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/EHPersonalities.h" 46 #include "llvm/Analysis/GlobalsModRef.h" 47 #include "llvm/Analysis/InstructionSimplify.h" 48 #include "llvm/Analysis/LoopInfo.h" 49 #include "llvm/Analysis/MemoryBuiltins.h" 50 #include "llvm/Analysis/TargetLibraryInfo.h" 51 #include "llvm/Analysis/ValueTracking.h" 52 #include "llvm/IR/CFG.h" 53 #include "llvm/IR/DataLayout.h" 54 #include "llvm/IR/Dominators.h" 55 #include "llvm/IR/GetElementPtrTypeIterator.h" 56 #include "llvm/IR/IntrinsicInst.h" 57 #include "llvm/IR/PatternMatch.h" 58 #include "llvm/IR/ValueHandle.h" 59 #include "llvm/Support/CommandLine.h" 60 #include "llvm/Support/Debug.h" 61 #include "llvm/Support/raw_ostream.h" 62 #include "llvm/Transforms/Scalar.h" 63 #include "llvm/Transforms/Utils/Local.h" 64 #include <algorithm> 65 #include <climits> 66 using namespace llvm; 67 using namespace llvm::PatternMatch; 68 69 #define DEBUG_TYPE "instcombine" 70 71 STATISTIC(NumCombined , "Number of insts combined"); 72 STATISTIC(NumConstProp, "Number of constant folds"); 73 STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 74 STATISTIC(NumSunkInst , "Number of instructions sunk"); 75 STATISTIC(NumExpand, "Number of expansions"); 76 STATISTIC(NumFactor , "Number of factorizations"); 77 STATISTIC(NumReassoc , "Number of reassociations"); 78 79 Value *InstCombiner::EmitGEPOffset(User *GEP) { 80 return llvm::EmitGEPOffset(Builder, DL, GEP); 81 } 82 83 /// Return true if it is desirable to convert an integer computation from a 84 /// given bit width to a new bit width. 85 /// We don't want to convert from a legal to an illegal type for example or from 86 /// a smaller to a larger illegal type. 87 bool InstCombiner::ShouldChangeType(unsigned FromWidth, 88 unsigned ToWidth) const { 89 bool FromLegal = DL.isLegalInteger(FromWidth); 90 bool ToLegal = DL.isLegalInteger(ToWidth); 91 92 // If this is a legal integer from type, and the result would be an illegal 93 // type, don't do the transformation. 94 if (FromLegal && !ToLegal) 95 return false; 96 97 // Otherwise, if both are illegal, do not increase the size of the result. We 98 // do allow things like i160 -> i64, but not i64 -> i160. 99 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 100 return false; 101 102 return true; 103 } 104 105 /// Return true if it is desirable to convert a computation from 'From' to 'To'. 106 /// We don't want to convert from a legal to an illegal type for example or from 107 /// a smaller to a larger illegal type. 108 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const { 109 assert(From->isIntegerTy() && To->isIntegerTy()); 110 111 unsigned FromWidth = From->getPrimitiveSizeInBits(); 112 unsigned ToWidth = To->getPrimitiveSizeInBits(); 113 return ShouldChangeType(FromWidth, ToWidth); 114 } 115 116 // Return true, if No Signed Wrap should be maintained for I. 117 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C", 118 // where both B and C should be ConstantInts, results in a constant that does 119 // not overflow. This function only handles the Add and Sub opcodes. For 120 // all other opcodes, the function conservatively returns false. 121 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { 122 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I); 123 if (!OBO || !OBO->hasNoSignedWrap()) { 124 return false; 125 } 126 127 // We reason about Add and Sub Only. 128 Instruction::BinaryOps Opcode = I.getOpcode(); 129 if (Opcode != Instruction::Add && 130 Opcode != Instruction::Sub) { 131 return false; 132 } 133 134 ConstantInt *CB = dyn_cast<ConstantInt>(B); 135 ConstantInt *CC = dyn_cast<ConstantInt>(C); 136 137 if (!CB || !CC) { 138 return false; 139 } 140 141 const APInt &BVal = CB->getValue(); 142 const APInt &CVal = CC->getValue(); 143 bool Overflow = false; 144 145 if (Opcode == Instruction::Add) { 146 BVal.sadd_ov(CVal, Overflow); 147 } else { 148 BVal.ssub_ov(CVal, Overflow); 149 } 150 151 return !Overflow; 152 } 153 154 /// Conservatively clears subclassOptionalData after a reassociation or 155 /// commutation. We preserve fast-math flags when applicable as they can be 156 /// preserved. 157 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { 158 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I); 159 if (!FPMO) { 160 I.clearSubclassOptionalData(); 161 return; 162 } 163 164 FastMathFlags FMF = I.getFastMathFlags(); 165 I.clearSubclassOptionalData(); 166 I.setFastMathFlags(FMF); 167 } 168 169 /// This performs a few simplifications for operators that are associative or 170 /// commutative: 171 /// 172 /// Commutative operators: 173 /// 174 /// 1. Order operands such that they are listed from right (least complex) to 175 /// left (most complex). This puts constants before unary operators before 176 /// binary operators. 177 /// 178 /// Associative operators: 179 /// 180 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 181 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 182 /// 183 /// Associative and commutative operators: 184 /// 185 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 186 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 187 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 188 /// if C1 and C2 are constants. 189 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 190 Instruction::BinaryOps Opcode = I.getOpcode(); 191 bool Changed = false; 192 193 do { 194 // Order operands such that they are listed from right (least complex) to 195 // left (most complex). This puts constants before unary operators before 196 // binary operators. 197 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 198 getComplexity(I.getOperand(1))) 199 Changed = !I.swapOperands(); 200 201 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 202 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 203 204 if (I.isAssociative()) { 205 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 206 if (Op0 && Op0->getOpcode() == Opcode) { 207 Value *A = Op0->getOperand(0); 208 Value *B = Op0->getOperand(1); 209 Value *C = I.getOperand(1); 210 211 // Does "B op C" simplify? 212 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) { 213 // It simplifies to V. Form "A op V". 214 I.setOperand(0, A); 215 I.setOperand(1, V); 216 // Conservatively clear the optional flags, since they may not be 217 // preserved by the reassociation. 218 if (MaintainNoSignedWrap(I, B, C) && 219 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) { 220 // Note: this is only valid because SimplifyBinOp doesn't look at 221 // the operands to Op0. 222 I.clearSubclassOptionalData(); 223 I.setHasNoSignedWrap(true); 224 } else { 225 ClearSubclassDataAfterReassociation(I); 226 } 227 228 Changed = true; 229 ++NumReassoc; 230 continue; 231 } 232 } 233 234 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 235 if (Op1 && Op1->getOpcode() == Opcode) { 236 Value *A = I.getOperand(0); 237 Value *B = Op1->getOperand(0); 238 Value *C = Op1->getOperand(1); 239 240 // Does "A op B" simplify? 241 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) { 242 // It simplifies to V. Form "V op C". 243 I.setOperand(0, V); 244 I.setOperand(1, C); 245 // Conservatively clear the optional flags, since they may not be 246 // preserved by the reassociation. 247 ClearSubclassDataAfterReassociation(I); 248 Changed = true; 249 ++NumReassoc; 250 continue; 251 } 252 } 253 } 254 255 if (I.isAssociative() && I.isCommutative()) { 256 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 257 if (Op0 && Op0->getOpcode() == Opcode) { 258 Value *A = Op0->getOperand(0); 259 Value *B = Op0->getOperand(1); 260 Value *C = I.getOperand(1); 261 262 // Does "C op A" simplify? 263 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 264 // It simplifies to V. Form "V op B". 265 I.setOperand(0, V); 266 I.setOperand(1, B); 267 // Conservatively clear the optional flags, since they may not be 268 // preserved by the reassociation. 269 ClearSubclassDataAfterReassociation(I); 270 Changed = true; 271 ++NumReassoc; 272 continue; 273 } 274 } 275 276 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 277 if (Op1 && Op1->getOpcode() == Opcode) { 278 Value *A = I.getOperand(0); 279 Value *B = Op1->getOperand(0); 280 Value *C = Op1->getOperand(1); 281 282 // Does "C op A" simplify? 283 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 284 // It simplifies to V. Form "B op V". 285 I.setOperand(0, B); 286 I.setOperand(1, V); 287 // Conservatively clear the optional flags, since they may not be 288 // preserved by the reassociation. 289 ClearSubclassDataAfterReassociation(I); 290 Changed = true; 291 ++NumReassoc; 292 continue; 293 } 294 } 295 296 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 297 // if C1 and C2 are constants. 298 if (Op0 && Op1 && 299 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 300 isa<Constant>(Op0->getOperand(1)) && 301 isa<Constant>(Op1->getOperand(1)) && 302 Op0->hasOneUse() && Op1->hasOneUse()) { 303 Value *A = Op0->getOperand(0); 304 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 305 Value *B = Op1->getOperand(0); 306 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 307 308 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 309 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B); 310 if (isa<FPMathOperator>(New)) { 311 FastMathFlags Flags = I.getFastMathFlags(); 312 Flags &= Op0->getFastMathFlags(); 313 Flags &= Op1->getFastMathFlags(); 314 New->setFastMathFlags(Flags); 315 } 316 InsertNewInstWith(New, I); 317 New->takeName(Op1); 318 I.setOperand(0, New); 319 I.setOperand(1, Folded); 320 // Conservatively clear the optional flags, since they may not be 321 // preserved by the reassociation. 322 ClearSubclassDataAfterReassociation(I); 323 324 Changed = true; 325 continue; 326 } 327 } 328 329 // No further simplifications. 330 return Changed; 331 } while (1); 332 } 333 334 /// Return whether "X LOp (Y ROp Z)" is always equal to 335 /// "(X LOp Y) ROp (X LOp Z)". 336 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp, 337 Instruction::BinaryOps ROp) { 338 switch (LOp) { 339 default: 340 return false; 341 342 case Instruction::And: 343 // And distributes over Or and Xor. 344 switch (ROp) { 345 default: 346 return false; 347 case Instruction::Or: 348 case Instruction::Xor: 349 return true; 350 } 351 352 case Instruction::Mul: 353 // Multiplication distributes over addition and subtraction. 354 switch (ROp) { 355 default: 356 return false; 357 case Instruction::Add: 358 case Instruction::Sub: 359 return true; 360 } 361 362 case Instruction::Or: 363 // Or distributes over And. 364 switch (ROp) { 365 default: 366 return false; 367 case Instruction::And: 368 return true; 369 } 370 } 371 } 372 373 /// Return whether "(X LOp Y) ROp Z" is always equal to 374 /// "(X ROp Z) LOp (Y ROp Z)". 375 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, 376 Instruction::BinaryOps ROp) { 377 if (Instruction::isCommutative(ROp)) 378 return LeftDistributesOverRight(ROp, LOp); 379 380 switch (LOp) { 381 default: 382 return false; 383 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. 384 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts. 385 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts. 386 case Instruction::And: 387 case Instruction::Or: 388 case Instruction::Xor: 389 switch (ROp) { 390 default: 391 return false; 392 case Instruction::Shl: 393 case Instruction::LShr: 394 case Instruction::AShr: 395 return true; 396 } 397 } 398 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", 399 // but this requires knowing that the addition does not overflow and other 400 // such subtleties. 401 return false; 402 } 403 404 /// This function returns identity value for given opcode, which can be used to 405 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1). 406 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) { 407 if (isa<Constant>(V)) 408 return nullptr; 409 410 if (OpCode == Instruction::Mul) 411 return ConstantInt::get(V->getType(), 1); 412 413 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc. 414 415 return nullptr; 416 } 417 418 /// This function factors binary ops which can be combined using distributive 419 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable 420 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called 421 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms 422 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and 423 /// RHS to 4. 424 static Instruction::BinaryOps 425 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode, 426 BinaryOperator *Op, Value *&LHS, Value *&RHS) { 427 if (!Op) 428 return Instruction::BinaryOpsEnd; 429 430 LHS = Op->getOperand(0); 431 RHS = Op->getOperand(1); 432 433 switch (TopLevelOpcode) { 434 default: 435 return Op->getOpcode(); 436 437 case Instruction::Add: 438 case Instruction::Sub: 439 if (Op->getOpcode() == Instruction::Shl) { 440 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) { 441 // The multiplier is really 1 << CST. 442 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST); 443 return Instruction::Mul; 444 } 445 } 446 return Op->getOpcode(); 447 } 448 449 // TODO: We can add other conversions e.g. shr => div etc. 450 } 451 452 /// This tries to simplify binary operations by factorizing out common terms 453 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 454 static Value *tryFactorization(InstCombiner::BuilderTy *Builder, 455 const DataLayout &DL, BinaryOperator &I, 456 Instruction::BinaryOps InnerOpcode, Value *A, 457 Value *B, Value *C, Value *D) { 458 459 // If any of A, B, C, D are null, we can not factor I, return early. 460 // Checking A and C should be enough. 461 if (!A || !C || !B || !D) 462 return nullptr; 463 464 Value *V = nullptr; 465 Value *SimplifiedInst = nullptr; 466 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 467 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); 468 469 // Does "X op' Y" always equal "Y op' X"? 470 bool InnerCommutative = Instruction::isCommutative(InnerOpcode); 471 472 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? 473 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) 474 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the 475 // commutative case, "(A op' B) op (C op' A)"? 476 if (A == C || (InnerCommutative && A == D)) { 477 if (A != C) 478 std::swap(C, D); 479 // Consider forming "A op' (B op D)". 480 // If "B op D" simplifies then it can be formed with no cost. 481 V = SimplifyBinOp(TopLevelOpcode, B, D, DL); 482 // If "B op D" doesn't simplify then only go on if both of the existing 483 // operations "A op' B" and "C op' D" will be zapped as no longer used. 484 if (!V && LHS->hasOneUse() && RHS->hasOneUse()) 485 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName()); 486 if (V) { 487 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V); 488 } 489 } 490 491 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? 