InstructionCombining.cpp revision 263508
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> 58using namespace llvm; 59using namespace llvm::PatternMatch; 60 61STATISTIC(NumCombined , "Number of insts combined"); 62STATISTIC(NumConstProp, "Number of constant folds"); 63STATISTIC(NumDeadInst , "Number of dead inst eliminated"); 64STATISTIC(NumSunkInst , "Number of instructions sunk"); 65STATISTIC(NumExpand, "Number of expansions"); 66STATISTIC(NumFactor , "Number of factorizations"); 67STATISTIC(NumReassoc , "Number of reassociations"); 68 69static 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 75void llvm::initializeInstCombine(PassRegistry &Registry) { 76 initializeInstCombinerPass(Registry); 77} 78 79void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { 80 initializeInstCombine(*unwrap(R)); 81} 82 83char InstCombiner::ID = 0; 84INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine", 85 "Combine redundant instructions", false, false) 86INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 87INITIALIZE_PASS_END(InstCombiner, "instcombine", 88 "Combine redundant instructions", false, false) 89 90void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 91 AU.setPreservesCFG(); 92 AU.addRequired<TargetLibraryInfo>(); 93} 94 95 96Value *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. 103bool 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. 132static 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. 168static 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// 201bool 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)". 342static 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)". 381static 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. 396Value *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// 515Value *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// 534Value *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 549static 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. 585Instruction *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/// 622Instruction *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 // Beware of ConstantExpr: it may eventually evaluate to getNullValue, 703 // even if currently isNullValue gives false. 704 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)); 705 if (InC && !isa<ConstantExpr>(InC)) 706 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; 707 else 708 InV = Builder->CreateSelect(PN->getIncomingValue(i), 709 TrueVInPred, FalseVInPred, "phitmp"); 710 NewPN->addIncoming(InV, ThisBB); 711 } 712 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) { 713 Constant *C = cast<Constant>(I.getOperand(1)); 714 for (unsigned i = 0; i != NumPHIValues; ++i) { 715 Value *InV = 0; 716 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 717 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); 718 else if (isa<ICmpInst>(CI)) 719 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i), 720 C, "phitmp"); 721 else 722 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i), 723 C, "phitmp"); 724 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 725 } 726 } else if (I.getNumOperands() == 2) { 727 Constant *C = cast<Constant>(I.getOperand(1)); 728 for (unsigned i = 0; i != NumPHIValues; ++i) { 729 Value *InV = 0; 730 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 731 InV = ConstantExpr::get(I.getOpcode(), InC, C); 732 else 733 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(), 734 PN->getIncomingValue(i), C, "phitmp"); 735 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 736 } 737 } else { 738 CastInst *CI = cast<CastInst>(&I); 739 Type *RetTy = CI->getType(); 740 for (unsigned i = 0; i != NumPHIValues; ++i) { 741 Value *InV; 742 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) 743 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); 744 else 745 InV = Builder->CreateCast(CI->getOpcode(), 746 PN->getIncomingValue(i), I.getType(), "phitmp"); 747 NewPN->addIncoming(InV, PN->getIncomingBlock(i)); 748 } 749 } 750 751 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); 752 UI != E; ) { 753 Instruction *User = cast<Instruction>(*UI++); 754 if (User == &I) continue; 755 ReplaceInstUsesWith(*User, NewPN); 756 EraseInstFromFunction(*User); 757 } 758 return ReplaceInstUsesWith(I, NewPN); 759} 760 761/// FindElementAtOffset - Given a pointer type and a constant offset, determine 762/// whether or not there is a sequence of GEP indices into the pointed type that 763/// will land us at the specified offset. If so, fill them into NewIndices and 764/// return the resultant element type, otherwise return null. 765Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset, 766 SmallVectorImpl<Value*> &NewIndices) { 767 assert(PtrTy->isPtrOrPtrVectorTy()); 768 769 if (!TD) 770 return 0; 771 772 Type *Ty = PtrTy->getPointerElementType(); 773 if (!Ty->isSized()) 774 return 0; 775 776 // Start with the index over the outer type. Note that the type size 777 // might be zero (even if the offset isn't zero) if the indexed type 778 // is something like [0 x {int, int}] 779 Type *IntPtrTy = TD->getIntPtrType(PtrTy); 780 int64_t FirstIdx = 0; 781 if (int64_t TySize = TD->getTypeAllocSize(Ty)) { 782 FirstIdx = Offset/TySize; 783 Offset -= FirstIdx*TySize; 784 785 // Handle hosts where % returns negative instead of values [0..TySize). 786 if (Offset < 0) { 787 --FirstIdx; 788 Offset += TySize; 789 assert(Offset >= 0); 790 } 791 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); 792 } 793 794 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); 795 796 // Index into the types. If we fail, set OrigBase to null. 797 while (Offset) { 798 // Indexing into tail padding between struct/array elements. 799 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty)) 800 return 0; 801 802 if (StructType *STy = dyn_cast<StructType>(Ty)) { 803 const StructLayout *SL = TD->getStructLayout(STy); 804 assert(Offset < (int64_t)SL->getSizeInBytes() && 805 "Offset must stay within the indexed type"); 806 807 unsigned Elt = SL->getElementContainingOffset(Offset); 808 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 809 Elt)); 810 811 Offset -= SL->getElementOffset(Elt); 812 Ty = STy->getElementType(Elt); 813 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 814 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType()); 815 assert(EltSize && "Cannot index into a zero-sized array"); 816 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); 817 Offset %= EltSize; 818 Ty = AT->getElementType(); 819 } else { 820 // Otherwise, we can't index into the middle of this atomic type, bail. 821 return 0; 822 } 823 } 824 825 return Ty; 826} 827 828static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { 829 // If this GEP has only 0 indices, it is the same pointer as 830 // Src. If Src is not a trivial GEP too, don't combine 831 // the indices. 832 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() && 833 !Src.hasOneUse()) 834 return false; 835 return true; 836} 837 838/// Descale - Return a value X such that Val = X * Scale, or null if none. If 839/// the multiplication is known not to overflow then NoSignedWrap is set. 840Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { 841 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); 842 assert(cast<IntegerType>(Val->getType())->getBitWidth() == 843 Scale.getBitWidth() && "Scale not compatible with value!"); 844 845 // If Val is zero or Scale is one then Val = Val * Scale. 846 if (match(Val, m_Zero()) || Scale == 1) { 847 NoSignedWrap = true; 848 return Val; 849 } 850 851 // If Scale is zero then it does not divide Val. 852 if (Scale.isMinValue()) 853 return 0; 854 855 // Look through chains of multiplications, searching for a constant that is 856 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 857 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by 858 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore 859 // down from Val: 860 // 861 // Val = M1 * X || Analysis starts here and works down 862 // M1 = M2 * Y || Doesn't descend into terms with more 863 // M2 = Z * 4 \/ than one use 864 // 865 // Then to modify a term at the bottom: 866 // 867 // Val = M1 * X 868 // M1 = Z * Y || Replaced M2 with Z 869 // 870 // Then to work back up correcting nsw flags. 871 872 // Op - the term we are currently analyzing. Starts at Val then drills down. 873 // Replaced with its descaled value before exiting from the drill down loop. 874 Value *Op = Val; 875 876 // Parent - initially null, but after drilling down notes where Op came from. 877 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the 878 // 0'th operand of Val. 879 std::pair<Instruction*, unsigned> Parent; 880 881 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper 882 // levels that doesn't overflow. 883 bool RequireNoSignedWrap = false; 884 885 // logScale - log base 2 of the scale. Negative if not a power of 2. 