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 #include "llvm/Transforms/Scalar.h" 37 #include "InstCombine.h" 38 #include "llvm-c/Initialization.h" 39 #include "llvm/ADT/SmallPtrSet.h" 40 #include "llvm/ADT/Statistic.h" 41 #include "llvm/ADT/StringSwitch.h" 42 #include "llvm/Analysis/ConstantFolding.h" 43 #include "llvm/Analysis/InstructionSimplify.h" 44 #include "llvm/Analysis/MemoryBuiltins.h" 45 #include "llvm/Analysis/ValueTracking.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> 58 using namespace llvm; 59 using namespace llvm::PatternMatch; 60 61 #define DEBUG_TYPE "instcombine" 62 63 STATISTIC(NumCombined , "Number of insts combined"); 64 STATISTIC(NumConstProp, "Number of constant folds"); 65 STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 66 STATISTIC(NumSunkInst , "Number of instructions sunk"); 67 STATISTIC(NumExpand, "Number of expansions"); 68 STATISTIC(NumFactor , "Number of factorizations"); 69 STATISTIC(NumReassoc , "Number of reassociations"); 70 71 static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden, 72 cl::init(false), 73 cl::desc("Enable unsafe double to float " 74 "shrinking for math lib calls")); 75 76 // Initialization Routines 77 void llvm::initializeInstCombine(PassRegistry &Registry) { 78 initializeInstCombinerPass(Registry); 79 } 80 81 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 82 initializeInstCombine(*unwrap(R)); 83 } 84 85 char InstCombiner::ID = 0; 86 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine", 87 "Combine redundant instructions", false, false) 88 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 89 INITIALIZE_PASS_END(InstCombiner, "instcombine", 90 "Combine redundant instructions", false, false) 91 92 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 93 AU.setPreservesCFG(); 94 AU.addRequired<TargetLibraryInfo>(); 95 } 96 97 98 Value *InstCombiner::EmitGEPOffset(User *GEP) { 99 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP); 100 } 101 102 /// ShouldChangeType - Return true if it is desirable to convert a computation 103 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal 104 /// type for example, or from a smaller to a larger illegal type. 105 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const { 106 assert(From->isIntegerTy() && To->isIntegerTy()); 107 108 // If we don't have DL, we don't know if the source/dest are legal. 109 if (!DL) return false; 110 111 unsigned FromWidth = From->getPrimitiveSizeInBits(); 112 unsigned ToWidth = To->getPrimitiveSizeInBits(); 113 bool FromLegal = DL->isLegalInteger(FromWidth); 114 bool ToLegal = DL->isLegalInteger(ToWidth); 115 116 // If this is a legal integer from type, and the result would be an illegal 117 // type, don't do the transformation. 118 if (FromLegal && !ToLegal) 119 return false; 120 121 // Otherwise, if both are illegal, do not increase the size of the result. We 122 // do allow things like i160 -> i64, but not i64 -> i160. 123 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 124 return false; 125 126 return true; 127 } 128 129 // Return true, if No Signed Wrap should be maintained for I. 130 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C", 131 // where both B and C should be ConstantInts, results in a constant that does 132 // not overflow. This function only handles the Add and Sub opcodes. For 133 // all other opcodes, the function conservatively returns false. 134 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { 135 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I); 136 if (!OBO || !OBO->hasNoSignedWrap()) { 137 return false; 138 } 139 140 // We reason about Add and Sub Only. 141 Instruction::BinaryOps Opcode = I.getOpcode(); 142 if (Opcode != Instruction::Add && 143 Opcode != Instruction::Sub) { 144 return false; 145 } 146 147 ConstantInt *CB = dyn_cast<ConstantInt>(B); 148 ConstantInt *CC = dyn_cast<ConstantInt>(C); 149 150 if (!CB || !CC) { 151 return false; 152 } 153 154 const APInt &BVal = CB->getValue(); 155 const APInt &CVal = CC->getValue(); 156 bool Overflow = false; 157 158 if (Opcode == Instruction::Add) { 159 BVal.sadd_ov(CVal, Overflow); 160 } else { 161 BVal.ssub_ov(CVal, Overflow); 162 } 163 164 return !Overflow; 165 } 166 167 /// Conservatively clears subclassOptionalData after a reassociation or 168 /// commutation. We preserve fast-math flags when applicable as they can be 169 /// preserved. 170 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { 171 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I); 172 if (!FPMO) { 173 I.clearSubclassOptionalData(); 174 return; 175 } 176 177 FastMathFlags FMF = I.getFastMathFlags(); 178 I.clearSubclassOptionalData(); 179 I.setFastMathFlags(FMF); 180 } 181 182 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for 183 /// operators which are associative or commutative: 184 // 185 // Commutative operators: 186 // 187 // 1. Order operands such that they are listed from right (least complex) to 188 // left (most complex). This puts constants before unary operators before 189 // binary operators. 190 // 191 // Associative operators: 192 // 193 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 194 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 195 // 196 // Associative and commutative operators: 197 // 198 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 199 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 200 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 201 // if C1 and C2 are constants. 202 // 203 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 204 Instruction::BinaryOps Opcode = I.getOpcode(); 205 bool Changed = false; 206 207 do { 208 // Order operands such that they are listed from right (least complex) to 209 // left (most complex). This puts constants before unary operators before 210 // binary operators. 211 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 212 getComplexity(I.getOperand(1))) 213 Changed = !I.swapOperands(); 214 215 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 216 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 217 218 if (I.isAssociative()) { 219 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 220 if (Op0 && Op0->getOpcode() == Opcode) { 221 Value *A = Op0->getOperand(0); 222 Value *B = Op0->getOperand(1); 223 Value *C = I.getOperand(1); 224 225 // Does "B op C" simplify? 226 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) { 227 // It simplifies to V. Form "A op V". 228 I.setOperand(0, A); 229 I.setOperand(1, V); 230 // Conservatively clear the optional flags, since they may not be 231 // preserved by the reassociation. 232 if (MaintainNoSignedWrap(I, B, C) && 233 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) { 234 // Note: this is only valid because SimplifyBinOp doesn't look at 235 // the operands to Op0. 236 I.clearSubclassOptionalData(); 237 I.setHasNoSignedWrap(true); 238 } else { 239 ClearSubclassDataAfterReassociation(I); 240 } 241 242 Changed = true; 243 ++NumReassoc; 244 continue; 245 } 246 } 247 248 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 249 if (Op1 && Op1->getOpcode() == Opcode) { 250 Value *A = I.getOperand(0); 251 Value *B = Op1->getOperand(0); 252 Value *C = Op1->getOperand(1); 253 254 // Does "A op B" simplify? 255 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) { 256 // It simplifies to V. Form "V op C". 257 I.setOperand(0, V); 258 I.setOperand(1, C); 259 // Conservatively clear the optional flags, since they may not be 260 // preserved by the reassociation. 261 ClearSubclassDataAfterReassociation(I); 262 Changed = true; 263 ++NumReassoc; 264 continue; 265 } 266 } 267 } 268 269 if (I.isAssociative() && I.isCommutative()) { 270 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 271 if (Op0 && Op0->getOpcode() == Opcode) { 272 Value *A = Op0->getOperand(0); 273 Value *B = Op0->getOperand(1); 274 Value *C = I.getOperand(1); 275 276 // Does "C op A" simplify? 277 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 278 // It simplifies to V. Form "V op B". 279 I.setOperand(0, V); 280 I.setOperand(1, B); 281 // Conservatively clear the optional flags, since they may not be 282 // preserved by the reassociation. 283 ClearSubclassDataAfterReassociation(I); 284 Changed = true; 285 ++NumReassoc; 286 continue; 287 } 288 } 289 290 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 291 if (Op1 && Op1->getOpcode() == Opcode) { 292 Value *A = I.getOperand(0); 293 Value *B = Op1->getOperand(0); 294 Value *C = Op1->getOperand(1); 295 296 // Does "C op A" simplify? 297 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 298 // It simplifies to V. Form "B op V". 299 I.setOperand(0, B); 300 I.setOperand(1, V); 301 // Conservatively clear the optional flags, since they may not be 302 // preserved by the reassociation. 303 ClearSubclassDataAfterReassociation(I); 304 Changed = true; 305 ++NumReassoc; 306 continue; 307 } 308 } 309 310 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 311 // if C1 and C2 are constants. 312 if (Op0 && Op1 && 313 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 314 isa<Constant>(Op0->getOperand(1)) && 315 isa<Constant>(Op1->getOperand(1)) && 316 Op0->hasOneUse() && Op1->hasOneUse()) { 317 Value *A = Op0->getOperand(0); 318 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 319 Value *B = Op1->getOperand(0); 320 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 321 322 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 323 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B); 324 if (isa<FPMathOperator>(New)) { 325 FastMathFlags Flags = I.getFastMathFlags(); 326 Flags &= Op0->getFastMathFlags(); 327 Flags &= Op1->getFastMathFlags(); 328 New->setFastMathFlags(Flags); 329 } 330 InsertNewInstWith(New, I); 331 New->takeName(Op1); 332 I.setOperand(0, New); 333 I.setOperand(1, Folded); 334 // Conservatively clear the optional flags, since they may not be 335 // preserved by the reassociation. 336 ClearSubclassDataAfterReassociation(I); 337 338 Changed = true; 339 continue; 340 } 341 } 342 343 // No further simplifications. 344 return Changed; 345 } while (1); 346 } 347 348 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to 349 /// "(X LOp Y) ROp (X LOp Z)". 350 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp, 351 Instruction::BinaryOps ROp) { 352 switch (LOp) { 353 default: 354 return false; 355 356 case Instruction::And: 357 // And distributes over Or and Xor. 358 switch (ROp) { 359 default: 360 return false; 361 case Instruction::Or: 362 case Instruction::Xor: 363 return true; 364 } 365 366 case Instruction::Mul: 367 // Multiplication distributes over addition and subtraction. 368 switch (ROp) { 369 default: 370 return false; 371 case Instruction::Add: 372 case Instruction::Sub: 373 return true; 374 } 375 376 case Instruction::Or: 377 // Or distributes over And. 378 switch (ROp) { 379 default: 380 return false; 381 case Instruction::And: 382 return true; 383 } 384 } 385 } 386 387 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to 388 /// "(X ROp Z) LOp (Y ROp Z)". 389 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, 390 Instruction::BinaryOps ROp) { 391 if (Instruction::isCommutative(ROp)) 392 return LeftDistributesOverRight(ROp, LOp); 393 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", 394 // but this requires knowing that the addition does not overflow and other 395 // such subtleties. 396 return false; 397 } 398 399 /// This function returns identity value for given opcode, which can be used to 400 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1). 401 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) { 402 if (isa<Constant>(V)) 403 return nullptr; 404 405 if (OpCode == Instruction::Mul) 406 return ConstantInt::get(V->getType(), 1); 407 408 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc. 409 410 return nullptr; 411 } 412 413 /// This function factors binary ops which can be combined using distributive 414 /// laws. This also factor SHL as MUL e.g. SHL(X, 2) ==> MUL(X, 4). 415 static Instruction::BinaryOps 416 getBinOpsForFactorization(BinaryOperator *Op, Value *&LHS, Value *&RHS) { 417 if (!Op) 418 return Instruction::BinaryOpsEnd; 419 420 if (Op->getOpcode() == Instruction::Shl) { 421 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) { 422 // The multiplier is really 1 << CST. 423 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST); 424 LHS = Op->getOperand(0); 425 return Instruction::Mul; 426 } 427 } 428 429 // TODO: We can add other conversions e.g. shr => div etc. 430 431 LHS = Op->getOperand(0); 432 RHS = Op->getOperand(1); 433 return Op->getOpcode(); 434 } 435 436 /// This tries to simplify binary operations by factorizing out common terms 437 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 438 static Value *tryFactorization(InstCombiner::BuilderTy *Builder, 439 const DataLayout *DL, BinaryOperator &I, 440 Instruction::BinaryOps InnerOpcode, Value *A, 441 Value *B, Value *C, Value *D) { 442 443 // If any of A, B, C, D are null, we can not factor I, return early. 444 // Checking A and C should be enough. 445 if (!A || !C || !B || !D) 446 return nullptr; 447 448 Value *SimplifiedInst = nullptr; 449 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 450 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); 451 452 // Does "X op' Y" always equal "Y op' X"? 