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/InstCombine/InstCombine.h" 37 #include "InstCombineInternal.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/AssumptionCache.h" 43 #include "llvm/Analysis/CFG.h" 44 #include "llvm/Analysis/ConstantFolding.h" 45 #include "llvm/Analysis/InstructionSimplify.h" 46 #include "llvm/Analysis/LibCallSemantics.h" 47 #include "llvm/Analysis/LoopInfo.h" 48 #include "llvm/Analysis/MemoryBuiltins.h" 49 #include "llvm/Analysis/TargetLibraryInfo.h" 50 #include "llvm/Analysis/ValueTracking.h" 51 #include "llvm/IR/CFG.h" 52 #include "llvm/IR/DataLayout.h" 53 #include "llvm/IR/Dominators.h" 54 #include "llvm/IR/GetElementPtrTypeIterator.h" 55 #include "llvm/IR/IntrinsicInst.h" 56 #include "llvm/IR/PatternMatch.h" 57 #include "llvm/IR/ValueHandle.h" 58 #include "llvm/Support/CommandLine.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Support/raw_ostream.h" 61 #include "llvm/Transforms/Scalar.h" 62 #include "llvm/Transforms/Utils/Local.h" 63 #include <algorithm> 64 #include <climits> 65 using namespace llvm; 66 using namespace llvm::PatternMatch; 67 68 #define DEBUG_TYPE "instcombine" 69 70 STATISTIC(NumCombined , "Number of insts combined"); 71 STATISTIC(NumConstProp, "Number of constant folds"); 72 STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 73 STATISTIC(NumSunkInst , "Number of instructions sunk"); 74 STATISTIC(NumExpand, "Number of expansions"); 75 STATISTIC(NumFactor , "Number of factorizations"); 76 STATISTIC(NumReassoc , "Number of reassociations"); 77 78 Value *InstCombiner::EmitGEPOffset(User *GEP) { 79 return llvm::EmitGEPOffset(Builder, DL, GEP); 80 } 81 82 /// ShouldChangeType - Return true if it is desirable to convert a computation 83 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal 84 /// type for example, or from a smaller to a larger illegal type. 85 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const { 86 assert(From->isIntegerTy() && To->isIntegerTy()); 87 88 unsigned FromWidth = From->getPrimitiveSizeInBits(); 89 unsigned ToWidth = To->getPrimitiveSizeInBits(); 90 bool FromLegal = DL.isLegalInteger(FromWidth); 91 bool ToLegal = DL.isLegalInteger(ToWidth); 92 93 // If this is a legal integer from type, and the result would be an illegal 94 // type, don't do the transformation. 95 if (FromLegal && !ToLegal) 96 return false; 97 98 // Otherwise, if both are illegal, do not increase the size of the result. We 99 // do allow things like i160 -> i64, but not i64 -> i160. 100 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 101 return false; 102 103 return true; 104 } 105 106 // Return true, if No Signed Wrap should be maintained for I. 107 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C", 108 // where both B and C should be ConstantInts, results in a constant that does 109 // not overflow. This function only handles the Add and Sub opcodes. For 110 // all other opcodes, the function conservatively returns false. 111 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { 112 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I); 113 if (!OBO || !OBO->hasNoSignedWrap()) { 114 return false; 115 } 116 117 // We reason about Add and Sub Only. 118 Instruction::BinaryOps Opcode = I.getOpcode(); 119 if (Opcode != Instruction::Add && 120 Opcode != Instruction::Sub) { 121 return false; 122 } 123 124 ConstantInt *CB = dyn_cast<ConstantInt>(B); 125 ConstantInt *CC = dyn_cast<ConstantInt>(C); 126 127 if (!CB || !CC) { 128 return false; 129 } 130 131 const APInt &BVal = CB->getValue(); 132 const APInt &CVal = CC->getValue(); 133 bool Overflow = false; 134 135 if (Opcode == Instruction::Add) { 136 BVal.sadd_ov(CVal, Overflow); 137 } else { 138 BVal.ssub_ov(CVal, Overflow); 139 } 140 141 return !Overflow; 142 } 143 144 /// Conservatively clears subclassOptionalData after a reassociation or 145 /// commutation. We preserve fast-math flags when applicable as they can be 146 /// preserved. 147 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { 148 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I); 149 if (!FPMO) { 150 I.clearSubclassOptionalData(); 151 return; 152 } 153 154 FastMathFlags FMF = I.getFastMathFlags(); 155 I.clearSubclassOptionalData(); 156 I.setFastMathFlags(FMF); 157 } 158 159 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for 160 /// operators which are associative or commutative: 161 // 162 // Commutative operators: 163 // 164 // 1. Order operands such that they are listed from right (least complex) to 165 // left (most complex). This puts constants before unary operators before 166 // binary operators. 167 // 168 // Associative operators: 169 // 170 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 171 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 172 // 173 // Associative and commutative operators: 174 // 175 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 176 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 177 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 178 // if C1 and C2 are constants. 179 // 180 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 181 Instruction::BinaryOps Opcode = I.getOpcode(); 182 bool Changed = false; 183 184 do { 185 // Order operands such that they are listed from right (least complex) to 186 // left (most complex). This puts constants before unary operators before 187 // binary operators. 188 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 189 getComplexity(I.getOperand(1))) 190 Changed = !I.swapOperands(); 191 192 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 193 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 194 195 if (I.isAssociative()) { 196 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 197 if (Op0 && Op0->getOpcode() == Opcode) { 198 Value *A = Op0->getOperand(0); 199 Value *B = Op0->getOperand(1); 200 Value *C = I.getOperand(1); 201 202 // Does "B op C" simplify? 203 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) { 204 // It simplifies to V. Form "A op V". 205 I.setOperand(0, A); 206 I.setOperand(1, V); 207 // Conservatively clear the optional flags, since they may not be 208 // preserved by the reassociation. 209 if (MaintainNoSignedWrap(I, B, C) && 210 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) { 211 // Note: this is only valid because SimplifyBinOp doesn't look at 212 // the operands to Op0. 213 I.clearSubclassOptionalData(); 214 I.setHasNoSignedWrap(true); 215 } else { 216 ClearSubclassDataAfterReassociation(I); 217 } 218 219 Changed = true; 220 ++NumReassoc; 221 continue; 222 } 223 } 224 225 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 226 if (Op1 && Op1->getOpcode() == Opcode) { 227 Value *A = I.getOperand(0); 228 Value *B = Op1->getOperand(0); 229 Value *C = Op1->getOperand(1); 230 231 // Does "A op B" simplify? 232 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) { 233 // It simplifies to V. Form "V op C". 234 I.setOperand(0, V); 235 I.setOperand(1, C); 236 // Conservatively clear the optional flags, since they may not be 237 // preserved by the reassociation. 238 ClearSubclassDataAfterReassociation(I); 239 Changed = true; 240 ++NumReassoc; 241 continue; 242 } 243 } 244 } 245 246 if (I.isAssociative() && I.isCommutative()) { 247 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 248 if (Op0 && Op0->getOpcode() == Opcode) { 249 Value *A = Op0->getOperand(0); 250 Value *B = Op0->getOperand(1); 251 Value *C = I.getOperand(1); 252 253 // Does "C op A" simplify? 254 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 255 // It simplifies to V. Form "V op B". 256 I.setOperand(0, V); 257 I.setOperand(1, B); 258 // Conservatively clear the optional flags, since they may not be 259 // preserved by the reassociation. 260 ClearSubclassDataAfterReassociation(I); 261 Changed = true; 262 ++NumReassoc; 263 continue; 264 } 265 } 266 267 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 268 if (Op1 && Op1->getOpcode() == Opcode) { 269 Value *A = I.getOperand(0); 270 Value *B = Op1->getOperand(0); 271 Value *C = Op1->getOperand(1); 272 273 // Does "C op A" simplify? 274 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 275 // It simplifies to V. Form "B op V". 276 I.setOperand(0, B); 277 I.setOperand(1, V); 278 // Conservatively clear the optional flags, since they may not be 279 // preserved by the reassociation. 280 ClearSubclassDataAfterReassociation(I); 281 Changed = true; 282 ++NumReassoc; 283 continue; 284 } 285 } 286 287 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 288 // if C1 and C2 are constants. 289 if (Op0 && Op1 && 290 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 291 isa<Constant>(Op0->getOperand(1)) && 292 isa<Constant>(Op1->getOperand(1)) && 293 Op0->hasOneUse() && Op1->hasOneUse()) { 294 Value *A = Op0->getOperand(0); 295 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 296 Value *B = Op1->getOperand(0); 297 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 298 299 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 300 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B); 301 if (isa<FPMathOperator>(New)) { 302 FastMathFlags Flags = I.getFastMathFlags(); 303 Flags &= Op0->getFastMathFlags(); 304 Flags &= Op1->getFastMathFlags(); 305 New->setFastMathFlags(Flags); 306 } 307 InsertNewInstWith(New, I); 308 New->takeName(Op1); 309 I.setOperand(0, New); 310 I.setOperand(1, Folded); 311 // Conservatively clear the optional flags, since they may not be 312 // preserved by the reassociation. 313 ClearSubclassDataAfterReassociation(I); 314 315 Changed = true; 316 continue; 317 } 318 } 319 320 // No further simplifications. 321 return Changed; 322 } while (1); 323 } 324 325 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to 326 /// "(X LOp Y) ROp (X LOp Z)". 327 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp, 328 Instruction::BinaryOps ROp) { 329 switch (LOp) { 330 default: 331 return false; 332 333 case Instruction::And: 334 // And distributes over Or and Xor. 335 switch (ROp) { 336 default: 337 return false; 338 case Instruction::Or: 339 case Instruction::Xor: 340 return true; 341 } 342 343 case Instruction::Mul: 344 // Multiplication distributes over addition and subtraction. 345 switch (ROp) { 346 default: 347 return false; 348 case Instruction::Add: 349 case Instruction::Sub: 350 return true; 351 } 352 353 case Instruction::Or: 354 // Or distributes over And. 355 switch (ROp) { 356 default: 357 return false; 358 case Instruction::And: 359 return true; 360 } 361 } 362 } 363 364 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to 365 /// "(X ROp Z) LOp (Y ROp Z)". 366 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, 367 Instruction::BinaryOps ROp) { 368 if (Instruction::isCommutative(ROp)) 369 return LeftDistributesOverRight(ROp, LOp); 370 371 switch (LOp) { 372 default: 373 return false; 374 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. 375 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts. 376 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts. 377 case Instruction::And: 378 case Instruction::Or: 379 case Instruction::Xor: 380 switch (ROp) { 381 default: 382 return false; 383 case Instruction::Shl: 384 case Instruction::LShr: 385 case Instruction::AShr: 386 return true; 387 } 388 } 389 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", 390 // but this requires knowing that the addition does not overflow and other 391 // such subtleties. 392 return false; 393 } 394 395 /// This function returns identity value for given opcode, which can be used to 396 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1). 397 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) { 398 if (isa<Constant>(V)) 399 return nullptr; 400 401 if (OpCode == Instruction::Mul) 402 return ConstantInt::get(V->getType(), 1); 403 404 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc. 405 406 return nullptr; 407 } 408 409 /// This function factors binary ops which can be combined using distributive 410 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable 411 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called 412 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms 413 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and 414 /// RHS to 4. 415 static Instruction::BinaryOps 416 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode, 417 BinaryOperator *Op, Value *&LHS, Value *&RHS) { 418 if (!Op) 419 return Instruction::BinaryOpsEnd; 420 421 LHS = Op->getOperand(0); 422 RHS = Op->getOperand(1); 423 424 switch (TopLevelOpcode) { 425 default: 426 return Op->getOpcode(); 427 428 case Instruction::Add: 429 case Instruction::Sub: 430 if (Op->getOpcode() == Instruction::Shl) { 431 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) { 432 // The multiplier is really 1 << CST. 433 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST); 434 return Instruction::Mul; 435 } 436 } 437 return Op->getOpcode(); 438 } 439 440 // TODO: We can add other conversions e.g. shr => div etc. 441 } 442 443 /// This tries to simplify binary operations by factorizing out common terms 444 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 445 static Value *tryFactorization(InstCombiner::BuilderTy *Builder, 446 const DataLayout &DL, BinaryOperator &I, 447 Instruction::BinaryOps InnerOpcode, Value *A, 448 Value *B, Value *C, Value *D) { 449 450 // If any of A, B, C, D are null, we can not factor I, return early. 451 // Checking A and C should be enough. 452 if (!A || !C || !B || !D) 453 return nullptr; 454 455 Value *SimplifiedInst = nullptr; 456 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 457 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); 458 459 // Does "X op' Y" always equal "Y op' X"? 460 bool InnerCommutative = Instruction::isCommutative(InnerOpcode); 461 462 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? 463 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) 464 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the 465 // commutative case, "(A op' B) op (C op' A)"? 466 if (A == C || (InnerCommutative && A == D)) { 467 if (A != C) 468 std::swap(C, D); 469 // Consider forming "A op' (B op D)". 