InstCombineAddSub.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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" 20using namespace llvm; 21using namespace PatternMatch; 22 23namespace { 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, bool NoNumber = false); 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//===----------------------------------------------------------------------===// 202FAddendCoef::~FAddendCoef() { 203 if (BufHasFpVal) 204 getFpValPtr()->~APFloat(); 205} 206 207void 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 220void 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 234APFloat 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 244void FAddendCoef::operator=(const FAddendCoef &That) { 245 if (That.isInt()) 246 set(That.IntVal); 247 else 248 set(That.getFpVal()); 249} 250 251void 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 272void 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 293void 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 325void FAddendCoef::negate() { 326 if (isInt()) 327 IntVal = 0 - IntVal; 328 else 329 getFpVal().changeSign(); 330} 331 332Value *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// 349unsigned 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// 413unsigned 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// 438Value *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 FastMathFlags Flags; 487 Flags.setUnsafeAlgebra(); 488 if (I0) Flags &= I->getFastMathFlags(); 489 if (I1) Flags &= I->getFastMathFlags(); 490 491 // Create expression "NewAddSub = AddSub0 +/- AddsSub1" 492 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ? 493 createFAdd(AddSub0, AddSub1) : 494 createFSub(AddSub0, AddSub1); 495 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) { 496 const APFloat &F = CFP->getValueAPF(); 497 if (!F.isNormal()) 498 return 0; 499 } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub)) 500 II->setFastMathFlags(Flags); 501 502 if (isMpy) { 503 Value *RI = createFMul(Factor, NewAddSub); 504 if (Instruction *II = dyn_cast<Instruction>(RI)) 505 II->setFastMathFlags(Flags); 506 return RI; 507 } 508 509 Value *RI = createFDiv(NewAddSub, Factor); 510 if (Instruction *II = dyn_cast<Instruction>(RI)) 511 II->setFastMathFlags(Flags); 512 return RI; 513} 514 515Value *FAddCombine::simplify(Instruction *I) { 516 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode"); 517 518 // Currently we are not able to handle vector type. 519 if (I->getType()->isVectorTy()) 520 return 0; 521 522 assert((I->getOpcode() == Instruction::FAdd || 523 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 524 525 // Save the instruction before calling other member-functions. 526 Instr = I; 527 528 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1; 529 530 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1); 531 532 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1. 533 unsigned Opnd0_ExpNum = 0; 534 unsigned Opnd1_ExpNum = 0; 535 536 if (!Opnd0.isConstant()) 537 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1); 538 539 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1. 540 if (OpndNum == 2 && !Opnd1.isConstant()) 541 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1); 542 543 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1 544 if (Opnd0_ExpNum && Opnd1_ExpNum) { 545 AddendVect AllOpnds; 546 AllOpnds.push_back(&Opnd0_0); 547 AllOpnds.push_back(&Opnd1_0); 548 if (Opnd0_ExpNum == 2) 549 AllOpnds.push_back(&Opnd0_1); 550 if (Opnd1_ExpNum == 2) 551 AllOpnds.push_back(&Opnd1_1); 552 553 // Compute instruction quota. We should save at least one instruction. 554 unsigned InstQuota = 0; 555 556 Value *V0 = I->getOperand(0); 557 Value *V1 = I->getOperand(1); 558 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) && 559 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1; 560 561 if (Value *R = simplifyFAdd(AllOpnds, InstQuota)) 562 return R; 563 } 564 565 if (OpndNum != 2) { 566 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be 567 // splitted into two addends, say "V = X - Y", the instruction would have 568 // been optimized into "I = Y - X" in the previous steps. 569 // 570 const FAddendCoef &CE = Opnd0.getCoef(); 571 return CE.isOne() ? Opnd0.getSymVal() : 0; 572 } 573 574 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1] 575 if (Opnd1_ExpNum) { 576 AddendVect AllOpnds; 577 AllOpnds.push_back(&Opnd0); 578 AllOpnds.push_back(&Opnd1_0); 579 if (Opnd1_ExpNum == 2) 580 AllOpnds.push_back(&Opnd1_1); 581 582 if (Value *R = simplifyFAdd(AllOpnds, 1)) 583 return R; 584 } 585 586 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1] 587 if (Opnd0_ExpNum) { 588 AddendVect AllOpnds; 589 AllOpnds.push_back(&Opnd1); 590 AllOpnds.push_back(&Opnd0_0); 591 if (Opnd0_ExpNum == 2) 592 AllOpnds.push_back(&Opnd0_1); 593 594 if (Value *R = simplifyFAdd(AllOpnds, 1)) 595 return R; 596 } 597 598 // step 6: Try factorization as the last resort, 599 return performFactorization(I); 600} 601 602Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) { 603 604 unsigned AddendNum = Addends.size(); 605 assert(AddendNum <= 4 && "Too many addends"); 606 607 // For saving intermediate results; 608 unsigned NextTmpIdx = 0; 609 FAddend TmpResult[3]; 610 611 // Points to the constant addend of the resulting simplified expression. 