 } 2511}; 2512} 2513 2514bool InstCombiner::runOnFunction(Function &F) { 2515 if (skipOptnoneFunction(F)) 2516 return false; 2517 2518 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 2519 DL = DLP ? &DLP->getDataLayout() : 0; 2520 TLI = &getAnalysis<TargetLibraryInfo>(); 2521 // Minimizing size? 2522 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 2523 Attribute::MinSize); 2524 2525 /// Builder - This is an IRBuilder that automatically inserts new 2526 /// instructions into the worklist when they are created. 2527 IRBuilder<true, TargetFolder, InstCombineIRInserter> 2528 TheBuilder(F.getContext(), TargetFolder(DL), 2529 InstCombineIRInserter(Worklist)); 2530 Builder = &TheBuilder; 2531 2532 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this); 2533 Simplifier = &TheSimplifier; 2534 2535 bool EverMadeChange = false; 2536 2537 // Lower dbg.declare intrinsics otherwise their value may be clobbered 2538 // by instcombiner. 2539 EverMadeChange = LowerDbgDeclare(F); 2540 2541 // Iterate while there is work to do. 2542 unsigned Iteration = 0; 2543 while (DoOneIteration(F, Iteration++)) 2544 EverMadeChange = true; 2545 2546 Builder = 0; 2547 return EverMadeChange; 2548} 2549 2550FunctionPass *llvm::createInstructionCombiningPass() { 2551 return new InstCombiner(); 2552} 2553