886 int32_t logScale = Scale.exactLogBase2(); 887 888 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down 889 890 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 891 // If Op is a constant divisible by Scale then descale to the quotient. 892 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. 893 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); 894 if (!Remainder.isMinValue()) 895 // Not divisible by Scale. 896 return 0; 897 // Replace with the quotient in the parent. 898 Op = ConstantInt::get(CI->getType(), Quotient); 899 NoSignedWrap = true; 900 break; 901 } 902 903 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { 904 905 if (BO->getOpcode() == Instruction::Mul) { 906 // Multiplication. 907 NoSignedWrap = BO->hasNoSignedWrap(); 908 if (RequireNoSignedWrap && !NoSignedWrap) 909 return 0; 910 911 // There are three cases for multiplication: multiplication by exactly 912 // the scale, multiplication by a constant different to the scale, and 913 // multiplication by something else. 914 Value *LHS = BO->getOperand(0); 915 Value *RHS = BO->getOperand(1); 916 917 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 918 // Multiplication by a constant. 919 if (CI->getValue() == Scale) { 920 // Multiplication by exactly the scale, replace the multiplication 921 // by its left-hand side in the parent. 922 Op = LHS; 923 break; 924 } 925 926 // Otherwise drill down into the constant. 927 if (!Op->hasOneUse()) 928 return 0; 929 930 Parent = std::make_pair(BO, 1); 931 continue; 932 } 933 934 // Multiplication by something else. Drill down into the left-hand side 935 // since that's where the reassociate pass puts the good stuff. 936 if (!Op->hasOneUse()) 937 return 0; 938 939 Parent = std::make_pair(BO, 0); 940 continue; 941 } 942 943 if (logScale > 0 && BO->getOpcode() == Instruction::Shl && 944 isa<ConstantInt>(BO->getOperand(1))) { 945 // Multiplication by a power of 2. 946 NoSignedWrap = BO->hasNoSignedWrap(); 947 if (RequireNoSignedWrap && !NoSignedWrap) 948 return 0; 949 950 Value *LHS = BO->getOperand(0); 951 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> 952 getLimitedValue(Scale.getBitWidth()); 953 // Op = LHS << Amt. 954 955 if (Amt == logScale) { 956 // Multiplication by exactly the scale, replace the multiplication 957 // by its left-hand side in the parent. 958 Op = LHS; 959 break; 960 } 961 if (Amt < logScale || !Op->hasOneUse()) 962 return 0; 963 964 // Multiplication by more than the scale. Reduce the multiplying amount 965 // by the scale in the parent. 966 Parent = std::make_pair(BO, 1); 967 Op = ConstantInt::get(BO->getType(), Amt - logScale); 968 break; 969 } 970 } 971 972 if (!Op->hasOneUse()) 973 return 0; 974 975 if (CastInst *Cast = dyn_cast<CastInst>(Op)) { 976 if (Cast->getOpcode() == Instruction::SExt) { 977 // Op is sign-extended from a smaller type, descale in the smaller type. 978 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 979 APInt SmallScale = Scale.trunc(SmallSize); 980 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to 981 // descale Op as (sext Y) * Scale. In order to have 982 // sext (Y * SmallScale) = (sext Y) * Scale 983 // some conditions need to hold however: SmallScale must sign-extend to 984 // Scale and the multiplication Y * SmallScale should not overflow. 985 if (SmallScale.sext(Scale.getBitWidth()) != Scale) 986 // SmallScale does not sign-extend to Scale. 987 return 0; 988 assert(SmallScale.exactLogBase2() == logScale); 989 // Require that Y * SmallScale must not overflow. 990 RequireNoSignedWrap = true; 991 992 // Drill down through the cast. 993 Parent = std::make_pair(Cast, 0); 994 Scale = SmallScale; 995 continue; 996 } 997 998 if (Cast->getOpcode() == Instruction::Trunc) { 999 // Op is truncated from a larger type, descale in the larger type. 1000 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then 1001 // trunc (Y * sext Scale) = (trunc Y) * Scale 1002 // always holds. However (trunc Y) * Scale may overflow even if 1003 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared 1004 // from this point up in the expression (see later). 1005 if (RequireNoSignedWrap) 1006 return 0; 1007 1008 // Drill down through the cast. 1009 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); 1010 Parent = std::make_pair(Cast, 0); 1011 Scale = Scale.sext(LargeSize); 1012 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) 1013 logScale = -1; 1014 assert(Scale.exactLogBase2() == logScale); 1015 continue; 1016 } 1017 } 1018 1019 // Unsupported expression, bail out. 1020 return 0; 1021 } 1022 1023 // We know that we can successfully descale, so from here on we can safely 1024 // modify the IR. Op holds the descaled version of the deepest term in the 1025 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known 1026 // not to overflow. 1027 1028 if (!Parent.first) 1029 // The expression only had one term. 1030 return Op; 1031 1032 // Rewrite the parent using the descaled version of its operand. 1033 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); 1034 assert(Op != Parent.first->getOperand(Parent.second) && 1035 "Descaling was a no-op?"); 1036 Parent.first->setOperand(Parent.second, Op); 1037 Worklist.Add(Parent.first); 1038 1039 // Now work back up the expression correcting nsw flags. The logic is based 1040 // on the following observation: if X * Y is known not to overflow as a signed 1041 // multiplication, and Y is replaced by a value Z with smaller absolute value, 1042 // then X * Z will not overflow as a signed multiplication either. As we work 1043 // our way up, having NoSignedWrap 'true' means that the descaled value at the 1044 // current level has strictly smaller absolute value than the original. 1045 Instruction *Ancestor = Parent.first; 1046 do { 1047 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { 1048 // If the multiplication wasn't nsw then we can't say anything about the 1049 // value of the descaled multiplication, and we have to clear nsw flags 1050 // from this point on up. 1051 bool OpNoSignedWrap = BO->hasNoSignedWrap(); 1052 NoSignedWrap &= OpNoSignedWrap; 1053 if (NoSignedWrap != OpNoSignedWrap) { 1054 BO->setHasNoSignedWrap(NoSignedWrap); 1055 Worklist.Add(Ancestor); 1056 } 1057 } else if (Ancestor->getOpcode() == Instruction::Trunc) { 1058 // The fact that the descaled input to the trunc has smaller absolute 1059 // value than the original input doesn't tell us anything useful about 1060 // the absolute values of the truncations. 1061 NoSignedWrap = false; 1062 } 1063 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && 1064 "Failed to keep proper track of nsw flags while drilling down?"); 1065 1066 if (Ancestor == Val) 1067 // Got to the top, all done! 1068 return Val; 1069 1070 // Move up one level in the expression. 1071 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); 1072 Ancestor = Ancestor->use_back(); 1073 } while (1); 1074} 1075 1076Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { 1077 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); 1078 1079 if (Value *V = SimplifyGEPInst(Ops, TD)) 1080 return ReplaceInstUsesWith(GEP, V); 1081 1082 Value *PtrOp = GEP.getOperand(0); 1083 1084 // Eliminate unneeded casts for indices, and replace indices which displace 1085 // by multiples of a zero size type with zero. 1086 if (TD) { 1087 bool MadeChange = false; 1088 Type *IntPtrTy = TD->getIntPtrType(GEP.getPointerOperandType()); 1089 1090 gep_type_iterator GTI = gep_type_begin(GEP); 1091 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); 1092 I != E; ++I, ++GTI) { 1093 // Skip indices into struct types. 1094 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI); 1095 if (!SeqTy) continue; 1096 1097 // If the element type has zero size then any index over it is equivalent 1098 // to an index of zero, so replace it with zero if it is not zero already. 1099 if (SeqTy->getElementType()->isSized() && 1100 TD->getTypeAllocSize(SeqTy->getElementType()) == 0) 1101 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) { 1102 *I = Constant::getNullValue(IntPtrTy); 1103 MadeChange = true; 1104 } 1105 1106 Type *IndexTy = (*I)->getType(); 1107 if (IndexTy != IntPtrTy) { 1108 // If we are using a wider index than needed for this platform, shrink 1109 // it to what we need. If narrower, sign-extend it to what we need. 1110 // This explicit cast can make subsequent optimizations more obvious. 1111 *I = Builder->CreateIntCast(*I, IntPtrTy, true); 1112 MadeChange = true; 1113 } 1114 } 1115 if (MadeChange) return &GEP; 1116 } 1117 1118 // Combine Indices - If the source pointer to this getelementptr instruction 1119 // is a getelementptr instruction, combine the indices of the two 1120 // getelementptr instructions into a single instruction. 1121 // 1122 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) { 1123 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) 1124 return 0; 1125 1126 // Note that if our source is a gep chain itself then we wait for that 1127 // chain to be resolved before we perform this transformation. This 1128 // avoids us creating a TON of code in some cases. 