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/Support/GetElementPtrTypeIterator.h" 19 #include "llvm/Support/PatternMatch.h" 20 using namespace llvm; 21 using namespace PatternMatch; 22 23 namespace { 24 25 /// Class representing coefficient of floating-point addend. 26 /// This class needs to be highly efficient, which is especially true for 27 /// the constructor. As of I write this comment, the cost of the default 28 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to 29 /// perform write-merging). 30 /// 31 class FAddendCoef { 32 public: 33 // The constructor has to initialize a APFloat, which is uncessary for 34 // most addends which have coefficient either 1 or -1. So, the constructor 35 // is expensive. In order to avoid the cost of the constructor, we should 36 // reuse some instances whenever possible. The pre-created instances 37 // FAddCombine::Add[0-5] embodies this idea. 38 // 39 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {} 40 ~FAddendCoef(); 41 42 void set(short C) { 43 assert(!insaneIntVal(C) && "Insane coefficient"); 44 IsFp = false; IntVal = C; 45 } 46 47 void set(const APFloat& C); 48 49 void negate(); 50 51 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); } 52 Value *getValue(Type *) const; 53 54 // If possible, don't define operator+/operator- etc because these 55 // operators inevitably call FAddendCoef's constructor which is not cheap. 56 void operator=(const FAddendCoef &A); 57 void operator+=(const FAddendCoef &A); 58 void operator-=(const FAddendCoef &A); 59 void operator*=(const FAddendCoef &S); 60 61 bool isOne() const { return isInt() && IntVal == 1; } 62 bool isTwo() const { return isInt() && IntVal == 2; } 63 bool isMinusOne() const { return isInt() && IntVal == -1; } 64 bool isMinusTwo() const { return isInt() && IntVal == -2; } 65 66 private: 67 bool insaneIntVal(int V) { return V > 4 || V < -4; } 68 APFloat *getFpValPtr(void) 69 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); } 70 const APFloat *getFpValPtr(void) const 71 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); } 72 73 const APFloat &getFpVal(void) const { 74 assert(IsFp && BufHasFpVal && "Incorret state"); 75 return *getFpValPtr(); 76 } 77 78 APFloat &getFpVal(void) { 79 assert(IsFp && BufHasFpVal && "Incorret state"); 80 return *getFpValPtr(); 81 } 82 83 bool isInt() const { return !IsFp; } 84 85 // If the coefficient is represented by an integer, promote it to a 86 // floating point. 87 void convertToFpType(const fltSemantics &Sem); 88 89 // Construct an APFloat from a signed integer. 90 // TODO: We should get rid of this function when APFloat can be constructed 91 // from an *SIGNED* integer. 92 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val); 93 private: 94 95 bool IsFp; 96 97 // True iff FpValBuf contains an instance of APFloat. 98 bool BufHasFpVal; 99 100 // The integer coefficient of an individual addend is either 1 or -1, 101 // and we try to simplify at most 4 addends from neighboring at most 102 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt 103 // is overkill of this end. 104 short IntVal; 105 106 AlignedCharArrayUnion<APFloat> FpValBuf; 107 }; 108 109 /// FAddend is used to represent floating-point addend. An addend is 110 /// represented as <C, V>, where the V is a symbolic value, and C is a 111 /// constant coefficient. A constant addend is represented as <C, 0>. 112 /// 113 class FAddend { 114 public: 115 FAddend() { Val = 0; } 116 117 Value *getSymVal (void) const { return Val; } 118 const FAddendCoef &getCoef(void) const { return Coeff; } 119 120 bool isConstant() const { return Val == 0; } 121 bool isZero() const { return Coeff.isZero(); } 122 123 void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; } 124 void set(const APFloat& Coefficient, Value *V) 125 { Coeff.set(Coefficient); Val = V; } 126 void set(const ConstantFP* Coefficient, Value *V) 127 { Coeff.set(Coefficient->getValueAPF()); Val = V; } 128 129 void negate() { Coeff.negate(); } 130 131 /// Drill down the U-D chain one step to find the definition of V, and 132 /// try to break the definition into one or two addends. 133 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1); 134 135 /// Similar to FAddend::drillDownOneStep() except that the value being 136 /// splitted is the addend itself. 137 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const; 138 139 void operator+=(const FAddend &T) { 140 assert((Val == T.Val) && "Symbolic-values disagree"); 141 Coeff += T.Coeff; 142 } 143 144 private: 145 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; } 146 147 // This addend has the value of "Coeff * Val". 148 Value *Val; 149 FAddendCoef Coeff; 150 }; 151 152 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along 153 /// with its neighboring at most two instructions. 154 /// 155 class FAddCombine { 156 public: 157 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(0) {} 158 Value *simplify(Instruction *FAdd); 159 160 private: 161 typedef SmallVector<const FAddend*, 4> AddendVect; 162 163 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota); 164 165 Value *performFactorization(Instruction *I); 166 167 /// Convert given addend to a Value 168 Value *createAddendVal(const FAddend &A, bool& NeedNeg); 169 170 /// Return the number of instructions needed to emit the N-ary addition. 171 unsigned calcInstrNumber(const AddendVect& Vect); 172 Value *createFSub(Value *Opnd0, Value *Opnd1); 173 Value *createFAdd(Value *Opnd0, Value *Opnd1); 174 Value *createFMul(Value *Opnd0, Value *Opnd1); 175 Value *createFDiv(Value *Opnd0, Value *Opnd1); 176 Value *createFNeg(Value *V); 177 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota); 178 void createInstPostProc(Instruction *NewInst); 179 180 InstCombiner::BuilderTy *Builder; 181 Instruction *Instr; 182 183 private: 184 // Debugging stuff are clustered here. 185 #ifndef NDEBUG 186 unsigned CreateInstrNum; 187 void initCreateInstNum() { CreateInstrNum = 0; } 188 void incCreateInstNum() { CreateInstrNum++; } 189 #else 190 void initCreateInstNum() {} 191 void incCreateInstNum() {} 192 #endif 193 }; 194 } 195 196 //===----------------------------------------------------------------------===// 197 // 198 // Implementation of 199 // {FAddendCoef, FAddend, FAddition, FAddCombine}. 