InstCombineAddSub.cpp revision 4ee576fac3a84553c9342faea87ff0e13e8eb48d
1//===- InstCombineAddSub.cpp ----------------------------------------------===// 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// This file implements the visit functions for add, fadd, sub, and fsub. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Analysis/InstructionSimplify.h" 16#include "llvm/IR/DataLayout.h" 17#include "llvm/Support/GetElementPtrTypeIterator.h" 18#include "llvm/Support/PatternMatch.h" 19using namespace llvm; 20using namespace PatternMatch; 21 22namespace { 23 24 /// Class representing coefficient of floating-point addend. 25 /// This class needs to be highly efficient, which is especially true for 26 /// the constructor. As of I write this comment, the cost of the default 27 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to 28 /// perform write-merging). 29 /// 30 class FAddendCoef { 31 public: 32 // The constructor has to initialize a APFloat, which is uncessary for 33 // most addends which have coefficient either 1 or -1. So, the constructor 34 // is expensive. In order to avoid the cost of the constructor, we should 35 // reuse some instances whenever possible. The pre-created instances 36 // FAddCombine::Add[0-5] embodies this idea. 37 // 38 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {} 39 ~FAddendCoef(); 40 41 void set(short C) { 42 assert(!insaneIntVal(C) && "Insane coefficient"); 43 IsFp = false; IntVal = C; 44 } 45 46 void set(const APFloat& C); 47 48 void negate(); 49 50 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); } 51 Value *getValue(Type *) const; 52 53 // If possible, don't define operator+/operator- etc because these 54 // operators inevitably call FAddendCoef's constructor which is not cheap. 55 void operator=(const FAddendCoef &A); 56 void operator+=(const FAddendCoef &A); 57 void operator-=(const FAddendCoef &A); 58 void operator*=(const FAddendCoef &S); 59 60 bool isOne() const { return isInt() && IntVal == 1; } 61 bool isTwo() const { return isInt() && IntVal == 2; } 62 bool isMinusOne() const { return isInt() && IntVal == -1; } 63 bool isMinusTwo() const { return isInt() && IntVal == -2; } 64 65 private: 66 bool insaneIntVal(int V) { return V > 4 || V < -4; } 67 APFloat *getFpValPtr(void) 68 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); } 69 const APFloat *getFpValPtr(void) const 70 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); } 71 72 const APFloat &getFpVal(void) const { 73 assert(IsFp && BufHasFpVal && "Incorret state"); 74 return *getFpValPtr(); 75 } 76 77 APFloat &getFpVal(void) 78 { assert(IsFp && BufHasFpVal && "Incorret state"); return *getFpValPtr(); } 79 80 bool isInt() const { return !IsFp; } 81 82 private: 83 84 bool IsFp; 85 86 // True iff FpValBuf contains an instance of APFloat. 87 bool BufHasFpVal; 88 89 // The integer coefficient of an individual addend is either 1 or -1, 90 // and we try to simplify at most 4 addends from neighboring at most 91 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt 92 // is overkill of this end. 93 short IntVal; 94 95 AlignedCharArrayUnion<APFloat> FpValBuf; 96 }; 97 98 /// FAddend is used to represent floating-point addend. An addend is 99 /// represented as <C, V>, where the V is a symbolic value, and C is a 100 /// constant coefficient. A constant addend is represented as <C, 0>. 101 /// 102 class FAddend { 103 public: 104 FAddend() { Val = 0; } 105 106 Value *getSymVal (void) const { return Val; } 107 const FAddendCoef &getCoef(void) const { return Coeff; } 108 109 bool isConstant() const { return Val == 0; } 110 bool isZero() const { return Coeff.isZero(); } 111 112 void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; } 113 void set(const APFloat& Coefficient, Value *V) 114 { Coeff.set(Coefficient); Val = V; } 115 void set(const ConstantFP* Coefficient, Value *V) 116 { Coeff.set(Coefficient->getValueAPF()); Val = V; } 117 118 void negate() { Coeff.negate(); } 119 120 /// Drill down the U-D chain one step to find the definition of V, and 121 /// try to break the definition into one or two addends. 122 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1); 123 124 /// Similar to FAddend::drillDownOneStep() except that the value being 125 /// splitted is the addend itself. 126 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const; 127 128 void operator+=(const FAddend &T) { 129 assert((Val == T.Val) && "Symbolic-values disagree"); 130 Coeff += T.Coeff; 131 } 132 133 private: 134 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; } 135 136 // This addend has the value of "Coeff * Val". 137 Value *Val; 138 FAddendCoef Coeff; 139 }; 140 141 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along 142 /// with its neighboring at most two instructions. 143 /// 144 class FAddCombine { 145 public: 146 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(0) {} 147 Value *simplify(Instruction *FAdd); 148 149 private: 150 typedef SmallVector<const FAddend*, 4> AddendVect; 151 152 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota); 153 154 /// Convert given addend to a Value 155 Value *createAddendVal(const FAddend &A, bool& NeedNeg); 156 157 /// Return the number of instructions needed to emit the N-ary addition. 