InstCombineAddSub.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
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 23#define DEBUG_TYPE "instcombine" 24 25namespace { 26 27 /// Class representing coefficient of floating-point addend. 28 /// This class needs to be highly efficient, which is especially true for 29 /// the constructor. As of I write this comment, the cost of the default 30 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to 31 /// perform write-merging). 32 /// 33 class FAddendCoef { 34 public: 35 // The constructor has to initialize a APFloat, which is uncessary for 36 // most addends which have coefficient either 1 or -1. So, the constructor 37 // is expensive. In order to avoid the cost of the constructor, we should 38 // reuse some instances whenever possible. The pre-created instances 39 // FAddCombine::Add[0-5] embodies this idea. 40 // 41 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {} 42 ~FAddendCoef(); 43 44 void set(short C) { 45 assert(!insaneIntVal(C) && "Insane coefficient"); 46 IsFp = false; IntVal = C; 47 } 48 49 void set(const APFloat& C); 50 51 void negate(); 52 53 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); } 54 Value *getValue(Type *) const; 55 56 // If possible, don't define operator+/operator- etc because these 57 // operators inevitably call FAddendCoef's constructor which is not cheap. 58 void operator=(const FAddendCoef &A); 59 void operator+=(const FAddendCoef &A); 60 void operator-=(const FAddendCoef &A); 61 void operator*=(const FAddendCoef &S); 62 63 bool isOne() const { return isInt() && IntVal == 1; } 64 bool isTwo() const { return isInt() && IntVal == 2; } 65 bool isMinusOne() const { return isInt() && IntVal == -1; } 66 bool isMinusTwo() const { return isInt() && IntVal == -2; } 67 68 private: 69 bool insaneIntVal(int V) { return V > 4 || V < -4; } 70 APFloat *getFpValPtr(void) 71 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); } 72 const APFloat *getFpValPtr(void) const 73 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); } 74 75 const APFloat &getFpVal(void) const { 76 assert(IsFp && BufHasFpVal && "Incorret state"); 77 return *getFpValPtr(); 78 } 79 80 APFloat &getFpVal(void) { 81 assert(IsFp && BufHasFpVal && "Incorret state"); 82 return *getFpValPtr(); 83 } 84 85 bool isInt() const { return !IsFp; } 86 87 // If the coefficient is represented by an integer, promote it to a 88 // floating point. 89 void convertToFpType(const fltSemantics &Sem); 90 91 // Construct an APFloat from a signed integer. 92 // TODO: We should get rid of this function when APFloat can be constructed 93 // from an *SIGNED* integer. 94 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val); 95 private: 96 97 bool IsFp; 98 99 // True iff FpValBuf contains an instance of APFloat. 100 bool BufHasFpVal; 101 102 // The integer coefficient of an individual addend is either 1 or -1, 103 // and we try to simplify at most 4 addends from neighboring at most 104 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt 105 // is overkill of this end. 106 short IntVal; 107 108 AlignedCharArrayUnion<APFloat> FpValBuf; 109 }; 110 111 /// FAddend is used to represent floating-point addend. An addend is 112 /// represented as <C, V>, where the V is a symbolic value, and C is a 113 /// constant coefficient. A constant addend is represented as <C, 0>. 114 /// 115 class FAddend { 116 public: 117 FAddend() { Val = nullptr; } 118 119 Value *getSymVal (void) const { return Val; } 120 const FAddendCoef &getCoef(void) const { return Coeff; } 121 122 bool isConstant() const { return Val == nullptr; } 123 bool isZero() const { return Coeff.isZero(); } 124 125 void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; } 126 void set(const APFloat& Coefficient, Value *V) 127 { Coeff.set(Coefficient); Val = V; } 128 void set(const ConstantFP* Coefficient, Value *V) 129 { Coeff.set(Coefficient->getValueAPF()); Val = V; } 130 131 void negate() { Coeff.negate(); } 132 133 /// Drill down the U-D chain one step to find the definition of V, and 134 /// try to break the definition into one or two addends. 135 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1); 136 137 /// Similar to FAddend::drillDownOneStep() except that the value being 138 /// splitted is the addend itself. 139 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const; 140 141 void operator+=(const FAddend &T) { 142 assert((Val == T.Val) && "Symbolic-values disagree"); 143 Coeff += T.Coeff; 144 } 145 146 private: 147 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; } 148 149 // This addend has the value of "Coeff * Val". 150 Value *Val; 151 FAddendCoef Coeff; 152 }; 153 154 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along 155 /// with its neighboring at most two instructions. 156 /// 157 class FAddCombine { 158 public: 159 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(nullptr) {} 160 Value *simplify(Instruction *FAdd); 161 162 private: 163 typedef SmallVector<const FAddend*, 4> AddendVect; 164 165 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota); 166 167 Value *performFactorization(Instruction *I); 168 169 /// Convert given addend to a Value 170 Value *createAddendVal(const FAddend &A, bool& NeedNeg); 171 172 /// Return the number of instructions needed to emit the N-ary addition. 173 unsigned calcInstrNumber(const AddendVect& Vect); 174 Value *createFSub(Value *Opnd0, Value *Opnd1); 175 Value *createFAdd(Value *Opnd0, Value *Opnd1); 176 Value *createFMul(Value *Opnd0, Value *Opnd1); 177 Value *createFDiv(Value *Opnd0, Value *Opnd1); 178 Value *createFNeg(Value *V); 179 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota); 180 void createInstPostProc(Instruction *NewInst, bool NoNumber = false); 181 182 InstCombiner::BuilderTy *Builder; 183 Instruction *Instr; 184 185 private: 186 // Debugging stuff are clustered here. 187 #ifndef NDEBUG 188 unsigned CreateInstrNum; 189 void initCreateInstNum() { CreateInstrNum = 0; } 190 void incCreateInstNum() { CreateInstrNum++; } 191 #else 192 void initCreateInstNum() {} 193 void incCreateInstNum() {} 194 #endif 195 }; 196} 197 198//===----------------------------------------------------------------------===// 199// 200// Implementation of 201// {FAddendCoef, FAddend, FAddition, FAddCombine}. 