200 // 201 //===----------------------------------------------------------------------===// 202 FAddendCoef::~FAddendCoef() { 203 if (BufHasFpVal) 204 getFpValPtr()->~APFloat(); 205 } 206 207 void FAddendCoef::set(const APFloat& C) { 208 APFloat *P = getFpValPtr(); 209 210 if (isInt()) { 211 // As the buffer is meanless byte stream, we cannot call 212 // APFloat::operator=(). 213 new(P) APFloat(C); 214 } else 215 *P = C; 216 217 IsFp = BufHasFpVal = true; 218 } 219 220 void FAddendCoef::convertToFpType(const fltSemantics &Sem) { 221 if (!isInt()) 222 return; 223 224 APFloat *P = getFpValPtr(); 225 if (IntVal > 0) 226 new(P) APFloat(Sem, IntVal); 227 else { 228 new(P) APFloat(Sem, 0 - IntVal); 229 P->changeSign(); 230 } 231 IsFp = BufHasFpVal = true; 232 } 233 234 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) { 235 if (Val >= 0) 236 return APFloat(Sem, Val); 237 238 APFloat T(Sem, 0 - Val); 239 T.changeSign(); 240 241 return T; 242 } 243 244 void FAddendCoef::operator=(const FAddendCoef &That) { 245 if (That.isInt()) 246 set(That.IntVal); 247 else 248 set(That.getFpVal()); 249 } 250 251 void FAddendCoef::operator+=(const FAddendCoef &That) { 252 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 253 if (isInt() == That.isInt()) { 254 if (isInt()) 255 IntVal += That.IntVal; 256 else 257 getFpVal().add(That.getFpVal(), RndMode); 258 return; 259 } 260 261 if (isInt()) { 262 const APFloat &T = That.getFpVal(); 263 convertToFpType(T.getSemantics()); 264 getFpVal().add(T, RndMode); 265 return; 266 } 267 268 APFloat &T = getFpVal(); 269 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode); 270 } 271 272 void FAddendCoef::operator-=(const FAddendCoef &That) { 273 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 274 if (isInt() == That.isInt()) { 275 if (isInt()) 276 IntVal -= That.IntVal; 277 else 278 getFpVal().subtract(That.getFpVal(), RndMode); 279 return; 280 } 281 282 if (isInt()) { 283 const APFloat &T = That.getFpVal(); 284 convertToFpType(T.getSemantics()); 285 getFpVal().subtract(T, RndMode); 286 return; 287 } 288 289 APFloat &T = getFpVal(); 290 T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode); 291 } 292 293 void FAddendCoef::operator*=(const FAddendCoef &That) { 294 if (That.isOne()) 295 return; 296 297 if (That.isMinusOne()) { 298 negate(); 299 return; 300 } 301 302 if (isInt() && That.isInt()) { 303 int Res = IntVal * (int)That.IntVal; 304 assert(!insaneIntVal(Res) && "Insane int value"); 305 IntVal = Res; 306 return; 307 } 308 309 const fltSemantics &Semantic = 310 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics(); 311 312 if (isInt()) 313 convertToFpType(Semantic); 314 APFloat &F0 = getFpVal(); 315 316 if (That.isInt()) 317 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal), 318 APFloat::rmNearestTiesToEven); 319 else 320 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven); 321 322 return; 323 } 324 325 void FAddendCoef::negate() { 326 if (isInt()) 327 IntVal = 0 - IntVal; 328 else 329 getFpVal().changeSign(); 330 } 331 332 Value *FAddendCoef::getValue(Type *Ty) const { 333 return isInt() ? 334 ConstantFP::get(Ty, float(IntVal)) : 335 ConstantFP::get(Ty->getContext(), getFpVal()); 336 } 337 338 // The definition of <Val> Addends 339 // ========================================= 340 // A + B <1, A>, <1,B> 341 // A - B <1, A>, <1,B> 342 // 0 - B <-1, B> 343 // C * A, <C, A> 344 // A + C <1, A> <C, NULL> 345 // 0 +/- 0 <0, NULL> (corner case) 346 // 347 // Legend: A and B are not constant, C is constant 348 // 349 unsigned FAddend::drillValueDownOneStep 350 (Value *Val, FAddend &Addend0, FAddend &Addend1) { 351 Instruction *I = 0; 352 if (Val == 0 || !(I = dyn_cast<Instruction>(Val))) 353 return 0; 354 355 unsigned Opcode = I->getOpcode(); 356 357 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) { 358 ConstantFP *C0, *C1; 359 Value *Opnd0 = I->getOperand(0); 360 Value *Opnd1 = I->getOperand(1); 361 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero()) 362 Opnd0 = 0; 363 364 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero()) 365 Opnd1 = 0; 366 367 if (Opnd0) { 368 if (!C0) 369 Addend0.set(1, Opnd0); 370 else 371 Addend0.set(C0, 0); 372 } 373 374 if (Opnd1) { 375 FAddend &Addend = Opnd0 ? Addend1 : Addend0; 376 if (!C1) 377 Addend.set(1, Opnd1); 378 else 379 Addend.set(C1, 0); 380 if (Opcode == Instruction::FSub) 381 Addend.negate(); 382 } 383 384 if (Opnd0 || Opnd1) 385 return Opnd0 && Opnd1 ? 2 : 1; 386 387 // Both operands are zero. Weird! 388 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), 0); 389 return 1; 390 } 391 392 if (I->getOpcode() == Instruction::FMul) { 393 Value *V0 = I->getOperand(0); 394 Value *V1 = I->getOperand(1); 395 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) { 396 Addend0.set(C, V1); 397 return 1; 398 } 399 400 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) { 401 Addend0.set(C, V0); 402 return 1; 403 } 404 } 405 406 return 0; 407 } 408 409 // Try to break *this* addend into two addends. e.g. Suppose this addend is 410 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends, 411 // i.e. <2.3, X> and <2.3, Y>. 412 // 413 unsigned FAddend::drillAddendDownOneStep 414 (FAddend &Addend0, FAddend &Addend1) const { 415 if (isConstant()) 416 return 0; 417 418 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1); 419 if (!BreakNum || Coeff.isOne()) 420 return BreakNum; 421 422 Addend0.Scale(Coeff); 423 424 if (BreakNum == 2) 425 Addend1.Scale(Coeff); 426 427 return BreakNum; 428 } 429 430 // Try to perform following optimization on the input instruction I. Return the 431 // simplified expression if was successful; otherwise, return 0. 