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/ADT/STLExtras.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/IR/DataLayout.h" 18 #include "llvm/IR/GetElementPtrTypeIterator.h" 19 #include "llvm/IR/PatternMatch.h" 20 using namespace llvm; 21 using namespace PatternMatch; 22 23 #define DEBUG_TYPE "instcombine" 24 25 namespace { 26 27 /// Class representing coefficient of floating-point addend. 28 /// This class needs to be highly efficient, which is especially true for 29 /// the constructor. As of I write this comment, the cost of the default 30 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to 31 /// perform write-merging). 32 /// 33 class FAddendCoef { 34 public: 35 // The constructor has to initialize a APFloat, which is unnecessary for 36 // most addends which have coefficient either 1 or -1. So, the constructor 37 // is expensive. In order to avoid the cost of the constructor, we should 38 // reuse some instances whenever possible. The pre-created instances 39 // FAddCombine::Add[0-5] embodies this idea. 40 // 41 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {} 42 ~FAddendCoef(); 43 44 void set(short C) { 45 assert(!insaneIntVal(C) && "Insane coefficient"); 46 IsFp = false; IntVal = C; 47 } 48 49 void set(const APFloat& C); 50 51 void negate(); 52 53 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); } 54 Value *getValue(Type *) const; 55 56 // If possible, don't define operator+/operator- etc because these 57 // operators inevitably call FAddendCoef's constructor which is not cheap. 58 void operator=(const FAddendCoef &A); 59 void operator+=(const FAddendCoef &A); 60 void operator-=(const FAddendCoef &A); 61 void operator*=(const FAddendCoef &S); 62 63 bool isOne() const { return isInt() && IntVal == 1; } 64 bool isTwo() const { return isInt() && IntVal == 2; } 65 bool isMinusOne() const { return isInt() && IntVal == -1; } 66 bool isMinusTwo() const { return isInt() && IntVal == -2; } 67 68 private: 69 bool insaneIntVal(int V) { return V > 4 || V < -4; } 70 APFloat *getFpValPtr(void) 71 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); } 72 const APFloat *getFpValPtr(void) const 73 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); } 74 75 const APFloat &getFpVal(void) const { 76 assert(IsFp && BufHasFpVal && "Incorret state"); 77 return *getFpValPtr(); 78 } 79 80 APFloat &getFpVal(void) { 81 assert(IsFp && BufHasFpVal && "Incorret state"); 82 return *getFpValPtr(); 83 } 84 85 bool isInt() const { return !IsFp; } 86 87 // If the coefficient is represented by an integer, promote it to a 88 // floating point. 89 void convertToFpType(const fltSemantics &Sem); 90 91 // Construct an APFloat from a signed integer. 92 // TODO: We should get rid of this function when APFloat can be constructed 93 // from an *SIGNED* integer. 94 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val); 95 private: 96 97 bool IsFp; 98 99 // True iff FpValBuf contains an instance of APFloat. 100 bool BufHasFpVal; 101 102 // The integer coefficient of an individual addend is either 1 or -1, 103 // and we try to simplify at most 4 addends from neighboring at most 104 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt 105 // is overkill of this end. 106 short IntVal; 107 108 AlignedCharArrayUnion<APFloat> FpValBuf; 109 }; 110 111 /// FAddend is used to represent floating-point addend. An addend is 112 /// represented as <C, V>, where the V is a symbolic value, and C is a 113 /// constant coefficient. A constant addend is represented as <C, 0>. 114 /// 115 class FAddend { 116 public: 117 FAddend() { Val = nullptr; } 118 119 Value *getSymVal (void) const { return Val; } 120 const FAddendCoef &getCoef(void) const { return Coeff; } 121 122 bool isConstant() const { return Val == nullptr; } 123 bool isZero() const { return Coeff.isZero(); } 124 125 void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; } 126 void set(const APFloat& Coefficient, Value *V) 127 { Coeff.set(Coefficient); Val = V; } 128 void set(const ConstantFP* Coefficient, Value *V) 129 { Coeff.set(Coefficient->getValueAPF()); Val = V; } 130 131 void negate() { Coeff.negate(); } 132 133 /// Drill down the U-D chain one step to find the definition of V, and 134 /// try to break the definition into one or two addends. 135 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1); 136 137 /// Similar to FAddend::drillDownOneStep() except that the value being 138 /// splitted is the addend itself. 139 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const; 140 141 void operator+=(const FAddend &T) { 142 assert((Val == T.Val) && "Symbolic-values disagree"); 143 Coeff += T.Coeff; 144 } 145 146 private: 147 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; } 148 149 // This addend has the value of "Coeff * Val". 150 Value *Val; 151 FAddendCoef Coeff; 152 }; 153 154 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along 155 /// with its neighboring at most two instructions. 156 /// 157 class FAddCombine { 158 public: 159 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(nullptr) {} 160 Value *simplify(Instruction *FAdd); 161 162 private: 163 typedef SmallVector<const FAddend*, 4> AddendVect; 164 165 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota); 166 167 Value *performFactorization(Instruction *I); 168 169 /// Convert given addend to a Value 170 Value *createAddendVal(const FAddend &A, bool& NeedNeg); 171 172 /// Return the number of instructions needed to emit the N-ary addition. 173 unsigned calcInstrNumber(const AddendVect& Vect); 174 Value *createFSub(Value *Opnd0, Value *Opnd1); 175 Value *createFAdd(Value *Opnd0, Value *Opnd1); 176 Value *createFMul(Value *Opnd0, Value *Opnd1); 177 Value *createFDiv(Value *Opnd0, Value *Opnd1); 178 Value *createFNeg(Value *V); 179 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota); 180 void createInstPostProc(Instruction *NewInst, bool NoNumber = false); 181 182 InstCombiner::BuilderTy *Builder; 183 Instruction *Instr; 184 185 private: 186 // Debugging stuff are clustered here. 187 #ifndef NDEBUG 188 unsigned CreateInstrNum; 189 void initCreateInstNum() { CreateInstrNum = 0; } 190 void incCreateInstNum() { CreateInstrNum++; } 191 #else 192 void initCreateInstNum() {} 193 void incCreateInstNum() {} 194 #endif 195 }; 196 } 197 198 //===----------------------------------------------------------------------===// 199 // 200 // Implementation of 201 // {FAddendCoef, FAddend, FAddition, FAddCombine}. 202 // 203 //===----------------------------------------------------------------------===// 204 FAddendCoef::~FAddendCoef() { 205 if (BufHasFpVal) 206 getFpValPtr()->~APFloat(); 207 } 208 209 void FAddendCoef::set(const APFloat& C) { 210 APFloat *P = getFpValPtr(); 211 212 if (isInt()) { 213 // As the buffer is meanless byte stream, we cannot call 214 // APFloat::operator=(). 