1129 if (GEPOperator *SrcGEP = 1130 dyn_cast<GEPOperator>(Src->getOperand(0))) 1131 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) 1132 return 0; // Wait until our source is folded to completion. 1133 1134 SmallVector<Value*, 8> Indices; 1135 1136 // Find out whether the last index in the source GEP is a sequential idx. 1137 bool EndsWithSequential = false; 1138 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src); 1139 I != E; ++I) 1140 EndsWithSequential = !(*I)->isStructTy(); 1141 1142 // Can we combine the two pointer arithmetics offsets? 1143 if (EndsWithSequential) { 1144 // Replace: gep (gep %P, long B), long A, ... 1145 // With: T = long A+B; gep %P, T, ... 1146 // 1147 Value *Sum; 1148 Value *SO1 = Src->getOperand(Src->getNumOperands()-1); 1149 Value *GO1 = GEP.getOperand(1); 1150 if (SO1 == Constant::getNullValue(SO1->getType())) { 1151 Sum = GO1; 1152 } else if (GO1 == Constant::getNullValue(GO1->getType())) { 1153 Sum = SO1; 1154 } else { 1155 // If they aren't the same type, then the input hasn't been processed 1156 // by the loop above yet (which canonicalizes sequential index types to 1157 // intptr_t). Just avoid transforming this until the input has been 1158 // normalized. 1159 if (SO1->getType() != GO1->getType()) 1160 return 0; 1161 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); 1162 } 1163 1164 // Update the GEP in place if possible. 1165 if (Src->getNumOperands() == 2) { 1166 GEP.setOperand(0, Src->getOperand(0)); 1167 GEP.setOperand(1, Sum); 1168 return &GEP; 1169 } 1170 Indices.append(Src->op_begin()+1, Src->op_end()-1); 1171 Indices.push_back(Sum); 1172 Indices.append(GEP.op_begin()+2, GEP.op_end()); 1173 } else if (isa<Constant>(*GEP.idx_begin()) && 1174 cast<Constant>(*GEP.idx_begin())->isNullValue() && 1175 Src->getNumOperands() != 1) { 1176 // Otherwise we can do the fold if the first index of the GEP is a zero 1177 Indices.append(Src->op_begin()+1, Src->op_end()); 1178 Indices.append(GEP.idx_begin()+1, GEP.idx_end()); 1179 } 1180 1181 if (!Indices.empty()) 1182 return (GEP.isInBounds() && Src->isInBounds()) ? 1183 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices, 1184 GEP.getName()) : 1185 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName()); 1186 } 1187 1188 // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y)) 1189 // The GEP pattern is emitted by the SCEV expander for certain kinds of 1190 // pointer arithmetic. 1191 if (TD && GEP.getNumIndices() == 1 && 1192 match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) { 1193 unsigned AS = GEP.getPointerAddressSpace(); 1194 if (GEP.getType() == Builder->getInt8PtrTy(AS) && 1195 GEP.getOperand(1)->getType()->getScalarSizeInBits() == 1196 TD->getPointerSizeInBits(AS)) { 1197 Operator *Index = cast<Operator>(GEP.getOperand(1)); 1198 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); 1199 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); 1200 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); 1201 } 1202 } 1203 1204 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). 1205 Value *StrippedPtr = PtrOp->stripPointerCasts(); 1206 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType()); 1207 1208 // We do not handle pointer-vector geps here. 1209 if (!StrippedPtrTy) 1210 return 0; 1211 1212 if (StrippedPtr != PtrOp && 1213 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 1214 1215 bool HasZeroPointerIndex = false; 1216 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1))) 1217 HasZeroPointerIndex = C->isZero(); 1218 1219 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... 1220 // into : GEP [10 x i8]* X, i32 0, ... 1221 // 1222 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ... 1223 // into : GEP i8* X, ... 1224 // 1225 // This occurs when the program declares an array extern like "int X[];" 1226 if (HasZeroPointerIndex) { 1227 PointerType *CPTy = cast<PointerType>(PtrOp->getType()); 1228 if (ArrayType *CATy = 1229 dyn_cast<ArrayType>(CPTy->getElementType())) { 1230 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ? 1231 if (CATy->getElementType() == StrippedPtrTy->getElementType()) { 1232 // -> GEP i8* X, ... 1233 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end()); 1234 GetElementPtrInst *Res = 1235 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName()); 1236 Res->setIsInBounds(GEP.isInBounds()); 1237 return Res; 1238 } 1239 1240 if (ArrayType *XATy = 1241 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){ 1242 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? 1243 if (CATy->getElementType() == XATy->getElementType()) { 1244 // -> GEP [10 x i8]* X, i32 0, ... 1245 // At this point, we know that the cast source type is a pointer 1246 // to an array of the same type as the destination pointer 1247 // array. Because the array type is never stepped over (there 1248 // is a leading zero) we can fold the cast into this GEP. 1249 GEP.setOperand(0, StrippedPtr); 1250 return &GEP; 1251 } 1252 } 1253 } 1254 } else if (GEP.getNumOperands() == 2) { 1255 // Transform things like: 1256 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V 1257 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast 1258 Type *SrcElTy = StrippedPtrTy->getElementType(); 1259 Type *ResElTy = PtrOp->getType()->getPointerElementType(); 1260 if (TD && SrcElTy->isArrayTy() && 1261 TD->getTypeAllocSize(SrcElTy->getArrayElementType()) == 1262 TD->getTypeAllocSize(ResElTy)) { 1263 Type *IdxType = TD->getIntPtrType(GEP.getType()); 1264 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; 1265 Value *NewGEP = GEP.isInBounds() ? 1266 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : 1267 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); 1268 // V and GEP are both pointer types --> BitCast 1269 return new BitCastInst(NewGEP, GEP.getType()); 1270 } 1271 1272 // Transform things like: 1273 // %V = mul i64 %N, 4 1274 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V 1275 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast 1276 if (TD && ResElTy->isSized() && SrcElTy->isSized()) { 1277 // Check that changing the type amounts to dividing the index by a scale 1278 // factor. 1279 uint64_t ResSize = TD->getTypeAllocSize(ResElTy); 1280 uint64_t SrcSize = TD->getTypeAllocSize(SrcElTy); 1281 if (ResSize && SrcSize % ResSize == 0) { 1282 Value *Idx = GEP.getOperand(1); 1283 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1284 uint64_t Scale = SrcSize / ResSize; 1285 1286 // Earlier transforms ensure that the index has type IntPtrType, which 1287 // considerably simplifies the logic by eliminating implicit casts. 1288 assert(Idx->getType() == TD->getIntPtrType(GEP.getType()) && 1289 "Index not cast to pointer width?"); 1290 1291 bool NSW; 1292 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1293 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1294 // If the multiplication NewIdx * Scale may overflow then the new 1295 // GEP may not be "inbounds". 1296 Value *NewGEP = GEP.isInBounds() && NSW ? 1297 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) : 1298 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName()); 1299 // The NewGEP must be pointer typed, so must the old one -> BitCast 1300 return new BitCastInst(NewGEP, GEP.getType()); 1301 } 1302 } 1303 } 1304 1305 // Similarly, transform things like: 1306 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp 1307 // (where tmp = 8*tmp2) into: 1308 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast 1309 if (TD && ResElTy->isSized() && SrcElTy->isSized() && 1310 SrcElTy->isArrayTy()) { 1311 // Check that changing to the array element type amounts to dividing the 1312 // index by a scale factor. 1313 uint64_t ResSize = TD->getTypeAllocSize(ResElTy); 1314 uint64_t ArrayEltSize 1315 = TD->getTypeAllocSize(SrcElTy->getArrayElementType()); 1316 if (ResSize && ArrayEltSize % ResSize == 0) { 1317 Value *Idx = GEP.getOperand(1); 1318 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); 1319 uint64_t Scale = ArrayEltSize / ResSize; 1320 1321 // Earlier transforms ensure that the index has type IntPtrType, which 1322 // considerably simplifies the logic by eliminating implicit casts. 1323 assert(Idx->getType() == TD->getIntPtrType(GEP.getType()) && 1324 "Index not cast to pointer width?"); 1325 1326 bool NSW; 1327 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { 1328 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. 1329 // If the multiplication NewIdx * Scale may overflow then the new 1330 // GEP may not be "inbounds". 1331 Value *Off[2] = { 1332 Constant::getNullValue(TD->getIntPtrType(GEP.getType())), 1333 NewIdx 1334 }; 1335 1336 Value *NewGEP = GEP.isInBounds() && NSW ? 1337 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) : 1338 Builder->CreateGEP(StrippedPtr, Off, GEP.getName()); 1339 // The NewGEP must be pointer typed, so must the old one -> BitCast 1340 return new BitCastInst(NewGEP, GEP.getType()); 1341 } 1342 } 1343 } 1344 } 1345 } 1346 1347 if (!TD) 1348 return 0; 1349 1350 /// See if we can simplify: 1351 /// X = bitcast A* to B* 1352 /// Y = gep X, <...constant indices...