470 // If "B op D" simplifies then it can be formed with no cost. 471 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL); 472 // If "B op D" doesn't simplify then only go on if both of the existing 473 // operations "A op' B" and "C op' D" will be zapped as no longer used. 474 if (!V && LHS->hasOneUse() && RHS->hasOneUse()) 475 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName()); 476 if (V) { 477 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V); 478 } 479 } 480 481 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? 482 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) 483 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the 484 // commutative case, "(A op' B) op (B op' D)"? 485 if (B == D || (InnerCommutative && B == C)) { 486 if (B != D) 487 std::swap(C, D); 488 // Consider forming "(A op C) op' B". 489 // If "A op C" simplifies then it can be formed with no cost. 490 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL); 491 492 // If "A op C" doesn't simplify then only go on if both of the existing 493 // operations "A op' B" and "C op' D" will be zapped as no longer used. 494 if (!V && LHS->hasOneUse() && RHS->hasOneUse()) 495 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName()); 496 if (V) { 497 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B); 498 } 499 } 500 501 if (SimplifiedInst) { 502 ++NumFactor; 503 SimplifiedInst->takeName(&I); 504 505 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag. 506 // TODO: Check for NUW. 507 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) { 508 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) { 509 bool HasNSW = false; 510 if (isa<OverflowingBinaryOperator>(&I)) 511 HasNSW = I.hasNoSignedWrap(); 512 513 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 514 if (isa<OverflowingBinaryOperator>(Op0)) 515 HasNSW &= Op0->hasNoSignedWrap(); 516 517 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 518 if (isa<OverflowingBinaryOperator>(Op1)) 519 HasNSW &= Op1->hasNoSignedWrap(); 520 BO->setHasNoSignedWrap(HasNSW); 521 } 522 } 523 } 524 return SimplifiedInst; 525 } 526 527 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations 528 /// which some other binary operation distributes over either by factorizing 529 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this 530 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is 531 /// a win). Returns the simplified value, or null if it didn't simplify. 532 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { 533 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 534 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 535 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 536 537 // Factorization. 538 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 539 auto TopLevelOpcode = I.getOpcode(); 540 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B); 541 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D); 542 543 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize 544 // a common term. 545 if (LHSOpcode == RHSOpcode) { 546 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D)) 547 return V; 548 } 549 550 // The instruction has the form "(A op' B) op (C)". Try to factorize common 551 // term. 552 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS, 553 getIdentityValue(LHSOpcode, RHS))) 554 return V; 555 556 // The instruction has the form "(B) op (C op' D)". Try to factorize common 557 // term. 558 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS, 559 getIdentityValue(RHSOpcode, LHS), C, D)) 560 return V; 561 562 // Expansion. 563 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { 564 // The instruction has the form "(A op' B) op C". See if expanding it out 565 // to "(A op C) op' (B op C)" results in simplifications. 566 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 567 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 568 569 // Do "A op C" and "B op C" both simplify? 570 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL)) 571 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) { 572 // They do! Return "L op' R". 573 ++NumExpand; 574 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 575 if ((L == A && R == B) || 576 (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) 577 return Op0; 578 // Otherwise return "L op' R" if it simplifies. 579 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 580 return V; 581 // Otherwise, create a new instruction. 582 C = Builder->CreateBinOp(InnerOpcode, L, R); 583 C->takeName(&I); 584 return C; 585 } 586 } 587 588 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) { 589 // The instruction has the form "A op (B op' C)". See if expanding it out 590 // to "(A op B) op' (A op C)" results in simplifications. 591 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 592 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' 593 594 // Do "A op B" and "A op C" both simplify? 595 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL)) 596 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) { 597 // They do! Return "L op' R". 598 ++NumExpand; 599 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 600 if ((L == B && R == C) || 601 (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) 602 return Op1; 603 // Otherwise return "L op' R" if it simplifies. 604 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 605 return V; 606 // Otherwise, create a new instruction. 607 A = Builder->CreateBinOp(InnerOpcode, L, R); 608 A->takeName(&I); 609 return A; 610 } 611 } 612 613 return nullptr; 614 } 615 616 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 617 // if the LHS is a constant zero (which is the 'negate' form). 618 // 619 Value *InstCombiner::dyn_castNegVal(Value *V) const { 620 if (BinaryOperator::isNeg(V)) 621 return BinaryOperator::getNegArgument(V); 622 623 // Constants can be considered to be negated values if they can be folded. 624 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 625 return ConstantExpr::getNeg(C); 626 627 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 628 if (C->getType()->getElementType()->isIntegerTy()) 629 return ConstantExpr::getNeg(C); 630 631 return nullptr; 632 } 633 634 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 635 // instruction if the LHS is a constant negative zero (which is the 'negate' 636 // form). 637 // 638 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const { 639 if (BinaryOperator::isFNeg(V, IgnoreZeroSign)) 640 return BinaryOperator::getFNegArgument(V); 641 642 // Constants can be considered to be negated values if they can be folded. 643 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 644 return ConstantExpr::getFNeg(C); 645 646 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 647 if (C->getType()->getElementType()->isFloatingPointTy()) 648 return ConstantExpr::getFNeg(C); 649 650 return nullptr; 651 } 652 653 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 654 InstCombiner *IC) { 655 if (CastInst *CI = dyn_cast<CastInst>(&I)) { 656 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 657 } 658 659 // Figure out if the constant is the left or the right argument. 660 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 661 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 662 663 if (Constant *SOC = dyn_cast<Constant>(SO)) { 664 if (ConstIsRHS) 665 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 666 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 667 } 668 669 Value *Op0 = SO, *Op1 = ConstOperand; 670 if (!ConstIsRHS) 671 std::swap(Op0, Op1); 672 673 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) { 674 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 675 SO->getName()+".op"); 676 Instruction *FPInst = dyn_cast<Instruction>(RI); 677 if (FPInst && isa<FPMathOperator>(FPInst)) 678 FPInst->copyFastMathFlags(BO); 679 return RI; 680 } 681 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 682 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 683 SO->getName()+".cmp"); 684 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 685 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 686 SO->getName()+".cmp"); 687 llvm_unreachable("Unknown binary instruction type!"); 688 } 689 690 // FoldOpIntoSelect - Given an instruction with a select as one operand and a 691 // constant as the other operand, try to fold the binary operator into the 692 // select arguments. This also works for Cast instructions, which obviously do 693 // not have a second operand. 694 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 695 // Don't modify shared select instructions 696 if (!SI->hasOneUse()) return nullptr; 697 Value *TV = SI->getOperand(1); 698 Value *FV = SI->getOperand(2); 699 700 if (isa<Constant>(TV) || isa<Constant>(FV)) { 701 // Bool selects with constant operands can be folded to logical ops. 702 if (SI->getType()->isIntegerTy(1)) return nullptr; 703 704 // If it's a bitcast involving vectors, make sure it has the same number of 705 // elements on both sides. 706 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) { 707 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy()); 708 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy()); 709 710 // Verify that either both or neither are vectors. 711 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr; 712 // If vectors, verify that they have the same number of elements. 713 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) 714 return nullptr; 715 } 716 717 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 718 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 719 720 return SelectInst::Create(SI->getCondition(), 721 SelectTrueVal, SelectFalseVal); 722 } 723 return nullptr; 724 } 725 726 727 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 728 /// has a PHI node as operand #0, see if we can fold the instruction into the 729 /// PHI (which is only possible if all operands to the PHI are constants). 730 /// 731 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { 732 PHINode *PN = cast<PHINode>(I.getOperand(0)); 733 unsigned NumPHIValues = PN->getNumIncomingValues(); 734 if (NumPHIValues == 0) 735 return nullptr; 736 737 // We normally only transform phis with a single use. However, if a PHI has 738 // multiple uses and they are all the same operation, we can fold *all* of the 739 // uses into the PHI. 740 if (!PN->hasOneUse()) { 741 // Walk the use list for the instruction, comparing them to I. 742 for (User *U : PN->users()) { 743 Instruction *UI = cast<Instruction>(U); 744 if (UI != &I && !I.isIdenticalTo(UI)) 745 return nullptr; 746 } 747 // Otherwise, we can replace *all* users with the new PHI we form. 748 } 749 750 // Check to see if all of the operands of the PHI are simple constants 751 // (constantint/constantfp/undef). If there is one non-constant value, 752 // remember the BB it is in. If there is more than one or if *it* is a PHI, 753 // bail out. We don't do arbitrary constant expressions here because moving 754 // their computation can be expensive without a cost model. 755 BasicBlock *NonConstBB = nullptr; 756 for (unsigned i = 0; i != NumPHIValues; ++i) { 757 Value *InVal = PN->getIncomingValue(i); 758 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal)) 759 continue; 760 761 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi. 762 if (NonConstBB) return nullptr; // More than one non-const value. 763 764 NonConstBB = PN->getIncomingBlock(i); 765 766 // If the InVal is an invoke at the end of the pred block, then we can't 767 // insert a computation after it without breaking the edge. 768 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 769 if (II->getParent() == NonConstBB) 770 return nullptr; 771 772 // If the incoming non-constant value is in I's block, we will remove one 773 // instruction, but insert another equivalent one, leading to infinite 774 // instcombine. 775 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI)) 776 return nullptr; 777 } 778 779 // If there is exactly one non-constant value, we can insert a copy of the 780 // operation in that block. However, if this is a critical edge, we would be 781 // inserting the computation on some other paths (e.g. inside a loop). Only 782 // do this if the pred block is unconditionally branching into the phi block. 783 if (NonConstBB != nullptr) { 784 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 785 if (!BI || !BI->isUnconditional()) return nullptr; 786 } 787 788 // Okay, we can do the transformation: create the new PHI node. 789 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues()); 790 InsertNewInstBefore(NewPN, *PN); 791 NewPN->takeName(PN); 792 793 // If we are going to have to insert a new computation, do so right before the 794 // predecessors terminator. 795 if (NonConstBB) 796 Builder->SetInsertPoint(NonConstBB->getTerminator()); 797 798 // Next, add all of the operands to the PHI. 799 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 800 // We only currently try to fold the condition of a select when it is a phi, 801 // not the true/false values. 802 Value *TrueV = SI->getTrueValue(); 803 Value *FalseV = SI->getFalseValue(); 804 BasicBlock *PhiTransBB = PN->getParent(); 805 for (unsigned i = 0; i != NumPHIValues; ++i) { 806 BasicBlock *ThisBB = PN->getIncomingBlock(i); 807 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 808 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 809 Value *InV = nullptr; 810 // Beware of ConstantExpr: it may eventually evaluate to getNullValue, 811 // even if currently isNullValue gives false. 