215 new(P) APFloat(C); 216 } else 217 *P = C; 218 219 IsFp = BufHasFpVal = true; 220 } 221 222 void FAddendCoef::convertToFpType(const fltSemantics &Sem) { 223 if (!isInt()) 224 return; 225 226 APFloat *P = getFpValPtr(); 227 if (IntVal > 0) 228 new(P) APFloat(Sem, IntVal); 229 else { 230 new(P) APFloat(Sem, 0 - IntVal); 231 P->changeSign(); 232 } 233 IsFp = BufHasFpVal = true; 234 } 235 236 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) { 237 if (Val >= 0) 238 return APFloat(Sem, Val); 239 240 APFloat T(Sem, 0 - Val); 241 T.changeSign(); 242 243 return T; 244 } 245 246 void FAddendCoef::operator=(const FAddendCoef &That) { 247 if (That.isInt()) 248 set(That.IntVal); 249 else 250 set(That.getFpVal()); 251 } 252 253 void FAddendCoef::operator+=(const FAddendCoef &That) { 254 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 255 if (isInt() == That.isInt()) { 256 if (isInt()) 257 IntVal += That.IntVal; 258 else 259 getFpVal().add(That.getFpVal(), RndMode); 260 return; 261 } 262 263 if (isInt()) { 264 const APFloat &T = That.getFpVal(); 265 convertToFpType(T.getSemantics()); 266 getFpVal().add(T, RndMode); 267 return; 268 } 269 270 APFloat &T = getFpVal(); 271 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode); 272 } 273 274 void FAddendCoef::operator-=(const FAddendCoef &That) { 275 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 276 if (isInt() == That.isInt()) { 277 if (isInt()) 278 IntVal -= That.IntVal; 279 else 280 getFpVal().subtract(That.getFpVal(), RndMode); 281 return; 282 } 283 284 if (isInt()) { 285 const APFloat &T = That.getFpVal(); 286 convertToFpType(T.getSemantics()); 287 getFpVal().subtract(T, RndMode); 288 return; 289 } 290 291 APFloat &T = getFpVal(); 292 T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode); 293 } 294 295 void FAddendCoef::operator*=(const FAddendCoef &That) { 296 if (That.isOne()) 297 return; 298 299 if (That.isMinusOne()) { 300 negate(); 301 return; 302 } 303 304 if (isInt() && That.isInt()) { 305 int Res = IntVal * (int)That.IntVal; 306 assert(!insaneIntVal(Res) && "Insane int value"); 307 IntVal = Res; 308 return; 309 } 310 311 const fltSemantics &Semantic = 312 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics(); 313 314 if (isInt()) 315 convertToFpType(Semantic); 316 APFloat &F0 = getFpVal(); 317 318 if (That.isInt()) 319 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal), 320 APFloat::rmNearestTiesToEven); 321 else 322 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven); 323 324 return; 325 } 326 327 void FAddendCoef::negate() { 328 if (isInt()) 329 IntVal = 0 - IntVal; 330 else 331 getFpVal().changeSign(); 332 } 333 334 Value *FAddendCoef::getValue(Type *Ty) const { 335 return isInt() ? 336 ConstantFP::get(Ty, float(IntVal)) : 337 ConstantFP::get(Ty->getContext(), getFpVal()); 338 } 339 340 // The definition of <Val> Addends 341 // ========================================= 342 // A + B <1, A>, <1,B> 343 // A - B <1, A>, <1,B> 344 // 0 - B <-1, B> 345 // C * A, <C, A> 346 // A + C <1, A> <C, NULL> 347 // 0 +/- 0 <0, NULL> (corner case) 348 // 349 // Legend: A and B are not constant, C is constant 350 // 351 unsigned FAddend::drillValueDownOneStep 352 (Value *Val, FAddend &Addend0, FAddend &Addend1) { 353 Instruction *I = nullptr; 354 if (!Val || !(I = dyn_cast<Instruction>(Val))) 355 return 0; 356 357 unsigned Opcode = I->getOpcode(); 358 359 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) { 360 ConstantFP *C0, *C1; 361 Value *Opnd0 = I->getOperand(0); 362 Value *Opnd1 = I->getOperand(1); 363 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero()) 364 Opnd0 = nullptr; 365 366 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero()) 367 Opnd1 = nullptr; 368 369 if (Opnd0) { 370 if (!C0) 371 Addend0.set(1, Opnd0); 372 else 373 Addend0.set(C0, nullptr); 374 } 375 376 if (Opnd1) { 377 FAddend &Addend = Opnd0 ? Addend1 : Addend0; 378 if (!C1) 379 Addend.set(1, Opnd1); 380 else 381 Addend.set(C1, nullptr); 382 if (Opcode == Instruction::FSub) 383 Addend.negate(); 384 } 385 386 if (Opnd0 || Opnd1) 387 return Opnd0 && Opnd1 ? 2 : 1; 388 389 // Both operands are zero. Weird! 390 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr); 391 return 1; 392 } 393 394 if (I->getOpcode() == Instruction::FMul) { 395 Value *V0 = I->getOperand(0); 396 Value *V1 = I->getOperand(1); 397 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) { 398 Addend0.set(C, V1); 399 return 1; 400 } 401 402 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) { 403 Addend0.set(C, V0); 404 return 1; 405 } 406 } 407 408 return 0; 409 } 410 411 // Try to break *this* addend into two addends. e.g. Suppose this addend is 412 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends, 413 // i.e. <2.3, X> and <2.3, Y>. 414 // 415 unsigned FAddend::drillAddendDownOneStep 416 (FAddend &Addend0, FAddend &Addend1) const { 417 if (isConstant()) 418 return 0; 419 420 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1); 421 if (!BreakNum || Coeff.isOne()) 422 return BreakNum; 423 424 Addend0.Scale(Coeff); 425 426 if (BreakNum == 2) 427 Addend1.Scale(Coeff); 428 429 return BreakNum; 430 } 431 432 // Try to perform following optimization on the input instruction I. Return the 433 // simplified expression if was successful; otherwise, return 0. 434 // 435 // Instruction "I" is Simplified into 436 // ------------------------------------------------------- 437 // (x * y) +/- (x * z) x * (y +/- z) 438 // (y / x) +/- (z / x) (y +/- z) / x 439 // 440 Value *FAddCombine::performFactorization(Instruction *I) { 441 assert((I->getOpcode() == Instruction::FAdd || 442 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 443 444 Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0)); 445 Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1)); 446 447 if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode()) 448 return nullptr; 449 450 bool isMpy = false; 451 if (I0->getOpcode() == Instruction::FMul) 452 isMpy = true; 453 else if (I0->getOpcode() != Instruction::FDiv) 454 return nullptr; 455 456 Value *Opnd0_0 = I0->getOperand(0); 457 Value *Opnd0_1 = I0->getOperand(1); 458 Value *Opnd1_0 = I1->getOperand(0); 459 Value *Opnd1_1 = I1->getOperand(1); 460 461 // Input Instr I Factor AddSub0 AddSub1 462 // ---------------------------------------------- 463 // (x*y) +/- (x*z) x y z 464 // (y/x) +/- (z/x) x y z 465 // 466 Value *Factor = nullptr; 467 Value *AddSub0 = nullptr, *AddSub1 = nullptr; 468 469 if (isMpy) { 470 if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1) 471 Factor = Opnd0_0; 472 else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1) 473 Factor = Opnd0_1; 474 475 if (Factor) { 476 AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0; 477 AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0; 478 } 479 } else if (Opnd0_1 == Opnd1_1) { 480 Factor = Opnd0_1; 481 AddSub0 = Opnd0_0; 482 AddSub1 = Opnd1_0; 483 } 484 485 if (!Factor) 486 return nullptr; 487 488 FastMathFlags Flags; 489 Flags.setUnsafeAlgebra(); 490 if (I0) Flags &= I->getFastMathFlags(); 491 if (I1) Flags &= I->getFastMathFlags(); 492 493 // Create expression "NewAddSub = AddSub0 +/- AddsSub1" 494 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ? 