> 1353 /// into a gep of the original struct. This is important for SROA and alias 1354 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. 1355 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) { 1356 Value *Operand = BCI->getOperand(0); 1357 PointerType *OpType = cast<PointerType>(Operand->getType()); 1358 unsigned OffsetBits = TD->getPointerTypeSizeInBits(OpType); 1359 APInt Offset(OffsetBits, 0); 1360 if (!isa<BitCastInst>(Operand) && 1361 GEP.accumulateConstantOffset(*TD, Offset) && 1362 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { 1363 1364 // If this GEP instruction doesn't move the pointer, just replace the GEP 1365 // with a bitcast of the real input to the dest type. 1366 if (!Offset) { 1367 // If the bitcast is of an allocation, and the allocation will be 1368 // converted to match the type of the cast, don't touch this. 1369 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) { 1370 // See if the bitcast simplifies, if so, don't nuke this GEP yet. 1371 if (Instruction *I = visitBitCast(*BCI)) { 1372 if (I != BCI) { 1373 I->takeName(BCI); 1374 BCI->getParent()->getInstList().insert(BCI, I); 1375 ReplaceInstUsesWith(*BCI, I); 1376 } 1377 return &GEP; 1378 } 1379 } 1380 return new BitCastInst(Operand, GEP.getType()); 1381 } 1382 1383 // Otherwise, if the offset is non-zero, we need to find out if there is a 1384 // field at Offset in 'A's type. If so, we can pull the cast through the 1385 // GEP. 1386 SmallVector<Value*, 8> NewIndices; 1387 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { 1388 Value *NGEP = GEP.isInBounds() ? 1389 Builder->CreateInBoundsGEP(Operand, NewIndices) : 1390 Builder->CreateGEP(Operand, NewIndices); 1391 1392 if (NGEP->getType() == GEP.getType()) 1393 return ReplaceInstUsesWith(GEP, NGEP); 1394 NGEP->takeName(&GEP); 1395 return new BitCastInst(NGEP, GEP.getType()); 1396 } 1397 } 1398 } 1399 1400 return 0; 1401} 1402 1403static bool 1404isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, 1405 const TargetLibraryInfo *TLI) { 1406 SmallVector<Instruction*, 4> Worklist; 1407 Worklist.push_back(AI); 1408 1409 do { 1410 Instruction *PI = Worklist.pop_back_val(); 1411 for (Value::use_iterator UI = PI->use_begin(), UE = PI->use_end(); UI != UE; 1412 ++UI) { 1413 Instruction *I = cast<Instruction>(*UI); 1414 switch (I->getOpcode()) { 1415 default: 1416 // Give up the moment we see something we can't handle. 1417 return false; 1418 1419 case Instruction::BitCast: 1420 case Instruction::GetElementPtr: 1421 Users.push_back(I); 1422 Worklist.push_back(I); 1423 continue; 1424 1425 case Instruction::ICmp: { 1426 ICmpInst *ICI = cast<ICmpInst>(I); 1427 // We can fold eq/ne comparisons with null to false/true, respectively. 1428 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1))) 1429 return false; 1430 Users.push_back(I); 1431 continue; 1432 } 1433 1434 case Instruction::Call: 1435 // Ignore no-op and store intrinsics. 1436 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1437 switch (II->getIntrinsicID()) { 1438 default: 1439 return false; 1440 1441 case Intrinsic::memmove: 1442 case Intrinsic::memcpy: 1443 case Intrinsic::memset: { 1444 MemIntrinsic *MI = cast<MemIntrinsic>(II); 1445 if (MI->isVolatile() || MI->getRawDest() != PI) 1446 return false; 1447 } 1448 // fall through 1449 case Intrinsic::dbg_declare: 1450 case Intrinsic::dbg_value: 1451 case Intrinsic::invariant_start: 1452 case Intrinsic::invariant_end: 1453 case Intrinsic::lifetime_start: 1454 case Intrinsic::lifetime_end: 1455 case Intrinsic::objectsize: 1456 Users.push_back(I); 1457 continue; 1458 } 1459 } 1460 1461 if (isFreeCall(I, TLI)) { 1462 Users.push_back(I); 1463 continue; 1464 } 1465 return false; 1466 1467 case Instruction::Store: { 1468 StoreInst *SI = cast<StoreInst>(I); 1469 if (SI->isVolatile() || SI->getPointerOperand() != PI) 1470 return false; 1471 Users.push_back(I); 1472 continue; 1473 } 1474 } 1475 llvm_unreachable("missing a return?"); 1476 } 1477 } while (!Worklist.empty()); 1478 return true; 1479} 1480 1481Instruction *InstCombiner::visitAllocSite(Instruction &MI) { 1482 // If we have a malloc call which is only used in any amount of comparisons 1483 // to null and free calls, delete the calls and replace the comparisons with 1484 // true or false as appropriate. 1485 SmallVector<WeakVH, 64> Users; 1486 if (isAllocSiteRemovable(&MI, Users, TLI)) { 1487 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 1488 Instruction *I = cast_or_null<Instruction>(&*Users[i]); 1489 if (!I) continue; 1490 1491 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) { 1492 ReplaceInstUsesWith(*C, 1493 ConstantInt::get(Type::getInt1Ty(C->getContext()), 1494 C->isFalseWhenEqual())); 1495 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 1496 ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); 1497 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1498 if (II->getIntrinsicID() == Intrinsic::objectsize) { 1499 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1)); 1500 uint64_t DontKnow = CI->isZero() ? -1ULL : 0; 1501 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); 1502 } 1503 } 1504 EraseInstFromFunction(*I); 1505 } 1506 1507 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) { 1508 // Replace invoke with a NOP intrinsic to maintain the original CFG 1509 Module *M = II->getParent()->getParent()->getParent(); 1510 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); 1511 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), 1512 None, "", II->getParent()); 1513 } 1514 return EraseInstFromFunction(MI); 1515 } 1516 return 0; 1517} 1518 1519/// \brief Move the call to free before a NULL test. 1520/// 1521/// Check if this free is accessed after its argument has been test 1522/// against NULL (property 0). 1523/// If yes, it is legal to move this call in its predecessor block. 1524/// 1525/// The move is performed only if the block containing the call to free 1526/// will be removed, i.e.: 1527/// 1. it has only one predecessor P, and P has two successors 1528/// 2. it contains the call and an unconditional branch 1529/// 3. its successor is the same as its predecessor's successor 1530/// 1531/// The profitability is out-of concern here and this function should 1532/// be called only if the caller knows this transformation would be 1533/// profitable (e.g., for code size). 1534static Instruction * 1535tryToMoveFreeBeforeNullTest(CallInst &FI) { 1536 Value *Op = FI.getArgOperand(0); 1537 BasicBlock *FreeInstrBB = FI.getParent(); 1538 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); 1539 1540 // Validate part of constraint #1: Only one predecessor 1541 // FIXME: We can extend the number of predecessor, but in that case, we 1542 // would duplicate the call to free in each predecessor and it may 1543 // not be profitable even for code size. 1544 if (!PredBB) 1545 return 0; 1546 1547 // Validate constraint #2: Does this block contains only the call to 1548 // free and an unconditional branch? 1549 // FIXME: We could check if we can speculate everything in the 1550 // predecessor block 1551 if (FreeInstrBB->size() != 2) 1552 return 0; 1553 BasicBlock *SuccBB; 1554 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) 1555 return 0; 1556 1557 // Validate the rest of constraint #1 by matching on the pred branch. 1558 TerminatorInst *TI = PredBB->getTerminator(); 1559 BasicBlock *TrueBB, *FalseBB; 1560 ICmpInst::Predicate Pred; 1561 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) 1562 return 0; 1563 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) 1564 return 0; 1565 1566 // Validate constraint #3: Ensure the null case just falls through. 1567 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) 1568 return 0; 1569 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && 1570 "Broken CFG: missing edge from predecessor to successor"); 1571 1572 FI.moveBefore(TI); 1573 return &FI; 1574} 1575 1576 1577Instruction *InstCombiner::visitFree(CallInst &FI) { 1578 Value *Op = FI.getArgOperand(0); 1579 1580 // free undef -> unreachable. 1581 if (isa<UndefValue>(Op)) { 1582 // Insert a new store to null because we cannot modify the CFG here. 1583 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()), 1584 UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); 1585 return EraseInstFromFunction(FI); 1586 } 1587 1588 // If we have 'free null' delete the instruction. This can happen in stl code 1589 // when lots of inlining happens. 