612 // If the resulting expr has constant-addend, this constant-addend is 613 // desirable to reside at the top of the resulting expression tree. Placing 614 // constant close to supper-expr(s) will potentially reveal some optimization 615 // opportunities in super-expr(s). 616 // 617 const FAddend *ConstAdd = 0; 618 619 // Simplified addends are placed <SimpVect>. 620 AddendVect SimpVect; 621 622 // The outer loop works on one symbolic-value at a time. Suppose the input 623 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ... 624 // The symbolic-values will be processed in this order: x, y, z. 625 // 626 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) { 627 628 const FAddend *ThisAddend = Addends[SymIdx]; 629 if (!ThisAddend) { 630 // This addend was processed before. 631 continue; 632 } 633 634 Value *Val = ThisAddend->getSymVal(); 635 unsigned StartIdx = SimpVect.size(); 636 SimpVect.push_back(ThisAddend); 637 638 // The inner loop collects addends sharing same symbolic-value, and these 639 // addends will be later on folded into a single addend. Following above 640 // example, if the symbolic value "y" is being processed, the inner loop 641 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will 642 // be later on folded into "<b1+b2, y>". 643 // 644 for (unsigned SameSymIdx = SymIdx + 1; 645 SameSymIdx < AddendNum; SameSymIdx++) { 646 const FAddend *T = Addends[SameSymIdx]; 647 if (T && T->getSymVal() == Val) { 648 // Set null such that next iteration of the outer loop will not process 649 // this addend again. 650 Addends[SameSymIdx] = 0; 651 SimpVect.push_back(T); 652 } 653 } 654 655 // If multiple addends share same symbolic value, fold them together. 656 if (StartIdx + 1 != SimpVect.size()) { 657 FAddend &R = TmpResult[NextTmpIdx ++]; 658 R = *SimpVect[StartIdx]; 659 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++) 660 R += *SimpVect[Idx]; 661 662 // Pop all addends being folded and push the resulting folded addend. 663 SimpVect.resize(StartIdx); 664 if (Val != 0) { 665 if (!R.isZero()) { 666 SimpVect.push_back(&R); 667 } 668 } else { 669 // Don't push constant addend at this time. It will be the last element 670 // of <SimpVect>. 671 ConstAdd = &R; 672 } 673 } 674 } 675 676 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) && 677 "out-of-bound access"); 678 679 if (ConstAdd) 680 SimpVect.push_back(ConstAdd); 681 682 Value *Result; 683 if (!SimpVect.empty()) 684 Result = createNaryFAdd(SimpVect, InstrQuota); 685 else { 686 // The addition is folded to 0.0. 687 Result = ConstantFP::get(Instr->getType(), 0.0); 688 } 689 690 return Result; 691} 692 693Value *FAddCombine::createNaryFAdd 694 (const AddendVect &Opnds, unsigned InstrQuota) { 695 assert(!Opnds.empty() && "Expect at least one addend"); 696 697 // Step 1: Check if the # of instructions needed exceeds the quota. 698 // 699 unsigned InstrNeeded = calcInstrNumber(Opnds); 700 if (InstrNeeded > InstrQuota) 701 return 0; 702 703 initCreateInstNum(); 704 705 // step 2: Emit the N-ary addition. 706 // Note that at most three instructions are involved in Fadd-InstCombine: the 707 // addition in question, and at most two neighboring instructions. 708 // The resulting optimized addition should have at least one less instruction 709 // than the original addition expression tree. This implies that the resulting 710 // N-ary addition has at most two instructions, and we don't need to worry 711 // about tree-height when constructing the N-ary addition. 712 713 Value *LastVal = 0; 714 bool LastValNeedNeg = false; 715 716 // Iterate the addends, creating fadd/fsub using adjacent two addends. 