432 // 433 // Instruction "I" is Simplified into 434 // ------------------------------------------------------- 435 // (x * y) +/- (x * z) x * (y +/- z) 436 // (y / x) +/- (z / x) (y +/- z) / x 437 // 438 Value *FAddCombine::performFactorization(Instruction *I) { 439 assert((I->getOpcode() == Instruction::FAdd || 440 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 441 442 Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0)); 443 Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1)); 444 445 if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode()) 446 return 0; 447 448 bool isMpy = false; 449 if (I0->getOpcode() == Instruction::FMul) 450 isMpy = true; 451 else if (I0->getOpcode() != Instruction::FDiv) 452 return 0; 453 454 Value *Opnd0_0 = I0->getOperand(0); 455 Value *Opnd0_1 = I0->getOperand(1); 456 Value *Opnd1_0 = I1->getOperand(0); 457 Value *Opnd1_1 = I1->getOperand(1); 458 459 // Input Instr I Factor AddSub0 AddSub1 460 // ---------------------------------------------- 461 // (x*y) +/- (x*z) x y z 462 // (y/x) +/- (z/x) x y z 463 // 464 Value *Factor = 0; 465 Value *AddSub0 = 0, *AddSub1 = 0; 466 467 if (isMpy) { 468 if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1) 469 Factor = Opnd0_0; 470 else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1) 471 Factor = Opnd0_1; 472 473 if (Factor) { 474 AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0; 475 AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0; 476 } 477 } else if (Opnd0_1 == Opnd1_1) { 478 Factor = Opnd0_1; 479 AddSub0 = Opnd0_0; 480 AddSub1 = Opnd1_0; 481 } 482 483 if (!Factor) 484 return 0; 485 486 // Create expression "NewAddSub = AddSub0 +/- AddsSub1" 487 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ? 488 createFAdd(AddSub0, AddSub1) : 489 createFSub(AddSub0, AddSub1); 490 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) { 491 const APFloat &F = CFP->getValueAPF(); 492 if (!F.isNormal()) 493 return 0; 494 } 495 496 if (isMpy) 497 return createFMul(Factor, NewAddSub); 498 499 return createFDiv(NewAddSub, Factor); 500 } 501 502 Value *FAddCombine::simplify(Instruction *I) { 503 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode"); 504 505 // Currently we are not able to handle vector type. 506 if (I->getType()->isVectorTy()) 507 return 0; 508 509 assert((I->getOpcode() == Instruction::FAdd || 510 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 511 512 // Save the instruction before calling other member-functions. 513 Instr = I; 514 515 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1; 516 517 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1); 518 519 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1. 520 unsigned Opnd0_ExpNum = 0; 521 unsigned Opnd1_ExpNum = 0; 522 523 if (!Opnd0.isConstant()) 524 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1); 525 526 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1. 527 if (OpndNum == 2 && !Opnd1.isConstant()) 528 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1); 529 530 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1 531 if (Opnd0_ExpNum && Opnd1_ExpNum) { 532 AddendVect AllOpnds; 533 AllOpnds.push_back(&Opnd0_0); 534 AllOpnds.push_back(&Opnd1_0); 535 if (Opnd0_ExpNum == 2) 536 AllOpnds.push_back(&Opnd0_1); 537 if (Opnd1_ExpNum == 2) 538 AllOpnds.push_back(&Opnd1_1); 539 540 // Compute instruction quota. We should save at least one instruction. 541 unsigned InstQuota = 0; 542 543 Value *V0 = I->getOperand(0); 544 Value *V1 = I->getOperand(1); 545 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) && 546 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1; 547 548 if (Value *R = simplifyFAdd(AllOpnds, InstQuota)) 549 return R; 550 } 551 552 if (OpndNum != 2) { 553 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be 554 // splitted into two addends, say "V = X - Y", the instruction would have 555 // been optimized into "I = Y - X" in the previous steps. 556 // 557 const FAddendCoef &CE = Opnd0.getCoef(); 558 return CE.isOne() ? Opnd0.getSymVal() : 0; 559 } 560 561 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1] 562 if (Opnd1_ExpNum) { 563 AddendVect AllOpnds; 564 AllOpnds.push_back(&Opnd0); 565 AllOpnds.push_back(&Opnd1_0); 566 if (Opnd1_ExpNum == 2) 567 AllOpnds.push_back(&Opnd1_1); 568 569 if (Value *R = simplifyFAdd(AllOpnds, 1)) 570 return R; 571 } 572 573 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1] 574 if (Opnd0_ExpNum) { 575 AddendVect AllOpnds; 576 AllOpnds.push_back(&Opnd1); 577 AllOpnds.push_back(&Opnd0_0); 578 if (Opnd0_ExpNum == 2) 579 AllOpnds.push_back(&Opnd0_1); 580 581 if (Value *R = simplifyFAdd(AllOpnds, 1)) 582 return R; 583 } 584 585 // step 6: Try factorization as the last resort, 586 return performFactorization(I); 587 } 588 589 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) { 590 591 unsigned AddendNum = Addends.size(); 592 assert(AddendNum <= 4 && "Too many addends"); 593 594 // For saving intermediate results; 595 unsigned NextTmpIdx = 0; 596 FAddend TmpResult[3]; 597 598 // Points to the constant addend of the resulting simplified expression. 599 // If the resulting expr has constant-addend, this constant-addend is 600 // desirable to reside at the top of the resulting expression tree. Placing 601 // constant close to supper-expr(s) will potentially reveal some optimization 602 // opportunities in super-expr(s). 603 // 604 const FAddend *ConstAdd = 0; 605 606 // Simplified addends are placed <SimpVect>. 607 AddendVect SimpVect; 608 609 // The outer loop works on one symbolic-value at a time. Suppose the input 610 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ... 611 // The symbolic-values will be processed in this order: x, y, z. 612 // 613 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) { 614 615 const FAddend *ThisAddend = Addends[SymIdx]; 616 if (!ThisAddend) { 617 // This addend was processed before. 618 continue; 619 } 620 621 Value *Val = ThisAddend->getSymVal(); 622 unsigned StartIdx = SimpVect.size(); 623 SimpVect.push_back(ThisAddend); 624 625 // The inner loop collects addends sharing same symbolic-value, and these 626 // addends will be later on folded into a single addend. Following above 627 // example, if the symbolic value "y" is being processed, the inner loop 628 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will 629 // be later on folded into "<b1+b2, y>". 630 // 631 for (unsigned SameSymIdx = SymIdx + 1; 632 SameSymIdx < AddendNum; SameSymIdx++) { 633 const FAddend *T = Addends[SameSymIdx]; 634 if (T && T->getSymVal() == Val) { 635 // Set null such that next iteration of the outer loop will not process 636 // this addend again. 637 Addends[SameSymIdx] = 0; 638 SimpVect.push_back(T); 639 } 640 } 641 642 // If multiple addends share same symbolic value, fold them together. 