1 //===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for add, fadd, sub, and fsub.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/GetElementPtrTypeIterator.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/Support/KnownBits.h"
23 using namespace PatternMatch;
25 #define DEBUG_TYPE "instcombine"
29 /// Class representing coefficient of floating-point addend.
30 /// This class needs to be highly efficient, which is especially true for
31 /// the constructor. As of I write this comment, the cost of the default
32 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
33 /// perform write-merging).
37 // The constructor has to initialize a APFloat, which is unnecessary for
38 // most addends which have coefficient either 1 or -1. So, the constructor
39 // is expensive. In order to avoid the cost of the constructor, we should
40 // reuse some instances whenever possible. The pre-created instances
41 // FAddCombine::Add[0-5] embodies this idea.
43 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {}
47 assert(!insaneIntVal(C) && "Insane coefficient");
48 IsFp = false; IntVal = C;
51 void set(const APFloat& C);
55 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
56 Value *getValue(Type *) const;
58 // If possible, don't define operator+/operator- etc because these
59 // operators inevitably call FAddendCoef's constructor which is not cheap.
60 void operator=(const FAddendCoef &A);
61 void operator+=(const FAddendCoef &A);
62 void operator*=(const FAddendCoef &S);
64 bool isOne() const { return isInt() && IntVal == 1; }
65 bool isTwo() const { return isInt() && IntVal == 2; }
66 bool isMinusOne() const { return isInt() && IntVal == -1; }
67 bool isMinusTwo() const { return isInt() && IntVal == -2; }
70 bool insaneIntVal(int V) { return V > 4 || V < -4; }
71 APFloat *getFpValPtr()
72 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); }
73 const APFloat *getFpValPtr() const
74 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); }
76 const APFloat &getFpVal() const {
77 assert(IsFp && BufHasFpVal && "Incorret state");
78 return *getFpValPtr();
82 assert(IsFp && BufHasFpVal && "Incorret state");
83 return *getFpValPtr();
86 bool isInt() const { return !IsFp; }
88 // If the coefficient is represented by an integer, promote it to a
90 void convertToFpType(const fltSemantics &Sem);
92 // Construct an APFloat from a signed integer.
93 // TODO: We should get rid of this function when APFloat can be constructed
94 // from an *SIGNED* integer.
95 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
100 // True iff FpValBuf contains an instance of APFloat.
103 // The integer coefficient of an individual addend is either 1 or -1,
104 // and we try to simplify at most 4 addends from neighboring at most
105 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
106 // is overkill of this end.
109 AlignedCharArrayUnion<APFloat> FpValBuf;
112 /// FAddend is used to represent floating-point addend. An addend is
113 /// represented as <C, V>, where the V is a symbolic value, and C is a
114 /// constant coefficient. A constant addend is represented as <C, 0>.
118 FAddend() : Val(nullptr) {}
120 Value *getSymVal() const { return Val; }
121 const FAddendCoef &getCoef() const { return Coeff; }
123 bool isConstant() const { return Val == nullptr; }
124 bool isZero() const { return Coeff.isZero(); }
126 void set(short Coefficient, Value *V) {
127 Coeff.set(Coefficient);
130 void set(const APFloat &Coefficient, Value *V) {
131 Coeff.set(Coefficient);
134 void set(const ConstantFP *Coefficient, Value *V) {
135 Coeff.set(Coefficient->getValueAPF());
139 void negate() { Coeff.negate(); }
141 /// Drill down the U-D chain one step to find the definition of V, and
142 /// try to break the definition into one or two addends.
143 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
145 /// Similar to FAddend::drillDownOneStep() except that the value being
146 /// splitted is the addend itself.
147 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
149 void operator+=(const FAddend &T) {
150 assert((Val == T.Val) && "Symbolic-values disagree");
155 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
157 // This addend has the value of "Coeff * Val".
162 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
163 /// with its neighboring at most two instructions.
167 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(nullptr) {}
168 Value *simplify(Instruction *FAdd);
171 typedef SmallVector<const FAddend*, 4> AddendVect;
173 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
175 Value *performFactorization(Instruction *I);
177 /// Convert given addend to a Value
178 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
180 /// Return the number of instructions needed to emit the N-ary addition.
181 unsigned calcInstrNumber(const AddendVect& Vect);
182 Value *createFSub(Value *Opnd0, Value *Opnd1);
183 Value *createFAdd(Value *Opnd0, Value *Opnd1);
184 Value *createFMul(Value *Opnd0, Value *Opnd1);
185 Value *createFDiv(Value *Opnd0, Value *Opnd1);
186 Value *createFNeg(Value *V);
187 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
188 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
190 InstCombiner::BuilderTy *Builder;
193 // Debugging stuff are clustered here.
195 unsigned CreateInstrNum;
196 void initCreateInstNum() { CreateInstrNum = 0; }
197 void incCreateInstNum() { CreateInstrNum++; }
199 void initCreateInstNum() {}
200 void incCreateInstNum() {}
204 } // anonymous namespace
206 //===----------------------------------------------------------------------===//
209 // {FAddendCoef, FAddend, FAddition, FAddCombine}.
211 //===----------------------------------------------------------------------===//
212 FAddendCoef::~FAddendCoef() {
214 getFpValPtr()->~APFloat();
217 void FAddendCoef::set(const APFloat& C) {
218 APFloat *P = getFpValPtr();
221 // As the buffer is meanless byte stream, we cannot call
222 // APFloat::operator=().