717 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 718 I != E; I++) { 719 bool NeedNeg; 720 Value *V = createAddendVal(**I, NeedNeg); 721 if (!LastVal) { 722 LastVal = V; 723 LastValNeedNeg = NeedNeg; 724 continue; 725 } 726 727 if (LastValNeedNeg == NeedNeg) { 728 LastVal = createFAdd(LastVal, V); 729 continue; 730 } 731 732 if (LastValNeedNeg) 733 LastVal = createFSub(V, LastVal); 734 else 735 LastVal = createFSub(LastVal, V); 736 737 LastValNeedNeg = false; 738 } 739 740 if (LastValNeedNeg) { 741 LastVal = createFNeg(LastVal); 742 } 743 744 #ifndef NDEBUG 745 assert(CreateInstrNum == InstrNeeded && 746 "Inconsistent in instruction numbers"); 747 #endif 748 749 return LastVal; 750} 751 752Value *FAddCombine::createFSub 753 (Value *Opnd0, Value *Opnd1) { 754 Value *V = Builder->CreateFSub(Opnd0, Opnd1); 755 if (Instruction *I = dyn_cast<Instruction>(V)) 756 createInstPostProc(I); 757 return V; 758} 759 760Value *FAddCombine::createFNeg(Value *V) { 761 Value *Zero = cast<Value>(ConstantFP::get(V->getType(), 0.0)); 762 Value *NewV = createFSub(Zero, V); 763 if (Instruction *I = dyn_cast<Instruction>(NewV)) 764 createInstPostProc(I, true); // fneg's don't receive instruction numbers. 765 return NewV; 766} 767 768Value *FAddCombine::createFAdd 769 (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 776Value *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 783Value *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 790void FAddCombine::createInstPostProc(Instruction *NewInstr, 791 bool NoNumber) { 792 NewInstr->setDebugLoc(Instr->getDebugLoc()); 793 794 // Keep track of the number of instruction created. 795 if (!NoNumber) 796 incCreateInstNum(); 797 798 // Propagate fast-math flags 799 NewInstr->setFastMathFlags(Instr->getFastMathFlags()); 800} 801 802// Return the number of instruction needed to emit the N-ary addition. 803// NOTE: Keep this function in sync with createAddendVal(). 804unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) { 805 unsigned OpndNum = Opnds.size(); 806 unsigned InstrNeeded = OpndNum - 1; 807 808 // The number of addends in the form of "(-1)*x". 809 unsigned NegOpndNum = 0; 810 811 // Adjust the number of instructions needed to emit the N-ary add. 812 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 813 I != E; I++) { 814 const FAddend *Opnd = *I; 815 if (Opnd->isConstant()) 816 continue; 817 818 const FAddendCoef &CE = Opnd->getCoef(); 819 if (CE.isMinusOne() || CE.isMinusTwo()) 820 NegOpndNum++; 821 822 // Let the addend be "c * x". If "c == +/-1", the value of the addend 823 // is immediately available; otherwise, it needs exactly one instruction 824 // to evaluate the value. 825 if (!CE.isMinusOne() && !CE.isOne()) 826 InstrNeeded++; 827 } 828 if (NegOpndNum == OpndNum) 829 InstrNeeded++; 830 return InstrNeeded; 831} 832 833// Input Addend Value NeedNeg(output) 834// ================================================================ 835// Constant C C false 836// <+/-1, V> V coefficient is -1 837// <2/-2, V> "fadd V, V" coefficient is -2 838// <C, V> "fmul V, C" false 839// 840// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber. 841Value *FAddCombine::createAddendVal 842 (const FAddend &Opnd, bool &NeedNeg) { 843 const FAddendCoef &Coeff = Opnd.getCoef(); 844 845 if (Opnd.isConstant()) { 846 NeedNeg = false; 847 return Coeff.getValue(Instr->getType()); 848 } 849 850 Value *OpndVal = Opnd.getSymVal(); 851 852 if (Coeff.isMinusOne() || Coeff.isOne()) { 853 NeedNeg = Coeff.isMinusOne(); 854 return OpndVal; 855 } 856 857 if (Coeff.isTwo() || Coeff.isMinusTwo()) { 858 NeedNeg = Coeff.isMinusTwo(); 859 return createFAdd(OpndVal, OpndVal); 860 } 861 862 NeedNeg = false; 863 return createFMul(OpndVal, Coeff.getValue(Instr->getType())); 864} 865 866// dyn_castFoldableMul - If this value is a multiply that can be folded into 867// other computations (because it has a constant operand), return the 868// non-constant operand of the multiply, and set CST to point to the multiplier. 869// Otherwise, return null. 870// 871static inline Value *dyn_castFoldableMul(Value *V, Constant *&CST) { 872 if (!V->hasOneUse() || !V->getType()->isIntOrIntVectorTy()) 873 return 0; 874 875 Instruction *I = dyn_cast<Instruction>(V); 876 if (I == 0) return 0; 877 878 if (I->getOpcode() == Instruction::Mul) 879 if ((CST = dyn_cast<Constant>(I->getOperand(1)))) 880 return I->getOperand(0); 881 if (I->getOpcode() == Instruction::Shl) 882 if ((CST = dyn_cast<Constant>(I->getOperand(1)))) { 883 // The multiplier is really 1 << CST. 884 CST = ConstantExpr::getShl(ConstantInt::get(V->getType(), 1), CST); 885 return I->getOperand(0); 886 } 887 return 0; 888} 889 890 891/// WillNotOverflowSignedAdd - Return true if we can prove that: 892/// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS)) 893/// This basically requires proving that the add in the original type would not 894/// overflow to change the sign bit or have a carry out. 895bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) { 896 // There are different heuristics we can use for this. Here are some simple 897 // ones. 898 899 // Add has the property that adding any two 2's complement numbers can only 900 // have one carry bit which can change a sign. As such, if LHS and RHS each 901 // have at least two sign bits, we know that the addition of the two values 902 // will sign extend fine. 903 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1) 904 return true; 905 906 907 // If one of the operands only has one non-zero bit, and if the other operand 908 // has a known-zero bit in a more significant place than it (not including the 909 // sign bit) the ripple may go up to and fill the zero, but won't change the 910 // sign. For example, (X & ~4) + 1. 