453 bool InnerCommutative = Instruction::isCommutative(InnerOpcode); 454 455 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? 456 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) 457 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the 458 // commutative case, "(A op' B) op (C op' A)"? 459 if (A == C || (InnerCommutative && A == D)) { 460 if (A != C) 461 std::swap(C, D); 462 // Consider forming "A op' (B op D)". 463 // If "B op D" simplifies then it can be formed with no cost. 464 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL); 465 // If "B op D" doesn't simplify then only go on if both of the existing 466 // operations "A op' B" and "C op' D" will be zapped as no longer used. 467 if (!V && LHS->hasOneUse() && RHS->hasOneUse()) 468 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName()); 469 if (V) { 470 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V); 471 } 472 } 473 474 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? 475 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) 476 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the 477 // commutative case, "(A op' B) op (B op' D)"? 478 if (B == D || (InnerCommutative && B == C)) { 479 if (B != D) 480 std::swap(C, D); 481 // Consider forming "(A op C) op' B". 482 // If "A op C" simplifies then it can be formed with no cost. 483 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL); 484 485 // If "A op C" doesn't simplify then only go on if both of the existing 486 // operations "A op' B" and "C op' D" will be zapped as no longer used. 487 if (!V && LHS->hasOneUse() && RHS->hasOneUse()) 488 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName()); 489 if (V) { 490 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B); 491 } 492 } 493 494 if (SimplifiedInst) { 495 ++NumFactor; 496 SimplifiedInst->takeName(&I); 497 498 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag. 499 // TODO: Check for NUW. 500 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) { 501 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) { 502 bool HasNSW = false; 503 if (isa<OverflowingBinaryOperator>(&I)) 504 HasNSW = I.hasNoSignedWrap(); 505 506 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 507 if (isa<OverflowingBinaryOperator>(Op0)) 508 HasNSW &= Op0->hasNoSignedWrap(); 509 510 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 511 if (isa<OverflowingBinaryOperator>(Op1)) 512 HasNSW &= Op1->hasNoSignedWrap(); 513 BO->setHasNoSignedWrap(HasNSW); 514 } 515 } 516 } 517 return SimplifiedInst; 518 } 519 520 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations 521 /// which some other binary operation distributes over either by factorizing 522 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this 523 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is 524 /// a win). Returns the simplified value, or null if it didn't simplify. 525 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { 526 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 527 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 528 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 529 530 // Factorization. 531 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 532 Instruction::BinaryOps LHSOpcode = getBinOpsForFactorization(Op0, A, B); 533 Instruction::BinaryOps RHSOpcode = getBinOpsForFactorization(Op1, C, D); 534 535 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize 536 // a common term. 537 if (LHSOpcode == RHSOpcode) { 538 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D)) 539 return V; 540 } 541 542 // The instruction has the form "(A op' B) op (C)". Try to factorize common 543 // term. 544 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS, 545 getIdentityValue(LHSOpcode, RHS))) 546 return V; 547 548 // The instruction has the form "(B) op (C op' D)". Try to factorize common 549 // term. 550 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS, 551 getIdentityValue(RHSOpcode, LHS), C, D)) 552 return V; 553 554 // Expansion. 555 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); 556 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { 557 // The instruction has the form "(A op' B) op C". See if expanding it out 558 // to "(A op C) op' (B op C)" results in simplifications. 559 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 560 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 561 562 // Do "A op C" and "B op C" both simplify? 563 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL)) 564 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) { 565 // They do! Return "L op' R". 566 ++NumExpand; 567 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 568 if ((L == A && R == B) || 569 (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) 570 return Op0; 571 // Otherwise return "L op' R" if it simplifies. 572 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 573 return V; 574 // Otherwise, create a new instruction. 575 C = Builder->CreateBinOp(InnerOpcode, L, R); 576 C->takeName(&I); 577 return C; 578 } 579 } 580 581 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) { 582 // The instruction has the form "A op (B op' C)". See if expanding it out 583 // to "(A op B) op' (A op C)" results in simplifications. 584 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 585 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' 586 587 // Do "A op B" and "A op C" both simplify? 588 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL)) 589 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) { 590 // They do! Return "L op' R". 591 ++NumExpand; 592 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 593 if ((L == B && R == C) || 594 (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) 595 return Op1; 596 // Otherwise return "L op' R" if it simplifies. 597 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 598 return V; 599 // Otherwise, create a new instruction. 600 A = Builder->CreateBinOp(InnerOpcode, L, R); 601 A->takeName(&I); 602 return A; 603 } 604 } 605 606 return nullptr; 607 } 608 609 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 610 // if the LHS is a constant zero (which is the 'negate' form). 611 // 612 Value *InstCombiner::dyn_castNegVal(Value *V) const { 613 if (BinaryOperator::isNeg(V)) 614 return BinaryOperator::getNegArgument(V); 615 616 // Constants can be considered to be negated values if they can be folded. 617 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 618 return ConstantExpr::getNeg(C); 619 620 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 621 if (C->getType()->getElementType()->isIntegerTy()) 622 return ConstantExpr::getNeg(C); 623 624 return nullptr; 625 } 626 627 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 628 // instruction if the LHS is a constant negative zero (which is the 'negate' 629 // form). 630 // 631 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const { 632 if (BinaryOperator::isFNeg(V, IgnoreZeroSign)) 633 return BinaryOperator::getFNegArgument(V); 634 635 // Constants can be considered to be negated values if they can be folded. 636 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 637 return ConstantExpr::getFNeg(C); 638 639 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 640 if (C->getType()->getElementType()->isFloatingPointTy()) 641 return ConstantExpr::getFNeg(C); 642 643 return nullptr; 644 } 645 646 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 647 InstCombiner *IC) { 648 if (CastInst *CI = dyn_cast<CastInst>(&I)) { 649 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 650 } 651 652 // Figure out if the constant is the left or the right argument. 653 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 654 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 655 656 if (Constant *SOC = dyn_cast<Constant>(SO)) { 657 if (ConstIsRHS) 658 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 659 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 660 } 661 662 Value *Op0 = SO, *Op1 = ConstOperand; 663 if (!ConstIsRHS) 664 std::swap(Op0, Op1); 665 666 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) { 667 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 668 SO->getName()+".op"); 669 Instruction *FPInst = dyn_cast<Instruction>(RI); 670 if (FPInst && isa<FPMathOperator>(FPInst)) 671 FPInst->copyFastMathFlags(BO); 672 return RI; 673 } 674 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 675 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 676 SO->getName()+".cmp"); 677 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 678 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 679 SO->getName()+".cmp"); 680 llvm_unreachable("Unknown binary instruction type!"); 681 } 682 683 // FoldOpIntoSelect - Given an instruction with a select as one operand and a 684 // constant as the other operand, try to fold the binary operator into the 685 // select arguments. This also works for Cast instructions, which obviously do 686 // not have a second operand. 687 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 688 // Don't modify shared select instructions 689 if (!SI->hasOneUse()) return nullptr; 690 Value *TV = SI->getOperand(1); 691 Value *FV = SI->getOperand(2); 692 693 if (isa<Constant>(TV) || isa<Constant>(FV)) { 694 // Bool selects with constant operands can be folded to logical ops. 695 if (SI->getType()->isIntegerTy(1)) return nullptr; 696 697 // If it's a bitcast involving vectors, make sure it has the same number of 698 // elements on both sides. 699 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) { 700 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy()); 701 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy()); 702 703 // Verify that either both or neither are vectors. 704 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr; 705 // If vectors, verify that they have the same number of elements. 706 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) 707 return nullptr; 708 } 709 710 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 711 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 712 713 return SelectInst::Create(SI->getCondition(), 714 SelectTrueVal, SelectFalseVal); 715 } 716 return nullptr; 717 } 718 719 720 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 721 /// has a PHI node as operand #0, see if we can fold the instruction into the 722 /// PHI (which is only possible if all operands to the PHI are constants). 723 /// 724 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { 725 PHINode *PN = cast<PHINode>(I.getOperand(0)); 726 unsigned NumPHIValues = PN->getNumIncomingValues(); 727 if (NumPHIValues == 0) 728 return nullptr; 729 730 // We normally only transform phis with a single use. However, if a PHI has 731 // multiple uses and they are all the same operation, we can fold *all* of the 732 // uses into the PHI. 733 if (!PN->hasOneUse()) { 734 // Walk the use list for the instruction, comparing them to I. 735 for (User *U : PN->users()) { 736 Instruction *UI = cast<Instruction>(U); 737 if (UI != &I && !I.isIdenticalTo(UI)) 738 return nullptr; 739 } 740 // Otherwise, we can replace *all* users with the new PHI we form. 741 } 742 743 // Check to see if all of the operands of the PHI are simple constants 744 // (constantint/constantfp/undef). If there is one non-constant value, 745 // remember the BB it is in. If there is more than one or if *it* is a PHI, 746 // bail out. We don't do arbitrary constant expressions here because moving 747 // their computation can be expensive without a cost model. 748 BasicBlock *NonConstBB = nullptr; 749 for (unsigned i = 0; i != NumPHIValues; ++i) { 750 Value *InVal = PN->getIncomingValue(i); 751 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal)) 752 continue; 753 754 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi. 755 if (NonConstBB) return nullptr; // More than one non-const value. 756 757 NonConstBB = PN->getIncomingBlock(i); 758 759 // If the InVal is an invoke at the end of the pred block, then we can't 760 // insert a computation after it without breaking the edge. 761 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 762 if (II->getParent() == NonConstBB) 763 return nullptr; 764 765 // If the incoming non-constant value is in I's block, we will remove one 766 // instruction, but insert another equivalent one, leading to infinite 767 // instcombine. 768 if (NonConstBB == I.getParent()) 769 return nullptr; 770 } 771 772 // If there is exactly one non-constant value, we can insert a copy of the 773 // operation in that block. However, if this is a critical edge, we would be 774 // inserting the computation one some other paths (e.g. inside a loop). Only 775 // do this if the pred block is unconditionally branching into the phi block. 776 if (NonConstBB != nullptr) { 777 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 778 if (!BI || !BI->isUnconditional()) return nullptr; 779 } 780 781 // Okay, we can do the transformation: create the new PHI node. 782 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues()); 783 InsertNewInstBefore(NewPN, *PN); 784 NewPN->takeName(PN); 785 786 // If we are going to have to insert a new computation, do so right before the 787 // predecessors terminator. 