643 if (StartIdx + 1 != SimpVect.size()) { 644 FAddend &R = TmpResult[NextTmpIdx ++]; 645 R = *SimpVect[StartIdx]; 646 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++) 647 R += *SimpVect[Idx]; 648 649 // Pop all addends being folded and push the resulting folded addend. 650 SimpVect.resize(StartIdx); 651 if (Val != 0) { 652 if (!R.isZero()) { 653 SimpVect.push_back(&R); 654 } 655 } else { 656 // Don't push constant addend at this time. It will be the last element 657 // of <SimpVect>. 658 ConstAdd = &R; 659 } 660 } 661 } 662 663 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) && 664 "out-of-bound access"); 665 666 if (ConstAdd) 667 SimpVect.push_back(ConstAdd); 668 669 Value *Result; 670 if (!SimpVect.empty()) 671 Result = createNaryFAdd(SimpVect, InstrQuota); 672 else { 673 // The addition is folded to 0.0. 674 Result = ConstantFP::get(Instr->getType(), 0.0); 675 } 676 677 return Result; 678 } 679 680 Value *FAddCombine::createNaryFAdd 681 (const AddendVect &Opnds, unsigned InstrQuota) { 682 assert(!Opnds.empty() && "Expect at least one addend"); 683 684 // Step 1: Check if the # of instructions needed exceeds the quota. 685 // 686 unsigned InstrNeeded = calcInstrNumber(Opnds); 687 if (InstrNeeded > InstrQuota) 688 return 0; 689 690 initCreateInstNum(); 691 692 // step 2: Emit the N-ary addition. 693 // Note that at most three instructions are involved in Fadd-InstCombine: the 694 // addition in question, and at most two neighboring instructions. 695 // The resulting optimized addition should have at least one less instruction 696 // than the original addition expression tree. This implies that the resulting 697 // N-ary addition has at most two instructions, and we don't need to worry 698 // about tree-height when constructing the N-ary addition. 699 700 Value *LastVal = 0; 701 bool LastValNeedNeg = false; 702 703 // Iterate the addends, creating fadd/fsub using adjacent two addends. 704 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 705 I != E; I++) { 706 bool NeedNeg; 707 Value *V = createAddendVal(**I, NeedNeg); 708 if (!LastVal) { 709 LastVal = V; 710 LastValNeedNeg = NeedNeg; 711 continue; 712 } 713 714 if (LastValNeedNeg == NeedNeg) { 715 LastVal = createFAdd(LastVal, V); 716 continue; 717 } 718 719 if (LastValNeedNeg) 720 LastVal = createFSub(V, LastVal); 721 else 722 LastVal = createFSub(LastVal, V); 723 724 LastValNeedNeg = false; 725 } 726 727 if (LastValNeedNeg) { 728 LastVal = createFNeg(LastVal); 729 } 730 731 #ifndef NDEBUG 732 assert(CreateInstrNum == InstrNeeded && 733 "Inconsistent in instruction numbers"); 734 #endif 735 736 return LastVal; 737 } 738 739 Value *FAddCombine::createFSub 740 (Value *Opnd0, Value *Opnd1) { 741 Value *V = Builder->CreateFSub(Opnd0, Opnd1); 742 if (Instruction *I = dyn_cast<Instruction>(V)) 743 createInstPostProc(I); 744 return V; 745 } 746 747 Value *FAddCombine::createFNeg(Value *V) { 748 Value *Zero = cast<Value>(ConstantFP::get(V->getType(), 0.0)); 749 return createFSub(Zero, V); 750 } 751 752 Value *FAddCombine::createFAdd 753 (Value *Opnd0, Value *Opnd1) { 754 Value *V = Builder->CreateFAdd(Opnd0, Opnd1); 755 if (Instruction *I = dyn_cast<Instruction>(V)) 756 createInstPostProc(I); 757 return V; 758 } 759 760 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) { 761 Value *V = Builder->CreateFMul(Opnd0, Opnd1); 762 if (Instruction *I = dyn_cast<Instruction>(V)) 763 createInstPostProc(I); 764 return V; 765 } 766 767 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) { 768 Value *V = Builder->CreateFDiv(Opnd0, Opnd1); 769 if (Instruction *I = dyn_cast<Instruction>(V)) 770 createInstPostProc(I); 771 return V; 772 } 773 774 void FAddCombine::createInstPostProc(Instruction *NewInstr) { 775 NewInstr->setDebugLoc(Instr->getDebugLoc()); 776 777 // Keep track of the number of instruction created. 778 incCreateInstNum(); 779 780 // Propagate fast-math flags 781 NewInstr->setFastMathFlags(Instr->getFastMathFlags()); 782 } 783 784 // Return the number of instruction needed to emit the N-ary addition. 785 // NOTE: Keep this function in sync with createAddendVal(). 786 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) { 787 unsigned OpndNum = Opnds.size(); 788 unsigned InstrNeeded = OpndNum - 1; 789 790 // The number of addends in the form of "(-1)*x". 791 unsigned NegOpndNum = 0; 792 793 // Adjust the number of instructions needed to emit the N-ary add. 794 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 795 I != E; I++) { 796 const FAddend *Opnd = *I; 797 if (Opnd->isConstant()) 798 continue; 799 800 const FAddendCoef &CE = Opnd->getCoef(); 801 if (CE.isMinusOne() || CE.isMinusTwo()) 802 NegOpndNum++; 803 804 // Let the addend be "c * x". If "c == +/-1", the value of the addend 805 // is immediately available; otherwise, it needs exactly one instruction 806 // to evaluate the value. 807 if (!CE.isMinusOne() && !CE.isOne()) 808 InstrNeeded++; 809 } 810 if (NegOpndNum == OpndNum) 811 InstrNeeded++; 812 return InstrNeeded; 813 } 814 815 // Input Addend Value NeedNeg(output) 816 // ================================================================ 817 // Constant C C false 818 // <+/-1, V> V coefficient is -1 819 // <2/-2, V> "fadd V, V" coefficient is -2 820 // <C, V> "fmul V, C" false 821 // 822 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber. 823 Value *FAddCombine::createAddendVal 824 (const FAddend &Opnd, bool &NeedNeg) { 825 const FAddendCoef &Coeff = Opnd.getCoef(); 826 827 if (Opnd.isConstant()) { 828 NeedNeg = false; 829 return Coeff.getValue(Instr->getType()); 830 } 831 832 Value *OpndVal = Opnd.getSymVal(); 833 834 if (Coeff.isMinusOne() || Coeff.isOne()) { 835 NeedNeg = Coeff.isMinusOne(); 836 return OpndVal; 837 } 838 839 if (Coeff.isTwo() || Coeff.isMinusTwo()) { 840 NeedNeg = Coeff.isMinusTwo(); 841 return createFAdd(OpndVal, OpndVal); 842 } 843 844 NeedNeg = false; 845 return createFMul(OpndVal, Coeff.getValue(Instr->getType())); 846 } 847 848 /// AddOne - Add one to a ConstantInt. 849 static Constant *AddOne(Constant *C) { 850 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 851 } 852 853 /// SubOne - Subtract one from a ConstantInt. 854 static Constant *SubOne(ConstantInt *C) { 855 return ConstantInt::get(C->getContext(), C->getValue()-1); 856 } 857 858 859 // dyn_castFoldableMul - If this value is a multiply that can be folded into 860 // other computations (because it has a constant operand), return the 861 // non-constant operand of the multiply, and set CST to point to the multiplier. 