495 createFAdd(AddSub0, AddSub1) : 496 createFSub(AddSub0, AddSub1); 497 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) { 498 const APFloat &F = CFP->getValueAPF(); 499 if (!F.isNormal()) 500 return nullptr; 501 } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub)) 502 II->setFastMathFlags(Flags); 503 504 if (isMpy) { 505 Value *RI = createFMul(Factor, NewAddSub); 506 if (Instruction *II = dyn_cast<Instruction>(RI)) 507 II->setFastMathFlags(Flags); 508 return RI; 509 } 510 511 Value *RI = createFDiv(NewAddSub, Factor); 512 if (Instruction *II = dyn_cast<Instruction>(RI)) 513 II->setFastMathFlags(Flags); 514 return RI; 515 } 516 517 Value *FAddCombine::simplify(Instruction *I) { 518 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode"); 519 520 // Currently we are not able to handle vector type. 521 if (I->getType()->isVectorTy()) 522 return nullptr; 523 524 assert((I->getOpcode() == Instruction::FAdd || 525 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 526 527 // Save the instruction before calling other member-functions. 528 Instr = I; 529 530 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1; 531 532 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1); 533 534 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1. 535 unsigned Opnd0_ExpNum = 0; 536 unsigned Opnd1_ExpNum = 0; 537 538 if (!Opnd0.isConstant()) 539 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1); 540 541 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1. 542 if (OpndNum == 2 && !Opnd1.isConstant()) 543 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1); 544 545 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1 546 if (Opnd0_ExpNum && Opnd1_ExpNum) { 547 AddendVect AllOpnds; 548 AllOpnds.push_back(&Opnd0_0); 549 AllOpnds.push_back(&Opnd1_0); 550 if (Opnd0_ExpNum == 2) 551 AllOpnds.push_back(&Opnd0_1); 552 if (Opnd1_ExpNum == 2) 553 AllOpnds.push_back(&Opnd1_1); 554 555 // Compute instruction quota. We should save at least one instruction. 556 unsigned InstQuota = 0; 557 558 Value *V0 = I->getOperand(0); 559 Value *V1 = I->getOperand(1); 560 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) && 561 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1; 562 563 if (Value *R = simplifyFAdd(AllOpnds, InstQuota)) 564 return R; 565 } 566 567 if (OpndNum != 2) { 568 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be 569 // splitted into two addends, say "V = X - Y", the instruction would have 570 // been optimized into "I = Y - X" in the previous steps. 571 // 572 const FAddendCoef &CE = Opnd0.getCoef(); 573 return CE.isOne() ? Opnd0.getSymVal() : nullptr; 574 } 575 576 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1] 577 if (Opnd1_ExpNum) { 578 AddendVect AllOpnds; 579 AllOpnds.push_back(&Opnd0); 580 AllOpnds.push_back(&Opnd1_0); 581 if (Opnd1_ExpNum == 2) 582 AllOpnds.push_back(&Opnd1_1); 583 584 if (Value *R = simplifyFAdd(AllOpnds, 1)) 585 return R; 586 } 587 588 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1] 589 if (Opnd0_ExpNum) { 590 AddendVect AllOpnds; 591 AllOpnds.push_back(&Opnd1); 592 AllOpnds.push_back(&Opnd0_0); 593 if (Opnd0_ExpNum == 2) 594 AllOpnds.push_back(&Opnd0_1); 595 596 if (Value *R = simplifyFAdd(AllOpnds, 1)) 597 return R; 598 } 599 600 // step 6: Try factorization as the last resort, 601 return performFactorization(I); 602 } 603 604 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) { 605 606 unsigned AddendNum = Addends.size(); 607 assert(AddendNum <= 4 && "Too many addends"); 608 609 // For saving intermediate results; 610 unsigned NextTmpIdx = 0; 611 FAddend TmpResult[3]; 612 613 // Points to the constant addend of the resulting simplified expression. 614 // If the resulting expr has constant-addend, this constant-addend is 615 // desirable to reside at the top of the resulting expression tree. Placing 616 // constant close to supper-expr(s) will potentially reveal some optimization 617 // opportunities in super-expr(s). 618 // 619 const FAddend *ConstAdd = nullptr; 620 621 // Simplified addends are placed <SimpVect>. 622 AddendVect SimpVect; 623 624 // The outer loop works on one symbolic-value at a time. Suppose the input 625 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ... 626 // The symbolic-values will be processed in this order: x, y, z. 627 // 628 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) { 629 630 const FAddend *ThisAddend = Addends[SymIdx]; 631 if (!ThisAddend) { 632 // This addend was processed before. 633 continue; 634 } 635 636 Value *Val = ThisAddend->getSymVal(); 637 unsigned StartIdx = SimpVect.size(); 638 SimpVect.push_back(ThisAddend); 639 640 // The inner loop collects addends sharing same symbolic-value, and these 641 // addends will be later on folded into a single addend. Following above 642 // example, if the symbolic value "y" is being processed, the inner loop 643 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will 644 // be later on folded into "<b1+b2, y>". 645 // 646 for (unsigned SameSymIdx = SymIdx + 1; 647 SameSymIdx < AddendNum; SameSymIdx++) { 648 const FAddend *T = Addends[SameSymIdx]; 649 if (T && T->getSymVal() == Val) { 650 // Set null such that next iteration of the outer loop will not process 651 // this addend again. 652 Addends[SameSymIdx] = nullptr; 653 SimpVect.push_back(T); 654 } 655 } 656 657 // If multiple addends share same symbolic value, fold them together. 658 if (StartIdx + 1 != SimpVect.size()) { 659 FAddend &R = TmpResult[NextTmpIdx ++]; 660 R = *SimpVect[StartIdx]; 661 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++) 662 R += *SimpVect[Idx]; 663 664 // Pop all addends being folded and push the resulting folded addend. 665 SimpVect.resize(StartIdx); 666 if (Val) { 667 if (!R.isZero()) { 668 SimpVect.push_back(&R); 669 } 670 } else { 671 // Don't push constant addend at this time. It will be the last element 672 // of <SimpVect>. 673 ConstAdd = &R; 674 } 675 } 676 } 677 678 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) && 679 "out-of-bound access"); 680 681 if (ConstAdd) 682 SimpVect.push_back(ConstAdd); 683 684 Value *Result; 685 if (!SimpVect.empty()) 686 Result = createNaryFAdd(SimpVect, InstrQuota); 687 else { 688 // The addition is folded to 0.0. 689 Result = ConstantFP::get(Instr->getType(), 0.0); 690 } 691 692 return Result; 693 } 694 695 Value *FAddCombine::createNaryFAdd 696 (const AddendVect &Opnds, unsigned InstrQuota) { 697 assert(!Opnds.empty() && "Expect at least one addend"); 698 699 // Step 1: Check if the # of instructions needed exceeds the quota. 700 // 701 unsigned InstrNeeded = calcInstrNumber(Opnds); 702 if (InstrNeeded > InstrQuota) 703 return nullptr; 704 705 initCreateInstNum(); 706 707 // step 2: Emit the N-ary addition. 708 // Note that at most three instructions are involved in Fadd-InstCombine: the 709 // addition in question, and at most two neighboring instructions. 710 // The resulting optimized addition should have at least one less instruction 711 // than the original addition expression tree. This implies that the resulting 712 // N-ary addition has at most two instructions, and we don't need to worry 713 // about tree-height when constructing the N-ary addition. 714 715 Value *LastVal = nullptr; 716 bool LastValNeedNeg = false; 717 718 // Iterate the addends, creating fadd/fsub using adjacent two addends. 