158 unsigned calcInstrNumber(const AddendVect& Vect); 159 Value *createFSub(Value *Opnd0, Value *Opnd1); 160 Value *createFAdd(Value *Opnd0, Value *Opnd1); 161 Value *createFMul(Value *Opnd0, Value *Opnd1); 162 Value *createFNeg(Value *V); 163 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota); 164 void createInstPostProc(Instruction *NewInst); 165 166 InstCombiner::BuilderTy *Builder; 167 Instruction *Instr; 168 169 private: 170 // Debugging stuff are clustered here. 171 #ifndef NDEBUG 172 unsigned CreateInstrNum; 173 void initCreateInstNum() { CreateInstrNum = 0; } 174 void incCreateInstNum() { CreateInstrNum++; } 175 #else 176 void initCreateInstNum() {} 177 void incCreateInstNum() {} 178 #endif 179 }; 180} 181 182//===----------------------------------------------------------------------===// 183// 184// Implementation of 185// {FAddendCoef, FAddend, FAddition, FAddCombine}. 186// 187//===----------------------------------------------------------------------===// 188FAddendCoef::~FAddendCoef() { 189 if (BufHasFpVal) 190 getFpValPtr()->~APFloat(); 191} 192 193void FAddendCoef::set(const APFloat& C) { 194 APFloat *P = getFpValPtr(); 195 196 if (isInt()) { 197 // As the buffer is meanless byte stream, we cannot call 198 // APFloat::operator=(). 199 new(P) APFloat(C); 200 } else 201 *P = C; 202 203 IsFp = BufHasFpVal = true; 204} 205 206void FAddendCoef::operator=(const FAddendCoef& That) { 207 if (That.isInt()) 208 set(That.IntVal); 209 else 210 set(That.getFpVal()); 211} 212 213void FAddendCoef::operator+=(const FAddendCoef &That) { 214 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 215 if (isInt() == That.isInt()) { 216 if (isInt()) 217 IntVal += That.IntVal; 218 else 219 getFpVal().add(That.getFpVal(), RndMode); 220 return; 221 } 222 223 if (isInt()) { 224 const APFloat &T = That.getFpVal(); 225 set(T); 226 getFpVal().add(APFloat(T.getSemantics(), IntVal), RndMode); 227 return; 228 } 229 230 APFloat &T = getFpVal(); 231 T.add(APFloat(T.getSemantics(), That.IntVal), RndMode); 232} 233 234void FAddendCoef::operator-=(const FAddendCoef &That) { 235 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 236 if (isInt() == That.isInt()) { 237 if (isInt()) 238 IntVal -= That.IntVal; 239 else 240 getFpVal().subtract(That.getFpVal(), RndMode); 241 return; 242 } 243 244 if (isInt()) { 245 const APFloat &T = That.getFpVal(); 246 set(T); 247 getFpVal().subtract(APFloat(T.getSemantics(), IntVal), RndMode); 248 return; 249 } 250 251 APFloat &T = getFpVal(); 252 T.subtract(APFloat(T.getSemantics(), IntVal), RndMode); 253} 254 255void FAddendCoef::operator*=(const FAddendCoef &That) { 256 if (That.isOne()) 257 return; 258 259 if (That.isMinusOne()) { 260 negate(); 261 return; 262 } 263 264 if (isInt() && That.isInt()) { 265 int Res = IntVal * (int)That.IntVal; 266 assert(!insaneIntVal(Res) && "Insane int value"); 267 IntVal = Res; 268 return; 269 } 270 271 const fltSemantics &Semantic = 272 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics(); 273 274 if (isInt()) 275 set(APFloat(Semantic, IntVal)); 276 APFloat &F0 = getFpVal(); 277 278 if (That.isInt()) 279 F0.multiply(APFloat(Semantic, That.IntVal), APFloat::rmNearestTiesToEven); 280 else 281 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven); 282 283 return; 284} 285 286void FAddendCoef::negate() { 287 if (isInt()) 288 IntVal = 0 - IntVal; 289 else 290 getFpVal().changeSign(); 291} 292 293Value *FAddendCoef::getValue(Type *Ty) const { 294 return isInt() ? 295 ConstantFP::get(Ty, float(IntVal)) : 296 ConstantFP::get(Ty->getContext(), getFpVal()); 297} 298 299// The definition of <Val> Addends 300// ========================================= 301// A + B <1, A>, <1,B> 302// A - B <1, A>, <1,B> 303// 0 - B <-1, B> 304// C * A, <C, A> 305// A + C <1, A> <C, NULL> 306// 0 +/- 0 <0, NULL> (corner case) 307// 308// Legend: A and B are not constant, C is constant 309// 310unsigned FAddend::drillValueDownOneStep 311 (Value *Val, FAddend &Addend0, FAddend &Addend1) { 312 Instruction *I = 0; 313 if (Val == 0 || !(I = dyn_cast<Instruction>(Val))) 314 return 0; 315 316 unsigned Opcode = I->getOpcode(); 317 318 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) { 319 ConstantFP *C0, *C1; 320 Value *Opnd0 = I->getOperand(0); 321 Value *Opnd1 = I->getOperand(1); 322 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero()) 323 Opnd0 = 0; 324 325 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero()) 326 Opnd1 = 0; 327 328 if (Opnd0) { 329 if (!C0) 330 Addend0.set(1, Opnd0); 331 else 332 Addend0.set(C0, 0); 333 } 334 335 if (Opnd1) { 336 FAddend &Addend = Opnd0 ? Addend1 : Addend0; 337 if (!C1) 338 Addend.set(1, Opnd1); 339 else 340 Addend.set(C1, 0); 341 if (Opcode == Instruction::FSub) 342 Addend.negate(); 343 } 344 345 if (Opnd0 || Opnd1) 346 return Opnd0 && Opnd1 ? 2 : 1; 347 348 // Both operands are zero. Weird! 349 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), 0); 350 return 1; 351 } 352 353 if (I->getOpcode() == Instruction::FMul) { 354 Value *V0 = I->getOperand(0); 355 Value *V1 = I->getOperand(1); 356 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) { 357 Addend0.set(C, V1); 358 return 1; 359 } 360 361 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) { 362 Addend0.set(C, V0); 363 return 1; 364 } 365 } 366 367 return 0; 368} 369 370// Try to break *this* addend into two addends. e.g. Suppose this addend is 371// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends, 372// i.e. <2.3, X> and <2.3, Y>. 