492 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) 493 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the 494 // commutative case, "(A op' B) op (B op' D)"? 495 if (B == D || (InnerCommutative && B == C)) { 496 if (B != D) 497 std::swap(C, D); 498 // Consider forming "(A op C) op' B". 499 // If "A op C" simplifies then it can be formed with no cost. 500 V = SimplifyBinOp(TopLevelOpcode, A, C, DL); 501 502 // If "A op C" doesn't simplify then only go on if both of the existing 503 // operations "A op' B" and "C op' D" will be zapped as no longer used. 504 if (!V && LHS->hasOneUse() && RHS->hasOneUse()) 505 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName()); 506 if (V) { 507 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B); 508 } 509 } 510 511 if (SimplifiedInst) { 512 ++NumFactor; 513 SimplifiedInst->takeName(&I); 514 515 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag. 516 // TODO: Check for NUW. 517 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) { 518 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) { 519 bool HasNSW = false; 520 if (isa<OverflowingBinaryOperator>(&I)) 521 HasNSW = I.hasNoSignedWrap(); 522 523 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 524 if (isa<OverflowingBinaryOperator>(Op0)) 525 HasNSW &= Op0->hasNoSignedWrap(); 526 527 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 528 if (isa<OverflowingBinaryOperator>(Op1)) 529 HasNSW &= Op1->hasNoSignedWrap(); 530 531 // We can propagate 'nsw' if we know that 532 // %Y = mul nsw i16 %X, C 533 // %Z = add nsw i16 %Y, %X 534 // => 535 // %Z = mul nsw i16 %X, C+1 536 // 537 // iff C+1 isn't INT_MIN 538 const APInt *CInt; 539 if (TopLevelOpcode == Instruction::Add && 540 InnerOpcode == Instruction::Mul) 541 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue()) 542 BO->setHasNoSignedWrap(HasNSW); 543 } 544 } 545 } 546 return SimplifiedInst; 547 } 548 549 /// This tries to simplify binary operations which some other binary operation 550 /// distributes over either by factorizing out common terms 551 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in 552 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win). 553 /// Returns the simplified value, or null if it didn't simplify. 554 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { 555 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 556 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 557 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 558 559 // Factorization. 560 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 561 auto TopLevelOpcode = I.getOpcode(); 562 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B); 563 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D); 564 565 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize 566 // a common term. 567 if (LHSOpcode == RHSOpcode) { 568 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D)) 569 return V; 570 } 571 572 // The instruction has the form "(A op' B) op (C)". Try to factorize common 573 // term. 574 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS, 575 getIdentityValue(LHSOpcode, RHS))) 576 return V; 577 578 // The instruction has the form "(B) op (C op' D)". Try to factorize common 579 // term. 580 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS, 581 getIdentityValue(RHSOpcode, LHS), C, D)) 582 return V; 583 584 // Expansion. 585 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { 586 // The instruction has the form "(A op' B) op C". See if expanding it out 587 // to "(A op C) op' (B op C)" results in simplifications. 588 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 589 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 590 591 // Do "A op C" and "B op C" both simplify? 592 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL)) 593 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) { 594 // They do! Return "L op' R". 595 ++NumExpand; 596 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 597 if ((L == A && R == B) || 598 (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) 599 return Op0; 600 // Otherwise return "L op' R" if it simplifies. 601 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 602 return V; 603 // Otherwise, create a new instruction. 604 C = Builder->CreateBinOp(InnerOpcode, L, R); 605 C->takeName(&I); 606 return C; 607 } 608 } 609 610 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) { 611 // The instruction has the form "A op (B op' C)". See if expanding it out 612 // to "(A op B) op' (A op C)" results in simplifications. 613 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 614 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' 615 616 // Do "A op B" and "A op C" both simplify? 617 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL)) 618 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) { 619 // They do! Return "L op' R". 620 ++NumExpand; 621 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 622 if ((L == B && R == C) || 623 (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) 624 return Op1; 625 // Otherwise return "L op' R" if it simplifies. 626 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 627 return V; 628 // Otherwise, create a new instruction. 629 A = Builder->CreateBinOp(InnerOpcode, L, R); 630 A->takeName(&I); 631 return A; 632 } 633 } 634 635 // (op (select (a, c, b)), (select (a, d, b))) -> (select (a, (op c, d), 0)) 636 // (op (select (a, b, c)), (select (a, b, d))) -> (select (a, 0, (op c, d))) 637 if (auto *SI0 = dyn_cast<SelectInst>(LHS)) { 638 if (auto *SI1 = dyn_cast<SelectInst>(RHS)) { 639 if (SI0->getCondition() == SI1->getCondition()) { 640 Value *SI = nullptr; 641 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getFalseValue(), 642 SI1->getFalseValue(), DL, TLI, DT, AC)) 643 SI = Builder->CreateSelect(SI0->getCondition(), 644 Builder->CreateBinOp(TopLevelOpcode, 645 SI0->getTrueValue(), 646 SI1->getTrueValue()), 647 V); 648 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getTrueValue(), 649 SI1->getTrueValue(), DL, TLI, DT, AC)) 650 SI = Builder->CreateSelect( 651 SI0->getCondition(), V, 652 Builder->CreateBinOp(TopLevelOpcode, SI0->getFalseValue(), 653 SI1->getFalseValue())); 654 if (SI) { 655 SI->takeName(&I); 656 return SI; 657 } 658 } 659 } 660 } 661 662 return nullptr; 663 } 664 665 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a 666 /// constant zero (which is the 'negate' form). 667 Value *InstCombiner::dyn_castNegVal(Value *V) const { 668 if (BinaryOperator::isNeg(V)) 669 return BinaryOperator::getNegArgument(V); 670 671 // Constants can be considered to be negated values if they can be folded. 672 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 673 return ConstantExpr::getNeg(C); 674 675 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 676 if (C->getType()->getElementType()->isIntegerTy()) 677 return ConstantExpr::getNeg(C); 678 679 return nullptr; 680 } 681 682 /// Given a 'fsub' instruction, return the RHS of the instruction if the LHS is 683 /// a constant negative zero (which is the 'negate' form). 684 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const { 685 if (BinaryOperator::isFNeg(V, IgnoreZeroSign)) 686 return BinaryOperator::getFNegArgument(V); 687 688 // Constants can be considered to be negated values if they can be folded. 689 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 690 return ConstantExpr::getFNeg(C); 691 692 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 693 if (C->getType()->getElementType()->isFloatingPointTy()) 694 return ConstantExpr::getFNeg(C); 695 696 return nullptr; 697 } 698 699 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 700 InstCombiner *IC) { 701 if (CastInst *CI = dyn_cast<CastInst>(&I)) { 702 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 703 } 704 705 // Figure out if the constant is the left or the right argument. 706 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 707 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 708 709 if (Constant *SOC = dyn_cast<Constant>(SO)) { 710 if (ConstIsRHS) 711 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 712 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 713 } 714 715 Value *Op0 = SO, *Op1 = ConstOperand; 716 if (!ConstIsRHS) 717 std::swap(Op0, Op1); 718 719 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) { 720 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 721 SO->getName()+".op"); 722 Instruction *FPInst = dyn_cast<Instruction>(RI); 723 if (FPInst && isa<FPMathOperator>(FPInst)) 724 FPInst->copyFastMathFlags(BO); 725 return RI; 726 } 727 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 728 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 729 SO->getName()+".cmp"); 730 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 731 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 732 SO->getName()+".cmp"); 733 llvm_unreachable("Unknown binary instruction type!"); 734 } 735 736 /// Given an instruction with a select as one operand and a constant as the 737 /// other operand, try to fold the binary operator into the select arguments. 738 /// This also works for Cast instructions, which obviously do not have a second 739 /// operand. 740 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 741 // Don't modify shared select instructions 742 if (!SI->hasOneUse()) return nullptr; 743 Value *TV = SI->getOperand(1); 744 Value *FV = SI->getOperand(2); 745 746 if (isa<Constant>(TV) || isa<Constant>(FV)) { 747 // Bool selects with constant operands can be folded to logical ops. 748 if (SI->getType()->isIntegerTy(1)) return nullptr; 749 750 // If it's a bitcast involving vectors, make sure it has the same number of 751 // elements on both sides. 752 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) { 753 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy()); 754 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy()); 755 756 // Verify that either both or neither are vectors. 757 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr; 758 // If vectors, verify that they have the same number of elements. 759 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) 760 return nullptr; 761 } 762 763 // Test if a CmpInst instruction is used exclusively by a select as 764 // part of a minimum or maximum operation. If so, refrain from doing 765 // any other folding. This helps out other analyses which understand 766 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 767 // and CodeGen. And in this case, at least one of the comparison 768 // operands has at least one user besides the compare (the select), 769 // which would often largely negate the benefit of folding anyway. 770 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) { 771 if (CI->hasOneUse()) { 772 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1); 773 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 774 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 775 return nullptr; 776 } 777 } 778 779 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 780 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 781 782 return SelectInst::Create(SI->getCondition(), 783 SelectTrueVal, SelectFalseVal); 784 } 785 return nullptr; 786 } 787 788 /// Given a binary operator, cast instruction, or select which has a PHI node as 789 /// operand #0, see if we can fold the instruction into the PHI (which is only 790 /// possible if all operands to the PHI are constants). 791 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { 792 PHINode *PN = cast<PHINode>(I.getOperand(0)); 793 unsigned NumPHIValues = PN->getNumIncomingValues(); 794 if (NumPHIValues == 0) 795 return nullptr; 796 797 // We normally only transform phis with a single use. However, if a PHI has 798 // multiple uses and they are all the same operation, we can fold *all* of the 799 // uses into the PHI. 800 if (!PN->hasOneUse()) { 801 // Walk the use list for the instruction, comparing them to I. 802 for (User *U : PN->users()) { 803 Instruction *UI = cast<Instruction>(U); 804 if (UI != &I && !I.isIdenticalTo(UI)) 805 return nullptr; 806 } 807 // Otherwise, we can replace *all* users with the new PHI we form. 808 } 809 810 // Check to see if all of the operands of the PHI are simple constants 811 // (constantint/constantfp/undef). If there is one non-constant value, 812 // remember the BB it is in. If there is more than one or if *it* is a PHI, 813 // bail out. We don't do arbitrary constant expressions here because moving 814 // their computation can be expensive without a cost model. 815 BasicBlock *NonConstBB = nullptr; 816 for (unsigned i = 0; i != NumPHIValues; ++i) { 817 Value *InVal = PN->getIncomingValue(i); 818 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal)) 819 continue; 820 821 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi. 822 if (NonConstBB) return nullptr; // More than one non-const value. 823 824 NonConstBB = PN->getIncomingBlock(i); 825 826 // If the InVal is an invoke at the end of the pred block, then we can't 827 // insert a computation after it without breaking the edge. 828 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 829 if (II->getParent() == NonConstBB) 830 return nullptr; 831 832 // If the incoming non-constant value is in I's block, we will remove one 833 // instruction, but insert another equivalent one, leading to infinite 834 // instcombine. 835 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI)) 836 return nullptr; 837 } 838 839 // If there is exactly one non-constant value, we can insert a copy of the 840 // operation in that block. However, if this is a critical edge, we would be 841 // inserting the computation on some other paths (e.g. inside a loop). Only 842 // do this if the pred block is unconditionally branching into the phi block. 843 if (NonConstBB != nullptr) { 844 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 845 if (!BI || !BI->isUnconditional()) return nullptr; 846 } 847 848 // Okay, we can do the transformation: create the new PHI node. 849 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues()); 850 InsertNewInstBefore(NewPN, *PN); 851 NewPN->takeName(PN); 852 853 // If we are going to have to insert a new computation, do so right before the 854 // predecessor's terminator. 855 if (NonConstBB) 856 Builder->SetInsertPoint(NonConstBB->getTerminator()); 857 858 // Next, add all of the operands to the PHI. 859 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 860 // We only currently try to fold the condition of a select when it is a phi, 861 // not the true/false values. 