788 if (NonConstBB) 789 Builder->SetInsertPoint(NonConstBB->getTerminator()); 790 791 // Next, add all of the operands to the PHI. 792 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 793 // We only currently try to fold the condition of a select when it is a phi, 794 // not the true/false values. 795 Value *TrueV = SI->getTrueValue(); 796 Value *FalseV = SI->getFalseValue(); 797 BasicBlock *PhiTransBB = PN->getParent(); 798 for (unsigned i = 0; i != NumPHIValues; ++i) { 799 BasicBlock *ThisBB = PN->getIncomingBlock(i); 800 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 801 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 802 Value *InV = nullptr; 803 // Beware of ConstantExpr: it may eventually evaluate to getNullValue, 804 // even if currently isNullValue gives false. 805 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)); 806 if (InC && !isa<ConstantExpr>(InC)) 807 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 808 else 809 InV = Builder->CreateSelect(PN->getIncomingValue(i), 810 TrueVInPred, FalseVInPred, "phitmp"); 811 NewPN->addIncoming(InV, ThisBB); 812 } 813 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 814 Constant *C = cast<Constant>(I.getOperand(1)); 815 for (unsigned i = 0; i != NumPHIValues; ++i) { 816 Value *InV = nullptr; 817 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 818 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 819 else if (isa<ICmpInst>(CI)) 820 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 821 C, "phitmp"); 822 else 823 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 824 C, "phitmp"); 825 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 826 } 827 } else if (I.getNumOperands() == 2) { 828 Constant *C = cast<Constant>(I.getOperand(1)); 829 for (unsigned i = 0; i != NumPHIValues; ++i) { 830 Value *InV = nullptr; 831 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 832 InV = ConstantExpr::get(I.getOpcode(), InC, C); 833 else 834 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 835 PN->getIncomingValue(i), C, "phitmp"); 836 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 837 } 838 } else { 839 CastInst *CI = cast<CastInst>(&I); 840 Type *RetTy = CI->getType(); 841 for (unsigned i = 0; i != NumPHIValues; ++i) { 842 Value *InV; 843 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 844 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 845 else 846 InV = Builder->CreateCast(CI->getOpcode(), 847 PN->getIncomingValue(i), I.getType(), "phitmp"); 848 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 849 } 850 } 851 852 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { 853 Instruction *User = cast<Instruction>(*UI++); 854 if (User == &I) continue; 855 ReplaceInstUsesWith(*User, NewPN); 856 EraseInstFromFunction(*User); 857 } 858 return ReplaceInstUsesWith(I, NewPN); 859 } 860 861 /// FindElementAtOffset - Given a pointer type and a constant offset, determine 862 /// whether or not there is a sequence of GEP indices into the pointed type that 863 /// will land us at the specified offset. If so, fill them into NewIndices and 864 /// return the resultant element type, otherwise return null. 865 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset, 866 SmallVectorImpl<Value*> &NewIndices) { 867 assert(PtrTy->isPtrOrPtrVectorTy()); 868 869 if (!DL) 870 return nullptr; 871 872 Type *Ty = PtrTy->getPointerElementType(); 873 if (!Ty->isSized()) 874 return nullptr; 875 876 // Start with the index over the outer type. Note that the type size 877 // might be zero (even if the offset isn't zero) if the indexed type 878 // is something like [0 x {int, int}] 879 Type *IntPtrTy = DL->getIntPtrType(PtrTy); 880 int64_t FirstIdx = 0; 881 if (int64_t TySize = DL->getTypeAllocSize(Ty)) { 882 FirstIdx = Offset/TySize; 883 Offset -= FirstIdx*TySize; 884 885 // Handle hosts where % returns negative instead of values [0..TySize). 886 if (Offset < 0) { 887 --FirstIdx; 888 Offset += TySize; 889 assert(Offset >= 0); 890 } 891 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 892 } 893 894 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 895 896 // Index into the types. If we fail, set OrigBase to null. 897 while (Offset) { 898 // Indexing into tail padding between struct/array elements. 899 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty)) 900 return nullptr; 901 902 if (StructType *STy = dyn_cast<StructType>(Ty)) { 903 const StructLayout *SL = DL->getStructLayout(STy); 904 assert(Offset < (int64_t)SL->getSizeInBytes() && 905 "Offset must stay within the indexed type"); 906 907 unsigned Elt = SL->getElementContainingOffset(Offset); 908 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 909 Elt)); 910 911 Offset -= SL->getElementOffset(Elt); 912 Ty = STy->getElementType(Elt); 913 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 914 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType()); 915 assert(EltSize && "Cannot index into a zero-sized array"); 916 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 917 Offset %= EltSize; 918 Ty = AT->getElementType(); 919 } else { 920 // Otherwise, we can't index into the middle of this atomic type, bail. 921 return nullptr; 922 } 923 } 924 925 return Ty; 926 } 927 928 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { 929 // If this GEP has only 0 indices, it is the same pointer as 930 // Src. If Src is not a trivial GEP too, don't combine 931 // the indices. 932 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() && 933 !Src.hasOneUse()) 934 return false; 935 return true; 936 } 937 938 /// Descale - Return a value X such that Val = X * Scale, or null if none. If 939 /// the multiplication is known not to overflow then NoSignedWrap is set. 940 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { 941 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); 942 assert(cast<IntegerType>(Val->getType())->getBitWidth() == 943 Scale.getBitWidth() && "Scale not compatible with value!"); 944 945 // If Val is zero or Scale is one then Val = Val * Scale. 946 if (match(Val, m_Zero()) || Scale == 1) { 947 NoSignedWrap = true; 948 return Val; 949 } 950 951 // If Scale is zero then it does not divide Val. 952 if (Scale.isMinValue()) 953 return nullptr; 954 955 // Look through chains of multiplications, searching for a constant that is 956 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 957 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by 958 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore 959 // down from Val: 960 // 961 // Val = M1 * X || Analysis starts here and works down 962 // M1 = M2 * Y || Doesn't descend into terms with more 963 // M2 = Z * 4 \/ than one use 964 // 965 // Then to modify a term at the bottom: 966 // 967 // Val = M1 * X 968 // M1 = Z * Y || Replaced M2 with Z 969 // 970 // Then to work back up correcting nsw flags. 971 972 // Op - the term we are currently analyzing. Starts at Val then drills down. 973 // Replaced with its descaled value before exiting from the drill down loop. 974 Value *Op = Val; 975 976 // Parent - initially null, but after drilling down notes where Op came from. 977 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the 978 // 0'th operand of Val. 979 std::pair<Instruction*, unsigned> Parent; 980 981 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper 982 // levels that doesn't overflow. 983 bool RequireNoSignedWrap = false; 984 985 // logScale - log base 2 of the scale. Negative if not a power of 2. 986 int32_t logScale = Scale.exactLogBase2(); 987 988 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down 989 990 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 991 // If Op is a constant divisible by Scale then descale to the quotient. 992 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. 993 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); 994 if (!Remainder.isMinValue()) 995 // Not divisible by Scale. 996 return nullptr; 997 // Replace with the quotient in the parent. 998 Op = ConstantInt::get(CI->getType(), Quotient); 999 NoSignedWrap = true; 1000 break; 1001 } 1002 1003 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { 1004 1005 if (BO->getOpcode() == Instruction::Mul) { 1006 // Multiplication. 1007 NoSignedWrap = BO->hasNoSignedWrap(); 1008 if (RequireNoSignedWrap && !NoSignedWrap) 1009 return nullptr; 1010 1011 // There are three cases for multiplication: multiplication by exactly 1012 // the scale, multiplication by a constant different to the scale, and 1013 // multiplication by something else. 1014 Value *LHS = BO->getOperand(0); 1015 Value *RHS = BO->getOperand(1); 1016 1017 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 1018 // Multiplication by a constant. 1019 if (CI->getValue() == Scale) { 1020 // Multiplication by exactly the scale, replace the multiplication 1021 // by its left-hand side in the parent. 1022 Op = LHS; 1023 break; 1024 } 1025 1026 // Otherwise drill down into the constant. 1027 if (!Op->hasOneUse()) 1028 return nullptr; 1029 1030 Parent = std::make_pair(BO, 1); 1031 continue; 1032 } 1033 1034 // Multiplication by something else. Drill down into the left-hand side 1035 // since that's where the reassociate pass puts the good stuff. 1036 if (!Op->hasOneUse()) 1037 return nullptr; 1038 1039 Parent = std::make_pair(BO, 0); 1040 continue; 1041 } 1042 1043 if (logScale > 0 && BO->getOpcode() == Instruction::Shl && 1044 isa<ConstantInt>(BO->getOperand(1))) { 1045 // Multiplication by a power of 2. 1046 NoSignedWrap = BO->hasNoSignedWrap(); 1047 if (RequireNoSignedWrap && !NoSignedWrap) 1048 return nullptr; 1049 1050 Value *LHS = BO->getOperand(0); 1051 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> 1052 getLimitedValue(Scale.getBitWidth()); 1053 // Op = LHS << Amt. 1054 1055 if (Amt == logScale) { 1056 // Multiplication by exactly the scale, replace the multiplication 1057 // by its left-hand side in the parent. 1058 Op = LHS; 1059 break; 1060 } 1061 if (Amt < logScale || !Op->hasOneUse()) 1062 return nullptr; 1063 1064 // Multiplication by more than the scale. Reduce the multiplying amount 1065 // by the scale in the parent. 1066 Parent = std::make_pair(BO, 1); 1067 Op = ConstantInt::get(BO->getType(), Amt - logScale); 1068 break; 1069 } 1070 } 1071 1072 if (!Op->hasOneUse()) 1073 return nullptr; 1074 1075 if (CastInst *Cast = dyn_cast<CastInst>(Op)) { 1076 if (Cast->getOpcode() == Instruction::SExt) { 1077 // Op is sign-extended from a smaller type, descale in the smaller type. 1078 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1079 APInt SmallScale = Scale.trunc(SmallSize); 1080 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to 1081 // descale Op as (sext Y) * Scale. In order to have 1082 // sext (Y * SmallScale) = (sext Y) * Scale 1083 // some conditions need to hold however: SmallScale must sign-extend to 1084 // Scale and the multiplication Y * SmallScale should not overflow. 1085 if (SmallScale.sext(Scale.getBitWidth()) != Scale) 1086 // SmallScale does not sign-extend to Scale. 1087 return nullptr; 1088 assert(SmallScale.exactLogBase2() == logScale); 1089 // Require that Y * SmallScale must not overflow. 1090 RequireNoSignedWrap = true; 1091 1092 // Drill down through the cast. 1093 Parent = std::make_pair(Cast, 0); 1094 Scale = SmallScale; 1095 continue; 1096 } 1097 1098 if (Cast->getOpcode() == Instruction::Trunc) { 1099 // Op is truncated from a larger type, descale in the larger type. 1100 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then 1101 // trunc (Y * sext Scale) = (trunc Y) * Scale 1102 // always holds. However (trunc Y) * Scale may overflow even if 1103 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared 1104 // from this point up in the expression (see later). 1105 if (RequireNoSignedWrap) 1106 return nullptr; 1107 1108 // Drill down through the cast. 1109 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1110 Parent = std::make_pair(Cast, 0); 1111 Scale = Scale.sext(LargeSize); 1112 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) 1113 logScale = -1; 1114 assert(Scale.exactLogBase2() == logScale); 1115 continue; 1116 } 1117 } 1118 1119 // Unsupported expression, bail out. 1120 return nullptr; 1121 } 1122 1123 // If Op is zero then Val = Op * Scale. 1124 if (match(Op, m_Zero())) { 1125 NoSignedWrap = true; 1126 return Op; 1127 } 1128 1129 // We know that we can successfully descale, so from here on we can safely 1130 // modify the IR. Op holds the descaled version of the deepest term in the 1131 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known 1132 // not to overflow. 1133 1134 if (!Parent.first) 1135 // The expression only had one term. 1136 return Op; 1137 1138 // Rewrite the parent using the descaled version of its operand. 1139 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); 1140 assert(Op != Parent.first->getOperand(Parent.second) && 1141 "Descaling was a no-op?"); 1142 Parent.first->setOperand(Parent.second, Op); 1143 Worklist.Add(Parent.first); 1144 1145 // Now work back up the expression correcting nsw flags. The logic is based 1146 // on the following observation: if X * Y is known not to overflow as a signed 1147 // multiplication, and Y is replaced by a value Z with smaller absolute value, 1148 // then X * Z will not overflow as a signed multiplication either. As we work 1149 // our way up, having NoSignedWrap 'true' means that the descaled value at the 1150 // current level has strictly smaller absolute value than the original. 