InstructionCombining.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===- InstructionCombining.cpp - Combine multiple instructions -----------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// InstructionCombining - Combine instructions to form fewer, simple 11// instructions. This pass does not modify the CFG. This pass is where 12// algebraic simplification happens. 13// 14// This pass combines things like: 15// %Y = add i32 %X, 1 16// %Z = add i32 %Y, 1 17// into: 18// %Z = add i32 %X, 2 19// 20// This is a simple worklist driven algorithm. 21// 22// This pass guarantees that the following canonicalizations are performed on 23// the program: 24// 1. If a binary operator has a constant operand, it is moved to the RHS 25// 2. Bitwise operators with constant operands are always grouped so that 26// shifts are performed first, then or's, then and's, then xor's. 27// 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible 28// 4. All cmp instructions on boolean values are replaced with logical ops 29// 5. add X, X is represented as (X*2) => (X << 1) 30// 6. Multiplies with a power-of-two constant argument are transformed into 31// shifts. 32// ... etc. 33// 34//===----------------------------------------------------------------------===// 35 36#include "llvm/Transforms/Scalar.h" 37#include "InstCombine.h" 38#include "llvm-c/Initialization.h" 39#include "llvm/ADT/SmallPtrSet.h" 40#include "llvm/ADT/Statistic.h" 41#include "llvm/ADT/StringSwitch.h" 42#include "llvm/Analysis/ConstantFolding.h" 43#include "llvm/Analysis/InstructionSimplify.h" 44#include "llvm/Analysis/MemoryBuiltins.h" 45#include "llvm/IR/CFG.h" 46#include "llvm/IR/DataLayout.h" 47#include "llvm/IR/GetElementPtrTypeIterator.h" 48#include "llvm/IR/IntrinsicInst.h" 49#include "llvm/IR/PatternMatch.h" 50#include "llvm/IR/ValueHandle.h" 51#include "llvm/Support/CommandLine.h" 52#include "llvm/Support/Debug.h" 53#include "llvm/Target/TargetLibraryInfo.h" 54#include "llvm/Transforms/Utils/Local.h" 55#include <algorithm> 56#include <climits> 57using namespace llvm; 58using namespace llvm::PatternMatch; 59 60#define DEBUG_TYPE "instcombine" 61 62STATISTIC(NumCombined , "Number of insts combined"); 63STATISTIC(NumConstProp, "Number of constant folds"); 64STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 65STATISTIC(NumSunkInst , "Number of instructions sunk"); 66STATISTIC(NumExpand, "Number of expansions"); 67STATISTIC(NumFactor , "Number of factorizations"); 68STATISTIC(NumReassoc , "Number of reassociations"); 69 70static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden, 71 cl::init(false), 72 cl::desc("Enable unsafe double to float " 73 "shrinking for math lib calls")); 74 75// Initialization Routines 76void llvm::initializeInstCombine(PassRegistry &Registry) { 77 initializeInstCombinerPass(Registry); 78} 79 80void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 81 initializeInstCombine(*unwrap(R)); 82} 83 84char InstCombiner::ID = 0; 85INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine", 86 "Combine redundant instructions", false, false) 87INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 88INITIALIZE_PASS_END(InstCombiner, "instcombine", 89 "Combine redundant instructions", false, false) 90 91void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 92 AU.setPreservesCFG(); 93 AU.addRequired<TargetLibraryInfo>(); 94} 95 96 97Value *InstCombiner::EmitGEPOffset(User *GEP) { 98 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP); 99} 100 101/// ShouldChangeType - Return true if it is desirable to convert a computation 102/// from 'From' to 'To'. We don't want to convert from a legal to an illegal 103/// type for example, or from a smaller to a larger illegal type. 104bool InstCombiner::ShouldChangeType(Type *From, Type *To) const { 105 assert(From->isIntegerTy() && To->isIntegerTy()); 106 107 // If we don't have DL, we don't know if the source/dest are legal. 108 if (!DL) return false; 109 110 unsigned FromWidth = From->getPrimitiveSizeInBits(); 111 unsigned ToWidth = To->getPrimitiveSizeInBits(); 112 bool FromLegal = DL->isLegalInteger(FromWidth); 113 bool ToLegal = DL->isLegalInteger(ToWidth); 114 115 // If this is a legal integer from type, and the result would be an illegal 116 // type, don't do the transformation. 117 if (FromLegal && !ToLegal) 118 return false; 119 120 // Otherwise, if both are illegal, do not increase the size of the result. We 121 // do allow things like i160 -> i64, but not i64 -> i160. 122 if (!FromLegal && !ToLegal && ToWidth > FromWidth) 123 return false; 124 125 return true; 126} 127 128// Return true, if No Signed Wrap should be maintained for I. 129// The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C", 130// where both B and C should be ConstantInts, results in a constant that does 131// not overflow. This function only handles the Add and Sub opcodes. For 132// all other opcodes, the function conservatively returns false. 133static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { 134 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I); 135 if (!OBO || !OBO->hasNoSignedWrap()) { 136 return false; 137 } 138 139 // We reason about Add and Sub Only. 140 Instruction::BinaryOps Opcode = I.getOpcode(); 141 if (Opcode != Instruction::Add && 142 Opcode != Instruction::Sub) { 143 return false; 144 } 145 146 ConstantInt *CB = dyn_cast<ConstantInt>(B); 147 ConstantInt *CC = dyn_cast<ConstantInt>(C); 148 149 if (!CB || !CC) { 150 return false; 151 } 152 153 const APInt &BVal = CB->getValue(); 154 const APInt &CVal = CC->getValue(); 155 bool Overflow = false; 156 157 if (Opcode == Instruction::Add) { 158 BVal.sadd_ov(CVal, Overflow); 159 } else { 160 BVal.ssub_ov(CVal, Overflow); 161 } 162 163 return !Overflow; 164} 165 166/// Conservatively clears subclassOptionalData after a reassociation or 167/// commutation. We preserve fast-math flags when applicable as they can be 168/// preserved. 169static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { 170 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I); 171 if (!FPMO) { 172 I.clearSubclassOptionalData(); 173 return; 174 } 175 176 FastMathFlags FMF = I.getFastMathFlags(); 177 I.clearSubclassOptionalData(); 178 I.setFastMathFlags(FMF); 179} 180 181/// SimplifyAssociativeOrCommutative - This performs a few simplifications for 182/// operators which are associative or commutative: 183// 184// Commutative operators: 185// 186// 1. Order operands such that they are listed from right (least complex) to 187// left (most complex). This puts constants before unary operators before 188// binary operators. 189// 190// Associative operators: 191// 192// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 193// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 194// 195// Associative and commutative operators: 196// 197// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 198// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 199// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 200// if C1 and C2 are constants. 201// 202bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { 203 Instruction::BinaryOps Opcode = I.getOpcode(); 204 bool Changed = false; 205 206 do { 207 // Order operands such that they are listed from right (least complex) to 208 // left (most complex). This puts constants before unary operators before 209 // binary operators. 210 if (I.isCommutative() && getComplexity(I.getOperand(0)) < 211 getComplexity(I.getOperand(1))) 212 Changed = !I.swapOperands(); 213 214 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); 215 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); 216 217 if (I.isAssociative()) { 218 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies. 219 if (Op0 && Op0->getOpcode() == Opcode) { 220 Value *A = Op0->getOperand(0); 221 Value *B = Op0->getOperand(1); 222 Value *C = I.getOperand(1); 223 224 // Does "B op C" simplify? 225 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) { 226 // It simplifies to V. Form "A op V". 227 I.setOperand(0, A); 228 I.setOperand(1, V); 229 // Conservatively clear the optional flags, since they may not be 230 // preserved by the reassociation. 231 if (MaintainNoSignedWrap(I, B, C) && 232 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) { 233 // Note: this is only valid because SimplifyBinOp doesn't look at 234 // the operands to Op0. 235 I.clearSubclassOptionalData(); 236 I.setHasNoSignedWrap(true); 237 } else { 238 ClearSubclassDataAfterReassociation(I); 239 } 240 241 Changed = true; 242 ++NumReassoc; 243 continue; 244 } 245 } 246 247 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies. 248 if (Op1 && Op1->getOpcode() == Opcode) { 249 Value *A = I.getOperand(0); 250 Value *B = Op1->getOperand(0); 251 Value *C = Op1->getOperand(1); 252 253 // Does "A op B" simplify? 254 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) { 255 // It simplifies to V. Form "V op C". 256 I.setOperand(0, V); 257 I.setOperand(1, C); 258 // Conservatively clear the optional flags, since they may not be 259 // preserved by the reassociation. 260 ClearSubclassDataAfterReassociation(I); 261 Changed = true; 262 ++NumReassoc; 263 continue; 264 } 265 } 266 } 267 268 if (I.isAssociative() && I.isCommutative()) { 269 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies. 270 if (Op0 && Op0->getOpcode() == Opcode) { 271 Value *A = Op0->getOperand(0); 272 Value *B = Op0->getOperand(1); 273 Value *C = I.getOperand(1); 274 275 // Does "C op A" simplify? 276 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 277 // It simplifies to V. Form "V op B". 278 I.setOperand(0, V); 279 I.setOperand(1, B); 280 // Conservatively clear the optional flags, since they may not be 281 // preserved by the reassociation. 282 ClearSubclassDataAfterReassociation(I); 283 Changed = true; 284 ++NumReassoc; 285 continue; 286 } 287 } 288 289 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies. 290 if (Op1 && Op1->getOpcode() == Opcode) { 291 Value *A = I.getOperand(0); 292 Value *B = Op1->getOperand(0); 293 Value *C = Op1->getOperand(1); 294 295 // Does "C op A" simplify? 296 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { 297 // It simplifies to V. Form "B op V". 298 I.setOperand(0, B); 299 I.setOperand(1, V); 300 // Conservatively clear the optional flags, since they may not be 301 // preserved by the reassociation. 302 ClearSubclassDataAfterReassociation(I); 303 Changed = true; 304 ++NumReassoc; 305 continue; 306 } 307 } 308 309 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)" 310 // if C1 and C2 are constants. 311 if (Op0 && Op1 && 312 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode && 313 isa<Constant>(Op0->getOperand(1)) && 314 isa<Constant>(Op1->getOperand(1)) && 315 Op0->hasOneUse() && Op1->hasOneUse()) { 316 Value *A = Op0->getOperand(0); 317 Constant *C1 = cast<Constant>(Op0->getOperand(1)); 318 Value *B = Op1->getOperand(0); 319 Constant *C2 = cast<Constant>(Op1->getOperand(1)); 320 321 Constant *Folded = ConstantExpr::get(Opcode, C1, C2); 322 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B); 323 if (isa<FPMathOperator>(New)) { 324 FastMathFlags Flags = I.getFastMathFlags(); 325 Flags &= Op0->getFastMathFlags(); 326 Flags &= Op1->getFastMathFlags(); 327 New->setFastMathFlags(Flags); 328 } 329 InsertNewInstWith(New, I); 330 New->takeName(Op1); 331 I.setOperand(0, New); 332 I.setOperand(1, Folded); 333 // Conservatively clear the optional flags, since they may not be 334 // preserved by the reassociation. 335 ClearSubclassDataAfterReassociation(I); 336 337 Changed = true; 338 continue; 339 } 340 } 341 342 // No further simplifications. 343 return Changed; 344 } while (1); 345} 346 347/// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to 348/// "(X LOp Y) ROp (X LOp Z)". 349static bool LeftDistributesOverRight(Instruction::BinaryOps LOp, 350 Instruction::BinaryOps ROp) { 351 switch (LOp) { 352 default: 353 return false; 354 355 case Instruction::And: 356 // And distributes over Or and Xor. 357 switch (ROp) { 358 default: 359 return false; 360 case Instruction::Or: 361 case Instruction::Xor: 362 return true; 363 } 364 365 case Instruction::Mul: 366 // Multiplication distributes over addition and subtraction. 367 switch (ROp) { 368 default: 369 return false; 370 case Instruction::Add: 371 case Instruction::Sub: 372 return true; 373 } 374 375 case Instruction::Or: 376 // Or distributes over And. 377 switch (ROp) { 378 default: 379 return false; 380 case Instruction::And: 381 return true; 382 } 383 } 384} 385 386/// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to 387/// "(X ROp Z) LOp (Y ROp Z)". 