911 912 // TODO: Implement. 913 914 return false; 915} 916 917Instruction *InstCombiner::visitAdd(BinaryOperator &I) { 918 bool Changed = SimplifyAssociativeOrCommutative(I); 919 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 920 921 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(), 922 I.hasNoUnsignedWrap(), DL)) 923 return ReplaceInstUsesWith(I, V); 924 925 // (A*B)+(A*C) -> A*(B+C) etc 926 if (Value *V = SimplifyUsingDistributiveLaws(I)) 927 return ReplaceInstUsesWith(I, V); 928 929 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 930 // X + (signbit) --> X ^ signbit 931 const APInt &Val = CI->getValue(); 932 if (Val.isSignBit()) 933 return BinaryOperator::CreateXor(LHS, RHS); 934 935 // See if SimplifyDemandedBits can simplify this. This handles stuff like 936 // (X & 254)+1 -> (X&254)|1 937 if (SimplifyDemandedInstructionBits(I)) 938 return &I; 939 940 // zext(bool) + C -> bool ? C + 1 : C 941 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS)) 942 if (ZI->getSrcTy()->isIntegerTy(1)) 943 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI); 944 945 Value *XorLHS = 0; ConstantInt *XorRHS = 0; 946 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) { 947 uint32_t TySizeBits = I.getType()->getScalarSizeInBits(); 948 const APInt &RHSVal = CI->getValue(); 949 unsigned ExtendAmt = 0; 950 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. 951 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. 952 if (XorRHS->getValue() == -RHSVal) { 953 if (RHSVal.isPowerOf2()) 954 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1; 955 else if (XorRHS->getValue().isPowerOf2()) 956 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1; 957 } 958 959 if (ExtendAmt) { 960 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt); 961 if (!MaskedValueIsZero(XorLHS, Mask)) 962 ExtendAmt = 0; 963 } 964 965 if (ExtendAmt) { 966 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt); 967 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext"); 968 return BinaryOperator::CreateAShr(NewShl, ShAmt); 969 } 970 971 // If this is a xor that was canonicalized from a sub, turn it back into 972 // a sub and fuse this add with it. 973 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) { 974 IntegerType *IT = cast<IntegerType>(I.getType()); 975 APInt LHSKnownOne(IT->getBitWidth(), 0); 976 APInt LHSKnownZero(IT->getBitWidth(), 0); 977 ComputeMaskedBits(XorLHS, LHSKnownZero, LHSKnownOne); 978 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue()) 979 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI), 980 XorLHS); 981 } 982 // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C, 983 // transform them into (X + (signbit ^ C)) 984 if (XorRHS->getValue().isSignBit()) 985 return BinaryOperator::CreateAdd(XorLHS, 986 ConstantExpr::getXor(XorRHS, CI)); 987 } 988 } 989 990 if (isa<Constant>(RHS) && isa<PHINode>(LHS)) 991 if (Instruction *NV = FoldOpIntoPhi(I)) 992 return NV; 993 994 if (I.getType()->getScalarType()->isIntegerTy(1)) 995 return BinaryOperator::CreateXor(LHS, RHS); 996 997 // X + X --> X << 1 998 if (LHS == RHS) { 999 BinaryOperator *New = 1000 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1)); 1001 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 1002 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1003 return New; 1004 } 1005 1006 // -A + B --> B - A 1007 // -A + -B --> -(A + B) 1008 if (Value *LHSV = dyn_castNegVal(LHS)) { 1009 if (!isa<Constant>(RHS)) 1010 if (Value *RHSV = dyn_castNegVal(RHS)) { 1011 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum"); 1012 return BinaryOperator::CreateNeg(NewAdd); 1013 } 1014 1015 return BinaryOperator::CreateSub(RHS, LHSV); 1016 } 1017 1018 // A + -B --> A - B 1019 if (!isa<Constant>(RHS)) 1020 if (Value *V = dyn_castNegVal(RHS)) 1021 return BinaryOperator::CreateSub(LHS, V); 1022 1023 1024 { 1025 Constant *C2; 1026 if (Value *X = dyn_castFoldableMul(LHS, C2)) { 1027 if (X == RHS) // X*C + X --> X * (C+1) 1028 return BinaryOperator::CreateMul(RHS, AddOne(C2)); 1029 1030 // X*C1 + X*C2 --> X * (C1+C2) 1031 Constant *C1; 1032 if (X == dyn_castFoldableMul(RHS, C1)) 1033 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2)); 1034 } 1035 1036 // X + X*C --> X * (C+1) 1037 if (dyn_castFoldableMul(RHS, C2) == LHS) 1038 return BinaryOperator::CreateMul(LHS, AddOne(C2)); 1039 } 1040 1041 // A+B --> A|B iff A and B have no bits set in common. 