719 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 720 I != E; I++) { 721 bool NeedNeg; 722 Value *V = createAddendVal(**I, NeedNeg); 723 if (!LastVal) { 724 LastVal = V; 725 LastValNeedNeg = NeedNeg; 726 continue; 727 } 728 729 if (LastValNeedNeg == NeedNeg) { 730 LastVal = createFAdd(LastVal, V); 731 continue; 732 } 733 734 if (LastValNeedNeg) 735 LastVal = createFSub(V, LastVal); 736 else 737 LastVal = createFSub(LastVal, V); 738 739 LastValNeedNeg = false; 740 } 741 742 if (LastValNeedNeg) { 743 LastVal = createFNeg(LastVal); 744 } 745 746 #ifndef NDEBUG 747 assert(CreateInstrNum == InstrNeeded && 748 "Inconsistent in instruction numbers"); 749 #endif 750 751 return LastVal; 752 } 753 754 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) { 755 Value *V = Builder->CreateFSub(Opnd0, Opnd1); 756 if (Instruction *I = dyn_cast<Instruction>(V)) 757 createInstPostProc(I); 758 return V; 759 } 760 761 Value *FAddCombine::createFNeg(Value *V) { 762 Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType())); 763 Value *NewV = createFSub(Zero, V); 764 if (Instruction *I = dyn_cast<Instruction>(NewV)) 765 createInstPostProc(I, true); // fneg's don't receive instruction numbers. 766 return NewV; 767 } 768 769 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) { 770 Value *V = Builder->CreateFAdd(Opnd0, Opnd1); 771 if (Instruction *I = dyn_cast<Instruction>(V)) 772 createInstPostProc(I); 773 return V; 774 } 775 776 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) { 777 Value *V = Builder->CreateFMul(Opnd0, Opnd1); 778 if (Instruction *I = dyn_cast<Instruction>(V)) 779 createInstPostProc(I); 780 return V; 781 } 782 783 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) { 784 Value *V = Builder->CreateFDiv(Opnd0, Opnd1); 785 if (Instruction *I = dyn_cast<Instruction>(V)) 786 createInstPostProc(I); 787 return V; 788 } 789 790 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) { 791 NewInstr->setDebugLoc(Instr->getDebugLoc()); 792 793 // Keep track of the number of instruction created. 794 if (!NoNumber) 795 incCreateInstNum(); 796 797 // Propagate fast-math flags 798 NewInstr->setFastMathFlags(Instr->getFastMathFlags()); 799 } 800 801 // Return the number of instruction needed to emit the N-ary addition. 802 // NOTE: Keep this function in sync with createAddendVal(). 803 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) { 804 unsigned OpndNum = Opnds.size(); 805 unsigned InstrNeeded = OpndNum - 1; 806 807 // The number of addends in the form of "(-1)*x". 808 unsigned NegOpndNum = 0; 809 810 // Adjust the number of instructions needed to emit the N-ary add. 811 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 812 I != E; I++) { 813 const FAddend *Opnd = *I; 814 if (Opnd->isConstant()) 815 continue; 816 817 const FAddendCoef &CE = Opnd->getCoef(); 818 if (CE.isMinusOne() || CE.isMinusTwo()) 819 NegOpndNum++; 820 821 // Let the addend be "c * x". If "c == +/-1", the value of the addend 822 // is immediately available; otherwise, it needs exactly one instruction 823 // to evaluate the value. 824 if (!CE.isMinusOne() && !CE.isOne()) 825 InstrNeeded++; 826 } 827 if (NegOpndNum == OpndNum) 828 InstrNeeded++; 829 return InstrNeeded; 830 } 831 832 // Input Addend Value NeedNeg(output) 833 // ================================================================ 834 // Constant C C false 835 // <+/-1, V> V coefficient is -1 836 // <2/-2, V> "fadd V, V" coefficient is -2 837 // <C, V> "fmul V, C" false 838 // 839 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber. 840 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) { 841 const FAddendCoef &Coeff = Opnd.getCoef(); 842 843 if (Opnd.isConstant()) { 844 NeedNeg = false; 845 return Coeff.getValue(Instr->getType()); 846 } 847 848 Value *OpndVal = Opnd.getSymVal(); 849 850 if (Coeff.isMinusOne() || Coeff.isOne()) { 851 NeedNeg = Coeff.isMinusOne(); 852 return OpndVal; 853 } 854 855 if (Coeff.isTwo() || Coeff.isMinusTwo()) { 856 NeedNeg = Coeff.isMinusTwo(); 857 return createFAdd(OpndVal, OpndVal); 858 } 859 860 NeedNeg = false; 861 return createFMul(OpndVal, Coeff.getValue(Instr->getType())); 862 } 863 864 // If one of the operands only has one non-zero bit, and if the other 865 // operand has a known-zero bit in a more significant place than it (not 866 // including the sign bit) the ripple may go up to and fill the zero, but 867 // won't change the sign. For example, (X & ~4) + 1. 868 static bool checkRippleForAdd(const APInt &Op0KnownZero, 869 const APInt &Op1KnownZero) { 870 APInt Op1MaybeOne = ~Op1KnownZero; 871 // Make sure that one of the operand has at most one bit set to 1. 872 if (Op1MaybeOne.countPopulation() != 1) 873 return false; 874 875 // Find the most significant known 0 other than the sign bit. 876 int BitWidth = Op0KnownZero.getBitWidth(); 877 APInt Op0KnownZeroTemp(Op0KnownZero); 878 Op0KnownZeroTemp.clearBit(BitWidth - 1); 879 int Op0ZeroPosition = BitWidth - Op0KnownZeroTemp.countLeadingZeros() - 1; 880 881 int Op1OnePosition = BitWidth - Op1MaybeOne.countLeadingZeros() - 1; 882 assert(Op1OnePosition >= 0); 883 884 // This also covers the case of no known zero, since in that case 885 // Op0ZeroPosition is -1. 886 return Op0ZeroPosition >= Op1OnePosition; 887 } 888 889 /// WillNotOverflowSignedAdd - Return true if we can prove that: 890 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS)) 891 /// This basically requires proving that the add in the original type would not 892 /// overflow to change the sign bit or have a carry out. 893 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS, 894 Instruction *CxtI) { 895 // There are different heuristics we can use for this. Here are some simple 896 // ones. 897 898 // If LHS and RHS each have at least two sign bits, the addition will look 899 // like 900 // 901 // XX..... + 902 // YY..... 903 // 904 // If the carry into the most significant position is 0, X and Y can't both 905 // be 1 and therefore the carry out of the addition is also 0. 906 // 907 // If the carry into the most significant position is 1, X and Y can't both 908 // be 0 and therefore the carry out of the addition is also 1. 909 // 910 // Since the carry into the most significant position is always equal to 911 // the carry out of the addition, there is no signed overflow. 912 if (ComputeNumSignBits(LHS, 0, CxtI) > 1 && 913 ComputeNumSignBits(RHS, 0, CxtI) > 1) 914 return true; 915 916 unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); 917 APInt LHSKnownZero(BitWidth, 0); 918 APInt LHSKnownOne(BitWidth, 0); 919 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, CxtI); 920 921 APInt RHSKnownZero(BitWidth, 0); 922 APInt RHSKnownOne(BitWidth, 0); 923 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, CxtI); 924 925 // Addition of two 2's compliment numbers having opposite signs will never 926 // overflow. 927 if ((LHSKnownOne[BitWidth - 1] && RHSKnownZero[BitWidth - 1]) || 928 (LHSKnownZero[BitWidth - 1] && RHSKnownOne[BitWidth - 1])) 929 return true; 930 931 // Check if carry bit of addition will not cause overflow. 932 if (checkRippleForAdd(LHSKnownZero, RHSKnownZero)) 933 return true; 934 if (checkRippleForAdd(RHSKnownZero, LHSKnownZero)) 935 return true; 936 937 return false; 938 } 939 940 /// \brief Return true if we can prove that: 941 /// (sub LHS, RHS) === (sub nsw LHS, RHS) 942 /// This basically requires proving that the add in the original type would not 943 /// overflow to change the sign bit or have a carry out. 944 /// TODO: Handle this for Vectors. 945 bool InstCombiner::WillNotOverflowSignedSub(Value *LHS, Value *RHS, 946 Instruction *CxtI) { 947 // If LHS and RHS each have at least two sign bits, the subtraction 948 // cannot overflow. 