227 IsFp = BufHasFpVal = true;
230 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
234 APFloat *P = getFpValPtr();
236 new(P) APFloat(Sem, IntVal);
238 new(P) APFloat(Sem, 0 - IntVal);
241 IsFp = BufHasFpVal = true;
244 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
246 return APFloat(Sem, Val);
248 APFloat T(Sem, 0 - Val);
254 void FAddendCoef::operator=(const FAddendCoef &That) {
258 set(That.getFpVal());
261 void FAddendCoef::operator+=(const FAddendCoef &That) {
262 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
263 if (isInt() == That.isInt()) {
265 IntVal += That.IntVal;
267 getFpVal().add(That.getFpVal(), RndMode);
272 const APFloat &T = That.getFpVal();
273 convertToFpType(T.getSemantics());
274 getFpVal().add(T, RndMode);
278 APFloat &T = getFpVal();
279 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
282 void FAddendCoef::operator*=(const FAddendCoef &That) {
286 if (That.isMinusOne()) {
291 if (isInt() && That.isInt()) {
292 int Res = IntVal * (int)That.IntVal;
293 assert(!insaneIntVal(Res) && "Insane int value");
298 const fltSemantics &Semantic =
299 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
302 convertToFpType(Semantic);
303 APFloat &F0 = getFpVal();
306 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
307 APFloat::rmNearestTiesToEven);
309 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
312 void FAddendCoef::negate() {
316 getFpVal().changeSign();
319 Value *FAddendCoef::getValue(Type *Ty) const {
321 ConstantFP::get(Ty, float(IntVal)) :
322 ConstantFP::get(Ty->getContext(), getFpVal());
325 // The definition of <Val> Addends
326 // =========================================
327 // A + B <1, A>, <1,B>
328 // A - B <1, A>, <1,B>
331 // A + C <1, A> <C, NULL>
332 // 0 +/- 0 <0, NULL> (corner case)
334 // Legend: A and B are not constant, C is constant
336 unsigned FAddend::drillValueDownOneStep
337 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
338 Instruction *I = nullptr;
339 if (!Val || !(I = dyn_cast<Instruction>(Val)))
342 unsigned Opcode = I->getOpcode();
344 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
346 Value *Opnd0 = I->getOperand(0);
347 Value *Opnd1 = I->getOperand(1);
348 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
351 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
356 Addend0.set(1, Opnd0);
358 Addend0.set(C0, nullptr);
362 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
364 Addend.set(1, Opnd1);
366 Addend.set(C1, nullptr);
367 if (Opcode == Instruction::FSub)
372 return Opnd0 && Opnd1 ? 2 : 1;
374 // Both operands are zero. Weird!
375 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
379 if (I->getOpcode() == Instruction::FMul) {
380 Value *V0 = I->getOperand(0);
381 Value *V1 = I->getOperand(1);
382 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
387 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
396 // Try to break *this* addend into two addends. e.g. Suppose this addend is
397 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
398 // i.e. <2.3, X> and <2.3, Y>.
400 unsigned FAddend::drillAddendDownOneStep
401 (FAddend &Addend0, FAddend &Addend1) const {
405 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
406 if (!BreakNum || Coeff.isOne())
409 Addend0.Scale(Coeff);
412 Addend1.Scale(Coeff);
417 // Try to perform following optimization on the input instruction I. Return the
418 // simplified expression if was successful; otherwise, return 0.
420 // Instruction "I" is Simplified into
421 // -------------------------------------------------------
422 // (x * y) +/- (x * z) x * (y +/- z)
423 // (y / x) +/- (z / x) (y +/- z) / x
425 Value *FAddCombine::performFactorization(Instruction *I) {
426 assert((I->getOpcode() == Instruction::FAdd ||
427 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
429 Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0));
430 Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1));
432 if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
436 if (I0->getOpcode() == Instruction::FMul)
438 else if (I0->getOpcode() != Instruction::FDiv)
441 Value *Opnd0_0 = I0->getOperand(0);
442 Value *Opnd0_1 = I0->getOperand(1);
443 Value *Opnd1_0 = I1->getOperand(0);
444 Value *Opnd1_1 = I1->getOperand(1);
446 // Input Instr I Factor AddSub0 AddSub1
447 // ----------------------------------------------
448 // (x*y) +/- (x*z) x y z
449 // (y/x) +/- (z/x) x y z
451 Value *Factor = nullptr;
452 Value *AddSub0 = nullptr, *AddSub1 = nullptr;
455 if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
457 else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
461 AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
462 AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
464 } else if (Opnd0_1 == Opnd1_1) {
474 Flags.setUnsafeAlgebra();
475 if (I0) Flags &= I->getFastMathFlags();
476 if (I1) Flags &= I->getFastMathFlags();
478 // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
479 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
480 createFAdd(AddSub0, AddSub1) :
481 createFSub(AddSub0, AddSub1);
482 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
483 const APFloat &F = CFP->getValueAPF();
486 } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
487 II->setFastMathFlags(Flags);
490 Value *RI = createFMul(Factor, NewAddSub);
491 if (Instruction *II = dyn_cast<Instruction>(RI))
492 II->setFastMathFlags(Flags);
496 Value *RI = createFDiv(NewAddSub, Factor);
497 if (Instruction *II = dyn_cast<Instruction>(RI))
498 II->setFastMathFlags(Flags);
502 Value *FAddCombine::simplify(Instruction *I) {
503 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode");
505 // Currently we are not able to handle vector type.
506 if (I->getType()->isVectorTy())
509 assert((I->getOpcode() == Instruction::FAdd ||
510 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
512 // Save the instruction before calling other member-functions.
515 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
517 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
519 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
520 unsigned Opnd0_ExpNum = 0;
521 unsigned Opnd1_ExpNum = 0;
523 if (!Opnd0.isConstant())
524 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
526 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
527 if (OpndNum == 2 && !Opnd1.isConstant())
528 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
530 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
531 if (Opnd0_ExpNum && Opnd1_ExpNum) {
533 AllOpnds.push_back(&Opnd0_0);
534 AllOpnds.push_back(&Opnd1_0);
535 if (Opnd0_ExpNum == 2)
536 AllOpnds.push_back(&Opnd0_1);
537 if (Opnd1_ExpNum == 2)
538 AllOpnds.push_back(&Opnd1_1);
540 // Compute instruction quota. We should save at least one instruction.