862 Value *TrueV = SI->getTrueValue(); 863 Value *FalseV = SI->getFalseValue(); 864 BasicBlock *PhiTransBB = PN->getParent(); 865 for (unsigned i = 0; i != NumPHIValues; ++i) { 866 BasicBlock *ThisBB = PN->getIncomingBlock(i); 867 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 868 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 869 Value *InV = nullptr; 870 // Beware of ConstantExpr: it may eventually evaluate to getNullValue, 871 // even if currently isNullValue gives false. 872 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)); 873 if (InC && !isa<ConstantExpr>(InC)) 874 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 875 else 876 InV = Builder->CreateSelect(PN->getIncomingValue(i), 877 TrueVInPred, FalseVInPred, "phitmp"); 878 NewPN->addIncoming(InV, ThisBB); 879 } 880 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 881 Constant *C = cast<Constant>(I.getOperand(1)); 882 for (unsigned i = 0; i != NumPHIValues; ++i) { 883 Value *InV = nullptr; 884 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 885 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 886 else if (isa<ICmpInst>(CI)) 887 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 888 C, "phitmp"); 889 else 890 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 891 C, "phitmp"); 892 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 893 } 894 } else if (I.getNumOperands() == 2) { 895 Constant *C = cast<Constant>(I.getOperand(1)); 896 for (unsigned i = 0; i != NumPHIValues; ++i) { 897 Value *InV = nullptr; 898 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 899 InV = ConstantExpr::get(I.getOpcode(), InC, C); 900 else 901 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 902 PN->getIncomingValue(i), C, "phitmp"); 903 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 904 } 905 } else { 906 CastInst *CI = cast<CastInst>(&I); 907 Type *RetTy = CI->getType(); 908 for (unsigned i = 0; i != NumPHIValues; ++i) { 909 Value *InV; 910 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 911 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 912 else 913 InV = Builder->CreateCast(CI->getOpcode(), 914 PN->getIncomingValue(i), I.getType(), "phitmp"); 915 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 916 } 917 } 918 919 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { 920 Instruction *User = cast<Instruction>(*UI++); 921 if (User == &I) continue; 922 ReplaceInstUsesWith(*User, NewPN); 923 EraseInstFromFunction(*User); 924 } 925 return ReplaceInstUsesWith(I, NewPN); 926 } 927 928 /// Given a pointer type and a constant offset, determine whether or not there 929 /// is a sequence of GEP indices into the pointed type that will land us at the 930 /// specified offset. If so, fill them into NewIndices and return the resultant 931 /// element type, otherwise return null. 932 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset, 933 SmallVectorImpl<Value *> &NewIndices) { 934 Type *Ty = PtrTy->getElementType(); 935 if (!Ty->isSized()) 936 return nullptr; 937 938 // Start with the index over the outer type. Note that the type size 939 // might be zero (even if the offset isn't zero) if the indexed type 940 // is something like [0 x {int, int}] 941 Type *IntPtrTy = DL.getIntPtrType(PtrTy); 942 int64_t FirstIdx = 0; 943 if (int64_t TySize = DL.getTypeAllocSize(Ty)) { 944 FirstIdx = Offset/TySize; 945 Offset -= FirstIdx*TySize; 946 947 // Handle hosts where % returns negative instead of values [0..TySize). 948 if (Offset < 0) { 949 --FirstIdx; 950 Offset += TySize; 951 assert(Offset >= 0); 952 } 953 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 954 } 955 956 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 957 958 // Index into the types. If we fail, set OrigBase to null. 959 while (Offset) { 960 // Indexing into tail padding between struct/array elements. 961 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty)) 962 return nullptr; 963 964 if (StructType *STy = dyn_cast<StructType>(Ty)) { 965 const StructLayout *SL = DL.getStructLayout(STy); 966 assert(Offset < (int64_t)SL->getSizeInBytes() && 967 "Offset must stay within the indexed type"); 968 969 unsigned Elt = SL->getElementContainingOffset(Offset); 970 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 971 Elt)); 972 973 Offset -= SL->getElementOffset(Elt); 974 Ty = STy->getElementType(Elt); 975 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 976 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType()); 977 assert(EltSize && "Cannot index into a zero-sized array"); 978 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 979 Offset %= EltSize; 980 Ty = AT->getElementType(); 981 } else { 982 // Otherwise, we can't index into the middle of this atomic type, bail. 983 return nullptr; 984 } 985 } 986 987 return Ty; 988 } 989 990 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { 991 // If this GEP has only 0 indices, it is the same pointer as 992 // Src. If Src is not a trivial GEP too, don't combine 993 // the indices. 994 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() && 995 !Src.hasOneUse()) 996 return false; 997 return true; 998 } 999 1000 /// Return a value X such that Val = X * Scale, or null if none. 1001 /// If the multiplication is known not to overflow, then NoSignedWrap is set. 1002 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { 1003 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); 1004 assert(cast<IntegerType>(Val->getType())->getBitWidth() == 1005 Scale.getBitWidth() && "Scale not compatible with value!"); 1006 1007 // If Val is zero or Scale is one then Val = Val * Scale. 1008 if (match(Val, m_Zero()) || Scale == 1) { 1009 NoSignedWrap = true; 1010 return Val; 1011 } 1012 1013 // If Scale is zero then it does not divide Val. 1014 if (Scale.isMinValue()) 1015 return nullptr; 1016 1017 // Look through chains of multiplications, searching for a constant that is 1018 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 1019 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by 1020 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore 1021 // down from Val: 1022 // 1023 // Val = M1 * X || Analysis starts here and works down 1024 // M1 = M2 * Y || Doesn't descend into terms with more 1025 // M2 = Z * 4 \/ than one use 1026 // 1027 // Then to modify a term at the bottom: 1028 // 1029 // Val = M1 * X 1030 // M1 = Z * Y || Replaced M2 with Z 1031 // 1032 // Then to work back up correcting nsw flags. 1033 1034 // Op - the term we are currently analyzing. Starts at Val then drills down. 1035 // Replaced with its descaled value before exiting from the drill down loop. 1036 Value *Op = Val; 1037 1038 // Parent - initially null, but after drilling down notes where Op came from. 1039 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the 1040 // 0'th operand of Val. 1041 std::pair<Instruction*, unsigned> Parent; 1042 1043 // Set if the transform requires a descaling at deeper levels that doesn't 1044 // overflow. 1045 bool RequireNoSignedWrap = false; 1046 1047 // Log base 2 of the scale. Negative if not a power of 2. 1048 int32_t logScale = Scale.exactLogBase2(); 1049 1050 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down 1051 1052 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 1053 // If Op is a constant divisible by Scale then descale to the quotient. 1054 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. 1055 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); 1056 if (!Remainder.isMinValue()) 1057 // Not divisible by Scale. 1058 return nullptr; 1059 // Replace with the quotient in the parent. 1060 Op = ConstantInt::get(CI->getType(), Quotient); 1061 NoSignedWrap = true; 1062 break; 1063 } 1064 1065 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { 1066 1067 if (BO->getOpcode() == Instruction::Mul) { 1068 // Multiplication. 1069 NoSignedWrap = BO->hasNoSignedWrap(); 1070 if (RequireNoSignedWrap && !NoSignedWrap) 1071 return nullptr; 1072 1073 // There are three cases for multiplication: multiplication by exactly 1074 // the scale, multiplication by a constant different to the scale, and 1075 // multiplication by something else. 1076 Value *LHS = BO->getOperand(0); 1077 Value *RHS = BO->getOperand(1); 1078 1079 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 1080 // Multiplication by a constant. 1081 if (CI->getValue() == Scale) { 1082 // Multiplication by exactly the scale, replace the multiplication 1083 // by its left-hand side in the parent. 1084 Op = LHS; 1085 break; 1086 } 1087 1088 // Otherwise drill down into the constant. 1089 if (!Op->hasOneUse()) 1090 return nullptr; 1091 1092 Parent = std::make_pair(BO, 1); 1093 continue; 1094 } 1095 1096 // Multiplication by something else. Drill down into the left-hand side 1097 // since that's where the reassociate pass puts the good stuff. 1098 if (!Op->hasOneUse()) 1099 return nullptr; 1100 1101 Parent = std::make_pair(BO, 0); 1102 continue; 1103 } 1104 1105 if (logScale > 0 && BO->getOpcode() == Instruction::Shl && 1106 isa<ConstantInt>(BO->getOperand(1))) { 1107 // Multiplication by a power of 2. 1108 NoSignedWrap = BO->hasNoSignedWrap(); 1109 if (RequireNoSignedWrap && !NoSignedWrap) 1110 return nullptr; 1111 1112 Value *LHS = BO->getOperand(0); 1113 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> 1114 getLimitedValue(Scale.getBitWidth()); 1115 // Op = LHS << Amt. 1116 1117 if (Amt == logScale) { 1118 // Multiplication by exactly the scale, replace the multiplication 1119 // by its left-hand side in the parent. 1120 Op = LHS; 1121 break; 1122 } 1123 if (Amt < logScale || !Op->hasOneUse()) 1124 return nullptr; 1125 1126 // Multiplication by more than the scale. Reduce the multiplying amount 1127 // by the scale in the parent. 1128 Parent = std::make_pair(BO, 1); 1129 Op = ConstantInt::get(BO->getType(), Amt - logScale); 1130 break; 1131 } 1132 } 1133 1134 if (!Op->hasOneUse()) 1135 return nullptr; 1136 1137 if (CastInst *Cast = dyn_cast<CastInst>(Op)) { 1138 if (Cast->getOpcode() == Instruction::SExt) { 1139 // Op is sign-extended from a smaller type, descale in the smaller type. 1140 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1141 APInt SmallScale = Scale.trunc(SmallSize); 1142 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to 1143 // descale Op as (sext Y) * Scale. In order to have 1144 // sext (Y * SmallScale) = (sext Y) * Scale 1145 // some conditions need to hold however: SmallScale must sign-extend to 1146 // Scale and the multiplication Y * SmallScale should not overflow. 1147 if (SmallScale.sext(Scale.getBitWidth()) != Scale) 1148 // SmallScale does not sign-extend to Scale. 1149 return nullptr; 1150 assert(SmallScale.exactLogBase2() == logScale); 1151 // Require that Y * SmallScale must not overflow. 1152 RequireNoSignedWrap = true; 1153 1154 // Drill down through the cast. 1155 Parent = std::make_pair(Cast, 0); 1156 Scale = SmallScale; 1157 continue; 1158 } 1159 1160 if (Cast->getOpcode() == Instruction::Trunc) { 1161 // Op is truncated from a larger type, descale in the larger type. 1162 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then 1163 // trunc (Y * sext Scale) = (trunc Y) * Scale 1164 // always holds. However (trunc Y) * Scale may overflow even if 1165 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared 1166 // from this point up in the expression (see later). 1167 if (RequireNoSignedWrap) 1168 return nullptr; 1169 1170 // Drill down through the cast. 1171 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1172 Parent = std::make_pair(Cast, 0); 1173 Scale = Scale.sext(LargeSize); 1174 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) 1175 logScale = -1; 1176 assert(Scale.exactLogBase2() == logScale); 1177 continue; 1178 } 1179 } 1180 1181 // Unsupported expression, bail out. 1182 return nullptr; 1183 } 1184 1185 // If Op is zero then Val = Op * Scale. 1186 if (match(Op, m_Zero())) { 1187 NoSignedWrap = true; 1188 return Op; 1189 } 1190 1191 // We know that we can successfully descale, so from here on we can safely 1192 // modify the IR. Op holds the descaled version of the deepest term in the 1193 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known 1194 // not to overflow. 1195 1196 if (!Parent.first) 1197 // The expression only had one term. 1198 return Op; 1199 1200 // Rewrite the parent using the descaled version of its operand. 1201 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); 1202 assert(Op != Parent.first->getOperand(Parent.second) && 1203 "Descaling was a no-op?"); 1204 Parent.first->setOperand(Parent.second, Op); 1205 Worklist.Add(Parent.first); 1206 1207 // Now work back up the expression correcting nsw flags. The logic is based 1208 // on the following observation: if X * Y is known not to overflow as a signed 1209 // multiplication, and Y is replaced by a value Z with smaller absolute value, 1210 // then X * Z will not overflow as a signed multiplication either. As we work 1211 // our way up, having NoSignedWrap 'true' means that the descaled value at the 1212 // current level has strictly smaller absolute value than the original. 1213 Instruction *Ancestor = Parent.first; 1214 do { 1215 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { 1216 // If the multiplication wasn't nsw then we can't say anything about the 1217 // value of the descaled multiplication, and we have to clear nsw flags 1218 // from this point on up. 1219 bool OpNoSignedWrap = BO->hasNoSignedWrap(); 1220 NoSignedWrap &= OpNoSignedWrap; 1221 if (NoSignedWrap != OpNoSignedWrap) { 1222 BO->setHasNoSignedWrap(NoSignedWrap); 1223 Worklist.Add(Ancestor); 1224 } 1225 } else if (Ancestor->getOpcode() == Instruction::Trunc) { 1226 // The fact that the descaled input to the trunc has smaller absolute 1227 // value than the original input doesn't tell us anything useful about 1228 // the absolute values of the truncations. 1229 NoSignedWrap = false; 1230 } 1231 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && 1232 "Failed to keep proper track of nsw flags while drilling down?"); 1233 1234 if (Ancestor == Val) 1235 // Got to the top, all done! 1236 return Val; 1237 1238 // Move up one level in the expression. 1239 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); 1240 Ancestor = Ancestor->user_back(); 1241 } while (1); 1242 } 1243 1244 /// \brief Creates node of binary operation with the same attributes as the 1245 /// specified one but with other operands. 1246 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS, 1247 InstCombiner::BuilderTy *B) { 1248 Value *BO = B->CreateBinOp(Inst.getOpcode(), LHS, RHS); 1249 // If LHS and RHS are constant, BO won't be a binary operator. 1250 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BO)) 1251 NewBO->copyIRFlags(&Inst); 1252 return BO; 1253 } 1254 1255 /// \brief Makes transformation of binary operation specific for vector types. 