812 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)); 813 if (InC && !isa<ConstantExpr>(InC)) 814 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 815 else 816 InV = Builder->CreateSelect(PN->getIncomingValue(i), 817 TrueVInPred, FalseVInPred, "phitmp"); 818 NewPN->addIncoming(InV, ThisBB); 819 } 820 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 821 Constant *C = cast<Constant>(I.getOperand(1)); 822 for (unsigned i = 0; i != NumPHIValues; ++i) { 823 Value *InV = nullptr; 824 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 825 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 826 else if (isa<ICmpInst>(CI)) 827 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 828 C, "phitmp"); 829 else 830 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 831 C, "phitmp"); 832 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 833 } 834 } else if (I.getNumOperands() == 2) { 835 Constant *C = cast<Constant>(I.getOperand(1)); 836 for (unsigned i = 0; i != NumPHIValues; ++i) { 837 Value *InV = nullptr; 838 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 839 InV = ConstantExpr::get(I.getOpcode(), InC, C); 840 else 841 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 842 PN->getIncomingValue(i), C, "phitmp"); 843 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 844 } 845 } else { 846 CastInst *CI = cast<CastInst>(&I); 847 Type *RetTy = CI->getType(); 848 for (unsigned i = 0; i != NumPHIValues; ++i) { 849 Value *InV; 850 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 851 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 852 else 853 InV = Builder->CreateCast(CI->getOpcode(), 854 PN->getIncomingValue(i), I.getType(), "phitmp"); 855 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 856 } 857 } 858 859 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { 860 Instruction *User = cast<Instruction>(*UI++); 861 if (User == &I) continue; 862 ReplaceInstUsesWith(*User, NewPN); 863 EraseInstFromFunction(*User); 864 } 865 return ReplaceInstUsesWith(I, NewPN); 866 } 867 868 /// FindElementAtOffset - Given a pointer type and a constant offset, determine 869 /// whether or not there is a sequence of GEP indices into the pointed type that 870 /// will land us at the specified offset. If so, fill them into NewIndices and 871 /// return the resultant element type, otherwise return null. 872 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset, 873 SmallVectorImpl<Value *> &NewIndices) { 874 Type *Ty = PtrTy->getElementType(); 875 if (!Ty->isSized()) 876 return nullptr; 877 878 // Start with the index over the outer type. Note that the type size 879 // might be zero (even if the offset isn't zero) if the indexed type 880 // is something like [0 x {int, int}] 881 Type *IntPtrTy = DL.getIntPtrType(PtrTy); 882 int64_t FirstIdx = 0; 883 if (int64_t TySize = DL.getTypeAllocSize(Ty)) { 884 FirstIdx = Offset/TySize; 885 Offset -= FirstIdx*TySize; 886 887 // Handle hosts where % returns negative instead of values [0..TySize). 888 if (Offset < 0) { 889 --FirstIdx; 890 Offset += TySize; 891 assert(Offset >= 0); 892 } 893 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 894 } 895 896 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 897 898 // Index into the types. If we fail, set OrigBase to null. 899 while (Offset) { 900 // Indexing into tail padding between struct/array elements. 901 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty)) 902 return nullptr; 903 904 if (StructType *STy = dyn_cast<StructType>(Ty)) { 905 const StructLayout *SL = DL.getStructLayout(STy); 906 assert(Offset < (int64_t)SL->getSizeInBytes() && 907 "Offset must stay within the indexed type"); 908 909 unsigned Elt = SL->getElementContainingOffset(Offset); 910 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 911 Elt)); 912 913 Offset -= SL->getElementOffset(Elt); 914 Ty = STy->getElementType(Elt); 915 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 916 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType()); 917 assert(EltSize && "Cannot index into a zero-sized array"); 918 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 919 Offset %= EltSize; 920 Ty = AT->getElementType(); 921 } else { 922 // Otherwise, we can't index into the middle of this atomic type, bail. 923 return nullptr; 924 } 925 } 926 927 return Ty; 928 } 929 930 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { 931 // If this GEP has only 0 indices, it is the same pointer as 932 // Src. If Src is not a trivial GEP too, don't combine 933 // the indices. 934 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() && 935 !Src.hasOneUse()) 936 return false; 937 return true; 938 } 939 940 /// Descale - Return a value X such that Val = X * Scale, or null if none. If 941 /// the multiplication is known not to overflow then NoSignedWrap is set. 942 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { 943 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); 944 assert(cast<IntegerType>(Val->getType())->getBitWidth() == 945 Scale.getBitWidth() && "Scale not compatible with value!"); 946 947 // If Val is zero or Scale is one then Val = Val * Scale. 948 if (match(Val, m_Zero()) || Scale == 1) { 949 NoSignedWrap = true; 950 return Val; 951 } 952 953 // If Scale is zero then it does not divide Val. 954 if (Scale.isMinValue()) 955 return nullptr; 956 957 // Look through chains of multiplications, searching for a constant that is 958 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 959 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by 960 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore 961 // down from Val: 962 // 963 // Val = M1 * X || Analysis starts here and works down 964 // M1 = M2 * Y || Doesn't descend into terms with more 965 // M2 = Z * 4 \/ than one use 966 // 967 // Then to modify a term at the bottom: 968 // 969 // Val = M1 * X 970 // M1 = Z * Y || Replaced M2 with Z 971 // 972 // Then to work back up correcting nsw flags. 973 974 // Op - the term we are currently analyzing. Starts at Val then drills down. 975 // Replaced with its descaled value before exiting from the drill down loop. 976 Value *Op = Val; 977 978 // Parent - initially null, but after drilling down notes where Op came from. 979 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the 980 // 0'th operand of Val. 981 std::pair<Instruction*, unsigned> Parent; 982 983 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper 984 // levels that doesn't overflow. 985 bool RequireNoSignedWrap = false; 986 987 // logScale - log base 2 of the scale. Negative if not a power of 2. 988 int32_t logScale = Scale.exactLogBase2(); 989 990 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down 991 992 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 993 // If Op is a constant divisible by Scale then descale to the quotient. 994 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. 995 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); 996 if (!Remainder.isMinValue()) 997 // Not divisible by Scale. 998 return nullptr; 999 // Replace with the quotient in the parent. 1000 Op = ConstantInt::get(CI->getType(), Quotient); 1001 NoSignedWrap = true; 1002 break; 1003 } 1004 1005 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { 1006 1007 if (BO->getOpcode() == Instruction::Mul) { 1008 // Multiplication. 1009 NoSignedWrap = BO->hasNoSignedWrap(); 1010 if (RequireNoSignedWrap && !NoSignedWrap) 1011 return nullptr; 1012 1013 // There are three cases for multiplication: multiplication by exactly 1014 // the scale, multiplication by a constant different to the scale, and 1015 // multiplication by something else. 1016 Value *LHS = BO->getOperand(0); 1017 Value *RHS = BO->getOperand(1); 1018 1019 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 1020 // Multiplication by a constant. 1021 if (CI->getValue() == Scale) { 1022 // Multiplication by exactly the scale, replace the multiplication 1023 // by its left-hand side in the parent. 1024 Op = LHS; 1025 break; 1026 } 1027 1028 // Otherwise drill down into the constant. 1029 if (!Op->hasOneUse()) 1030 return nullptr; 1031 1032 Parent = std::make_pair(BO, 1); 1033 continue; 1034 } 1035 1036 // Multiplication by something else. Drill down into the left-hand side 1037 // since that's where the reassociate pass puts the good stuff. 1038 if (!Op->hasOneUse()) 1039 return nullptr; 1040 1041 Parent = std::make_pair(BO, 0); 1042 continue; 1043 } 1044 1045 if (logScale > 0 && BO->getOpcode() == Instruction::Shl && 1046 isa<ConstantInt>(BO->getOperand(1))) { 1047 // Multiplication by a power of 2. 1048 NoSignedWrap = BO->hasNoSignedWrap(); 1049 if (RequireNoSignedWrap && !NoSignedWrap) 1050 return nullptr; 1051 1052 Value *LHS = BO->getOperand(0); 1053 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> 1054 getLimitedValue(Scale.getBitWidth()); 1055 // Op = LHS << Amt. 1056 1057 if (Amt == logScale) { 1058 // Multiplication by exactly the scale, replace the multiplication 1059 // by its left-hand side in the parent. 1060 Op = LHS; 1061 break; 1062 } 1063 if (Amt < logScale || !Op->hasOneUse()) 1064 return nullptr; 1065 1066 // Multiplication by more than the scale. Reduce the multiplying amount 1067 // by the scale in the parent. 1068 Parent = std::make_pair(BO, 1); 1069 Op = ConstantInt::get(BO->getType(), Amt - logScale); 1070 break; 1071 } 1072 } 1073 1074 if (!Op->hasOneUse()) 1075 return nullptr; 1076 1077 if (CastInst *Cast = dyn_cast<CastInst>(Op)) { 1078 if (Cast->getOpcode() == Instruction::SExt) { 1079 // Op is sign-extended from a smaller type, descale in the smaller type. 1080 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1081 APInt SmallScale = Scale.trunc(SmallSize); 1082 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to 1083 // descale Op as (sext Y) * Scale. In order to have 1084 // sext (Y * SmallScale) = (sext Y) * Scale 1085 // some conditions need to hold however: SmallScale must sign-extend to 1086 // Scale and the multiplication Y * SmallScale should not overflow. 1087 if (SmallScale.sext(Scale.getBitWidth()) != Scale) 1088 // SmallScale does not sign-extend to Scale. 1089 return nullptr; 1090 assert(SmallScale.exactLogBase2() == logScale); 1091 // Require that Y * SmallScale must not overflow. 1092 RequireNoSignedWrap = true; 1093 1094 // Drill down through the cast. 1095 Parent = std::make_pair(Cast, 0); 1096 Scale = SmallScale; 1097 continue; 1098 } 1099 1100 if (Cast->getOpcode() == Instruction::Trunc) { 1101 // Op is truncated from a larger type, descale in the larger type. 1102 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then 1103 // trunc (Y * sext Scale) = (trunc Y) * Scale 1104 // always holds. However (trunc Y) * Scale may overflow even if 1105 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared 1106 // from this point up in the expression (see later). 1107 if (RequireNoSignedWrap) 1108 return nullptr; 1109 1110 // Drill down through the cast. 1111 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1112 Parent = std::make_pair(Cast, 0); 1113 Scale = Scale.sext(LargeSize); 1114 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) 1115 logScale = -1; 1116 assert(Scale.exactLogBase2() == logScale); 1117 continue; 1118 } 1119 } 1120 1121 // Unsupported expression, bail out. 1122 return nullptr; 1123 } 1124 1125 // If Op is zero then Val = Op * Scale. 1126 if (match(Op, m_Zero())) { 1127 NoSignedWrap = true; 1128 return Op; 1129 } 1130 1131 // We know that we can successfully descale, so from here on we can safely 1132 // modify the IR. Op holds the descaled version of the deepest term in the 1133 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known 1134 // not to overflow. 1135 1136 if (!Parent.first) 1137 // The expression only had one term. 1138 return Op; 1139 1140 // Rewrite the parent using the descaled version of its operand. 1141 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); 1142 assert(Op != Parent.first->getOperand(Parent.second) && 1143 "Descaling was a no-op?"); 1144 Parent.first->setOperand(Parent.second, Op); 1145 Worklist.Add(Parent.first); 1146 1147 // Now work back up the expression correcting nsw flags. The logic is based 1148 // on the following observation: if X * Y is known not to overflow as a signed 1149 // multiplication, and Y is replaced by a value Z with smaller absolute value, 1150 // then X * Z will not overflow as a signed multiplication either. As we work 1151 // our way up, having NoSignedWrap 'true' means that the descaled value at the 1152 // current level has strictly smaller absolute value than the original. 1153 Instruction *Ancestor = Parent.first; 1154 do { 1155 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { 1156 // If the multiplication wasn't nsw then we can't say anything about the 1157 // value of the descaled multiplication, and we have to clear nsw flags 1158 // from this point on up. 1159 bool OpNoSignedWrap = BO->hasNoSignedWrap(); 1160 NoSignedWrap &= OpNoSignedWrap; 1161 if (NoSignedWrap != OpNoSignedWrap) { 1162 BO->setHasNoSignedWrap(NoSignedWrap); 1163 Worklist.Add(Ancestor); 1164 } 1165 } else if (Ancestor->getOpcode() == Instruction::Trunc) { 1166 // The fact that the descaled input to the trunc has smaller absolute 1167 // value than the original input doesn't tell us anything useful about 1168 // the absolute values of the truncations. 1169 NoSignedWrap = false; 1170 } 1171 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && 1172 "Failed to keep proper track of nsw flags while drilling down?"); 1173 1174 if (Ancestor == Val) 1175 // Got to the top, all done! 1176 return Val; 1177 1178 // Move up one level in the expression. 1179 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); 1180 Ancestor = Ancestor->user_back(); 1181 } while (1); 1182 } 1183 1184 /// \brief Creates node of binary operation with the same attributes as the 1185 /// specified one but with other operands. 1186 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS, 1187 InstCombiner::BuilderTy *B) { 1188 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS); 1189 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) { 1190 if (isa<OverflowingBinaryOperator>(NewBO)) { 1191 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap()); 1192 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap()); 1193 } 1194 if (isa<PossiblyExactOperator>(NewBO)) 1195 NewBO->setIsExact(Inst.isExact()); 1196 } 1197 return BORes; 1198 } 1199 1200 /// \brief Makes transformation of binary operation specific for vector types. 1201 /// \param Inst Binary operator to transform. 1202 /// \return Pointer to node that must replace the original binary operator, or 1203 /// null pointer if no transformation was made. 1204 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) { 1205 if (!Inst.getType()->isVectorTy()) return nullptr; 1206 1207 // It may not be safe to reorder shuffles and things like div, urem, etc. 1208 // because we may trap when executing those ops on unknown vector elements. 