541 unsigned InstQuota = 0;
543 Value *V0 = I->getOperand(0);
544 Value *V1 = I->getOperand(1);
545 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
546 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
548 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
553 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
554 // splitted into two addends, say "V = X - Y", the instruction would have
555 // been optimized into "I = Y - X" in the previous steps.
557 const FAddendCoef &CE = Opnd0.getCoef();
558 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
561 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
564 AllOpnds.push_back(&Opnd0);
565 AllOpnds.push_back(&Opnd1_0);
566 if (Opnd1_ExpNum == 2)
567 AllOpnds.push_back(&Opnd1_1);
569 if (Value *R = simplifyFAdd(AllOpnds, 1))
573 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
576 AllOpnds.push_back(&Opnd1);
577 AllOpnds.push_back(&Opnd0_0);
578 if (Opnd0_ExpNum == 2)
579 AllOpnds.push_back(&Opnd0_1);
581 if (Value *R = simplifyFAdd(AllOpnds, 1))
585 // step 6: Try factorization as the last resort,
586 return performFactorization(I);
589 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
590 unsigned AddendNum = Addends.size();
591 assert(AddendNum <= 4 && "Too many addends");
593 // For saving intermediate results;
594 unsigned NextTmpIdx = 0;
595 FAddend TmpResult[3];
597 // Points to the constant addend of the resulting simplified expression.
598 // If the resulting expr has constant-addend, this constant-addend is
599 // desirable to reside at the top of the resulting expression tree. Placing
600 // constant close to supper-expr(s) will potentially reveal some optimization
601 // opportunities in super-expr(s).
603 const FAddend *ConstAdd = nullptr;
605 // Simplified addends are placed <SimpVect>.
608 // The outer loop works on one symbolic-value at a time. Suppose the input
609 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
610 // The symbolic-values will be processed in this order: x, y, z.
612 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
614 const FAddend *ThisAddend = Addends[SymIdx];
616 // This addend was processed before.
620 Value *Val = ThisAddend->getSymVal();
621 unsigned StartIdx = SimpVect.size();
622 SimpVect.push_back(ThisAddend);
624 // The inner loop collects addends sharing same symbolic-value, and these
625 // addends will be later on folded into a single addend. Following above
626 // example, if the symbolic value "y" is being processed, the inner loop
627 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
628 // be later on folded into "<b1+b2, y>".
630 for (unsigned SameSymIdx = SymIdx + 1;
631 SameSymIdx < AddendNum; SameSymIdx++) {
632 const FAddend *T = Addends[SameSymIdx];
633 if (T && T->getSymVal() == Val) {
634 // Set null such that next iteration of the outer loop will not process
635 // this addend again.
636 Addends[SameSymIdx] = nullptr;
637 SimpVect.push_back(T);
641 // If multiple addends share same symbolic value, fold them together.
642 if (StartIdx + 1 != SimpVect.size()) {
643 FAddend &R = TmpResult[NextTmpIdx ++];
644 R = *SimpVect[StartIdx];
645 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
648 // Pop all addends being folded and push the resulting folded addend.
649 SimpVect.resize(StartIdx);
652 SimpVect.push_back(&R);
655 // Don't push constant addend at this time. It will be the last element
662 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
663 "out-of-bound access");
666 SimpVect.push_back(ConstAdd);
669 if (!SimpVect.empty())
670 Result = createNaryFAdd(SimpVect, InstrQuota);
672 // The addition is folded to 0.0.
673 Result = ConstantFP::get(Instr->getType(), 0.0);
679 Value *FAddCombine::createNaryFAdd
680 (const AddendVect &Opnds, unsigned InstrQuota) {
681 assert(!Opnds.empty() && "Expect at least one addend");
683 // Step 1: Check if the # of instructions needed exceeds the quota.
685 unsigned InstrNeeded = calcInstrNumber(Opnds);
686 if (InstrNeeded > InstrQuota)
691 // step 2: Emit the N-ary addition.
692 // Note that at most three instructions are involved in Fadd-InstCombine: the
693 // addition in question, and at most two neighboring instructions.
694 // The resulting optimized addition should have at least one less instruction
695 // than the original addition expression tree. This implies that the resulting
696 // N-ary addition has at most two instructions, and we don't need to worry
697 // about tree-height when constructing the N-ary addition.
699 Value *LastVal = nullptr;
700 bool LastValNeedNeg = false;
702 // Iterate the addends, creating fadd/fsub using adjacent two addends.
703 for (const FAddend *Opnd : Opnds) {
705 Value *V = createAddendVal(*Opnd, NeedNeg);
708 LastValNeedNeg = NeedNeg;
712 if (LastValNeedNeg == NeedNeg) {
713 LastVal = createFAdd(LastVal, V);
718 LastVal = createFSub(V, LastVal);
720 LastVal = createFSub(LastVal, V);
722 LastValNeedNeg = false;
725 if (LastValNeedNeg) {
726 LastVal = createFNeg(LastVal);
730 assert(CreateInstrNum == InstrNeeded &&
731 "Inconsistent in instruction numbers");
737 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
738 Value *V = Builder->CreateFSub(Opnd0, Opnd1);
739 if (Instruction *I = dyn_cast<Instruction>(V))
740 createInstPostProc(I);
744 Value *FAddCombine::createFNeg(Value *V) {
745 Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType()));
746 Value *NewV = createFSub(Zero, V);
747 if (Instruction *I = dyn_cast<Instruction>(NewV))
748 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
752 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
753 Value *V = Builder->CreateFAdd(Opnd0, Opnd1);
754 if (Instruction *I = dyn_cast<Instruction>(V))
755 createInstPostProc(I);
759 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
760 Value *V = Builder->CreateFMul(Opnd0, Opnd1);
761 if (Instruction *I = dyn_cast<Instruction>(V))
762 createInstPostProc(I);
766 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) {
767 Value *V = Builder->CreateFDiv(Opnd0, Opnd1);
768 if (Instruction *I = dyn_cast<Instruction>(V))
769 createInstPostProc(I);
773 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
774 NewInstr->setDebugLoc(Instr->getDebugLoc());
776 // Keep track of the number of instruction created.