388static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, 389 Instruction::BinaryOps ROp) { 390 if (Instruction::isCommutative(ROp)) 391 return LeftDistributesOverRight(ROp, LOp); 392 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", 393 // but this requires knowing that the addition does not overflow and other 394 // such subtleties. 395 return false; 396} 397 398/// SimplifyUsingDistributiveLaws - This tries to simplify binary operations 399/// which some other binary operation distributes over either by factorizing 400/// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this 401/// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is 402/// a win). Returns the simplified value, or null if it didn't simplify. 403Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { 404 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 405 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 406 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 407 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op 408 409 // Factorization. 410 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) { 411 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize 412 // a common term. 413 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1); 414 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1); 415 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 416 417 // Does "X op' Y" always equal "Y op' X"? 418 bool InnerCommutative = Instruction::isCommutative(InnerOpcode); 419 420 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? 421 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) 422 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the 423 // commutative case, "(A op' B) op (C op' A)"? 424 if (A == C || (InnerCommutative && A == D)) { 425 if (A != C) 426 std::swap(C, D); 427 // Consider forming "A op' (B op D)". 428 // If "B op D" simplifies then it can be formed with no cost. 429 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL); 430 // If "B op D" doesn't simplify then only go on if both of the existing 431 // operations "A op' B" and "C op' D" will be zapped as no longer used. 432 if (!V && Op0->hasOneUse() && Op1->hasOneUse()) 433 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName()); 434 if (V) { 435 ++NumFactor; 436 V = Builder->CreateBinOp(InnerOpcode, A, V); 437 V->takeName(&I); 438 return V; 439 } 440 } 441 442 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? 443 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) 444 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the 445 // commutative case, "(A op' B) op (B op' D)"? 446 if (B == D || (InnerCommutative && B == C)) { 447 if (B != D) 448 std::swap(C, D); 449 // Consider forming "(A op C) op' B". 450 // If "A op C" simplifies then it can be formed with no cost. 451 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL); 452 // If "A op C" doesn't simplify then only go on if both of the existing 453 // operations "A op' B" and "C op' D" will be zapped as no longer used. 454 if (!V && Op0->hasOneUse() && Op1->hasOneUse()) 455 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName()); 456 if (V) { 457 ++NumFactor; 458 V = Builder->CreateBinOp(InnerOpcode, V, B); 459 V->takeName(&I); 460 return V; 461 } 462 } 463 } 464 465 // Expansion. 466 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { 467 // The instruction has the form "(A op' B) op C". See if expanding it out 468 // to "(A op C) op' (B op C)" results in simplifications. 469 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 470 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' 471 472 // Do "A op C" and "B op C" both simplify? 473 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL)) 474 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) { 475 // They do! Return "L op' R". 476 ++NumExpand; 477 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 478 if ((L == A && R == B) || 479 (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) 480 return Op0; 481 // Otherwise return "L op' R" if it simplifies. 482 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 483 return V; 484 // Otherwise, create a new instruction. 485 C = Builder->CreateBinOp(InnerOpcode, L, R); 486 C->takeName(&I); 487 return C; 488 } 489 } 490 491 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) { 492 // The instruction has the form "A op (B op' C)". See if expanding it out 493 // to "(A op B) op' (A op C)" results in simplifications. 494 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 495 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' 496 497 // Do "A op B" and "A op C" both simplify? 498 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL)) 499 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) { 500 // They do! Return "L op' R". 501 ++NumExpand; 502 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 503 if ((L == B && R == C) || 504 (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) 505 return Op1; 506 // Otherwise return "L op' R" if it simplifies. 507 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) 508 return V; 509 // Otherwise, create a new instruction. 510 A = Builder->CreateBinOp(InnerOpcode, L, R); 511 A->takeName(&I); 512 return A; 513 } 514 } 515 516 return nullptr; 517} 518 519// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction 520// if the LHS is a constant zero (which is the 'negate' form). 521// 522Value *InstCombiner::dyn_castNegVal(Value *V) const { 523 if (BinaryOperator::isNeg(V)) 524 return BinaryOperator::getNegArgument(V); 525 526 // Constants can be considered to be negated values if they can be folded. 527 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 528 return ConstantExpr::getNeg(C); 529 530 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 531 if (C->getType()->getElementType()->isIntegerTy()) 532 return ConstantExpr::getNeg(C); 533 534 return nullptr; 535} 536 537// dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the 538// instruction if the LHS is a constant negative zero (which is the 'negate' 539// form). 540// 541Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const { 542 if (BinaryOperator::isFNeg(V, IgnoreZeroSign)) 543 return BinaryOperator::getFNegArgument(V); 544 545 // Constants can be considered to be negated values if they can be folded. 546 if (ConstantFP *C = dyn_cast<ConstantFP>(V)) 547 return ConstantExpr::getFNeg(C); 548 549 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V)) 550 if (C->getType()->getElementType()->isFloatingPointTy()) 551 return ConstantExpr::getFNeg(C); 552 553 return nullptr; 554} 555 556static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 557 InstCombiner *IC) { 558 if (CastInst *CI = dyn_cast<CastInst>(&I)) { 559 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType()); 560 } 561 562 // Figure out if the constant is the left or the right argument. 563 bool ConstIsRHS = isa<Constant>(I.getOperand(1)); 564 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS)); 565 566 if (Constant *SOC = dyn_cast<Constant>(SO)) { 567 if (ConstIsRHS) 568 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand); 569 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC); 570 } 571 572 Value *Op0 = SO, *Op1 = ConstOperand; 573 if (!ConstIsRHS) 574 std::swap(Op0, Op1); 575 576 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) { 577 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, 578 SO->getName()+".op"); 579 Instruction *FPInst = dyn_cast<Instruction>(RI); 580 if (FPInst && isa<FPMathOperator>(FPInst)) 581 FPInst->copyFastMathFlags(BO); 582 return RI; 583 } 584 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I)) 585 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 586 SO->getName()+".cmp"); 587 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I)) 588 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, 589 SO->getName()+".cmp"); 590 llvm_unreachable("Unknown binary instruction type!"); 591} 592 593// FoldOpIntoSelect - Given an instruction with a select as one operand and a 594// constant as the other operand, try to fold the binary operator into the 595// select arguments. This also works for Cast instructions, which obviously do 596// not have a second operand. 597Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { 598 // Don't modify shared select instructions 599 if (!SI->hasOneUse()) return nullptr; 600 Value *TV = SI->getOperand(1); 601 Value *FV = SI->getOperand(2); 602 603 if (isa<Constant>(TV) || isa<Constant>(FV)) { 604 // Bool selects with constant operands can be folded to logical ops. 605 if (SI->getType()->isIntegerTy(1)) return nullptr; 606 607 // If it's a bitcast involving vectors, make sure it has the same number of 608 // elements on both sides. 609 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) { 610 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy()); 611 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy()); 612 613 // Verify that either both or neither are vectors. 614 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr; 615 // If vectors, verify that they have the same number of elements. 616 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) 617 return nullptr; 618 } 619 620 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); 621 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); 622 623 return SelectInst::Create(SI->getCondition(), 624 SelectTrueVal, SelectFalseVal); 625 } 626 return nullptr; 627} 628 629 630/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which 631/// has a PHI node as operand #0, see if we can fold the instruction into the 632/// PHI (which is only possible if all operands to the PHI are constants). 633/// 634Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { 635 PHINode *PN = cast<PHINode>(I.getOperand(0)); 636 unsigned NumPHIValues = PN->getNumIncomingValues(); 637 if (NumPHIValues == 0) 638 return nullptr; 639 640 // We normally only transform phis with a single use. However, if a PHI has 641 // multiple uses and they are all the same operation, we can fold *all* of the 642 // uses into the PHI. 643 if (!PN->hasOneUse()) { 644 // Walk the use list for the instruction, comparing them to I. 645 for (User *U : PN->users()) { 646 Instruction *UI = cast<Instruction>(U); 647 if (UI != &I && !I.isIdenticalTo(UI)) 648 return nullptr; 649 } 650 // Otherwise, we can replace *all* users with the new PHI we form. 651 } 652 653 // Check to see if all of the operands of the PHI are simple constants 654 // (constantint/constantfp/undef). If there is one non-constant value, 655 // remember the BB it is in. If there is more than one or if *it* is a PHI, 656 // bail out. We don't do arbitrary constant expressions here because moving 657 // their computation can be expensive without a cost model. 658 BasicBlock *NonConstBB = nullptr; 659 for (unsigned i = 0; i != NumPHIValues; ++i) { 660 Value *InVal = PN->getIncomingValue(i); 661 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal)) 662 continue; 663 664 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi. 665 if (NonConstBB) return nullptr; // More than one non-const value. 666 667 NonConstBB = PN->getIncomingBlock(i); 668 669 // If the InVal is an invoke at the end of the pred block, then we can't 670 // insert a computation after it without breaking the edge. 671 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 672 if (II->getParent() == NonConstBB) 673 return nullptr; 674 675 // If the incoming non-constant value is in I's block, we will remove one 676 // instruction, but insert another equivalent one, leading to infinite 677 // instcombine. 678 if (NonConstBB == I.getParent()) 679 return nullptr; 680 } 681 682 // If there is exactly one non-constant value, we can insert a copy of the 683 // operation in that block. However, if this is a critical edge, we would be 684 // inserting the computation one some other paths (e.g. inside a loop). Only 685 // do this if the pred block is unconditionally branching into the phi block. 686 if (NonConstBB != nullptr) { 687 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator()); 688 if (!BI || !BI->isUnconditional()) return nullptr; 689 } 690 691 // Okay, we can do the transformation: create the new PHI node. 692 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues()); 693 InsertNewInstBefore(NewPN, *PN); 694 NewPN->takeName(PN); 695 696 // If we are going to have to insert a new computation, do so right before the 697 // predecessors terminator. 