1256 /// \param Inst Binary operator to transform. 1257 /// \return Pointer to node that must replace the original binary operator, or 1258 /// null pointer if no transformation was made. 1259 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) { 1260 if (!Inst.getType()->isVectorTy()) return nullptr; 1261 1262 // It may not be safe to reorder shuffles and things like div, urem, etc. 1263 // because we may trap when executing those ops on unknown vector elements. 1264 // See PR20059. 1265 if (!isSafeToSpeculativelyExecute(&Inst)) 1266 return nullptr; 1267 1268 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements(); 1269 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1); 1270 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth); 1271 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth); 1272 1273 // If both arguments of binary operation are shuffles, which use the same 1274 // mask and shuffle within a single vector, it is worthwhile to move the 1275 // shuffle after binary operation: 1276 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m) 1277 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) { 1278 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS); 1279 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS); 1280 if (isa<UndefValue>(LShuf->getOperand(1)) && 1281 isa<UndefValue>(RShuf->getOperand(1)) && 1282 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() && 1283 LShuf->getMask() == RShuf->getMask()) { 1284 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0), 1285 RShuf->getOperand(0), Builder); 1286 return Builder->CreateShuffleVector(NewBO, 1287 UndefValue::get(NewBO->getType()), LShuf->getMask()); 1288 } 1289 } 1290 1291 // If one argument is a shuffle within one vector, the other is a constant, 1292 // try moving the shuffle after the binary operation. 1293 ShuffleVectorInst *Shuffle = nullptr; 1294 Constant *C1 = nullptr; 1295 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS); 1296 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS); 1297 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS); 1298 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS); 1299 if (Shuffle && C1 && 1300 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) && 1301 isa<UndefValue>(Shuffle->getOperand(1)) && 1302 Shuffle->getType() == Shuffle->getOperand(0)->getType()) { 1303 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask(); 1304 // Find constant C2 that has property: 1305 // shuffle(C2, ShMask) = C1 1306 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>) 1307 // reorder is not possible. 1308 SmallVector<Constant*, 16> C2M(VWidth, 1309 UndefValue::get(C1->getType()->getScalarType())); 1310 bool MayChange = true; 1311 for (unsigned I = 0; I < VWidth; ++I) { 1312 if (ShMask[I] >= 0) { 1313 assert(ShMask[I] < (int)VWidth); 1314 if (!isa<UndefValue>(C2M[ShMask[I]])) { 1315 MayChange = false; 1316 break; 1317 } 1318 C2M[ShMask[I]] = C1->getAggregateElement(I); 1319 } 1320 } 1321 if (MayChange) { 1322 Constant *C2 = ConstantVector::get(C2M); 1323 Value *NewLHS = isa<Constant>(LHS) ? C2 : Shuffle->getOperand(0); 1324 Value *NewRHS = isa<Constant>(LHS) ? Shuffle->getOperand(0) : C2; 1325 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder); 1326 return Builder->CreateShuffleVector(NewBO, 1327 UndefValue::get(Inst.getType()), Shuffle->getMask()); 1328 } 1329 } 1330 1331 return nullptr; 1332 } 1333 1334 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 1335 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 1336 1337 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AC)) 1338 return ReplaceInstUsesWith(GEP, V); 1339 1340 Value *PtrOp = GEP.getOperand(0); 1341 1342 // Eliminate unneeded casts for indices, and replace indices which displace 1343 // by multiples of a zero size type with zero. 1344 bool MadeChange = false; 1345 Type *IntPtrTy = 1346 DL.getIntPtrType(GEP.getPointerOperandType()->getScalarType()); 1347 1348 gep_type_iterator GTI = gep_type_begin(GEP); 1349 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E; 1350 ++I, ++GTI) { 1351 // Skip indices into struct types. 1352 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 1353 if (!SeqTy) 1354 continue; 1355 1356 // Index type should have the same width as IntPtr 1357 Type *IndexTy = (*I)->getType(); 1358 Type *NewIndexType = IndexTy->isVectorTy() ? 1359 VectorType::get(IntPtrTy, IndexTy->getVectorNumElements()) : IntPtrTy; 1360 1361 // If the element type has zero size then any index over it is equivalent 1362 // to an index of zero, so replace it with zero if it is not zero already. 1363 if (SeqTy->getElementType()->isSized() && 1364 DL.getTypeAllocSize(SeqTy->getElementType()) == 0) 1365 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 1366 *I = Constant::getNullValue(NewIndexType); 1367 MadeChange = true; 1368 } 1369 1370 if (IndexTy != NewIndexType) { 1371 // If we are using a wider index than needed for this platform, shrink 1372 // it to what we need. If narrower, sign-extend it to what we need. 1373 // This explicit cast can make subsequent optimizations more obvious. 1374 *I = Builder->CreateIntCast(*I, NewIndexType, true); 1375 MadeChange = true; 1376 } 1377 } 1378 if (MadeChange) 1379 return &GEP; 1380 1381 // Check to see if the inputs to the PHI node are getelementptr instructions. 1382 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) { 1383 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0)); 1384 if (!Op1) 1385 return nullptr; 1386 1387 // Don't fold a GEP into itself through a PHI node. This can only happen 1388 // through the back-edge of a loop. Folding a GEP into itself means that 1389 // the value of the previous iteration needs to be stored in the meantime, 1390 // thus requiring an additional register variable to be live, but not 1391 // actually achieving anything (the GEP still needs to be executed once per 1392 // loop iteration). 1393 if (Op1 == &GEP) 1394 return nullptr; 1395 1396 signed DI = -1; 1397 1398 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) { 1399 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I); 1400 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands()) 1401 return nullptr; 1402 1403 // As for Op1 above, don't try to fold a GEP into itself. 1404 if (Op2 == &GEP) 1405 return nullptr; 1406 1407 // Keep track of the type as we walk the GEP. 1408 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType(); 1409 1410 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) { 1411 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType()) 1412 return nullptr; 1413 1414 if (Op1->getOperand(J) != Op2->getOperand(J)) { 1415 if (DI == -1) { 1416 // We have not seen any differences yet in the GEPs feeding the 1417 // PHI yet, so we record this one if it is allowed to be a 1418 // variable. 1419 1420 // The first two arguments can vary for any GEP, the rest have to be 1421 // static for struct slots 1422 if (J > 1 && CurTy->isStructTy()) 1423 return nullptr; 1424 1425 DI = J; 1426 } else { 1427 // The GEP is different by more than one input. While this could be 1428 // extended to support GEPs that vary by more than one variable it 1429 // doesn't make sense since it greatly increases the complexity and 1430 // would result in an R+R+R addressing mode which no backend 1431 // directly supports and would need to be broken into several 1432 // simpler instructions anyway. 1433 return nullptr; 1434 } 1435 } 1436 1437 // Sink down a layer of the type for the next iteration. 1438 if (J > 0) { 1439 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) { 1440 CurTy = CT->getTypeAtIndex(Op1->getOperand(J)); 1441 } else { 1442 CurTy = nullptr; 1443 } 1444 } 1445 } 1446 } 1447 1448 // If not all GEPs are identical we'll have to create a new PHI node. 1449 // Check that the old PHI node has only one use so that it will get 1450 // removed. 1451 if (DI != -1 && !PN->hasOneUse()) 1452 return nullptr; 1453 1454 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone()); 1455 if (DI == -1) { 1456 // All the GEPs feeding the PHI are identical. Clone one down into our 1457 // BB so that it can be merged with the current GEP. 1458 GEP.getParent()->getInstList().insert( 1459 GEP.getParent()->getFirstInsertionPt(), NewGEP); 1460 } else { 1461 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP 1462 // into the current block so it can be merged, and create a new PHI to 1463 // set that index. 1464 PHINode *NewPN; 1465 { 1466 IRBuilderBase::InsertPointGuard Guard(*Builder); 1467 Builder->SetInsertPoint(PN); 1468 NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(), 1469 PN->getNumOperands()); 1470 } 1471 1472 for (auto &I : PN->operands()) 1473 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI), 1474 PN->getIncomingBlock(I)); 1475 1476 NewGEP->setOperand(DI, NewPN); 1477 GEP.getParent()->getInstList().insert( 1478 GEP.getParent()->getFirstInsertionPt(), NewGEP); 1479 NewGEP->setOperand(DI, NewPN); 1480 } 1481 1482 GEP.setOperand(0, NewGEP); 1483 PtrOp = NewGEP; 1484 } 1485 1486 // Combine Indices - If the source pointer to this getelementptr instruction 1487 // is a getelementptr instruction, combine the indices of the two 1488 // getelementptr instructions into a single instruction. 1489 // 1490 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 1491 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) 1492 return nullptr; 1493 1494 // Note that if our source is a gep chain itself then we wait for that 1495 // chain to be resolved before we perform this transformation. This 1496 // avoids us creating a TON of code in some cases. 1497 if (GEPOperator *SrcGEP = 1498 dyn_cast<GEPOperator>(Src->getOperand(0))) 1499 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) 1500 return nullptr; // Wait until our source is folded to completion. 1501 1502 SmallVector<Value*, 8> Indices; 1503 1504 // Find out whether the last index in the source GEP is a sequential idx. 1505 bool EndsWithSequential = false; 1506 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 1507 I != E; ++I) 1508 EndsWithSequential = !(*I)->isStructTy(); 1509 1510 // Can we combine the two pointer arithmetics offsets? 1511 if (EndsWithSequential) { 1512 // Replace: gep (gep %P, long B), long A, ... 1513 // With: T = long A+B; gep %P, T, ... 1514 // 1515 Value *Sum; 1516 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 1517 Value *GO1 = GEP.getOperand(1); 1518 if (SO1 == Constant::getNullValue(SO1->getType())) { 1519 Sum = GO1; 1520 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 1521 Sum = SO1; 1522 } else { 1523 // If they aren't the same type, then the input hasn't been processed 1524 // by the loop above yet (which canonicalizes sequential index types to 1525 // intptr_t). Just avoid transforming this until the input has been 1526 // normalized. 1527 if (SO1->getType() != GO1->getType()) 1528 return nullptr; 1529 // Only do the combine when GO1 and SO1 are both constants. Only in 1530 // this case, we are sure the cost after the merge is never more than 1531 // that before the merge. 1532 if (!isa<Constant>(GO1) || !isa<Constant>(SO1)) 1533 return nullptr; 1534 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 1535 } 1536 1537 // Update the GEP in place if possible. 1538 if (Src->getNumOperands() == 2) { 1539 GEP.setOperand(0, Src->getOperand(0)); 1540 GEP.setOperand(1, Sum); 1541 return &GEP; 1542 } 1543 Indices.append(Src->op_begin()+1, Src->op_end()-1); 1544 Indices.push_back(Sum); 1545 Indices.append(GEP.op_begin()+2, GEP.op_end()); 1546 } else if (isa<Constant>(*GEP.idx_begin()) && 1547 cast<Constant>(*GEP.idx_begin())->isNullValue() && 1548 Src->getNumOperands() != 1) { 1549 // Otherwise we can do the fold if the first index of the GEP is a zero 1550 Indices.append(Src->op_begin()+1, Src->op_end()); 1551 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 1552 } 1553 1554 if (!Indices.empty()) 1555 return GEP.isInBounds() && Src->isInBounds() 1556 ? GetElementPtrInst::CreateInBounds( 1557 Src->getSourceElementType(), Src->getOperand(0), Indices, 1558 GEP.getName()) 1559 : GetElementPtrInst::Create(Src->getSourceElementType(), 1560 Src->getOperand(0), Indices, 1561 GEP.getName()); 1562 } 1563 1564 if (GEP.getNumIndices() == 1) { 1565 unsigned AS = GEP.getPointerAddressSpace(); 1566 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() == 1567 DL.getPointerSizeInBits(AS)) { 1568 Type *PtrTy = GEP.getPointerOperandType(); 1569 Type *Ty = PtrTy->getPointerElementType(); 1570 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty); 1571 1572 bool Matched = false; 1573 uint64_t C; 1574 Value *V = nullptr; 1575 if (TyAllocSize == 1) { 1576 V = GEP.getOperand(1); 1577 Matched = true; 1578 } else if (match(GEP.getOperand(1), 1579 m_AShr(m_Value(V), m_ConstantInt(C)))) { 1580 if (TyAllocSize == 1ULL << C) 1581 Matched = true; 1582 } else if (match(GEP.getOperand(1), 1583 m_SDiv(m_Value(V), m_ConstantInt(C)))) { 1584 if (TyAllocSize == C) 1585 Matched = true; 1586 } 1587 1588 if (Matched) { 1589 // Canonicalize (gep i8* X, -(ptrtoint Y)) 1590 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y))) 1591 // The GEP pattern is emitted by the SCEV expander for certain kinds of 1592 // pointer arithmetic. 1593 if (match(V, m_Neg(m_PtrToInt(m_Value())))) { 1594 Operator *Index = cast<Operator>(V); 1595 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); 1596 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); 1597 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); 1598 } 1599 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) 1600 // to (bitcast Y) 1601 Value *Y; 1602 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)), 1603 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) { 1604 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, 1605 GEP.getType()); 1606 } 1607 } 1608 } 1609 } 1610 1611 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 1612 Value *StrippedPtr = PtrOp->stripPointerCasts(); 1613 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType()); 1614 1615 // We do not handle pointer-vector geps here. 1616 if (!StrippedPtrTy) 1617 return nullptr; 1618 1619 if (StrippedPtr != PtrOp) { 1620 bool HasZeroPointerIndex = false; 1621 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 1622 HasZeroPointerIndex = C->isZero(); 1623 1624 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 1625 // into : GEP [10 x i8]* X, i32 0, ... 1626 // 1627 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 1628 // into : GEP i8* X, ... 