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