1590 if (isa<ConstantPointerNull>(Op)) 1591 return EraseInstFromFunction(FI); 1592 1593 // If we optimize for code size, try to move the call to free before the null 1594 // test so that simplify cfg can remove the empty block and dead code 1595 // elimination the branch. I.e., helps to turn something like: 1596 // if (foo) free(foo); 1597 // into 1598 // free(foo); 1599 if (MinimizeSize) 1600 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) 1601 return I; 1602 1603 return 0; 1604} 1605 1606 1607 1608Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 1609 // Change br (not X), label True, label False to: br X, label False, True 1610 Value *X = 0; 1611 BasicBlock *TrueDest; 1612 BasicBlock *FalseDest; 1613 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && 1614 !isa<Constant>(X)) { 1615 // Swap Destinations and condition... 1616 BI.setCondition(X); 1617 BI.swapSuccessors(); 1618 return &BI; 1619 } 1620 1621 // Cannonicalize fcmp_one -> fcmp_oeq 1622 FCmpInst::Predicate FPred; Value *Y; 1623 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), 1624 TrueDest, FalseDest)) && 1625 BI.getCondition()->hasOneUse()) 1626 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || 1627 FPred == FCmpInst::FCMP_OGE) { 1628 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition()); 1629 Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); 1630 1631 // Swap Destinations and condition. 1632 BI.swapSuccessors(); 1633 Worklist.Add(Cond); 1634 return &BI; 1635 } 1636 1637 // Cannonicalize icmp_ne -> icmp_eq 1638 ICmpInst::Predicate IPred; 1639 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), 1640 TrueDest, FalseDest)) && 1641 BI.getCondition()->hasOneUse()) 1642 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE || 1643 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE || 1644 IPred == ICmpInst::ICMP_SGE) { 1645 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition()); 1646 Cond->setPredicate(ICmpInst::getInversePredicate(IPred)); 1647 // Swap Destinations and condition. 1648 BI.swapSuccessors(); 1649 Worklist.Add(Cond); 1650 return &BI; 1651 } 1652 1653 return 0; 1654} 1655 1656Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { 1657 Value *Cond = SI.getCondition(); 1658 if (Instruction *I = dyn_cast<Instruction>(Cond)) { 1659 if (I->getOpcode() == Instruction::Add) 1660 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 1661 // change 'switch (X+4) case 1:' into 'switch (X) case -3' 1662 // Skip the first item since that's the default case. 1663 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); 1664 i != e; ++i) { 1665 ConstantInt* CaseVal = i.getCaseValue(); 1666 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal), 1667 AddRHS); 1668 assert(isa<ConstantInt>(NewCaseVal) && 1669 "Result of expression should be constant"); 1670 i.setValue(cast<ConstantInt>(NewCaseVal)); 1671 } 1672 SI.setCondition(I->getOperand(0)); 1673 Worklist.Add(I); 1674 return &SI; 1675 } 1676 } 1677 return 0; 1678} 1679 1680Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { 1681 Value *Agg = EV.getAggregateOperand(); 1682 1683 if (!EV.hasIndices()) 1684 return ReplaceInstUsesWith(EV, Agg); 1685 1686 if (Constant *C = dyn_cast<Constant>(Agg)) { 1687 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) { 1688 if (EV.getNumIndices() == 0) 1689 return ReplaceInstUsesWith(EV, C2); 1690 // Extract the remaining indices out of the constant indexed by the 1691 // first index 1692 return ExtractValueInst::Create(C2, EV.getIndices().slice(1)); 1693 } 1694 return 0; // Can't handle other constants 1695 } 1696 1697 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) { 1698 // We're extracting from an insertvalue instruction, compare the indices 1699 const unsigned *exti, *exte, *insi, *inse; 1700 for (exti = EV.idx_begin(), insi = IV->idx_begin(), 1701 exte = EV.idx_end(), inse = IV->idx_end(); 1702 exti != exte && insi != inse; 1703 ++exti, ++insi) { 1704 if (*insi != *exti) 1705 // The insert and extract both reference distinctly different elements. 1706 // This means the extract is not influenced by the insert, and we can 1707 // replace the aggregate operand of the extract with the aggregate 1708 // operand of the insert. i.e., replace 1709 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1710 // %E = extractvalue { i32, { i32 } } %I, 0 1711 // with 1712 // %E = extractvalue { i32, { i32 } } %A, 0 1713 return ExtractValueInst::Create(IV->getAggregateOperand(), 1714 EV.getIndices()); 1715 } 1716 if (exti == exte && insi == inse) 1717 // Both iterators are at the end: Index lists are identical. Replace 1718 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1719 // %C = extractvalue { i32, { i32 } } %B, 1, 0 1720 // with "i32 42" 1721 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand()); 1722 if (exti == exte) { 1723 // The extract list is a prefix of the insert list. i.e. replace 1724 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0 1725 // %E = extractvalue { i32, { i32 } } %I, 1 1726 // with 1727 // %X = extractvalue { i32, { i32 } } %A, 1 1728 // %E = insertvalue { i32 } %X, i32 42, 0 1729 // by switching the order of the insert and extract (though the 1730 // insertvalue should be left in, since it may have other uses). 1731 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(), 1732 EV.getIndices()); 1733 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(), 1734 makeArrayRef(insi, inse)); 1735 } 1736 if (insi == inse) 1737 // The insert list is a prefix of the extract list 1738 // We can simply remove the common indices from the extract and make it 1739 // operate on the inserted value instead of the insertvalue result. 1740 // i.e., replace 1741 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1 1742 // %E = extractvalue { i32, { i32 } } %I, 1, 0 1743 // with 1744 // %E extractvalue { i32 } { i32 42 }, 0 1745 return ExtractValueInst::Create(IV->getInsertedValueOperand(), 1746 makeArrayRef(exti, exte)); 1747 } 1748 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) { 1749 // We're extracting from an intrinsic, see if we're the only user, which 1750 // allows us to simplify multiple result intrinsics to simpler things that 1751 // just get one value. 1752 if (II->hasOneUse()) { 1753 // Check if we're grabbing the overflow bit or the result of a 'with 1754 // overflow' intrinsic. If it's the latter we can remove the intrinsic 1755 // and replace it with a traditional binary instruction. 1756 switch (II->getIntrinsicID()) { 1757 case Intrinsic::uadd_with_overflow: 1758 case Intrinsic::sadd_with_overflow: 1759 if (*EV.idx_begin() == 0) { // Normal result. 1760 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1761 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1762 EraseInstFromFunction(*II); 1763 return BinaryOperator::CreateAdd(LHS, RHS); 1764 } 1765 1766 // If the normal result of the add is dead, and the RHS is a constant, 1767 // we can transform this into a range comparison. 1768 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 1769 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow) 1770 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1))) 1771 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0), 1772 ConstantExpr::getNot(CI)); 1773 break; 1774 case Intrinsic::usub_with_overflow: 1775 case Intrinsic::ssub_with_overflow: 1776 if (*EV.idx_begin() == 0) { // Normal result. 1777 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1778 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1779 EraseInstFromFunction(*II); 1780 return BinaryOperator::CreateSub(LHS, RHS); 1781 } 1782 break; 1783 case Intrinsic::umul_with_overflow: 1784 case Intrinsic::smul_with_overflow: 1785 if (*EV.idx_begin() == 0) { // Normal result. 1786 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 1787 ReplaceInstUsesWith(*II, UndefValue::get(II->getType())); 1788 EraseInstFromFunction(*II); 1789 return BinaryOperator::CreateMul(LHS, RHS); 1790 } 1791 break; 1792 default: 1793 break; 1794 } 1795 } 1796 } 1797 if (LoadInst *L = dyn_cast<LoadInst>(Agg)) 1798 // If the (non-volatile) load only has one use, we can rewrite this to a 1799 // load from a GEP. This reduces the size of the load. 1800 // FIXME: If a load is used only by extractvalue instructions then this 1801 // could be done regardless of having multiple uses. 1802 if (L->isSimple() && L->hasOneUse()) { 1803 // extractvalue has integer indices, getelementptr has Value*s. Convert. 1804 SmallVector<Value*, 4> Indices; 1805 // Prefix an i32 0 since we need the first element. 1806 Indices.push_back(Builder->getInt32(0)); 1807 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end(); 1808 I != E; ++I) 1809 Indices.push_back(Builder->getInt32(*I)); 1810 1811 // We need to insert these at the location of the old load, not at that of 1812 // the extractvalue. 1813 Builder->SetInsertPoint(L->getParent(), L); 1814 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices); 1815 // Returning the load directly will cause the main loop to insert it in 1816 // the wrong spot, so use ReplaceInstUsesWith(). 