1151 Instruction *Ancestor = Parent.first; 1152 do { 1153 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { 1154 // If the multiplication wasn't nsw then we can't say anything about the 1155 // value of the descaled multiplication, and we have to clear nsw flags 1156 // from this point on up. 1157 bool OpNoSignedWrap = BO->hasNoSignedWrap(); 1158 NoSignedWrap &= OpNoSignedWrap; 1159 if (NoSignedWrap != OpNoSignedWrap) { 1160 BO->setHasNoSignedWrap(NoSignedWrap); 1161 Worklist.Add(Ancestor); 1162 } 1163 } else if (Ancestor->getOpcode() == Instruction::Trunc) { 1164 // The fact that the descaled input to the trunc has smaller absolute 1165 // value than the original input doesn't tell us anything useful about 1166 // the absolute values of the truncations. 1167 NoSignedWrap = false; 1168 } 1169 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && 1170 "Failed to keep proper track of nsw flags while drilling down?"); 1171 1172 if (Ancestor == Val) 1173 // Got to the top, all done! 1174 return Val; 1175 1176 // Move up one level in the expression. 1177 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); 1178 Ancestor = Ancestor->user_back(); 1179 } while (1); 1180 } 1181 1182 /// \brief Creates node of binary operation with the same attributes as the 1183 /// specified one but with other operands. 1184 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS, 1185 InstCombiner::BuilderTy *B) { 1186 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS); 1187 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) { 1188 if (isa<OverflowingBinaryOperator>(NewBO)) { 1189 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap()); 1190 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap()); 1191 } 1192 if (isa<PossiblyExactOperator>(NewBO)) 1193 NewBO->setIsExact(Inst.isExact()); 1194 } 1195 return BORes; 1196 } 1197 1198 /// \brief Makes transformation of binary operation specific for vector types. 1199 /// \param Inst Binary operator to transform. 1200 /// \return Pointer to node that must replace the original binary operator, or 1201 /// null pointer if no transformation was made. 1202 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) { 1203 if (!Inst.getType()->isVectorTy()) return nullptr; 1204 1205 // It may not be safe to reorder shuffles and things like div, urem, etc. 1206 // because we may trap when executing those ops on unknown vector elements. 1207 // See PR20059. 1208 if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr; 1209 1210 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements(); 1211 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1); 1212 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth); 1213 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth); 1214 1215 // If both arguments of binary operation are shuffles, which use the same 1216 // mask and shuffle within a single vector, it is worthwhile to move the 1217 // shuffle after binary operation: 1218 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m) 1219 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) { 1220 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS); 1221 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS); 1222 if (isa<UndefValue>(LShuf->getOperand(1)) && 1223 isa<UndefValue>(RShuf->getOperand(1)) && 1224 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() && 1225 LShuf->getMask() == RShuf->getMask()) { 1226 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0), 1227 RShuf->getOperand(0), Builder); 1228 Value *Res = Builder->CreateShuffleVector(NewBO, 1229 UndefValue::get(NewBO->getType()), LShuf->getMask()); 1230 return Res; 1231 } 1232 } 1233 1234 // If one argument is a shuffle within one vector, the other is a constant, 1235 // try moving the shuffle after the binary operation. 1236 ShuffleVectorInst *Shuffle = nullptr; 1237 Constant *C1 = nullptr; 1238 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS); 1239 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS); 1240 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS); 1241 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS); 1242 if (Shuffle && C1 && 1243 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) && 1244 isa<UndefValue>(Shuffle->getOperand(1)) && 1245 Shuffle->getType() == Shuffle->getOperand(0)->getType()) { 1246 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask(); 1247 // Find constant C2 that has property: 1248 // shuffle(C2, ShMask) = C1 1249 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>) 1250 // reorder is not possible. 1251 SmallVector<Constant*, 16> C2M(VWidth, 1252 UndefValue::get(C1->getType()->getScalarType())); 1253 bool MayChange = true; 1254 for (unsigned I = 0; I < VWidth; ++I) { 1255 if (ShMask[I] >= 0) { 1256 assert(ShMask[I] < (int)VWidth); 1257 if (!isa<UndefValue>(C2M[ShMask[I]])) { 1258 MayChange = false; 1259 break; 1260 } 1261 C2M[ShMask[I]] = C1->getAggregateElement(I); 1262 } 1263 } 1264 if (MayChange) { 1265 Constant *C2 = ConstantVector::get(C2M); 1266 Value *NewLHS, *NewRHS; 1267 if (isa<Constant>(LHS)) { 1268 NewLHS = C2; 1269 NewRHS = Shuffle->getOperand(0); 1270 } else { 1271 NewLHS = Shuffle->getOperand(0); 1272 NewRHS = C2; 1273 } 1274 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder); 1275 Value *Res = Builder->CreateShuffleVector(NewBO, 1276 UndefValue::get(Inst.getType()), Shuffle->getMask()); 1277 return Res; 1278 } 1279 } 1280 1281 return nullptr; 1282 } 1283 1284 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 1285 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 1286 1287 if (Value *V = SimplifyGEPInst(Ops, DL)) 1288 return ReplaceInstUsesWith(GEP, V); 1289 1290 Value *PtrOp = GEP.getOperand(0); 1291 1292 // Eliminate unneeded casts for indices, and replace indices which displace 1293 // by multiples of a zero size type with zero. 1294 if (DL) { 1295 bool MadeChange = false; 1296 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType()); 1297 1298 gep_type_iterator GTI = gep_type_begin(GEP); 1299 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 1300 I != E; ++I, ++GTI) { 1301 // Skip indices into struct types. 1302 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 1303 if (!SeqTy) continue; 1304 1305 // If the element type has zero size then any index over it is equivalent 1306 // to an index of zero, so replace it with zero if it is not zero already. 1307 if (SeqTy->getElementType()->isSized() && 1308 DL->getTypeAllocSize(SeqTy->getElementType()) == 0) 1309 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 1310 *I = Constant::getNullValue(IntPtrTy); 1311 MadeChange = true; 1312 } 1313 1314 Type *IndexTy = (*I)->getType(); 1315 if (IndexTy != IntPtrTy) { 1316 // If we are using a wider index than needed for this platform, shrink 1317 // it to what we need. If narrower, sign-extend it to what we need. 1318 // This explicit cast can make subsequent optimizations more obvious. 1319 *I = Builder->CreateIntCast(*I, IntPtrTy, true); 1320 MadeChange = true; 1321 } 1322 } 1323 if (MadeChange) return &GEP; 1324 } 1325 1326 // Check to see if the inputs to the PHI node are getelementptr instructions. 1327 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) { 1328 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0)); 1329 if (!Op1) 1330 return nullptr; 1331 1332 signed DI = -1; 1333 1334 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) { 1335 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I); 1336 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands()) 1337 return nullptr; 1338 1339 // Keep track of the type as we walk the GEP. 1340 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType(); 1341 1342 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) { 1343 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType()) 1344 return nullptr; 1345 1346 if (Op1->getOperand(J) != Op2->getOperand(J)) { 1347 if (DI == -1) { 1348 // We have not seen any differences yet in the GEPs feeding the 1349 // PHI yet, so we record this one if it is allowed to be a 1350 // variable. 1351 1352 // The first two arguments can vary for any GEP, the rest have to be 1353 // static for struct slots 1354 if (J > 1 && CurTy->isStructTy()) 1355 return nullptr; 1356 1357 DI = J; 1358 } else { 1359 // The GEP is different by more than one input. While this could be 1360 // extended to support GEPs that vary by more than one variable it 1361 // doesn't make sense since it greatly increases the complexity and 1362 // would result in an R+R+R addressing mode which no backend 1363 // directly supports and would need to be broken into several 1364 // simpler instructions anyway. 1365 return nullptr; 1366 } 1367 } 1368 1369 // Sink down a layer of the type for the next iteration. 1370 if (J > 0) { 1371 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) { 1372 CurTy = CT->getTypeAtIndex(Op1->getOperand(J)); 1373 } else { 1374 CurTy = nullptr; 1375 } 1376 } 1377 } 1378 } 1379 1380 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone()); 1381 1382 if (DI == -1) { 1383 // All the GEPs feeding the PHI are identical. Clone one down into our 1384 // BB so that it can be merged with the current GEP. 1385 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(), 1386 NewGEP); 1387 } else { 1388 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP 1389 // into the current block so it can be merged, and create a new PHI to 1390 // set that index. 1391 Instruction *InsertPt = Builder->GetInsertPoint(); 1392 Builder->SetInsertPoint(PN); 1393 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(), 1394 PN->getNumOperands()); 1395 Builder->SetInsertPoint(InsertPt); 1396 1397 for (auto &I : PN->operands()) 1398 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI), 1399 PN->getIncomingBlock(I)); 1400 1401 NewGEP->setOperand(DI, NewPN); 1402 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(), 1403 NewGEP); 1404 NewGEP->setOperand(DI, NewPN); 1405 } 1406 1407 GEP.setOperand(0, NewGEP); 1408 PtrOp = NewGEP; 1409 } 1410 1411 // Combine Indices - If the source pointer to this getelementptr instruction 1412 // is a getelementptr instruction, combine the indices of the two 1413 // getelementptr instructions into a single instruction. 1414 // 1415 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 1416 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) 1417 return nullptr; 1418 1419 // Note that if our source is a gep chain itself then we wait for that 1420 // chain to be resolved before we perform this transformation. This 1421 // avoids us creating a TON of code in some cases. 1422 if (GEPOperator *SrcGEP = 1423 dyn_cast<GEPOperator>(Src->getOperand(0))) 1424 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) 1425 return nullptr; // Wait until our source is folded to completion. 1426 1427 SmallVector<Value*, 8> Indices; 1428 1429 // Find out whether the last index in the source GEP is a sequential idx. 1430 bool EndsWithSequential = false; 1431 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 1432 I != E; ++I) 1433 EndsWithSequential = !(*I)->isStructTy(); 1434 1435 // Can we combine the two pointer arithmetics offsets? 1436 if (EndsWithSequential) { 1437 // Replace: gep (gep %P, long B), long A, ... 1438 // With: T = long A+B; gep %P, T, ... 1439 // 1440 Value *Sum; 1441 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 1442 Value *GO1 = GEP.getOperand(1); 1443 if (SO1 == Constant::getNullValue(SO1->getType())) { 1444 Sum = GO1; 1445 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 1446 Sum = SO1; 1447 } else { 1448 // If they aren't the same type, then the input hasn't been processed 1449 // by the loop above yet (which canonicalizes sequential index types to 1450 // intptr_t). Just avoid transforming this until the input has been 1451 // normalized. 1452 if (SO1->getType() != GO1->getType()) 1453 return nullptr; 1454 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 1455 } 1456 1457 // Update the GEP in place if possible. 1458 if (Src->getNumOperands() == 2) { 1459 GEP.setOperand(0, Src->getOperand(0)); 1460 GEP.setOperand(1, Sum); 1461 return &GEP; 1462 } 1463 Indices.append(Src->op_begin()+1, Src->op_end()-1); 1464 Indices.push_back(Sum); 1465 Indices.append(GEP.op_begin()+2, GEP.op_end()); 1466 } else if (isa<Constant>(*GEP.idx_begin()) && 1467 cast<Constant>(*GEP.idx_begin())->isNullValue() && 1468 Src->getNumOperands() != 1) { 1469 // Otherwise we can do the fold if the first index of the GEP is a zero 1470 Indices.append(Src->op_begin()+1, Src->op_end()); 1471 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 1472 } 1473 1474 if (!Indices.empty()) 1475 return (GEP.isInBounds() && Src->isInBounds()) ? 1476 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices, 1477 GEP.getName()) : 1478 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName()); 1479 } 1480 1481 // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y)) 1482 // The GEP pattern is emitted by the SCEV expander for certain kinds of 1483 // pointer arithmetic. 