202// 203//===----------------------------------------------------------------------===// 204FAddendCoef::~FAddendCoef() { 205 if (BufHasFpVal) 206 getFpValPtr()->~APFloat(); 207} 208 209void FAddendCoef::set(const APFloat& C) { 210 APFloat *P = getFpValPtr(); 211 212 if (isInt()) { 213 // As the buffer is meanless byte stream, we cannot call 214 // APFloat::operator=(). 215 new(P) APFloat(C); 216 } else 217 *P = C; 218 219 IsFp = BufHasFpVal = true; 220} 221 222void FAddendCoef::convertToFpType(const fltSemantics &Sem) { 223 if (!isInt()) 224 return; 225 226 APFloat *P = getFpValPtr(); 227 if (IntVal > 0) 228 new(P) APFloat(Sem, IntVal); 229 else { 230 new(P) APFloat(Sem, 0 - IntVal); 231 P->changeSign(); 232 } 233 IsFp = BufHasFpVal = true; 234} 235 236APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) { 237 if (Val >= 0) 238 return APFloat(Sem, Val); 239 240 APFloat T(Sem, 0 - Val); 241 T.changeSign(); 242 243 return T; 244} 245 246void FAddendCoef::operator=(const FAddendCoef &That) { 247 if (That.isInt()) 248 set(That.IntVal); 249 else 250 set(That.getFpVal()); 251} 252 253void FAddendCoef::operator+=(const FAddendCoef &That) { 254 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 255 if (isInt() == That.isInt()) { 256 if (isInt()) 257 IntVal += That.IntVal; 258 else 259 getFpVal().add(That.getFpVal(), RndMode); 260 return; 261 } 262 263 if (isInt()) { 264 const APFloat &T = That.getFpVal(); 265 convertToFpType(T.getSemantics()); 266 getFpVal().add(T, RndMode); 267 return; 268 } 269 270 APFloat &T = getFpVal(); 271 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode); 272} 273 274void FAddendCoef::operator-=(const FAddendCoef &That) { 275 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven; 276 if (isInt() == That.isInt()) { 277 if (isInt()) 278 IntVal -= That.IntVal; 279 else 280 getFpVal().subtract(That.getFpVal(), RndMode); 281 return; 282 } 283 284 if (isInt()) { 285 const APFloat &T = That.getFpVal(); 286 convertToFpType(T.getSemantics()); 287 getFpVal().subtract(T, RndMode); 288 return; 289 } 290 291 APFloat &T = getFpVal(); 292 T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode); 293} 294 295void FAddendCoef::operator*=(const FAddendCoef &That) { 296 if (That.isOne()) 297 return; 298 299 if (That.isMinusOne()) { 300 negate(); 301 return; 302 } 303 304 if (isInt() && That.isInt()) { 305 int Res = IntVal * (int)That.IntVal; 306 assert(!insaneIntVal(Res) && "Insane int value"); 307 IntVal = Res; 308 return; 309 } 310 311 const fltSemantics &Semantic = 312 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics(); 313 314 if (isInt()) 315 convertToFpType(Semantic); 316 APFloat &F0 = getFpVal(); 317 318 if (That.isInt()) 319 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal), 320 APFloat::rmNearestTiesToEven); 321 else 322 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven); 323 324 return; 325} 326 327void FAddendCoef::negate() { 328 if (isInt()) 329 IntVal = 0 - IntVal; 330 else 331 getFpVal().changeSign(); 332} 333 334Value *FAddendCoef::getValue(Type *Ty) const { 335 return isInt() ? 336 ConstantFP::get(Ty, float(IntVal)) : 337 ConstantFP::get(Ty->getContext(), getFpVal()); 338} 339 340// The definition of <Val> Addends 341// ========================================= 342// A + B <1, A>, <1,B> 343// A - B <1, A>, <1,B> 344// 0 - B <-1, B> 345// C * A, <C, A> 346// A + C <1, A> <C, NULL> 347// 0 +/- 0 <0, NULL> (corner case) 348// 349// Legend: A and B are not constant, C is constant 350// 351unsigned FAddend::drillValueDownOneStep 352 (Value *Val, FAddend &Addend0, FAddend &Addend1) { 353 Instruction *I = nullptr; 354 if (!Val || !(I = dyn_cast<Instruction>(Val))) 355 return 0; 356 357 unsigned Opcode = I->getOpcode(); 358 359 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) { 360 ConstantFP *C0, *C1; 361 Value *Opnd0 = I->getOperand(0); 362 Value *Opnd1 = I->getOperand(1); 363 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero()) 364 Opnd0 = nullptr; 365 366 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero()) 367 Opnd1 = nullptr; 368 369 if (Opnd0) { 370 if (!C0) 371 Addend0.set(1, Opnd0); 372 else 373 Addend0.set(C0, nullptr); 374 } 375 376 if (Opnd1) { 377 FAddend &Addend = Opnd0 ? Addend1 : Addend0; 378 if (!C1) 379 Addend.set(1, Opnd1); 380 else 381 Addend.set(C1, nullptr); 382 if (Opcode == Instruction::FSub) 383 Addend.negate(); 384 } 385 386 if (Opnd0 || Opnd1) 387 return Opnd0 && Opnd1 ? 2 : 1; 388 389 // Both operands are zero. Weird! 390 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr); 391 return 1; 392 } 393 394 if (I->getOpcode() == Instruction::FMul) { 395 Value *V0 = I->getOperand(0); 396 Value *V1 = I->getOperand(1); 397 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) { 398 Addend0.set(C, V1); 399 return 1; 400 } 401 402 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) { 403 Addend0.set(C, V0); 404 return 1; 405 } 406 } 407 408 return 0; 409} 410 411// Try to break *this* addend into two addends. e.g. Suppose this addend is 412// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends, 413// i.e. <2.3, X> and <2.3, Y>. 414// 415unsigned FAddend::drillAddendDownOneStep 416 (FAddend &Addend0, FAddend &Addend1) const { 417 if (isConstant()) 418 return 0; 419 420 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1); 421 if (!BreakNum || Coeff.isOne()) 422 return BreakNum; 423 424 Addend0.Scale(Coeff); 425 426 if (BreakNum == 2) 427 Addend1.Scale(Coeff); 428 429 return BreakNum; 430} 431 432// Try to perform following optimization on the input instruction I. Return the 433// simplified expression if was successful; otherwise, return 0. 