1629 // 1630 // This occurs when the program declares an array extern like "int X[];" 1631 if (HasZeroPointerIndex) { 1632 PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 1633 if (ArrayType *CATy = 1634 dyn_cast<ArrayType>(CPTy->getElementType())) { 1635 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 1636 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 1637 // -> GEP i8* X, ... 1638 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 1639 GetElementPtrInst *Res = GetElementPtrInst::Create( 1640 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName()); 1641 Res->setIsInBounds(GEP.isInBounds()); 1642 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) 1643 return Res; 1644 // Insert Res, and create an addrspacecast. 1645 // e.g., 1646 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ... 1647 // -> 1648 // %0 = GEP i8 addrspace(1)* X, ... 1649 // addrspacecast i8 addrspace(1)* %0 to i8* 1650 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType()); 1651 } 1652 1653 if (ArrayType *XATy = 1654 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 1655 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 1656 if (CATy->getElementType() == XATy->getElementType()) { 1657 // -> GEP [10 x i8]* X, i32 0, ... 1658 // At this point, we know that the cast source type is a pointer 1659 // to an array of the same type as the destination pointer 1660 // array. Because the array type is never stepped over (there 1661 // is a leading zero) we can fold the cast into this GEP. 1662 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) { 1663 GEP.setOperand(0, StrippedPtr); 1664 GEP.setSourceElementType(XATy); 1665 return &GEP; 1666 } 1667 // Cannot replace the base pointer directly because StrippedPtr's 1668 // address space is different. Instead, create a new GEP followed by 1669 // an addrspacecast. 1670 // e.g., 1671 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*), 1672 // i32 0, ... 1673 // -> 1674 // %0 = GEP [10 x i8] addrspace(1)* X, ... 1675 // addrspacecast i8 addrspace(1)* %0 to i8* 1676 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end()); 1677 Value *NewGEP = GEP.isInBounds() 1678 ? Builder->CreateInBoundsGEP( 1679 nullptr, StrippedPtr, Idx, GEP.getName()) 1680 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, 1681 GEP.getName()); 1682 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1683 } 1684 } 1685 } 1686 } else if (GEP.getNumOperands() == 2) { 1687 // Transform things like: 1688 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 1689 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 1690 Type *SrcElTy = StrippedPtrTy->getElementType(); 1691 Type *ResElTy = PtrOp->getType()->getPointerElementType(); 1692 if (SrcElTy->isArrayTy() && 1693 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) == 1694 DL.getTypeAllocSize(ResElTy)) { 1695 Type *IdxType = DL.getIntPtrType(GEP.getType()); 1696 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; 1697 Value *NewGEP = 1698 GEP.isInBounds() 1699 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, Idx, 1700 GEP.getName()) 1701 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName()); 1702 1703 // V and GEP are both pointer types --> BitCast 1704 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, 1705 GEP.getType()); 1706 } 1707 1708 // Transform things like: 1709 // %V = mul i64 %N, 4 1710 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V 1711 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast 1712 if (ResElTy->isSized() && SrcElTy->isSized()) { 1713 // Check that changing the type amounts to dividing the index by a scale 1714 // factor. 1715 uint64_t ResSize = DL.getTypeAllocSize(ResElTy); 1716 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy); 1717 if (ResSize && SrcSize % ResSize == 0) { 1718 Value *Idx = GEP.getOperand(1); 1719 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1720 uint64_t Scale = SrcSize / ResSize; 1721 1722 // Earlier transforms ensure that the index has type IntPtrType, which 1723 // considerably simplifies the logic by eliminating implicit casts. 1724 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) && 1725 "Index not cast to pointer width?"); 1726 1727 bool NSW; 1728 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1729 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1730 // If the multiplication NewIdx * Scale may overflow then the new 1731 // GEP may not be "inbounds". 1732 Value *NewGEP = 1733 GEP.isInBounds() && NSW 1734 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx, 1735 GEP.getName()) 1736 : Builder->CreateGEP(nullptr, StrippedPtr, NewIdx, 1737 GEP.getName()); 1738 1739 // The NewGEP must be pointer typed, so must the old one -> BitCast 1740 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, 1741 GEP.getType()); 1742 } 1743 } 1744 } 1745 1746 // Similarly, transform things like: 1747 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 1748 // (where tmp = 8*tmp2) into: 1749 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 1750 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) { 1751 // Check that changing to the array element type amounts to dividing the 1752 // index by a scale factor. 1753 uint64_t ResSize = DL.getTypeAllocSize(ResElTy); 1754 uint64_t ArrayEltSize = 1755 DL.getTypeAllocSize(SrcElTy->getArrayElementType()); 1756 if (ResSize && ArrayEltSize % ResSize == 0) { 1757 Value *Idx = GEP.getOperand(1); 1758 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1759 uint64_t Scale = ArrayEltSize / ResSize; 1760 1761 // Earlier transforms ensure that the index has type IntPtrType, which 1762 // considerably simplifies the logic by eliminating implicit casts. 1763 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) && 1764 "Index not cast to pointer width?"); 1765 1766 bool NSW; 1767 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1768 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1769 // If the multiplication NewIdx * Scale may overflow then the new 1770 // GEP may not be "inbounds". 1771 Value *Off[2] = { 1772 Constant::getNullValue(DL.getIntPtrType(GEP.getType())), 1773 NewIdx}; 1774 1775 Value *NewGEP = GEP.isInBounds() && NSW 1776 ? Builder->CreateInBoundsGEP( 1777 SrcElTy, StrippedPtr, Off, GEP.getName()) 1778 : Builder->CreateGEP(SrcElTy, StrippedPtr, Off, 1779 GEP.getName()); 1780 // The NewGEP must be pointer typed, so must the old one -> BitCast 1781 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, 1782 GEP.getType()); 1783 } 1784 } 1785 } 1786 } 1787 } 1788 1789 // addrspacecast between types is canonicalized as a bitcast, then an 1790 // addrspacecast. To take advantage of the below bitcast + struct GEP, look 1791 // through the addrspacecast. 1792 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) { 1793 // X = bitcast A addrspace(1)* to B addrspace(1)* 1794 // Y = addrspacecast A addrspace(1)* to B addrspace(2)* 1795 // Z = gep Y, <...constant indices...> 1796 // Into an addrspacecasted GEP of the struct. 1797 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0))) 1798 PtrOp = BC; 1799 } 1800 1801 /// See if we can simplify: 1802 /// X = bitcast A* to B* 1803 /// Y = gep X, <...constant indices...> 1804 /// into a gep of the original struct. This is important for SROA and alias 1805 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 1806 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 1807 Value *Operand = BCI->getOperand(0); 1808 PointerType *OpType = cast<PointerType>(Operand->getType()); 1809 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType()); 1810 APInt Offset(OffsetBits, 0); 1811 if (!isa<BitCastInst>(Operand) && 1812 GEP.accumulateConstantOffset(DL, Offset)) { 1813 1814 // If this GEP instruction doesn't move the pointer, just replace the GEP 1815 // with a bitcast of the real input to the dest type. 1816 if (!Offset) { 1817 // If the bitcast is of an allocation, and the allocation will be 1818 // converted to match the type of the cast, don't touch this. 1819 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) { 1820 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1821 if (Instruction *I = visitBitCast(*BCI)) { 1822 if (I != BCI) { 1823 I->takeName(BCI); 1824 BCI->getParent()->getInstList().insert(BCI->getIterator(), I); 1825 ReplaceInstUsesWith(*BCI, I); 1826 } 1827 return &GEP; 1828 } 1829 } 1830 1831 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace()) 1832 return new AddrSpaceCastInst(Operand, GEP.getType()); 1833 return new BitCastInst(Operand, GEP.getType()); 1834 } 1835 1836 // Otherwise, if the offset is non-zero, we need to find out if there is a 1837 // field at Offset in 'A's type. If so, we can pull the cast through the 1838 // GEP. 1839 SmallVector<Value*, 8> NewIndices; 1840 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { 1841 Value *NGEP = 1842 GEP.isInBounds() 1843 ? Builder->CreateInBoundsGEP(nullptr, Operand, NewIndices) 1844 : Builder->CreateGEP(nullptr, Operand, NewIndices); 1845 1846 if (NGEP->getType() == GEP.getType()) 1847 return ReplaceInstUsesWith(GEP, NGEP); 1848 NGEP->takeName(&GEP); 1849 1850 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace()) 1851 return new AddrSpaceCastInst(NGEP, GEP.getType()); 1852 return new BitCastInst(NGEP, GEP.getType()); 1853 } 1854 } 1855 } 1856 1857 return nullptr; 1858 } 1859 1860 static bool 1861 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, 1862 const TargetLibraryInfo *TLI) { 1863 SmallVector<Instruction*, 4> Worklist; 1864 Worklist.push_back(AI); 1865 1866 do { 1867 Instruction *PI = Worklist.pop_back_val(); 1868 for (User *U : PI->users()) { 1869 Instruction *I = cast<Instruction>(U); 1870 switch (I->getOpcode()) { 1871 default: 1872 // Give up the moment we see something we can't handle. 1873 return false; 1874 1875 case Instruction::BitCast: 1876 case Instruction::GetElementPtr: 1877 Users.emplace_back(I); 1878 Worklist.push_back(I); 1879 continue; 1880 1881 case Instruction::ICmp: { 1882 ICmpInst *ICI = cast<ICmpInst>(I); 1883 // We can fold eq/ne comparisons with null to false/true, respectively. 1884 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1))) 1885 return false; 1886 Users.emplace_back(I); 1887 continue; 1888 } 1889 1890 case Instruction::Call: 1891 // Ignore no-op and store intrinsics. 1892 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1893 switch (II->getIntrinsicID()) { 1894 default: 1895 return false; 1896 1897 case Intrinsic::memmove: 1898 case Intrinsic::memcpy: 1899 case Intrinsic::memset: { 1900 MemIntrinsic *MI = cast<MemIntrinsic>(II); 1901 if (MI->isVolatile() || MI->getRawDest() != PI) 1902 return false; 1903 } 1904 // fall through 1905 case Intrinsic::dbg_declare: 1906 case Intrinsic::dbg_value: 1907 case Intrinsic::invariant_start: 1908 case Intrinsic::invariant_end: 1909 case Intrinsic::lifetime_start: 1910 case Intrinsic::lifetime_end: 1911 case Intrinsic::objectsize: 1912 Users.emplace_back(I); 1913 continue; 1914 } 1915 } 1916 1917 if (isFreeCall(I, TLI)) { 1918 Users.emplace_back(I); 1919 continue; 1920 } 1921 return false; 1922 1923 case Instruction::Store: { 1924 StoreInst *SI = cast<StoreInst>(I); 1925 if (SI->isVolatile() || SI->getPointerOperand() != PI) 1926 return false; 1927 Users.emplace_back(I); 1928 continue; 1929 } 1930 } 1931 llvm_unreachable("missing a return?"); 1932 } 1933 } while (!Worklist.empty()); 1934 return true; 1935 } 1936 1937 Instruction *InstCombiner::visitAllocSite(Instruction &MI) { 1938 // If we have a malloc call which is only used in any amount of comparisons 1939 // to null and free calls, delete the calls and replace the comparisons with 1940 // true or false as appropriate. 1941 SmallVector<WeakVH, 64> Users; 1942 if (isAllocSiteRemovable(&MI, Users, TLI)) { 1943 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 1944 Instruction *I = cast_or_null<Instruction>(&*Users[i]); 1945 if (!I) continue; 1946 1947 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) { 1948 ReplaceInstUsesWith(*C, 1949 ConstantInt::get(Type::getInt1Ty(C->getContext()), 1950 C->isFalseWhenEqual())); 1951 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 1952 ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); 1953 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1954 if (II->getIntrinsicID() == Intrinsic::objectsize) { 1955 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1)); 1956 uint64_t DontKnow = CI->isZero() ? -1ULL : 0; 1957 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); 1958 } 1959 } 1960 EraseInstFromFunction(*I); 1961 } 1962 1963 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) { 1964 // Replace invoke with a NOP intrinsic to maintain the original CFG 1965 Module *M = II->getModule(); 1966 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); 1967 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), 1968 None, "", II->getParent()); 1969 } 1970 return EraseInstFromFunction(MI); 1971 } 1972 return nullptr; 1973 } 1974 1975 /// \brief Move the call to free before a NULL test. 1976 /// 1977 /// Check if this free is accessed after its argument has been test 1978 /// against NULL (property 0). 1979 /// If yes, it is legal to move this call in its predecessor block. 1980 /// 1981 /// The move is performed only if the block containing the call to free 1982 /// will be removed, i.e.: 1983 /// 1. it has only one predecessor P, and P has two successors 1984 /// 2. it contains the call and an unconditional branch 1985 /// 3. its successor is the same as its predecessor's successor 1986 /// 1987 /// The profitability is out-of concern here and this function should 1988 /// be called only if the caller knows this transformation would be 1989 /// profitable (e.g., for code size). 1990 static Instruction * 1991 tryToMoveFreeBeforeNullTest(CallInst &FI) { 1992 Value *Op = FI.getArgOperand(0); 1993 BasicBlock *FreeInstrBB = FI.getParent(); 1994 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); 1995 1996 // Validate part of constraint #1: Only one predecessor 1997 // FIXME: We can extend the number of predecessor, but in that case, we 1998 // would duplicate the call to free in each predecessor and it may 1999 // not be profitable even for code size. 2000 if (!PredBB) 2001 return nullptr; 2002 2003 // Validate constraint #2: Does this block contains only the call to 2004 // free and an unconditional branch? 2005 // FIXME: We could check if we can speculate everything in the 2006 // predecessor block 2007 if (FreeInstrBB->size() != 2) 2008 return nullptr; 2009 BasicBlock *SuccBB; 2010 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) 2011 return nullptr; 2012 2013 // Validate the rest of constraint #1 by matching on the pred branch. 2014 TerminatorInst *TI = PredBB->getTerminator(); 2015 BasicBlock *TrueBB, *FalseBB; 2016 ICmpInst::Predicate Pred; 2017 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) 2018 return nullptr; 2019 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) 2020 return nullptr; 2021 2022 // Validate constraint #3: Ensure the null case just falls through. 