373// 374unsigned FAddend::drillAddendDownOneStep 375 (FAddend &Addend0, FAddend &Addend1) const { 376 if (isConstant()) 377 return 0; 378 379 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1); 380 if (!BreakNum || Coeff.isOne()) 381 return BreakNum; 382 383 Addend0.Scale(Coeff); 384 385 if (BreakNum == 2) 386 Addend1.Scale(Coeff); 387 388 return BreakNum; 389} 390 391Value *FAddCombine::simplify(Instruction *I) { 392 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode"); 393 394 // Currently we are not able to handle vector type. 395 if (I->getType()->isVectorTy()) 396 return 0; 397 398 assert((I->getOpcode() == Instruction::FAdd || 399 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 400 401 // Save the instruction before calling other member-functions. 402 Instr = I; 403 404 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1; 405 406 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1); 407 408 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1. 409 unsigned Opnd0_ExpNum = 0; 410 unsigned Opnd1_ExpNum = 0; 411 412 if (!Opnd0.isConstant()) 413 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1); 414 415 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1. 416 if (OpndNum == 2 && !Opnd1.isConstant()) 417 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1); 418 419 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1 420 if (Opnd0_ExpNum && Opnd1_ExpNum) { 421 AddendVect AllOpnds; 422 AllOpnds.push_back(&Opnd0_0); 423 AllOpnds.push_back(&Opnd1_0); 424 if (Opnd0_ExpNum == 2) 425 AllOpnds.push_back(&Opnd0_1); 426 if (Opnd1_ExpNum == 2) 427 AllOpnds.push_back(&Opnd1_1); 428 429 // Compute instruction quota. We should save at least one instruction. 430 unsigned InstQuota = 0; 431 432 Value *V0 = I->getOperand(0); 433 Value *V1 = I->getOperand(1); 434 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) && 435 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1; 436 437 if (Value *R = simplifyFAdd(AllOpnds, InstQuota)) 438 return R; 439 } 440 441 if (OpndNum != 2) { 442 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be 443 // splitted into two addends, say "V = X - Y", the instruction would have 444 // been optimized into "I = Y - X" in the previous steps. 445 // 446 const FAddendCoef &CE = Opnd0.getCoef(); 447 return CE.isOne() ? Opnd0.getSymVal() : 0; 448 } 449 450 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1] 451 if (Opnd1_ExpNum) { 452 AddendVect AllOpnds; 453 AllOpnds.push_back(&Opnd0); 454 AllOpnds.push_back(&Opnd1_0); 455 if (Opnd1_ExpNum == 2) 456 AllOpnds.push_back(&Opnd1_1); 457 458 if (Value *R = simplifyFAdd(AllOpnds, 1)) 459 return R; 460 } 461 462 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1] 463 if (Opnd0_ExpNum) { 464 AddendVect AllOpnds; 465 AllOpnds.push_back(&Opnd1); 466 AllOpnds.push_back(&Opnd0_0); 467 if (Opnd0_ExpNum == 2) 468 AllOpnds.push_back(&Opnd0_1); 469 470 if (Value *R = simplifyFAdd(AllOpnds, 1)) 471 return R; 472 } 473 474 return 0; 475} 476 477Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) { 478 479 unsigned AddendNum = Addends.size(); 480 assert(AddendNum <= 4 && "Too many addends"); 481 482 // For saving intermediate results; 483 unsigned NextTmpIdx = 0; 484 FAddend TmpResult[3]; 485 486 // Points to the constant addend of the resulting simplified expression. 487 // If the resulting expr has constant-addend, this constant-addend is 488 // desirable to reside at the top of the resulting expression tree. Placing 489 // constant close to supper-expr(s) will potentially reveal some optimization 490 // opportunities in super-expr(s). 491 // 492 const FAddend *ConstAdd = 0; 493 494 // Simplified addends are placed <SimpVect>. 495 AddendVect SimpVect; 496 497 // The outer loop works on one symbolic-value at a time. Suppose the input 498 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ... 499 // The symbolic-values will be processed in this order: x, y, z. 500 // 501 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) { 502 503 const FAddend *ThisAddend = Addends[SymIdx]; 504 if (!ThisAddend) { 505 // This addend was processed before. 506 continue; 507 } 508 509 Value *Val = ThisAddend->getSymVal(); 510 unsigned StartIdx = SimpVect.size(); 511 SimpVect.push_back(ThisAddend); 512 513 // The inner loop collects addends sharing same symbolic-value, and these 514 // addends will be later on folded into a single addend. Following above 515 // example, if the symbolic value "y" is being processed, the inner loop 516 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will 517 // be later on folded into "<b1+b2, y>". 518 // 519 for (unsigned SameSymIdx = SymIdx + 1; 520 SameSymIdx < AddendNum; SameSymIdx++) { 521 const FAddend *T = Addends[SameSymIdx]; 522 if (T && T->getSymVal() == Val) { 523 // Set null such that next iteration of the outer loop will not process 524 // this addend again. 525 Addends[SameSymIdx] = 0; 526 SimpVect.push_back(T); 527 } 528 } 529 530 // If multiple addends share same symbolic value, fold them together. 531 if (StartIdx + 1 != SimpVect.size()) { 532 FAddend &R = TmpResult[NextTmpIdx ++]; 533 R = *SimpVect[StartIdx]; 534 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++) 535 R += *SimpVect[Idx]; 536 537 // Pop all addends being folded and push the resulting folded addend. 538 SimpVect.resize(StartIdx); 539 if (Val != 0) { 540 if (!R.isZero()) { 541 SimpVect.push_back(&R); 542 } 543 } else { 544 // Don't push constant addend at this time. It will be the last element 545 // of <SimpVect>. 546 ConstAdd = &R; 547 } 548 } 549 } 550 551 assert((NextTmpIdx <= sizeof(TmpResult)/sizeof(TmpResult[0]) + 1) && 552 "out-of-bound access"); 553 554 if (ConstAdd) 555 SimpVect.push_back(ConstAdd); 556 557 Value *Result; 558 if (!SimpVect.empty()) 559 Result = createNaryFAdd(SimpVect, InstrQuota); 560 else { 561 // The addition is folded to 0.0. 