434// 435// Instruction "I" is Simplified into 436// ------------------------------------------------------- 437// (x * y) +/- (x * z) x * (y +/- z) 438// (y / x) +/- (z / x) (y +/- z) / x 439// 440Value *FAddCombine::performFactorization(Instruction *I) { 441 assert((I->getOpcode() == Instruction::FAdd || 442 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 443 444 Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0)); 445 Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1)); 446 447 if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode()) 448 return nullptr; 449 450 bool isMpy = false; 451 if (I0->getOpcode() == Instruction::FMul) 452 isMpy = true; 453 else if (I0->getOpcode() != Instruction::FDiv) 454 return nullptr; 455 456 Value *Opnd0_0 = I0->getOperand(0); 457 Value *Opnd0_1 = I0->getOperand(1); 458 Value *Opnd1_0 = I1->getOperand(0); 459 Value *Opnd1_1 = I1->getOperand(1); 460 461 // Input Instr I Factor AddSub0 AddSub1 462 // ---------------------------------------------- 463 // (x*y) +/- (x*z) x y z 464 // (y/x) +/- (z/x) x y z 465 // 466 Value *Factor = nullptr; 467 Value *AddSub0 = nullptr, *AddSub1 = nullptr; 468 469 if (isMpy) { 470 if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1) 471 Factor = Opnd0_0; 472 else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1) 473 Factor = Opnd0_1; 474 475 if (Factor) { 476 AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0; 477 AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0; 478 } 479 } else if (Opnd0_1 == Opnd1_1) { 480 Factor = Opnd0_1; 481 AddSub0 = Opnd0_0; 482 AddSub1 = Opnd1_0; 483 } 484 485 if (!Factor) 486 return nullptr; 487 488 FastMathFlags Flags; 489 Flags.setUnsafeAlgebra(); 490 if (I0) Flags &= I->getFastMathFlags(); 491 if (I1) Flags &= I->getFastMathFlags(); 492 493 // Create expression "NewAddSub = AddSub0 +/- AddsSub1" 494 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ? 495 createFAdd(AddSub0, AddSub1) : 496 createFSub(AddSub0, AddSub1); 497 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) { 498 const APFloat &F = CFP->getValueAPF(); 499 if (!F.isNormal()) 500 return nullptr; 501 } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub)) 502 II->setFastMathFlags(Flags); 503 504 if (isMpy) { 505 Value *RI = createFMul(Factor, NewAddSub); 506 if (Instruction *II = dyn_cast<Instruction>(RI)) 507 II->setFastMathFlags(Flags); 508 return RI; 509 } 510 511 Value *RI = createFDiv(NewAddSub, Factor); 512 if (Instruction *II = dyn_cast<Instruction>(RI)) 513 II->setFastMathFlags(Flags); 514 return RI; 515} 516 517Value *FAddCombine::simplify(Instruction *I) { 518 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode"); 519 520 // Currently we are not able to handle vector type. 521 if (I->getType()->isVectorTy()) 522 return nullptr; 523 524 assert((I->getOpcode() == Instruction::FAdd || 525 I->getOpcode() == Instruction::FSub) && "Expect add/sub"); 526 527 // Save the instruction before calling other member-functions. 528 Instr = I; 529 530 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1; 531 532 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1); 533 534 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1. 535 unsigned Opnd0_ExpNum = 0; 536 unsigned Opnd1_ExpNum = 0; 537 538 if (!Opnd0.isConstant()) 539 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1); 540 541 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1. 542 if (OpndNum == 2 && !Opnd1.isConstant()) 543 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1); 544 545 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1 546 if (Opnd0_ExpNum && Opnd1_ExpNum) { 547 AddendVect AllOpnds; 548 AllOpnds.push_back(&Opnd0_0); 549 AllOpnds.push_back(&Opnd1_0); 550 if (Opnd0_ExpNum == 2) 551 AllOpnds.push_back(&Opnd0_1); 552 if (Opnd1_ExpNum == 2) 553 AllOpnds.push_back(&Opnd1_1); 554 555 // Compute instruction quota. We should save at least one instruction. 556 unsigned InstQuota = 0; 557 558 Value *V0 = I->getOperand(0); 559 Value *V1 = I->getOperand(1); 560 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) && 561 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1; 562 563 if (Value *R = simplifyFAdd(AllOpnds, InstQuota)) 564 return R; 565 } 566 567 if (OpndNum != 2) { 568 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be 569 // splitted into two addends, say "V = X - Y", the instruction would have 570 // been optimized into "I = Y - X" in the previous steps. 571 // 572 const FAddendCoef &CE = Opnd0.getCoef(); 573 return CE.isOne() ? Opnd0.getSymVal() : nullptr; 574 } 575 576 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1] 577 if (Opnd1_ExpNum) { 578 AddendVect AllOpnds; 579 AllOpnds.push_back(&Opnd0); 580 AllOpnds.push_back(&Opnd1_0); 581 if (Opnd1_ExpNum == 2) 582 AllOpnds.push_back(&Opnd1_1); 583 584 if (Value *R = simplifyFAdd(AllOpnds, 1)) 585 return R; 586 } 587 588 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1] 589 if (Opnd0_ExpNum) { 590 AddendVect AllOpnds; 591 AllOpnds.push_back(&Opnd1); 592 AllOpnds.push_back(&Opnd0_0); 593 if (Opnd0_ExpNum == 2) 594 AllOpnds.push_back(&Opnd0_1); 595 596 if (Value *R = simplifyFAdd(AllOpnds, 1)) 597 return R; 598 } 599 600 // step 6: Try factorization as the last resort, 601 return performFactorization(I); 602} 603 604Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) { 605 606 unsigned AddendNum = Addends.size(); 607 assert(AddendNum <= 4 && "Too many addends"); 608 609 // For saving intermediate results; 610 unsigned NextTmpIdx = 0; 611 FAddend TmpResult[3]; 612 613 // Points to the constant addend of the resulting simplified expression. 614 // If the resulting expr has constant-addend, this constant-addend is 615 // desirable to reside at the top of the resulting expression tree. Placing 616 // constant close to supper-expr(s) will potentially reveal some optimization 617 // opportunities in super-expr(s). 618 // 619 const FAddend *ConstAdd = nullptr; 620 621 // Simplified addends are placed <SimpVect>. 622 AddendVect SimpVect; 623 624 // The outer loop works on one symbolic-value at a time. Suppose the input 625 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ... 626 // The symbolic-values will be processed in this order: x, y, z. 627 // 628 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) { 629 630 const FAddend *ThisAddend = Addends[SymIdx]; 631 if (!ThisAddend) { 632 // This addend was processed before. 633 continue; 634 } 635 636 Value *Val = ThisAddend->getSymVal(); 637 unsigned StartIdx = SimpVect.size(); 638 SimpVect.push_back(ThisAddend); 639 640 // The inner loop collects addends sharing same symbolic-value, and these 641 // addends will be later on folded into a single addend. Following above 642 // example, if the symbolic value "y" is being processed, the inner loop 643 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will 644 // be later on folded into "<b1+b2, y>". 645 // 646 for (unsigned SameSymIdx = SymIdx + 1; 647 SameSymIdx < AddendNum; SameSymIdx++) { 648 const FAddend *T = Addends[SameSymIdx]; 649 if (T && T->getSymVal() == Val) { 650 // Set null such that next iteration of the outer loop will not process 651 // this addend again. 652 Addends[SameSymIdx] = nullptr; 653 SimpVect.push_back(T); 654 } 655 } 656 657 // If multiple addends share same symbolic value, fold them together. 