698 if (NonConstBB) 699 Builder->SetInsertPoint(NonConstBB->getTerminator()); 700 701 // Next, add all of the operands to the PHI. 702 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) { 703 // We only currently try to fold the condition of a select when it is a phi, 704 // not the true/false values. 705 Value *TrueV = SI->getTrueValue(); 706 Value *FalseV = SI->getFalseValue(); 707 BasicBlock *PhiTransBB = PN->getParent(); 708 for (unsigned i = 0; i != NumPHIValues; ++i) { 709 BasicBlock *ThisBB = PN->getIncomingBlock(i); 710 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); 711 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); 712 Value *InV = nullptr; 713 // Beware of ConstantExpr: it may eventually evaluate to getNullValue, 714 // even if currently isNullValue gives false. 715 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)); 716 if (InC && !isa<ConstantExpr>(InC)) 717 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 718 else 719 InV = Builder->CreateSelect(PN->getIncomingValue(i), 720 TrueVInPred, FalseVInPred, "phitmp"); 721 NewPN->addIncoming(InV, ThisBB); 722 } 723 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 724 Constant *C = cast<Constant>(I.getOperand(1)); 725 for (unsigned i = 0; i != NumPHIValues; ++i) { 726 Value *InV = nullptr; 727 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 728 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 729 else if (isa<ICmpInst>(CI)) 730 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 731 C, "phitmp"); 732 else 733 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 734 C, "phitmp"); 735 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 736 } 737 } else if (I.getNumOperands() == 2) { 738 Constant *C = cast<Constant>(I.getOperand(1)); 739 for (unsigned i = 0; i != NumPHIValues; ++i) { 740 Value *InV = nullptr; 741 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 742 InV = ConstantExpr::get(I.getOpcode(), InC, C); 743 else 744 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 745 PN->getIncomingValue(i), C, "phitmp"); 746 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 747 } 748 } else { 749 CastInst *CI = cast<CastInst>(&I); 750 Type *RetTy = CI->getType(); 751 for (unsigned i = 0; i != NumPHIValues; ++i) { 752 Value *InV; 753 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 754 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 755 else 756 InV = Builder->CreateCast(CI->getOpcode(), 757 PN->getIncomingValue(i), I.getType(), "phitmp"); 758 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 759 } 760 } 761 762 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { 763 Instruction *User = cast<Instruction>(*UI++); 764 if (User == &I) continue; 765 ReplaceInstUsesWith(*User, NewPN); 766 EraseInstFromFunction(*User); 767 } 768 return ReplaceInstUsesWith(I, NewPN); 769} 770 771/// FindElementAtOffset - Given a pointer type and a constant offset, determine 772/// whether or not there is a sequence of GEP indices into the pointed type that 773/// will land us at the specified offset. If so, fill them into NewIndices and 774/// return the resultant element type, otherwise return null. 775Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset, 776 SmallVectorImpl<Value*> &NewIndices) { 777 assert(PtrTy->isPtrOrPtrVectorTy()); 778 779 if (!DL) 780 return nullptr; 781 782 Type *Ty = PtrTy->getPointerElementType(); 783 if (!Ty->isSized()) 784 return nullptr; 785 786 // Start with the index over the outer type. Note that the type size 787 // might be zero (even if the offset isn't zero) if the indexed type 788 // is something like [0 x {int, int}] 789 Type *IntPtrTy = DL->getIntPtrType(PtrTy); 790 int64_t FirstIdx = 0; 791 if (int64_t TySize = DL->getTypeAllocSize(Ty)) { 792 FirstIdx = Offset/TySize; 793 Offset -= FirstIdx*TySize; 794 795 // Handle hosts where % returns negative instead of values [0..TySize). 796 if (Offset < 0) { 797 --FirstIdx; 798 Offset += TySize; 799 assert(Offset >= 0); 800 } 801 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 802 } 803 804 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 805 806 // Index into the types. If we fail, set OrigBase to null. 807 while (Offset) { 808 // Indexing into tail padding between struct/array elements. 809 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty)) 810 return nullptr; 811 812 if (StructType *STy = dyn_cast<StructType>(Ty)) { 813 const StructLayout *SL = DL->getStructLayout(STy); 814 assert(Offset < (int64_t)SL->getSizeInBytes() && 815 "Offset must stay within the indexed type"); 816 817 unsigned Elt = SL->getElementContainingOffset(Offset); 818 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 819 Elt)); 820 821 Offset -= SL->getElementOffset(Elt); 822 Ty = STy->getElementType(Elt); 823 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 824 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType()); 825 assert(EltSize && "Cannot index into a zero-sized array"); 826 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 827 Offset %= EltSize; 828 Ty = AT->getElementType(); 829 } else { 830 // Otherwise, we can't index into the middle of this atomic type, bail. 831 return nullptr; 832 } 833 } 834 835 return Ty; 836} 837 838static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { 839 // If this GEP has only 0 indices, it is the same pointer as 840 // Src. If Src is not a trivial GEP too, don't combine 841 // the indices. 842 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() && 843 !Src.hasOneUse()) 844 return false; 845 return true; 846} 847 848/// Descale - Return a value X such that Val = X * Scale, or null if none. If 849/// the multiplication is known not to overflow then NoSignedWrap is set. 850Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { 851 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); 852 assert(cast<IntegerType>(Val->getType())->getBitWidth() == 853 Scale.getBitWidth() && "Scale not compatible with value!"); 854 855 // If Val is zero or Scale is one then Val = Val * Scale. 856 if (match(Val, m_Zero()) || Scale == 1) { 857 NoSignedWrap = true; 858 return Val; 859 } 860 861 // If Scale is zero then it does not divide Val. 862 if (Scale.isMinValue()) 863 return nullptr; 864 865 // Look through chains of multiplications, searching for a constant that is 866 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 867 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by 868 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore 869 // down from Val: 870 // 871 // Val = M1 * X || Analysis starts here and works down 872 // M1 = M2 * Y || Doesn't descend into terms with more 873 // M2 = Z * 4 \/ than one use 874 // 875 // Then to modify a term at the bottom: 876 // 877 // Val = M1 * X 878 // M1 = Z * Y || Replaced M2 with Z 879 // 880 // Then to work back up correcting nsw flags. 881 882 // Op - the term we are currently analyzing. Starts at Val then drills down. 883 // Replaced with its descaled value before exiting from the drill down loop. 884 Value *Op = Val; 885 886 // Parent - initially null, but after drilling down notes where Op came from. 887 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the 888 // 0'th operand of Val. 889 std::pair<Instruction*, unsigned> Parent; 890 891 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper 892 // levels that doesn't overflow. 893 bool RequireNoSignedWrap = false; 894 895 // logScale - log base 2 of the scale. Negative if not a power of 2. 896 int32_t logScale = Scale.exactLogBase2(); 897 898 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down 899 900 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 901 // If Op is a constant divisible by Scale then descale to the quotient. 902 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. 903 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); 904 if (!Remainder.isMinValue()) 905 // Not divisible by Scale. 906 return nullptr; 907 // Replace with the quotient in the parent. 908 Op = ConstantInt::get(CI->getType(), Quotient); 909 NoSignedWrap = true; 910 break; 911 } 912 913 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { 914 915 if (BO->getOpcode() == Instruction::Mul) { 916 // Multiplication. 917 NoSignedWrap = BO->hasNoSignedWrap(); 918 if (RequireNoSignedWrap && !NoSignedWrap) 919 return nullptr; 920 921 // There are three cases for multiplication: multiplication by exactly 922 // the scale, multiplication by a constant different to the scale, and 923 // multiplication by something else. 924 Value *LHS = BO->getOperand(0); 925 Value *RHS = BO->getOperand(1); 926 927 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 928 // Multiplication by a constant. 929 if (CI->getValue() == Scale) { 930 // Multiplication by exactly the scale, replace the multiplication 931 // by its left-hand side in the parent. 932 Op = LHS; 933 break; 934 } 935 936 // Otherwise drill down into the constant. 937 if (!Op->hasOneUse()) 938 return nullptr; 939 940 Parent = std::make_pair(BO, 1); 941 continue; 942 } 943 944 // Multiplication by something else. Drill down into the left-hand side 945 // since that's where the reassociate pass puts the good stuff. 946 if (!Op->hasOneUse()) 947 return nullptr; 948 949 Parent = std::make_pair(BO, 0); 950 continue; 951 } 952 953 if (logScale > 0 && BO->getOpcode() == Instruction::Shl && 954 isa<ConstantInt>(BO->getOperand(1))) { 955 // Multiplication by a power of 2. 956 NoSignedWrap = BO->hasNoSignedWrap(); 957 if (RequireNoSignedWrap && !NoSignedWrap) 958 return nullptr; 959 960 Value *LHS = BO->getOperand(0); 961 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> 962 getLimitedValue(Scale.getBitWidth()); 963 // Op = LHS << Amt. 964 965 if (Amt == logScale) { 966 // Multiplication by exactly the scale, replace the multiplication 967 // by its left-hand side in the parent. 968 Op = LHS; 969 break; 970 } 971 if (Amt < logScale || !Op->hasOneUse()) 972 return nullptr; 973 974 // Multiplication by more than the scale. Reduce the multiplying amount 975 // by the scale in the parent. 976 Parent = std::make_pair(BO, 1); 977 Op = ConstantInt::get(BO->getType(), Amt - logScale); 978 break; 979 } 980 } 981 982 if (!Op->hasOneUse()) 983 return nullptr; 984 985 if (CastInst *Cast = dyn_cast<CastInst>(Op)) { 986 if (Cast->getOpcode() == Instruction::SExt) { 987 // Op is sign-extended from a smaller type, descale in the smaller type. 988 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 989 APInt SmallScale = Scale.trunc(SmallSize); 990 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to 991 // descale Op as (sext Y) * Scale. In order to have 992 // sext (Y * SmallScale) = (sext Y) * Scale 993 // some conditions need to hold however: SmallScale must sign-extend to 994 // Scale and the multiplication Y * SmallScale should not overflow. 995 if (SmallScale.sext(Scale.getBitWidth()) != Scale) 996 // SmallScale does not sign-extend to Scale. 997 return nullptr; 998 assert(SmallScale.exactLogBase2() == logScale); 999 // Require that Y * SmallScale must not overflow. 1000 RequireNoSignedWrap = true; 1001 1002 // Drill down through the cast. 1003 Parent = std::make_pair(Cast, 0); 1004 Scale = SmallScale; 1005 continue; 1006 } 1007 1008 if (Cast->getOpcode() == Instruction::Trunc) { 1009 // Op is truncated from a larger type, descale in the larger type. 1010 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then 1011 // trunc (Y * sext Scale) = (trunc Y) * Scale 1012 // always holds. However (trunc Y) * Scale may overflow even if 1013 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared 1014 // from this point up in the expression (see later). 1015 if (RequireNoSignedWrap) 1016 return nullptr; 1017 1018 // Drill down through the cast. 1019 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1020 Parent = std::make_pair(Cast, 0); 1021 Scale = Scale.sext(LargeSize); 1022 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) 1023 logScale = -1; 1024 assert(Scale.exactLogBase2() == logScale); 1025 continue; 1026 } 1027 } 1028 1029 // Unsupported expression, bail out. 1030 return nullptr; 1031 } 1032 1033 // We know that we can successfully descale, so from here on we can safely 1034 // modify the IR. Op holds the descaled version of the deepest term in the 1035 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known 1036 // not to overflow. 1037 1038 if (!Parent.first) 1039 // The expression only had one term. 1040 return Op; 1041 1042 // Rewrite the parent using the descaled version of its operand. 