949 if (ComputeNumSignBits(LHS, 0, CxtI) > 1 && 950 ComputeNumSignBits(RHS, 0, CxtI) > 1) 951 return true; 952 953 unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); 954 APInt LHSKnownZero(BitWidth, 0); 955 APInt LHSKnownOne(BitWidth, 0); 956 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, CxtI); 957 958 APInt RHSKnownZero(BitWidth, 0); 959 APInt RHSKnownOne(BitWidth, 0); 960 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, CxtI); 961 962 // Subtraction of two 2's compliment numbers having identical signs will 963 // never overflow. 964 if ((LHSKnownOne[BitWidth - 1] && RHSKnownOne[BitWidth - 1]) || 965 (LHSKnownZero[BitWidth - 1] && RHSKnownZero[BitWidth - 1])) 966 return true; 967 968 // TODO: implement logic similar to checkRippleForAdd 969 return false; 970 } 971 972 /// \brief Return true if we can prove that: 973 /// (sub LHS, RHS) === (sub nuw LHS, RHS) 974 bool InstCombiner::WillNotOverflowUnsignedSub(Value *LHS, Value *RHS, 975 Instruction *CxtI) { 976 // If the LHS is negative and the RHS is non-negative, no unsigned wrap. 977 bool LHSKnownNonNegative, LHSKnownNegative; 978 bool RHSKnownNonNegative, RHSKnownNegative; 979 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, /*Depth=*/0, CxtI); 980 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, /*Depth=*/0, CxtI); 981 if (LHSKnownNegative && RHSKnownNonNegative) 982 return true; 983 984 return false; 985 } 986 987 // Checks if any operand is negative and we can convert add to sub. 988 // This function checks for following negative patterns 989 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C)) 990 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C)) 991 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even 992 static Value *checkForNegativeOperand(BinaryOperator &I, 993 InstCombiner::BuilderTy *Builder) { 994 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 995 996 // This function creates 2 instructions to replace ADD, we need at least one 997 // of LHS or RHS to have one use to ensure benefit in transform. 998 if (!LHS->hasOneUse() && !RHS->hasOneUse()) 999 return nullptr; 1000 1001 Value *X = nullptr, *Y = nullptr, *Z = nullptr; 1002 const APInt *C1 = nullptr, *C2 = nullptr; 1003 1004 // if ONE is on other side, swap 1005 if (match(RHS, m_Add(m_Value(X), m_One()))) 1006 std::swap(LHS, RHS); 1007 1008 if (match(LHS, m_Add(m_Value(X), m_One()))) { 1009 // if XOR on other side, swap 1010 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1)))) 1011 std::swap(X, RHS); 1012 1013 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) { 1014 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1)) 1015 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1)) 1016 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) { 1017 Value *NewAnd = Builder->CreateAnd(Z, *C1); 1018 return Builder->CreateSub(RHS, NewAnd, "sub"); 1019 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) { 1020 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1)) 1021 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1)) 1022 Value *NewOr = Builder->CreateOr(Z, ~(*C1)); 1023 return Builder->CreateSub(RHS, NewOr, "sub"); 1024 } 1025 } 1026 } 1027 1028 // Restore LHS and RHS 1029 LHS = I.getOperand(0); 1030 RHS = I.getOperand(1); 1031 1032 // if XOR is on other side, swap 1033 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1)))) 1034 std::swap(LHS, RHS); 1035 1036 // C2 is ODD 1037 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2)) 1038 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2)) 1039 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1)))) 1040 if (C1->countTrailingZeros() == 0) 1041 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) { 1042 Value *NewOr = Builder->CreateOr(Z, ~(*C2)); 1043 return Builder->CreateSub(RHS, NewOr, "sub"); 1044 } 1045 return nullptr; 1046 } 1047 1048 Instruction *InstCombiner::visitAdd(BinaryOperator &I) { 1049 bool Changed = SimplifyAssociativeOrCommutative(I); 1050 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 1051 1052 if (Value *V = SimplifyVectorOp(I)) 1053 return ReplaceInstUsesWith(I, V); 1054 1055 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(), 1056 I.hasNoUnsignedWrap(), DL, TLI, DT, AC)) 1057 return ReplaceInstUsesWith(I, V); 1058 1059 // (A*B)+(A*C) -> A*(B+C) etc 1060 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1061 return ReplaceInstUsesWith(I, V); 1062 1063 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 1064 // X + (signbit) --> X ^ signbit 1065 const APInt &Val = CI->getValue(); 1066 if (Val.isSignBit()) 1067 return BinaryOperator::CreateXor(LHS, RHS); 1068 1069 // See if SimplifyDemandedBits can simplify this. This handles stuff like 1070 // (X & 254)+1 -> (X&254)|1 1071 if (SimplifyDemandedInstructionBits(I)) 1072 return &I; 1073 1074 // zext(bool) + C -> bool ? C + 1 : C 1075 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS)) 1076 if (ZI->getSrcTy()->isIntegerTy(1)) 1077 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI); 1078 1079 Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr; 1080 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) { 1081 uint32_t TySizeBits = I.getType()->getScalarSizeInBits(); 1082 const APInt &RHSVal = CI->getValue(); 1083 unsigned ExtendAmt = 0; 1084 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. 1085 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. 1086 if (XorRHS->getValue() == -RHSVal) { 1087 if (RHSVal.isPowerOf2()) 1088 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1; 1089 else if (XorRHS->getValue().isPowerOf2()) 1090 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1; 1091 } 1092 1093 if (ExtendAmt) { 1094 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt); 1095 if (!MaskedValueIsZero(XorLHS, Mask, 0, &I)) 1096 ExtendAmt = 0; 1097 } 1098 1099 if (ExtendAmt) { 1100 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt); 1101 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext"); 1102 return BinaryOperator::CreateAShr(NewShl, ShAmt); 1103 } 1104 1105 // If this is a xor that was canonicalized from a sub, turn it back into 1106 // a sub and fuse this add with it. 1107 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) { 1108 IntegerType *IT = cast<IntegerType>(I.getType()); 1109 APInt LHSKnownOne(IT->getBitWidth(), 0); 1110 APInt LHSKnownZero(IT->getBitWidth(), 0); 1111 computeKnownBits(XorLHS, LHSKnownZero, LHSKnownOne, 0, &I); 1112 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue()) 1113 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI), 1114 XorLHS); 1115 } 1116 // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C, 1117 // transform them into (X + (signbit ^ C)) 1118 if (XorRHS->getValue().isSignBit()) 1119 return BinaryOperator::CreateAdd(XorLHS, 1120 ConstantExpr::getXor(XorRHS, CI)); 1121 } 1122 } 1123 1124 if (isa<Constant>(RHS) && isa<PHINode>(LHS)) 1125 if (Instruction *NV = FoldOpIntoPhi(I)) 1126 return NV; 1127 1128 if (I.getType()->getScalarType()->isIntegerTy(1)) 1129 return BinaryOperator::CreateXor(LHS, RHS); 1130 1131 // X + X --> X << 1 1132 if (LHS == RHS) { 1133 BinaryOperator *New = 1134 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1)); 1135 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 1136 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1137 return New; 1138 } 1139 1140 // -A + B --> B - A 1141 // -A + -B --> -(A + B) 1142 if (Value *LHSV = dyn_castNegVal(LHS)) { 1143 if (!isa<Constant>(RHS)) 1144 if (Value *RHSV = dyn_castNegVal(RHS)) { 1145 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum"); 1146 return BinaryOperator::CreateNeg(NewAdd); 1147 } 1148 1149 return BinaryOperator::CreateSub(RHS, LHSV); 1150 } 1151 1152 // A + -B --> A - B 1153 if (!isa<Constant>(RHS)) 1154 if (Value *V = dyn_castNegVal(RHS)) 1155 return BinaryOperator::CreateSub(LHS, V); 1156 1157 if (Value *V = checkForNegativeOperand(I, Builder)) 1158 return ReplaceInstUsesWith(I, V); 1159 1160 // A+B --> A|B iff A and B have no bits set in common. 