658 if (StartIdx + 1 != SimpVect.size()) { 659 FAddend &R = TmpResult[NextTmpIdx ++]; 660 R = *SimpVect[StartIdx]; 661 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++) 662 R += *SimpVect[Idx]; 663 664 // Pop all addends being folded and push the resulting folded addend. 665 SimpVect.resize(StartIdx); 666 if (Val) { 667 if (!R.isZero()) { 668 SimpVect.push_back(&R); 669 } 670 } else { 671 // Don't push constant addend at this time. It will be the last element 672 // of <SimpVect>. 673 ConstAdd = &R; 674 } 675 } 676 } 677 678 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) && 679 "out-of-bound access"); 680 681 if (ConstAdd) 682 SimpVect.push_back(ConstAdd); 683 684 Value *Result; 685 if (!SimpVect.empty()) 686 Result = createNaryFAdd(SimpVect, InstrQuota); 687 else { 688 // The addition is folded to 0.0. 689 Result = ConstantFP::get(Instr->getType(), 0.0); 690 } 691 692 return Result; 693} 694 695Value *FAddCombine::createNaryFAdd 696 (const AddendVect &Opnds, unsigned InstrQuota) { 697 assert(!Opnds.empty() && "Expect at least one addend"); 698 699 // Step 1: Check if the # of instructions needed exceeds the quota. 700 // 701 unsigned InstrNeeded = calcInstrNumber(Opnds); 702 if (InstrNeeded > InstrQuota) 703 return nullptr; 704 705 initCreateInstNum(); 706 707 // step 2: Emit the N-ary addition. 708 // Note that at most three instructions are involved in Fadd-InstCombine: the 709 // addition in question, and at most two neighboring instructions. 710 // The resulting optimized addition should have at least one less instruction 711 // than the original addition expression tree. This implies that the resulting 712 // N-ary addition has at most two instructions, and we don't need to worry 713 // about tree-height when constructing the N-ary addition. 714 715 Value *LastVal = nullptr; 716 bool LastValNeedNeg = false; 717 718 // Iterate the addends, creating fadd/fsub using adjacent two addends. 719 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 720 I != E; I++) { 721 bool NeedNeg; 722 Value *V = createAddendVal(**I, NeedNeg); 723 if (!LastVal) { 724 LastVal = V; 725 LastValNeedNeg = NeedNeg; 726 continue; 727 } 728 729 if (LastValNeedNeg == NeedNeg) { 730 LastVal = createFAdd(LastVal, V); 731 continue; 732 } 733 734 if (LastValNeedNeg) 735 LastVal = createFSub(V, LastVal); 736 else 737 LastVal = createFSub(LastVal, V); 738 739 LastValNeedNeg = false; 740 } 741 742 if (LastValNeedNeg) { 743 LastVal = createFNeg(LastVal); 744 } 745 746 #ifndef NDEBUG 747 assert(CreateInstrNum == InstrNeeded && 748 "Inconsistent in instruction numbers"); 749 #endif 750 751 return LastVal; 752} 753 754Value *FAddCombine::createFSub 755 (Value *Opnd0, Value *Opnd1) { 756 Value *V = Builder->CreateFSub(Opnd0, Opnd1); 757 if (Instruction *I = dyn_cast<Instruction>(V)) 758 createInstPostProc(I); 759 return V; 760} 761 762Value *FAddCombine::createFNeg(Value *V) { 763 Value *Zero = cast<Value>(ConstantFP::get(V->getType(), 0.0)); 764 Value *NewV = createFSub(Zero, V); 765 if (Instruction *I = dyn_cast<Instruction>(NewV)) 766 createInstPostProc(I, true); // fneg's don't receive instruction numbers. 767 return NewV; 768} 769 770Value *FAddCombine::createFAdd 771 (Value *Opnd0, Value *Opnd1) { 772 Value *V = Builder->CreateFAdd(Opnd0, Opnd1); 773 if (Instruction *I = dyn_cast<Instruction>(V)) 774 createInstPostProc(I); 775 return V; 776} 777 778Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) { 779 Value *V = Builder->CreateFMul(Opnd0, Opnd1); 780 if (Instruction *I = dyn_cast<Instruction>(V)) 781 createInstPostProc(I); 782 return V; 783} 784 785Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) { 786 Value *V = Builder->CreateFDiv(Opnd0, Opnd1); 787 if (Instruction *I = dyn_cast<Instruction>(V)) 788 createInstPostProc(I); 789 return V; 790} 791 792void FAddCombine::createInstPostProc(Instruction *NewInstr, 793 bool NoNumber) { 794 NewInstr->setDebugLoc(Instr->getDebugLoc()); 795 796 // Keep track of the number of instruction created. 797 if (!NoNumber) 798 incCreateInstNum(); 799 800 // Propagate fast-math flags 801 NewInstr->setFastMathFlags(Instr->getFastMathFlags()); 802} 803 804// Return the number of instruction needed to emit the N-ary addition. 805// NOTE: Keep this function in sync with createAddendVal(). 806unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) { 807 unsigned OpndNum = Opnds.size(); 808 unsigned InstrNeeded = OpndNum - 1; 809 810 // The number of addends in the form of "(-1)*x". 811 unsigned NegOpndNum = 0; 812 813 // Adjust the number of instructions needed to emit the N-ary add. 814 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end(); 815 I != E; I++) { 816 const FAddend *Opnd = *I; 817 if (Opnd->isConstant()) 818 continue; 819 820 const FAddendCoef &CE = Opnd->getCoef(); 821 if (CE.isMinusOne() || CE.isMinusTwo()) 822 NegOpndNum++; 823 824 // Let the addend be "c * x". If "c == +/-1", the value of the addend 825 // is immediately available; otherwise, it needs exactly one instruction 826 // to evaluate the value. 827 if (!CE.isMinusOne() && !CE.isOne()) 828 InstrNeeded++; 829 } 830 if (NegOpndNum == OpndNum) 831 InstrNeeded++; 832 return InstrNeeded; 833} 834 835// Input Addend Value NeedNeg(output) 836// ================================================================ 837// Constant C C false 838// <+/-1, V> V coefficient is -1 839// <2/-2, V> "fadd V, V" coefficient is -2 840// <C, V> "fmul V, C" false 841// 842// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber. 843Value *FAddCombine::createAddendVal 844 (const FAddend &Opnd, bool &NeedNeg) { 845 const FAddendCoef &Coeff = Opnd.getCoef(); 846 847 if (Opnd.isConstant()) { 848 NeedNeg = false; 849 return Coeff.getValue(Instr->getType()); 850 } 851 852 Value *OpndVal = Opnd.getSymVal(); 853 854 if (Coeff.isMinusOne() || Coeff.isOne()) { 855 NeedNeg = Coeff.isMinusOne(); 856 return OpndVal; 857 } 858 859 if (Coeff.isTwo() || Coeff.isMinusTwo()) { 860 NeedNeg = Coeff.isMinusTwo(); 861 return createFAdd(OpndVal, OpndVal); 862 } 863 864 NeedNeg = false; 865 return createFMul(OpndVal, Coeff.getValue(Instr->getType())); 866} 867 868// dyn_castFoldableMul - If this value is a multiply that can be folded into 869// other computations (because it has a constant operand), return the 870// non-constant operand of the multiply, and set CST to point to the multiplier. 