780 // Propagate fast-math flags
781 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
784 // Return the number of instruction needed to emit the N-ary addition.
785 // NOTE: Keep this function in sync with createAddendVal().
786 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
787 unsigned OpndNum = Opnds.size();
788 unsigned InstrNeeded = OpndNum - 1;
790 // The number of addends in the form of "(-1)*x".
791 unsigned NegOpndNum = 0;
793 // Adjust the number of instructions needed to emit the N-ary add.
794 for (const FAddend *Opnd : Opnds) {
795 if (Opnd->isConstant())
798 // The constant check above is really for a few special constant
800 if (isa<UndefValue>(Opnd->getSymVal()))
803 const FAddendCoef &CE = Opnd->getCoef();
804 if (CE.isMinusOne() || CE.isMinusTwo())
807 // Let the addend be "c * x". If "c == +/-1", the value of the addend
808 // is immediately available; otherwise, it needs exactly one instruction
809 // to evaluate the value.
810 if (!CE.isMinusOne() && !CE.isOne())
813 if (NegOpndNum == OpndNum)
818 // Input Addend Value NeedNeg(output)
819 // ================================================================
820 // Constant C C false
821 // <+/-1, V> V coefficient is -1
822 // <2/-2, V> "fadd V, V" coefficient is -2
823 // <C, V> "fmul V, C" false
825 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
826 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
827 const FAddendCoef &Coeff = Opnd.getCoef();
829 if (Opnd.isConstant()) {
831 return Coeff.getValue(Instr->getType());
834 Value *OpndVal = Opnd.getSymVal();
836 if (Coeff.isMinusOne() || Coeff.isOne()) {
837 NeedNeg = Coeff.isMinusOne();
841 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
842 NeedNeg = Coeff.isMinusTwo();
843 return createFAdd(OpndVal, OpndVal);
847 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
850 /// \brief Return true if we can prove that:
851 /// (sub LHS, RHS) === (sub nsw LHS, RHS)
852 /// This basically requires proving that the add in the original type would not
853 /// overflow to change the sign bit or have a carry out.
854 /// TODO: Handle this for Vectors.
855 bool InstCombiner::willNotOverflowSignedSub(const Value *LHS,
857 const Instruction &CxtI) const {
858 // If LHS and RHS each have at least two sign bits, the subtraction
860 if (ComputeNumSignBits(LHS, 0, &CxtI) > 1 &&
861 ComputeNumSignBits(RHS, 0, &CxtI) > 1)
864 KnownBits LHSKnown = computeKnownBits(LHS, 0, &CxtI);
866 KnownBits RHSKnown = computeKnownBits(RHS, 0, &CxtI);
868 // Subtraction of two 2's complement numbers having identical signs will
870 if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
871 (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
874 // TODO: implement logic similar to checkRippleForAdd
878 /// \brief Return true if we can prove that:
879 /// (sub LHS, RHS) === (sub nuw LHS, RHS)
880 bool InstCombiner::willNotOverflowUnsignedSub(const Value *LHS,
882 const Instruction &CxtI) const {
883 // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
884 KnownBits LHSKnown = computeKnownBits(LHS, /*Depth=*/0, &CxtI);
885 KnownBits RHSKnown = computeKnownBits(RHS, /*Depth=*/0, &CxtI);
886 if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
892 // Checks if any operand is negative and we can convert add to sub.
893 // This function checks for following negative patterns
894 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
895 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
896 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
897 static Value *checkForNegativeOperand(BinaryOperator &I,
898 InstCombiner::BuilderTy *Builder) {
899 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
901 // This function creates 2 instructions to replace ADD, we need at least one
902 // of LHS or RHS to have one use to ensure benefit in transform.
903 if (!LHS->hasOneUse() && !RHS->hasOneUse())
906 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
907 const APInt *C1 = nullptr, *C2 = nullptr;
909 // if ONE is on other side, swap
910 if (match(RHS, m_Add(m_Value(X), m_One())))
913 if (match(LHS, m_Add(m_Value(X), m_One()))) {
914 // if XOR on other side, swap
915 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
918 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
919 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
920 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
921 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
922 Value *NewAnd = Builder->CreateAnd(Z, *C1);
923 return Builder->CreateSub(RHS, NewAnd, "sub");
924 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
925 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
926 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
927 Value *NewOr = Builder->CreateOr(Z, ~(*C1));
928 return Builder->CreateSub(RHS, NewOr, "sub");
933 // Restore LHS and RHS
934 LHS = I.getOperand(0);
935 RHS = I.getOperand(1);
937 // if XOR is on other side, swap
938 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
942 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
943 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
944 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
945 if (C1->countTrailingZeros() == 0)
946 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
947 Value *NewOr = Builder->CreateOr(Z, ~(*C2));
948 return Builder->CreateSub(RHS, NewOr, "sub");
953 static Instruction *foldAddWithConstant(BinaryOperator &Add,
954 InstCombiner::BuilderTy &Builder) {
955 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
957 if (!match(Op1, m_APInt(C)))
960 if (C->isSignMask()) {
961 // If wrapping is not allowed, then the addition must set the sign bit:
962 // X + (signmask) --> X | signmask
963 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
964 return BinaryOperator::CreateOr(Op0, Op1);
966 // If wrapping is allowed, then the addition flips the sign bit of LHS:
967 // X + (signmask) --> X ^ signmask
968 return BinaryOperator::CreateXor(Op0, Op1);
973 Type *Ty = Add.getType();
975 // Is this add the last step in a convoluted sext?