1043 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); 1044 assert(Op != Parent.first->getOperand(Parent.second) && 1045 "Descaling was a no-op?"); 1046 Parent.first->setOperand(Parent.second, Op); 1047 Worklist.Add(Parent.first); 1048 1049 // Now work back up the expression correcting nsw flags. The logic is based 1050 // on the following observation: if X * Y is known not to overflow as a signed 1051 // multiplication, and Y is replaced by a value Z with smaller absolute value, 1052 // then X * Z will not overflow as a signed multiplication either. As we work 1053 // our way up, having NoSignedWrap 'true' means that the descaled value at the 1054 // current level has strictly smaller absolute value than the original. 1055 Instruction *Ancestor = Parent.first; 1056 do { 1057 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { 1058 // If the multiplication wasn't nsw then we can't say anything about the 1059 // value of the descaled multiplication, and we have to clear nsw flags 1060 // from this point on up. 1061 bool OpNoSignedWrap = BO->hasNoSignedWrap(); 1062 NoSignedWrap &= OpNoSignedWrap; 1063 if (NoSignedWrap != OpNoSignedWrap) { 1064 BO->setHasNoSignedWrap(NoSignedWrap); 1065 Worklist.Add(Ancestor); 1066 } 1067 } else if (Ancestor->getOpcode() == Instruction::Trunc) { 1068 // The fact that the descaled input to the trunc has smaller absolute 1069 // value than the original input doesn't tell us anything useful about 1070 // the absolute values of the truncations. 1071 NoSignedWrap = false; 1072 } 1073 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && 1074 "Failed to keep proper track of nsw flags while drilling down?"); 1075 1076 if (Ancestor == Val) 1077 // Got to the top, all done! 1078 return Val; 1079 1080 // Move up one level in the expression. 1081 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); 1082 Ancestor = Ancestor->user_back(); 1083 } while (1); 1084} 1085 1086/// \brief Creates node of binary operation with the same attributes as the 1087/// specified one but with other operands. 1088static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS, 1089 InstCombiner::BuilderTy *B) { 1090 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS); 1091 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) { 1092 if (isa<OverflowingBinaryOperator>(NewBO)) { 1093 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap()); 1094 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap()); 1095 } 1096 if (isa<PossiblyExactOperator>(NewBO)) 1097 NewBO->setIsExact(Inst.isExact()); 1098 } 1099 return BORes; 1100} 1101 1102/// \brief Makes transformation of binary operation specific for vector types. 1103/// \param Inst Binary operator to transform. 1104/// \return Pointer to node that must replace the original binary operator, or 1105/// null pointer if no transformation was made. 1106Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) { 1107 if (!Inst.getType()->isVectorTy()) return nullptr; 1108 1109 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements(); 1110 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1); 1111 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth); 1112 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth); 1113 1114 // If both arguments of binary operation are shuffles, which use the same 1115 // mask and shuffle within a single vector, it is worthwhile to move the 1116 // shuffle after binary operation: 1117 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m) 1118 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) { 1119 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS); 1120 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS); 1121 if (isa<UndefValue>(LShuf->getOperand(1)) && 1122 isa<UndefValue>(RShuf->getOperand(1)) && 1123 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() && 1124 LShuf->getMask() == RShuf->getMask()) { 1125 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0), 1126 RShuf->getOperand(0), Builder); 1127 Value *Res = Builder->CreateShuffleVector(NewBO, 1128 UndefValue::get(NewBO->getType()), LShuf->getMask()); 1129 return Res; 1130 } 1131 } 1132 1133 // If one argument is a shuffle within one vector, the other is a constant, 1134 // try moving the shuffle after the binary operation. 1135 ShuffleVectorInst *Shuffle = nullptr; 1136 Constant *C1 = nullptr; 1137 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS); 1138 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS); 1139 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS); 1140 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS); 1141 if (Shuffle && C1 && isa<UndefValue>(Shuffle->getOperand(1)) && 1142 Shuffle->getType() == Shuffle->getOperand(0)->getType()) { 1143 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask(); 1144 // Find constant C2 that has property: 1145 // shuffle(C2, ShMask) = C1 1146 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>) 1147 // reorder is not possible. 1148 SmallVector<Constant*, 16> C2M(VWidth, 1149 UndefValue::get(C1->getType()->getScalarType())); 1150 bool MayChange = true; 1151 for (unsigned I = 0; I < VWidth; ++I) { 1152 if (ShMask[I] >= 0) { 1153 assert(ShMask[I] < (int)VWidth); 1154 if (!isa<UndefValue>(C2M[ShMask[I]])) { 1155 MayChange = false; 1156 break; 1157 } 1158 C2M[ShMask[I]] = C1->getAggregateElement(I); 1159 } 1160 } 1161 if (MayChange) { 1162 Constant *C2 = ConstantVector::get(C2M); 1163 Value *NewLHS, *NewRHS; 1164 if (isa<Constant>(LHS)) { 1165 NewLHS = C2; 1166 NewRHS = Shuffle->getOperand(0); 1167 } else { 1168 NewLHS = Shuffle->getOperand(0); 1169 NewRHS = C2; 1170 } 1171 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder); 1172 Value *Res = Builder->CreateShuffleVector(NewBO, 1173 UndefValue::get(Inst.getType()), Shuffle->getMask()); 1174 return Res; 1175 } 1176 } 1177 1178 return nullptr; 1179} 1180 1181Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 1182 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 1183 1184 if (Value *V = SimplifyGEPInst(Ops, DL)) 1185 return ReplaceInstUsesWith(GEP, V); 1186 1187 Value *PtrOp = GEP.getOperand(0); 1188 1189 // Eliminate unneeded casts for indices, and replace indices which displace 1190 // by multiples of a zero size type with zero. 1191 if (DL) { 1192 bool MadeChange = false; 1193 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType()); 1194 1195 gep_type_iterator GTI = gep_type_begin(GEP); 1196 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 1197 I != E; ++I, ++GTI) { 1198 // Skip indices into struct types. 1199 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 1200 if (!SeqTy) continue; 1201 1202 // If the element type has zero size then any index over it is equivalent 1203 // to an index of zero, so replace it with zero if it is not zero already. 1204 if (SeqTy->getElementType()->isSized() && 1205 DL->getTypeAllocSize(SeqTy->getElementType()) == 0) 1206 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 1207 *I = Constant::getNullValue(IntPtrTy); 1208 MadeChange = true; 1209 } 1210 1211 Type *IndexTy = (*I)->getType(); 1212 if (IndexTy != IntPtrTy) { 1213 // If we are using a wider index than needed for this platform, shrink 1214 // it to what we need. If narrower, sign-extend it to what we need. 1215 // This explicit cast can make subsequent optimizations more obvious. 1216 *I = Builder->CreateIntCast(*I, IntPtrTy, true); 1217 MadeChange = true; 1218 } 1219 } 1220 if (MadeChange) return &GEP; 1221 } 1222 1223 // Combine Indices - If the source pointer to this getelementptr instruction 1224 // is a getelementptr instruction, combine the indices of the two 1225 // getelementptr instructions into a single instruction. 1226 // 1227 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 1228 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) 1229 return nullptr; 1230 1231 // Note that if our source is a gep chain itself then we wait for that 1232 // chain to be resolved before we perform this transformation. This 1233 // avoids us creating a TON of code in some cases. 1234 if (GEPOperator *SrcGEP = 1235 dyn_cast<GEPOperator>(Src->getOperand(0))) 1236 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) 1237 return nullptr; // Wait until our source is folded to completion. 1238 1239 SmallVector<Value*, 8> Indices; 1240 1241 // Find out whether the last index in the source GEP is a sequential idx. 1242 bool EndsWithSequential = false; 1243 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 1244 I != E; ++I) 1245 EndsWithSequential = !(*I)->isStructTy(); 1246 1247 // Can we combine the two pointer arithmetics offsets? 1248 if (EndsWithSequential) { 1249 // Replace: gep (gep %P, long B), long A, ... 1250 // With: T = long A+B; gep %P, T, ... 1251 // 1252 Value *Sum; 1253 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 1254 Value *GO1 = GEP.getOperand(1); 1255 if (SO1 == Constant::getNullValue(SO1->getType())) { 1256 Sum = GO1; 1257 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 1258 Sum = SO1; 1259 } else { 1260 // If they aren't the same type, then the input hasn't been processed 1261 // by the loop above yet (which canonicalizes sequential index types to 1262 // intptr_t). Just avoid transforming this until the input has been 1263 // normalized. 1264 if (SO1->getType() != GO1->getType()) 1265 return nullptr; 1266 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 1267 } 1268 1269 // Update the GEP in place if possible. 1270 if (Src->getNumOperands() == 2) { 1271 GEP.setOperand(0, Src->getOperand(0)); 1272 GEP.setOperand(1, Sum); 1273 return &GEP; 1274 } 1275 Indices.append(Src->op_begin()+1, Src->op_end()-1); 1276 Indices.push_back(Sum); 1277 Indices.append(GEP.op_begin()+2, GEP.op_end()); 1278 } else if (isa<Constant>(*GEP.idx_begin()) && 1279 cast<Constant>(*GEP.idx_begin())->isNullValue() && 1280 Src->getNumOperands() != 1) { 1281 // Otherwise we can do the fold if the first index of the GEP is a zero 1282 Indices.append(Src->op_begin()+1, Src->op_end()); 1283 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 1284 } 1285 1286 if (!Indices.empty()) 1287 return (GEP.isInBounds() && Src->isInBounds()) ? 1288 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices, 1289 GEP.getName()) : 1290 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName()); 1291 } 1292 1293 // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y)) 1294 // The GEP pattern is emitted by the SCEV expander for certain kinds of 1295 // pointer arithmetic. 1296 if (DL && GEP.getNumIndices() == 1 && 1297 match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) { 1298 unsigned AS = GEP.getPointerAddressSpace(); 1299 if (GEP.getType() == Builder->getInt8PtrTy(AS) && 1300 GEP.getOperand(1)->getType()->getScalarSizeInBits() == 1301 DL->getPointerSizeInBits(AS)) { 1302 Operator *Index = cast<Operator>(GEP.getOperand(1)); 1303 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); 1304 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); 1305 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); 1306 } 1307 } 1308 1309 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 1310 Value *StrippedPtr = PtrOp->stripPointerCasts(); 1311 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType()); 1312 1313 // We do not handle pointer-vector geps here. 1314 if (!StrippedPtrTy) 1315 return nullptr; 1316 1317 if (StrippedPtr != PtrOp) { 1318 bool HasZeroPointerIndex = false; 1319 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 1320 HasZeroPointerIndex = C->isZero(); 1321 1322 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 1323 // into : GEP [10 x i8]* X, i32 0, ... 1324 // 1325 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 1326 // into : GEP i8* X, ... 1327 // 1328 // This occurs when the program declares an array extern like "int X[];" 1329 if (HasZeroPointerIndex) { 1330 PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 1331 if (ArrayType *CATy = 1332 dyn_cast<ArrayType>(CPTy->getElementType())) { 1333 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 1334 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 1335 // -> GEP i8* X, ... 1336 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 1337 GetElementPtrInst *Res = 1338 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName()); 1339 Res->setIsInBounds(GEP.isInBounds()); 1340 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) 1341 return Res; 1342 // Insert Res, and create an addrspacecast. 1343 // e.g., 1344 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ... 1345 // -> 1346 // %0 = GEP i8 addrspace(1)* X, ... 1347 // addrspacecast i8 addrspace(1)* %0 to i8* 1348 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType()); 1349 } 1350 1351 if (ArrayType *XATy = 1352 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 1353 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 1354 if (CATy->getElementType() == XATy->getElementType()) { 1355 // -> GEP [10 x i8]* X, i32 0, ... 