862 // Otherwise, return null. 863 // 864 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) { 865 if (!V->hasOneUse() || !V->getType()->isIntegerTy()) 866 return 0; 867 868 Instruction *I = dyn_cast<Instruction>(V); 869 if (I == 0) return 0; 870 871 if (I->getOpcode() == Instruction::Mul) 872 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) 873 return I->getOperand(0); 874 if (I->getOpcode() == Instruction::Shl) 875 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) { 876 // The multiplier is really 1 << CST. 877 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth(); 878 uint32_t CSTVal = CST->getLimitedValue(BitWidth); 879 CST = ConstantInt::get(V->getType()->getContext(), 880 APInt::getOneBitSet(BitWidth, CSTVal)); 881 return I->getOperand(0); 882 } 883 return 0; 884 } 885 886 887 /// WillNotOverflowSignedAdd - Return true if we can prove that: 888 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS)) 889 /// This basically requires proving that the add in the original type would not 890 /// overflow to change the sign bit or have a carry out. 891 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) { 892 // There are different heuristics we can use for this. Here are some simple 893 // ones. 894 895 // Add has the property that adding any two 2's complement numbers can only 896 // have one carry bit which can change a sign. As such, if LHS and RHS each 897 // have at least two sign bits, we know that the addition of the two values 898 // will sign extend fine. 899 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1) 900 return true; 901 902 903 // If one of the operands only has one non-zero bit, and if the other operand 904 // has a known-zero bit in a more significant place than it (not including the 905 // sign bit) the ripple may go up to and fill the zero, but won't change the 906 // sign. For example, (X & ~4) + 1. 907 908 // TODO: Implement. 909 910 return false; 911 } 912 913 Instruction *InstCombiner::visitAdd(BinaryOperator &I) { 914 bool Changed = SimplifyAssociativeOrCommutative(I); 915 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 916 917 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(), 918 I.hasNoUnsignedWrap(), TD)) 919 return ReplaceInstUsesWith(I, V); 920 921 // (A*B)+(A*C) -> A*(B+C) etc 922 if (Value *V = SimplifyUsingDistributiveLaws(I)) 923 return ReplaceInstUsesWith(I, V); 924 925 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 926 // X + (signbit) --> X ^ signbit 927 const APInt &Val = CI->getValue(); 928 if (Val.isSignBit()) 929 return BinaryOperator::CreateXor(LHS, RHS); 930 931 // See if SimplifyDemandedBits can simplify this. This handles stuff like 932 // (X & 254)+1 -> (X&254)|1 933 if (SimplifyDemandedInstructionBits(I)) 934 return &I; 935 936 // zext(bool) + C -> bool ? C + 1 : C 937 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS)) 938 if (ZI->getSrcTy()->isIntegerTy(1)) 939 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI); 940 941 Value *XorLHS = 0; ConstantInt *XorRHS = 0; 942 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) { 943 uint32_t TySizeBits = I.getType()->getScalarSizeInBits(); 944 const APInt &RHSVal = CI->getValue(); 945 unsigned ExtendAmt = 0; 946 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. 947 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. 948 if (XorRHS->getValue() == -RHSVal) { 949 if (RHSVal.isPowerOf2()) 950 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1; 951 else if (XorRHS->getValue().isPowerOf2()) 952 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1; 953 } 954 955 if (ExtendAmt) { 956 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt); 957 if (!MaskedValueIsZero(XorLHS, Mask)) 958 ExtendAmt = 0; 959 } 960 961 if (ExtendAmt) { 962 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt); 963 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext"); 964 return BinaryOperator::CreateAShr(NewShl, ShAmt); 965 } 966 967 // If this is a xor that was canonicalized from a sub, turn it back into 968 // a sub and fuse this add with it. 969 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) { 970 IntegerType *IT = cast<IntegerType>(I.getType()); 971 APInt LHSKnownOne(IT->getBitWidth(), 0); 972 APInt LHSKnownZero(IT->getBitWidth(), 0); 973 ComputeMaskedBits(XorLHS, LHSKnownZero, LHSKnownOne); 974 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue()) 975 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI), 976 XorLHS); 977 } 978 // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C, 979 // transform them into (X + (signbit ^ C)) 980 if (XorRHS->getValue().isSignBit()) 981 return BinaryOperator::CreateAdd(XorLHS, 982 ConstantExpr::getXor(XorRHS, CI)); 983 } 984 } 985 986 if (isa<Constant>(RHS) && isa<PHINode>(LHS)) 987 if (Instruction *NV = FoldOpIntoPhi(I)) 988 return NV; 989 990 if (I.getType()->isIntegerTy(1)) 991 return BinaryOperator::CreateXor(LHS, RHS); 992 993 // X + X --> X << 1 994 if (LHS == RHS) { 995 BinaryOperator *New = 996 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1)); 997 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 998 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 999 return New; 1000 } 1001 1002 // -A + B --> B - A 1003 // -A + -B --> -(A + B) 1004 if (Value *LHSV = dyn_castNegVal(LHS)) { 1005 if (!isa<Constant>(RHS)) 1006 if (Value *RHSV = dyn_castNegVal(RHS)) { 1007 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum"); 1008 return BinaryOperator::CreateNeg(NewAdd); 1009 } 1010 1011 return BinaryOperator::CreateSub(RHS, LHSV); 1012 } 1013 1014 // A + -B --> A - B 1015 if (!isa<Constant>(RHS)) 1016 if (Value *V = dyn_castNegVal(RHS)) 1017 return BinaryOperator::CreateSub(LHS, V); 1018 1019 1020 ConstantInt *C2; 1021 if (Value *X = dyn_castFoldableMul(LHS, C2)) { 1022 if (X == RHS) // X*C + X --> X * (C+1) 1023 return BinaryOperator::CreateMul(RHS, AddOne(C2)); 1024 1025 // X*C1 + X*C2 --> X * (C1+C2) 1026 ConstantInt *C1; 1027 if (X == dyn_castFoldableMul(RHS, C1)) 1028 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2)); 1029 } 1030 1031 // X + X*C --> X * (C+1) 1032 if (dyn_castFoldableMul(RHS, C2) == LHS) 1033 return BinaryOperator::CreateMul(LHS, AddOne(C2)); 1034 1035 // A+B --> A|B iff A and B have no bits set in common. 