1484 if (DL && GEP.getNumIndices() == 1 && 1485 match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) { 1486 unsigned AS = GEP.getPointerAddressSpace(); 1487 if (GEP.getType() == Builder->getInt8PtrTy(AS) && 1488 GEP.getOperand(1)->getType()->getScalarSizeInBits() == 1489 DL->getPointerSizeInBits(AS)) { 1490 Operator *Index = cast<Operator>(GEP.getOperand(1)); 1491 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); 1492 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); 1493 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); 1494 } 1495 } 1496 1497 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 1498 Value *StrippedPtr = PtrOp->stripPointerCasts(); 1499 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType()); 1500 1501 // We do not handle pointer-vector geps here. 1502 if (!StrippedPtrTy) 1503 return nullptr; 1504 1505 if (StrippedPtr != PtrOp) { 1506 bool HasZeroPointerIndex = false; 1507 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 1508 HasZeroPointerIndex = C->isZero(); 1509 1510 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 1511 // into : GEP [10 x i8]* X, i32 0, ... 1512 // 1513 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 1514 // into : GEP i8* X, ... 1515 // 1516 // This occurs when the program declares an array extern like "int X[];" 1517 if (HasZeroPointerIndex) { 1518 PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 1519 if (ArrayType *CATy = 1520 dyn_cast<ArrayType>(CPTy->getElementType())) { 1521 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 1522 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 1523 // -> GEP i8* X, ... 1524 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 1525 GetElementPtrInst *Res = 1526 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName()); 1527 Res->setIsInBounds(GEP.isInBounds()); 1528 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) 1529 return Res; 1530 // Insert Res, and create an addrspacecast. 1531 // e.g., 1532 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ... 1533 // -> 1534 // %0 = GEP i8 addrspace(1)* X, ... 1535 // addrspacecast i8 addrspace(1)* %0 to i8* 1536 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType()); 1537 } 1538 1539 if (ArrayType *XATy = 1540 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 1541 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 1542 if (CATy->getElementType() == XATy->getElementType()) { 1543 // -> GEP [10 x i8]* X, i32 0, ... 1544 // At this point, we know that the cast source type is a pointer 1545 // to an array of the same type as the destination pointer 1546 // array. Because the array type is never stepped over (there 1547 // is a leading zero) we can fold the cast into this GEP. 1548 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) { 1549 GEP.setOperand(0, StrippedPtr); 1550 return &GEP; 1551 } 1552 // Cannot replace the base pointer directly because StrippedPtr's 1553 // address space is different. Instead, create a new GEP followed by 1554 // an addrspacecast. 1555 // e.g., 1556 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*), 1557 // i32 0, ... 1558 // -> 1559 // %0 = GEP [10 x i8] addrspace(1)* X, ... 1560 // addrspacecast i8 addrspace(1)* %0 to i8* 1561 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end()); 1562 Value *NewGEP = GEP.isInBounds() ? 1563 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : 1564 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); 1565 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1566 } 1567 } 1568 } 1569 } else if (GEP.getNumOperands() == 2) { 1570 // Transform things like: 1571 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 1572 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 1573 Type *SrcElTy = StrippedPtrTy->getElementType(); 1574 Type *ResElTy = PtrOp->getType()->getPointerElementType(); 1575 if (DL && SrcElTy->isArrayTy() && 1576 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) == 1577 DL->getTypeAllocSize(ResElTy)) { 1578 Type *IdxType = DL->getIntPtrType(GEP.getType()); 1579 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; 1580 Value *NewGEP = GEP.isInBounds() ? 1581 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : 1582 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); 1583 1584 // V and GEP are both pointer types --> BitCast 1585 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1586 return new BitCastInst(NewGEP, GEP.getType()); 1587 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1588 } 1589 1590 // Transform things like: 1591 // %V = mul i64 %N, 4 1592 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V 1593 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast 1594 if (DL && ResElTy->isSized() && SrcElTy->isSized()) { 1595 // Check that changing the type amounts to dividing the index by a scale 1596 // factor. 1597 uint64_t ResSize = DL->getTypeAllocSize(ResElTy); 1598 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy); 1599 if (ResSize && SrcSize % ResSize == 0) { 1600 Value *Idx = GEP.getOperand(1); 1601 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1602 uint64_t Scale = SrcSize / ResSize; 1603 1604 // Earlier transforms ensure that the index has type IntPtrType, which 1605 // considerably simplifies the logic by eliminating implicit casts. 1606 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && 1607 "Index not cast to pointer width?"); 1608 1609 bool NSW; 1610 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1611 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1612 // If the multiplication NewIdx * Scale may overflow then the new 1613 // GEP may not be "inbounds". 1614 Value *NewGEP = GEP.isInBounds() && NSW ? 1615 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) : 1616 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName()); 1617 1618 // The NewGEP must be pointer typed, so must the old one -> BitCast 1619 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1620 return new BitCastInst(NewGEP, GEP.getType()); 1621 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1622 } 1623 } 1624 } 1625 1626 // Similarly, transform things like: 1627 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 1628 // (where tmp = 8*tmp2) into: 1629 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 1630 if (DL && ResElTy->isSized() && SrcElTy->isSized() && 1631 SrcElTy->isArrayTy()) { 1632 // Check that changing to the array element type amounts to dividing the 1633 // index by a scale factor. 1634 uint64_t ResSize = DL->getTypeAllocSize(ResElTy); 1635 uint64_t ArrayEltSize 1636 = DL->getTypeAllocSize(SrcElTy->getArrayElementType()); 1637 if (ResSize && ArrayEltSize % ResSize == 0) { 1638 Value *Idx = GEP.getOperand(1); 1639 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1640 uint64_t Scale = ArrayEltSize / ResSize; 1641 1642 // Earlier transforms ensure that the index has type IntPtrType, which 1643 // considerably simplifies the logic by eliminating implicit casts. 1644 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && 1645 "Index not cast to pointer width?"); 1646 1647 bool NSW; 1648 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1649 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1650 // If the multiplication NewIdx * Scale may overflow then the new 1651 // GEP may not be "inbounds". 1652 Value *Off[2] = { 1653 Constant::getNullValue(DL->getIntPtrType(GEP.getType())), 1654 NewIdx 1655 }; 1656 1657 Value *NewGEP = GEP.isInBounds() && NSW ? 1658 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) : 1659 Builder->CreateGEP(StrippedPtr, Off, GEP.getName()); 1660 // The NewGEP must be pointer typed, so must the old one -> BitCast 1661 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1662 return new BitCastInst(NewGEP, GEP.getType()); 1663 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1664 } 1665 } 1666 } 1667 } 1668 } 1669 1670 if (!DL) 1671 return nullptr; 1672 1673 /// See if we can simplify: 1674 /// X = bitcast A* to B* 1675 /// Y = gep X, <...constant indices...> 1676 /// into a gep of the original struct. This is important for SROA and alias 1677 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 1678 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 1679 Value *Operand = BCI->getOperand(0); 1680 PointerType *OpType = cast<PointerType>(Operand->getType()); 1681 unsigned OffsetBits = DL->getPointerTypeSizeInBits(OpType); 1682 APInt Offset(OffsetBits, 0); 1683 if (!isa<BitCastInst>(Operand) && 1684 GEP.accumulateConstantOffset(*DL, Offset) && 1685 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 1686 1687 // If this GEP instruction doesn't move the pointer, just replace the GEP 1688 // with a bitcast of the real input to the dest type. 1689 if (!Offset) { 1690 // If the bitcast is of an allocation, and the allocation will be 1691 // converted to match the type of the cast, don't touch this. 1692 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) { 1693 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1694 if (Instruction *I = visitBitCast(*BCI)) { 1695 if (I != BCI) { 1696 I->takeName(BCI); 1697 BCI->getParent()->getInstList().insert(BCI, I); 1698 ReplaceInstUsesWith(*BCI, I); 1699 } 1700 return &GEP; 1701 } 1702 } 1703 return new BitCastInst(Operand, GEP.getType()); 1704 } 1705 1706 // Otherwise, if the offset is non-zero, we need to find out if there is a 1707 // field at Offset in 'A's type. If so, we can pull the cast through the 1708 // GEP. 1709 SmallVector<Value*, 8> NewIndices; 1710 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { 1711 Value *NGEP = GEP.isInBounds() ? 1712 Builder->CreateInBoundsGEP(Operand, NewIndices) : 1713 Builder->CreateGEP(Operand, NewIndices); 1714 1715 if (NGEP->getType() == GEP.getType()) 1716 return ReplaceInstUsesWith(GEP, NGEP); 1717 NGEP->takeName(&GEP); 1718 return new BitCastInst(NGEP, GEP.getType()); 1719 } 1720 } 1721 } 1722 1723 return nullptr; 1724 } 1725 1726 static bool 1727 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, 1728 const TargetLibraryInfo *TLI) { 1729 SmallVector<Instruction*, 4> Worklist; 1730 Worklist.push_back(AI); 1731 1732 do { 1733 Instruction *PI = Worklist.pop_back_val(); 1734 for (User *U : PI->users()) { 1735 Instruction *I = cast<Instruction>(U); 1736 switch (I->getOpcode()) { 1737 default: 1738 // Give up the moment we see something we can't handle. 1739 return false; 1740 1741 case Instruction::BitCast: 1742 case Instruction::GetElementPtr: 1743 Users.push_back(I); 1744 Worklist.push_back(I); 1745 continue; 1746 1747 case Instruction::ICmp: { 1748 ICmpInst *ICI = cast<ICmpInst>(I); 1749 // We can fold eq/ne comparisons with null to false/true, respectively. 1750 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1))) 1751 return false; 1752 Users.push_back(I); 1753 continue; 1754 } 1755 1756 case Instruction::Call: 1757 // Ignore no-op and store intrinsics. 1758 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1759 switch (II->getIntrinsicID()) { 1760 default: 1761 return false; 1762 1763 case Intrinsic::memmove: 1764 case Intrinsic::memcpy: 1765 case Intrinsic::memset: { 1766 MemIntrinsic *MI = cast<MemIntrinsic>(II); 1767 if (MI->isVolatile() || MI->getRawDest() != PI) 1768 return false; 1769 } 1770 // fall through 1771 case Intrinsic::dbg_declare: 1772 case Intrinsic::dbg_value: 1773 case Intrinsic::invariant_start: 1774 case Intrinsic::invariant_end: 1775 case Intrinsic::lifetime_start: 1776 case Intrinsic::lifetime_end: 1777 case Intrinsic::objectsize: 1778 Users.push_back(I); 1779 continue; 1780 } 1781 } 1782 1783 if (isFreeCall(I, TLI)) { 1784 Users.push_back(I); 1785 continue; 1786 } 1787 return false; 1788 1789 case Instruction::Store: { 1790 StoreInst *SI = cast<StoreInst>(I); 1791 if (SI->isVolatile() || SI->getPointerOperand() != PI) 1792 return false; 1793 Users.push_back(I); 1794 continue; 1795 } 1796 } 1797 llvm_unreachable("missing a return?"); 1798 } 1799 } while (!Worklist.empty()); 1800 return true; 1801 } 1802 1803 Instruction *InstCombiner::visitAllocSite(Instruction &MI) { 1804 // If we have a malloc call which is only used in any amount of comparisons 1805 // to null and free calls, delete the calls and replace the comparisons with 1806 // true or false as appropriate. 1807 SmallVector<WeakVH, 64> Users; 1808 if (isAllocSiteRemovable(&MI, Users, TLI)) { 1809 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 1810 Instruction *I = cast_or_null<Instruction>(&*Users[i]); 1811 if (!I) continue; 1812 1813 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) { 1814 ReplaceInstUsesWith(*C, 1815 ConstantInt::get(Type::getInt1Ty(C->getContext()), 1816 C->isFalseWhenEqual())); 1817 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 1818 ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); 1819 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1820 if (II->getIntrinsicID() == Intrinsic::objectsize) { 1821 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1)); 1822 uint64_t DontKnow = CI->isZero() ? -1ULL : 0; 1823 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); 1824 } 1825 } 1826 EraseInstFromFunction(*I); 1827 } 1828 1829 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) { 1830 // Replace invoke with a NOP intrinsic to maintain the original CFG 1831 Module *M = II->getParent()->getParent()->getParent(); 1832 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); 1833 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), 1834 None, "", II->getParent()); 1835 } 1836 return EraseInstFromFunction(MI); 1837 } 1838 return nullptr; 1839 } 1840 1841 /// \brief Move the call to free before a NULL test. 1842 /// 1843 /// Check if this free is accessed after its argument has been test 1844 /// against NULL (property 0). 