1356 // At this point, we know that the cast source type is a pointer 1357 // to an array of the same type as the destination pointer 1358 // array. Because the array type is never stepped over (there 1359 // is a leading zero) we can fold the cast into this GEP. 1360 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) { 1361 GEP.setOperand(0, StrippedPtr); 1362 return &GEP; 1363 } 1364 // Cannot replace the base pointer directly because StrippedPtr's 1365 // address space is different. Instead, create a new GEP followed by 1366 // an addrspacecast. 1367 // e.g., 1368 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*), 1369 // i32 0, ... 1370 // -> 1371 // %0 = GEP [10 x i8] addrspace(1)* X, ... 1372 // addrspacecast i8 addrspace(1)* %0 to i8* 1373 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end()); 1374 Value *NewGEP = GEP.isInBounds() ? 1375 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : 1376 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); 1377 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1378 } 1379 } 1380 } 1381 } else if (GEP.getNumOperands() == 2) { 1382 // Transform things like: 1383 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 1384 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 1385 Type *SrcElTy = StrippedPtrTy->getElementType(); 1386 Type *ResElTy = PtrOp->getType()->getPointerElementType(); 1387 if (DL && SrcElTy->isArrayTy() && 1388 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) == 1389 DL->getTypeAllocSize(ResElTy)) { 1390 Type *IdxType = DL->getIntPtrType(GEP.getType()); 1391 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; 1392 Value *NewGEP = GEP.isInBounds() ? 1393 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : 1394 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); 1395 1396 // V and GEP are both pointer types --> BitCast 1397 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1398 return new BitCastInst(NewGEP, GEP.getType()); 1399 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1400 } 1401 1402 // Transform things like: 1403 // %V = mul i64 %N, 4 1404 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V 1405 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast 1406 if (DL && ResElTy->isSized() && SrcElTy->isSized()) { 1407 // Check that changing the type amounts to dividing the index by a scale 1408 // factor. 1409 uint64_t ResSize = DL->getTypeAllocSize(ResElTy); 1410 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy); 1411 if (ResSize && SrcSize % ResSize == 0) { 1412 Value *Idx = GEP.getOperand(1); 1413 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1414 uint64_t Scale = SrcSize / ResSize; 1415 1416 // Earlier transforms ensure that the index has type IntPtrType, which 1417 // considerably simplifies the logic by eliminating implicit casts. 1418 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && 1419 "Index not cast to pointer width?"); 1420 1421 bool NSW; 1422 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1423 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1424 // If the multiplication NewIdx * Scale may overflow then the new 1425 // GEP may not be "inbounds". 1426 Value *NewGEP = GEP.isInBounds() && NSW ? 1427 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) : 1428 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName()); 1429 1430 // The NewGEP must be pointer typed, so must the old one -> BitCast 1431 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1432 return new BitCastInst(NewGEP, GEP.getType()); 1433 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1434 } 1435 } 1436 } 1437 1438 // Similarly, transform things like: 1439 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 1440 // (where tmp = 8*tmp2) into: 1441 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 1442 if (DL && ResElTy->isSized() && SrcElTy->isSized() && 1443 SrcElTy->isArrayTy()) { 1444 // Check that changing to the array element type amounts to dividing the 1445 // index by a scale factor. 1446 uint64_t ResSize = DL->getTypeAllocSize(ResElTy); 1447 uint64_t ArrayEltSize 1448 = DL->getTypeAllocSize(SrcElTy->getArrayElementType()); 1449 if (ResSize && ArrayEltSize % ResSize == 0) { 1450 Value *Idx = GEP.getOperand(1); 1451 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1452 uint64_t Scale = ArrayEltSize / ResSize; 1453 1454 // Earlier transforms ensure that the index has type IntPtrType, which 1455 // considerably simplifies the logic by eliminating implicit casts. 1456 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && 1457 "Index not cast to pointer width?"); 1458 1459 bool NSW; 1460 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1461 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1462 // If the multiplication NewIdx * Scale may overflow then the new 1463 // GEP may not be "inbounds". 1464 Value *Off[2] = { 1465 Constant::getNullValue(DL->getIntPtrType(GEP.getType())), 1466 NewIdx 1467 }; 1468 1469 Value *NewGEP = GEP.isInBounds() && NSW ? 1470 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) : 1471 Builder->CreateGEP(StrippedPtr, Off, GEP.getName()); 1472 // The NewGEP must be pointer typed, so must the old one -> BitCast 1473 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) 1474 return new BitCastInst(NewGEP, GEP.getType()); 1475 return new AddrSpaceCastInst(NewGEP, GEP.getType()); 1476 } 1477 } 1478 } 1479 } 1480 } 1481 1482 if (!DL) 1483 return nullptr; 1484 1485 /// See if we can simplify: 1486 /// X = bitcast A* to B* 1487 /// Y = gep X, <...constant indices...> 1488 /// into a gep of the original struct. This is important for SROA and alias 1489 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 1490 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 1491 Value *Operand = BCI->getOperand(0); 1492 PointerType *OpType = cast<PointerType>(Operand->getType()); 1493 unsigned OffsetBits = DL->getPointerTypeSizeInBits(OpType); 1494 APInt Offset(OffsetBits, 0); 1495 if (!isa<BitCastInst>(Operand) && 1496 GEP.accumulateConstantOffset(*DL, Offset) && 1497 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 1498 1499 // If this GEP instruction doesn't move the pointer, just replace the GEP 1500 // with a bitcast of the real input to the dest type. 1501 if (!Offset) { 1502 // If the bitcast is of an allocation, and the allocation will be 1503 // converted to match the type of the cast, don't touch this. 1504 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) { 1505 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1506 if (Instruction *I = visitBitCast(*BCI)) { 1507 if (I != BCI) { 1508 I->takeName(BCI); 1509 BCI->getParent()->getInstList().insert(BCI, I); 1510 ReplaceInstUsesWith(*BCI, I); 1511 } 1512 return &GEP; 1513 } 1514 } 1515 return new BitCastInst(Operand, GEP.getType()); 1516 } 1517 1518 // Otherwise, if the offset is non-zero, we need to find out if there is a 1519 // field at Offset in 'A's type. If so, we can pull the cast through the 1520 // GEP. 1521 SmallVector<Value*, 8> NewIndices; 1522 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { 1523 Value *NGEP = GEP.isInBounds() ? 1524 Builder->CreateInBoundsGEP(Operand, NewIndices) : 1525 Builder->CreateGEP(Operand, NewIndices); 1526 1527 if (NGEP->getType() == GEP.getType()) 1528 return ReplaceInstUsesWith(GEP, NGEP); 1529 NGEP->takeName(&GEP); 1530 return new BitCastInst(NGEP, GEP.getType()); 1531 } 1532 } 1533 } 1534 1535 return nullptr; 1536} 1537 1538static bool 1539isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, 1540 const TargetLibraryInfo *TLI) { 1541 SmallVector<Instruction*, 4> Worklist; 1542 Worklist.push_back(AI); 1543 1544 do { 1545 Instruction *PI = Worklist.pop_back_val(); 1546 for (User *U : PI->users()) { 1547 Instruction *I = cast<Instruction>(U); 1548 switch (I->getOpcode()) { 1549 default: 1550 // Give up the moment we see something we can't handle. 1551 return false; 1552 1553 case Instruction::BitCast: 1554 case Instruction::GetElementPtr: 1555 Users.push_back(I); 1556 Worklist.push_back(I); 1557 continue; 1558 1559 case Instruction::ICmp: { 1560 ICmpInst *ICI = cast<ICmpInst>(I); 1561 // We can fold eq/ne comparisons with null to false/true, respectively. 1562 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1))) 1563 return false; 1564 Users.push_back(I); 1565 continue; 1566 } 1567 1568 case Instruction::Call: 1569 // Ignore no-op and store intrinsics. 1570 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1571 switch (II->getIntrinsicID()) { 1572 default: 1573 return false; 1574 1575 case Intrinsic::memmove: 1576 case Intrinsic::memcpy: 1577 case Intrinsic::memset: { 1578 MemIntrinsic *MI = cast<MemIntrinsic>(II); 1579 if (MI->isVolatile() || MI->getRawDest() != PI) 1580 return false; 1581 } 1582 // fall through 1583 case Intrinsic::dbg_declare: 1584 case Intrinsic::dbg_value: 1585 case Intrinsic::invariant_start: 1586 case Intrinsic::invariant_end: 1587 case Intrinsic::lifetime_start: 1588 case Intrinsic::lifetime_end: 1589 case Intrinsic::objectsize: 1590 Users.push_back(I); 1591 continue; 1592 } 1593 } 1594 1595 if (isFreeCall(I, TLI)) { 1596 Users.push_back(I); 1597 continue; 1598 } 1599 return false; 1600 1601 case Instruction::Store: { 1602 StoreInst *SI = cast<StoreInst>(I); 1603 if (SI->isVolatile() || SI->getPointerOperand() != PI) 1604 return false; 1605 Users.push_back(I); 1606 continue; 1607 } 1608 } 1609 llvm_unreachable("missing a return?"); 1610 } 1611 } while (!Worklist.empty()); 1612 return true; 1613} 1614 1615Instruction *InstCombiner::visitAllocSite(Instruction &MI) { 1616 // If we have a malloc call which is only used in any amount of comparisons 1617 // to null and free calls, delete the calls and replace the comparisons with 1618 // true or false as appropriate. 1619 SmallVector<WeakVH, 64> Users; 1620 if (isAllocSiteRemovable(&MI, Users, TLI)) { 1621 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 1622 Instruction *I = cast_or_null<Instruction>(&*Users[i]); 1623 if (!I) continue; 1624 1625 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) { 1626 ReplaceInstUsesWith(*C, 1627 ConstantInt::get(Type::getInt1Ty(C->getContext()), 1628 C->isFalseWhenEqual())); 1629 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 1630 ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); 1631 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1632 if (II->getIntrinsicID() == Intrinsic::objectsize) { 1633 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1)); 1634 uint64_t DontKnow = CI->isZero() ? -1ULL : 0; 1635 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); 1636 } 1637 } 1638 EraseInstFromFunction(*I); 1639 } 1640 1641 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) { 1642 // Replace invoke with a NOP intrinsic to maintain the original CFG 1643 Module *M = II->getParent()->getParent()->getParent(); 1644 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); 1645 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), 1646 None, "", II->getParent()); 1647 } 1648 return EraseInstFromFunction(MI); 1649 } 1650 return nullptr; 1651} 1652 1653/// \brief Move the call to free before a NULL test. 1654/// 1655/// Check if this free is accessed after its argument has been test 1656/// against NULL (property 0). 1657/// If yes, it is legal to move this call in its predecessor block. 1658/// 1659/// The move is performed only if the block containing the call to free 1660/// will be removed, i.e.: 1661/// 1. it has only one predecessor P, and P has two successors 1662/// 2. it contains the call and an unconditional branch 1663/// 3. its successor is the same as its predecessor's successor 1664/// 1665/// The profitability is out-of concern here and this function should 1666/// be called only if the caller knows this transformation would be 1667/// profitable (e.g., for code size). 1668static Instruction * 1669tryToMoveFreeBeforeNullTest(CallInst &FI) { 1670 Value *Op = FI.getArgOperand(0); 1671 BasicBlock *FreeInstrBB = FI.getParent(); 1672 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); 1673 1674 // Validate part of constraint #1: Only one predecessor 1675 // FIXME: We can extend the number of predecessor, but in that case, we 1676 // would duplicate the call to free in each predecessor and it may 1677 // not be profitable even for code size. 1678 if (!PredBB) 1679 return nullptr; 1680 1681 // Validate constraint #2: Does this block contains only the call to 1682 // free and an unconditional branch? 