871// Otherwise, return null. 872// 873static inline Value *dyn_castFoldableMul(Value *V, Constant *&CST) { 874 if (!V->hasOneUse() || !V->getType()->isIntOrIntVectorTy()) 875 return nullptr; 876 877 Instruction *I = dyn_cast<Instruction>(V); 878 if (!I) return nullptr; 879 880 if (I->getOpcode() == Instruction::Mul) 881 if ((CST = dyn_cast<Constant>(I->getOperand(1)))) 882 return I->getOperand(0); 883 if (I->getOpcode() == Instruction::Shl) 884 if ((CST = dyn_cast<Constant>(I->getOperand(1)))) { 885 // The multiplier is really 1 << CST. 886 CST = ConstantExpr::getShl(ConstantInt::get(V->getType(), 1), CST); 887 return I->getOperand(0); 888 } 889 return nullptr; 890} 891 892 893/// WillNotOverflowSignedAdd - Return true if we can prove that: 894/// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS)) 895/// This basically requires proving that the add in the original type would not 896/// overflow to change the sign bit or have a carry out. 897bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) { 898 // There are different heuristics we can use for this. Here are some simple 899 // ones. 900 901 // Add has the property that adding any two 2's complement numbers can only 902 // have one carry bit which can change a sign. As such, if LHS and RHS each 903 // have at least two sign bits, we know that the addition of the two values 904 // will sign extend fine. 905 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1) 906 return true; 907 908 909 // If one of the operands only has one non-zero bit, and if the other operand 910 // has a known-zero bit in a more significant place than it (not including the 911 // sign bit) the ripple may go up to and fill the zero, but won't change the 912 // sign. For example, (X & ~4) + 1. 913 914 // TODO: Implement. 915 916 return false; 917} 918 919Instruction *InstCombiner::visitAdd(BinaryOperator &I) { 920 bool Changed = SimplifyAssociativeOrCommutative(I); 921 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 922 923 if (Value *V = SimplifyVectorOp(I)) 924 return ReplaceInstUsesWith(I, V); 925 926 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(), 927 I.hasNoUnsignedWrap(), DL)) 928 return ReplaceInstUsesWith(I, V); 929 930 // (A*B)+(A*C) -> A*(B+C) etc 931 if (Value *V = SimplifyUsingDistributiveLaws(I)) 932 return ReplaceInstUsesWith(I, V); 933 934 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 935 // X + (signbit) --> X ^ signbit 936 const APInt &Val = CI->getValue(); 937 if (Val.isSignBit()) 938 return BinaryOperator::CreateXor(LHS, RHS); 939 940 // See if SimplifyDemandedBits can simplify this. This handles stuff like 941 // (X & 254)+1 -> (X&254)|1 942 if (SimplifyDemandedInstructionBits(I)) 943 return &I; 944 945 // zext(bool) + C -> bool ? C + 1 : C 946 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS)) 947 if (ZI->getSrcTy()->isIntegerTy(1)) 948 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI); 949 950 Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr; 951 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) { 952 uint32_t TySizeBits = I.getType()->getScalarSizeInBits(); 953 const APInt &RHSVal = CI->getValue(); 954 unsigned ExtendAmt = 0; 955 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. 956 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. 957 if (XorRHS->getValue() == -RHSVal) { 958 if (RHSVal.isPowerOf2()) 959 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1; 960 else if (XorRHS->getValue().isPowerOf2()) 961 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1; 962 } 963 964 if (ExtendAmt) { 965 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt); 966 if (!MaskedValueIsZero(XorLHS, Mask)) 967 ExtendAmt = 0; 968 } 969 970 if (ExtendAmt) { 971 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt); 972 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext"); 973 return BinaryOperator::CreateAShr(NewShl, ShAmt); 974 } 975 976 // If this is a xor that was canonicalized from a sub, turn it back into 977 // a sub and fuse this add with it. 978 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) { 979 IntegerType *IT = cast<IntegerType>(I.getType()); 980 APInt LHSKnownOne(IT->getBitWidth(), 0); 981 APInt LHSKnownZero(IT->getBitWidth(), 0); 982 computeKnownBits(XorLHS, LHSKnownZero, LHSKnownOne); 983 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue()) 984 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI), 985 XorLHS); 986 } 987 // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C, 988 // transform them into (X + (signbit ^ C)) 989 if (XorRHS->getValue().isSignBit()) 990 return BinaryOperator::CreateAdd(XorLHS, 991 ConstantExpr::getXor(XorRHS, CI)); 992 } 993 } 994 995 if (isa<Constant>(RHS) && isa<PHINode>(LHS)) 996 if (Instruction *NV = FoldOpIntoPhi(I)) 997 return NV; 998 999 if (I.getType()->getScalarType()->isIntegerTy(1)) 1000 return BinaryOperator::CreateXor(LHS, RHS); 1001 1002 // X + X --> X << 1 1003 if (LHS == RHS) { 1004 BinaryOperator *New = 1005 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1)); 1006 New->setHasNoSignedWrap(I.hasNoSignedWrap()); 1007 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1008 return New; 1009 } 1010 1011 // -A + B --> B - A 1012 // -A + -B --> -(A + B) 1013 if (Value *LHSV = dyn_castNegVal(LHS)) { 1014 if (!isa<Constant>(RHS)) 1015 if (Value *RHSV = dyn_castNegVal(RHS)) { 1016 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum"); 1017 return BinaryOperator::CreateNeg(NewAdd); 1018 } 1019 1020 return BinaryOperator::CreateSub(RHS, LHSV); 1021 } 1022 1023 // A + -B --> A - B 1024 if (!isa<Constant>(RHS)) 1025 if (Value *V = dyn_castNegVal(RHS)) 1026 return BinaryOperator::CreateSub(LHS, V); 1027 1028 1029 { 1030 Constant *C2; 1031 if (Value *X = dyn_castFoldableMul(LHS, C2)) { 1032 if (X == RHS) // X*C + X --> X * (C+1) 1033 return BinaryOperator::CreateMul(RHS, AddOne(C2)); 1034 1035 // X*C1 + X*C2 --> X * (C1+C2) 1036 Constant *C1; 1037 if (X == dyn_castFoldableMul(RHS, C1)) 1038 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2)); 1039 } 1040 1041 // X + X*C --> X * (C+1) 1042 if (dyn_castFoldableMul(RHS, C2) == LHS) 1043 return BinaryOperator::CreateMul(LHS, AddOne(C2)); 1044 } 1045 1046 // A+B --> A|B iff A and B have no bits set in common. 