976 // add(zext(xor i16 X, -32768), -32768) --> sext X
977 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
978 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
979 return CastInst::Create(Instruction::SExt, X, Ty);
981 // (add (zext (add nuw X, C2)), C) --> (zext (add nuw X, C2 + C))
982 // FIXME: This should check hasOneUse to not increase the instruction count?
983 if (C->isNegative() &&
984 match(Op0, m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2)))) &&
985 C->sge(-C2->sext(C->getBitWidth()))) {
987 ConstantInt::get(X->getType(), *C2 + C->trunc(C2->getBitWidth()));
988 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
991 if (C->isOneValue() && Op0->hasOneUse()) {
992 // add (sext i1 X), 1 --> zext (not X)
993 // TODO: The smallest IR representation is (select X, 0, 1), and that would
994 // not require the one-use check. But we need to remove a transform in
995 // visitSelect and make sure that IR value tracking for select is equal or
996 // better than for these ops.
997 if (match(Op0, m_SExt(m_Value(X))) &&
998 X->getType()->getScalarSizeInBits() == 1)
999 return new ZExtInst(Builder.CreateNot(X), Ty);
1001 // Shifts and add used to flip and mask off the low bit:
1002 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
1004 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
1005 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
1006 Value *NotX = Builder.CreateNot(X);
1007 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
1014 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1015 bool Changed = SimplifyAssociativeOrCommutative(I);
1016 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1018 if (Value *V = SimplifyVectorOp(I))
1019 return replaceInstUsesWith(I, V);
1022 SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1023 SQ.getWithInstruction(&I)))
1024 return replaceInstUsesWith(I, V);
1026 // (A*B)+(A*C) -> A*(B+C) etc
1027 if (Value *V = SimplifyUsingDistributiveLaws(I))
1028 return replaceInstUsesWith(I, V);
1030 if (Instruction *X = foldAddWithConstant(I, *Builder))
1033 // FIXME: This should be moved into the above helper function to allow these
1034 // transforms for splat vectors.
1035 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1036 // zext(bool) + C -> bool ? C + 1 : C
1037 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
1038 if (ZI->getSrcTy()->isIntegerTy(1))
1039 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
1041 Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1042 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1043 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
1044 const APInt &RHSVal = CI->getValue();
1045 unsigned ExtendAmt = 0;
1046 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1047 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1048 if (XorRHS->getValue() == -RHSVal) {
1049 if (RHSVal.isPowerOf2())
1050 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1051 else if (XorRHS->getValue().isPowerOf2())
1052 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1056 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1057 if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1062 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt);
1063 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext");
1064 return BinaryOperator::CreateAShr(NewShl, ShAmt);
1067 // If this is a xor that was canonicalized from a sub, turn it back into
1068 // a sub and fuse this add with it.
1069 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1070 KnownBits LHSKnown = computeKnownBits(XorLHS, 0, &I);
1071 if ((XorRHS->getValue() | LHSKnown.Zero).isAllOnesValue())
1072 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1075 // (X + signmask) + C could have gotten canonicalized to (X^signmask) + C,
1076 // transform them into (X + (signmask ^ C))
1077 if (XorRHS->getValue().isSignMask())
1078 return BinaryOperator::CreateAdd(XorLHS,
1079 ConstantExpr::getXor(XorRHS, CI));
1083 if (isa<Constant>(RHS))
1084 if (Instruction *NV = foldOpWithConstantIntoOperand(I))
1087 if (I.getType()->getScalarType()->isIntegerTy(1))
1088 return BinaryOperator::CreateXor(LHS, RHS);
1092 BinaryOperator *New =
1093 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1));
1094 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1095 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1100 // -A + -B --> -(A + B)
1101 if (Value *LHSV = dyn_castNegVal(LHS)) {
1102 if (!isa<Constant>(RHS))
1103 if (Value *RHSV = dyn_castNegVal(RHS)) {
1104 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
1105 return BinaryOperator::CreateNeg(NewAdd);
1108 return BinaryOperator::CreateSub(RHS, LHSV);
1112 if (!isa<Constant>(RHS))
1113 if (Value *V = dyn_castNegVal(RHS))
1114 return BinaryOperator::CreateSub(LHS, V);
1116 if (Value *V = checkForNegativeOperand(I, Builder))
1117 return replaceInstUsesWith(I, V);
1119 // A+B --> A|B iff A and B have no bits set in common.
1120 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1121 return BinaryOperator::CreateOr(LHS, RHS);
1123 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
1125 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
1126 return BinaryOperator::CreateSub(SubOne(CRHS), X);
1129 // FIXME: We already did a check for ConstantInt RHS above this.
1130 // FIXME: Is this pattern covered by another fold? No regression tests fail on
1132 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1133 // (X & FF00) + xx00 -> (X+xx00) & FF00
1136 if (LHS->hasOneUse() &&
1137 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1138 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1139 // See if all bits from the first bit set in the Add RHS up are included
1140 // in the mask. First, get the rightmost bit.
1141 const APInt &AddRHSV = CRHS->getValue();
1143 // Form a mask of all bits from the lowest bit added through the top.
1144 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1146 // See if the and mask includes all of these bits.
1147 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1149 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1150 // Okay, the xform is safe. Insert the new add pronto.
1151 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
1152 return BinaryOperator::CreateAnd(NewAdd, C2);
1157 // add (select X 0 (sub n A)) A --> select X A n
1159 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1162 SI = dyn_cast<SelectInst>(RHS);
1165 if (SI && SI->hasOneUse()) {
1166 Value *TV = SI->getTrueValue();
1167 Value *FV = SI->getFalseValue();
1170 // Can we fold the add into the argument of the select?