562 Result = ConstantFP::get(Instr->getType(), 0.0); 563 } 564 565 return Result; 566} 567 568Value *FAddCombine::createNaryFAdd 569 (const AddendVect &Opnds, unsigned InstrQuota) { 570 assert(!Opnds.empty() && "Expect at least one addend"); 571 572 // Step 1: Check if the # of instructions needed exceeds the quota. 573 // 574 unsigned InstrNeeded = calcInstrNumber(Opnds); 575 if (InstrNeeded > InstrQuota) 576 return 0; 577 578 initCreateInstNum(); 579 580 // step 2: Emit the N-ary addition. 581 // Note that at most three instructions are involved in Fadd-InstCombine: the 582 // addition in question, and at most two neighboring instructions. 583 // The resulting optimized addition should have at least one less instruction 584 // than the original addition expression tree. This implies that the resulting 585 // N-ary addition has at most two instructions, and we don't need to worry 586 // about tree-height when constructing the N-ary addition. 587 588 Value *LastVal = 0; 589 bool LastValNeedNeg = false; 590 591 // Iterate the addends, creating fadd/fsub using adjacent two addends. 592 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 593 I != E; I++) { 594 bool NeedNeg; 595 Value *V = createAddendVal(**I, NeedNeg); 596 if (!LastVal) { 597 LastVal = V; 598 LastValNeedNeg = NeedNeg; 599 continue; 600 } 601 602 if (LastValNeedNeg == NeedNeg) { 603 LastVal = createFAdd(LastVal, V); 604 continue; 605 } 606 607 if (LastValNeedNeg) 608 LastVal = createFSub(V, LastVal); 609 else 610 LastVal = createFSub(LastVal, V); 611 612 LastValNeedNeg = false; 613 } 614 615 if (LastValNeedNeg) { 616 LastVal = createFNeg(LastVal); 617 } 618 619 #ifndef NDEBUG 620 assert(CreateInstrNum == InstrNeeded && 621 "Inconsistent in instruction numbers"); 622 #endif 623 624 return LastVal; 625} 626 627Value *FAddCombine::createFSub 628 (Value *Opnd0, Value *Opnd1) { 629 Value *V = Builder->CreateFSub(Opnd0, Opnd1); 630 createInstPostProc(cast<Instruction>(V)); 631 return V; 632} 633 634Value *FAddCombine::createFNeg(Value *V) { 635 Value *Zero = cast<Value>(ConstantFP::get(V->getType(), 0.0)); 636 return createFSub(Zero, V); 637} 638 639Value *FAddCombine::createFAdd 640 (Value *Opnd0, Value *Opnd1) { 641 Value *V = Builder->CreateFAdd(Opnd0, Opnd1); 642 createInstPostProc(cast<Instruction>(V)); 643 return V; 644} 645 646Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) { 647 Value *V = Builder->CreateFMul(Opnd0, Opnd1); 648 createInstPostProc(cast<Instruction>(V)); 649 return V; 650} 651 652void FAddCombine::createInstPostProc(Instruction *NewInstr) { 653 NewInstr->setDebugLoc(Instr->getDebugLoc()); 654 655 // Keep track of the number of instruction created. 656 incCreateInstNum(); 657 658 // Propagate fast-math flags 659 NewInstr->setFastMathFlags(Instr->getFastMathFlags()); 660} 661 662// Return the number of instruction needed to emit the N-ary addition. 663// NOTE: Keep this function in sync with createAddendVal(). 664unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) { 665 unsigned OpndNum = Opnds.size(); 666 unsigned InstrNeeded = OpndNum - 1; 667 668 // The number of addends in the form of "(-1)*x". 669 unsigned NegOpndNum = 0; 670 671 // Adjust the number of instructions needed to emit the N-ary add. 672 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 673 I != E; I++) { 674 const FAddend *Opnd = *I; 675 if (Opnd->isConstant()) 676 continue; 677 678 const FAddendCoef &CE = Opnd->getCoef(); 679 if (CE.isMinusOne() || CE.isMinusTwo()) 680 NegOpndNum++; 681 682 // Let the addend be "c * x". If "c == +/-1", the value of the addend 683 // is immediately available; otherwise, it needs exactly one instruction 684 // to evaluate the value. 685 if (!CE.isMinusOne() && !CE.isOne()) 686 InstrNeeded++; 687 } 688 if (NegOpndNum == OpndNum) 689 InstrNeeded++; 690 return InstrNeeded; 691} 692 693// Input Addend Value NeedNeg(output) 694// ================================================================ 695// Constant C C false 696// <+/-1, V> V coefficient is -1 697// <2/-2, V> "fadd V, V" coefficient is -2 698// <C, V> "fmul V, C" false 699// 700// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber. 701Value *FAddCombine::createAddendVal 702 (const FAddend &Opnd, bool &NeedNeg) { 703 const FAddendCoef &Coeff = Opnd.getCoef(); 704 705 if (Opnd.isConstant()) { 706 NeedNeg = false; 707 return Coeff.getValue(Instr->getType()); 708 } 709 710 Value *OpndVal = Opnd.getSymVal(); 711 712 if (Coeff.isMinusOne() || Coeff.isOne()) { 713 NeedNeg = Coeff.isMinusOne(); 714 return OpndVal; 715 } 716 717 if (Coeff.isTwo() || Coeff.isMinusTwo()) { 718 NeedNeg = Coeff.isMinusTwo(); 719 return createFAdd(OpndVal, OpndVal); 720 } 721 722 NeedNeg = false; 723 return createFMul(OpndVal, Coeff.getValue(Instr->getType())); 724} 725 726/// AddOne - Add one to a ConstantInt. 727static Constant *AddOne(Constant *C) { 728 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 729} 730 731/// SubOne - Subtract one from a ConstantInt. 732static Constant *SubOne(ConstantInt *C) { 733 return ConstantInt::get(C->getContext(), C->getValue()-1); 734} 735 736 737// dyn_castFoldableMul - If this value is a multiply that can be folded into 738// other computations (because it has a constant operand), return the 739// non-constant operand of the multiply, and set CST to point to the multiplier. 