1042 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) { 1043 APInt LHSKnownOne(IT->getBitWidth(), 0); 1044 APInt LHSKnownZero(IT->getBitWidth(), 0); 1045 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 1046 if (LHSKnownZero != 0) { 1047 APInt RHSKnownOne(IT->getBitWidth(), 0); 1048 APInt RHSKnownZero(IT->getBitWidth(), 0); 1049 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 1050 1051 // No bits in common -> bitwise or. 1052 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue()) 1053 return BinaryOperator::CreateOr(LHS, RHS); 1054 } 1055 } 1056 1057 // W*X + Y*Z --> W * (X+Z) iff W == Y 1058 { 1059 Value *W, *X, *Y, *Z; 1060 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) && 1061 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) { 1062 if (W != Y) { 1063 if (W == Z) { 1064 std::swap(Y, Z); 1065 } else if (Y == X) { 1066 std::swap(W, X); 1067 } else if (X == Z) { 1068 std::swap(Y, Z); 1069 std::swap(W, X); 1070 } 1071 } 1072 1073 if (W == Y) { 1074 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName()); 1075 return BinaryOperator::CreateMul(W, NewAdd); 1076 } 1077 } 1078 } 1079 1080 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 1081 Value *X; 1082 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X 1083 return BinaryOperator::CreateSub(SubOne(CRHS), X); 1084 } 1085 1086 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) { 1087 // (X & FF00) + xx00 -> (X+xx00) & FF00 1088 Value *X; 1089 ConstantInt *C2; 1090 if (LHS->hasOneUse() && 1091 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) && 1092 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) { 1093 // See if all bits from the first bit set in the Add RHS up are included 1094 // in the mask. First, get the rightmost bit. 1095 const APInt &AddRHSV = CRHS->getValue(); 1096 1097 // Form a mask of all bits from the lowest bit added through the top. 1098 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1)); 1099 1100 // See if the and mask includes all of these bits. 1101 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue()); 1102 1103 if (AddRHSHighBits == AddRHSHighBitsAnd) { 1104 // Okay, the xform is safe. Insert the new add pronto. 1105 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName()); 1106 return BinaryOperator::CreateAnd(NewAdd, C2); 1107 } 1108 } 1109 1110 // Try to fold constant add into select arguments. 1111 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1112 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1113 return R; 1114 } 1115 1116 // add (select X 0 (sub n A)) A --> select X A n 1117 { 1118 SelectInst *SI = dyn_cast<SelectInst>(LHS); 1119 Value *A = RHS; 1120 if (!SI) { 1121 SI = dyn_cast<SelectInst>(RHS); 1122 A = LHS; 1123 } 1124 if (SI && SI->hasOneUse()) { 1125 Value *TV = SI->getTrueValue(); 1126 Value *FV = SI->getFalseValue(); 1127 Value *N; 1128 1129 // Can we fold the add into the argument of the select? 1130 // We check both true and false select arguments for a matching subtract. 1131 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A)))) 1132 // Fold the add into the true select value. 1133 return SelectInst::Create(SI->getCondition(), N, A); 1134 1135 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A)))) 1136 // Fold the add into the false select value. 1137 return SelectInst::Create(SI->getCondition(), A, N); 1138 } 1139 } 1140 1141 // Check for (add (sext x), y), see if we can merge this into an 1142 // integer add followed by a sext. 1143 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) { 1144 // (add (sext x), cst) --> (sext (add x, cst')) 1145 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 1146 Constant *CI = 1147 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType()); 1148 if (LHSConv->hasOneUse() && 1149 ConstantExpr::getSExt(CI, I.getType()) == RHSC && 1150 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1151 // Insert the new, smaller add. 1152 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1153 CI, "addconv"); 1154 return new SExtInst(NewAdd, I.getType()); 1155 } 1156 } 1157 1158 // (add (sext x), (sext y)) --> (sext (add int x, y)) 1159 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) { 1160 // Only do this if x/y have the same type, if at last one of them has a 1161 // single use (so we don't increase the number of sexts), and if the 1162 // integer add will not overflow. 