1161 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) { 1162 APInt LHSKnownOne(IT->getBitWidth(), 0); 1163 APInt LHSKnownZero(IT->getBitWidth(), 0); 1164 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, &I); 1165 if (LHSKnownZero != 0) { 1166 APInt RHSKnownOne(IT->getBitWidth(), 0); 1167 APInt RHSKnownZero(IT->getBitWidth(), 0); 1168 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, &I); 1169 1170 // No bits in common -> bitwise or. 1171 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue()) 1172 return BinaryOperator::CreateOr(LHS, RHS); 1173 } 1174 } 1175 1176 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 1177 Value *X; 1178 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X 1179 return BinaryOperator::CreateSub(SubOne(CRHS), X); 1180 } 1181 1182 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) { 1183 // (X & FF00) + xx00 -> (X+xx00) & FF00 1184 Value *X; 1185 ConstantInt *C2; 1186 if (LHS->hasOneUse() && 1187 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) && 1188 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) { 1189 // See if all bits from the first bit set in the Add RHS up are included 1190 // in the mask. First, get the rightmost bit. 1191 const APInt &AddRHSV = CRHS->getValue(); 1192 1193 // Form a mask of all bits from the lowest bit added through the top. 1194 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1)); 1195 1196 // See if the and mask includes all of these bits. 1197 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue()); 1198 1199 if (AddRHSHighBits == AddRHSHighBitsAnd) { 1200 // Okay, the xform is safe. Insert the new add pronto. 1201 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName()); 1202 return BinaryOperator::CreateAnd(NewAdd, C2); 1203 } 1204 } 1205 1206 // Try to fold constant add into select arguments. 1207 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1208 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1209 return R; 1210 } 1211 1212 // add (select X 0 (sub n A)) A --> select X A n 1213 { 1214 SelectInst *SI = dyn_cast<SelectInst>(LHS); 1215 Value *A = RHS; 1216 if (!SI) { 1217 SI = dyn_cast<SelectInst>(RHS); 1218 A = LHS; 1219 } 1220 if (SI && SI->hasOneUse()) { 1221 Value *TV = SI->getTrueValue(); 1222 Value *FV = SI->getFalseValue(); 1223 Value *N; 1224 1225 // Can we fold the add into the argument of the select? 1226 // We check both true and false select arguments for a matching subtract. 1227 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A)))) 1228 // Fold the add into the true select value. 1229 return SelectInst::Create(SI->getCondition(), N, A); 1230 1231 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A)))) 1232 // Fold the add into the false select value. 1233 return SelectInst::Create(SI->getCondition(), A, N); 1234 } 1235 } 1236 1237 // Check for (add (sext x), y), see if we can merge this into an 1238 // integer add followed by a sext. 1239 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) { 1240 // (add (sext x), cst) --> (sext (add x, cst')) 1241 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 1242 Constant *CI = 1243 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType()); 1244 if (LHSConv->hasOneUse() && 1245 ConstantExpr::getSExt(CI, I.getType()) == RHSC && 1246 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI, &I)) { 1247 // Insert the new, smaller add. 1248 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1249 CI, "addconv"); 1250 return new SExtInst(NewAdd, I.getType()); 1251 } 1252 } 1253 1254 // (add (sext x), (sext y)) --> (sext (add int x, y)) 1255 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) { 1256 // Only do this if x/y have the same type, if at last one of them has a 1257 // single use (so we don't increase the number of sexts), and if the 1258 // integer add will not overflow. 1259 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1260 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1261 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1262 RHSConv->getOperand(0), &I)) { 1263 // Insert the new integer add. 1264 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1265 RHSConv->getOperand(0), "addconv"); 1266 return new SExtInst(NewAdd, I.getType()); 1267 } 1268 } 1269 } 1270 1271 // (add (xor A, B) (and A, B)) --> (or A, B) 1272 { 1273 Value *A = nullptr, *B = nullptr; 1274 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) && 1275 (match(LHS, m_And(m_Specific(A), m_Specific(B))) || 1276 match(LHS, m_And(m_Specific(B), m_Specific(A))))) 1277 return BinaryOperator::CreateOr(A, B); 1278 1279 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) && 1280 (match(RHS, m_And(m_Specific(A), m_Specific(B))) || 1281 match(RHS, m_And(m_Specific(B), m_Specific(A))))) 1282 return BinaryOperator::CreateOr(A, B); 1283 } 1284 1285 // (add (or A, B) (and A, B)) --> (add A, B) 1286 { 1287 Value *A = nullptr, *B = nullptr; 1288 if (match(RHS, m_Or(m_Value(A), m_Value(B))) && 1289 (match(LHS, m_And(m_Specific(A), m_Specific(B))) || 1290 match(LHS, m_And(m_Specific(B), m_Specific(A))))) { 1291 auto *New = BinaryOperator::CreateAdd(A, B); 1292 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 1293 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1294 return New; 1295 } 1296 1297 if (match(LHS, m_Or(m_Value(A), m_Value(B))) && 1298 (match(RHS, m_And(m_Specific(A), m_Specific(B))) || 1299 match(RHS, m_And(m_Specific(B), m_Specific(A))))) { 1300 auto *New = BinaryOperator::CreateAdd(A, B); 1301 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 1302 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1303 return New; 1304 } 1305 } 1306 1307 // TODO(jingyue): Consider WillNotOverflowSignedAdd and 1308 // WillNotOverflowUnsignedAdd to reduce the number of invocations of 1309 // computeKnownBits. 