1209 // See PR20059. 1210 if (!isSafeToSpeculativelyExecute(&Inst)) 1211 return nullptr; 1212 1213 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements(); 1214 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1); 1215 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth); 1216 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth); 1217 1218 // If both arguments of binary operation are shuffles, which use the same 1219 // mask and shuffle within a single vector, it is worthwhile to move the 1220 // shuffle after binary operation: 1221 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m) 1222 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) { 1223 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS); 1224 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS); 1225 if (isa<UndefValue>(LShuf->getOperand(1)) && 1226 isa<UndefValue>(RShuf->getOperand(1)) && 1227 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() && 1228 LShuf->getMask() == RShuf->getMask()) { 1229 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0), 1230 RShuf->getOperand(0), Builder); 1231 Value *Res = Builder->CreateShuffleVector(NewBO, 1232 UndefValue::get(NewBO->getType()), LShuf->getMask()); 1233 return Res; 1234 } 1235 } 1236 1237 // If one argument is a shuffle within one vector, the other is a constant, 1238 // try moving the shuffle after the binary operation. 1239 ShuffleVectorInst *Shuffle = nullptr; 1240 Constant *C1 = nullptr; 1241 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS); 1242 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS); 1243 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS); 1244 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS); 1245 if (Shuffle && C1 && 1246 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) && 1247 isa<UndefValue>(Shuffle->getOperand(1)) && 1248 Shuffle->getType() == Shuffle->getOperand(0)->getType()) { 1249 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask(); 1250 // Find constant C2 that has property: 1251 // shuffle(C2, ShMask) = C1 1252 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>) 1253 // reorder is not possible. 1254 SmallVector<Constant*, 16> C2M(VWidth, 1255 UndefValue::get(C1->getType()->getScalarType())); 1256 bool MayChange = true; 1257 for (unsigned I = 0; I < VWidth; ++I) { 1258 if (ShMask[I] >= 0) { 1259 assert(ShMask[I] < (int)VWidth); 1260 if (!isa<UndefValue>(C2M[ShMask[I]])) { 1261 MayChange = false; 1262 break; 1263 } 1264 C2M[ShMask[I]] = C1->getAggregateElement(I); 1265 } 1266 } 1267 if (MayChange) { 1268 Constant *C2 = ConstantVector::get(C2M); 1269 Value *NewLHS, *NewRHS; 1270 if (isa<Constant>(LHS)) { 1271 NewLHS = C2; 1272 NewRHS = Shuffle->getOperand(0); 1273 } else { 1274 NewLHS = Shuffle->getOperand(0); 1275 NewRHS = C2; 1276 } 1277 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder); 1278 Value *Res = Builder->CreateShuffleVector(NewBO, 1279 UndefValue::get(Inst.getType()), Shuffle->getMask()); 1280 return Res; 1281 } 1282 } 1283 1284 return nullptr; 1285 } 1286 1287 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 1288 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 1289 1290 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AC)) 1291 return ReplaceInstUsesWith(GEP, V); 1292 1293 Value *PtrOp = GEP.getOperand(0); 1294 1295 // Eliminate unneeded casts for indices, and replace indices which displace 1296 // by multiples of a zero size type with zero. 1297 bool MadeChange = false; 1298 Type *IntPtrTy = DL.getIntPtrType(GEP.getPointerOperandType()); 1299 1300 gep_type_iterator GTI = gep_type_begin(GEP); 1301 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E; 1302 ++I, ++GTI) { 1303 // Skip indices into struct types. 1304 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 1305 if (!SeqTy) 1306 continue; 1307 1308 // If the element type has zero size then any index over it is equivalent 1309 // to an index of zero, so replace it with zero if it is not zero already. 1310 if (SeqTy->getElementType()->isSized() && 1311 DL.getTypeAllocSize(SeqTy->getElementType()) == 0) 1312 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 1313 *I = Constant::getNullValue(IntPtrTy); 1314 MadeChange = true; 1315 } 1316 1317 Type *IndexTy = (*I)->getType(); 1318 if (IndexTy != IntPtrTy) { 1319 // If we are using a wider index than needed for this platform, shrink 1320 // it to what we need. If narrower, sign-extend it to what we need. 1321 // This explicit cast can make subsequent optimizations more obvious. 1322 *I = Builder->CreateIntCast(*I, IntPtrTy, true); 1323 MadeChange = true; 1324 } 1325 } 1326 if (MadeChange) 1327 return &GEP; 1328 1329 // Check to see if the inputs to the PHI node are getelementptr instructions. 1330 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) { 1331 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0)); 1332 if (!Op1) 1333 return nullptr; 1334 1335 // Don't fold a GEP into itself through a PHI node. This can only happen 1336 // through the back-edge of a loop. Folding a GEP into itself means that 1337 // the value of the previous iteration needs to be stored in the meantime, 1338 // thus requiring an additional register variable to be live, but not 1339 // actually achieving anything (the GEP still needs to be executed once per 1340 // loop iteration). 1341 if (Op1 == &GEP) 1342 return nullptr; 1343 1344 signed DI = -1; 1345 1346 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) { 1347 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I); 1348 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands()) 1349 return nullptr; 1350 1351 // As for Op1 above, don't try to fold a GEP into itself. 1352 if (Op2 == &GEP) 1353 return nullptr; 1354 1355 // Keep track of the type as we walk the GEP. 1356 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType(); 1357 1358 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) { 1359 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType()) 1360 return nullptr; 1361 1362 if (Op1->getOperand(J) != Op2->getOperand(J)) { 1363 if (DI == -1) { 1364 // We have not seen any differences yet in the GEPs feeding the 1365 // PHI yet, so we record this one if it is allowed to be a 1366 // variable. 1367 1368 // The first two arguments can vary for any GEP, the rest have to be 1369 // static for struct slots 1370 if (J > 1 && CurTy->isStructTy()) 1371 return nullptr; 1372 1373 DI = J; 1374 } else { 1375 // The GEP is different by more than one input. While this could be 1376 // extended to support GEPs that vary by more than one variable it 1377 // doesn't make sense since it greatly increases the complexity and 1378 // would result in an R+R+R addressing mode which no backend 1379 // directly supports and would need to be broken into several 1380 // simpler instructions anyway. 1381 return nullptr; 1382 } 1383 } 1384 1385 // Sink down a layer of the type for the next iteration. 1386 if (J > 0) { 1387 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) { 1388 CurTy = CT->getTypeAtIndex(Op1->getOperand(J)); 1389 } else { 1390 CurTy = nullptr; 1391 } 1392 } 1393 } 1394 } 1395 1396 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone()); 1397 1398 if (DI == -1) { 1399 // All the GEPs feeding the PHI are identical. Clone one down into our 1400 // BB so that it can be merged with the current GEP. 1401 GEP.getParent()->getInstList().insert( 1402 GEP.getParent()->getFirstInsertionPt(), NewGEP); 1403 } else { 1404 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP 1405 // into the current block so it can be merged, and create a new PHI to 1406 // set that index. 1407 Instruction *InsertPt = Builder->GetInsertPoint(); 1408 Builder->SetInsertPoint(PN); 1409 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(), 1410 PN->getNumOperands()); 1411 Builder->SetInsertPoint(InsertPt); 1412 1413 for (auto &I : PN->operands()) 1414 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI), 1415 PN->getIncomingBlock(I)); 1416 1417 NewGEP->setOperand(DI, NewPN); 1418 GEP.getParent()->getInstList().insert( 1419 GEP.getParent()->getFirstInsertionPt(), NewGEP); 1420 NewGEP->setOperand(DI, NewPN); 1421 } 1422 1423 GEP.setOperand(0, NewGEP); 1424 PtrOp = NewGEP; 1425 } 1426 1427 // Combine Indices - If the source pointer to this getelementptr instruction 1428 // is a getelementptr instruction, combine the indices of the two 1429 // getelementptr instructions into a single instruction. 1430 // 1431 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 1432 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) 1433 return nullptr; 1434 1435 // Note that if our source is a gep chain itself then we wait for that 1436 // chain to be resolved before we perform this transformation. This 1437 // avoids us creating a TON of code in some cases. 1438 if (GEPOperator *SrcGEP = 1439 dyn_cast<GEPOperator>(Src->getOperand(0))) 1440 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) 1441 return nullptr; // Wait until our source is folded to completion. 1442 1443 SmallVector<Value*, 8> Indices; 1444 1445 // Find out whether the last index in the source GEP is a sequential idx. 1446 bool EndsWithSequential = false; 1447 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 1448 I != E; ++I) 1449 EndsWithSequential = !(*I)->isStructTy(); 1450 1451 // Can we combine the two pointer arithmetics offsets? 1452 if (EndsWithSequential) { 1453 // Replace: gep (gep %P, long B), long A, ... 1454 // With: T = long A+B; gep %P, T, ... 1455 // 1456 Value *Sum; 1457 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 1458 Value *GO1 = GEP.getOperand(1); 1459 if (SO1 == Constant::getNullValue(SO1->getType())) { 1460 Sum = GO1; 1461 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 1462 Sum = SO1; 1463 } else { 1464 // If they aren't the same type, then the input hasn't been processed 1465 // by the loop above yet (which canonicalizes sequential index types to 1466 // intptr_t). Just avoid transforming this until the input has been 1467 // normalized. 1468 if (SO1->getType() != GO1->getType()) 1469 return nullptr; 1470 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 1471 } 1472 1473 // Update the GEP in place if possible. 1474 if (Src->getNumOperands() == 2) { 1475 GEP.setOperand(0, Src->getOperand(0)); 1476 GEP.setOperand(1, Sum); 1477 return &GEP; 1478 } 1479 Indices.append(Src->op_begin()+1, Src->op_end()-1); 1480 Indices.push_back(Sum); 1481 Indices.append(GEP.op_begin()+2, GEP.op_end()); 1482 } else if (isa<Constant>(*GEP.idx_begin()) && 1483 cast<Constant>(*GEP.idx_begin())->isNullValue() && 1484 Src->getNumOperands() != 1) { 1485 // Otherwise we can do the fold if the first index of the GEP is a zero 1486 Indices.append(Src->op_begin()+1, Src->op_end()); 1487 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 1488 } 1489 1490 if (!Indices.empty()) 1491 return GEP.isInBounds() && Src->isInBounds() 1492 ? GetElementPtrInst::CreateInBounds( 1493 Src->getSourceElementType(), Src->getOperand(0), Indices, 1494 GEP.getName()) 1495 : GetElementPtrInst::Create(Src->getSourceElementType(), 1496 Src->getOperand(0), Indices, 1497 GEP.getName()); 1498 } 1499 1500 if (GEP.getNumIndices() == 1) { 1501 unsigned AS = GEP.getPointerAddressSpace(); 1502 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() == 1503 DL.getPointerSizeInBits(AS)) { 1504 Type *PtrTy = GEP.getPointerOperandType(); 1505 Type *Ty = PtrTy->getPointerElementType(); 1506 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty); 1507 1508 bool Matched = false; 1509 uint64_t C; 1510 Value *V = nullptr; 1511 if (TyAllocSize == 1) { 1512 V = GEP.getOperand(1); 1513 Matched = true; 1514 } else if (match(GEP.getOperand(1), 1515 m_AShr(m_Value(V), m_ConstantInt(C)))) { 1516 if (TyAllocSize == 1ULL << C) 1517 Matched = true; 1518 } else if (match(GEP.getOperand(1), 1519 m_SDiv(m_Value(V), m_ConstantInt(C)))) { 1520 if (TyAllocSize == C) 1521 Matched = true; 1522 } 1523 1524 if (Matched) { 1525 // Canonicalize (gep i8* X, -(ptrtoint Y)) 1526 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y))) 1527 // The GEP pattern is emitted by the SCEV expander for certain kinds of 1528 // pointer arithmetic. 1529 if (match(V, m_Neg(m_PtrToInt(m_Value())))) { 1530 Operator *Index = cast<Operator>(V); 1531 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); 1532 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); 1533 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); 1534 } 1535 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) 1536 // to (bitcast Y) 1537 Value *Y; 1538 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)), 1539 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) { 1540 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, 1541 GEP.getType()); 1542 } 1543 } 1544 } 1545 } 1546 1547 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 1548 Value *StrippedPtr = PtrOp->stripPointerCasts(); 1549 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType()); 1550 1551 // We do not handle pointer-vector geps here. 1552 if (!StrippedPtrTy) 1553 return nullptr; 1554 1555 if (StrippedPtr != PtrOp) { 1556 bool HasZeroPointerIndex = false; 1557 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 1558 HasZeroPointerIndex = C->isZero(); 1559 1560 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 1561 // into : GEP [10 x i8]* X, i32 0, ... 1562 // 1563 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 1564 // into : GEP i8* X, ... 1565 // 1566 // This occurs when the program declares an array extern like "int X[];" 1567 if (HasZeroPointerIndex) { 1568 PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 1569 if (ArrayType *CATy = 1570 dyn_cast<ArrayType>(CPTy->getElementType())) { 1571 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 1572 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 1573 // -> GEP i8* X, ... 1574 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 1575 GetElementPtrInst *Res = GetElementPtrInst::Create( 1576 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName()); 1577 Res->setIsInBounds(GEP.isInBounds()); 1578 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) 1579 return Res; 1580 // Insert Res, and create an addrspacecast. 