1683 // FIXME: We could check if we can speculate everything in the 1684 // predecessor block 1685 if (FreeInstrBB->size() != 2) 1686 return nullptr; 1687 BasicBlock *SuccBB; 1688 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) 1689 return nullptr; 1690 1691 // Validate the rest of constraint #1 by matching on the pred branch. 1692 TerminatorInst *TI = PredBB->getTerminator(); 1693 BasicBlock *TrueBB, *FalseBB; 1694 ICmpInst::Predicate Pred; 1695 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) 1696 return nullptr; 1697 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) 1698 return nullptr; 1699 1700 // Validate constraint #3: Ensure the null case just falls through. 1701 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) 1702 return nullptr; 1703 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && 1704 "Broken CFG: missing edge from predecessor to successor"); 1705 1706 FI.moveBefore(TI); 1707 return &FI; 1708} 1709 1710 1711Instruction *InstCombiner::visitFree(CallInst &FI) { 1712 Value *Op = FI.getArgOperand(0); 1713 1714 // free undef -> unreachable. 1715 if (isa<UndefValue>(Op)) { 1716 // Insert a new store to null because we cannot modify the CFG here. 1717 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()), 1718 UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); 1719 return EraseInstFromFunction(FI); 1720 } 1721 1722 // If we have 'free null' delete the instruction. This can happen in stl code 1723 // when lots of inlining happens. 1724 if (isa<ConstantPointerNull>(Op)) 1725 return EraseInstFromFunction(FI); 1726 1727 // If we optimize for code size, try to move the call to free before the null 1728 // test so that simplify cfg can remove the empty block and dead code 1729 // elimination the branch. I.e., helps to turn something like: 1730 // if (foo) free(foo); 1731 // into 1732 // free(foo); 1733 if (MinimizeSize) 1734 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) 1735 return I; 1736 1737 return nullptr; 1738} 1739 1740 1741 1742Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 1743 // Change br (not X), label True, label False to: br X, label False, True 1744 Value *X = nullptr; 1745 BasicBlock *TrueDest; 1746 BasicBlock *FalseDest; 1747 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 1748 !isa<Constant>(X)) { 1749 // Swap Destinations and condition... 1750 BI.setCondition(X); 1751 BI.swapSuccessors(); 1752 return &BI; 1753 } 1754 1755 // Canonicalize fcmp_one -> fcmp_oeq 1756 FCmpInst::Predicate FPred; Value *Y; 1757 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 1758 TrueDest, FalseDest)) && 1759 BI.getCondition()->hasOneUse()) 1760 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 1761 FPred == FCmpInst::FCMP_OGE) { 1762 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 1763 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 1764 1765 // Swap Destinations and condition. 1766 BI.swapSuccessors(); 1767 Worklist.Add(Cond); 1768 return &BI; 1769 } 1770 1771 // Canonicalize icmp_ne -> icmp_eq 1772 ICmpInst::Predicate IPred; 1773 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 1774 TrueDest, FalseDest)) && 1775 BI.getCondition()->hasOneUse()) 1776 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 1777 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 1778 IPred == ICmpInst::ICMP_SGE) { 1779 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 1780 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 1781 // Swap Destinations and condition. 1782 BI.swapSuccessors(); 1783 Worklist.Add(Cond); 1784 return &BI; 1785 } 1786 1787 return nullptr; 1788} 1789 1790Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 1791 Value *Cond = SI.getCondition(); 1792 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 1793 if (I->getOpcode() == Instruction::Add) 1794 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 1795 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 1796 // Skip the first item since that's the default case. 1797 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); 1798 i != e; ++i) { 1799 ConstantInt* CaseVal = i.getCaseValue(); 1800 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal), 1801 AddRHS); 1802 assert(isa<ConstantInt>(NewCaseVal) && 1803 "Result of expression should be constant"); 1804 i.setValue(cast<ConstantInt>(NewCaseVal)); 1805 } 1806 SI.setCondition(I->getOperand(0)); 1807 Worklist.Add(I); 1808 return &SI; 1809 } 1810 } 1811 return nullptr; 1812} 1813 1814Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 1815 Value *Agg = EV.getAggregateOperand(); 1816 1817 if (!EV.hasIndices()) 1818 return ReplaceInstUsesWith(EV, Agg); 1819 1820 if (Constant *C = dyn_cast<Constant>(Agg)) { 1821 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) { 1822 if (EV.getNumIndices() == 0) 1823 return ReplaceInstUsesWith(EV, C2); 1824 // Extract the remaining indices out of the constant indexed by the 1825 // first index 1826 return ExtractValueInst::Create(C2, EV.getIndices().slice(1)); 1827 } 1828 return nullptr; // Can't handle other constants 1829 } 1830 1831 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 1832 // We're extracting from an insertvalue instruction, compare the indices 1833 const unsigned *exti, *exte, *insi, *inse; 1834 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 1835 exte = EV.idx_end(), inse = IV->idx_end(); 1836 exti != exte && insi != inse; 1837 ++exti, ++insi) { 1838 if (*insi != *exti) 1839 // The insert and extract both reference distinctly different elements. 1840 // This means the extract is not influenced by the insert, and we can 1841 // replace the aggregate operand of the extract with the aggregate 1842 // operand of the insert. i.e., replace 1843 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1844 // %E = extractvalue { i32, { i32 } } %I, 0 1845 // with 1846 // %E = extractvalue { i32, { i32 } } %A, 0 1847 return ExtractValueInst::Create(IV->getAggregateOperand(), 1848 EV.getIndices()); 1849 } 1850 if (exti == exte && insi == inse) 1851 // Both iterators are at the end: Index lists are identical. Replace 1852 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1853 // %C = extractvalue { i32, { i32 } } %B, 1, 0 1854 // with "i32 42" 1855 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 1856 if (exti == exte) { 1857 // The extract list is a prefix of the insert list. i.e. replace 1858 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1859 // %E = extractvalue { i32, { i32 } } %I, 1 1860 // with 1861 // %X = extractvalue { i32, { i32 } } %A, 1 1862 // %E = insertvalue { i32 } %X, i32 42, 0 1863 // by switching the order of the insert and extract (though the 1864 // insertvalue should be left in, since it may have other uses). 1865 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 1866 EV.getIndices()); 1867 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 1868 makeArrayRef(insi, inse)); 1869 } 1870 if (insi == inse) 1871 // The insert list is a prefix of the extract list 1872 // We can simply remove the common indices from the extract and make it 1873 // operate on the inserted value instead of the insertvalue result. 1874 // i.e., replace 1875 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1876 // %E = extractvalue { i32, { i32 } } %I, 1, 0 1877 // with 1878 // %E extractvalue { i32 } { i32 42 }, 0 1879 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 1880 makeArrayRef(exti, exte)); 1881 } 1882 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 1883 // We're extracting from an intrinsic, see if we're the only user, which 1884 // allows us to simplify multiple result intrinsics to simpler things that 1885 // just get one value. 1886 if (II->hasOneUse()) { 1887 // Check if we're grabbing the overflow bit or the result of a 'with 1888 // overflow' intrinsic. If it's the latter we can remove the intrinsic 1889 // and replace it with a traditional binary instruction. 1890 switch (II->getIntrinsicID()) { 1891 case Intrinsic::uadd_with_overflow: 1892 case Intrinsic::sadd_with_overflow: 1893 if (*EV.idx_begin() == 0) { // Normal result. 1894 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1895 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1896 EraseInstFromFunction(*II); 1897 return BinaryOperator::CreateAdd(LHS, RHS); 1898 } 1899 1900 // If the normal result of the add is dead, and the RHS is a constant, 1901 // we can transform this into a range comparison. 1902 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 1903 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 1904 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 1905 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 1906 ConstantExpr::getNot(CI)); 1907 break; 1908 case Intrinsic::usub_with_overflow: 1909 case Intrinsic::ssub_with_overflow: 1910 if (*EV.idx_begin() == 0) { // Normal result. 1911 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1912 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1913 EraseInstFromFunction(*II); 1914 return BinaryOperator::CreateSub(LHS, RHS); 1915 } 1916 break; 1917 case Intrinsic::umul_with_overflow: 1918 case Intrinsic::smul_with_overflow: 1919 if (*EV.idx_begin() == 0) { // Normal result. 1920 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1921 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1922 EraseInstFromFunction(*II); 1923 return BinaryOperator::CreateMul(LHS, RHS); 1924 } 1925 break; 1926 default: 1927 break; 1928 } 1929 } 1930 } 1931 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 1932 // If the (non-volatile) load only has one use, we can rewrite this to a 1933 // load from a GEP. This reduces the size of the load. 1934 // FIXME: If a load is used only by extractvalue instructions then this 1935 // could be done regardless of having multiple uses. 1936 if (L->isSimple() && L->hasOneUse()) { 1937 // extractvalue has integer indices, getelementptr has Value*s. Convert. 1938 SmallVector<Value*, 4> Indices; 1939 // Prefix an i32 0 since we need the first element. 1940 Indices.push_back(Builder->getInt32(0)); 1941 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 1942 I != E; ++I) 1943 Indices.push_back(Builder->getInt32(*I)); 1944 1945 // We need to insert these at the location of the old load, not at that of 1946 // the extractvalue. 1947 Builder->SetInsertPoint(L->getParent(), L); 1948 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices); 1949 // Returning the load directly will cause the main loop to insert it in 1950 // the wrong spot, so use ReplaceInstUsesWith(). 1951 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 1952 } 1953 // We could simplify extracts from other values. Note that nested extracts may 1954 // already be simplified implicitly by the above: extract (extract (insert) ) 1955 // will be translated into extract ( insert ( extract ) ) first and then just 1956 // the value inserted, if appropriate. Similarly for extracts from single-use 1957 // loads: extract (extract (load)) will be translated to extract (load (gep)) 1958 // and if again single-use then via load (gep (gep)) to load (gep). 1959 // However, double extracts from e.g. function arguments or return values 1960 // aren't handled yet. 1961 return nullptr; 1962} 1963 1964enum Personality_Type { 1965 Unknown_Personality, 1966 GNU_Ada_Personality, 1967 GNU_CXX_Personality, 1968 GNU_ObjC_Personality 1969}; 1970 1971/// RecognizePersonality - See if the given exception handling personality 1972/// function is one that we understand. If so, return a description of it; 1973/// otherwise return Unknown_Personality. 1974static Personality_Type RecognizePersonality(Value *Pers) { 1975 Function *F = dyn_cast<Function>(Pers->stripPointerCasts()); 1976 if (!F) 1977 return Unknown_Personality; 1978 return StringSwitch<Personality_Type>(F->getName()) 1979 .Case("__gnat_eh_personality", GNU_Ada_Personality) 1980 .Case("__gxx_personality_v0", GNU_CXX_Personality) 1981 .Case("__objc_personality_v0", GNU_ObjC_Personality) 1982 .Default(Unknown_Personality); 1983} 1984 1985/// isCatchAll - Return 'true' if the given typeinfo will match anything. 1986static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) { 1987 switch (Personality) { 1988 case Unknown_Personality: 1989 return false; 1990 case GNU_Ada_Personality: 1991 // While __gnat_all_others_value will match any Ada exception, it doesn't 1992 // match foreign exceptions (or didn't, before gcc-4.7). 