1163 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1164 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1165 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1166 RHSConv->getOperand(0))) { 1167 // Insert the new integer add. 1168 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1169 RHSConv->getOperand(0), "addconv"); 1170 return new SExtInst(NewAdd, I.getType()); 1171 } 1172 } 1173 } 1174 1175 // Check for (x & y) + (x ^ y) 1176 { 1177 Value *A = 0, *B = 0; 1178 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) && 1179 (match(LHS, m_And(m_Specific(A), m_Specific(B))) || 1180 match(LHS, m_And(m_Specific(B), m_Specific(A))))) 1181 return BinaryOperator::CreateOr(A, B); 1182 1183 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) && 1184 (match(RHS, m_And(m_Specific(A), m_Specific(B))) || 1185 match(RHS, m_And(m_Specific(B), m_Specific(A))))) 1186 return BinaryOperator::CreateOr(A, B); 1187 } 1188 1189 return Changed ? &I : 0; 1190} 1191 1192Instruction *InstCombiner::visitFAdd(BinaryOperator &I) { 1193 bool Changed = SimplifyAssociativeOrCommutative(I); 1194 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 1195 1196 if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL)) 1197 return ReplaceInstUsesWith(I, V); 1198 1199 if (isa<Constant>(RHS)) { 1200 if (isa<PHINode>(LHS)) 1201 if (Instruction *NV = FoldOpIntoPhi(I)) 1202 return NV; 1203 1204 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1205 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1206 return NV; 1207 } 1208 1209 // -A + B --> B - A 1210 // -A + -B --> -(A + B) 1211 if (Value *LHSV = dyn_castFNegVal(LHS)) { 1212 Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV); 1213 RI->copyFastMathFlags(&I); 1214 return RI; 1215 } 1216 1217 // A + -B --> A - B 1218 if (!isa<Constant>(RHS)) 1219 if (Value *V = dyn_castFNegVal(RHS)) { 1220 Instruction *RI = BinaryOperator::CreateFSub(LHS, V); 1221 RI->copyFastMathFlags(&I); 1222 return RI; 1223 } 1224 1225 // Check for (fadd double (sitofp x), y), see if we can merge this into an 1226 // integer add followed by a promotion. 1227 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) { 1228 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst)) 1229 // ... if the constant fits in the integer value. This is useful for things 1230 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer 1231 // requires a constant pool load, and generally allows the add to be better 1232 // instcombined. 1233 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 1234 Constant *CI = 1235 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType()); 1236 if (LHSConv->hasOneUse() && 1237 ConstantExpr::getSIToFP(CI, I.getType()) == CFP && 1238 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1239 // Insert the new integer add. 1240 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1241 CI, "addconv"); 1242 return new SIToFPInst(NewAdd, I.getType()); 1243 } 1244 } 1245 1246 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y)) 1247 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) { 1248 // Only do this if x/y have the same type, if at last one of them has a 1249 // single use (so we don't increase the number of int->fp conversions), 1250 // and if the integer add will not overflow. 1251 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1252 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1253 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1254 RHSConv->getOperand(0))) { 1255 // Insert the new integer add. 1256 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1257 RHSConv->getOperand(0),"addconv"); 1258 return new SIToFPInst(NewAdd, I.getType()); 1259 } 1260 } 1261 } 1262 1263 // select C, 0, B + select C, A, 0 -> select C, A, B 1264 { 1265 Value *A1, *B1, *C1, *A2, *B2, *C2; 1266 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) && 1267 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) { 1268 if (C1 == C2) { 1269 Constant *Z1=0, *Z2=0; 1270 Value *A, *B, *C=C1; 1271 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) { 1272 Z1 = dyn_cast<Constant>(A1); A = A2; 1273 Z2 = dyn_cast<Constant>(B2); B = B1; 1274 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) { 1275 Z1 = dyn_cast<Constant>(B1); B = B2; 1276 Z2 = dyn_cast<Constant>(A2); A = A1; 1277 } 1278 1279 if (Z1 && Z2 && 1280 (I.hasNoSignedZeros() || 1281 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) { 1282 return SelectInst::Create(C, A, B); 1283 } 1284 } 1285 } 1286 } 1287 1288 if (I.hasUnsafeAlgebra()) { 1289 if (Value *V = FAddCombine(Builder).simplify(&I)) 1290 return ReplaceInstUsesWith(I, V); 1291 } 1292 1293 return Changed ? &I : 0; 1294} 1295 1296 1297/// Optimize pointer differences into the same array into a size. Consider: 1298/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer 1299/// operands to the ptrtoint instructions for the LHS/RHS of the subtract. 