740// Otherwise, return null. 741// 742static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) { 743 if (!V->hasOneUse() || !V->getType()->isIntegerTy()) 744 return 0; 745 746 Instruction *I = dyn_cast<Instruction>(V); 747 if (I == 0) return 0; 748 749 if (I->getOpcode() == Instruction::Mul) 750 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) 751 return I->getOperand(0); 752 if (I->getOpcode() == Instruction::Shl) 753 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) { 754 // The multiplier is really 1 << CST. 755 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth(); 756 uint32_t CSTVal = CST->getLimitedValue(BitWidth); 757 CST = ConstantInt::get(V->getType()->getContext(), 758 APInt(BitWidth, 1).shl(CSTVal)); 759 return I->getOperand(0); 760 } 761 return 0; 762} 763 764 765/// WillNotOverflowSignedAdd - Return true if we can prove that: 766/// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS)) 767/// This basically requires proving that the add in the original type would not 768/// overflow to change the sign bit or have a carry out. 769bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) { 770 // There are different heuristics we can use for this. Here are some simple 771 // ones. 772 773 // Add has the property that adding any two 2's complement numbers can only 774 // have one carry bit which can change a sign. As such, if LHS and RHS each 775 // have at least two sign bits, we know that the addition of the two values 776 // will sign extend fine. 777 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1) 778 return true; 779 780 781 // If one of the operands only has one non-zero bit, and if the other operand 782 // has a known-zero bit in a more significant place than it (not including the 783 // sign bit) the ripple may go up to and fill the zero, but won't change the 784 // sign. For example, (X & ~4) + 1. 785 786 // TODO: Implement. 787 788 return false; 789} 790 791Instruction *InstCombiner::visitAdd(BinaryOperator &I) { 792 bool Changed = SimplifyAssociativeOrCommutative(I); 793 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 794 795 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(), 796 I.hasNoUnsignedWrap(), TD)) 797 return ReplaceInstUsesWith(I, V); 798 799 // (A*B)+(A*C) -> A*(B+C) etc 800 if (Value *V = SimplifyUsingDistributiveLaws(I)) 801 return ReplaceInstUsesWith(I, V); 802 803 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 804 // X + (signbit) --> X ^ signbit 805 const APInt &Val = CI->getValue(); 806 if (Val.isSignBit()) 807 return BinaryOperator::CreateXor(LHS, RHS); 808 809 // See if SimplifyDemandedBits can simplify this. This handles stuff like 810 // (X & 254)+1 -> (X&254)|1 811 if (SimplifyDemandedInstructionBits(I)) 812 return &I; 813 814 // zext(bool) + C -> bool ? C + 1 : C 815 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS)) 816 if (ZI->getSrcTy()->isIntegerTy(1)) 817 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI); 818 819 Value *XorLHS = 0; ConstantInt *XorRHS = 0; 820 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) { 821 uint32_t TySizeBits = I.getType()->getScalarSizeInBits(); 822 const APInt &RHSVal = CI->getValue(); 823 unsigned ExtendAmt = 0; 824 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. 825 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. 826 if (XorRHS->getValue() == -RHSVal) { 827 if (RHSVal.isPowerOf2()) 828 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1; 829 else if (XorRHS->getValue().isPowerOf2()) 830 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1; 831 } 832 833 if (ExtendAmt) { 834 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt); 835 if (!MaskedValueIsZero(XorLHS, Mask)) 836 ExtendAmt = 0; 837 } 838 839 if (ExtendAmt) { 840 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt); 841 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext"); 842 return BinaryOperator::CreateAShr(NewShl, ShAmt); 843 } 844 845 // If this is a xor that was canonicalized from a sub, turn it back into 846 // a sub and fuse this add with it. 847 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) { 848 IntegerType *IT = cast<IntegerType>(I.getType()); 849 APInt LHSKnownOne(IT->getBitWidth(), 0); 850 APInt LHSKnownZero(IT->getBitWidth(), 0); 851 ComputeMaskedBits(XorLHS, LHSKnownZero, LHSKnownOne); 852 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue()) 853 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI), 854 XorLHS); 855 } 856 } 857 } 858 859 if (isa<Constant>(RHS) && isa<PHINode>(LHS)) 860 if (Instruction *NV = FoldOpIntoPhi(I)) 861 return NV; 862 863 if (I.getType()->isIntegerTy(1)) 864 return BinaryOperator::CreateXor(LHS, RHS); 865 866 // X + X --> X << 1 867 if (LHS == RHS) { 868 BinaryOperator *New = 869 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1)); 870 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 871 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 872 return New; 873 } 874 875 // -A + B --> B - A 876 // -A + -B --> -(A + B) 877 if (Value *LHSV = dyn_castNegVal(LHS)) { 878 if (!isa<Constant>(RHS)) 879 if (Value *RHSV = dyn_castNegVal(RHS)) { 880 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum"); 881 return BinaryOperator::CreateNeg(NewAdd); 882 } 883 884 return BinaryOperator::CreateSub(RHS, LHSV); 885 } 886 887 // A + -B --> A - B 888 if (!isa<Constant>(RHS)) 889 if (Value *V = dyn_castNegVal(RHS)) 890 return BinaryOperator::CreateSub(LHS, V); 891 892 893 ConstantInt *C2; 894 if (Value *X = dyn_castFoldableMul(LHS, C2)) { 895 if (X == RHS) // X*C + X --> X * (C+1) 896 return BinaryOperator::CreateMul(RHS, AddOne(C2)); 897 898 // X*C1 + X*C2 --> X * (C1+C2) 899 ConstantInt *C1; 900 if (X == dyn_castFoldableMul(RHS, C1)) 901 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2)); 902 } 903 904 // X + X*C --> X * (C+1) 905 if (dyn_castFoldableMul(RHS, C2) == LHS) 906 return BinaryOperator::CreateMul(LHS, AddOne(C2)); 907 908 // A+B --> A|B iff A and B have no bits set in common. 