2023 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) 2024 return nullptr; 2025 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && 2026 "Broken CFG: missing edge from predecessor to successor"); 2027 2028 FI.moveBefore(TI); 2029 return &FI; 2030 } 2031 2032 2033 Instruction *InstCombiner::visitFree(CallInst &FI) { 2034 Value *Op = FI.getArgOperand(0); 2035 2036 // free undef -> unreachable. 2037 if (isa<UndefValue>(Op)) { 2038 // Insert a new store to null because we cannot modify the CFG here. 2039 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()), 2040 UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); 2041 return EraseInstFromFunction(FI); 2042 } 2043 2044 // If we have 'free null' delete the instruction. This can happen in stl code 2045 // when lots of inlining happens. 2046 if (isa<ConstantPointerNull>(Op)) 2047 return EraseInstFromFunction(FI); 2048 2049 // If we optimize for code size, try to move the call to free before the null 2050 // test so that simplify cfg can remove the empty block and dead code 2051 // elimination the branch. I.e., helps to turn something like: 2052 // if (foo) free(foo); 2053 // into 2054 // free(foo); 2055 if (MinimizeSize) 2056 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) 2057 return I; 2058 2059 return nullptr; 2060 } 2061 2062 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) { 2063 if (RI.getNumOperands() == 0) // ret void 2064 return nullptr; 2065 2066 Value *ResultOp = RI.getOperand(0); 2067 Type *VTy = ResultOp->getType(); 2068 if (!VTy->isIntegerTy()) 2069 return nullptr; 2070 2071 // There might be assume intrinsics dominating this return that completely 2072 // determine the value. If so, constant fold it. 2073 unsigned BitWidth = VTy->getPrimitiveSizeInBits(); 2074 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 2075 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI); 2076 if ((KnownZero|KnownOne).isAllOnesValue()) 2077 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne)); 2078 2079 return nullptr; 2080 } 2081 2082 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 2083 // Change br (not X), label True, label False to: br X, label False, True 2084 Value *X = nullptr; 2085 BasicBlock *TrueDest; 2086 BasicBlock *FalseDest; 2087 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 2088 !isa<Constant>(X)) { 2089 // Swap Destinations and condition... 2090 BI.setCondition(X); 2091 BI.swapSuccessors(); 2092 return &BI; 2093 } 2094 2095 // If the condition is irrelevant, remove the use so that other 2096 // transforms on the condition become more effective. 2097 if (BI.isConditional() && 2098 BI.getSuccessor(0) == BI.getSuccessor(1) && 2099 !isa<UndefValue>(BI.getCondition())) { 2100 BI.setCondition(UndefValue::get(BI.getCondition()->getType())); 2101 return &BI; 2102 } 2103 2104 // Canonicalize fcmp_one -> fcmp_oeq 2105 FCmpInst::Predicate FPred; Value *Y; 2106 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 2107 TrueDest, FalseDest)) && 2108 BI.getCondition()->hasOneUse()) 2109 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 2110 FPred == FCmpInst::FCMP_OGE) { 2111 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 2112 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 2113 2114 // Swap Destinations and condition. 2115 BI.swapSuccessors(); 2116 Worklist.Add(Cond); 2117 return &BI; 2118 } 2119 2120 // Canonicalize icmp_ne -> icmp_eq 2121 ICmpInst::Predicate IPred; 2122 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 2123 TrueDest, FalseDest)) && 2124 BI.getCondition()->hasOneUse()) 2125 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 2126 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 2127 IPred == ICmpInst::ICMP_SGE) { 2128 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 2129 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 2130 // Swap Destinations and condition. 2131 BI.swapSuccessors(); 2132 Worklist.Add(Cond); 2133 return &BI; 2134 } 2135 2136 return nullptr; 2137 } 2138 2139 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 2140 Value *Cond = SI.getCondition(); 2141 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth(); 2142 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 2143 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI); 2144 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes(); 2145 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes(); 2146 2147 // Compute the number of leading bits we can ignore. 2148 for (auto &C : SI.cases()) { 2149 LeadingKnownZeros = std::min( 2150 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros()); 2151 LeadingKnownOnes = std::min( 2152 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes()); 2153 } 2154 2155 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes); 2156 2157 // Truncate the condition operand if the new type is equal to or larger than 2158 // the largest legal integer type. We need to be conservative here since 2159 // x86 generates redundant zero-extension instructions if the operand is 2160 // truncated to i8 or i16. 2161 bool TruncCond = false; 2162 if (NewWidth > 0 && BitWidth > NewWidth && 2163 NewWidth >= DL.getLargestLegalIntTypeSize()) { 2164 TruncCond = true; 2165 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth); 2166 Builder->SetInsertPoint(&SI); 2167 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc"); 2168 SI.setCondition(NewCond); 2169 2170 for (auto &C : SI.cases()) 2171 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get( 2172 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth))); 2173 } 2174 2175 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 2176 if (I->getOpcode() == Instruction::Add) 2177 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 2178 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 2179 // Skip the first item since that's the default case. 2180 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); 2181 i != e; ++i) { 2182 ConstantInt* CaseVal = i.getCaseValue(); 2183 Constant *LHS = CaseVal; 2184 if (TruncCond) 2185 LHS = LeadingKnownZeros 2186 ? ConstantExpr::getZExt(CaseVal, Cond->getType()) 2187 : ConstantExpr::getSExt(CaseVal, Cond->getType()); 2188 Constant* NewCaseVal = ConstantExpr::getSub(LHS, AddRHS); 2189 assert(isa<ConstantInt>(NewCaseVal) && 2190 "Result of expression should be constant"); 2191 i.setValue(cast<ConstantInt>(NewCaseVal)); 2192 } 2193 SI.setCondition(I->getOperand(0)); 2194 Worklist.Add(I); 2195 return &SI; 2196 } 2197 } 2198 2199 return TruncCond ? &SI : nullptr; 2200 } 2201 2202 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 2203 Value *Agg = EV.getAggregateOperand(); 2204 2205 if (!EV.hasIndices()) 2206 return ReplaceInstUsesWith(EV, Agg); 2207 2208 if (Value *V = 2209 SimplifyExtractValueInst(Agg, EV.getIndices(), DL, TLI, DT, AC)) 2210 return ReplaceInstUsesWith(EV, V); 2211 2212 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 2213 // We're extracting from an insertvalue instruction, compare the indices 2214 const unsigned *exti, *exte, *insi, *inse; 2215 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 2216 exte = EV.idx_end(), inse = IV->idx_end(); 2217 exti != exte && insi != inse; 2218 ++exti, ++insi) { 2219 if (*insi != *exti) 2220 // The insert and extract both reference distinctly different elements. 2221 // This means the extract is not influenced by the insert, and we can 2222 // replace the aggregate operand of the extract with the aggregate 2223 // operand of the insert. i.e., replace 2224 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 2225 // %E = extractvalue { i32, { i32 } } %I, 0 2226 // with 2227 // %E = extractvalue { i32, { i32 } } %A, 0 2228 return ExtractValueInst::Create(IV->getAggregateOperand(), 2229 EV.getIndices()); 2230 } 2231 if (exti == exte && insi == inse) 2232 // Both iterators are at the end: Index lists are identical. Replace 2233 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 2234 // %C = extractvalue { i32, { i32 } } %B, 1, 0 2235 // with "i32 42" 2236 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 2237 if (exti == exte) { 2238 // The extract list is a prefix of the insert list. i.e. replace 2239 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 2240 // %E = extractvalue { i32, { i32 } } %I, 1 2241 // with 2242 // %X = extractvalue { i32, { i32 } } %A, 1 2243 // %E = insertvalue { i32 } %X, i32 42, 0 2244 // by switching the order of the insert and extract (though the 2245 // insertvalue should be left in, since it may have other uses). 2246 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 2247 EV.getIndices()); 2248 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 2249 makeArrayRef(insi, inse)); 2250 } 2251 if (insi == inse) 2252 // The insert list is a prefix of the extract list 2253 // We can simply remove the common indices from the extract and make it 2254 // operate on the inserted value instead of the insertvalue result. 2255 // i.e., replace 2256 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 2257 // %E = extractvalue { i32, { i32 } } %I, 1, 0 2258 // with 2259 // %E extractvalue { i32 } { i32 42 }, 0 2260 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 2261 makeArrayRef(exti, exte)); 2262 } 2263 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 2264 // We're extracting from an intrinsic, see if we're the only user, which 2265 // allows us to simplify multiple result intrinsics to simpler things that 2266 // just get one value. 2267 if (II->hasOneUse()) { 2268 // Check if we're grabbing the overflow bit or the result of a 'with 2269 // overflow' intrinsic. If it's the latter we can remove the intrinsic 2270 // and replace it with a traditional binary instruction. 2271 switch (II->getIntrinsicID()) { 2272 case Intrinsic::uadd_with_overflow: 2273 case Intrinsic::sadd_with_overflow: 2274 if (*EV.idx_begin() == 0) { // Normal result. 2275 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2276 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2277 EraseInstFromFunction(*II); 2278 return BinaryOperator::CreateAdd(LHS, RHS); 2279 } 2280 2281 // If the normal result of the add is dead, and the RHS is a constant, 2282 // we can transform this into a range comparison. 2283 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 2284 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 2285 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 2286 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 2287 ConstantExpr::getNot(CI)); 2288 break; 2289 case Intrinsic::usub_with_overflow: 2290 case Intrinsic::ssub_with_overflow: 2291 if (*EV.idx_begin() == 0) { // Normal result. 2292 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2293 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2294 EraseInstFromFunction(*II); 2295 return BinaryOperator::CreateSub(LHS, RHS); 2296 } 2297 break; 2298 case Intrinsic::umul_with_overflow: 2299 case Intrinsic::smul_with_overflow: 2300 if (*EV.idx_begin() == 0) { // Normal result. 2301 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 2302 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 2303 EraseInstFromFunction(*II); 2304 return BinaryOperator::CreateMul(LHS, RHS); 2305 } 2306 break; 2307 default: 2308 break; 2309 } 2310 } 2311 } 2312 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 2313 // If the (non-volatile) load only has one use, we can rewrite this to a 2314 // load from a GEP. This reduces the size of the load. If a load is used 2315 // only by extractvalue instructions then this either must have been 2316 // optimized before, or it is a struct with padding, in which case we 2317 // don't want to do the transformation as it loses padding knowledge. 2318 if (L->isSimple() && L->hasOneUse()) { 2319 // extractvalue has integer indices, getelementptr has Value*s. Convert. 2320 SmallVector<Value*, 4> Indices; 2321 // Prefix an i32 0 since we need the first element. 2322 Indices.push_back(Builder->getInt32(0)); 2323 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 2324 I != E; ++I) 2325 Indices.push_back(Builder->getInt32(*I)); 2326 2327 // We need to insert these at the location of the old load, not at that of 2328 // the extractvalue. 2329 Builder->SetInsertPoint(L); 2330 Value *GEP = Builder->CreateInBoundsGEP(L->getType(), 2331 L->getPointerOperand(), Indices); 2332 // Returning the load directly will cause the main loop to insert it in 2333 // the wrong spot, so use ReplaceInstUsesWith(). 2334 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 2335 } 2336 // We could simplify extracts from other values. Note that nested extracts may 2337 // already be simplified implicitly by the above: extract (extract (insert) ) 2338 // will be translated into extract ( insert ( extract ) ) first and then just 2339 // the value inserted, if appropriate. Similarly for extracts from single-use 2340 // loads: extract (extract (load)) will be translated to extract (load (gep)) 2341 // and if again single-use then via load (gep (gep)) to load (gep). 2342 // However, double extracts from e.g. function arguments or return values 2343 // aren't handled yet. 2344 return nullptr; 2345 } 2346 2347 /// Return 'true' if the given typeinfo will match anything. 2348 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) { 2349 switch (Personality) { 2350 case EHPersonality::GNU_C: 2351 // The GCC C EH personality only exists to support cleanups, so it's not 2352 // clear what the semantics of catch clauses are. 2353 return false; 2354 case EHPersonality::Unknown: 2355 return false; 2356 case EHPersonality::GNU_Ada: 2357 // While __gnat_all_others_value will match any Ada exception, it doesn't 2358 // match foreign exceptions (or didn't, before gcc-4.7). 2359 return false; 2360 case EHPersonality::GNU_CXX: 2361 case EHPersonality::GNU_ObjC: 2362 case EHPersonality::MSVC_X86SEH: 2363 case EHPersonality::MSVC_Win64SEH: 2364 case EHPersonality::MSVC_CXX: 2365 case EHPersonality::CoreCLR: 2366 return TypeInfo->isNullValue(); 2367 } 2368 llvm_unreachable("invalid enum"); 2369 } 2370 2371 static bool shorter_filter(const Value *LHS, const Value *RHS) { 2372 return 2373 cast<ArrayType>(LHS->getType())->getNumElements() 2374 < 2375 cast<ArrayType>(RHS->getType())->getNumElements(); 2376 } 2377 2378 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { 2379 // The logic here should be correct for any real-world personality function. 2380 // However if that turns out not to be true, the offending logic can always 2381 // be conditioned on the personality function, like the catch-all logic is. 2382 EHPersonality Personality = 2383 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn()); 2384 2385 // Simplify the list of clauses, eg by removing repeated catch clauses 2386 // (these are often created by inlining). 