1300/// 1301Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS, 1302 Type *Ty) { 1303 assert(DL && "Must have target data info for this"); 1304 1305 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize 1306 // this. 1307 bool Swapped = false; 1308 GEPOperator *GEP1 = 0, *GEP2 = 0; 1309 1310 // For now we require one side to be the base pointer "A" or a constant 1311 // GEP derived from it. 1312 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1313 // (gep X, ...) - X 1314 if (LHSGEP->getOperand(0) == RHS) { 1315 GEP1 = LHSGEP; 1316 Swapped = false; 1317 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1318 // (gep X, ...) - (gep X, ...) 1319 if (LHSGEP->getOperand(0)->stripPointerCasts() == 1320 RHSGEP->getOperand(0)->stripPointerCasts()) { 1321 GEP2 = RHSGEP; 1322 GEP1 = LHSGEP; 1323 Swapped = false; 1324 } 1325 } 1326 } 1327 1328 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1329 // X - (gep X, ...) 1330 if (RHSGEP->getOperand(0) == LHS) { 1331 GEP1 = RHSGEP; 1332 Swapped = true; 1333 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1334 // (gep X, ...) - (gep X, ...) 1335 if (RHSGEP->getOperand(0)->stripPointerCasts() == 1336 LHSGEP->getOperand(0)->stripPointerCasts()) { 1337 GEP2 = LHSGEP; 1338 GEP1 = RHSGEP; 1339 Swapped = true; 1340 } 1341 } 1342 } 1343 1344 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and 1345 // multiple users. 1346 if (GEP1 == 0 || 1347 (GEP2 != 0 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse())) 1348 return 0; 1349 1350 // Emit the offset of the GEP and an intptr_t. 1351 Value *Result = EmitGEPOffset(GEP1); 1352 1353 // If we had a constant expression GEP on the other side offsetting the 1354 // pointer, subtract it from the offset we have. 1355 if (GEP2) { 1356 Value *Offset = EmitGEPOffset(GEP2); 1357 Result = Builder->CreateSub(Result, Offset); 1358 } 1359 1360 // If we have p - gep(p, ...) then we have to negate the result. 1361 if (Swapped) 1362 Result = Builder->CreateNeg(Result, "diff.neg"); 1363 1364 return Builder->CreateIntCast(Result, Ty, true); 1365} 1366 1367 1368Instruction *InstCombiner::visitSub(BinaryOperator &I) { 1369 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1370 1371 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(), 1372 I.hasNoUnsignedWrap(), DL)) 1373 return ReplaceInstUsesWith(I, V); 1374 1375 // (A*B)-(A*C) -> A*(B-C) etc 1376 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1377 return ReplaceInstUsesWith(I, V); 1378 1379 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW. 1380 if (Value *V = dyn_castNegVal(Op1)) { 1381 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V); 1382 Res->setHasNoSignedWrap(I.hasNoSignedWrap()); 1383 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1384 return Res; 1385 } 1386 1387 if (I.getType()->isIntegerTy(1)) 1388 return BinaryOperator::CreateXor(Op0, Op1); 1389 1390 // Replace (-1 - A) with (~A). 1391 if (match(Op0, m_AllOnes())) 1392 return BinaryOperator::CreateNot(Op1); 1393 1394 if (Constant *C = dyn_cast<Constant>(Op0)) { 1395 // C - ~X == X + (1+C) 1396 Value *X = 0; 1397 if (match(Op1, m_Not(m_Value(X)))) 1398 return BinaryOperator::CreateAdd(X, AddOne(C)); 1399 1400 // Try to fold constant sub into select arguments. 1401 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1402 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1403 return R; 1404 1405 // C-(X+C2) --> (C-C2)-X 1406 Constant *C2; 1407 if (match(Op1, m_Add(m_Value(X), m_Constant(C2)))) 1408 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X); 1409 1410 if (SimplifyDemandedInstructionBits(I)) 1411 return &I; 1412 1413 // Fold (sub 0, (zext bool to B)) --> (sext bool to B) 1414 if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X)))) 1415 if (X->getType()->getScalarType()->isIntegerTy(1)) 1416 return CastInst::CreateSExtOrBitCast(X, Op1->getType()); 1417 1418 // Fold (sub 0, (sext bool to B)) --> (zext bool to B) 1419 if (C->isNullValue() && match(Op1, m_SExt(m_Value(X)))) 1420 if (X->getType()->getScalarType()->isIntegerTy(1)) 1421 return CastInst::CreateZExtOrBitCast(X, Op1->getType()); 1422 } 1423 1424 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) { 1425 // -(X >>u 31) -> (X >>s 31) 1426 // -(X >>s 31) -> (X >>u 31) 1427 if (C->isZero()) { 1428 Value *X; ConstantInt *CI; 1429 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) && 1430 // Verify we are shifting out everything but the sign bit. 1431 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1432 return BinaryOperator::CreateAShr(X, CI); 1433 1434 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) && 1435 // Verify we are shifting out everything but the sign bit. 1436 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1437 return BinaryOperator::CreateLShr(X, CI); 1438 } 1439 } 1440 1441 1442 { Value *Y; 1443 // X-(X+Y) == -Y X-(Y+X) == -Y 1444 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) || 1445 match(Op1, m_Add(m_Value(Y), m_Specific(Op0)))) 1446 return BinaryOperator::CreateNeg(Y); 1447 1448 // (X-Y)-X == -Y 1449 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y)))) 1450 return BinaryOperator::CreateNeg(Y); 1451 } 1452 1453 if (Op1->hasOneUse()) { 1454 Value *X = 0, *Y = 0, *Z = 0; 1455 Constant *C = 0; 1456 Constant *CI = 0; 1457 1458 // (X - (Y - Z)) --> (X + (Z - Y)). 