1047 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) { 1048 APInt LHSKnownOne(IT->getBitWidth(), 0); 1049 APInt LHSKnownZero(IT->getBitWidth(), 0); 1050 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne); 1051 if (LHSKnownZero != 0) { 1052 APInt RHSKnownOne(IT->getBitWidth(), 0); 1053 APInt RHSKnownZero(IT->getBitWidth(), 0); 1054 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne); 1055 1056 // No bits in common -> bitwise or. 1057 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue()) 1058 return BinaryOperator::CreateOr(LHS, RHS); 1059 } 1060 } 1061 1062 // W*X + Y*Z --> W * (X+Z) iff W == Y 1063 { 1064 Value *W, *X, *Y, *Z; 1065 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) && 1066 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) { 1067 if (W != Y) { 1068 if (W == Z) { 1069 std::swap(Y, Z); 1070 } else if (Y == X) { 1071 std::swap(W, X); 1072 } else if (X == Z) { 1073 std::swap(Y, Z); 1074 std::swap(W, X); 1075 } 1076 } 1077 1078 if (W == Y) { 1079 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName()); 1080 return BinaryOperator::CreateMul(W, NewAdd); 1081 } 1082 } 1083 } 1084 1085 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 1086 Value *X; 1087 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X 1088 return BinaryOperator::CreateSub(SubOne(CRHS), X); 1089 } 1090 1091 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) { 1092 // (X & FF00) + xx00 -> (X+xx00) & FF00 1093 Value *X; 1094 ConstantInt *C2; 1095 if (LHS->hasOneUse() && 1096 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) && 1097 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) { 1098 // See if all bits from the first bit set in the Add RHS up are included 1099 // in the mask. First, get the rightmost bit. 1100 const APInt &AddRHSV = CRHS->getValue(); 1101 1102 // Form a mask of all bits from the lowest bit added through the top. 1103 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1)); 1104 1105 // See if the and mask includes all of these bits. 1106 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue()); 1107 1108 if (AddRHSHighBits == AddRHSHighBitsAnd) { 1109 // Okay, the xform is safe. Insert the new add pronto. 1110 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName()); 1111 return BinaryOperator::CreateAnd(NewAdd, C2); 1112 } 1113 } 1114 1115 // Try to fold constant add into select arguments. 1116 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1117 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1118 return R; 1119 } 1120 1121 // add (select X 0 (sub n A)) A --> select X A n 1122 { 1123 SelectInst *SI = dyn_cast<SelectInst>(LHS); 1124 Value *A = RHS; 1125 if (!SI) { 1126 SI = dyn_cast<SelectInst>(RHS); 1127 A = LHS; 1128 } 1129 if (SI && SI->hasOneUse()) { 1130 Value *TV = SI->getTrueValue(); 1131 Value *FV = SI->getFalseValue(); 1132 Value *N; 1133 1134 // Can we fold the add into the argument of the select? 1135 // We check both true and false select arguments for a matching subtract. 1136 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A)))) 1137 // Fold the add into the true select value. 1138 return SelectInst::Create(SI->getCondition(), N, A); 1139 1140 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A)))) 1141 // Fold the add into the false select value. 1142 return SelectInst::Create(SI->getCondition(), A, N); 1143 } 1144 } 1145 1146 // Check for (add (sext x), y), see if we can merge this into an 1147 // integer add followed by a sext. 1148 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) { 1149 // (add (sext x), cst) --> (sext (add x, cst')) 1150 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 1151 Constant *CI = 1152 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType()); 1153 if (LHSConv->hasOneUse() && 1154 ConstantExpr::getSExt(CI, I.getType()) == RHSC && 1155 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1156 // Insert the new, smaller add. 1157 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1158 CI, "addconv"); 1159 return new SExtInst(NewAdd, I.getType()); 1160 } 1161 } 1162 1163 // (add (sext x), (sext y)) --> (sext (add int x, y)) 1164 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) { 1165 // Only do this if x/y have the same type, if at last one of them has a 1166 // single use (so we don't increase the number of sexts), and if the 1167 // integer add will not overflow. 1168 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1169 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1170 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1171 RHSConv->getOperand(0))) { 1172 // Insert the new integer add. 1173 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1174 RHSConv->getOperand(0), "addconv"); 1175 return new SExtInst(NewAdd, I.getType()); 1176 } 1177 } 1178 } 1179 1180 // Check for (x & y) + (x ^ y) 1181 { 1182 Value *A = nullptr, *B = nullptr; 1183 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) && 1184 (match(LHS, m_And(m_Specific(A), m_Specific(B))) || 1185 match(LHS, m_And(m_Specific(B), m_Specific(A))))) 1186 return BinaryOperator::CreateOr(A, B); 1187 1188 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) && 1189 (match(RHS, m_And(m_Specific(A), m_Specific(B))) || 1190 match(RHS, m_And(m_Specific(B), m_Specific(A))))) 1191 return BinaryOperator::CreateOr(A, B); 1192 } 1193 1194 return Changed ? &I : nullptr; 1195} 1196 1197Instruction *InstCombiner::visitFAdd(BinaryOperator &I) { 1198 bool Changed = SimplifyAssociativeOrCommutative(I); 1199 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 1200 1201 if (Value *V = SimplifyVectorOp(I)) 1202 return ReplaceInstUsesWith(I, V); 1203 1204 if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL)) 1205 return ReplaceInstUsesWith(I, V); 1206 1207 if (isa<Constant>(RHS)) { 1208 if (isa<PHINode>(LHS)) 1209 if (Instruction *NV = FoldOpIntoPhi(I)) 1210 return NV; 1211 1212 if (SelectInst *SI = dyn_cast<SelectInst>(LHS)) 1213 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1214 return NV; 1215 } 1216 1217 // -A + B --> B - A 1218 // -A + -B --> -(A + B) 1219 if (Value *LHSV = dyn_castFNegVal(LHS)) { 1220 Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV); 1221 RI->copyFastMathFlags(&I); 1222 return RI; 1223 } 1224 1225 // A + -B --> A - B 1226 if (!isa<Constant>(RHS)) 1227 if (Value *V = dyn_castFNegVal(RHS)) { 1228 Instruction *RI = BinaryOperator::CreateFSub(LHS, V); 1229 RI->copyFastMathFlags(&I); 1230 return RI; 1231 } 1232 1233 // Check for (fadd double (sitofp x), y), see if we can merge this into an 1234 // integer add followed by a promotion. 