2387 bool MakeNewInstruction = false; // If true, recreate using the following: 2388 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction; 2389 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. 2390 2391 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already. 2392 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) { 2393 bool isLastClause = i + 1 == e; 2394 if (LI.isCatch(i)) { 2395 // A catch clause. 2396 Constant *CatchClause = LI.getClause(i); 2397 Constant *TypeInfo = CatchClause->stripPointerCasts(); 2398 2399 // If we already saw this clause, there is no point in having a second 2400 // copy of it. 2401 if (AlreadyCaught.insert(TypeInfo).second) { 2402 // This catch clause was not already seen. 2403 NewClauses.push_back(CatchClause); 2404 } else { 2405 // Repeated catch clause - drop the redundant copy. 2406 MakeNewInstruction = true; 2407 } 2408 2409 // If this is a catch-all then there is no point in keeping any following 2410 // clauses or marking the landingpad as having a cleanup. 2411 if (isCatchAll(Personality, TypeInfo)) { 2412 if (!isLastClause) 2413 MakeNewInstruction = true; 2414 CleanupFlag = false; 2415 break; 2416 } 2417 } else { 2418 // A filter clause. If any of the filter elements were already caught 2419 // then they can be dropped from the filter. It is tempting to try to 2420 // exploit the filter further by saying that any typeinfo that does not 2421 // occur in the filter can't be caught later (and thus can be dropped). 2422 // However this would be wrong, since typeinfos can match without being 2423 // equal (for example if one represents a C++ class, and the other some 2424 // class derived from it). 2425 assert(LI.isFilter(i) && "Unsupported landingpad clause!"); 2426 Constant *FilterClause = LI.getClause(i); 2427 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType()); 2428 unsigned NumTypeInfos = FilterType->getNumElements(); 2429 2430 // An empty filter catches everything, so there is no point in keeping any 2431 // following clauses or marking the landingpad as having a cleanup. By 2432 // dealing with this case here the following code is made a bit simpler. 2433 if (!NumTypeInfos) { 2434 NewClauses.push_back(FilterClause); 2435 if (!isLastClause) 2436 MakeNewInstruction = true; 2437 CleanupFlag = false; 2438 break; 2439 } 2440 2441 bool MakeNewFilter = false; // If true, make a new filter. 2442 SmallVector<Constant *, 16> NewFilterElts; // New elements. 2443 if (isa<ConstantAggregateZero>(FilterClause)) { 2444 // Not an empty filter - it contains at least one null typeinfo. 2445 assert(NumTypeInfos > 0 && "Should have handled empty filter already!"); 2446 Constant *TypeInfo = 2447 Constant::getNullValue(FilterType->getElementType()); 2448 // If this typeinfo is a catch-all then the filter can never match. 2449 if (isCatchAll(Personality, TypeInfo)) { 2450 // Throw the filter away. 2451 MakeNewInstruction = true; 2452 continue; 2453 } 2454 2455 // There is no point in having multiple copies of this typeinfo, so 2456 // discard all but the first copy if there is more than one. 2457 NewFilterElts.push_back(TypeInfo); 2458 if (NumTypeInfos > 1) 2459 MakeNewFilter = true; 2460 } else { 2461 ConstantArray *Filter = cast<ConstantArray>(FilterClause); 2462 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements. 2463 NewFilterElts.reserve(NumTypeInfos); 2464 2465 // Remove any filter elements that were already caught or that already 2466 // occurred in the filter. While there, see if any of the elements are 2467 // catch-alls. If so, the filter can be discarded. 2468 bool SawCatchAll = false; 2469 for (unsigned j = 0; j != NumTypeInfos; ++j) { 2470 Constant *Elt = Filter->getOperand(j); 2471 Constant *TypeInfo = Elt->stripPointerCasts(); 2472 if (isCatchAll(Personality, TypeInfo)) { 2473 // This element is a catch-all. Bail out, noting this fact. 2474 SawCatchAll = true; 2475 break; 2476 } 2477 2478 // Even if we've seen a type in a catch clause, we don't want to 2479 // remove it from the filter. An unexpected type handler may be 2480 // set up for a call site which throws an exception of the same 2481 // type caught. In order for the exception thrown by the unexpected 2482 // handler to propogate correctly, the filter must be correctly 2483 // described for the call site. 2484 // 2485 // Example: 2486 // 2487 // void unexpected() { throw 1;} 2488 // void foo() throw (int) { 2489 // std::set_unexpected(unexpected); 2490 // try { 2491 // throw 2.0; 2492 // } catch (int i) {} 2493 // } 2494 2495 // There is no point in having multiple copies of the same typeinfo in 2496 // a filter, so only add it if we didn't already. 2497 if (SeenInFilter.insert(TypeInfo).second) 2498 NewFilterElts.push_back(cast<Constant>(Elt)); 2499 } 2500 // A filter containing a catch-all cannot match anything by definition. 2501 if (SawCatchAll) { 2502 // Throw the filter away. 2503 MakeNewInstruction = true; 2504 continue; 2505 } 2506 2507 // If we dropped something from the filter, make a new one. 2508 if (NewFilterElts.size() < NumTypeInfos) 2509 MakeNewFilter = true; 2510 } 2511 if (MakeNewFilter) { 2512 FilterType = ArrayType::get(FilterType->getElementType(), 2513 NewFilterElts.size()); 2514 FilterClause = ConstantArray::get(FilterType, NewFilterElts); 2515 MakeNewInstruction = true; 2516 } 2517 2518 NewClauses.push_back(FilterClause); 2519 2520 // If the new filter is empty then it will catch everything so there is 2521 // no point in keeping any following clauses or marking the landingpad 2522 // as having a cleanup. The case of the original filter being empty was 2523 // already handled above. 2524 if (MakeNewFilter && !NewFilterElts.size()) { 2525 assert(MakeNewInstruction && "New filter but not a new instruction!"); 2526 CleanupFlag = false; 2527 break; 2528 } 2529 } 2530 } 2531 2532 // If several filters occur in a row then reorder them so that the shortest 2533 // filters come first (those with the smallest number of elements). This is 2534 // advantageous because shorter filters are more likely to match, speeding up 2535 // unwinding, but mostly because it increases the effectiveness of the other 2536 // filter optimizations below. 2537 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) { 2538 unsigned j; 2539 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters. 2540 for (j = i; j != e; ++j) 2541 if (!isa<ArrayType>(NewClauses[j]->getType())) 2542 break; 2543 2544 // Check whether the filters are already sorted by length. We need to know 2545 // if sorting them is actually going to do anything so that we only make a 2546 // new landingpad instruction if it does. 2547 for (unsigned k = i; k + 1 < j; ++k) 2548 if (shorter_filter(NewClauses[k+1], NewClauses[k])) { 2549 // Not sorted, so sort the filters now. Doing an unstable sort would be 2550 // correct too but reordering filters pointlessly might confuse users. 2551 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j, 2552 shorter_filter); 2553 MakeNewInstruction = true; 2554 break; 2555 } 2556 2557 // Look for the next batch of filters. 2558 i = j + 1; 2559 } 2560 2561 // If typeinfos matched if and only if equal, then the elements of a filter L 2562 // that occurs later than a filter F could be replaced by the intersection of 2563 // the elements of F and L. In reality two typeinfos can match without being 2564 // equal (for example if one represents a C++ class, and the other some class 2565 // derived from it) so it would be wrong to perform this transform in general. 2566 // However the transform is correct and useful if F is a subset of L. In that 2567 // case L can be replaced by F, and thus removed altogether since repeating a 2568 // filter is pointless. So here we look at all pairs of filters F and L where 2569 // L follows F in the list of clauses, and remove L if every element of F is 2570 // an element of L. This can occur when inlining C++ functions with exception 2571 // specifications. 2572 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) { 2573 // Examine each filter in turn. 2574 Value *Filter = NewClauses[i]; 2575 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType()); 2576 if (!FTy) 2577 // Not a filter - skip it. 2578 continue; 2579 unsigned FElts = FTy->getNumElements(); 2580 // Examine each filter following this one. Doing this backwards means that 2581 // we don't have to worry about filters disappearing under us when removed. 2582 for (unsigned j = NewClauses.size() - 1; j != i; --j) { 2583 Value *LFilter = NewClauses[j]; 2584 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType()); 2585 if (!LTy) 2586 // Not a filter - skip it. 2587 continue; 2588 // If Filter is a subset of LFilter, i.e. every element of Filter is also 2589 // an element of LFilter, then discard LFilter. 2590 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j; 2591 // If Filter is empty then it is a subset of LFilter. 2592 if (!FElts) { 2593 // Discard LFilter. 2594 NewClauses.erase(J); 2595 MakeNewInstruction = true; 2596 // Move on to the next filter. 2597 continue; 2598 } 2599 unsigned LElts = LTy->getNumElements(); 2600 // If Filter is longer than LFilter then it cannot be a subset of it. 2601 if (FElts > LElts) 2602 // Move on to the next filter. 2603 continue; 2604 // At this point we know that LFilter has at least one element. 2605 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros. 2606 // Filter is a subset of LFilter iff Filter contains only zeros (as we 2607 // already know that Filter is not longer than LFilter). 2608 if (isa<ConstantAggregateZero>(Filter)) { 2609 assert(FElts <= LElts && "Should have handled this case earlier!"); 2610 // Discard LFilter. 2611 NewClauses.erase(J); 2612 MakeNewInstruction = true; 2613 } 2614 // Move on to the next filter. 2615 continue; 2616 } 2617 ConstantArray *LArray = cast<ConstantArray>(LFilter); 2618 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros. 2619 // Since Filter is non-empty and contains only zeros, it is a subset of 2620 // LFilter iff LFilter contains a zero. 2621 assert(FElts > 0 && "Should have eliminated the empty filter earlier!"); 2622 for (unsigned l = 0; l != LElts; ++l) 2623 if (LArray->getOperand(l)->isNullValue()) { 2624 // LFilter contains a zero - discard it. 2625 NewClauses.erase(J); 2626 MakeNewInstruction = true; 2627 break; 2628 } 2629 // Move on to the next filter. 2630 continue; 2631 } 2632 // At this point we know that both filters are ConstantArrays. Loop over 2633 // operands to see whether every element of Filter is also an element of 2634 // LFilter. Since filters tend to be short this is probably faster than 2635 // using a method that scales nicely. 2636 ConstantArray *FArray = cast<ConstantArray>(Filter); 2637 bool AllFound = true; 2638 for (unsigned f = 0; f != FElts; ++f) { 2639 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts(); 2640 AllFound = false; 2641 for (unsigned l = 0; l != LElts; ++l) { 2642 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts(); 2643 if (LTypeInfo == FTypeInfo) { 2644 AllFound = true; 2645 break; 2646 } 2647 } 2648 if (!AllFound) 2649 break; 2650 } 2651 if (AllFound) { 2652 // Discard LFilter. 2653 NewClauses.erase(J); 2654 MakeNewInstruction = true; 2655 } 2656 // Move on to the next filter. 2657 } 2658 } 2659 2660 // If we changed any of the clauses, replace the old landingpad instruction 2661 // with a new one. 2662 if (MakeNewInstruction) { 2663 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(), 2664 NewClauses.size()); 2665 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i) 2666 NLI->addClause(NewClauses[i]); 2667 // A landing pad with no clauses must have the cleanup flag set. It is 2668 // theoretically possible, though highly unlikely, that we eliminated all 2669 // clauses. If so, force the cleanup flag to true. 2670 if (NewClauses.empty()) 2671 CleanupFlag = true; 2672 NLI->setCleanup(CleanupFlag); 2673 return NLI; 2674 } 2675 2676 // Even if none of the clauses changed, we may nonetheless have understood 2677 // that the cleanup flag is pointless. Clear it if so. 2678 if (LI.isCleanup() != CleanupFlag) { 2679 assert(!CleanupFlag && "Adding a cleanup, not removing one?!"); 2680 LI.setCleanup(CleanupFlag); 2681 return &LI; 2682 } 2683 2684 return nullptr; 2685 } 2686 2687 /// Try to move the specified instruction from its current block into the 2688 /// beginning of DestBlock, which can only happen if it's safe to move the 2689 /// instruction past all of the instructions between it and the end of its 2690 /// block. 2691 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 2692 assert(I->hasOneUse() && "Invariants didn't hold!"); 2693 2694 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 2695 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() || 2696 isa<TerminatorInst>(I)) 2697 return false; 2698 2699 // Do not sink alloca instructions out of the entry block. 2700 if (isa<AllocaInst>(I) && I->getParent() == 2701 &DestBlock->getParent()->getEntryBlock()) 2702 return false; 2703 2704 // Do not sink convergent call instructions. 2705 if (auto *CI = dyn_cast<CallInst>(I)) { 2706 if (CI->isConvergent()) 2707 return false; 2708 } 2709 2710 // We can only sink load instructions if there is nothing between the load and 2711 // the end of block that could change the value. 2712 if (I->mayReadFromMemory()) { 2713 for (BasicBlock::iterator Scan = I->getIterator(), 2714 E = I->getParent()->end(); 2715 Scan != E; ++Scan) 2716 if (Scan->mayWriteToMemory()) 2717 return false; 2718 } 2719 2720 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt(); 2721 I->moveBefore(&*InsertPos); 2722 ++NumSunkInst; 2723 return true; 2724 } 2725 2726 bool InstCombiner::run() { 2727 while (!Worklist.isEmpty()) { 2728 Instruction *I = Worklist.RemoveOne(); 2729 if (I == nullptr) continue; // skip null values. 2730 2731 // Check to see if we can DCE the instruction. 2732 if (isInstructionTriviallyDead(I, TLI)) { 2733 DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); 2734 EraseInstFromFunction(*I); 2735 ++NumDeadInst; 2736 MadeIRChange = true; 2737 continue; 2738 } 2739 2740 // Instruction isn't dead, see if we can constant propagate it. 2741 if (!I->use_empty() && 2742 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) { 2743 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) { 2744 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 2745 2746 // Add operands to the worklist. 2747 ReplaceInstUsesWith(*I, C); 2748 ++NumConstProp; 2749 EraseInstFromFunction(*I); 2750 MadeIRChange = true; 2751 continue; 2752 } 2753 } 2754 2755 // In general, it is possible for computeKnownBits to determine all bits in a 2756 // value even when the operands are not all constants. 