1459 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z)))) 1460 return BinaryOperator::CreateAdd(Op0, 1461 Builder->CreateSub(Z, Y, Op1->getName())); 1462 1463 // (X - (X & Y)) --> (X & ~Y) 1464 // 1465 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) || 1466 match(Op1, m_And(m_Specific(Op0), m_Value(Y)))) 1467 return BinaryOperator::CreateAnd(Op0, 1468 Builder->CreateNot(Y, Y->getName() + ".not")); 1469 1470 // 0 - (X sdiv C) -> (X sdiv -C) 1471 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && 1472 match(Op0, m_Zero())) 1473 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C)); 1474 1475 // 0 - (X << Y) -> (-X << Y) when X is freely negatable. 1476 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero())) 1477 if (Value *XNeg = dyn_castNegVal(X)) 1478 return BinaryOperator::CreateShl(XNeg, Y); 1479 1480 // X - X*C --> X * (1-C) 1481 if (match(Op1, m_Mul(m_Specific(Op0), m_Constant(CI)))) { 1482 Constant *CP1 = ConstantExpr::getSub(ConstantInt::get(I.getType(),1), CI); 1483 return BinaryOperator::CreateMul(Op0, CP1); 1484 } 1485 1486 // X - X<<C --> X * (1-(1<<C)) 1487 if (match(Op1, m_Shl(m_Specific(Op0), m_Constant(CI)))) { 1488 Constant *One = ConstantInt::get(I.getType(), 1); 1489 C = ConstantExpr::getSub(One, ConstantExpr::getShl(One, CI)); 1490 return BinaryOperator::CreateMul(Op0, C); 1491 } 1492 1493 // X - A*-B -> X + A*B 1494 // X - -A*B -> X + A*B 1495 Value *A, *B; 1496 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) || 1497 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B)))) 1498 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B)); 1499 1500 // X - A*CI -> X + A*-CI 1501 // X - CI*A -> X + A*-CI 1502 if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) || 1503 match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) { 1504 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI)); 1505 return BinaryOperator::CreateAdd(Op0, NewMul); 1506 } 1507 } 1508 1509 Constant *C1; 1510 if (Value *X = dyn_castFoldableMul(Op0, C1)) { 1511 if (X == Op1) // X*C - X --> X * (C-1) 1512 return BinaryOperator::CreateMul(Op1, SubOne(C1)); 1513 1514 Constant *C2; // X*C1 - X*C2 -> X * (C1-C2) 1515 if (X == dyn_castFoldableMul(Op1, C2)) 1516 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2)); 1517 } 1518 1519 // Optimize pointer differences into the same array into a size. Consider: 1520 // &A[10] - &A[0]: we should compile this to "10". 1521 if (DL) { 1522 Value *LHSOp, *RHSOp; 1523 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) && 1524 match(Op1, m_PtrToInt(m_Value(RHSOp)))) 1525 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1526 return ReplaceInstUsesWith(I, Res); 1527 1528 // trunc(p)-trunc(q) -> trunc(p-q) 1529 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) && 1530 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp))))) 1531 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1532 return ReplaceInstUsesWith(I, Res); 1533 } 1534 1535 return 0; 1536} 1537 1538Instruction *InstCombiner::visitFSub(BinaryOperator &I) { 1539 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1540 1541 if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL)) 1542 return ReplaceInstUsesWith(I, V); 1543 1544 if (isa<Constant>(Op0)) 1545 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1546 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1547 return NV; 1548 1549 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking 1550 // through FP extensions/truncations along the way. 1551 if (Value *V = dyn_castFNegVal(Op1)) { 1552 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V); 1553 NewI->copyFastMathFlags(&I); 1554 return NewI; 1555 } 1556 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) { 1557 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) { 1558 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType()); 1559 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc); 1560 NewI->copyFastMathFlags(&I); 1561 return NewI; 1562 } 1563 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) { 1564 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) { 1565 Value *NewExt = Builder->CreateFPExt(V, I.getType()); 1566 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt); 1567 NewI->copyFastMathFlags(&I); 1568 return NewI; 1569 } 1570 } 1571 1572 if (I.hasUnsafeAlgebra()) { 1573 if (Value *V = FAddCombine(Builder).simplify(&I)) 1574 return ReplaceInstUsesWith(I, V); 1575 } 1576 1577 return 0; 1578} 1579