1036 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) { 1037 APInt LHSKnownOne(IT->getBitWidth(), 0); 1038 APInt LHSKnownZero(IT->getBitWidth(), 0); 1039 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 1040 if (LHSKnownZero != 0) { 1041 APInt RHSKnownOne(IT->getBitWidth(), 0); 1042 APInt RHSKnownZero(IT->getBitWidth(), 0); 1043 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 1044 1045 // No bits in common -> bitwise or. 1046 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue()) 1047 return BinaryOperator::CreateOr(LHS, RHS); 1048 } 1049 } 1050 1051 // W*X + Y*Z --> W * (X+Z) iff W == Y 1052 { 1053 Value *W, *X, *Y, *Z; 1054 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) && 1055 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) { 1056 if (W != Y) { 1057 if (W == Z) { 1058 std::swap(Y, Z); 1059 } else if (Y == X) { 1060 std::swap(W, X); 1061 } else if (X == Z) { 1062 std::swap(Y, Z); 1063 std::swap(W, X); 1064 } 1065 } 1066 1067 if (W == Y) { 1068 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName()); 1069 return BinaryOperator::CreateMul(W, NewAdd); 1070 } 1071 } 1072 } 1073 1074 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) { 1075 Value *X = 0; 1076 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X 1077 return BinaryOperator::CreateSub(SubOne(CRHS), X); 1078 1079 // (X & FF00) + xx00 -> (X+xx00) & FF00 1080 if (LHS->hasOneUse() && 1081 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) && 1082 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) { 1083 // See if all bits from the first bit set in the Add RHS up are included 1084 // in the mask. First, get the rightmost bit. 1085 const APInt &AddRHSV = CRHS->getValue(); 1086 1087 // Form a mask of all bits from the lowest bit added through the top. 1088 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1)); 1089 1090 // See if the and mask includes all of these bits. 1091 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue()); 1092 1093 if (AddRHSHighBits == AddRHSHighBitsAnd) { 1094 // Okay, the xform is safe. Insert the new add pronto. 1095 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName()); 1096 return BinaryOperator::CreateAnd(NewAdd, C2); 1097 } 1098 } 1099 1100 // Try to fold constant add into select arguments. 1101 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1102 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1103 return R; 1104 } 1105 1106 // add (select X 0 (sub n A)) A --> select X A n 1107 { 1108 SelectInst *SI = dyn_cast<SelectInst>(LHS); 1109 Value *A = RHS; 1110 if (!SI) { 1111 SI = dyn_cast<SelectInst>(RHS); 1112 A = LHS; 1113 } 1114 if (SI && SI->hasOneUse()) { 1115 Value *TV = SI->getTrueValue(); 1116 Value *FV = SI->getFalseValue(); 1117 Value *N; 1118 1119 // Can we fold the add into the argument of the select? 1120 // We check both true and false select arguments for a matching subtract. 1121 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A)))) 1122 // Fold the add into the true select value. 1123 return SelectInst::Create(SI->getCondition(), N, A); 1124 1125 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A)))) 1126 // Fold the add into the false select value. 1127 return SelectInst::Create(SI->getCondition(), A, N); 1128 } 1129 } 1130 1131 // Check for (add (sext x), y), see if we can merge this into an 1132 // integer add followed by a sext. 1133 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) { 1134 // (add (sext x), cst) --> (sext (add x, cst')) 1135 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 1136 Constant *CI = 1137 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType()); 1138 if (LHSConv->hasOneUse() && 1139 ConstantExpr::getSExt(CI, I.getType()) == RHSC && 1140 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1141 // Insert the new, smaller add. 1142 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1143 CI, "addconv"); 1144 return new SExtInst(NewAdd, I.getType()); 1145 } 1146 } 1147 1148 // (add (sext x), (sext y)) --> (sext (add int x, y)) 1149 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) { 1150 // Only do this if x/y have the same type, if at last one of them has a 1151 // single use (so we don't increase the number of sexts), and if the 1152 // integer add will not overflow. 1153 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1154 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1155 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1156 RHSConv->getOperand(0))) { 1157 // Insert the new integer add. 1158 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1159 RHSConv->getOperand(0), "addconv"); 1160 return new SExtInst(NewAdd, I.getType()); 1161 } 1162 } 1163 } 1164 1165 // Check for (x & y) + (x ^ y) 1166 { 1167 Value *A = 0, *B = 0; 1168 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) && 1169 (match(LHS, m_And(m_Specific(A), m_Specific(B))) || 1170 match(LHS, m_And(m_Specific(B), m_Specific(A))))) 1171 return BinaryOperator::CreateOr(A, B); 1172 1173 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) && 1174 (match(RHS, m_And(m_Specific(A), m_Specific(B))) || 1175 match(RHS, m_And(m_Specific(B), m_Specific(A))))) 1176 return BinaryOperator::CreateOr(A, B); 1177 } 1178 1179 return Changed ? &I : 0; 1180 } 1181 1182 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) { 1183 bool Changed = SimplifyAssociativeOrCommutative(I); 1184 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 1185 1186 if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), TD)) 1187 return ReplaceInstUsesWith(I, V); 1188 1189 if (isa<Constant>(RHS)) { 1190 if (isa<PHINode>(LHS)) 1191 if (Instruction *NV = FoldOpIntoPhi(I)) 1192 return NV; 1193 1194 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1195 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1196 return NV; 1197 } 1198 1199 // -A + B --> B - A 1200 // -A + -B --> -(A + B) 1201 if (Value *LHSV = dyn_castFNegVal(LHS)) 1202 return BinaryOperator::CreateFSub(RHS, LHSV); 1203 1204 // A + -B --> A - B 1205 if (!isa<Constant>(RHS)) 1206 if (Value *V = dyn_castFNegVal(RHS)) 1207 return BinaryOperator::CreateFSub(LHS, V); 1208 1209 // Check for (fadd double (sitofp x), y), see if we can merge this into an 1210 // integer add followed by a promotion. 