1993 return false; 1994 case GNU_CXX_Personality: 1995 case GNU_ObjC_Personality: 1996 return TypeInfo->isNullValue(); 1997 } 1998 llvm_unreachable("Unknown personality!"); 1999} 2000 2001static bool shorter_filter(const Value *LHS, const Value *RHS) { 2002 return 2003 cast<ArrayType>(LHS->getType())->getNumElements() 2004 < 2005 cast<ArrayType>(RHS->getType())->getNumElements(); 2006} 2007 2008Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { 2009 // The logic here should be correct for any real-world personality function. 2010 // However if that turns out not to be true, the offending logic can always 2011 // be conditioned on the personality function, like the catch-all logic is. 2012 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn()); 2013 2014 // Simplify the list of clauses, eg by removing repeated catch clauses 2015 // (these are often created by inlining). 2016 bool MakeNewInstruction = false; // If true, recreate using the following: 2017 SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction; 2018 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. 2019 2020 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already. 2021 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) { 2022 bool isLastClause = i + 1 == e; 2023 if (LI.isCatch(i)) { 2024 // A catch clause. 2025 Value *CatchClause = LI.getClause(i); 2026 Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts()); 2027 2028 // If we already saw this clause, there is no point in having a second 2029 // copy of it. 2030 if (AlreadyCaught.insert(TypeInfo)) { 2031 // This catch clause was not already seen. 2032 NewClauses.push_back(CatchClause); 2033 } else { 2034 // Repeated catch clause - drop the redundant copy. 2035 MakeNewInstruction = true; 2036 } 2037 2038 // If this is a catch-all then there is no point in keeping any following 2039 // clauses or marking the landingpad as having a cleanup. 2040 if (isCatchAll(Personality, TypeInfo)) { 2041 if (!isLastClause) 2042 MakeNewInstruction = true; 2043 CleanupFlag = false; 2044 break; 2045 } 2046 } else { 2047 // A filter clause. If any of the filter elements were already caught 2048 // then they can be dropped from the filter. It is tempting to try to 2049 // exploit the filter further by saying that any typeinfo that does not 2050 // occur in the filter can't be caught later (and thus can be dropped). 2051 // However this would be wrong, since typeinfos can match without being 2052 // equal (for example if one represents a C++ class, and the other some 2053 // class derived from it). 2054 assert(LI.isFilter(i) && "Unsupported landingpad clause!"); 2055 Value *FilterClause = LI.getClause(i); 2056 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType()); 2057 unsigned NumTypeInfos = FilterType->getNumElements(); 2058 2059 // An empty filter catches everything, so there is no point in keeping any 2060 // following clauses or marking the landingpad as having a cleanup. By 2061 // dealing with this case here the following code is made a bit simpler. 2062 if (!NumTypeInfos) { 2063 NewClauses.push_back(FilterClause); 2064 if (!isLastClause) 2065 MakeNewInstruction = true; 2066 CleanupFlag = false; 2067 break; 2068 } 2069 2070 bool MakeNewFilter = false; // If true, make a new filter. 2071 SmallVector<Constant *, 16> NewFilterElts; // New elements. 2072 if (isa<ConstantAggregateZero>(FilterClause)) { 2073 // Not an empty filter - it contains at least one null typeinfo. 2074 assert(NumTypeInfos > 0 && "Should have handled empty filter already!"); 2075 Constant *TypeInfo = 2076 Constant::getNullValue(FilterType->getElementType()); 2077 // If this typeinfo is a catch-all then the filter can never match. 2078 if (isCatchAll(Personality, TypeInfo)) { 2079 // Throw the filter away. 2080 MakeNewInstruction = true; 2081 continue; 2082 } 2083 2084 // There is no point in having multiple copies of this typeinfo, so 2085 // discard all but the first copy if there is more than one. 2086 NewFilterElts.push_back(TypeInfo); 2087 if (NumTypeInfos > 1) 2088 MakeNewFilter = true; 2089 } else { 2090 ConstantArray *Filter = cast<ConstantArray>(FilterClause); 2091 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements. 2092 NewFilterElts.reserve(NumTypeInfos); 2093 2094 // Remove any filter elements that were already caught or that already 2095 // occurred in the filter. While there, see if any of the elements are 2096 // catch-alls. If so, the filter can be discarded. 2097 bool SawCatchAll = false; 2098 for (unsigned j = 0; j != NumTypeInfos; ++j) { 2099 Value *Elt = Filter->getOperand(j); 2100 Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts()); 2101 if (isCatchAll(Personality, TypeInfo)) { 2102 // This element is a catch-all. Bail out, noting this fact. 2103 SawCatchAll = true; 2104 break; 2105 } 2106 if (AlreadyCaught.count(TypeInfo)) 2107 // Already caught by an earlier clause, so having it in the filter 2108 // is pointless. 2109 continue; 2110 // There is no point in having multiple copies of the same typeinfo in 2111 // a filter, so only add it if we didn't already. 2112 if (SeenInFilter.insert(TypeInfo)) 2113 NewFilterElts.push_back(cast<Constant>(Elt)); 2114 } 2115 // A filter containing a catch-all cannot match anything by definition. 2116 if (SawCatchAll) { 2117 // Throw the filter away. 2118 MakeNewInstruction = true; 2119 continue; 2120 } 2121 2122 // If we dropped something from the filter, make a new one. 2123 if (NewFilterElts.size() < NumTypeInfos) 2124 MakeNewFilter = true; 2125 } 2126 if (MakeNewFilter) { 2127 FilterType = ArrayType::get(FilterType->getElementType(), 2128 NewFilterElts.size()); 2129 FilterClause = ConstantArray::get(FilterType, NewFilterElts); 2130 MakeNewInstruction = true; 2131 } 2132 2133 NewClauses.push_back(FilterClause); 2134 2135 // If the new filter is empty then it will catch everything so there is 2136 // no point in keeping any following clauses or marking the landingpad 2137 // as having a cleanup. The case of the original filter being empty was 2138 // already handled above. 2139 if (MakeNewFilter && !NewFilterElts.size()) { 2140 assert(MakeNewInstruction && "New filter but not a new instruction!"); 2141 CleanupFlag = false; 2142 break; 2143 } 2144 } 2145 } 2146 2147 // If several filters occur in a row then reorder them so that the shortest 2148 // filters come first (those with the smallest number of elements). This is 2149 // advantageous because shorter filters are more likely to match, speeding up 2150 // unwinding, but mostly because it increases the effectiveness of the other 2151 // filter optimizations below. 2152 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) { 2153 unsigned j; 2154 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters. 2155 for (j = i; j != e; ++j) 2156 if (!isa<ArrayType>(NewClauses[j]->getType())) 2157 break; 2158 2159 // Check whether the filters are already sorted by length. We need to know 2160 // if sorting them is actually going to do anything so that we only make a 2161 // new landingpad instruction if it does. 2162 for (unsigned k = i; k + 1 < j; ++k) 2163 if (shorter_filter(NewClauses[k+1], NewClauses[k])) { 2164 // Not sorted, so sort the filters now. Doing an unstable sort would be 2165 // correct too but reordering filters pointlessly might confuse users. 2166 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j, 2167 shorter_filter); 2168 MakeNewInstruction = true; 2169 break; 2170 } 2171 2172 // Look for the next batch of filters. 2173 i = j + 1; 2174 } 2175 2176 // If typeinfos matched if and only if equal, then the elements of a filter L 2177 // that occurs later than a filter F could be replaced by the intersection of 2178 // the elements of F and L. In reality two typeinfos can match without being 2179 // equal (for example if one represents a C++ class, and the other some class 2180 // derived from it) so it would be wrong to perform this transform in general. 2181 // However the transform is correct and useful if F is a subset of L. In that 2182 // case L can be replaced by F, and thus removed altogether since repeating a 2183 // filter is pointless. So here we look at all pairs of filters F and L where 2184 // L follows F in the list of clauses, and remove L if every element of F is 2185 // an element of L. This can occur when inlining C++ functions with exception 2186 // specifications. 2187 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) { 2188 // Examine each filter in turn. 2189 Value *Filter = NewClauses[i]; 2190 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType()); 2191 if (!FTy) 2192 // Not a filter - skip it. 2193 continue; 2194 unsigned FElts = FTy->getNumElements(); 2195 // Examine each filter following this one. Doing this backwards means that 2196 // we don't have to worry about filters disappearing under us when removed. 2197 for (unsigned j = NewClauses.size() - 1; j != i; --j) { 2198 Value *LFilter = NewClauses[j]; 2199 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType()); 2200 if (!LTy) 2201 // Not a filter - skip it. 2202 continue; 2203 // If Filter is a subset of LFilter, i.e. every element of Filter is also 2204 // an element of LFilter, then discard LFilter. 2205 SmallVectorImpl<Value *>::iterator J = NewClauses.begin() + j; 2206 // If Filter is empty then it is a subset of LFilter. 2207 if (!FElts) { 2208 // Discard LFilter. 2209 NewClauses.erase(J); 2210 MakeNewInstruction = true; 2211 // Move on to the next filter. 2212 continue; 2213 } 2214 unsigned LElts = LTy->getNumElements(); 2215 // If Filter is longer than LFilter then it cannot be a subset of it. 2216 if (FElts > LElts) 2217 // Move on to the next filter. 2218 continue; 2219 // At this point we know that LFilter has at least one element. 2220 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros. 2221 // Filter is a subset of LFilter iff Filter contains only zeros (as we 2222 // already know that Filter is not longer than LFilter). 2223 if (isa<ConstantAggregateZero>(Filter)) { 2224 assert(FElts <= LElts && "Should have handled this case earlier!"); 2225 // Discard LFilter. 2226 NewClauses.erase(J); 2227 MakeNewInstruction = true; 2228 } 2229 // Move on to the next filter. 2230 continue; 2231 } 2232 ConstantArray *LArray = cast<ConstantArray>(LFilter); 2233 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros. 2234 // Since Filter is non-empty and contains only zeros, it is a subset of 2235 // LFilter iff LFilter contains a zero. 2236 assert(FElts > 0 && "Should have eliminated the empty filter earlier!"); 2237 for (unsigned l = 0; l != LElts; ++l) 2238 if (LArray->getOperand(l)->isNullValue()) { 2239 // LFilter contains a zero - discard it. 2240 NewClauses.erase(J); 2241 MakeNewInstruction = true; 2242 break; 2243 } 2244 // Move on to the next filter. 2245 continue; 2246 } 2247 // At this point we know that both filters are ConstantArrays. Loop over 2248 // operands to see whether every element of Filter is also an element of 2249 // LFilter. Since filters tend to be short this is probably faster than 2250 // using a method that scales nicely. 2251 ConstantArray *FArray = cast<ConstantArray>(Filter); 2252 bool AllFound = true; 2253 for (unsigned f = 0; f != FElts; ++f) { 2254 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts(); 2255 AllFound = false; 2256 for (unsigned l = 0; l != LElts; ++l) { 2257 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts(); 2258 if (LTypeInfo == FTypeInfo) { 2259 AllFound = true; 2260 break; 2261 } 2262 } 2263 if (!AllFound) 2264 break; 2265 } 2266 if (AllFound) { 2267 // Discard LFilter. 2268 NewClauses.erase(J); 2269 MakeNewInstruction = true; 2270 } 2271 // Move on to the next filter. 2272 } 2273 } 2274 2275 // If we changed any of the clauses, replace the old landingpad instruction 2276 // with a new one. 2277 if (MakeNewInstruction) { 2278 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(), 2279 LI.getPersonalityFn(), 2280 NewClauses.size()); 2281 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i) 2282 NLI->addClause(NewClauses[i]); 2283 // A landing pad with no clauses must have the cleanup flag set. It is 2284 // theoretically possible, though highly unlikely, that we eliminated all 2285 // clauses. If so, force the cleanup flag to true. 2286 if (NewClauses.empty()) 2287 CleanupFlag = true; 2288 NLI->setCleanup(CleanupFlag); 2289 return NLI; 2290 } 2291 2292 // Even if none of the clauses changed, we may nonetheless have understood 2293 // that the cleanup flag is pointless. Clear it if so. 2294 if (LI.isCleanup() != CleanupFlag) { 2295 assert(!CleanupFlag && "Adding a cleanup, not removing one?!"); 2296 LI.setCleanup(CleanupFlag); 2297 return &LI; 2298 } 2299 2300 return nullptr; 2301} 2302 2303 2304 2305 2306/// TryToSinkInstruction - Try to move the specified instruction from its 2307/// current block into the beginning of DestBlock, which can only happen if it's 2308/// safe to move the instruction past all of the instructions between it and the 2309/// end of its block. 