1845 /// If yes, it is legal to move this call in its predecessor block. 1846 /// 1847 /// The move is performed only if the block containing the call to free 1848 /// will be removed, i.e.: 1849 /// 1. it has only one predecessor P, and P has two successors 1850 /// 2. it contains the call and an unconditional branch 1851 /// 3. its successor is the same as its predecessor's successor 1852 /// 1853 /// The profitability is out-of concern here and this function should 1854 /// be called only if the caller knows this transformation would be 1855 /// profitable (e.g., for code size). 1856 static Instruction * 1857 tryToMoveFreeBeforeNullTest(CallInst &FI) { 1858 Value *Op = FI.getArgOperand(0); 1859 BasicBlock *FreeInstrBB = FI.getParent(); 1860 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); 1861 1862 // Validate part of constraint #1: Only one predecessor 1863 // FIXME: We can extend the number of predecessor, but in that case, we 1864 // would duplicate the call to free in each predecessor and it may 1865 // not be profitable even for code size. 1866 if (!PredBB) 1867 return nullptr; 1868 1869 // Validate constraint #2: Does this block contains only the call to 1870 // free and an unconditional branch? 1871 // FIXME: We could check if we can speculate everything in the 1872 // predecessor block 1873 if (FreeInstrBB->size() != 2) 1874 return nullptr; 1875 BasicBlock *SuccBB; 1876 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) 1877 return nullptr; 1878 1879 // Validate the rest of constraint #1 by matching on the pred branch. 1880 TerminatorInst *TI = PredBB->getTerminator(); 1881 BasicBlock *TrueBB, *FalseBB; 1882 ICmpInst::Predicate Pred; 1883 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) 1884 return nullptr; 1885 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) 1886 return nullptr; 1887 1888 // Validate constraint #3: Ensure the null case just falls through. 1889 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) 1890 return nullptr; 1891 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && 1892 "Broken CFG: missing edge from predecessor to successor"); 1893 1894 FI.moveBefore(TI); 1895 return &FI; 1896 } 1897 1898 1899 Instruction *InstCombiner::visitFree(CallInst &FI) { 1900 Value *Op = FI.getArgOperand(0); 1901 1902 // free undef -> unreachable. 1903 if (isa<UndefValue>(Op)) { 1904 // Insert a new store to null because we cannot modify the CFG here. 1905 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()), 1906 UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); 1907 return EraseInstFromFunction(FI); 1908 } 1909 1910 // If we have 'free null' delete the instruction. This can happen in stl code 1911 // when lots of inlining happens. 1912 if (isa<ConstantPointerNull>(Op)) 1913 return EraseInstFromFunction(FI); 1914 1915 // If we optimize for code size, try to move the call to free before the null 1916 // test so that simplify cfg can remove the empty block and dead code 1917 // elimination the branch. I.e., helps to turn something like: 1918 // if (foo) free(foo); 1919 // into 1920 // free(foo); 1921 if (MinimizeSize) 1922 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) 1923 return I; 1924 1925 return nullptr; 1926 } 1927 1928 1929 1930 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 1931 // Change br (not X), label True, label False to: br X, label False, True 1932 Value *X = nullptr; 1933 BasicBlock *TrueDest; 1934 BasicBlock *FalseDest; 1935 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 1936 !isa<Constant>(X)) { 1937 // Swap Destinations and condition... 1938 BI.setCondition(X); 1939 BI.swapSuccessors(); 1940 return &BI; 1941 } 1942 1943 // Canonicalize fcmp_one -> fcmp_oeq 1944 FCmpInst::Predicate FPred; Value *Y; 1945 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 1946 TrueDest, FalseDest)) && 1947 BI.getCondition()->hasOneUse()) 1948 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 1949 FPred == FCmpInst::FCMP_OGE) { 1950 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 1951 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 1952 1953 // Swap Destinations and condition. 1954 BI.swapSuccessors(); 1955 Worklist.Add(Cond); 1956 return &BI; 1957 } 1958 1959 // Canonicalize icmp_ne -> icmp_eq 1960 ICmpInst::Predicate IPred; 1961 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 1962 TrueDest, FalseDest)) && 1963 BI.getCondition()->hasOneUse()) 1964 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 1965 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 1966 IPred == ICmpInst::ICMP_SGE) { 1967 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 1968 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 1969 // Swap Destinations and condition. 1970 BI.swapSuccessors(); 1971 Worklist.Add(Cond); 1972 return &BI; 1973 } 1974 1975 return nullptr; 1976 } 1977 1978 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 1979 Value *Cond = SI.getCondition(); 1980 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 1981 if (I->getOpcode() == Instruction::Add) 1982 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 1983 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 1984 // Skip the first item since that's the default case. 1985 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); 1986 i != e; ++i) { 1987 ConstantInt* CaseVal = i.getCaseValue(); 1988 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal), 1989 AddRHS); 1990 assert(isa<ConstantInt>(NewCaseVal) && 1991 "Result of expression should be constant"); 1992 i.setValue(cast<ConstantInt>(NewCaseVal)); 1993 } 1994 SI.setCondition(I->getOperand(0)); 1995 Worklist.Add(I); 1996 return &SI; 1997 } 1998 } 1999 return nullptr; 2000 } 2001 2002 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 2003 Value *Agg = EV.getAggregateOperand(); 2004 2005 if (!EV.hasIndices()) 2006 return ReplaceInstUsesWith(EV, Agg); 2007 2008 if (Constant *C = dyn_cast<Constant>(Agg)) { 2009 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) { 2010 if (EV.getNumIndices() == 0) 2011 return ReplaceInstUsesWith(EV, C2); 2012 // Extract the remaining indices out of the constant indexed by the 2013 // first index 2014 return ExtractValueInst::Create(C2, EV.getIndices().slice(1)); 2015 } 2016 return nullptr; // Can't handle other constants 2017 } 2018 2019 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 2020 // We're extracting from an insertvalue instruction, compare the indices 2021 const unsigned *exti, *exte, *insi, *inse; 2022 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 2023 exte = EV.idx_end(), inse = IV->idx_end(); 2024 exti != exte && insi != inse; 2025 ++exti, ++insi) { 2026 if (*insi != *exti) 2027 // The insert and extract both reference distinctly different elements. 2028 // This means the extract is not influenced by the insert, and we can 2029 // replace the aggregate operand of the extract with the aggregate 2030 // operand of the insert. i.e., replace 2031 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 2032 // %E = extractvalue { i32, { i32 } } %I, 0 2033 // with 2034 // %E = extractvalue { i32, { i32 } } %A, 0 2035 return ExtractValueInst::Create(IV->getAggregateOperand(), 2036 EV.getIndices()); 2037 } 2038 if (exti == exte && insi == inse) 2039 // Both iterators are at the end: Index lists are identical. Replace 2040 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 2041 // %C = extractvalue { i32, { i32 } } %B, 1, 0 2042 // with "i32 42" 2043 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 2044 if (exti == exte) { 2045 // The extract list is a prefix of the insert list. i.e. replace 2046 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 2047 // %E = extractvalue { i32, { i32 } } %I, 1 2048 // with 2049 // %X = extractvalue { i32, { i32 } } %A, 1 2050 // %E = insertvalue { i32 } %X, i32 42, 0 2051 // by switching the order of the insert and extract (though the 2052 // insertvalue should be left in, since it may have other uses). 2053 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 2054 EV.getIndices()); 2055 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 2056 makeArrayRef(insi, inse)); 2057 } 2058 if (insi == inse) 2059 // The insert list is a prefix of the extract list 2060 // We can simply remove the common indices from the extract and make it 2061 // operate on the inserted value instead of the insertvalue result. 2062 // i.e., replace 2063 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 2064 // %E = extractvalue { i32, { i32 } } %I, 1, 0 2065 // with 2066 // %E extractvalue { i32 } { i32 42 }, 0 2067 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 2068 makeArrayRef(exti, exte)); 2069 } 2070 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 2071 // We're extracting from an intrinsic, see if we're the only user, which 2072 // allows us to simplify multiple result intrinsics to simpler things that 2073 // just get one value. 2074 if (II->hasOneUse()) { 2075 // Check if we're grabbing the overflow bit or the result of a 'with 2076 // overflow' intrinsic. If it's the latter we can remove the intrinsic 2077 // and replace it with a traditional binary instruction. 2078 switch (II->getIntrinsicID()) { 2079 case Intrinsic::uadd_with_overflow: 2080 case Intrinsic::sadd_with_overflow: 2081 if (*EV.idx_begin() == 0) { // Normal result. 2082 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2083 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2084 EraseInstFromFunction(*II); 2085 return BinaryOperator::CreateAdd(LHS, RHS); 2086 } 2087 2088 // If the normal result of the add is dead, and the RHS is a constant, 2089 // we can transform this into a range comparison. 2090 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 2091 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 2092 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 2093 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 2094 ConstantExpr::getNot(CI)); 2095 break; 2096 case Intrinsic::usub_with_overflow: 2097 case Intrinsic::ssub_with_overflow: 2098 if (*EV.idx_begin() == 0) { // Normal result. 2099 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2100 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2101 EraseInstFromFunction(*II); 2102 return BinaryOperator::CreateSub(LHS, RHS); 2103 } 2104 break; 2105 case Intrinsic::umul_with_overflow: 2106 case Intrinsic::smul_with_overflow: 2107 if (*EV.idx_begin() == 0) { // Normal result. 2108 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2109 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2110 EraseInstFromFunction(*II); 2111 return BinaryOperator::CreateMul(LHS, RHS); 2112 } 2113 break; 2114 default: 2115 break; 2116 } 2117 } 2118 } 2119 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 2120 // If the (non-volatile) load only has one use, we can rewrite this to a 2121 // load from a GEP. This reduces the size of the load. 2122 // FIXME: If a load is used only by extractvalue instructions then this 2123 // could be done regardless of having multiple uses. 2124 if (L->isSimple() && L->hasOneUse()) { 2125 // extractvalue has integer indices, getelementptr has Value*s. Convert. 2126 SmallVector<Value*, 4> Indices; 2127 // Prefix an i32 0 since we need the first element. 2128 Indices.push_back(Builder->getInt32(0)); 2129 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 2130 I != E; ++I) 2131 Indices.push_back(Builder->getInt32(*I)); 2132 2133 // We need to insert these at the location of the old load, not at that of 2134 // the extractvalue. 2135 Builder->SetInsertPoint(L->getParent(), L); 2136 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices); 2137 // Returning the load directly will cause the main loop to insert it in 2138 // the wrong spot, so use ReplaceInstUsesWith(). 2139 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 2140 } 2141 // We could simplify extracts from other values. Note that nested extracts may 2142 // already be simplified implicitly by the above: extract (extract (insert) ) 2143 // will be translated into extract ( insert ( extract ) ) first and then just 2144 // the value inserted, if appropriate. Similarly for extracts from single-use 2145 // loads: extract (extract (load)) will be translated to extract (load (gep)) 2146 // and if again single-use then via load (gep (gep)) to load (gep). 2147 // However, double extracts from e.g. function arguments or return values 2148 // aren't handled yet. 2149 return nullptr; 2150 } 2151 2152 enum Personality_Type { 2153 Unknown_Personality, 2154 GNU_Ada_Personality, 2155 GNU_CXX_Personality, 2156 GNU_ObjC_Personality 2157 }; 2158 2159 /// RecognizePersonality - See if the given exception handling personality 2160 /// function is one that we understand. If so, return a description of it; 2161 /// otherwise return Unknown_Personality. 