1235 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) { 1236 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst)) 1237 // ... if the constant fits in the integer value. This is useful for things 1238 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer 1239 // requires a constant pool load, and generally allows the add to be better 1240 // instcombined. 1241 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 1242 Constant *CI = 1243 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType()); 1244 if (LHSConv->hasOneUse() && 1245 ConstantExpr::getSIToFP(CI, I.getType()) == CFP && 1246 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) { 1247 // Insert the new integer add. 1248 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1249 CI, "addconv"); 1250 return new SIToFPInst(NewAdd, I.getType()); 1251 } 1252 } 1253 1254 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y)) 1255 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) { 1256 // Only do this if x/y have the same type, if at last one of them has a 1257 // single use (so we don't increase the number of int->fp conversions), 1258 // and if the integer add will not overflow. 1259 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&& 1260 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && 1261 WillNotOverflowSignedAdd(LHSConv->getOperand(0), 1262 RHSConv->getOperand(0))) { 1263 // Insert the new integer add. 1264 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0), 1265 RHSConv->getOperand(0),"addconv"); 1266 return new SIToFPInst(NewAdd, I.getType()); 1267 } 1268 } 1269 } 1270 1271 // select C, 0, B + select C, A, 0 -> select C, A, B 1272 { 1273 Value *A1, *B1, *C1, *A2, *B2, *C2; 1274 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) && 1275 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) { 1276 if (C1 == C2) { 1277 Constant *Z1=nullptr, *Z2=nullptr; 1278 Value *A, *B, *C=C1; 1279 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) { 1280 Z1 = dyn_cast<Constant>(A1); A = A2; 1281 Z2 = dyn_cast<Constant>(B2); B = B1; 1282 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) { 1283 Z1 = dyn_cast<Constant>(B1); B = B2; 1284 Z2 = dyn_cast<Constant>(A2); A = A1; 1285 } 1286 1287 if (Z1 && Z2 && 1288 (I.hasNoSignedZeros() || 1289 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) { 1290 return SelectInst::Create(C, A, B); 1291 } 1292 } 1293 } 1294 } 1295 1296 if (I.hasUnsafeAlgebra()) { 1297 if (Value *V = FAddCombine(Builder).simplify(&I)) 1298 return ReplaceInstUsesWith(I, V); 1299 } 1300 1301 return Changed ? &I : nullptr; 1302} 1303 1304 1305/// Optimize pointer differences into the same array into a size. Consider: 1306/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer 1307/// operands to the ptrtoint instructions for the LHS/RHS of the subtract. 1308/// 1309Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS, 1310 Type *Ty) { 1311 assert(DL && "Must have target data info for this"); 1312 1313 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize 1314 // this. 1315 bool Swapped = false; 1316 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr; 1317 1318 // For now we require one side to be the base pointer "A" or a constant 1319 // GEP derived from it. 1320 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1321 // (gep X, ...) - X 1322 if (LHSGEP->getOperand(0) == RHS) { 1323 GEP1 = LHSGEP; 1324 Swapped = false; 1325 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1326 // (gep X, ...) - (gep X, ...) 1327 if (LHSGEP->getOperand(0)->stripPointerCasts() == 1328 RHSGEP->getOperand(0)->stripPointerCasts()) { 1329 GEP2 = RHSGEP; 1330 GEP1 = LHSGEP; 1331 Swapped = false; 1332 } 1333 } 1334 } 1335 1336 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) { 1337 // X - (gep X, ...) 1338 if (RHSGEP->getOperand(0) == LHS) { 1339 GEP1 = RHSGEP; 1340 Swapped = true; 1341 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) { 1342 // (gep X, ...) - (gep X, ...) 1343 if (RHSGEP->getOperand(0)->stripPointerCasts() == 1344 LHSGEP->getOperand(0)->stripPointerCasts()) { 1345 GEP2 = LHSGEP; 1346 GEP1 = RHSGEP; 1347 Swapped = true; 1348 } 1349 } 1350 } 1351 1352 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and 1353 // multiple users. 1354 if (!GEP1 || 1355 (GEP2 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse())) 1356 return nullptr; 1357 1358 // Emit the offset of the GEP and an intptr_t. 1359 Value *Result = EmitGEPOffset(GEP1); 1360 1361 // If we had a constant expression GEP on the other side offsetting the 1362 // pointer, subtract it from the offset we have. 1363 if (GEP2) { 1364 Value *Offset = EmitGEPOffset(GEP2); 1365 Result = Builder->CreateSub(Result, Offset); 1366 } 1367 1368 // If we have p - gep(p, ...) then we have to negate the result. 1369 if (Swapped) 1370 Result = Builder->CreateNeg(Result, "diff.neg"); 1371 1372 return Builder->CreateIntCast(Result, Ty, true); 1373} 1374 1375 1376Instruction *InstCombiner::visitSub(BinaryOperator &I) { 1377 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1378 1379 if (Value *V = SimplifyVectorOp(I)) 1380 return ReplaceInstUsesWith(I, V); 1381 1382 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(), 1383 I.hasNoUnsignedWrap(), DL)) 1384 return ReplaceInstUsesWith(I, V); 1385 1386 // (A*B)-(A*C) -> A*(B-C) etc 1387 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1388 return ReplaceInstUsesWith(I, V); 1389 1390 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW. 1391 if (Value *V = dyn_castNegVal(Op1)) { 1392 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V); 1393 Res->setHasNoSignedWrap(I.hasNoSignedWrap()); 1394 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 1395 return Res; 1396 } 1397 1398 if (I.getType()->isIntegerTy(1)) 1399 return BinaryOperator::CreateXor(Op0, Op1); 1400 1401 // Replace (-1 - A) with (~A). 1402 if (match(Op0, m_AllOnes())) 1403 return BinaryOperator::CreateNot(Op1); 1404 1405 if (Constant *C = dyn_cast<Constant>(Op0)) { 1406 // C - ~X == X + (1+C) 1407 Value *X = nullptr; 1408 if (match(Op1, m_Not(m_Value(X)))) 1409 return BinaryOperator::CreateAdd(X, AddOne(C)); 1410 1411 // Try to fold constant sub into select arguments. 