1171 // We check both true and false select arguments for a matching subtract.
1172 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1173 // Fold the add into the true select value.
1174 return SelectInst::Create(SI->getCondition(), N, A);
1176 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1177 // Fold the add into the false select value.
1178 return SelectInst::Create(SI->getCondition(), A, N);
1182 // Check for (add (sext x), y), see if we can merge this into an
1183 // integer add followed by a sext.
1184 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
1185 // (add (sext x), cst) --> (sext (add x, cst'))
1186 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1187 if (LHSConv->hasOneUse()) {
1189 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1190 if (ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
1191 willNotOverflowSignedAdd(LHSConv->getOperand(0), CI, I)) {
1192 // Insert the new, smaller add.
1194 Builder->CreateNSWAdd(LHSConv->getOperand(0), CI, "addconv");
1195 return new SExtInst(NewAdd, I.getType());
1200 // (add (sext x), (sext y)) --> (sext (add int x, y))
1201 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
1202 // Only do this if x/y have the same type, if at least one of them has a
1203 // single use (so we don't increase the number of sexts), and if the
1204 // integer add will not overflow.
1205 if (LHSConv->getOperand(0)->getType() ==
1206 RHSConv->getOperand(0)->getType() &&
1207 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1208 willNotOverflowSignedAdd(LHSConv->getOperand(0),
1209 RHSConv->getOperand(0), I)) {
1210 // Insert the new integer add.
1211 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1212 RHSConv->getOperand(0), "addconv");
1213 return new SExtInst(NewAdd, I.getType());
1218 // Check for (add (zext x), y), see if we can merge this into an
1219 // integer add followed by a zext.
1220 if (auto *LHSConv = dyn_cast<ZExtInst>(LHS)) {
1221 // (add (zext x), cst) --> (zext (add x, cst'))
1222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1223 if (LHSConv->hasOneUse()) {
1225 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1226 if (ConstantExpr::getZExt(CI, I.getType()) == RHSC &&
1227 willNotOverflowUnsignedAdd(LHSConv->getOperand(0), CI, I)) {
1228 // Insert the new, smaller add.
1230 Builder->CreateNUWAdd(LHSConv->getOperand(0), CI, "addconv");
1231 return new ZExtInst(NewAdd, I.getType());
1236 // (add (zext x), (zext y)) --> (zext (add int x, y))
1237 if (auto *RHSConv = dyn_cast<ZExtInst>(RHS)) {
1238 // Only do this if x/y have the same type, if at least one of them has a
1239 // single use (so we don't increase the number of zexts), and if the
1240 // integer add will not overflow.
1241 if (LHSConv->getOperand(0)->getType() ==
1242 RHSConv->getOperand(0)->getType() &&
1243 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1244 willNotOverflowUnsignedAdd(LHSConv->getOperand(0),
1245 RHSConv->getOperand(0), I)) {
1246 // Insert the new integer add.
1247 Value *NewAdd = Builder->CreateNUWAdd(
1248 LHSConv->getOperand(0), RHSConv->getOperand(0), "addconv");
1249 return new ZExtInst(NewAdd, I.getType());
1254 // (add (xor A, B) (and A, B)) --> (or A, B)
1256 Value *A = nullptr, *B = nullptr;
1257 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) &&
1258 match(LHS, m_c_And(m_Specific(A), m_Specific(B))))
1259 return BinaryOperator::CreateOr(A, B);
1261 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
1262 match(RHS, m_c_And(m_Specific(A), m_Specific(B))))
1263 return BinaryOperator::CreateOr(A, B);
1266 // (add (or A, B) (and A, B)) --> (add A, B)
1268 Value *A = nullptr, *B = nullptr;
1269 if (match(RHS, m_Or(m_Value(A), m_Value(B))) &&
1270 match(LHS, m_c_And(m_Specific(A), m_Specific(B)))) {
1271 auto *New = BinaryOperator::CreateAdd(A, B);
1272 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1273 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1277 if (match(LHS, m_Or(m_Value(A), m_Value(B))) &&
1278 match(RHS, m_c_And(m_Specific(A), m_Specific(B)))) {
1279 auto *New = BinaryOperator::CreateAdd(A, B);
1280 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1281 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1286 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1287 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1288 // computeKnownBits.
1289 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1291 I.setHasNoSignedWrap(true);
1293 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1295 I.setHasNoUnsignedWrap(true);
1298 return Changed ? &I : nullptr;
1301 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
1302 bool Changed = SimplifyAssociativeOrCommutative(I);
1303 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1305 if (Value *V = SimplifyVectorOp(I))
1306 return replaceInstUsesWith(I, V);
1308 if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(),
1309 SQ.getWithInstruction(&I)))
1310 return replaceInstUsesWith(I, V);
1312 if (isa<Constant>(RHS))
1313 if (Instruction *FoldedFAdd = foldOpWithConstantIntoOperand(I))
1317 // -A + -B --> -(A + B)
1318 if (Value *LHSV = dyn_castFNegVal(LHS)) {
1319 Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV);
1320 RI->copyFastMathFlags(&I);
1325 if (!isa<Constant>(RHS))
1326 if (Value *V = dyn_castFNegVal(RHS)) {
1327 Instruction *RI = BinaryOperator::CreateFSub(LHS, V);
1328 RI->copyFastMathFlags(&I);
1332 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1333 // integer add followed by a promotion.
1334 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1335 Value *LHSIntVal = LHSConv->getOperand(0);
1336 Type *FPType = LHSConv->getType();
1338 // TODO: This check is overly conservative. In many cases known bits
1339 // analysis can tell us that the result of the addition has less significant
1340 // bits than the integer type can hold.