1581 // e.g., 1582 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ... 1583 // -> 1584 // %0 = GEP i8 addrspace(1)* X, ... 1585 // addrspacecast i8 addrspace(1)* %0 to i8* 1586 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType()); 1587 } 1588 1589 if (ArrayType *XATy = 1590 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 1591 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 1592 if (CATy->getElementType() == XATy->getElementType()) { 1593 // -> GEP [10 x i8]* X, i32 0, ... 1594 // At this point, we know that the cast source type is a pointer 1595 // to an array of the same type as the destination pointer 1596 // array. Because the array type is never stepped over (there 1597 // is a leading zero) we can fold the cast into this GEP. 1598 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) { 1599 GEP.setOperand(0, StrippedPtr); 1600 return &GEP; 1601 } 1602 // Cannot replace the base pointer directly because StrippedPtr's 1603 // address space is different. Instead, create a new GEP followed by 1604 // an addrspacecast. 1605 // e.g., 1606 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*), 1607 // i32 0, ... 1608 // -> 1609 // %0 = GEP [10 x i8] addrspace(1)* X, ... 1610 // addrspacecast i8 addrspace(1)* %0 to i8* 1611 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end()); 1612 Value *NewGEP = GEP.isInBounds() 1613 ? Builder->CreateInBoundsGEP( 1614 nullptr, StrippedPtr, Idx, GEP.getName()) 1615 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, 1616 GEP.getName()); 1617 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1618 } 1619 } 1620 } 1621 } else if (GEP.getNumOperands() == 2) { 1622 // Transform things like: 1623 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 1624 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 1625 Type *SrcElTy = StrippedPtrTy->getElementType(); 1626 Type *ResElTy = PtrOp->getType()->getPointerElementType(); 1627 if (SrcElTy->isArrayTy() && 1628 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) == 1629 DL.getTypeAllocSize(ResElTy)) { 1630 Type *IdxType = DL.getIntPtrType(GEP.getType()); 1631 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; 1632 Value *NewGEP = 1633 GEP.isInBounds() 1634 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, Idx, 1635 GEP.getName()) 1636 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName()); 1637 1638 // V and GEP are both pointer types --> BitCast 1639 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, 1640 GEP.getType()); 1641 } 1642 1643 // Transform things like: 1644 // %V = mul i64 %N, 4 1645 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V 1646 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast 1647 if (ResElTy->isSized() && SrcElTy->isSized()) { 1648 // Check that changing the type amounts to dividing the index by a scale 1649 // factor. 1650 uint64_t ResSize = DL.getTypeAllocSize(ResElTy); 1651 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy); 1652 if (ResSize && SrcSize % ResSize == 0) { 1653 Value *Idx = GEP.getOperand(1); 1654 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1655 uint64_t Scale = SrcSize / ResSize; 1656 1657 // Earlier transforms ensure that the index has type IntPtrType, which 1658 // considerably simplifies the logic by eliminating implicit casts. 1659 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) && 1660 "Index not cast to pointer width?"); 1661 1662 bool NSW; 1663 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1664 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1665 // If the multiplication NewIdx * Scale may overflow then the new 1666 // GEP may not be "inbounds". 1667 Value *NewGEP = 1668 GEP.isInBounds() && NSW 1669 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx, 1670 GEP.getName()) 1671 : Builder->CreateGEP(nullptr, StrippedPtr, NewIdx, 1672 GEP.getName()); 1673 1674 // The NewGEP must be pointer typed, so must the old one -> BitCast 1675 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, 1676 GEP.getType()); 1677 } 1678 } 1679 } 1680 1681 // Similarly, transform things like: 1682 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 1683 // (where tmp = 8*tmp2) into: 1684 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 1685 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) { 1686 // Check that changing to the array element type amounts to dividing the 1687 // index by a scale factor. 1688 uint64_t ResSize = DL.getTypeAllocSize(ResElTy); 1689 uint64_t ArrayEltSize = 1690 DL.getTypeAllocSize(SrcElTy->getArrayElementType()); 1691 if (ResSize && ArrayEltSize % ResSize == 0) { 1692 Value *Idx = GEP.getOperand(1); 1693 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1694 uint64_t Scale = ArrayEltSize / ResSize; 1695 1696 // Earlier transforms ensure that the index has type IntPtrType, which 1697 // considerably simplifies the logic by eliminating implicit casts. 1698 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) && 1699 "Index not cast to pointer width?"); 1700 1701 bool NSW; 1702 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1703 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1704 // If the multiplication NewIdx * Scale may overflow then the new 1705 // GEP may not be "inbounds". 1706 Value *Off[2] = { 1707 Constant::getNullValue(DL.getIntPtrType(GEP.getType())), 1708 NewIdx}; 1709 1710 Value *NewGEP = GEP.isInBounds() && NSW 1711 ? Builder->CreateInBoundsGEP( 1712 SrcElTy, StrippedPtr, Off, GEP.getName()) 1713 : Builder->CreateGEP(SrcElTy, StrippedPtr, Off, 1714 GEP.getName()); 1715 // The NewGEP must be pointer typed, so must the old one -> BitCast 1716 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, 1717 GEP.getType()); 1718 } 1719 } 1720 } 1721 } 1722 } 1723 1724 // addrspacecast between types is canonicalized as a bitcast, then an 1725 // addrspacecast. To take advantage of the below bitcast + struct GEP, look 1726 // through the addrspacecast. 1727 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) { 1728 // X = bitcast A addrspace(1)* to B addrspace(1)* 1729 // Y = addrspacecast A addrspace(1)* to B addrspace(2)* 1730 // Z = gep Y, <...constant indices...> 1731 // Into an addrspacecasted GEP of the struct. 1732 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0))) 1733 PtrOp = BC; 1734 } 1735 1736 /// See if we can simplify: 1737 /// X = bitcast A* to B* 1738 /// Y = gep X, <...constant indices...> 1739 /// into a gep of the original struct. This is important for SROA and alias 1740 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 1741 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 1742 Value *Operand = BCI->getOperand(0); 1743 PointerType *OpType = cast<PointerType>(Operand->getType()); 1744 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType()); 1745 APInt Offset(OffsetBits, 0); 1746 if (!isa<BitCastInst>(Operand) && 1747 GEP.accumulateConstantOffset(DL, Offset)) { 1748 1749 // If this GEP instruction doesn't move the pointer, just replace the GEP 1750 // with a bitcast of the real input to the dest type. 1751 if (!Offset) { 1752 // If the bitcast is of an allocation, and the allocation will be 1753 // converted to match the type of the cast, don't touch this. 1754 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) { 1755 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1756 if (Instruction *I = visitBitCast(*BCI)) { 1757 if (I != BCI) { 1758 I->takeName(BCI); 1759 BCI->getParent()->getInstList().insert(BCI, I); 1760 ReplaceInstUsesWith(*BCI, I); 1761 } 1762 return &GEP; 1763 } 1764 } 1765 1766 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace()) 1767 return new AddrSpaceCastInst(Operand, GEP.getType()); 1768 return new BitCastInst(Operand, GEP.getType()); 1769 } 1770 1771 // Otherwise, if the offset is non-zero, we need to find out if there is a 1772 // field at Offset in 'A's type. If so, we can pull the cast through the 1773 // GEP. 1774 SmallVector<Value*, 8> NewIndices; 1775 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { 1776 Value *NGEP = 1777 GEP.isInBounds() 1778 ? Builder->CreateInBoundsGEP(nullptr, Operand, NewIndices) 1779 : Builder->CreateGEP(nullptr, Operand, NewIndices); 1780 1781 if (NGEP->getType() == GEP.getType()) 1782 return ReplaceInstUsesWith(GEP, NGEP); 1783 NGEP->takeName(&GEP); 1784 1785 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace()) 1786 return new AddrSpaceCastInst(NGEP, GEP.getType()); 1787 return new BitCastInst(NGEP, GEP.getType()); 1788 } 1789 } 1790 } 1791 1792 return nullptr; 1793 } 1794 1795 static bool 1796 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, 1797 const TargetLibraryInfo *TLI) { 1798 SmallVector<Instruction*, 4> Worklist; 1799 Worklist.push_back(AI); 1800 1801 do { 1802 Instruction *PI = Worklist.pop_back_val(); 1803 for (User *U : PI->users()) { 1804 Instruction *I = cast<Instruction>(U); 1805 switch (I->getOpcode()) { 1806 default: 1807 // Give up the moment we see something we can't handle. 1808 return false; 1809 1810 case Instruction::BitCast: 1811 case Instruction::GetElementPtr: 1812 Users.push_back(I); 1813 Worklist.push_back(I); 1814 continue; 1815 1816 case Instruction::ICmp: { 1817 ICmpInst *ICI = cast<ICmpInst>(I); 1818 // We can fold eq/ne comparisons with null to false/true, respectively. 1819 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1))) 1820 return false; 1821 Users.push_back(I); 1822 continue; 1823 } 1824 1825 case Instruction::Call: 1826 // Ignore no-op and store intrinsics. 1827 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1828 switch (II->getIntrinsicID()) { 1829 default: 1830 return false; 1831 1832 case Intrinsic::memmove: 1833 case Intrinsic::memcpy: 1834 case Intrinsic::memset: { 1835 MemIntrinsic *MI = cast<MemIntrinsic>(II); 1836 if (MI->isVolatile() || MI->getRawDest() != PI) 1837 return false; 1838 } 1839 // fall through 1840 case Intrinsic::dbg_declare: 1841 case Intrinsic::dbg_value: 1842 case Intrinsic::invariant_start: 1843 case Intrinsic::invariant_end: 1844 case Intrinsic::lifetime_start: 1845 case Intrinsic::lifetime_end: 1846 case Intrinsic::objectsize: 1847 Users.push_back(I); 1848 continue; 1849 } 1850 } 1851 1852 if (isFreeCall(I, TLI)) { 1853 Users.push_back(I); 1854 continue; 1855 } 1856 return false; 1857 1858 case Instruction::Store: { 1859 StoreInst *SI = cast<StoreInst>(I); 1860 if (SI->isVolatile() || SI->getPointerOperand() != PI) 1861 return false; 1862 Users.push_back(I); 1863 continue; 1864 } 1865 } 1866 llvm_unreachable("missing a return?"); 1867 } 1868 } while (!Worklist.empty()); 1869 return true; 1870 } 1871 1872 Instruction *InstCombiner::visitAllocSite(Instruction &MI) { 1873 // If we have a malloc call which is only used in any amount of comparisons 1874 // to null and free calls, delete the calls and replace the comparisons with 1875 // true or false as appropriate. 1876 SmallVector<WeakVH, 64> Users; 1877 if (isAllocSiteRemovable(&MI, Users, TLI)) { 1878 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 1879 Instruction *I = cast_or_null<Instruction>(&*Users[i]); 1880 if (!I) continue; 1881 1882 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) { 1883 ReplaceInstUsesWith(*C, 1884 ConstantInt::get(Type::getInt1Ty(C->getContext()), 1885 C->isFalseWhenEqual())); 1886 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 1887 ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); 1888 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1889 if (II->getIntrinsicID() == Intrinsic::objectsize) { 1890 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1)); 1891 uint64_t DontKnow = CI->isZero() ? -1ULL : 0; 1892 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); 1893 } 1894 } 1895 EraseInstFromFunction(*I); 1896 } 1897 1898 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) { 1899 // Replace invoke with a NOP intrinsic to maintain the original CFG 1900 Module *M = II->getParent()->getParent()->getParent(); 1901 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); 1902 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), 1903 None, "", II->getParent()); 1904 } 1905 return EraseInstFromFunction(MI); 1906 } 1907 return nullptr; 1908 } 1909 1910 /// \brief Move the call to free before a NULL test. 1911 /// 1912 /// Check if this free is accessed after its argument has been test 1913 /// against NULL (property 0). 1914 /// If yes, it is legal to move this call in its predecessor block. 1915 /// 1916 /// The move is performed only if the block containing the call to free 1917 /// will be removed, i.e.: 1918 /// 1. it has only one predecessor P, and P has two successors 1919 /// 2. it contains the call and an unconditional branch 1920 /// 3. its successor is the same as its predecessor's successor 1921 /// 1922 /// The profitability is out-of concern here and this function should 1923 /// be called only if the caller knows this transformation would be 1924 /// profitable (e.g., for code size). 1925 static Instruction * 1926 tryToMoveFreeBeforeNullTest(CallInst &FI) { 1927 Value *Op = FI.getArgOperand(0); 1928 BasicBlock *FreeInstrBB = FI.getParent(); 1929 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); 1930 1931 // Validate part of constraint #1: Only one predecessor 1932 // FIXME: We can extend the number of predecessor, but in that case, we 1933 // would duplicate the call to free in each predecessor and it may 1934 // not be profitable even for code size. 1935 if (!PredBB) 1936 return nullptr; 1937 1938 // Validate constraint #2: Does this block contains only the call to 1939 // free and an unconditional branch? 1940 // FIXME: We could check if we can speculate everything in the 1941 // predecessor block 1942 if (FreeInstrBB->size() != 2) 1943 return nullptr; 1944 BasicBlock *SuccBB; 1945 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) 1946 return nullptr; 1947 1948 // Validate the rest of constraint #1 by matching on the pred branch. 1949 TerminatorInst *TI = PredBB->getTerminator(); 1950 BasicBlock *TrueBB, *FalseBB; 1951 ICmpInst::Predicate Pred; 1952 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) 1953 return nullptr; 1954 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) 1955 return nullptr; 1956 1957 // Validate constraint #3: Ensure the null case just falls through. 