1211 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) { 1212 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst)) 1213 // ... if the constant fits in the integer value. This is useful for things 1214 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer 1215 // requires a constant pool load, and generally allows the add to be better 1216 // instcombined. 1217 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 1218 Constant *CI = 1219 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType()); 1220 if (LHSConv->hasOneUse() && 1221 ConstantExpr::getSIToFP(CI, I.getType()) == CFP && 1222 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1223 // Insert the new integer add. 1224 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1225 CI, "addconv"); 1226 return new SIToFPInst(NewAdd, I.getType()); 1227 } 1228 } 1229 1230 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y)) 1231 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) { 1232 // Only do this if x/y have the same type, if at last one of them has a 1233 // single use (so we don't increase the number of int->fp conversions), 1234 // and if the integer add will not overflow. 1235 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1236 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1237 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1238 RHSConv->getOperand(0))) { 1239 // Insert the new integer add. 1240 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1241 RHSConv->getOperand(0),"addconv"); 1242 return new SIToFPInst(NewAdd, I.getType()); 1243 } 1244 } 1245 } 1246 1247 // select C, 0, B + select C, A, 0 -> select C, A, B 1248 { 1249 Value *A1, *B1, *C1, *A2, *B2, *C2; 1250 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) && 1251 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) { 1252 if (C1 == C2) { 1253 Constant *Z1=0, *Z2=0; 1254 Value *A, *B, *C=C1; 1255 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) { 1256 Z1 = dyn_cast<Constant>(A1); A = A2; 1257 Z2 = dyn_cast<Constant>(B2); B = B1; 1258 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) { 1259 Z1 = dyn_cast<Constant>(B1); B = B2; 1260 Z2 = dyn_cast<Constant>(A2); A = A1; 1261 } 1262 1263 if (Z1 && Z2 && 1264 (I.hasNoSignedZeros() || 1265 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) { 1266 return SelectInst::Create(C, A, B); 1267 } 1268 } 1269 } 1270 } 1271 1272 if (I.hasUnsafeAlgebra()) { 1273 if (Value *V = FAddCombine(Builder).simplify(&I)) 1274 return ReplaceInstUsesWith(I, V); 1275 } 1276 1277 return Changed ? &I : 0; 1278 } 1279 1280 1281 /// Optimize pointer differences into the same array into a size. Consider: 1282 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer 1283 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract. 1284 /// 1285 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS, 1286 Type *Ty) { 1287 assert(TD && "Must have target data info for this"); 1288 1289 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize 1290 // this. 1291 bool Swapped = false; 1292 GEPOperator *GEP1 = 0, *GEP2 = 0; 1293 1294 // For now we require one side to be the base pointer "A" or a constant 1295 // GEP derived from it. 1296 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1297 // (gep X, ...) - X 1298 if (LHSGEP->getOperand(0) == RHS) { 1299 GEP1 = LHSGEP; 1300 Swapped = false; 1301 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1302 // (gep X, ...) - (gep X, ...) 1303 if (LHSGEP->getOperand(0)->stripPointerCasts() == 1304 RHSGEP->getOperand(0)->stripPointerCasts()) { 1305 GEP2 = RHSGEP; 1306 GEP1 = LHSGEP; 1307 Swapped = false; 1308 } 1309 } 1310 } 1311 1312 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1313 // X - (gep X, ...) 1314 if (RHSGEP->getOperand(0) == LHS) { 1315 GEP1 = RHSGEP; 1316 Swapped = true; 1317 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1318 // (gep X, ...) - (gep X, ...) 1319 if (RHSGEP->getOperand(0)->stripPointerCasts() == 1320 LHSGEP->getOperand(0)->stripPointerCasts()) { 1321 GEP2 = LHSGEP; 1322 GEP1 = RHSGEP; 1323 Swapped = true; 1324 } 1325 } 1326 } 1327 1328 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and 1329 // multiple users. 1330 if (GEP1 == 0 || 1331 (GEP2 != 0 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse())) 1332 return 0; 1333 1334 // Emit the offset of the GEP and an intptr_t. 1335 Value *Result = EmitGEPOffset(GEP1); 1336 1337 // If we had a constant expression GEP on the other side offsetting the 1338 // pointer, subtract it from the offset we have. 1339 if (GEP2) { 1340 Value *Offset = EmitGEPOffset(GEP2); 1341 Result = Builder->CreateSub(Result, Offset); 1342 } 1343 1344 // If we have p - gep(p, ...) then we have to negate the result. 1345 if (Swapped) 1346 Result = Builder->CreateNeg(Result, "diff.neg"); 1347 1348 return Builder->CreateIntCast(Result, Ty, true); 1349 } 1350 1351 1352 Instruction *InstCombiner::visitSub(BinaryOperator &I) { 1353 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1354 1355 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(), 1356 I.hasNoUnsignedWrap(), TD)) 1357 return ReplaceInstUsesWith(I, V); 1358 1359 // (A*B)-(A*C) -> A*(B-C) etc 1360 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1361 return ReplaceInstUsesWith(I, V); 1362 1363 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW. 1364 if (Value *V = dyn_castNegVal(Op1)) { 1365 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V); 1366 Res->setHasNoSignedWrap(I.hasNoSignedWrap()); 1367 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1368 return Res; 1369 } 1370 1371 if (I.getType()->isIntegerTy(1)) 1372 return BinaryOperator::CreateXor(Op0, Op1); 1373 1374 // Replace (-1 - A) with (~A). 1375 if (match(Op0, m_AllOnes())) 1376 return BinaryOperator::CreateNot(Op1); 1377 1378 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) { 1379 // C - ~X == X + (1+C) 1380 Value *X = 0; 1381 if (match(Op1, m_Not(m_Value(X)))) 1382 return BinaryOperator::CreateAdd(X, AddOne(C)); 1383 1384 // -(X >>u 31) -> (X >>s 31) 1385 // -(X >>s 31) -> (X >>u 31) 1386 if (C->isZero()) { 1387 Value *X; ConstantInt *CI; 1388 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) && 1389 // Verify we are shifting out everything but the sign bit. 1390 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1391 return BinaryOperator::CreateAShr(X, CI); 1392 1393 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) && 1394 // Verify we are shifting out everything but the sign bit. 1395 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1396 return BinaryOperator::CreateLShr(X, CI); 1397 } 1398 1399 // Try to fold constant sub into select arguments. 