909 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) { 910 APInt LHSKnownOne(IT->getBitWidth(), 0); 911 APInt LHSKnownZero(IT->getBitWidth(), 0); 912 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 913 if (LHSKnownZero != 0) { 914 APInt RHSKnownOne(IT->getBitWidth(), 0); 915 APInt RHSKnownZero(IT->getBitWidth(), 0); 916 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 917 918 // No bits in common -> bitwise or. 919 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue()) 920 return BinaryOperator::CreateOr(LHS, RHS); 921 } 922 } 923 924 // W*X + Y*Z --> W * (X+Z) iff W == Y 925 { 926 Value *W, *X, *Y, *Z; 927 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) && 928 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) { 929 if (W != Y) { 930 if (W == Z) { 931 std::swap(Y, Z); 932 } else if (Y == X) { 933 std::swap(W, X); 934 } else if (X == Z) { 935 std::swap(Y, Z); 936 std::swap(W, X); 937 } 938 } 939 940 if (W == Y) { 941 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName()); 942 return BinaryOperator::CreateMul(W, NewAdd); 943 } 944 } 945 } 946 947 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) { 948 Value *X = 0; 949 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X 950 return BinaryOperator::CreateSub(SubOne(CRHS), X); 951 952 // (X & FF00) + xx00 -> (X+xx00) & FF00 953 if (LHS->hasOneUse() && 954 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) && 955 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) { 956 // See if all bits from the first bit set in the Add RHS up are included 957 // in the mask. First, get the rightmost bit. 958 const APInt &AddRHSV = CRHS->getValue(); 959 960 // Form a mask of all bits from the lowest bit added through the top. 961 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1)); 962 963 // See if the and mask includes all of these bits. 964 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue()); 965 966 if (AddRHSHighBits == AddRHSHighBitsAnd) { 967 // Okay, the xform is safe. Insert the new add pronto. 968 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName()); 969 return BinaryOperator::CreateAnd(NewAdd, C2); 970 } 971 } 972 973 // Try to fold constant add into select arguments. 974 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 975 if (Instruction *R = FoldOpIntoSelect(I, SI)) 976 return R; 977 } 978 979 // add (select X 0 (sub n A)) A --> select X A n 980 { 981 SelectInst *SI = dyn_cast<SelectInst>(LHS); 982 Value *A = RHS; 983 if (!SI) { 984 SI = dyn_cast<SelectInst>(RHS); 985 A = LHS; 986 } 987 if (SI && SI->hasOneUse()) { 988 Value *TV = SI->getTrueValue(); 989 Value *FV = SI->getFalseValue(); 990 Value *N; 991 992 // Can we fold the add into the argument of the select? 993 // We check both true and false select arguments for a matching subtract. 994 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A)))) 995 // Fold the add into the true select value. 996 return SelectInst::Create(SI->getCondition(), N, A); 997 998 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A)))) 999 // Fold the add into the false select value. 1000 return SelectInst::Create(SI->getCondition(), A, N); 1001 } 1002 } 1003 1004 // Check for (add (sext x), y), see if we can merge this into an 1005 // integer add followed by a sext. 1006 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) { 1007 // (add (sext x), cst) --> (sext (add x, cst')) 1008 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 1009 Constant *CI = 1010 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType()); 1011 if (LHSConv->hasOneUse() && 1012 ConstantExpr::getSExt(CI, I.getType()) == RHSC && 1013 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1014 // Insert the new, smaller add. 1015 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1016 CI, "addconv"); 1017 return new SExtInst(NewAdd, I.getType()); 1018 } 1019 } 1020 1021 // (add (sext x), (sext y)) --> (sext (add int x, y)) 1022 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) { 1023 // Only do this if x/y have the same type, if at last one of them has a 1024 // single use (so we don't increase the number of sexts), and if the 1025 // integer add will not overflow. 1026 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1027 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1028 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1029 RHSConv->getOperand(0))) { 1030 // Insert the new integer add. 1031 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1032 RHSConv->getOperand(0), "addconv"); 1033 return new SExtInst(NewAdd, I.getType()); 1034 } 1035 } 1036 } 1037 1038 // Check for (x & y) + (x ^ y) 1039 { 1040 Value *A = 0, *B = 0; 1041 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) && 1042 (match(LHS, m_And(m_Specific(A), m_Specific(B))) || 1043 match(LHS, m_And(m_Specific(B), m_Specific(A))))) 1044 return BinaryOperator::CreateOr(A, B); 1045 1046 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) && 1047 (match(RHS, m_And(m_Specific(A), m_Specific(B))) || 1048 match(RHS, m_And(m_Specific(B), m_Specific(A))))) 1049 return BinaryOperator::CreateOr(A, B); 1050 } 1051 1052 return Changed ? &I : 0; 1053} 1054 1055Instruction *InstCombiner::visitFAdd(BinaryOperator &I) { 1056 bool Changed = SimplifyAssociativeOrCommutative(I); 1057 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 1058 1059 if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), TD)) 1060 return ReplaceInstUsesWith(I, V); 1061 1062 if (isa<Constant>(RHS) && isa<PHINode>(LHS)) 1063 if (Instruction *NV = FoldOpIntoPhi(I)) 1064 return NV; 1065 1066 // -A + B --> B - A 1067 // -A + -B --> -(A + B) 1068 if (Value *LHSV = dyn_castFNegVal(LHS)) 1069 return BinaryOperator::CreateFSub(RHS, LHSV); 1070 1071 // A + -B --> A - B 1072 if (!isa<Constant>(RHS)) 1073 if (Value *V = dyn_castFNegVal(RHS)) 1074 return BinaryOperator::CreateFSub(LHS, V); 1075 1076 // Check for (fadd double (sitofp x), y), see if we can merge this into an 1077 // integer add followed by a promotion. 