1958 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) 1959 return nullptr; 1960 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && 1961 "Broken CFG: missing edge from predecessor to successor"); 1962 1963 FI.moveBefore(TI); 1964 return &FI; 1965 } 1966 1967 1968 Instruction *InstCombiner::visitFree(CallInst &FI) { 1969 Value *Op = FI.getArgOperand(0); 1970 1971 // free undef -> unreachable. 1972 if (isa<UndefValue>(Op)) { 1973 // Insert a new store to null because we cannot modify the CFG here. 1974 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()), 1975 UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); 1976 return EraseInstFromFunction(FI); 1977 } 1978 1979 // If we have 'free null' delete the instruction. This can happen in stl code 1980 // when lots of inlining happens. 1981 if (isa<ConstantPointerNull>(Op)) 1982 return EraseInstFromFunction(FI); 1983 1984 // If we optimize for code size, try to move the call to free before the null 1985 // test so that simplify cfg can remove the empty block and dead code 1986 // elimination the branch. I.e., helps to turn something like: 1987 // if (foo) free(foo); 1988 // into 1989 // free(foo); 1990 if (MinimizeSize) 1991 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) 1992 return I; 1993 1994 return nullptr; 1995 } 1996 1997 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) { 1998 if (RI.getNumOperands() == 0) // ret void 1999 return nullptr; 2000 2001 Value *ResultOp = RI.getOperand(0); 2002 Type *VTy = ResultOp->getType(); 2003 if (!VTy->isIntegerTy()) 2004 return nullptr; 2005 2006 // There might be assume intrinsics dominating this return that completely 2007 // determine the value. If so, constant fold it. 2008 unsigned BitWidth = VTy->getPrimitiveSizeInBits(); 2009 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 2010 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI); 2011 if ((KnownZero|KnownOne).isAllOnesValue()) 2012 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne)); 2013 2014 return nullptr; 2015 } 2016 2017 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 2018 // Change br (not X), label True, label False to: br X, label False, True 2019 Value *X = nullptr; 2020 BasicBlock *TrueDest; 2021 BasicBlock *FalseDest; 2022 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 2023 !isa<Constant>(X)) { 2024 // Swap Destinations and condition... 2025 BI.setCondition(X); 2026 BI.swapSuccessors(); 2027 return &BI; 2028 } 2029 2030 // If the condition is irrelevant, remove the use so that other 2031 // transforms on the condition become more effective. 2032 if (BI.isConditional() && 2033 BI.getSuccessor(0) == BI.getSuccessor(1) && 2034 !isa<UndefValue>(BI.getCondition())) { 2035 BI.setCondition(UndefValue::get(BI.getCondition()->getType())); 2036 return &BI; 2037 } 2038 2039 // Canonicalize fcmp_one -> fcmp_oeq 2040 FCmpInst::Predicate FPred; Value *Y; 2041 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 2042 TrueDest, FalseDest)) && 2043 BI.getCondition()->hasOneUse()) 2044 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 2045 FPred == FCmpInst::FCMP_OGE) { 2046 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 2047 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 2048 2049 // Swap Destinations and condition. 2050 BI.swapSuccessors(); 2051 Worklist.Add(Cond); 2052 return &BI; 2053 } 2054 2055 // Canonicalize icmp_ne -> icmp_eq 2056 ICmpInst::Predicate IPred; 2057 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 2058 TrueDest, FalseDest)) && 2059 BI.getCondition()->hasOneUse()) 2060 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 2061 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 2062 IPred == ICmpInst::ICMP_SGE) { 2063 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 2064 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 2065 // Swap Destinations and condition. 2066 BI.swapSuccessors(); 2067 Worklist.Add(Cond); 2068 return &BI; 2069 } 2070 2071 return nullptr; 2072 } 2073 2074 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 2075 Value *Cond = SI.getCondition(); 2076 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth(); 2077 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 2078 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI); 2079 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes(); 2080 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes(); 2081 2082 // Compute the number of leading bits we can ignore. 2083 for (auto &C : SI.cases()) { 2084 LeadingKnownZeros = std::min( 2085 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros()); 2086 LeadingKnownOnes = std::min( 2087 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes()); 2088 } 2089 2090 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes); 2091 2092 // Truncate the condition operand if the new type is equal to or larger than 2093 // the largest legal integer type. We need to be conservative here since 2094 // x86 generates redundant zero-extenstion instructions if the operand is 2095 // truncated to i8 or i16. 2096 bool TruncCond = false; 2097 if (NewWidth > 0 && BitWidth > NewWidth && 2098 NewWidth >= DL.getLargestLegalIntTypeSize()) { 2099 TruncCond = true; 2100 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth); 2101 Builder->SetInsertPoint(&SI); 2102 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc"); 2103 SI.setCondition(NewCond); 2104 2105 for (auto &C : SI.cases()) 2106 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get( 2107 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth))); 2108 } 2109 2110 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 2111 if (I->getOpcode() == Instruction::Add) 2112 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 2113 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 2114 // Skip the first item since that's the default case. 2115 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); 2116 i != e; ++i) { 2117 ConstantInt* CaseVal = i.getCaseValue(); 2118 Constant *LHS = CaseVal; 2119 if (TruncCond) 2120 LHS = LeadingKnownZeros 2121 ? ConstantExpr::getZExt(CaseVal, Cond->getType()) 2122 : ConstantExpr::getSExt(CaseVal, Cond->getType()); 2123 Constant* NewCaseVal = ConstantExpr::getSub(LHS, AddRHS); 2124 assert(isa<ConstantInt>(NewCaseVal) && 2125 "Result of expression should be constant"); 2126 i.setValue(cast<ConstantInt>(NewCaseVal)); 2127 } 2128 SI.setCondition(I->getOperand(0)); 2129 Worklist.Add(I); 2130 return &SI; 2131 } 2132 } 2133 2134 return TruncCond ? &SI : nullptr; 2135 } 2136 2137 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 2138 Value *Agg = EV.getAggregateOperand(); 2139 2140 if (!EV.hasIndices()) 2141 return ReplaceInstUsesWith(EV, Agg); 2142 2143 if (Constant *C = dyn_cast<Constant>(Agg)) { 2144 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) { 2145 if (EV.getNumIndices() == 0) 2146 return ReplaceInstUsesWith(EV, C2); 2147 // Extract the remaining indices out of the constant indexed by the 2148 // first index 2149 return ExtractValueInst::Create(C2, EV.getIndices().slice(1)); 2150 } 2151 return nullptr; // Can't handle other constants 2152 } 2153 2154 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 2155 // We're extracting from an insertvalue instruction, compare the indices 2156 const unsigned *exti, *exte, *insi, *inse; 2157 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 2158 exte = EV.idx_end(), inse = IV->idx_end(); 2159 exti != exte && insi != inse; 2160 ++exti, ++insi) { 2161 if (*insi != *exti) 2162 // The insert and extract both reference distinctly different elements. 2163 // This means the extract is not influenced by the insert, and we can 2164 // replace the aggregate operand of the extract with the aggregate 2165 // operand of the insert. i.e., replace 2166 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 2167 // %E = extractvalue { i32, { i32 } } %I, 0 2168 // with 2169 // %E = extractvalue { i32, { i32 } } %A, 0 2170 return ExtractValueInst::Create(IV->getAggregateOperand(), 2171 EV.getIndices()); 2172 } 2173 if (exti == exte && insi == inse) 2174 // Both iterators are at the end: Index lists are identical. Replace 2175 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 2176 // %C = extractvalue { i32, { i32 } } %B, 1, 0 2177 // with "i32 42" 2178 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 2179 if (exti == exte) { 2180 // The extract list is a prefix of the insert list. i.e. replace 2181 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 2182 // %E = extractvalue { i32, { i32 } } %I, 1 2183 // with 2184 // %X = extractvalue { i32, { i32 } } %A, 1 2185 // %E = insertvalue { i32 } %X, i32 42, 0 2186 // by switching the order of the insert and extract (though the 2187 // insertvalue should be left in, since it may have other uses). 2188 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 2189 EV.getIndices()); 2190 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 2191 makeArrayRef(insi, inse)); 2192 } 2193 if (insi == inse) 2194 // The insert list is a prefix of the extract list 2195 // We can simply remove the common indices from the extract and make it 2196 // operate on the inserted value instead of the insertvalue result. 2197 // i.e., replace 2198 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 2199 // %E = extractvalue { i32, { i32 } } %I, 1, 0 2200 // with 2201 // %E extractvalue { i32 } { i32 42 }, 0 2202 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 2203 makeArrayRef(exti, exte)); 2204 } 2205 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 2206 // We're extracting from an intrinsic, see if we're the only user, which 2207 // allows us to simplify multiple result intrinsics to simpler things that 2208 // just get one value. 2209 if (II->hasOneUse()) { 2210 // Check if we're grabbing the overflow bit or the result of a 'with 2211 // overflow' intrinsic. If it's the latter we can remove the intrinsic 2212 // and replace it with a traditional binary instruction. 2213 switch (II->getIntrinsicID()) { 2214 case Intrinsic::uadd_with_overflow: 2215 case Intrinsic::sadd_with_overflow: 2216 if (*EV.idx_begin() == 0) { // Normal result. 2217 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2218 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2219 EraseInstFromFunction(*II); 2220 return BinaryOperator::CreateAdd(LHS, RHS); 2221 } 2222 2223 // If the normal result of the add is dead, and the RHS is a constant, 2224 // we can transform this into a range comparison. 2225 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 2226 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 2227 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 2228 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 2229 ConstantExpr::getNot(CI)); 2230 break; 2231 case Intrinsic::usub_with_overflow: 2232 case Intrinsic::ssub_with_overflow: 2233 if (*EV.idx_begin() == 0) { // Normal result. 2234 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2235 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2236 EraseInstFromFunction(*II); 2237 return BinaryOperator::CreateSub(LHS, RHS); 2238 } 2239 break; 2240 case Intrinsic::umul_with_overflow: 2241 case Intrinsic::smul_with_overflow: 2242 if (*EV.idx_begin() == 0) { // Normal result. 2243 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2244 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2245 EraseInstFromFunction(*II); 2246 return BinaryOperator::CreateMul(LHS, RHS); 2247 } 2248 break; 2249 default: 2250 break; 2251 } 2252 } 2253 } 2254 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 2255 // If the (non-volatile) load only has one use, we can rewrite this to a 2256 // load from a GEP. This reduces the size of the load. 2257 // FIXME: If a load is used only by extractvalue instructions then this 2258 // could be done regardless of having multiple uses. 2259 if (L->isSimple() && L->hasOneUse()) { 2260 // extractvalue has integer indices, getelementptr has Value*s. Convert. 2261 SmallVector<Value*, 4> Indices; 2262 // Prefix an i32 0 since we need the first element. 2263 Indices.push_back(Builder->getInt32(0)); 2264 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 2265 I != E; ++I) 2266 Indices.push_back(Builder->getInt32(*I)); 2267 2268 // We need to insert these at the location of the old load, not at that of 2269 // the extractvalue. 2270 Builder->SetInsertPoint(L->getParent(), L); 2271 Value *GEP = Builder->CreateInBoundsGEP(L->getType(), 2272 L->getPointerOperand(), Indices); 2273 // Returning the load directly will cause the main loop to insert it in 2274 // the wrong spot, so use ReplaceInstUsesWith(). 2275 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 2276 } 2277 // We could simplify extracts from other values. Note that nested extracts may 2278 // already be simplified implicitly by the above: extract (extract (insert) ) 2279 // will be translated into extract ( insert ( extract ) ) first and then just 2280 // the value inserted, if appropriate. Similarly for extracts from single-use 2281 // loads: extract (extract (load)) will be translated to extract (load (gep)) 2282 // and if again single-use then via load (gep (gep)) to load (gep). 2283 // However, double extracts from e.g. function arguments or return values 2284 // aren't handled yet. 2285 return nullptr; 2286 } 2287 2288 /// isCatchAll - Return 'true' if the given typeinfo will match anything. 2289 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) { 2290 switch (Personality) { 2291 case EHPersonality::GNU_C: 2292 // The GCC C EH personality only exists to support cleanups, so it's not 2293 // clear what the semantics of catch clauses are. 2294 return false; 2295 case EHPersonality::Unknown: 2296 return false; 2297 case EHPersonality::GNU_Ada: 2298 // While __gnat_all_others_value will match any Ada exception, it doesn't 2299 // match foreign exceptions (or didn't, before gcc-4.7). 