2310static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 2311 assert(I->hasOneUse() && "Invariants didn't hold!"); 2312 2313 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 2314 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() || 2315 isa<TerminatorInst>(I)) 2316 return false; 2317 2318 // Do not sink alloca instructions out of the entry block. 2319 if (isa<AllocaInst>(I) && I->getParent() == 2320 &DestBlock->getParent()->getEntryBlock()) 2321 return false; 2322 2323 // We can only sink load instructions if there is nothing between the load and 2324 // the end of block that could change the value. 2325 if (I->mayReadFromMemory()) { 2326 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 2327 Scan != E; ++Scan) 2328 if (Scan->mayWriteToMemory()) 2329 return false; 2330 } 2331 2332 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt(); 2333 I->moveBefore(InsertPos); 2334 ++NumSunkInst; 2335 return true; 2336} 2337 2338 2339/// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 2340/// all reachable code to the worklist. 2341/// 2342/// This has a couple of tricks to make the code faster and more powerful. In 2343/// particular, we constant fold and DCE instructions as we go, to avoid adding 2344/// them to the worklist (this significantly speeds up instcombine on code where 2345/// many instructions are dead or constant). Additionally, if we find a branch 2346/// whose condition is a known constant, we only visit the reachable successors. 2347/// 2348static bool AddReachableCodeToWorklist(BasicBlock *BB, 2349 SmallPtrSet<BasicBlock*, 64> &Visited, 2350 InstCombiner &IC, 2351 const DataLayout *DL, 2352 const TargetLibraryInfo *TLI) { 2353 bool MadeIRChange = false; 2354 SmallVector<BasicBlock*, 256> Worklist; 2355 Worklist.push_back(BB); 2356 2357 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 2358 DenseMap<ConstantExpr*, Constant*> FoldedConstants; 2359 2360 do { 2361 BB = Worklist.pop_back_val(); 2362 2363 // We have now visited this block! If we've already been here, ignore it. 2364 if (!Visited.insert(BB)) continue; 2365 2366 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 2367 Instruction *Inst = BBI++; 2368 2369 // DCE instruction if trivially dead. 2370 if (isInstructionTriviallyDead(Inst, TLI)) { 2371 ++NumDeadInst; 2372 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); 2373 Inst->eraseFromParent(); 2374 continue; 2375 } 2376 2377 // ConstantProp instruction if trivially constant. 2378 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 2379 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) { 2380 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " 2381 << *Inst << '\n'); 2382 Inst->replaceAllUsesWith(C); 2383 ++NumConstProp; 2384 Inst->eraseFromParent(); 2385 continue; 2386 } 2387 2388 if (DL) { 2389 // See if we can constant fold its operands. 2390 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 2391 i != e; ++i) { 2392 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 2393 if (CE == nullptr) continue; 2394 2395 Constant*& FoldRes = FoldedConstants[CE]; 2396 if (!FoldRes) 2397 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); 2398 if (!FoldRes) 2399 FoldRes = CE; 2400 2401 if (FoldRes != CE) { 2402 *i = FoldRes; 2403 MadeIRChange = true; 2404 } 2405 } 2406 } 2407 2408 InstrsForInstCombineWorklist.push_back(Inst); 2409 } 2410 2411 // Recursively visit successors. If this is a branch or switch on a 2412 // constant, only visit the reachable successor. 2413 TerminatorInst *TI = BB->getTerminator(); 2414 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2415 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 2416 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 2417 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 2418 Worklist.push_back(ReachableBB); 2419 continue; 2420 } 2421 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2422 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 2423 // See if this is an explicit destination. 2424 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2425 i != e; ++i) 2426 if (i.getCaseValue() == Cond) { 2427 BasicBlock *ReachableBB = i.getCaseSuccessor(); 2428 Worklist.push_back(ReachableBB); 2429 continue; 2430 } 2431 2432 // Otherwise it is the default destination. 2433 Worklist.push_back(SI->getDefaultDest()); 2434 continue; 2435 } 2436 } 2437 2438 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 2439 Worklist.push_back(TI->getSuccessor(i)); 2440 } while (!Worklist.empty()); 2441 2442 // Once we've found all of the instructions to add to instcombine's worklist, 2443 // add them in reverse order. This way instcombine will visit from the top 2444 // of the function down. This jives well with the way that it adds all uses 2445 // of instructions to the worklist after doing a transformation, thus avoiding 2446 // some N^2 behavior in pathological cases. 2447 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 2448 InstrsForInstCombineWorklist.size()); 2449 2450 return MadeIRChange; 2451} 2452 2453bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 2454 MadeIRChange = false; 2455 2456 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 2457 << F.getName() << "\n"); 2458 2459 { 2460 // Do a depth-first traversal of the function, populate the worklist with 2461 // the reachable instructions. Ignore blocks that are not reachable. Keep 2462 // track of which blocks we visit. 2463 SmallPtrSet<BasicBlock*, 64> Visited; 2464 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL, 2465 TLI); 2466 2467 // Do a quick scan over the function. If we find any blocks that are 2468 // unreachable, remove any instructions inside of them. This prevents 2469 // the instcombine code from having to deal with some bad special cases. 2470 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 2471 if (Visited.count(BB)) continue; 2472 2473 // Delete the instructions backwards, as it has a reduced likelihood of 2474 // having to update as many def-use and use-def chains. 2475 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 2476 while (EndInst != BB->begin()) { 2477 // Delete the next to last instruction. 2478 BasicBlock::iterator I = EndInst; 2479 Instruction *Inst = --I; 2480 if (!Inst->use_empty()) 2481 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 2482 if (isa<LandingPadInst>(Inst)) { 2483 EndInst = Inst; 2484 continue; 2485 } 2486 if (!isa<DbgInfoIntrinsic>(Inst)) { 2487 ++NumDeadInst; 2488 MadeIRChange = true; 2489 } 2490 Inst->eraseFromParent(); 2491 } 2492 } 2493 } 2494 2495 while (!Worklist.isEmpty()) { 2496 Instruction *I = Worklist.RemoveOne(); 2497 if (I == nullptr) continue; // skip null values. 2498 2499 // Check to see if we can DCE the instruction. 2500 if (isInstructionTriviallyDead(I, TLI)) { 2501 DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); 2502 EraseInstFromFunction(*I); 2503 ++NumDeadInst; 2504 MadeIRChange = true; 2505 continue; 2506 } 2507 2508 // Instruction isn't dead, see if we can constant propagate it. 2509 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 2510 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) { 2511 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 2512 2513 // Add operands to the worklist. 2514 ReplaceInstUsesWith(*I, C); 2515 ++NumConstProp; 2516 EraseInstFromFunction(*I); 2517 MadeIRChange = true; 2518 continue; 2519 } 2520 2521 // See if we can trivially sink this instruction to a successor basic block. 2522 if (I->hasOneUse()) { 2523 BasicBlock *BB = I->getParent(); 2524 Instruction *UserInst = cast<Instruction>(*I->user_begin()); 2525 BasicBlock *UserParent; 2526 2527 // Get the block the use occurs in. 2528 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 2529 UserParent = PN->getIncomingBlock(*I->use_begin()); 2530 else 2531 UserParent = UserInst->getParent(); 2532 2533 if (UserParent != BB) { 2534 bool UserIsSuccessor = false; 2535 // See if the user is one of our successors. 2536 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 2537 if (*SI == UserParent) { 2538 UserIsSuccessor = true; 2539 break; 2540 } 2541 2542 // If the user is one of our immediate successors, and if that successor 2543 // only has us as a predecessors (we'd have to split the critical edge 2544 // otherwise), we can keep going. 2545 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 2546 // Okay, the CFG is simple enough, try to sink this instruction. 2547 MadeIRChange |= TryToSinkInstruction(I, UserParent); 2548 } 2549 } 2550 2551 // Now that we have an instruction, try combining it to simplify it. 2552 Builder->SetInsertPoint(I->getParent(), I); 2553 Builder->SetCurrentDebugLocation(I->getDebugLoc()); 2554 2555#ifndef NDEBUG 2556 std::string OrigI; 2557#endif 2558 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 2559 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); 2560 2561 if (Instruction *Result = visit(*I)) { 2562 ++NumCombined; 2563 // Should we replace the old instruction with a new one? 2564 if (Result != I) { 2565 DEBUG(dbgs() << "IC: Old = " << *I << '\n' 2566 << " New = " << *Result << '\n'); 2567 2568 if (!I->getDebugLoc().isUnknown()) 2569 Result->setDebugLoc(I->getDebugLoc()); 2570 // Everything uses the new instruction now. 2571 I->replaceAllUsesWith(Result); 2572 2573 // Move the name to the new instruction first. 2574 Result->takeName(I); 2575 2576 // Push the new instruction and any users onto the worklist. 2577 Worklist.Add(Result); 2578 Worklist.AddUsersToWorkList(*Result); 2579 2580 // Insert the new instruction into the basic block... 2581 BasicBlock *InstParent = I->getParent(); 2582 BasicBlock::iterator InsertPos = I; 2583 2584 // If we replace a PHI with something that isn't a PHI, fix up the 2585 // insertion point. 2586 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos)) 2587 InsertPos = InstParent->getFirstInsertionPt(); 2588 2589 InstParent->getInstList().insert(InsertPos, Result); 2590 2591 EraseInstFromFunction(*I); 2592 } else { 2593#ifndef NDEBUG 2594 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' 2595 << " New = " << *I << '\n'); 2596#endif 2597 2598 // If the instruction was modified, it's possible that it is now dead. 2599 // if so, remove it. 2600 if (isInstructionTriviallyDead(I, TLI)) { 2601 EraseInstFromFunction(*I); 2602 } else { 2603 Worklist.Add(I); 2604 Worklist.AddUsersToWorkList(*I); 2605 } 2606 } 2607 MadeIRChange = true; 2608 } 2609 } 2610 2611 Worklist.Zap(); 2612 return MadeIRChange; 2613} 2614 2615namespace { 2616class InstCombinerLibCallSimplifier : public LibCallSimplifier { 2617 InstCombiner *IC; 2618public: 2619 InstCombinerLibCallSimplifier(const DataLayout *DL, 2620 const TargetLibraryInfo *TLI, 2621 InstCombiner *IC) 2622 : LibCallSimplifier(DL, TLI, UnsafeFPShrink) { 2623 this->IC = IC; 2624 } 2625 2626 /// replaceAllUsesWith - override so that instruction replacement 2627 /// can be defined in terms of the instruction combiner framework. 2628 void replaceAllUsesWith(Instruction *I, Value *With) const override { 2629 IC->ReplaceInstUsesWith(*I, With); 2630 } 2631}; 2632} 2633 2634bool InstCombiner::runOnFunction(Function &F) { 2635 if (skipOptnoneFunction(F)) 2636 return false; 2637 2638 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 2639 DL = DLP ? &DLP->getDataLayout() : nullptr; 2640 TLI = &getAnalysis<TargetLibraryInfo>(); 2641 // Minimizing size? 2642 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 2643 Attribute::MinSize); 2644 2645 /// Builder - This is an IRBuilder that automatically inserts new 2646 /// instructions into the worklist when they are created. 2647 IRBuilder<true, TargetFolder, InstCombineIRInserter> 2648 TheBuilder(F.getContext(), TargetFolder(DL), 2649 InstCombineIRInserter(Worklist)); 2650 Builder = &TheBuilder; 2651 2652 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this); 2653 Simplifier = &TheSimplifier; 2654 2655 bool EverMadeChange = false; 2656 2657 // Lower dbg.declare intrinsics otherwise their value may be clobbered 2658 // by instcombiner. 2659 EverMadeChange = LowerDbgDeclare(F); 2660 2661 // Iterate while there is work to do. 2662 unsigned Iteration = 0; 2663 while (DoOneIteration(F, Iteration++)) 2664 EverMadeChange = true; 2665 2666 Builder = nullptr; 2667 return EverMadeChange; 2668} 2669 2670FunctionPass *llvm::createInstructionCombiningPass() { 2671 return new InstCombiner(); 2672} 2673