2162 static Personality_Type RecognizePersonality(Value *Pers) { 2163 Function *F = dyn_cast<Function>(Pers->stripPointerCasts()); 2164 if (!F) 2165 return Unknown_Personality; 2166 return StringSwitch<Personality_Type>(F->getName()) 2167 .Case("__gnat_eh_personality", GNU_Ada_Personality) 2168 .Case("__gxx_personality_v0", GNU_CXX_Personality) 2169 .Case("__objc_personality_v0", GNU_ObjC_Personality) 2170 .Default(Unknown_Personality); 2171 } 2172 2173 /// isCatchAll - Return 'true' if the given typeinfo will match anything. 2174 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) { 2175 switch (Personality) { 2176 case Unknown_Personality: 2177 return false; 2178 case GNU_Ada_Personality: 2179 // While __gnat_all_others_value will match any Ada exception, it doesn't 2180 // match foreign exceptions (or didn't, before gcc-4.7). 2181 return false; 2182 case GNU_CXX_Personality: 2183 case GNU_ObjC_Personality: 2184 return TypeInfo->isNullValue(); 2185 } 2186 llvm_unreachable("Unknown personality!"); 2187 } 2188 2189 static bool shorter_filter(const Value *LHS, const Value *RHS) { 2190 return 2191 cast<ArrayType>(LHS->getType())->getNumElements() 2192 < 2193 cast<ArrayType>(RHS->getType())->getNumElements(); 2194 } 2195 2196 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { 2197 // The logic here should be correct for any real-world personality function. 2198 // However if that turns out not to be true, the offending logic can always 2199 // be conditioned on the personality function, like the catch-all logic is. 2200 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn()); 2201 2202 // Simplify the list of clauses, eg by removing repeated catch clauses 2203 // (these are often created by inlining). 2204 bool MakeNewInstruction = false; // If true, recreate using the following: 2205 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction; 2206 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. 2207 2208 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already. 2209 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) { 2210 bool isLastClause = i + 1 == e; 2211 if (LI.isCatch(i)) { 2212 // A catch clause. 2213 Constant *CatchClause = LI.getClause(i); 2214 Constant *TypeInfo = CatchClause->stripPointerCasts(); 2215 2216 // If we already saw this clause, there is no point in having a second 2217 // copy of it. 2218 if (AlreadyCaught.insert(TypeInfo)) { 2219 // This catch clause was not already seen. 2220 NewClauses.push_back(CatchClause); 2221 } else { 2222 // Repeated catch clause - drop the redundant copy. 2223 MakeNewInstruction = true; 2224 } 2225 2226 // If this is a catch-all then there is no point in keeping any following 2227 // clauses or marking the landingpad as having a cleanup. 2228 if (isCatchAll(Personality, TypeInfo)) { 2229 if (!isLastClause) 2230 MakeNewInstruction = true; 2231 CleanupFlag = false; 2232 break; 2233 } 2234 } else { 2235 // A filter clause. If any of the filter elements were already caught 2236 // then they can be dropped from the filter. It is tempting to try to 2237 // exploit the filter further by saying that any typeinfo that does not 2238 // occur in the filter can't be caught later (and thus can be dropped). 2239 // However this would be wrong, since typeinfos can match without being 2240 // equal (for example if one represents a C++ class, and the other some 2241 // class derived from it). 2242 assert(LI.isFilter(i) && "Unsupported landingpad clause!"); 2243 Constant *FilterClause = LI.getClause(i); 2244 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType()); 2245 unsigned NumTypeInfos = FilterType->getNumElements(); 2246 2247 // An empty filter catches everything, so there is no point in keeping any 2248 // following clauses or marking the landingpad as having a cleanup. By 2249 // dealing with this case here the following code is made a bit simpler. 2250 if (!NumTypeInfos) { 2251 NewClauses.push_back(FilterClause); 2252 if (!isLastClause) 2253 MakeNewInstruction = true; 2254 CleanupFlag = false; 2255 break; 2256 } 2257 2258 bool MakeNewFilter = false; // If true, make a new filter. 2259 SmallVector<Constant *, 16> NewFilterElts; // New elements. 2260 if (isa<ConstantAggregateZero>(FilterClause)) { 2261 // Not an empty filter - it contains at least one null typeinfo. 2262 assert(NumTypeInfos > 0 && "Should have handled empty filter already!"); 2263 Constant *TypeInfo = 2264 Constant::getNullValue(FilterType->getElementType()); 2265 // If this typeinfo is a catch-all then the filter can never match. 2266 if (isCatchAll(Personality, TypeInfo)) { 2267 // Throw the filter away. 2268 MakeNewInstruction = true; 2269 continue; 2270 } 2271 2272 // There is no point in having multiple copies of this typeinfo, so 2273 // discard all but the first copy if there is more than one. 2274 NewFilterElts.push_back(TypeInfo); 2275 if (NumTypeInfos > 1) 2276 MakeNewFilter = true; 2277 } else { 2278 ConstantArray *Filter = cast<ConstantArray>(FilterClause); 2279 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements. 2280 NewFilterElts.reserve(NumTypeInfos); 2281 2282 // Remove any filter elements that were already caught or that already 2283 // occurred in the filter. While there, see if any of the elements are 2284 // catch-alls. If so, the filter can be discarded. 2285 bool SawCatchAll = false; 2286 for (unsigned j = 0; j != NumTypeInfos; ++j) { 2287 Constant *Elt = Filter->getOperand(j); 2288 Constant *TypeInfo = Elt->stripPointerCasts(); 2289 if (isCatchAll(Personality, TypeInfo)) { 2290 // This element is a catch-all. Bail out, noting this fact. 2291 SawCatchAll = true; 2292 break; 2293 } 2294 if (AlreadyCaught.count(TypeInfo)) 2295 // Already caught by an earlier clause, so having it in the filter 2296 // is pointless. 2297 continue; 2298 // There is no point in having multiple copies of the same typeinfo in 2299 // a filter, so only add it if we didn't already. 2300 if (SeenInFilter.insert(TypeInfo)) 2301 NewFilterElts.push_back(cast<Constant>(Elt)); 2302 } 2303 // A filter containing a catch-all cannot match anything by definition. 2304 if (SawCatchAll) { 2305 // Throw the filter away. 2306 MakeNewInstruction = true; 2307 continue; 2308 } 2309 2310 // If we dropped something from the filter, make a new one. 2311 if (NewFilterElts.size() < NumTypeInfos) 2312 MakeNewFilter = true; 2313 } 2314 if (MakeNewFilter) { 2315 FilterType = ArrayType::get(FilterType->getElementType(), 2316 NewFilterElts.size()); 2317 FilterClause = ConstantArray::get(FilterType, NewFilterElts); 2318 MakeNewInstruction = true; 2319 } 2320 2321 NewClauses.push_back(FilterClause); 2322 2323 // If the new filter is empty then it will catch everything so there is 2324 // no point in keeping any following clauses or marking the landingpad 2325 // as having a cleanup. The case of the original filter being empty was 2326 // already handled above. 2327 if (MakeNewFilter && !NewFilterElts.size()) { 2328 assert(MakeNewInstruction && "New filter but not a new instruction!"); 2329 CleanupFlag = false; 2330 break; 2331 } 2332 } 2333 } 2334 2335 // If several filters occur in a row then reorder them so that the shortest 2336 // filters come first (those with the smallest number of elements). This is 2337 // advantageous because shorter filters are more likely to match, speeding up 2338 // unwinding, but mostly because it increases the effectiveness of the other 2339 // filter optimizations below. 2340 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) { 2341 unsigned j; 2342 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters. 2343 for (j = i; j != e; ++j) 2344 if (!isa<ArrayType>(NewClauses[j]->getType())) 2345 break; 2346 2347 // Check whether the filters are already sorted by length. We need to know 2348 // if sorting them is actually going to do anything so that we only make a 2349 // new landingpad instruction if it does. 2350 for (unsigned k = i; k + 1 < j; ++k) 2351 if (shorter_filter(NewClauses[k+1], NewClauses[k])) { 2352 // Not sorted, so sort the filters now. Doing an unstable sort would be 2353 // correct too but reordering filters pointlessly might confuse users. 2354 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j, 2355 shorter_filter); 2356 MakeNewInstruction = true; 2357 break; 2358 } 2359 2360 // Look for the next batch of filters. 2361 i = j + 1; 2362 } 2363 2364 // If typeinfos matched if and only if equal, then the elements of a filter L 2365 // that occurs later than a filter F could be replaced by the intersection of 2366 // the elements of F and L. In reality two typeinfos can match without being 2367 // equal (for example if one represents a C++ class, and the other some class 2368 // derived from it) so it would be wrong to perform this transform in general. 2369 // However the transform is correct and useful if F is a subset of L. In that 2370 // case L can be replaced by F, and thus removed altogether since repeating a 2371 // filter is pointless. So here we look at all pairs of filters F and L where 2372 // L follows F in the list of clauses, and remove L if every element of F is 2373 // an element of L. This can occur when inlining C++ functions with exception 2374 // specifications. 2375 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) { 2376 // Examine each filter in turn. 2377 Value *Filter = NewClauses[i]; 2378 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType()); 2379 if (!FTy) 2380 // Not a filter - skip it. 2381 continue; 2382 unsigned FElts = FTy->getNumElements(); 2383 // Examine each filter following this one. Doing this backwards means that 2384 // we don't have to worry about filters disappearing under us when removed. 2385 for (unsigned j = NewClauses.size() - 1; j != i; --j) { 2386 Value *LFilter = NewClauses[j]; 2387 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType()); 2388 if (!LTy) 2389 // Not a filter - skip it. 2390 continue; 2391 // If Filter is a subset of LFilter, i.e. every element of Filter is also 2392 // an element of LFilter, then discard LFilter. 2393 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j; 2394 // If Filter is empty then it is a subset of LFilter. 2395 if (!FElts) { 2396 // Discard LFilter. 2397 NewClauses.erase(J); 2398 MakeNewInstruction = true; 2399 // Move on to the next filter. 2400 continue; 2401 } 2402 unsigned LElts = LTy->getNumElements(); 2403 // If Filter is longer than LFilter then it cannot be a subset of it. 2404 if (FElts > LElts) 2405 // Move on to the next filter. 2406 continue; 2407 // At this point we know that LFilter has at least one element. 2408 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros. 2409 // Filter is a subset of LFilter iff Filter contains only zeros (as we 2410 // already know that Filter is not longer than LFilter). 2411 if (isa<ConstantAggregateZero>(Filter)) { 2412 assert(FElts <= LElts && "Should have handled this case earlier!"); 2413 // Discard LFilter. 2414 NewClauses.erase(J); 2415 MakeNewInstruction = true; 2416 } 2417 // Move on to the next filter. 2418 continue; 2419 } 2420 ConstantArray *LArray = cast<ConstantArray>(LFilter); 2421 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros. 2422 // Since Filter is non-empty and contains only zeros, it is a subset of 2423 // LFilter iff LFilter contains a zero. 2424 assert(FElts > 0 && "Should have eliminated the empty filter earlier!"); 2425 for (unsigned l = 0; l != LElts; ++l) 2426 if (LArray->getOperand(l)->isNullValue()) { 2427 // LFilter contains a zero - discard it. 2428 NewClauses.erase(J); 2429 MakeNewInstruction = true; 2430 break; 2431 } 2432 // Move on to the next filter. 2433 continue; 2434 } 2435 // At this point we know that both filters are ConstantArrays. Loop over 2436 // operands to see whether every element of Filter is also an element of 2437 // LFilter. Since filters tend to be short this is probably faster than 2438 // using a method that scales nicely. 2439 ConstantArray *FArray = cast<ConstantArray>(Filter); 2440 bool AllFound = true; 2441 for (unsigned f = 0; f != FElts; ++f) { 2442 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts(); 2443 AllFound = false; 2444 for (unsigned l = 0; l != LElts; ++l) { 2445 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts(); 2446 if (LTypeInfo == FTypeInfo) { 2447 AllFound = true; 2448 break; 2449 } 2450 } 2451 if (!AllFound) 2452 break; 2453 } 2454 if (AllFound) { 2455 // Discard LFilter. 2456 NewClauses.erase(J); 2457 MakeNewInstruction = true; 2458 } 2459 // Move on to the next filter. 