1310 if (!I.hasNoSignedWrap() && WillNotOverflowSignedAdd(LHS, RHS, &I)) { 1311 Changed = true; 1312 I.setHasNoSignedWrap(true); 1313 } 1314 if (!I.hasNoUnsignedWrap() && 1315 computeOverflowForUnsignedAdd(LHS, RHS, &I) == 1316 OverflowResult::NeverOverflows) { 1317 Changed = true; 1318 I.setHasNoUnsignedWrap(true); 1319 } 1320 1321 return Changed ? &I : nullptr; 1322 } 1323 1324 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) { 1325 bool Changed = SimplifyAssociativeOrCommutative(I); 1326 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 1327 1328 if (Value *V = SimplifyVectorOp(I)) 1329 return ReplaceInstUsesWith(I, V); 1330 1331 if (Value *V = 1332 SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL, TLI, DT, AC)) 1333 return ReplaceInstUsesWith(I, V); 1334 1335 if (isa<Constant>(RHS)) { 1336 if (isa<PHINode>(LHS)) 1337 if (Instruction *NV = FoldOpIntoPhi(I)) 1338 return NV; 1339 1340 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1341 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1342 return NV; 1343 } 1344 1345 // -A + B --> B - A 1346 // -A + -B --> -(A + B) 1347 if (Value *LHSV = dyn_castFNegVal(LHS)) { 1348 Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV); 1349 RI->copyFastMathFlags(&I); 1350 return RI; 1351 } 1352 1353 // A + -B --> A - B 1354 if (!isa<Constant>(RHS)) 1355 if (Value *V = dyn_castFNegVal(RHS)) { 1356 Instruction *RI = BinaryOperator::CreateFSub(LHS, V); 1357 RI->copyFastMathFlags(&I); 1358 return RI; 1359 } 1360 1361 // Check for (fadd double (sitofp x), y), see if we can merge this into an 1362 // integer add followed by a promotion. 1363 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) { 1364 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst)) 1365 // ... if the constant fits in the integer value. This is useful for things 1366 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer 1367 // requires a constant pool load, and generally allows the add to be better 1368 // instcombined. 1369 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 1370 Constant *CI = 1371 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType()); 1372 if (LHSConv->hasOneUse() && 1373 ConstantExpr::getSIToFP(CI, I.getType()) == CFP && 1374 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI, &I)) { 1375 // Insert the new integer add. 1376 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1377 CI, "addconv"); 1378 return new SIToFPInst(NewAdd, I.getType()); 1379 } 1380 } 1381 1382 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y)) 1383 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) { 1384 // Only do this if x/y have the same type, if at last one of them has a 1385 // single use (so we don't increase the number of int->fp conversions), 1386 // and if the integer add will not overflow. 1387 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1388 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1389 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1390 RHSConv->getOperand(0), &I)) { 1391 // Insert the new integer add. 1392 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1393 RHSConv->getOperand(0),"addconv"); 1394 return new SIToFPInst(NewAdd, I.getType()); 1395 } 1396 } 1397 } 1398 1399 // select C, 0, B + select C, A, 0 -> select C, A, B 1400 { 1401 Value *A1, *B1, *C1, *A2, *B2, *C2; 1402 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) && 1403 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) { 1404 if (C1 == C2) { 1405 Constant *Z1=nullptr, *Z2=nullptr; 1406 Value *A, *B, *C=C1; 1407 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) { 1408 Z1 = dyn_cast<Constant>(A1); A = A2; 1409 Z2 = dyn_cast<Constant>(B2); B = B1; 1410 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) { 1411 Z1 = dyn_cast<Constant>(B1); B = B2; 1412 Z2 = dyn_cast<Constant>(A2); A = A1; 1413 } 1414 1415 if (Z1 && Z2 && 1416 (I.hasNoSignedZeros() || 1417 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) { 1418 return SelectInst::Create(C, A, B); 1419 } 1420 } 1421 } 1422 } 1423 1424 if (I.hasUnsafeAlgebra()) { 1425 if (Value *V = FAddCombine(Builder).simplify(&I)) 1426 return ReplaceInstUsesWith(I, V); 1427 } 1428 1429 return Changed ? &I : nullptr; 1430 } 1431 1432 1433 /// Optimize pointer differences into the same array into a size. Consider: 1434 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer 1435 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract. 1436 /// 1437 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS, 1438 Type *Ty) { 1439 assert(DL && "Must have target data info for this"); 1440 1441 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize 1442 // this. 1443 bool Swapped = false; 1444 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr; 1445 1446 // For now we require one side to be the base pointer "A" or a constant 1447 // GEP derived from it. 1448 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1449 // (gep X, ...) - X 1450 if (LHSGEP->getOperand(0) == RHS) { 1451 GEP1 = LHSGEP; 1452 Swapped = false; 1453 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1454 // (gep X, ...) - (gep X, ...) 1455 if (LHSGEP->getOperand(0)->stripPointerCasts() == 1456 RHSGEP->getOperand(0)->stripPointerCasts()) { 1457 GEP2 = RHSGEP; 1458 GEP1 = LHSGEP; 1459 Swapped = false; 1460 } 1461 } 1462 } 1463 1464 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1465 // X - (gep X, ...) 1466 if (RHSGEP->getOperand(0) == LHS) { 1467 GEP1 = RHSGEP; 1468 Swapped = true; 1469 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1470 // (gep X, ...) - (gep X, ...) 1471 if (RHSGEP->getOperand(0)->stripPointerCasts() == 1472 LHSGEP->getOperand(0)->stripPointerCasts()) { 1473 GEP2 = LHSGEP; 1474 GEP1 = RHSGEP; 1475 Swapped = true; 1476 } 1477 } 1478 } 1479 1480 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and 1481 // multiple users. 1482 if (!GEP1 || 1483 (GEP2 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse())) 1484 return nullptr; 1485 1486 // Emit the offset of the GEP and an intptr_t. 1487 Value *Result = EmitGEPOffset(GEP1); 1488 1489 // If we had a constant expression GEP on the other side offsetting the 1490 // pointer, subtract it from the offset we have. 1491 if (GEP2) { 1492 Value *Offset = EmitGEPOffset(GEP2); 1493 Result = Builder->CreateSub(Result, Offset); 1494 } 1495 1496 // If we have p - gep(p, ...) then we have to negate the result. 1497 if (Swapped) 1498 Result = Builder->CreateNeg(Result, "diff.neg"); 1499 1500 return Builder->CreateIntCast(Result, Ty, true); 1501 } 1502 1503 Instruction *InstCombiner::visitSub(BinaryOperator &I) { 1504 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1505 1506 if (Value *V = SimplifyVectorOp(I)) 1507 return ReplaceInstUsesWith(I, V); 1508 1509 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(), 1510 I.hasNoUnsignedWrap(), DL, TLI, DT, AC)) 1511 return ReplaceInstUsesWith(I, V); 1512 1513 // (A*B)-(A*C) -> A*(B-C) etc 1514 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1515 return ReplaceInstUsesWith(I, V); 1516 1517 // If this is a 'B = x-(-A)', change to B = x+A. 1518 if (Value *V = dyn_castNegVal(Op1)) { 1519 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V); 1520 1521 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) { 1522 assert(BO->getOpcode() == Instruction::Sub && 1523 "Expected a subtraction operator!"); 1524 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap()) 1525 Res->setHasNoSignedWrap(true); 1526 } else { 1527 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap()) 1528 Res->setHasNoSignedWrap(true); 1529 } 1530 1531 return Res; 1532 } 1533 1534 if (I.getType()->isIntegerTy(1)) 1535 return BinaryOperator::CreateXor(Op0, Op1); 1536 1537 // Replace (-1 - A) with (~A). 1538 if (match(Op0, m_AllOnes())) 1539 return BinaryOperator::CreateNot(Op1); 1540 1541 if (Constant *C = dyn_cast<Constant>(Op0)) { 1542 // C - ~X == X + (1+C) 1543 Value *X = nullptr; 1544 if (match(Op1, m_Not(m_Value(X)))) 1545 return BinaryOperator::CreateAdd(X, AddOne(C)); 1546 1547 // Try to fold constant sub into select arguments. 