1078 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) { 1079 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst)) 1080 // ... if the constant fits in the integer value. This is useful for things 1081 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer 1082 // requires a constant pool load, and generally allows the add to be better 1083 // instcombined. 1084 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 1085 Constant *CI = 1086 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType()); 1087 if (LHSConv->hasOneUse() && 1088 ConstantExpr::getSIToFP(CI, I.getType()) == CFP && 1089 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1090 // Insert the new integer add. 1091 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1092 CI, "addconv"); 1093 return new SIToFPInst(NewAdd, I.getType()); 1094 } 1095 } 1096 1097 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y)) 1098 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) { 1099 // Only do this if x/y have the same type, if at last one of them has a 1100 // single use (so we don't increase the number of int->fp conversions), 1101 // and if the integer add will not overflow. 1102 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1103 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1104 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1105 RHSConv->getOperand(0))) { 1106 // Insert the new integer add. 1107 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1108 RHSConv->getOperand(0),"addconv"); 1109 return new SIToFPInst(NewAdd, I.getType()); 1110 } 1111 } 1112 } 1113 1114 if (I.hasUnsafeAlgebra()) { 1115 if (Value *V = FAddCombine(Builder).simplify(&I)) 1116 return ReplaceInstUsesWith(I, V); 1117 } 1118 1119 return Changed ? &I : 0; 1120} 1121 1122 1123/// Optimize pointer differences into the same array into a size. Consider: 1124/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer 1125/// operands to the ptrtoint instructions for the LHS/RHS of the subtract. 1126/// 1127Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS, 1128 Type *Ty) { 1129 assert(TD && "Must have target data info for this"); 1130 1131 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize 1132 // this. 1133 bool Swapped = false; 1134 GEPOperator *GEP1 = 0, *GEP2 = 0; 1135 1136 // For now we require one side to be the base pointer "A" or a constant 1137 // GEP derived from it. 1138 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1139 // (gep X, ...) - X 1140 if (LHSGEP->getOperand(0) == RHS) { 1141 GEP1 = LHSGEP; 1142 Swapped = false; 1143 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1144 // (gep X, ...) - (gep X, ...) 1145 if (LHSGEP->getOperand(0)->stripPointerCasts() == 1146 RHSGEP->getOperand(0)->stripPointerCasts()) { 1147 GEP2 = RHSGEP; 1148 GEP1 = LHSGEP; 1149 Swapped = false; 1150 } 1151 } 1152 } 1153 1154 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1155 // X - (gep X, ...) 1156 if (RHSGEP->getOperand(0) == LHS) { 1157 GEP1 = RHSGEP; 1158 Swapped = true; 1159 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1160 // (gep X, ...) - (gep X, ...) 1161 if (RHSGEP->getOperand(0)->stripPointerCasts() == 1162 LHSGEP->getOperand(0)->stripPointerCasts()) { 1163 GEP2 = LHSGEP; 1164 GEP1 = RHSGEP; 1165 Swapped = true; 1166 } 1167 } 1168 } 1169 1170 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and 1171 // multiple users. 1172 if (GEP1 == 0 || 1173 (GEP2 != 0 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse())) 1174 return 0; 1175 1176 // Emit the offset of the GEP and an intptr_t. 1177 Value *Result = EmitGEPOffset(GEP1); 1178 1179 // If we had a constant expression GEP on the other side offsetting the 1180 // pointer, subtract it from the offset we have. 1181 if (GEP2) { 1182 Value *Offset = EmitGEPOffset(GEP2); 1183 Result = Builder->CreateSub(Result, Offset); 1184 } 1185 1186 // If we have p - gep(p, ...) then we have to negate the result. 1187 if (Swapped) 1188 Result = Builder->CreateNeg(Result, "diff.neg"); 1189 1190 return Builder->CreateIntCast(Result, Ty, true); 1191} 1192 1193 1194Instruction *InstCombiner::visitSub(BinaryOperator &I) { 1195 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1196 1197 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(), 1198 I.hasNoUnsignedWrap(), TD)) 1199 return ReplaceInstUsesWith(I, V); 1200 1201 // (A*B)-(A*C) -> A*(B-C) etc 1202 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1203 return ReplaceInstUsesWith(I, V); 1204 1205 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW. 1206 if (Value *V = dyn_castNegVal(Op1)) { 1207 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V); 1208 Res->setHasNoSignedWrap(I.hasNoSignedWrap()); 1209 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1210 return Res; 1211 } 1212 1213 if (I.getType()->isIntegerTy(1)) 1214 return BinaryOperator::CreateXor(Op0, Op1); 1215 1216 // Replace (-1 - A) with (~A). 