1817 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP)); 1818 } 1819 // We could simplify extracts from other values. Note that nested extracts may 1820 // already be simplified implicitly by the above: extract (extract (insert) ) 1821 // will be translated into extract ( insert ( extract ) ) first and then just 1822 // the value inserted, if appropriate. Similarly for extracts from single-use 1823 // loads: extract (extract (load)) will be translated to extract (load (gep)) 1824 // and if again single-use then via load (gep (gep)) to load (gep). 1825 // However, double extracts from e.g. function arguments or return values 1826 // aren't handled yet. 1827 return 0; 1828} 1829 1830enum Personality_Type { 1831 Unknown_Personality, 1832 GNU_Ada_Personality, 1833 GNU_CXX_Personality, 1834 GNU_ObjC_Personality 1835}; 1836 1837/// RecognizePersonality - See if the given exception handling personality 1838/// function is one that we understand. If so, return a description of it; 1839/// otherwise return Unknown_Personality. 1840static Personality_Type RecognizePersonality(Value *Pers) { 1841 Function *F = dyn_cast<Function>(Pers->stripPointerCasts()); 1842 if (!F) 1843 return Unknown_Personality; 1844 return StringSwitch<Personality_Type>(F->getName()) 1845 .Case("__gnat_eh_personality", GNU_Ada_Personality) 1846 .Case("__gxx_personality_v0", GNU_CXX_Personality) 1847 .Case("__objc_personality_v0", GNU_ObjC_Personality) 1848 .Default(Unknown_Personality); 1849} 1850 1851/// isCatchAll - Return 'true' if the given typeinfo will match anything. 1852static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) { 1853 switch (Personality) { 1854 case Unknown_Personality: 1855 return false; 1856 case GNU_Ada_Personality: 1857 // While __gnat_all_others_value will match any Ada exception, it doesn't 1858 // match foreign exceptions (or didn't, before gcc-4.7). 1859 return false; 1860 case GNU_CXX_Personality: 1861 case GNU_ObjC_Personality: 1862 return TypeInfo->isNullValue(); 1863 } 1864 llvm_unreachable("Unknown personality!"); 1865} 1866 1867static bool shorter_filter(const Value *LHS, const Value *RHS) { 1868 return 1869 cast<ArrayType>(LHS->getType())->getNumElements() 1870 < 1871 cast<ArrayType>(RHS->getType())->getNumElements(); 1872} 1873 1874Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { 1875 // The logic here should be correct for any real-world personality function. 1876 // However if that turns out not to be true, the offending logic can always 1877 // be conditioned on the personality function, like the catch-all logic is. 1878 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn()); 1879 1880 // Simplify the list of clauses, eg by removing repeated catch clauses 1881 // (these are often created by inlining). 1882 bool MakeNewInstruction = false; // If true, recreate using the following: 1883 SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction; 1884 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. 1885 1886 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already. 1887 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) { 1888 bool isLastClause = i + 1 == e; 1889 if (LI.isCatch(i)) { 1890 // A catch clause. 1891 Value *CatchClause = LI.getClause(i); 1892 Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts()); 1893 1894 // If we already saw this clause, there is no point in having a second 1895 // copy of it. 1896 if (AlreadyCaught.insert(TypeInfo)) { 1897 // This catch clause was not already seen. 1898 NewClauses.push_back(CatchClause); 1899 } else { 1900 // Repeated catch clause - drop the redundant copy. 1901 MakeNewInstruction = true; 1902 } 1903 1904 // If this is a catch-all then there is no point in keeping any following 1905 // clauses or marking the landingpad as having a cleanup. 1906 if (isCatchAll(Personality, TypeInfo)) { 1907 if (!isLastClause) 1908 MakeNewInstruction = true; 1909 CleanupFlag = false; 1910 break; 1911 } 1912 } else { 1913 // A filter clause. If any of the filter elements were already caught 1914 // then they can be dropped from the filter. It is tempting to try to 1915 // exploit the filter further by saying that any typeinfo that does not 1916 // occur in the filter can't be caught later (and thus can be dropped). 1917 // However this would be wrong, since typeinfos can match without being 1918 // equal (for example if one represents a C++ class, and the other some 1919 // class derived from it). 1920 assert(LI.isFilter(i) && "Unsupported landingpad clause!"); 1921 Value *FilterClause = LI.getClause(i); 1922 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType()); 1923 unsigned NumTypeInfos = FilterType->getNumElements(); 1924 1925 // An empty filter catches everything, so there is no point in keeping any 1926 // following clauses or marking the landingpad as having a cleanup. By 1927 // dealing with this case here the following code is made a bit simpler. 1928 if (!NumTypeInfos) { 1929 NewClauses.push_back(FilterClause); 1930 if (!isLastClause) 1931 MakeNewInstruction = true; 1932 CleanupFlag = false; 1933 break; 1934 } 1935 1936 bool MakeNewFilter = false; // If true, make a new filter. 1937 SmallVector<Constant *, 16> NewFilterElts; // New elements. 1938 if (isa<ConstantAggregateZero>(FilterClause)) { 1939 // Not an empty filter - it contains at least one null typeinfo. 1940 assert(NumTypeInfos > 0 && "Should have handled empty filter already!"); 1941 Constant *TypeInfo = 1942 Constant::getNullValue(FilterType->getElementType()); 1943 // If this typeinfo is a catch-all then the filter can never match. 1944 if (isCatchAll(Personality, TypeInfo)) { 1945 // Throw the filter away. 1946 MakeNewInstruction = true; 1947 continue; 1948 } 1949 1950 // There is no point in having multiple copies of this typeinfo, so 1951 // discard all but the first copy if there is more than one. 1952 NewFilterElts.push_back(TypeInfo); 1953 if (NumTypeInfos > 1) 1954 MakeNewFilter = true; 1955 } else { 1956 ConstantArray *Filter = cast<ConstantArray>(FilterClause); 1957 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements. 1958 NewFilterElts.reserve(NumTypeInfos); 1959 1960 // Remove any filter elements that were already caught or that already 1961 // occurred in the filter. While there, see if any of the elements are 1962 // catch-alls. If so, the filter can be discarded. 1963 bool SawCatchAll = false; 1964 for (unsigned j = 0; j != NumTypeInfos; ++j) { 1965 Value *Elt = Filter->getOperand(j); 1966 Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts()); 1967 if (isCatchAll(Personality, TypeInfo)) { 1968 // This element is a catch-all. Bail out, noting this fact. 1969 SawCatchAll = true; 1970 break; 1971 } 1972 if (AlreadyCaught.count(TypeInfo)) 1973 // Already caught by an earlier clause, so having it in the filter 1974 // is pointless. 1975 continue; 1976 // There is no point in having multiple copies of the same typeinfo in 1977 // a filter, so only add it if we didn't already. 1978 if (SeenInFilter.insert(TypeInfo)) 1979 NewFilterElts.push_back(cast<Constant>(Elt)); 1980 } 1981 // A filter containing a catch-all cannot match anything by definition. 1982 if (SawCatchAll) { 1983 // Throw the filter away. 1984 MakeNewInstruction = true; 1985 continue; 1986 } 1987 1988 // If we dropped something from the filter, make a new one. 1989 if (NewFilterElts.size() < NumTypeInfos) 1990 MakeNewFilter = true; 1991 } 1992 if (MakeNewFilter) { 1993 FilterType = ArrayType::get(FilterType->getElementType(), 1994 NewFilterElts.size()); 1995 FilterClause = ConstantArray::get(FilterType, NewFilterElts); 1996 MakeNewInstruction = true; 1997 } 1998 1999 NewClauses.push_back(FilterClause); 2000 2001 // If the new filter is empty then it will catch everything so there is 2002 // no point in keeping any following clauses or marking the landingpad 2003 // as having a cleanup. The case of the original filter being empty was 2004 // already handled above. 2005 if (MakeNewFilter && !NewFilterElts.size()) { 2006 assert(MakeNewInstruction && "New filter but not a new instruction!"); 2007 CleanupFlag = false; 2008 break; 2009 } 2010 } 2011 } 2012 2013 // If several filters occur in a row then reorder them so that the shortest 2014 // filters come first (those with the smallest number of elements). This is 2015 // advantageous because shorter filters are more likely to match, speeding up 2016 // unwinding, but mostly because it increases the effectiveness of the other 2017 // filter optimizations below. 2018 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) { 2019 unsigned j; 2020 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters. 2021 for (j = i; j != e; ++j) 2022 if (!isa<ArrayType>(NewClauses[j]->getType())) 2023 break; 2024 2025 // Check whether the filters are already sorted by length. We need to know 2026 // if sorting them is actually going to do anything so that we only make a 2027 // new landingpad instruction if it does. 2028 for (unsigned k = i; k + 1 < j; ++k) 2029 if (shorter_filter(NewClauses[k+1], NewClauses[k])) { 2030 // Not sorted, so sort the filters now. Doing an unstable sort would be 2031 // correct too but reordering filters pointlessly might confuse users. 