1412 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1413 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1414 return R; 1415 1416 // C-(X+C2) --> (C-C2)-X 1417 Constant *C2; 1418 if (match(Op1, m_Add(m_Value(X), m_Constant(C2)))) 1419 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X); 1420 1421 if (SimplifyDemandedInstructionBits(I)) 1422 return &I; 1423 1424 // Fold (sub 0, (zext bool to B)) --> (sext bool to B) 1425 if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X)))) 1426 if (X->getType()->getScalarType()->isIntegerTy(1)) 1427 return CastInst::CreateSExtOrBitCast(X, Op1->getType()); 1428 1429 // Fold (sub 0, (sext bool to B)) --> (zext bool to B) 1430 if (C->isNullValue() && match(Op1, m_SExt(m_Value(X)))) 1431 if (X->getType()->getScalarType()->isIntegerTy(1)) 1432 return CastInst::CreateZExtOrBitCast(X, Op1->getType()); 1433 } 1434 1435 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) { 1436 // -(X >>u 31) -> (X >>s 31) 1437 // -(X >>s 31) -> (X >>u 31) 1438 if (C->isZero()) { 1439 Value *X; ConstantInt *CI; 1440 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) && 1441 // Verify we are shifting out everything but the sign bit. 1442 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1443 return BinaryOperator::CreateAShr(X, CI); 1444 1445 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) && 1446 // Verify we are shifting out everything but the sign bit. 1447 CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1) 1448 return BinaryOperator::CreateLShr(X, CI); 1449 } 1450 } 1451 1452 1453 { Value *Y; 1454 // X-(X+Y) == -Y X-(Y+X) == -Y 1455 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) || 1456 match(Op1, m_Add(m_Value(Y), m_Specific(Op0)))) 1457 return BinaryOperator::CreateNeg(Y); 1458 1459 // (X-Y)-X == -Y 1460 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y)))) 1461 return BinaryOperator::CreateNeg(Y); 1462 } 1463 1464 if (Op1->hasOneUse()) { 1465 Value *X = nullptr, *Y = nullptr, *Z = nullptr; 1466 Constant *C = nullptr; 1467 Constant *CI = nullptr; 1468 1469 // (X - (Y - Z)) --> (X + (Z - Y)). 1470 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z)))) 1471 return BinaryOperator::CreateAdd(Op0, 1472 Builder->CreateSub(Z, Y, Op1->getName())); 1473 1474 // (X - (X & Y)) --> (X & ~Y) 1475 // 1476 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) || 1477 match(Op1, m_And(m_Specific(Op0), m_Value(Y)))) 1478 return BinaryOperator::CreateAnd(Op0, 1479 Builder->CreateNot(Y, Y->getName() + ".not")); 1480 1481 // 0 - (X sdiv C) -> (X sdiv -C) 1482 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && 1483 match(Op0, m_Zero())) 1484 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C)); 1485 1486 // 0 - (X << Y) -> (-X << Y) when X is freely negatable. 1487 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero())) 1488 if (Value *XNeg = dyn_castNegVal(X)) 1489 return BinaryOperator::CreateShl(XNeg, Y); 1490 1491 // X - X*C --> X * (1-C) 1492 if (match(Op1, m_Mul(m_Specific(Op0), m_Constant(CI)))) { 1493 Constant *CP1 = ConstantExpr::getSub(ConstantInt::get(I.getType(),1), CI); 1494 return BinaryOperator::CreateMul(Op0, CP1); 1495 } 1496 1497 // X - X<<C --> X * (1-(1<<C)) 1498 if (match(Op1, m_Shl(m_Specific(Op0), m_Constant(CI)))) { 1499 Constant *One = ConstantInt::get(I.getType(), 1); 1500 C = ConstantExpr::getSub(One, ConstantExpr::getShl(One, CI)); 1501 return BinaryOperator::CreateMul(Op0, C); 1502 } 1503 1504 // X - A*-B -> X + A*B 1505 // X - -A*B -> X + A*B 1506 Value *A, *B; 1507 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) || 1508 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B)))) 1509 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B)); 1510 1511 // X - A*CI -> X + A*-CI 1512 // X - CI*A -> X + A*-CI 1513 if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) || 1514 match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) { 1515 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI)); 1516 return BinaryOperator::CreateAdd(Op0, NewMul); 1517 } 1518 } 1519 1520 Constant *C1; 1521 if (Value *X = dyn_castFoldableMul(Op0, C1)) { 1522 if (X == Op1) // X*C - X --> X * (C-1) 1523 return BinaryOperator::CreateMul(Op1, SubOne(C1)); 1524 1525 Constant *C2; // X*C1 - X*C2 -> X * (C1-C2) 1526 if (X == dyn_castFoldableMul(Op1, C2)) 1527 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2)); 1528 } 1529 1530 // Optimize pointer differences into the same array into a size. Consider: 1531 // &A[10] - &A[0]: we should compile this to "10". 1532 if (DL) { 1533 Value *LHSOp, *RHSOp; 1534 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) && 1535 match(Op1, m_PtrToInt(m_Value(RHSOp)))) 1536 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1537 return ReplaceInstUsesWith(I, Res); 1538 1539 // trunc(p)-trunc(q) -> trunc(p-q) 1540 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) && 1541 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp))))) 1542 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType())) 1543 return ReplaceInstUsesWith(I, Res); 1544 } 1545 1546 return nullptr; 1547} 1548 1549Instruction *InstCombiner::visitFSub(BinaryOperator &I) { 1550 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1551 1552 if (Value *V = SimplifyVectorOp(I)) 1553 return ReplaceInstUsesWith(I, V); 1554 1555 if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL)) 1556 return ReplaceInstUsesWith(I, V); 1557 1558 if (isa<Constant>(Op0)) 1559 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1560 if (Instruction *NV = FoldOpIntoSelect(I, SI)) 1561 return NV; 1562 1563 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking 1564 // through FP extensions/truncations along the way. 1565 if (Value *V = dyn_castFNegVal(Op1)) { 1566 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V); 1567 NewI->copyFastMathFlags(&I); 1568 return NewI; 1569 } 1570 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) { 1571 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) { 1572 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType()); 1573 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc); 1574 NewI->copyFastMathFlags(&I); 1575 return NewI; 1576 } 1577 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) { 1578 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) { 1579 Value *NewExt = Builder->CreateFPExt(V, I.getType()); 1580 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt); 1581 NewI->copyFastMathFlags(&I); 1582 return NewI; 1583 } 1584 } 1585 1586 if (I.hasUnsafeAlgebra()) { 1587 if (Value *V = FAddCombine(Builder).simplify(&I)) 1588 return ReplaceInstUsesWith(I, V); 1589 } 1590 1591 return nullptr; 1592} 1593