1341 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1342 Type *FScalarTy = FTy->getScalarType();
1343 Type *IScalarTy = ITy->getScalarType();
1345 // Do we have enough bits in the significand to represent the result of
1346 // the integer addition?
1347 unsigned MaxRepresentableBits =
1348 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1349 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1352 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1353 // ... if the constant fits in the integer value. This is useful for things
1354 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1355 // requires a constant pool load, and generally allows the add to be better
1357 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1358 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1360 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1361 if (LHSConv->hasOneUse() &&
1362 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1363 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1364 // Insert the new integer add.
1365 Value *NewAdd = Builder->CreateNSWAdd(LHSIntVal,
1367 return new SIToFPInst(NewAdd, I.getType());
1371 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1372 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1373 Value *RHSIntVal = RHSConv->getOperand(0);
1374 // It's enough to check LHS types only because we require int types to
1375 // be the same for this transform.
1376 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1377 // Only do this if x/y have the same type, if at least one of them has a
1378 // single use (so we don't increase the number of int->fp conversions),
1379 // and if the integer add will not overflow.
1380 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1381 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1382 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1383 // Insert the new integer add.
1384 Value *NewAdd = Builder->CreateNSWAdd(LHSIntVal,
1385 RHSIntVal, "addconv");
1386 return new SIToFPInst(NewAdd, I.getType());
1392 // select C, 0, B + select C, A, 0 -> select C, A, B
1394 Value *A1, *B1, *C1, *A2, *B2, *C2;
1395 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) &&
1396 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) {
1398 Constant *Z1=nullptr, *Z2=nullptr;
1399 Value *A, *B, *C=C1;
1400 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) {
1401 Z1 = dyn_cast<Constant>(A1); A = A2;
1402 Z2 = dyn_cast<Constant>(B2); B = B1;
1403 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) {
1404 Z1 = dyn_cast<Constant>(B1); B = B2;
1405 Z2 = dyn_cast<Constant>(A2); A = A1;
1409 (I.hasNoSignedZeros() ||
1410 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) {
1411 return SelectInst::Create(C, A, B);
1417 if (I.hasUnsafeAlgebra()) {
1418 if (Value *V = FAddCombine(Builder).simplify(&I))
1419 return replaceInstUsesWith(I, V);
1422 return Changed ? &I : nullptr;
1425 /// Optimize pointer differences into the same array into a size. Consider:
1426 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1427 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1429 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1431 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1433 bool Swapped = false;
1434 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1436 // For now we require one side to be the base pointer "A" or a constant
1437 // GEP derived from it.
1438 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1440 if (LHSGEP->getOperand(0) == RHS) {
1443 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1444 // (gep X, ...) - (gep X, ...)
1445 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1446 RHSGEP->getOperand(0)->stripPointerCasts()) {
1454 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1456 if (RHSGEP->getOperand(0) == LHS) {
1459 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1460 // (gep X, ...) - (gep X, ...)
1461 if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1462 LHSGEP->getOperand(0)->stripPointerCasts()) {
1470 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and
1473 (GEP2 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse()))
1476 // Emit the offset of the GEP and an intptr_t.
1477 Value *Result = EmitGEPOffset(GEP1);
1479 // If we had a constant expression GEP on the other side offsetting the
1480 // pointer, subtract it from the offset we have.
1482 Value *Offset = EmitGEPOffset(GEP2);
1483 Result = Builder->CreateSub(Result, Offset);
1486 // If we have p - gep(p, ...) then we have to negate the result.
1488 Result = Builder->CreateNeg(Result, "diff.neg");
1490 return Builder->CreateIntCast(Result, Ty, true);
1493 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1494 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1496 if (Value *V = SimplifyVectorOp(I))
1497 return replaceInstUsesWith(I, V);
1500 SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1501 SQ.getWithInstruction(&I)))
1502 return replaceInstUsesWith(I, V);
1504 // (A*B)-(A*C) -> A*(B-C) etc
1505 if (Value *V = SimplifyUsingDistributiveLaws(I))
1506 return replaceInstUsesWith(I, V);
1508 // If this is a 'B = x-(-A)', change to B = x+A.
1509 if (Value *V = dyn_castNegVal(Op1)) {
1510 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1512 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1513 assert(BO->getOpcode() == Instruction::Sub &&
1514 "Expected a subtraction operator!");
1515 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1516 Res->setHasNoSignedWrap(true);
1518 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1519 Res->setHasNoSignedWrap(true);
1525 if (I.getType()->getScalarType()->isIntegerTy(1))
1526 return BinaryOperator::CreateXor(Op0, Op1);
1528 // Replace (-1 - A) with (~A).
1529 if (match(Op0, m_AllOnes()))
1530 return BinaryOperator::CreateNot(Op1);
1532 if (Constant *C = dyn_cast<Constant>(Op0)) {
1533 // C - ~X == X + (1+C)
1535 if (match(Op1, m_Not(m_Value(X))))
1536 return BinaryOperator::CreateAdd(X, AddOne(C));
1538 // Try to fold constant sub into select arguments.
1539 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1540 if (Instruction *R = FoldOpIntoSelect(I, SI))
1543 // Try to fold constant sub into PHI values.