1217 if (match(Op0, m_AllOnes())) 1218 return BinaryOperator::CreateNot(Op1); 1219 1220 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) { 1221 // C - ~X == X + (1+C) 1222 Value *X = 0; 1223 if (match(Op1, m_Not(m_Value(X)))) 1224 return BinaryOperator::CreateAdd(X, AddOne(C)); 1225 1226 // -(X >>u 31) -> (X >>s 31) 1227 // -(X >>s 31) -> (X >>u 31) 1228 if (C->isZero()) { 1229 Value *X; ConstantInt *CI; 1230 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) && 1231 // Verify we are shifting out everything but the sign bit. 1232 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1233 return BinaryOperator::CreateAShr(X, CI); 1234 1235 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) && 1236 // Verify we are shifting out everything but the sign bit. 1237 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1238 return BinaryOperator::CreateLShr(X, CI); 1239 } 1240 1241 // Try to fold constant sub into select arguments. 1242 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1243 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1244 return R; 1245 1246 // C-(X+C2) --> (C-C2)-X 1247 ConstantInt *C2; 1248 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(C2)))) 1249 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X); 1250 1251 if (SimplifyDemandedInstructionBits(I)) 1252 return &I; 1253 } 1254 1255 1256 { Value *Y; 1257 // X-(X+Y) == -Y X-(Y+X) == -Y 1258 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) || 1259 match(Op1, m_Add(m_Value(Y), m_Specific(Op0)))) 1260 return BinaryOperator::CreateNeg(Y); 1261 1262 // (X-Y)-X == -Y 1263 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y)))) 1264 return BinaryOperator::CreateNeg(Y); 1265 } 1266 1267 if (Op1->hasOneUse()) { 1268 Value *X = 0, *Y = 0, *Z = 0; 1269 Constant *C = 0; 1270 ConstantInt *CI = 0; 1271 1272 // (X - (Y - Z)) --> (X + (Z - Y)). 1273 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z)))) 1274 return BinaryOperator::CreateAdd(Op0, 1275 Builder->CreateSub(Z, Y, Op1->getName())); 1276 1277 // (X - (X & Y)) --> (X & ~Y) 1278 // 1279 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) || 1280 match(Op1, m_And(m_Specific(Op0), m_Value(Y)))) 1281 return BinaryOperator::CreateAnd(Op0, 1282 Builder->CreateNot(Y, Y->getName() + ".not")); 1283 1284 // 0 - (X sdiv C) -> (X sdiv -C) 1285 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && 1286 match(Op0, m_Zero())) 1287 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C)); 1288 1289 // 0 - (X << Y) -> (-X << Y) when X is freely negatable. 1290 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero())) 1291 if (Value *XNeg = dyn_castNegVal(X)) 1292 return BinaryOperator::CreateShl(XNeg, Y); 1293 1294 // X - X*C --> X * (1-C) 1295 if (match(Op1, m_Mul(m_Specific(Op0), m_ConstantInt(CI)))) { 1296 Constant *CP1 = ConstantExpr::getSub(ConstantInt::get(I.getType(),1), CI); 1297 return BinaryOperator::CreateMul(Op0, CP1); 1298 } 1299 1300 // X - X<<C --> X * (1-(1<<C)) 1301 if (match(Op1, m_Shl(m_Specific(Op0), m_ConstantInt(CI)))) { 1302 Constant *One = ConstantInt::get(I.getType(), 1); 1303 C = ConstantExpr::getSub(One, ConstantExpr::getShl(One, CI)); 1304 return BinaryOperator::CreateMul(Op0, C); 1305 } 1306 1307 // X - A*-B -> X + A*B 1308 // X - -A*B -> X + A*B 1309 Value *A, *B; 1310 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) || 1311 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B)))) 1312 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B)); 1313 1314 // X - A*CI -> X + A*-CI 1315 // X - CI*A -> X + A*-CI 1316 if (match(Op1, m_Mul(m_Value(A), m_ConstantInt(CI))) || 1317 match(Op1, m_Mul(m_ConstantInt(CI), m_Value(A)))) { 1318 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI)); 1319 return BinaryOperator::CreateAdd(Op0, NewMul); 1320 } 1321 } 1322 1323 ConstantInt *C1; 1324 if (Value *X = dyn_castFoldableMul(Op0, C1)) { 1325 if (X == Op1) // X*C - X --> X * (C-1) 1326 return BinaryOperator::CreateMul(Op1, SubOne(C1)); 1327 1328 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2) 1329 if (X == dyn_castFoldableMul(Op1, C2)) 1330 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2)); 1331 } 1332 1333 // Optimize pointer differences into the same array into a size. Consider: 1334 // &A[10] - &A[0]: we should compile this to "10". 1335 if (TD) { 1336 Value *LHSOp, *RHSOp; 1337 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) && 1338 match(Op1, m_PtrToInt(m_Value(RHSOp)))) 1339 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1340 return ReplaceInstUsesWith(I, Res); 1341 1342 // trunc(p)-trunc(q) -> trunc(p-q) 1343 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) && 1344 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp))))) 1345 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1346 return ReplaceInstUsesWith(I, Res); 1347 } 1348 1349 return 0; 1350} 1351 1352Instruction *InstCombiner::visitFSub(BinaryOperator &I) { 1353 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1354 1355 if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), TD)) 1356 return ReplaceInstUsesWith(I, V); 1357 1358 // If this is a 'B = x-(-A)', change to B = x+A... 1359 if (Value *V = dyn_castFNegVal(Op1)) 1360 return BinaryOperator::CreateFAdd(Op0, V); 1361 1362 if (I.hasUnsafeAlgebra()) { 1363 if (Value *V = FAddCombine(Builder).simplify(&I)) 1364 return ReplaceInstUsesWith(I, V); 1365 } 1366 1367 return 0; 1368} 1369