2300 return false; 2301 case EHPersonality::GNU_CXX: 2302 case EHPersonality::GNU_ObjC: 2303 case EHPersonality::MSVC_X86SEH: 2304 case EHPersonality::MSVC_Win64SEH: 2305 case EHPersonality::MSVC_CXX: 2306 return TypeInfo->isNullValue(); 2307 } 2308 llvm_unreachable("invalid enum"); 2309 } 2310 2311 static bool shorter_filter(const Value *LHS, const Value *RHS) { 2312 return 2313 cast<ArrayType>(LHS->getType())->getNumElements() 2314 < 2315 cast<ArrayType>(RHS->getType())->getNumElements(); 2316 } 2317 2318 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { 2319 // The logic here should be correct for any real-world personality function. 2320 // However if that turns out not to be true, the offending logic can always 2321 // be conditioned on the personality function, like the catch-all logic is. 2322 EHPersonality Personality = classifyEHPersonality(LI.getPersonalityFn()); 2323 2324 // Simplify the list of clauses, eg by removing repeated catch clauses 2325 // (these are often created by inlining). 2326 bool MakeNewInstruction = false; // If true, recreate using the following: 2327 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction; 2328 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. 2329 2330 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already. 2331 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) { 2332 bool isLastClause = i + 1 == e; 2333 if (LI.isCatch(i)) { 2334 // A catch clause. 2335 Constant *CatchClause = LI.getClause(i); 2336 Constant *TypeInfo = CatchClause->stripPointerCasts(); 2337 2338 // If we already saw this clause, there is no point in having a second 2339 // copy of it. 2340 if (AlreadyCaught.insert(TypeInfo).second) { 2341 // This catch clause was not already seen. 2342 NewClauses.push_back(CatchClause); 2343 } else { 2344 // Repeated catch clause - drop the redundant copy. 2345 MakeNewInstruction = true; 2346 } 2347 2348 // If this is a catch-all then there is no point in keeping any following 2349 // clauses or marking the landingpad as having a cleanup. 2350 if (isCatchAll(Personality, TypeInfo)) { 2351 if (!isLastClause) 2352 MakeNewInstruction = true; 2353 CleanupFlag = false; 2354 break; 2355 } 2356 } else { 2357 // A filter clause. If any of the filter elements were already caught 2358 // then they can be dropped from the filter. It is tempting to try to 2359 // exploit the filter further by saying that any typeinfo that does not 2360 // occur in the filter can't be caught later (and thus can be dropped). 2361 // However this would be wrong, since typeinfos can match without being 2362 // equal (for example if one represents a C++ class, and the other some 2363 // class derived from it). 2364 assert(LI.isFilter(i) && "Unsupported landingpad clause!"); 2365 Constant *FilterClause = LI.getClause(i); 2366 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType()); 2367 unsigned NumTypeInfos = FilterType->getNumElements(); 2368 2369 // An empty filter catches everything, so there is no point in keeping any 2370 // following clauses or marking the landingpad as having a cleanup. By 2371 // dealing with this case here the following code is made a bit simpler. 2372 if (!NumTypeInfos) { 2373 NewClauses.push_back(FilterClause); 2374 if (!isLastClause) 2375 MakeNewInstruction = true; 2376 CleanupFlag = false; 2377 break; 2378 } 2379 2380 bool MakeNewFilter = false; // If true, make a new filter. 2381 SmallVector<Constant *, 16> NewFilterElts; // New elements. 2382 if (isa<ConstantAggregateZero>(FilterClause)) { 2383 // Not an empty filter - it contains at least one null typeinfo. 2384 assert(NumTypeInfos > 0 && "Should have handled empty filter already!"); 2385 Constant *TypeInfo = 2386 Constant::getNullValue(FilterType->getElementType()); 2387 // If this typeinfo is a catch-all then the filter can never match. 2388 if (isCatchAll(Personality, TypeInfo)) { 2389 // Throw the filter away. 2390 MakeNewInstruction = true; 2391 continue; 2392 } 2393 2394 // There is no point in having multiple copies of this typeinfo, so 2395 // discard all but the first copy if there is more than one. 2396 NewFilterElts.push_back(TypeInfo); 2397 if (NumTypeInfos > 1) 2398 MakeNewFilter = true; 2399 } else { 2400 ConstantArray *Filter = cast<ConstantArray>(FilterClause); 2401 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements. 2402 NewFilterElts.reserve(NumTypeInfos); 2403 2404 // Remove any filter elements that were already caught or that already 2405 // occurred in the filter. While there, see if any of the elements are 2406 // catch-alls. If so, the filter can be discarded. 2407 bool SawCatchAll = false; 2408 for (unsigned j = 0; j != NumTypeInfos; ++j) { 2409 Constant *Elt = Filter->getOperand(j); 2410 Constant *TypeInfo = Elt->stripPointerCasts(); 2411 if (isCatchAll(Personality, TypeInfo)) { 2412 // This element is a catch-all. Bail out, noting this fact. 2413 SawCatchAll = true; 2414 break; 2415 } 2416 if (AlreadyCaught.count(TypeInfo)) 2417 // Already caught by an earlier clause, so having it in the filter 2418 // is pointless. 2419 continue; 2420 // There is no point in having multiple copies of the same typeinfo in 2421 // a filter, so only add it if we didn't already. 2422 if (SeenInFilter.insert(TypeInfo).second) 2423 NewFilterElts.push_back(cast<Constant>(Elt)); 2424 } 2425 // A filter containing a catch-all cannot match anything by definition. 2426 if (SawCatchAll) { 2427 // Throw the filter away. 2428 MakeNewInstruction = true; 2429 continue; 2430 } 2431 2432 // If we dropped something from the filter, make a new one. 2433 if (NewFilterElts.size() < NumTypeInfos) 2434 MakeNewFilter = true; 2435 } 2436 if (MakeNewFilter) { 2437 FilterType = ArrayType::get(FilterType->getElementType(), 2438 NewFilterElts.size()); 2439 FilterClause = ConstantArray::get(FilterType, NewFilterElts); 2440 MakeNewInstruction = true; 2441 } 2442 2443 NewClauses.push_back(FilterClause); 2444 2445 // If the new filter is empty then it will catch everything so there is 2446 // no point in keeping any following clauses or marking the landingpad 2447 // as having a cleanup. The case of the original filter being empty was 2448 // already handled above. 2449 if (MakeNewFilter && !NewFilterElts.size()) { 2450 assert(MakeNewInstruction && "New filter but not a new instruction!"); 2451 CleanupFlag = false; 2452 break; 2453 } 2454 } 2455 } 2456 2457 // If several filters occur in a row then reorder them so that the shortest 2458 // filters come first (those with the smallest number of elements). This is 2459 // advantageous because shorter filters are more likely to match, speeding up 2460 // unwinding, but mostly because it increases the effectiveness of the other 2461 // filter optimizations below. 2462 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) { 2463 unsigned j; 2464 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters. 2465 for (j = i; j != e; ++j) 2466 if (!isa<ArrayType>(NewClauses[j]->getType())) 2467 break; 2468 2469 // Check whether the filters are already sorted by length. We need to know 2470 // if sorting them is actually going to do anything so that we only make a 2471 // new landingpad instruction if it does. 2472 for (unsigned k = i; k + 1 < j; ++k) 2473 if (shorter_filter(NewClauses[k+1], NewClauses[k])) { 2474 // Not sorted, so sort the filters now. Doing an unstable sort would be 2475 // correct too but reordering filters pointlessly might confuse users. 2476 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j, 2477 shorter_filter); 2478 MakeNewInstruction = true; 2479 break; 2480 } 2481 2482 // Look for the next batch of filters. 2483 i = j + 1; 2484 } 2485 2486 // If typeinfos matched if and only if equal, then the elements of a filter L 2487 // that occurs later than a filter F could be replaced by the intersection of 2488 // the elements of F and L. In reality two typeinfos can match without being 2489 // equal (for example if one represents a C++ class, and the other some class 2490 // derived from it) so it would be wrong to perform this transform in general. 2491 // However the transform is correct and useful if F is a subset of L. In that 2492 // case L can be replaced by F, and thus removed altogether since repeating a 2493 // filter is pointless. So here we look at all pairs of filters F and L where 2494 // L follows F in the list of clauses, and remove L if every element of F is 2495 // an element of L. This can occur when inlining C++ functions with exception 2496 // specifications. 2497 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) { 2498 // Examine each filter in turn. 2499 Value *Filter = NewClauses[i]; 2500 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType()); 2501 if (!FTy) 2502 // Not a filter - skip it. 2503 continue; 2504 unsigned FElts = FTy->getNumElements(); 2505 // Examine each filter following this one. Doing this backwards means that 2506 // we don't have to worry about filters disappearing under us when removed. 2507 for (unsigned j = NewClauses.size() - 1; j != i; --j) { 2508 Value *LFilter = NewClauses[j]; 2509 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType()); 2510 if (!LTy) 2511 // Not a filter - skip it. 2512 continue; 2513 // If Filter is a subset of LFilter, i.e. every element of Filter is also 2514 // an element of LFilter, then discard LFilter. 2515 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j; 2516 // If Filter is empty then it is a subset of LFilter. 2517 if (!FElts) { 2518 // Discard LFilter. 2519 NewClauses.erase(J); 2520 MakeNewInstruction = true; 2521 // Move on to the next filter. 2522 continue; 2523 } 2524 unsigned LElts = LTy->getNumElements(); 2525 // If Filter is longer than LFilter then it cannot be a subset of it. 2526 if (FElts > LElts) 2527 // Move on to the next filter. 2528 continue; 2529 // At this point we know that LFilter has at least one element. 2530 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros. 2531 // Filter is a subset of LFilter iff Filter contains only zeros (as we 2532 // already know that Filter is not longer than LFilter). 2533 if (isa<ConstantAggregateZero>(Filter)) { 2534 assert(FElts <= LElts && "Should have handled this case earlier!"); 2535 // Discard LFilter. 2536 NewClauses.erase(J); 2537 MakeNewInstruction = true; 2538 } 2539 // Move on to the next filter. 2540 continue; 2541 } 2542 ConstantArray *LArray = cast<ConstantArray>(LFilter); 2543 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros. 2544 // Since Filter is non-empty and contains only zeros, it is a subset of 2545 // LFilter iff LFilter contains a zero. 2546 assert(FElts > 0 && "Should have eliminated the empty filter earlier!"); 2547 for (unsigned l = 0; l != LElts; ++l) 2548 if (LArray->getOperand(l)->isNullValue()) { 2549 // LFilter contains a zero - discard it. 2550 NewClauses.erase(J); 2551 MakeNewInstruction = true; 2552 break; 2553 } 2554 // Move on to the next filter. 2555 continue; 2556 } 2557 // At this point we know that both filters are ConstantArrays. Loop over 2558 // operands to see whether every element of Filter is also an element of 2559 // LFilter. Since filters tend to be short this is probably faster than 2560 // using a method that scales nicely. 2561 ConstantArray *FArray = cast<ConstantArray>(Filter); 2562 bool AllFound = true; 2563 for (unsigned f = 0; f != FElts; ++f) { 2564 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts(); 2565 AllFound = false; 2566 for (unsigned l = 0; l != LElts; ++l) { 2567 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts(); 2568 if (LTypeInfo == FTypeInfo) { 2569 AllFound = true; 2570 break; 2571 } 2572 } 2573 if (!AllFound) 2574 break; 2575 } 2576 if (AllFound) { 2577 // Discard LFilter. 2578 NewClauses.erase(J); 2579 MakeNewInstruction = true; 2580 } 2581 // Move on to the next filter. 2582 } 2583 } 2584 2585 // If we changed any of the clauses, replace the old landingpad instruction 2586 // with a new one. 2587 if (MakeNewInstruction) { 2588 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(), 2589 LI.getPersonalityFn(), 2590 NewClauses.size()); 2591 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i) 2592 NLI->addClause(NewClauses[i]); 2593 // A landing pad with no clauses must have the cleanup flag set. It is 2594 // theoretically possible, though highly unlikely, that we eliminated all 2595 // clauses. If so, force the cleanup flag to true. 2596 if (NewClauses.empty()) 2597 CleanupFlag = true; 2598 NLI->setCleanup(CleanupFlag); 2599 return NLI; 2600 } 2601 2602 // Even if none of the clauses changed, we may nonetheless have understood 2603 // that the cleanup flag is pointless. Clear it if so. 2604 if (LI.isCleanup() != CleanupFlag) { 2605 assert(!CleanupFlag && "Adding a cleanup, not removing one?!"); 2606 LI.setCleanup(CleanupFlag); 2607 return &LI; 2608 } 2609 2610 return nullptr; 2611 } 2612 2613 /// TryToSinkInstruction - Try to move the specified instruction from its 2614 /// current block into the beginning of DestBlock, which can only happen if it's 2615 /// safe to move the instruction past all of the instructions between it and the 2616 /// end of its block. 2617 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 2618 assert(I->hasOneUse() && "Invariants didn't hold!"); 2619 2620 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 2621 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() || 2622 isa<TerminatorInst>(I)) 2623 return false; 2624 2625 // Do not sink alloca instructions out of the entry block. 2626 if (isa<AllocaInst>(I) && I->getParent() == 2627 &DestBlock->getParent()->getEntryBlock()) 2628 return false; 2629 2630 // We can only sink load instructions if there is nothing between the load and 2631 // the end of block that could change the value. 2632 if (I->mayReadFromMemory()) { 2633 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 2634 Scan != E; ++Scan) 2635 if (Scan->mayWriteToMemory()) 2636 return false; 2637 } 2638 2639 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt(); 2640 I->moveBefore(InsertPos); 2641 ++NumSunkInst; 2642 return true; 2643 } 2644 2645 bool InstCombiner::run() { 2646 while (!Worklist.isEmpty()) { 2647 Instruction *I = Worklist.RemoveOne(); 2648 if (I == nullptr) continue; // skip null values. 2649 2650 // Check to see if we can DCE the instruction. 