2460 } 2461 } 2462 2463 // If we changed any of the clauses, replace the old landingpad instruction 2464 // with a new one. 2465 if (MakeNewInstruction) { 2466 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(), 2467 LI.getPersonalityFn(), 2468 NewClauses.size()); 2469 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i) 2470 NLI->addClause(NewClauses[i]); 2471 // A landing pad with no clauses must have the cleanup flag set. It is 2472 // theoretically possible, though highly unlikely, that we eliminated all 2473 // clauses. If so, force the cleanup flag to true. 2474 if (NewClauses.empty()) 2475 CleanupFlag = true; 2476 NLI->setCleanup(CleanupFlag); 2477 return NLI; 2478 } 2479 2480 // Even if none of the clauses changed, we may nonetheless have understood 2481 // that the cleanup flag is pointless. Clear it if so. 2482 if (LI.isCleanup() != CleanupFlag) { 2483 assert(!CleanupFlag && "Adding a cleanup, not removing one?!"); 2484 LI.setCleanup(CleanupFlag); 2485 return &LI; 2486 } 2487 2488 return nullptr; 2489 } 2490 2491 2492 2493 2494 /// TryToSinkInstruction - Try to move the specified instruction from its 2495 /// current block into the beginning of DestBlock, which can only happen if it's 2496 /// safe to move the instruction past all of the instructions between it and the 2497 /// end of its block. 2498 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 2499 assert(I->hasOneUse() && "Invariants didn't hold!"); 2500 2501 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 2502 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() || 2503 isa<TerminatorInst>(I)) 2504 return false; 2505 2506 // Do not sink alloca instructions out of the entry block. 2507 if (isa<AllocaInst>(I) && I->getParent() == 2508 &DestBlock->getParent()->getEntryBlock()) 2509 return false; 2510 2511 // We can only sink load instructions if there is nothing between the load and 2512 // the end of block that could change the value. 2513 if (I->mayReadFromMemory()) { 2514 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 2515 Scan != E; ++Scan) 2516 if (Scan->mayWriteToMemory()) 2517 return false; 2518 } 2519 2520 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt(); 2521 I->moveBefore(InsertPos); 2522 ++NumSunkInst; 2523 return true; 2524 } 2525 2526 2527 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 2528 /// all reachable code to the worklist. 2529 /// 2530 /// This has a couple of tricks to make the code faster and more powerful. In 2531 /// particular, we constant fold and DCE instructions as we go, to avoid adding 2532 /// them to the worklist (this significantly speeds up instcombine on code where 2533 /// many instructions are dead or constant). Additionally, if we find a branch 2534 /// whose condition is a known constant, we only visit the reachable successors. 2535 /// 2536 static bool AddReachableCodeToWorklist(BasicBlock *BB, 2537 SmallPtrSet<BasicBlock*, 64> &Visited, 2538 InstCombiner &IC, 2539 const DataLayout *DL, 2540 const TargetLibraryInfo *TLI) { 2541 bool MadeIRChange = false; 2542 SmallVector<BasicBlock*, 256> Worklist; 2543 Worklist.push_back(BB); 2544 2545 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 2546 DenseMap<ConstantExpr*, Constant*> FoldedConstants; 2547 2548 do { 2549 BB = Worklist.pop_back_val(); 2550 2551 // We have now visited this block! If we've already been here, ignore it. 2552 if (!Visited.insert(BB)) continue; 2553 2554 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 2555 Instruction *Inst = BBI++; 2556 2557 // DCE instruction if trivially dead. 2558 if (isInstructionTriviallyDead(Inst, TLI)) { 2559 ++NumDeadInst; 2560 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); 2561 Inst->eraseFromParent(); 2562 continue; 2563 } 2564 2565 // ConstantProp instruction if trivially constant. 2566 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 2567 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) { 2568 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " 2569 << *Inst << '\n'); 2570 Inst->replaceAllUsesWith(C); 2571 ++NumConstProp; 2572 Inst->eraseFromParent(); 2573 continue; 2574 } 2575 2576 if (DL) { 2577 // See if we can constant fold its operands. 2578 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 2579 i != e; ++i) { 2580 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 2581 if (CE == nullptr) continue; 2582 2583 Constant*& FoldRes = FoldedConstants[CE]; 2584 if (!FoldRes) 2585 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); 2586 if (!FoldRes) 2587 FoldRes = CE; 2588 2589 if (FoldRes != CE) { 2590 *i = FoldRes; 2591 MadeIRChange = true; 2592 } 2593 } 2594 } 2595 2596 InstrsForInstCombineWorklist.push_back(Inst); 2597 } 2598 2599 // Recursively visit successors. If this is a branch or switch on a 2600 // constant, only visit the reachable successor. 2601 TerminatorInst *TI = BB->getTerminator(); 2602 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2603 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 2604 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 2605 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 2606 Worklist.push_back(ReachableBB); 2607 continue; 2608 } 2609 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2610 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 2611 // See if this is an explicit destination. 2612 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2613 i != e; ++i) 2614 if (i.getCaseValue() == Cond) { 2615 BasicBlock *ReachableBB = i.getCaseSuccessor(); 2616 Worklist.push_back(ReachableBB); 2617 continue; 2618 } 2619 2620 // Otherwise it is the default destination. 2621 Worklist.push_back(SI->getDefaultDest()); 2622 continue; 2623 } 2624 } 2625 2626 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 2627 Worklist.push_back(TI->getSuccessor(i)); 2628 } while (!Worklist.empty()); 2629 2630 // Once we've found all of the instructions to add to instcombine's worklist, 2631 // add them in reverse order. This way instcombine will visit from the top 2632 // of the function down. This jives well with the way that it adds all uses 2633 // of instructions to the worklist after doing a transformation, thus avoiding 2634 // some N^2 behavior in pathological cases. 2635 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 2636 InstrsForInstCombineWorklist.size()); 2637 2638 return MadeIRChange; 2639 } 2640 2641 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 2642 MadeIRChange = false; 2643 2644 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 2645 << F.getName() << "\n"); 2646 2647 { 2648 // Do a depth-first traversal of the function, populate the worklist with 2649 // the reachable instructions. Ignore blocks that are not reachable. Keep 2650 // track of which blocks we visit. 2651 SmallPtrSet<BasicBlock*, 64> Visited; 2652 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL, 2653 TLI); 2654 2655 // Do a quick scan over the function. If we find any blocks that are 2656 // unreachable, remove any instructions inside of them. This prevents 2657 // the instcombine code from having to deal with some bad special cases. 2658 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 2659 if (Visited.count(BB)) continue; 2660 2661 // Delete the instructions backwards, as it has a reduced likelihood of 2662 // having to update as many def-use and use-def chains. 2663 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 2664 while (EndInst != BB->begin()) { 2665 // Delete the next to last instruction. 2666 BasicBlock::iterator I = EndInst; 2667 Instruction *Inst = --I; 2668 if (!Inst->use_empty()) 2669 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 2670 if (isa<LandingPadInst>(Inst)) { 2671 EndInst = Inst; 2672 continue; 2673 } 2674 if (!isa<DbgInfoIntrinsic>(Inst)) { 2675 ++NumDeadInst; 2676 MadeIRChange = true; 2677 } 2678 Inst->eraseFromParent(); 2679 } 2680 } 2681 } 2682 2683 while (!Worklist.isEmpty()) { 2684 Instruction *I = Worklist.RemoveOne(); 2685 if (I == nullptr) continue; // skip null values. 2686 2687 // Check to see if we can DCE the instruction. 2688 if (isInstructionTriviallyDead(I, TLI)) { 2689 DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); 2690 EraseInstFromFunction(*I); 2691 ++NumDeadInst; 2692 MadeIRChange = true; 2693 continue; 2694 } 2695 2696 // Instruction isn't dead, see if we can constant propagate it. 2697 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 2698 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) { 2699 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 2700 2701 // Add operands to the worklist. 2702 ReplaceInstUsesWith(*I, C); 2703 ++NumConstProp; 2704 EraseInstFromFunction(*I); 2705 MadeIRChange = true; 2706 continue; 2707 } 2708 2709 // See if we can trivially sink this instruction to a successor basic block. 2710 if (I->hasOneUse()) { 2711 BasicBlock *BB = I->getParent(); 2712 Instruction *UserInst = cast<Instruction>(*I->user_begin()); 2713 BasicBlock *UserParent; 2714 2715 // Get the block the use occurs in. 2716 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 2717 UserParent = PN->getIncomingBlock(*I->use_begin()); 2718 else 2719 UserParent = UserInst->getParent(); 2720 2721 if (UserParent != BB) { 2722 bool UserIsSuccessor = false; 2723 // See if the user is one of our successors. 2724 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 2725 if (*SI == UserParent) { 2726 UserIsSuccessor = true; 2727 break; 2728 } 2729 2730 // If the user is one of our immediate successors, and if that successor 2731 // only has us as a predecessors (we'd have to split the critical edge 2732 // otherwise), we can keep going. 2733 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 2734 // Okay, the CFG is simple enough, try to sink this instruction. 2735 MadeIRChange |= TryToSinkInstruction(I, UserParent); 2736 } 2737 } 2738 2739 // Now that we have an instruction, try combining it to simplify it. 2740 Builder->SetInsertPoint(I->getParent(), I); 2741 Builder->SetCurrentDebugLocation(I->getDebugLoc()); 2742 2743 #ifndef NDEBUG 2744 std::string OrigI; 2745 #endif 2746 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 2747 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); 2748 2749 if (Instruction *Result = visit(*I)) { 2750 ++NumCombined; 2751 // Should we replace the old instruction with a new one? 2752 if (Result != I) { 2753 DEBUG(dbgs() << "IC: Old = " << *I << '\n' 2754 << " New = " << *Result << '\n'); 2755 2756 if (!I->getDebugLoc().isUnknown()) 2757 Result->setDebugLoc(I->getDebugLoc()); 2758 // Everything uses the new instruction now. 2759 I->replaceAllUsesWith(Result); 2760 2761 // Move the name to the new instruction first. 2762 Result->takeName(I); 2763 2764 // Push the new instruction and any users onto the worklist. 2765 Worklist.Add(Result); 2766 Worklist.AddUsersToWorkList(*Result); 2767 2768 // Insert the new instruction into the basic block... 2769 BasicBlock *InstParent = I->getParent(); 2770 BasicBlock::iterator InsertPos = I; 2771 2772 // If we replace a PHI with something that isn't a PHI, fix up the 2773 // insertion point. 2774 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos)) 2775 InsertPos = InstParent->getFirstInsertionPt(); 2776 2777 InstParent->getInstList().insert(InsertPos, Result); 2778 2779 EraseInstFromFunction(*I); 2780 } else { 2781 #ifndef NDEBUG 2782 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' 2783 << " New = " << *I << '\n'); 2784 #endif 2785 2786 // If the instruction was modified, it's possible that it is now dead. 2787 // if so, remove it. 2788 if (isInstructionTriviallyDead(I, TLI)) { 2789 EraseInstFromFunction(*I); 2790 } else { 2791 Worklist.Add(I); 2792 Worklist.AddUsersToWorkList(*I); 2793 } 2794 } 2795 MadeIRChange = true; 2796 } 2797 } 2798 2799 Worklist.Zap(); 2800 return MadeIRChange; 2801 } 2802 2803 namespace { 2804 class InstCombinerLibCallSimplifier : public LibCallSimplifier { 2805 InstCombiner *IC; 2806 public: 2807 InstCombinerLibCallSimplifier(const DataLayout *DL, 2808 const TargetLibraryInfo *TLI, 2809 InstCombiner *IC) 2810 : LibCallSimplifier(DL, TLI, UnsafeFPShrink) { 2811 this->IC = IC; 2812 } 2813 2814 /// replaceAllUsesWith - override so that instruction replacement 2815 /// can be defined in terms of the instruction combiner framework. 2816 void replaceAllUsesWith(Instruction *I, Value *With) const override { 2817 IC->ReplaceInstUsesWith(*I, With); 2818 } 2819 }; 2820 } 2821 2822 bool InstCombiner::runOnFunction(Function &F) { 2823 if (skipOptnoneFunction(F)) 2824 return false; 2825 2826 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 2827 DL = DLP ? &DLP->getDataLayout() : nullptr; 2828 TLI = &getAnalysis<TargetLibraryInfo>(); 2829 // Minimizing size? 2830 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 2831 Attribute::MinSize); 2832 2833 /// Builder - This is an IRBuilder that automatically inserts new 2834 /// instructions into the worklist when they are created. 2835 IRBuilder<true, TargetFolder, InstCombineIRInserter> 2836 TheBuilder(F.getContext(), TargetFolder(DL), 2837 InstCombineIRInserter(Worklist)); 2838 Builder = &TheBuilder; 2839 2840 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this); 2841 Simplifier = &TheSimplifier; 2842 2843 bool EverMadeChange = false; 2844 2845 // Lower dbg.declare intrinsics otherwise their value may be clobbered 2846 // by instcombiner. 2847 EverMadeChange = LowerDbgDeclare(F); 2848 2849 // Iterate while there is work to do. 2850 unsigned Iteration = 0; 2851 while (DoOneIteration(F, Iteration++)) 2852 EverMadeChange = true; 2853 2854 Builder = nullptr; 2855 return EverMadeChange; 2856 } 2857 2858 FunctionPass *llvm::createInstructionCombiningPass() { 2859 return new InstCombiner(); 2860 } 2861