2032 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j, 2033 shorter_filter); 2034 MakeNewInstruction = true; 2035 break; 2036 } 2037 2038 // Look for the next batch of filters. 2039 i = j + 1; 2040 } 2041 2042 // If typeinfos matched if and only if equal, then the elements of a filter L 2043 // that occurs later than a filter F could be replaced by the intersection of 2044 // the elements of F and L. In reality two typeinfos can match without being 2045 // equal (for example if one represents a C++ class, and the other some class 2046 // derived from it) so it would be wrong to perform this transform in general. 2047 // However the transform is correct and useful if F is a subset of L. In that 2048 // case L can be replaced by F, and thus removed altogether since repeating a 2049 // filter is pointless. So here we look at all pairs of filters F and L where 2050 // L follows F in the list of clauses, and remove L if every element of F is 2051 // an element of L. This can occur when inlining C++ functions with exception 2052 // specifications. 2053 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) { 2054 // Examine each filter in turn. 2055 Value *Filter = NewClauses[i]; 2056 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType()); 2057 if (!FTy) 2058 // Not a filter - skip it. 2059 continue; 2060 unsigned FElts = FTy->getNumElements(); 2061 // Examine each filter following this one. Doing this backwards means that 2062 // we don't have to worry about filters disappearing under us when removed. 2063 for (unsigned j = NewClauses.size() - 1; j != i; --j) { 2064 Value *LFilter = NewClauses[j]; 2065 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType()); 2066 if (!LTy) 2067 // Not a filter - skip it. 2068 continue; 2069 // If Filter is a subset of LFilter, i.e. every element of Filter is also 2070 // an element of LFilter, then discard LFilter. 2071 SmallVectorImpl<Value *>::iterator J = NewClauses.begin() + j; 2072 // If Filter is empty then it is a subset of LFilter. 2073 if (!FElts) { 2074 // Discard LFilter. 2075 NewClauses.erase(J); 2076 MakeNewInstruction = true; 2077 // Move on to the next filter. 2078 continue; 2079 } 2080 unsigned LElts = LTy->getNumElements(); 2081 // If Filter is longer than LFilter then it cannot be a subset of it. 2082 if (FElts > LElts) 2083 // Move on to the next filter. 2084 continue; 2085 // At this point we know that LFilter has at least one element. 2086 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros. 2087 // Filter is a subset of LFilter iff Filter contains only zeros (as we 2088 // already know that Filter is not longer than LFilter). 2089 if (isa<ConstantAggregateZero>(Filter)) { 2090 assert(FElts <= LElts && "Should have handled this case earlier!"); 2091 // Discard LFilter. 2092 NewClauses.erase(J); 2093 MakeNewInstruction = true; 2094 } 2095 // Move on to the next filter. 2096 continue; 2097 } 2098 ConstantArray *LArray = cast<ConstantArray>(LFilter); 2099 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros. 2100 // Since Filter is non-empty and contains only zeros, it is a subset of 2101 // LFilter iff LFilter contains a zero. 2102 assert(FElts > 0 && "Should have eliminated the empty filter earlier!"); 2103 for (unsigned l = 0; l != LElts; ++l) 2104 if (LArray->getOperand(l)->isNullValue()) { 2105 // LFilter contains a zero - discard it. 2106 NewClauses.erase(J); 2107 MakeNewInstruction = true; 2108 break; 2109 } 2110 // Move on to the next filter. 2111 continue; 2112 } 2113 // At this point we know that both filters are ConstantArrays. Loop over 2114 // operands to see whether every element of Filter is also an element of 2115 // LFilter. Since filters tend to be short this is probably faster than 2116 // using a method that scales nicely. 2117 ConstantArray *FArray = cast<ConstantArray>(Filter); 2118 bool AllFound = true; 2119 for (unsigned f = 0; f != FElts; ++f) { 2120 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts(); 2121 AllFound = false; 2122 for (unsigned l = 0; l != LElts; ++l) { 2123 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts(); 2124 if (LTypeInfo == FTypeInfo) { 2125 AllFound = true; 2126 break; 2127 } 2128 } 2129 if (!AllFound) 2130 break; 2131 } 2132 if (AllFound) { 2133 // Discard LFilter. 2134 NewClauses.erase(J); 2135 MakeNewInstruction = true; 2136 } 2137 // Move on to the next filter. 2138 } 2139 } 2140 2141 // If we changed any of the clauses, replace the old landingpad instruction 2142 // with a new one. 2143 if (MakeNewInstruction) { 2144 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(), 2145 LI.getPersonalityFn(), 2146 NewClauses.size()); 2147 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i) 2148 NLI->addClause(NewClauses[i]); 2149 // A landing pad with no clauses must have the cleanup flag set. It is 2150 // theoretically possible, though highly unlikely, that we eliminated all 2151 // clauses. If so, force the cleanup flag to true. 2152 if (NewClauses.empty()) 2153 CleanupFlag = true; 2154 NLI->setCleanup(CleanupFlag); 2155 return NLI; 2156 } 2157 2158 // Even if none of the clauses changed, we may nonetheless have understood 2159 // that the cleanup flag is pointless. Clear it if so. 2160 if (LI.isCleanup() != CleanupFlag) { 2161 assert(!CleanupFlag && "Adding a cleanup, not removing one?!"); 2162 LI.setCleanup(CleanupFlag); 2163 return &LI; 2164 } 2165 2166 return 0; 2167} 2168 2169 2170 2171 2172/// TryToSinkInstruction - Try to move the specified instruction from its 2173/// current block into the beginning of DestBlock, which can only happen if it's 2174/// safe to move the instruction past all of the instructions between it and the 2175/// end of its block. 2176static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { 2177 assert(I->hasOneUse() && "Invariants didn't hold!"); 2178 2179 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 2180 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() || 2181 isa<TerminatorInst>(I)) 2182 return false; 2183 2184 // Do not sink alloca instructions out of the entry block. 2185 if (isa<AllocaInst>(I) && I->getParent() == 2186 &DestBlock->getParent()->getEntryBlock()) 2187 return false; 2188 2189 // We can only sink load instructions if there is nothing between the load and 2190 // the end of block that could change the value. 2191 if (I->mayReadFromMemory()) { 2192 for (BasicBlock::iterator Scan = I, E = I->getParent()->end(); 2193 Scan != E; ++Scan) 2194 if (Scan->mayWriteToMemory()) 2195 return false; 2196 } 2197 2198 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt(); 2199 I->moveBefore(InsertPos); 2200 ++NumSunkInst; 2201 return true; 2202} 2203 2204 2205/// AddReachableCodeToWorklist - Walk the function in depth-first order, adding 2206/// all reachable code to the worklist. 2207/// 2208/// This has a couple of tricks to make the code faster and more powerful. In 2209/// particular, we constant fold and DCE instructions as we go, to avoid adding 2210/// them to the worklist (this significantly speeds up instcombine on code where 2211/// many instructions are dead or constant). Additionally, if we find a branch 2212/// whose condition is a known constant, we only visit the reachable successors. 2213/// 2214static bool AddReachableCodeToWorklist(BasicBlock *BB, 2215 SmallPtrSet<BasicBlock*, 64> &Visited, 2216 InstCombiner &IC, 2217 const DataLayout *TD, 2218 const TargetLibraryInfo *TLI) { 2219 bool MadeIRChange = false; 2220 SmallVector<BasicBlock*, 256> Worklist; 2221 Worklist.push_back(BB); 2222 2223 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist; 2224 DenseMap<ConstantExpr*, Constant*> FoldedConstants; 2225 2226 do { 2227 BB = Worklist.pop_back_val(); 2228 2229 // We have now visited this block! If we've already been here, ignore it. 2230 if (!Visited.insert(BB)) continue; 2231 2232 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 2233 Instruction *Inst = BBI++; 2234 2235 // DCE instruction if trivially dead. 2236 if (isInstructionTriviallyDead(Inst, TLI)) { 2237 ++NumDeadInst; 2238 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); 2239 Inst->eraseFromParent(); 2240 continue; 2241 } 2242 2243 // ConstantProp instruction if trivially constant. 2244 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0))) 2245 if (Constant *C = ConstantFoldInstruction(Inst, TD, TLI)) { 2246 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " 2247 << *Inst << '\n'); 2248 Inst->replaceAllUsesWith(C); 2249 ++NumConstProp; 2250 Inst->eraseFromParent(); 2251 continue; 2252 } 2253 2254 if (TD) { 2255 // See if we can constant fold its operands. 2256 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); 2257 i != e; ++i) { 2258 ConstantExpr *CE = dyn_cast<ConstantExpr>(i); 2259 if (CE == 0) continue; 2260 2261 Constant*& FoldRes = FoldedConstants[CE]; 2262 if (!FoldRes) 2263 FoldRes = ConstantFoldConstantExpression(CE, TD, TLI); 2264 if (!FoldRes) 2265 FoldRes = CE; 2266 2267 if (FoldRes != CE) { 2268 *i = FoldRes; 2269 MadeIRChange = true; 2270 } 2271 } 2272 } 2273 2274 InstrsForInstCombineWorklist.push_back(Inst); 2275 } 2276 2277 // Recursively visit successors. If this is a branch or switch on a 2278 // constant, only visit the reachable successor. 