1400 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1401 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1402 return R; 1403 1404 // C-(X+C2) --> (C-C2)-X 1405 ConstantInt *C2; 1406 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(C2)))) 1407 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X); 1408 1409 if (SimplifyDemandedInstructionBits(I)) 1410 return &I; 1411 1412 // Fold (sub 0, (zext bool to B)) --> (sext bool to B) 1413 if (C->isZero() && match(Op1, m_ZExt(m_Value(X)))) 1414 if (X->getType()->isIntegerTy(1)) 1415 return CastInst::CreateSExtOrBitCast(X, Op1->getType()); 1416 1417 // Fold (sub 0, (sext bool to B)) --> (zext bool to B) 1418 if (C->isZero() && match(Op1, m_SExt(m_Value(X)))) 1419 if (X->getType()->isIntegerTy(1)) 1420 return CastInst::CreateZExtOrBitCast(X, Op1->getType()); 1421 } 1422 1423 1424 { Value *Y; 1425 // X-(X+Y) == -Y X-(Y+X) == -Y 1426 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) || 1427 match(Op1, m_Add(m_Value(Y), m_Specific(Op0)))) 1428 return BinaryOperator::CreateNeg(Y); 1429 1430 // (X-Y)-X == -Y 1431 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y)))) 1432 return BinaryOperator::CreateNeg(Y); 1433 } 1434 1435 if (Op1->hasOneUse()) { 1436 Value *X = 0, *Y = 0, *Z = 0; 1437 Constant *C = 0; 1438 ConstantInt *CI = 0; 1439 1440 // (X - (Y - Z)) --> (X + (Z - Y)). 1441 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z)))) 1442 return BinaryOperator::CreateAdd(Op0, 1443 Builder->CreateSub(Z, Y, Op1->getName())); 1444 1445 // (X - (X & Y)) --> (X & ~Y) 1446 // 1447 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) || 1448 match(Op1, m_And(m_Specific(Op0), m_Value(Y)))) 1449 return BinaryOperator::CreateAnd(Op0, 1450 Builder->CreateNot(Y, Y->getName() + ".not")); 1451 1452 // 0 - (X sdiv C) -> (X sdiv -C) 1453 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && 1454 match(Op0, m_Zero())) 1455 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C)); 1456 1457 // 0 - (X << Y) -> (-X << Y) when X is freely negatable. 1458 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero())) 1459 if (Value *XNeg = dyn_castNegVal(X)) 1460 return BinaryOperator::CreateShl(XNeg, Y); 1461 1462 // X - X*C --> X * (1-C) 1463 if (match(Op1, m_Mul(m_Specific(Op0), m_ConstantInt(CI)))) { 1464 Constant *CP1 = ConstantExpr::getSub(ConstantInt::get(I.getType(),1), CI); 1465 return BinaryOperator::CreateMul(Op0, CP1); 1466 } 1467 1468 // X - X<<C --> X * (1-(1<<C)) 1469 if (match(Op1, m_Shl(m_Specific(Op0), m_ConstantInt(CI)))) { 1470 Constant *One = ConstantInt::get(I.getType(), 1); 1471 C = ConstantExpr::getSub(One, ConstantExpr::getShl(One, CI)); 1472 return BinaryOperator::CreateMul(Op0, C); 1473 } 1474 1475 // X - A*-B -> X + A*B 1476 // X - -A*B -> X + A*B 1477 Value *A, *B; 1478 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) || 1479 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B)))) 1480 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B)); 1481 1482 // X - A*CI -> X + A*-CI 1483 // X - CI*A -> X + A*-CI 1484 if (match(Op1, m_Mul(m_Value(A), m_ConstantInt(CI))) || 1485 match(Op1, m_Mul(m_ConstantInt(CI), m_Value(A)))) { 1486 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI)); 1487 return BinaryOperator::CreateAdd(Op0, NewMul); 1488 } 1489 } 1490 1491 ConstantInt *C1; 1492 if (Value *X = dyn_castFoldableMul(Op0, C1)) { 1493 if (X == Op1) // X*C - X --> X * (C-1) 1494 return BinaryOperator::CreateMul(Op1, SubOne(C1)); 1495 1496 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2) 1497 if (X == dyn_castFoldableMul(Op1, C2)) 1498 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2)); 1499 } 1500 1501 // Optimize pointer differences into the same array into a size. Consider: 1502 // &A[10] - &A[0]: we should compile this to "10". 1503 if (TD) { 1504 Value *LHSOp, *RHSOp; 1505 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) && 1506 match(Op1, m_PtrToInt(m_Value(RHSOp)))) 1507 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1508 return ReplaceInstUsesWith(I, Res); 1509 1510 // trunc(p)-trunc(q) -> trunc(p-q) 1511 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) && 1512 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp))))) 1513 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1514 return ReplaceInstUsesWith(I, Res); 1515 } 1516 1517 return 0; 1518 } 1519 1520 Instruction *InstCombiner::visitFSub(BinaryOperator &I) { 1521 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1522 1523 if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), TD)) 1524 return ReplaceInstUsesWith(I, V); 1525 1526 if (isa<Constant>(Op0)) 1527 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1528 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1529 return NV; 1530 1531 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking 1532 // through FP extensions/truncations along the way. 1533 if (Value *V = dyn_castFNegVal(Op1)) { 1534 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V); 1535 NewI->copyFastMathFlags(&I); 1536 return NewI; 1537 } 1538 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) { 1539 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) { 1540 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType()); 1541 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc); 1542 NewI->copyFastMathFlags(&I); 1543 return NewI; 1544 } 1545 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) { 1546 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) { 1547 Value *NewExt = Builder->CreateFPExt(V, I.getType()); 1548 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt); 1549 NewI->copyFastMathFlags(&I); 1550 return NewI; 1551 } 1552 } 1553 1554 if (I.hasUnsafeAlgebra()) { 1555 if (Value *V = FAddCombine(Builder).simplify(&I)) 1556 return ReplaceInstUsesWith(I, V); 1557 } 1558 1559 return 0; 1560 } 1561