1548 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1549 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1550 return R; 1551 1552 // C-(X+C2) --> (C-C2)-X 1553 Constant *C2; 1554 if (match(Op1, m_Add(m_Value(X), m_Constant(C2)))) 1555 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X); 1556 1557 if (SimplifyDemandedInstructionBits(I)) 1558 return &I; 1559 1560 // Fold (sub 0, (zext bool to B)) --> (sext bool to B) 1561 if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X)))) 1562 if (X->getType()->getScalarType()->isIntegerTy(1)) 1563 return CastInst::CreateSExtOrBitCast(X, Op1->getType()); 1564 1565 // Fold (sub 0, (sext bool to B)) --> (zext bool to B) 1566 if (C->isNullValue() && match(Op1, m_SExt(m_Value(X)))) 1567 if (X->getType()->getScalarType()->isIntegerTy(1)) 1568 return CastInst::CreateZExtOrBitCast(X, Op1->getType()); 1569 } 1570 1571 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) { 1572 // -(X >>u 31) -> (X >>s 31) 1573 // -(X >>s 31) -> (X >>u 31) 1574 if (C->isZero()) { 1575 Value *X; 1576 ConstantInt *CI; 1577 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) && 1578 // Verify we are shifting out everything but the sign bit. 1579 CI->getValue() == I.getType()->getPrimitiveSizeInBits() - 1) 1580 return BinaryOperator::CreateAShr(X, CI); 1581 1582 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) && 1583 // Verify we are shifting out everything but the sign bit. 1584 CI->getValue() == I.getType()->getPrimitiveSizeInBits() - 1) 1585 return BinaryOperator::CreateLShr(X, CI); 1586 } 1587 } 1588 1589 1590 { 1591 Value *Y; 1592 // X-(X+Y) == -Y X-(Y+X) == -Y 1593 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) || 1594 match(Op1, m_Add(m_Value(Y), m_Specific(Op0)))) 1595 return BinaryOperator::CreateNeg(Y); 1596 1597 // (X-Y)-X == -Y 1598 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y)))) 1599 return BinaryOperator::CreateNeg(Y); 1600 } 1601 1602 // (sub (or A, B) (xor A, B)) --> (and A, B) 1603 { 1604 Value *A = nullptr, *B = nullptr; 1605 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 1606 (match(Op0, m_Or(m_Specific(A), m_Specific(B))) || 1607 match(Op0, m_Or(m_Specific(B), m_Specific(A))))) 1608 return BinaryOperator::CreateAnd(A, B); 1609 } 1610 1611 if (Op0->hasOneUse()) { 1612 Value *Y = nullptr; 1613 // ((X | Y) - X) --> (~X & Y) 1614 if (match(Op0, m_Or(m_Value(Y), m_Specific(Op1))) || 1615 match(Op0, m_Or(m_Specific(Op1), m_Value(Y)))) 1616 return BinaryOperator::CreateAnd( 1617 Y, Builder->CreateNot(Op1, Op1->getName() + ".not")); 1618 } 1619 1620 if (Op1->hasOneUse()) { 1621 Value *X = nullptr, *Y = nullptr, *Z = nullptr; 1622 Constant *C = nullptr; 1623 Constant *CI = nullptr; 1624 1625 // (X - (Y - Z)) --> (X + (Z - Y)). 1626 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z)))) 1627 return BinaryOperator::CreateAdd(Op0, 1628 Builder->CreateSub(Z, Y, Op1->getName())); 1629 1630 // (X - (X & Y)) --> (X & ~Y) 1631 // 1632 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) || 1633 match(Op1, m_And(m_Specific(Op0), m_Value(Y)))) 1634 return BinaryOperator::CreateAnd(Op0, 1635 Builder->CreateNot(Y, Y->getName() + ".not")); 1636 1637 // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow. 1638 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) && 1639 C->isNotMinSignedValue() && !C->isOneValue()) 1640 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C)); 1641 1642 // 0 - (X << Y) -> (-X << Y) when X is freely negatable. 1643 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero())) 1644 if (Value *XNeg = dyn_castNegVal(X)) 1645 return BinaryOperator::CreateShl(XNeg, Y); 1646 1647 // X - A*-B -> X + A*B 1648 // X - -A*B -> X + A*B 1649 Value *A, *B; 1650 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) || 1651 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B)))) 1652 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B)); 1653 1654 // X - A*CI -> X + A*-CI 1655 // X - CI*A -> X + A*-CI 1656 if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) || 1657 match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) { 1658 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI)); 1659 return BinaryOperator::CreateAdd(Op0, NewMul); 1660 } 1661 } 1662 1663 // Optimize pointer differences into the same array into a size. Consider: 1664 // &A[10] - &A[0]: we should compile this to "10". 1665 if (DL) { 1666 Value *LHSOp, *RHSOp; 1667 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) && 1668 match(Op1, m_PtrToInt(m_Value(RHSOp)))) 1669 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1670 return ReplaceInstUsesWith(I, Res); 1671 1672 // trunc(p)-trunc(q) -> trunc(p-q) 1673 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) && 1674 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp))))) 1675 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1676 return ReplaceInstUsesWith(I, Res); 1677 } 1678 1679 bool Changed = false; 1680 if (!I.hasNoSignedWrap() && WillNotOverflowSignedSub(Op0, Op1, &I)) { 1681 Changed = true; 1682 I.setHasNoSignedWrap(true); 1683 } 1684 if (!I.hasNoUnsignedWrap() && WillNotOverflowUnsignedSub(Op0, Op1, &I)) { 1685 Changed = true; 1686 I.setHasNoUnsignedWrap(true); 1687 } 1688 1689 return Changed ? &I : nullptr; 1690 } 1691 1692 Instruction *InstCombiner::visitFSub(BinaryOperator &I) { 1693 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1694 1695 if (Value *V = SimplifyVectorOp(I)) 1696 return ReplaceInstUsesWith(I, V); 1697 1698 if (Value *V = 1699 SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL, TLI, DT, AC)) 1700 return ReplaceInstUsesWith(I, V); 1701 1702 // fsub nsz 0, X ==> fsub nsz -0.0, X 1703 if (I.getFastMathFlags().noSignedZeros() && match(Op0, m_Zero())) { 1704 // Subtraction from -0.0 is the canonical form of fneg. 1705 Instruction *NewI = BinaryOperator::CreateFNeg(Op1); 1706 NewI->copyFastMathFlags(&I); 1707 return NewI; 1708 } 1709 1710 if (isa<Constant>(Op0)) 1711 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1712 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1713 return NV; 1714 1715 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking 1716 // through FP extensions/truncations along the way. 1717 if (Value *V = dyn_castFNegVal(Op1)) { 1718 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V); 1719 NewI->copyFastMathFlags(&I); 1720 return NewI; 1721 } 1722 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) { 1723 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) { 1724 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType()); 1725 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc); 1726 NewI->copyFastMathFlags(&I); 1727 return NewI; 1728 } 1729 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) { 1730 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) { 1731 Value *NewExt = Builder->CreateFPExt(V, I.getType()); 1732 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt); 1733 NewI->copyFastMathFlags(&I); 1734 return NewI; 1735 } 1736 } 1737 1738 if (I.hasUnsafeAlgebra()) { 1739 if (Value *V = FAddCombine(Builder).simplify(&I)) 1740 return ReplaceInstUsesWith(I, V); 1741 } 1742 1743 return nullptr; 1744 } 1745