1544 if (PHINode *PN = dyn_cast<PHINode>(Op1))
1545 if (Instruction *R = foldOpIntoPhi(I, PN))
1548 // C-(X+C2) --> (C-C2)-X
1550 if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1551 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1553 // Fold (sub 0, (zext bool to B)) --> (sext bool to B)
1554 if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X))))
1555 if (X->getType()->getScalarType()->isIntegerTy(1))
1556 return CastInst::CreateSExtOrBitCast(X, Op1->getType());
1558 // Fold (sub 0, (sext bool to B)) --> (zext bool to B)
1559 if (C->isNullValue() && match(Op1, m_SExt(m_Value(X))))
1560 if (X->getType()->getScalarType()->isIntegerTy(1))
1561 return CastInst::CreateZExtOrBitCast(X, Op1->getType());
1565 if (match(Op0, m_APInt(Op0C))) {
1566 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1568 // -(X >>u 31) -> (X >>s 31)
1569 // -(X >>s 31) -> (X >>u 31)
1570 if (Op0C->isNullValue()) {
1573 if (match(Op1, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1574 *ShAmt == BitWidth - 1) {
1575 Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1576 return BinaryOperator::CreateAShr(X, ShAmtOp);
1578 if (match(Op1, m_AShr(m_Value(X), m_APInt(ShAmt))) &&
1579 *ShAmt == BitWidth - 1) {
1580 Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1581 return BinaryOperator::CreateLShr(X, ShAmtOp);
1585 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1587 if (Op0C->isMask()) {
1588 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1589 if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
1590 return BinaryOperator::CreateXor(Op1, Op0);
1596 // X-(X+Y) == -Y X-(Y+X) == -Y
1597 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1598 return BinaryOperator::CreateNeg(Y);
1601 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1602 return BinaryOperator::CreateNeg(Y);
1605 // (sub (or A, B) (xor A, B)) --> (and A, B)
1608 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1609 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1610 return BinaryOperator::CreateAnd(A, B);
1615 // ((X | Y) - X) --> (~X & Y)
1616 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1617 return BinaryOperator::CreateAnd(
1618 Y, Builder->CreateNot(Op1, Op1->getName() + ".not"));
1621 if (Op1->hasOneUse()) {
1622 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1623 Constant *C = nullptr;
1625 // (X - (Y - Z)) --> (X + (Z - Y)).
1626 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1627 return BinaryOperator::CreateAdd(Op0,
1628 Builder->CreateSub(Z, Y, Op1->getName()));
1630 // (X - (X & Y)) --> (X & ~Y)
1632 if (match(Op1, m_c_And(m_Value(Y), m_Specific(Op0))))
1633 return BinaryOperator::CreateAnd(Op0,
1634 Builder->CreateNot(Y, Y->getName() + ".not"));
1636 // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow.
1637 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1638 C->isNotMinSignedValue() && !C->isOneValue())
1639 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1641 // 0 - (X << Y) -> (-X << Y) when X is freely negatable.
1642 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1643 if (Value *XNeg = dyn_castNegVal(X))
1644 return BinaryOperator::CreateShl(XNeg, Y);
1646 // Subtracting -1/0 is the same as adding 1/0:
1647 // sub [nsw] Op0, sext(bool Y) -> add [nsw] Op0, zext(bool Y)
1648 // 'nuw' is dropped in favor of the canonical form.
1649 if (match(Op1, m_SExt(m_Value(Y))) &&
1650 Y->getType()->getScalarSizeInBits() == 1) {
1651 Value *Zext = Builder->CreateZExt(Y, I.getType());
1652 BinaryOperator *Add = BinaryOperator::CreateAdd(Op0, Zext);
1653 Add->setHasNoSignedWrap(I.hasNoSignedWrap());
1657 // X - A*-B -> X + A*B
1658 // X - -A*B -> X + A*B
1661 if (match(Op1, m_c_Mul(m_Value(A), m_Neg(m_Value(B)))))
1662 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B));
1664 // X - A*CI -> X + A*-CI
1665 // No need to handle commuted multiply because multiply handling will
1666 // ensure constant will be move to the right hand side.
1667 if (match(Op1, m_Mul(m_Value(A), m_Constant(CI)))) {
1668 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI));
1669 return BinaryOperator::CreateAdd(Op0, NewMul);
1673 // Optimize pointer differences into the same array into a size. Consider:
1674 // &A[10] - &A[0]: we should compile this to "10".
1675 Value *LHSOp, *RHSOp;
1676 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1677 match(Op1, m_PtrToInt(m_Value(RHSOp))))
1678 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1679 return replaceInstUsesWith(I, Res);
1681 // trunc(p)-trunc(q) -> trunc(p-q)
1682 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1683 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1684 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1685 return replaceInstUsesWith(I, Res);
1687 bool Changed = false;
1688 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1690 I.setHasNoSignedWrap(true);
1692 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1694 I.setHasNoUnsignedWrap(true);
1697 return Changed ? &I : nullptr;
1700 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
1701 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1703 if (Value *V = SimplifyVectorOp(I))
1704 return replaceInstUsesWith(I, V);
1706 if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(),
1707 SQ.getWithInstruction(&I)))
1708 return replaceInstUsesWith(I, V);
1710 // fsub nsz 0, X ==> fsub nsz -0.0, X
1711 if (I.getFastMathFlags().noSignedZeros() && match(Op0, m_Zero())) {
1712 // Subtraction from -0.0 is the canonical form of fneg.
1713 Instruction *NewI = BinaryOperator::CreateFNeg(Op1);
1714 NewI->copyFastMathFlags(&I);
1718 if (isa<Constant>(Op0))
1719 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1720 if (Instruction *NV = FoldOpIntoSelect(I, SI))
1723 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking
1724 // through FP extensions/truncations along the way.
1725 if (Value *V = dyn_castFNegVal(Op1)) {
1726 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V);
1727 NewI->copyFastMathFlags(&I);
1730 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) {
1731 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) {
1732 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType());
1733 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc);
1734 NewI->copyFastMathFlags(&I);
1737 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) {
1738 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
1739 Value *NewExt = Builder->CreateFPExt(V, I.getType());
1740 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt);
1741 NewI->copyFastMathFlags(&I);
1746 if (I.hasUnsafeAlgebra()) {
1747 if (Value *V = FAddCombine(Builder).simplify(&I))
1748 return replaceInstUsesWith(I, V);