2651 if (isInstructionTriviallyDead(I, TLI)) { 2652 DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); 2653 EraseInstFromFunction(*I); 2654 ++NumDeadInst; 2655 MadeIRChange = true; 2656 continue; 2657 } 2658 2659 // Instruction isn't dead, see if we can constant propagate it. 2660 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) { 2661 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) { 2662 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 2663 2664 // Add operands to the worklist. 2665 ReplaceInstUsesWith(*I, C); 2666 ++NumConstProp; 2667 EraseInstFromFunction(*I); 2668 MadeIRChange = true; 2669 continue; 2670 } 2671 } 2672 2673 // See if we can trivially sink this instruction to a successor basic block. 2674 if (I->hasOneUse()) { 2675 BasicBlock *BB = I->getParent(); 2676 Instruction *UserInst = cast<Instruction>(*I->user_begin()); 2677 BasicBlock *UserParent; 2678 2679 // Get the block the use occurs in. 2680 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 2681 UserParent = PN->getIncomingBlock(*I->use_begin()); 2682 else 2683 UserParent = UserInst->getParent(); 2684 2685 if (UserParent != BB) { 2686 bool UserIsSuccessor = false; 2687 // See if the user is one of our successors. 2688 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 2689 if (*SI == UserParent) { 2690 UserIsSuccessor = true; 2691 break; 2692 } 2693 2694 // If the user is one of our immediate successors, and if that successor 2695 // only has us as a predecessors (we'd have to split the critical edge 2696 // otherwise), we can keep going. 2697 if (UserIsSuccessor && UserParent->getSinglePredecessor()) { 2698 // Okay, the CFG is simple enough, try to sink this instruction. 2699 if (TryToSinkInstruction(I, UserParent)) { 2700 MadeIRChange = true; 2701 // We'll add uses of the sunk instruction below, but since sinking 2702 // can expose opportunities for it's *operands* add them to the 2703 // worklist 2704 for (Use &U : I->operands()) 2705 if (Instruction *OpI = dyn_cast<Instruction>(U.get())) 2706 Worklist.Add(OpI); 2707 } 2708 } 2709 } 2710 } 2711 2712 // Now that we have an instruction, try combining it to simplify it. 2713 Builder->SetInsertPoint(I->getParent(), I); 2714 Builder->SetCurrentDebugLocation(I->getDebugLoc()); 2715 2716 #ifndef NDEBUG 2717 std::string OrigI; 2718 #endif 2719 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 2720 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); 2721 2722 if (Instruction *Result = visit(*I)) { 2723 ++NumCombined; 2724 // Should we replace the old instruction with a new one? 2725 if (Result != I) { 2726 DEBUG(dbgs() << "IC: Old = " << *I << '\n' 2727 << " New = " << *Result << '\n'); 2728 2729 if (I->getDebugLoc()) 2730 Result->setDebugLoc(I->getDebugLoc()); 2731 // Everything uses the new instruction now. 2732 I->replaceAllUsesWith(Result); 2733 2734 // Move the name to the new instruction first. 2735 Result->takeName(I); 2736 2737 // Push the new instruction and any users onto the worklist. 2738 Worklist.Add(Result); 2739 Worklist.AddUsersToWorkList(*Result); 2740 2741 // Insert the new instruction into the basic block... 2742 BasicBlock *InstParent = I->getParent(); 2743 BasicBlock::iterator InsertPos = I; 2744 2745 // If we replace a PHI with something that isn't a PHI, fix up the 2746 // insertion point. 2747 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos)) 2748 InsertPos = InstParent->getFirstInsertionPt(); 2749 2750 InstParent->getInstList().insert(InsertPos, Result); 2751 2752 EraseInstFromFunction(*I); 2753 } else { 2754 #ifndef NDEBUG 2755 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' 2756 << " New = " << *I << '\n'); 2757 #endif 2758 2759 // If the instruction was modified, it's possible that it is now dead. 2760 // if so, remove it. 2761 if (isInstructionTriviallyDead(I, TLI)) { 2762 EraseInstFromFunction(*I); 2763 } else { 2764 Worklist.Add(I); 2765 Worklist.AddUsersToWorkList(*I); 2766 } 2767 } 2768 MadeIRChange = true; 2769 } 2770 } 2771 2772 Worklist.Zap(); 2773 return MadeIRChange; 2774 } 2775 2776 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 2777 /// all reachable code to the worklist. 2778 /// 2779 /// This has a couple of tricks to make the code faster and more powerful. In 2780 /// particular, we constant fold and DCE instructions as we go, to avoid adding 2781 /// them to the worklist (this significantly speeds up instcombine on code where 2782 /// many instructions are dead or constant). Additionally, if we find a branch 2783 /// whose condition is a known constant, we only visit the reachable successors. 2784 /// 2785 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL, 2786 SmallPtrSetImpl<BasicBlock *> &Visited, 2787 InstCombineWorklist &ICWorklist, 2788 const TargetLibraryInfo *TLI) { 2789 bool MadeIRChange = false; 2790 SmallVector<BasicBlock*, 256> Worklist; 2791 Worklist.push_back(BB); 2792 2793 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 2794 DenseMap<ConstantExpr*, Constant*> FoldedConstants; 2795 2796 do { 2797 BB = Worklist.pop_back_val(); 2798 2799 // We have now visited this block! If we've already been here, ignore it. 2800 if (!Visited.insert(BB).second) 2801 continue; 2802 2803 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 2804 Instruction *Inst = BBI++; 2805 2806 // DCE instruction if trivially dead. 2807 if (isInstructionTriviallyDead(Inst, TLI)) { 2808 ++NumDeadInst; 2809 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); 2810 Inst->eraseFromParent(); 2811 continue; 2812 } 2813 2814 // ConstantProp instruction if trivially constant. 2815 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 2816 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) { 2817 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " 2818 << *Inst << '\n'); 2819 Inst->replaceAllUsesWith(C); 2820 ++NumConstProp; 2821 Inst->eraseFromParent(); 2822 continue; 2823 } 2824 2825 // See if we can constant fold its operands. 2826 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e; 2827 ++i) { 2828 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 2829 if (CE == nullptr) 2830 continue; 2831 2832 Constant *&FoldRes = FoldedConstants[CE]; 2833 if (!FoldRes) 2834 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); 2835 if (!FoldRes) 2836 FoldRes = CE; 2837 2838 if (FoldRes != CE) { 2839 *i = FoldRes; 2840 MadeIRChange = true; 2841 } 2842 } 2843 2844 InstrsForInstCombineWorklist.push_back(Inst); 2845 } 2846 2847 // Recursively visit successors. If this is a branch or switch on a 2848 // constant, only visit the reachable successor. 2849 TerminatorInst *TI = BB->getTerminator(); 2850 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2851 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 2852 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 2853 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 2854 Worklist.push_back(ReachableBB); 2855 continue; 2856 } 2857 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2858 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 2859 // See if this is an explicit destination. 2860 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2861 i != e; ++i) 2862 if (i.getCaseValue() == Cond) { 2863 BasicBlock *ReachableBB = i.getCaseSuccessor(); 2864 Worklist.push_back(ReachableBB); 2865 continue; 2866 } 2867 2868 // Otherwise it is the default destination. 2869 Worklist.push_back(SI->getDefaultDest()); 2870 continue; 2871 } 2872 } 2873 2874 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 2875 Worklist.push_back(TI->getSuccessor(i)); 2876 } while (!Worklist.empty()); 2877 2878 // Once we've found all of the instructions to add to instcombine's worklist, 2879 // add them in reverse order. This way instcombine will visit from the top 2880 // of the function down. This jives well with the way that it adds all uses 2881 // of instructions to the worklist after doing a transformation, thus avoiding 2882 // some N^2 behavior in pathological cases. 2883 ICWorklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 2884 InstrsForInstCombineWorklist.size()); 2885 2886 return MadeIRChange; 2887 } 2888 2889 /// \brief Populate the IC worklist from a function, and prune any dead basic 2890 /// blocks discovered in the process. 2891 /// 2892 /// This also does basic constant propagation and other forward fixing to make 2893 /// the combiner itself run much faster. 2894 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL, 2895 TargetLibraryInfo *TLI, 2896 InstCombineWorklist &ICWorklist) { 2897 bool MadeIRChange = false; 2898 2899 // Do a depth-first traversal of the function, populate the worklist with 2900 // the reachable instructions. Ignore blocks that are not reachable. Keep 2901 // track of which blocks we visit. 2902 SmallPtrSet<BasicBlock *, 64> Visited; 2903 MadeIRChange |= 2904 AddReachableCodeToWorklist(F.begin(), DL, Visited, ICWorklist, TLI); 2905 2906 // Do a quick scan over the function. If we find any blocks that are 2907 // unreachable, remove any instructions inside of them. This prevents 2908 // the instcombine code from having to deal with some bad special cases. 2909 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 2910 if (Visited.count(BB)) 2911 continue; 2912 2913 // Delete the instructions backwards, as it has a reduced likelihood of 2914 // having to update as many def-use and use-def chains. 2915 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 2916 while (EndInst != BB->begin()) { 2917 // Delete the next to last instruction. 2918 BasicBlock::iterator I = EndInst; 2919 Instruction *Inst = --I; 2920 if (!Inst->use_empty()) 2921 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 2922 if (isa<LandingPadInst>(Inst)) { 2923 EndInst = Inst; 2924 continue; 2925 } 2926 if (!isa<DbgInfoIntrinsic>(Inst)) { 2927 ++NumDeadInst; 2928 MadeIRChange = true; 2929 } 2930 Inst->eraseFromParent(); 2931 } 2932 } 2933 2934 return MadeIRChange; 2935 } 2936 2937 static bool 2938 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist, 2939 AssumptionCache &AC, TargetLibraryInfo &TLI, 2940 DominatorTree &DT, LoopInfo *LI = nullptr) { 2941 // Minimizing size? 2942 bool MinimizeSize = F.hasFnAttribute(Attribute::MinSize); 2943 auto &DL = F.getParent()->getDataLayout(); 2944 2945 /// Builder - This is an IRBuilder that automatically inserts new 2946 /// instructions into the worklist when they are created. 2947 IRBuilder<true, TargetFolder, InstCombineIRInserter> Builder( 2948 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, &AC)); 2949 2950 // Lower dbg.declare intrinsics otherwise their value may be clobbered 2951 // by instcombiner. 2952 bool DbgDeclaresChanged = LowerDbgDeclare(F); 2953 2954 // Iterate while there is work to do. 2955 int Iteration = 0; 2956 for (;;) { 2957 ++Iteration; 2958 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 2959 << F.getName() << "\n"); 2960 2961 bool Changed = false; 2962 if (prepareICWorklistFromFunction(F, DL, &TLI, Worklist)) 2963 Changed = true; 2964 2965 InstCombiner IC(Worklist, &Builder, MinimizeSize, &AC, &TLI, &DT, DL, LI); 2966 if (IC.run()) 2967 Changed = true; 2968 2969 if (!Changed) 2970 break; 2971 } 2972 2973 return DbgDeclaresChanged || Iteration > 1; 2974 } 2975 2976 PreservedAnalyses InstCombinePass::run(Function &F, 2977 AnalysisManager<Function> *AM) { 2978 auto &AC = AM->getResult<AssumptionAnalysis>(F); 2979 auto &DT = AM->getResult<DominatorTreeAnalysis>(F); 2980 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F); 2981 2982 auto *LI = AM->getCachedResult<LoopAnalysis>(F); 2983 2984 if (!combineInstructionsOverFunction(F, Worklist, AC, TLI, DT, LI)) 2985 // No changes, all analyses are preserved. 2986 return PreservedAnalyses::all(); 2987 2988 // Mark all the analyses that instcombine updates as preserved. 2989 // FIXME: Need a way to preserve CFG analyses here! 2990 PreservedAnalyses PA; 2991 PA.preserve<DominatorTreeAnalysis>(); 2992 return PA; 2993 } 2994 2995 namespace { 2996 /// \brief The legacy pass manager's instcombine pass. 2997 /// 2998 /// This is a basic whole-function wrapper around the instcombine utility. It 2999 /// will try to combine all instructions in the function. 3000 class InstructionCombiningPass : public FunctionPass { 3001 InstCombineWorklist Worklist; 3002 3003 public: 3004 static char ID; // Pass identification, replacement for typeid 3005 3006 InstructionCombiningPass() : FunctionPass(ID) { 3007 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry()); 3008 } 3009 3010 void getAnalysisUsage(AnalysisUsage &AU) const override; 3011 bool runOnFunction(Function &F) override; 3012 }; 3013 } 3014 3015 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const { 3016 AU.setPreservesCFG(); 3017 AU.addRequired<AssumptionCacheTracker>(); 3018 AU.addRequired<TargetLibraryInfoWrapperPass>(); 3019 AU.addRequired<DominatorTreeWrapperPass>(); 3020 AU.addPreserved<DominatorTreeWrapperPass>(); 3021 } 3022 3023 bool InstructionCombiningPass::runOnFunction(Function &F) { 3024 if (skipOptnoneFunction(F)) 3025 return false; 3026 3027 // Required analyses. 3028 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 3029 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 3030 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 3031 3032 // Optional analyses. 3033 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); 3034 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr; 3035 3036 return combineInstructionsOverFunction(F, Worklist, AC, TLI, DT, LI); 3037 } 3038 3039 char InstructionCombiningPass::ID = 0; 3040 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine", 3041 "Combine redundant instructions", false, false) 3042 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 3043 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 3044 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 3045 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine", 3046 "Combine redundant instructions", false, false) 3047 3048 // Initialization Routines 3049 void llvm::initializeInstCombine(PassRegistry &Registry) { 3050 initializeInstructionCombiningPassPass(Registry); 3051 } 3052 3053 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 3054 initializeInstructionCombiningPassPass(*unwrap(R)); 3055 } 3056 3057 FunctionPass *llvm::createInstructionCombiningPass() { 3058 return new InstructionCombiningPass(); 3059 } 3060