2279 TerminatorInst *TI = BB->getTerminator(); 2280 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2281 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) { 2282 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue(); 2283 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal); 2284 Worklist.push_back(ReachableBB); 2285 continue; 2286 } 2287 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2288 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) { 2289 // See if this is an explicit destination. 2290 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2291 i != e; ++i) 2292 if (i.getCaseValue() == Cond) { 2293 BasicBlock *ReachableBB = i.getCaseSuccessor(); 2294 Worklist.push_back(ReachableBB); 2295 continue; 2296 } 2297 2298 // Otherwise it is the default destination. 2299 Worklist.push_back(SI->getDefaultDest()); 2300 continue; 2301 } 2302 } 2303 2304 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 2305 Worklist.push_back(TI->getSuccessor(i)); 2306 } while (!Worklist.empty()); 2307 2308 // Once we've found all of the instructions to add to instcombine's worklist, 2309 // add them in reverse order. This way instcombine will visit from the top 2310 // of the function down. This jives well with the way that it adds all uses 2311 // of instructions to the worklist after doing a transformation, thus avoiding 2312 // some N^2 behavior in pathological cases. 2313 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], 2314 InstrsForInstCombineWorklist.size()); 2315 2316 return MadeIRChange; 2317} 2318 2319bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 2320 MadeIRChange = false; 2321 2322 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " 2323 << F.getName() << "\n"); 2324 2325 { 2326 // Do a depth-first traversal of the function, populate the worklist with 2327 // the reachable instructions. Ignore blocks that are not reachable. Keep 2328 // track of which blocks we visit. 2329 SmallPtrSet<BasicBlock*, 64> Visited; 2330 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD, 2331 TLI); 2332 2333 // Do a quick scan over the function. If we find any blocks that are 2334 // unreachable, remove any instructions inside of them. This prevents 2335 // the instcombine code from having to deal with some bad special cases. 2336 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 2337 if (Visited.count(BB)) continue; 2338 2339 // Delete the instructions backwards, as it has a reduced likelihood of 2340 // having to update as many def-use and use-def chains. 2341 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 2342 while (EndInst != BB->begin()) { 2343 // Delete the next to last instruction. 2344 BasicBlock::iterator I = EndInst; 2345 Instruction *Inst = --I; 2346 if (!Inst->use_empty()) 2347 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 2348 if (isa<LandingPadInst>(Inst)) { 2349 EndInst = Inst; 2350 continue; 2351 } 2352 if (!isa<DbgInfoIntrinsic>(Inst)) { 2353 ++NumDeadInst; 2354 MadeIRChange = true; 2355 } 2356 Inst->eraseFromParent(); 2357 } 2358 } 2359 } 2360 2361 while (!Worklist.isEmpty()) { 2362 Instruction *I = Worklist.RemoveOne(); 2363 if (I == 0) continue; // skip null values. 2364 2365 // Check to see if we can DCE the instruction. 2366 if (isInstructionTriviallyDead(I, TLI)) { 2367 DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); 2368 EraseInstFromFunction(*I); 2369 ++NumDeadInst; 2370 MadeIRChange = true; 2371 continue; 2372 } 2373 2374 // Instruction isn't dead, see if we can constant propagate it. 2375 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) 2376 if (Constant *C = ConstantFoldInstruction(I, TD, TLI)) { 2377 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); 2378 2379 // Add operands to the worklist. 2380 ReplaceInstUsesWith(*I, C); 2381 ++NumConstProp; 2382 EraseInstFromFunction(*I); 2383 MadeIRChange = true; 2384 continue; 2385 } 2386 2387 // See if we can trivially sink this instruction to a successor basic block. 2388 if (I->hasOneUse()) { 2389 BasicBlock *BB = I->getParent(); 2390 Instruction *UserInst = cast<Instruction>(I->use_back()); 2391 BasicBlock *UserParent; 2392 2393 // Get the block the use occurs in. 2394 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) 2395 UserParent = PN->getIncomingBlock(I->use_begin().getUse()); 2396 else 2397 UserParent = UserInst->getParent(); 2398 2399 if (UserParent != BB) { 2400 bool UserIsSuccessor = false; 2401 // See if the user is one of our successors. 2402 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 2403 if (*SI == UserParent) { 2404 UserIsSuccessor = true; 2405 break; 2406 } 2407 2408 // If the user is one of our immediate successors, and if that successor 2409 // only has us as a predecessors (we'd have to split the critical edge 2410 // otherwise), we can keep going. 2411 if (UserIsSuccessor && UserParent->getSinglePredecessor()) 2412 // Okay, the CFG is simple enough, try to sink this instruction. 2413 MadeIRChange |= TryToSinkInstruction(I, UserParent); 2414 } 2415 } 2416 2417 // Now that we have an instruction, try combining it to simplify it. 2418 Builder->SetInsertPoint(I->getParent(), I); 2419 Builder->SetCurrentDebugLocation(I->getDebugLoc()); 2420 2421#ifndef NDEBUG 2422 std::string OrigI; 2423#endif 2424 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); 2425 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); 2426 2427 if (Instruction *Result = visit(*I)) { 2428 ++NumCombined; 2429 // Should we replace the old instruction with a new one? 2430 if (Result != I) { 2431 DEBUG(dbgs() << "IC: Old = " << *I << '\n' 2432 << " New = " << *Result << '\n'); 2433 2434 if (!I->getDebugLoc().isUnknown()) 2435 Result->setDebugLoc(I->getDebugLoc()); 2436 // Everything uses the new instruction now. 2437 I->replaceAllUsesWith(Result); 2438 2439 // Move the name to the new instruction first. 2440 Result->takeName(I); 2441 2442 // Push the new instruction and any users onto the worklist. 2443 Worklist.Add(Result); 2444 Worklist.AddUsersToWorkList(*Result); 2445 2446 // Insert the new instruction into the basic block... 2447 BasicBlock *InstParent = I->getParent(); 2448 BasicBlock::iterator InsertPos = I; 2449 2450 // If we replace a PHI with something that isn't a PHI, fix up the 2451 // insertion point. 2452 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos)) 2453 InsertPos = InstParent->getFirstInsertionPt(); 2454 2455 InstParent->getInstList().insert(InsertPos, Result); 2456 2457 EraseInstFromFunction(*I); 2458 } else { 2459#ifndef NDEBUG 2460 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' 2461 << " New = " << *I << '\n'); 2462#endif 2463 2464 // If the instruction was modified, it's possible that it is now dead. 2465 // if so, remove it. 2466 if (isInstructionTriviallyDead(I, TLI)) { 2467 EraseInstFromFunction(*I); 2468 } else { 2469 Worklist.Add(I); 2470 Worklist.AddUsersToWorkList(*I); 2471 } 2472 } 2473 MadeIRChange = true; 2474 } 2475 } 2476 2477 Worklist.Zap(); 2478 return MadeIRChange; 2479} 2480 2481namespace { 2482class InstCombinerLibCallSimplifier : public LibCallSimplifier { 2483 InstCombiner *IC; 2484public: 2485 InstCombinerLibCallSimplifier(const DataLayout *TD, 2486 const TargetLibraryInfo *TLI, 2487 InstCombiner *IC) 2488 : LibCallSimplifier(TD, TLI, UnsafeFPShrink) { 2489 this->IC = IC; 2490 } 2491 2492 /// replaceAllUsesWith - override so that instruction replacement 2493 /// can be defined in terms of the instruction combiner framework. 2494 virtual void replaceAllUsesWith(Instruction *I, Value *With) const { 2495 IC->ReplaceInstUsesWith(*I, With); 2496 } 2497}; 2498} 2499 2500bool InstCombiner::runOnFunction(Function &F) { 2501 TD = getAnalysisIfAvailable<DataLayout>(); 2502 TLI = &getAnalysis<TargetLibraryInfo>(); 2503 // Minimizing size? 2504 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 2505 Attribute::MinSize); 2506 2507 /// Builder - This is an IRBuilder that automatically inserts new 2508 /// instructions into the worklist when they are created. 2509 IRBuilder<true, TargetFolder, InstCombineIRInserter> 2510 TheBuilder(F.getContext(), TargetFolder(TD), 2511 InstCombineIRInserter(Worklist)); 2512 Builder = &TheBuilder; 2513 2514 InstCombinerLibCallSimplifier TheSimplifier(TD, TLI, this); 2515 Simplifier = &TheSimplifier; 2516 2517 bool EverMadeChange = false; 2518 2519 // Lower dbg.declare intrinsics otherwise their value may be clobbered 2520 // by instcombiner. 2521 EverMadeChange = LowerDbgDeclare(F); 2522 2523 // Iterate while there is work to do. 2524 unsigned Iteration = 0; 2525 while (DoOneIteration(F, Iteration++)) 2526 EverMadeChange = true; 2527 2528 Builder = 0; 2529 return EverMadeChange; 2530} 2531 2532FunctionPass *llvm::createInstructionCombiningPass() { 2533 return new InstCombiner(); 2534} 2535