2757 if (!I->use_empty() && I->getType()->isIntegerTy()) { 2758 unsigned BitWidth = I->getType()->getScalarSizeInBits(); 2759 APInt KnownZero(BitWidth, 0); 2760 APInt KnownOne(BitWidth, 0); 2761 computeKnownBits(I, KnownZero, KnownOne, /*Depth*/0, I); 2762 if ((KnownZero | KnownOne).isAllOnesValue()) { 2763 Constant *C = ConstantInt::get(I->getContext(), KnownOne); 2764 DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C << 2765 " from: " << *I << '\n'); 2766 2767 // Add operands to the worklist. 2768 ReplaceInstUsesWith(*I, C); 2769 ++NumConstProp; 2770 EraseInstFromFunction(*I); 2771 MadeIRChange = true; 2772 continue; 2773 } 2774 } 2775 2776 // See if we can trivially sink this instruction to a successor basic block. 2777 if (I->hasOneUse()) { 2778 BasicBlock *BB = I->getParent(); 2779 Instruction *UserInst = cast<Instruction>(*I->user_begin()); 2780 BasicBlock *UserParent; 2781 2782 // Get the block the use occurs in. 2783 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 2784 UserParent = PN->getIncomingBlock(*I->use_begin()); 2785 else 2786 UserParent = UserInst->getParent(); 2787 2788 if (UserParent != BB) { 2789 bool UserIsSuccessor = false; 2790 // See if the user is one of our successors. 2791 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 2792 if (*SI == UserParent) { 2793 UserIsSuccessor = true; 2794 break; 2795 } 2796 2797 // If the user is one of our immediate successors, and if that successor 2798 // only has us as a predecessors (we'd have to split the critical edge 2799 // otherwise), we can keep going. 2800 if (UserIsSuccessor && UserParent->getSinglePredecessor()) { 2801 // Okay, the CFG is simple enough, try to sink this instruction. 2802 if (TryToSinkInstruction(I, UserParent)) { 2803 MadeIRChange = true; 2804 // We'll add uses of the sunk instruction below, but since sinking 2805 // can expose opportunities for it's *operands* add them to the 2806 // worklist 2807 for (Use &U : I->operands()) 2808 if (Instruction *OpI = dyn_cast<Instruction>(U.get())) 2809 Worklist.Add(OpI); 2810 } 2811 } 2812 } 2813 } 2814 2815 // Now that we have an instruction, try combining it to simplify it. 2816 Builder->SetInsertPoint(I); 2817 Builder->SetCurrentDebugLocation(I->getDebugLoc()); 2818 2819 #ifndef NDEBUG 2820 std::string OrigI; 2821 #endif 2822 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 2823 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); 2824 2825 if (Instruction *Result = visit(*I)) { 2826 ++NumCombined; 2827 // Should we replace the old instruction with a new one? 2828 if (Result != I) { 2829 DEBUG(dbgs() << "IC: Old = " << *I << '\n' 2830 << " New = " << *Result << '\n'); 2831 2832 if (I->getDebugLoc()) 2833 Result->setDebugLoc(I->getDebugLoc()); 2834 // Everything uses the new instruction now. 2835 I->replaceAllUsesWith(Result); 2836 2837 // Move the name to the new instruction first. 2838 Result->takeName(I); 2839 2840 // Push the new instruction and any users onto the worklist. 2841 Worklist.Add(Result); 2842 Worklist.AddUsersToWorkList(*Result); 2843 2844 // Insert the new instruction into the basic block... 2845 BasicBlock *InstParent = I->getParent(); 2846 BasicBlock::iterator InsertPos = I->getIterator(); 2847 2848 // If we replace a PHI with something that isn't a PHI, fix up the 2849 // insertion point. 2850 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos)) 2851 InsertPos = InstParent->getFirstInsertionPt(); 2852 2853 InstParent->getInstList().insert(InsertPos, Result); 2854 2855 EraseInstFromFunction(*I); 2856 } else { 2857 #ifndef NDEBUG 2858 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' 2859 << " New = " << *I << '\n'); 2860 #endif 2861 2862 // If the instruction was modified, it's possible that it is now dead. 2863 // if so, remove it. 2864 if (isInstructionTriviallyDead(I, TLI)) { 2865 EraseInstFromFunction(*I); 2866 } else { 2867 Worklist.Add(I); 2868 Worklist.AddUsersToWorkList(*I); 2869 } 2870 } 2871 MadeIRChange = true; 2872 } 2873 } 2874 2875 Worklist.Zap(); 2876 return MadeIRChange; 2877 } 2878 2879 /// Walk the function in depth-first order, adding all reachable code to the 2880 /// worklist. 2881 /// 2882 /// This has a couple of tricks to make the code faster and more powerful. In 2883 /// particular, we constant fold and DCE instructions as we go, to avoid adding 2884 /// them to the worklist (this significantly speeds up instcombine on code where 2885 /// many instructions are dead or constant). Additionally, if we find a branch 2886 /// whose condition is a known constant, we only visit the reachable successors. 2887 /// 2888 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL, 2889 SmallPtrSetImpl<BasicBlock *> &Visited, 2890 InstCombineWorklist &ICWorklist, 2891 const TargetLibraryInfo *TLI) { 2892 bool MadeIRChange = false; 2893 SmallVector<BasicBlock*, 256> Worklist; 2894 Worklist.push_back(BB); 2895 2896 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 2897 DenseMap<ConstantExpr*, Constant*> FoldedConstants; 2898 2899 do { 2900 BB = Worklist.pop_back_val(); 2901 2902 // We have now visited this block! If we've already been here, ignore it. 2903 if (!Visited.insert(BB).second) 2904 continue; 2905 2906 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 2907 Instruction *Inst = &*BBI++; 2908 2909 // DCE instruction if trivially dead. 2910 if (isInstructionTriviallyDead(Inst, TLI)) { 2911 ++NumDeadInst; 2912 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); 2913 Inst->eraseFromParent(); 2914 continue; 2915 } 2916 2917 // ConstantProp instruction if trivially constant. 2918 if (!Inst->use_empty() && 2919 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0)))) 2920 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) { 2921 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " 2922 << *Inst << '\n'); 2923 Inst->replaceAllUsesWith(C); 2924 ++NumConstProp; 2925 Inst->eraseFromParent(); 2926 continue; 2927 } 2928 2929 // See if we can constant fold its operands. 2930 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e; 2931 ++i) { 2932 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 2933 if (CE == nullptr) 2934 continue; 2935 2936 Constant *&FoldRes = FoldedConstants[CE]; 2937 if (!FoldRes) 2938 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); 2939 if (!FoldRes) 2940 FoldRes = CE; 2941 2942 if (FoldRes != CE) { 2943 *i = FoldRes; 2944 MadeIRChange = true; 2945 } 2946 } 2947 2948 InstrsForInstCombineWorklist.push_back(Inst); 2949 } 2950 2951 // Recursively visit successors. If this is a branch or switch on a 2952 // constant, only visit the reachable successor. 2953 TerminatorInst *TI = BB->getTerminator(); 2954 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2955 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 2956 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 2957 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 2958 Worklist.push_back(ReachableBB); 2959 continue; 2960 } 2961 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2962 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 2963 // See if this is an explicit destination. 2964 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2965 i != e; ++i) 2966 if (i.getCaseValue() == Cond) { 2967 BasicBlock *ReachableBB = i.getCaseSuccessor(); 2968 Worklist.push_back(ReachableBB); 2969 continue; 2970 } 2971 2972 // Otherwise it is the default destination. 2973 Worklist.push_back(SI->getDefaultDest()); 2974 continue; 2975 } 2976 } 2977 2978 for (BasicBlock *SuccBB : TI->successors()) 2979 Worklist.push_back(SuccBB); 2980 } while (!Worklist.empty()); 2981 2982 // Once we've found all of the instructions to add to instcombine's worklist, 2983 // add them in reverse order. This way instcombine will visit from the top 2984 // of the function down. This jives well with the way that it adds all uses 2985 // of instructions to the worklist after doing a transformation, thus avoiding 2986 // some N^2 behavior in pathological cases. 2987 ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist); 2988 2989 return MadeIRChange; 2990 } 2991 2992 /// \brief Populate the IC worklist from a function, and prune any dead basic 2993 /// blocks discovered in the process. 2994 /// 2995 /// This also does basic constant propagation and other forward fixing to make 2996 /// the combiner itself run much faster. 2997 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL, 2998 TargetLibraryInfo *TLI, 2999 InstCombineWorklist &ICWorklist) { 3000 bool MadeIRChange = false; 3001 3002 // Do a depth-first traversal of the function, populate the worklist with 3003 // the reachable instructions. Ignore blocks that are not reachable. Keep 3004 // track of which blocks we visit. 3005 SmallPtrSet<BasicBlock *, 64> Visited; 3006 MadeIRChange |= 3007 AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI); 3008 3009 // Do a quick scan over the function. If we find any blocks that are 3010 // unreachable, remove any instructions inside of them. This prevents 3011 // the instcombine code from having to deal with some bad special cases. 3012 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 3013 if (Visited.count(&*BB)) 3014 continue; 3015 3016 // Delete the instructions backwards, as it has a reduced likelihood of 3017 // having to update as many def-use and use-def chains. 3018 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 3019 while (EndInst != BB->begin()) { 3020 // Delete the next to last instruction. 3021 Instruction *Inst = &*--EndInst->getIterator(); 3022 if (!Inst->use_empty() && !Inst->getType()->isTokenTy()) 3023 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 3024 if (Inst->isEHPad()) { 3025 EndInst = Inst; 3026 continue; 3027 } 3028 if (!isa<DbgInfoIntrinsic>(Inst)) { 3029 ++NumDeadInst; 3030 MadeIRChange = true; 3031 } 3032 if (!Inst->getType()->isTokenTy()) 3033 Inst->eraseFromParent(); 3034 } 3035 } 3036 3037 return MadeIRChange; 3038 } 3039 3040 static bool 3041 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist, 3042 AliasAnalysis *AA, AssumptionCache &AC, 3043 TargetLibraryInfo &TLI, DominatorTree &DT, 3044 LoopInfo *LI = nullptr) { 3045 auto &DL = F.getParent()->getDataLayout(); 3046 3047 /// Builder - This is an IRBuilder that automatically inserts new 3048 /// instructions into the worklist when they are created. 3049 IRBuilder<true, TargetFolder, InstCombineIRInserter> Builder( 3050 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, &AC)); 3051 3052 // Lower dbg.declare intrinsics otherwise their value may be clobbered 3053 // by instcombiner. 3054 bool DbgDeclaresChanged = LowerDbgDeclare(F); 3055 3056 // Iterate while there is work to do. 3057 int Iteration = 0; 3058 for (;;) { 3059 ++Iteration; 3060 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 3061 << F.getName() << "\n"); 3062 3063 bool Changed = false; 3064 if (prepareICWorklistFromFunction(F, DL, &TLI, Worklist)) 3065 Changed = true; 3066 3067 InstCombiner IC(Worklist, &Builder, F.optForMinSize(), 3068 AA, &AC, &TLI, &DT, DL, LI); 3069 if (IC.run()) 3070 Changed = true; 3071 3072 if (!Changed) 3073 break; 3074 } 3075 3076 return DbgDeclaresChanged || Iteration > 1; 3077 } 3078 3079 PreservedAnalyses InstCombinePass::run(Function &F, 3080 AnalysisManager<Function> *AM) { 3081 auto &AC = AM->getResult<AssumptionAnalysis>(F); 3082 auto &DT = AM->getResult<DominatorTreeAnalysis>(F); 3083 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F); 3084 3085 auto *LI = AM->getCachedResult<LoopAnalysis>(F); 3086 3087 // FIXME: The AliasAnalysis is not yet supported in the new pass manager 3088 if (!combineInstructionsOverFunction(F, Worklist, nullptr, AC, TLI, DT, LI)) 3089 // No changes, all analyses are preserved. 3090 return PreservedAnalyses::all(); 3091 3092 // Mark all the analyses that instcombine updates as preserved. 3093 // FIXME: Need a way to preserve CFG analyses here! 3094 PreservedAnalyses PA; 3095 PA.preserve<DominatorTreeAnalysis>(); 3096 return PA; 3097 } 3098 3099 namespace { 3100 /// \brief The legacy pass manager's instcombine pass. 3101 /// 3102 /// This is a basic whole-function wrapper around the instcombine utility. It 3103 /// will try to combine all instructions in the function. 3104 class InstructionCombiningPass : public FunctionPass { 3105 InstCombineWorklist Worklist; 3106 3107 public: 3108 static char ID; // Pass identification, replacement for typeid 3109 3110 InstructionCombiningPass() : FunctionPass(ID) { 3111 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry()); 3112 } 3113 3114 void getAnalysisUsage(AnalysisUsage &AU) const override; 3115 bool runOnFunction(Function &F) override; 3116 }; 3117 } 3118 3119 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const { 3120 AU.setPreservesCFG(); 3121 AU.addRequired<AAResultsWrapperPass>(); 3122 AU.addRequired<AssumptionCacheTracker>(); 3123 AU.addRequired<TargetLibraryInfoWrapperPass>(); 3124 AU.addRequired<DominatorTreeWrapperPass>(); 3125 AU.addPreserved<DominatorTreeWrapperPass>(); 3126 AU.addPreserved<GlobalsAAWrapperPass>(); 3127 } 3128 3129 bool InstructionCombiningPass::runOnFunction(Function &F) { 3130 if (skipOptnoneFunction(F)) 3131 return false; 3132 3133 // Required analyses. 3134 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 3135 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 3136 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 3137 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 3138 3139 // Optional analyses. 3140 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); 3141 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr; 3142 3143 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, LI); 3144 } 3145 3146 char InstructionCombiningPass::ID = 0; 3147 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine", 3148 "Combine redundant instructions", false, false) 3149 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 3150 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 3151 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 3152 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 3153 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) 3154 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine", 3155 "Combine redundant instructions", false, false) 3156 3157 // Initialization Routines 3158 void llvm::initializeInstCombine(PassRegistry &Registry) { 3159 initializeInstructionCombiningPassPass(Registry); 3160 } 3161 3162 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 3163 initializeInstructionCombiningPassPass(*unwrap(R)); 3164 } 3165 3166 FunctionPass *llvm::createInstructionCombiningPass() { 3167 return new InstructionCombiningPass(); 3168 } 3169