1 //===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements the visit functions for add, fadd, sub, and fsub.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APFloat.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/Constant.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/InstrTypes.h"
23 #include "llvm/IR/Instruction.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/Operator.h"
26 #include "llvm/IR/PatternMatch.h"
27 #include "llvm/IR/Type.h"
28 #include "llvm/IR/Value.h"
29 #include "llvm/Support/AlignOf.h"
30 #include "llvm/Support/Casting.h"
31 #include "llvm/Support/KnownBits.h"
32 #include "llvm/Transforms/InstCombine/InstCombiner.h"
37 using namespace PatternMatch;
39 #define DEBUG_TYPE "instcombine"
43 /// Class representing coefficient of floating-point addend.
44 /// This class needs to be highly efficient, which is especially true for
45 /// the constructor. As of I write this comment, the cost of the default
46 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
47 /// perform write-merging).
51 // The constructor has to initialize a APFloat, which is unnecessary for
52 // most addends which have coefficient either 1 or -1. So, the constructor
53 // is expensive. In order to avoid the cost of the constructor, we should
54 // reuse some instances whenever possible. The pre-created instances
55 // FAddCombine::Add[0-5] embodies this idea.
56 FAddendCoef() = default;
59 // If possible, don't define operator+/operator- etc because these
60 // operators inevitably call FAddendCoef's constructor which is not cheap.
61 void operator=(const FAddendCoef &A);
62 void operator+=(const FAddendCoef &A);
63 void operator*=(const FAddendCoef &S);
66 assert(!insaneIntVal(C) && "Insane coefficient");
67 IsFp = false; IntVal = C;
70 void set(const APFloat& C);
74 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75 Value *getValue(Type *) const;
77 bool isOne() const { return isInt() && IntVal == 1; }
78 bool isTwo() const { return isInt() && IntVal == 2; }
79 bool isMinusOne() const { return isInt() && IntVal == -1; }
80 bool isMinusTwo() const { return isInt() && IntVal == -2; }
83 bool insaneIntVal(int V) { return V > 4 || V < -4; }
85 APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
91 const APFloat &getFpVal() const {
92 assert(IsFp && BufHasFpVal && "Incorret state");
93 return *getFpValPtr();
97 assert(IsFp && BufHasFpVal && "Incorret state");
98 return *getFpValPtr();
101 bool isInt() const { return !IsFp; }
103 // If the coefficient is represented by an integer, promote it to a
105 void convertToFpType(const fltSemantics &Sem);
107 // Construct an APFloat from a signed integer.
108 // TODO: We should get rid of this function when APFloat can be constructed
109 // from an *SIGNED* integer.
110 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
114 // True iff FpValBuf contains an instance of APFloat.
115 bool BufHasFpVal = false;
117 // The integer coefficient of an individual addend is either 1 or -1,
118 // and we try to simplify at most 4 addends from neighboring at most
119 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120 // is overkill of this end.
123 AlignedCharArrayUnion<APFloat> FpValBuf;
126 /// FAddend is used to represent floating-point addend. An addend is
127 /// represented as <C, V>, where the V is a symbolic value, and C is a
128 /// constant coefficient. A constant addend is represented as <C, 0>.
133 void operator+=(const FAddend &T) {
134 assert((Val == T.Val) && "Symbolic-values disagree");
138 Value *getSymVal() const { return Val; }
139 const FAddendCoef &getCoef() const { return Coeff; }
141 bool isConstant() const { return Val == nullptr; }
142 bool isZero() const { return Coeff.isZero(); }
144 void set(short Coefficient, Value *V) {
145 Coeff.set(Coefficient);
148 void set(const APFloat &Coefficient, Value *V) {
149 Coeff.set(Coefficient);
152 void set(const ConstantFP *Coefficient, Value *V) {
153 Coeff.set(Coefficient->getValueAPF());
157 void negate() { Coeff.negate(); }
159 /// Drill down the U-D chain one step to find the definition of V, and
160 /// try to break the definition into one or two addends.
161 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
163 /// Similar to FAddend::drillDownOneStep() except that the value being
164 /// splitted is the addend itself.
165 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
168 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
170 // This addend has the value of "Coeff * Val".
171 Value *Val = nullptr;
175 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176 /// with its neighboring at most two instructions.
180 FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
182 Value *simplify(Instruction *FAdd);
185 using AddendVect = SmallVector<const FAddend *, 4>;
187 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
189 /// Convert given addend to a Value
190 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
192 /// Return the number of instructions needed to emit the N-ary addition.
193 unsigned calcInstrNumber(const AddendVect& Vect);
195 Value *createFSub(Value *Opnd0, Value *Opnd1);
196 Value *createFAdd(Value *Opnd0, Value *Opnd1);
197 Value *createFMul(Value *Opnd0, Value *Opnd1);
198 Value *createFNeg(Value *V);
199 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
202 // Debugging stuff are clustered here.
204 unsigned CreateInstrNum;
205 void initCreateInstNum() { CreateInstrNum = 0; }
206 void incCreateInstNum() { CreateInstrNum++; }
208 void initCreateInstNum() {}
209 void incCreateInstNum() {}
212 InstCombiner::BuilderTy &Builder;
213 Instruction *Instr = nullptr;
216 } // end anonymous namespace
218 //===----------------------------------------------------------------------===//
221 // {FAddendCoef, FAddend, FAddition, FAddCombine}.
223 //===----------------------------------------------------------------------===//
224 FAddendCoef::~FAddendCoef() {
226 getFpValPtr()->~APFloat();
229 void FAddendCoef::set(const APFloat& C) {
230 APFloat *P = getFpValPtr();
233 // As the buffer is meanless byte stream, we cannot call
234 // APFloat::operator=().
239 IsFp = BufHasFpVal = true;
242 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
246 APFloat *P = getFpValPtr();
248 new(P) APFloat(Sem, IntVal);
250 new(P) APFloat(Sem, 0 - IntVal);
253 IsFp = BufHasFpVal = true;
256 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
258 return APFloat(Sem, Val);
260 APFloat T(Sem, 0 - Val);
266 void FAddendCoef::operator=(const FAddendCoef &That) {
270 set(That.getFpVal());
273 void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
277 IntVal += That.IntVal;
279 getFpVal().add(That.getFpVal(), RndMode);
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
294 void FAddendCoef::operator*=(const FAddendCoef &That) {
298 if (That.isMinusOne()) {
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value");
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
324 void FAddendCoef::negate() {
328 getFpVal().changeSign();
331 Value *FAddendCoef::getValue(Type *Ty) const {
333 ConstantFP::get(Ty, float(IntVal)) :
334 ConstantFP::get(Ty->getContext(), getFpVal());
337 // The definition of <Val> Addends
338 // =========================================
339 // A + B <1, A>, <1,B>
340 // A - B <1, A>, <1,B>
343 // A + C <1, A> <C, NULL>
344 // 0 +/- 0 <0, NULL> (corner case)
346 // Legend: A and B are not constant, C is constant
347 unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
353 unsigned Opcode = I->getOpcode();
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
357 Value *Opnd0 = I->getOperand(0);
358 Value *Opnd1 = I->getOperand(1);
359 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
362 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
367 Addend0.set(1, Opnd0);
369 Addend0.set(C0, nullptr);
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
375 Addend.set(1, Opnd1);
377 Addend.set(C1, nullptr);
378 if (Opcode == Instruction::FSub)
383 return Opnd0 && Opnd1 ? 2 : 1;
385 // Both operands are zero. Weird!
386 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(0);
392 Value *V1 = I->getOperand(1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
407 // Try to break *this* addend into two addends. e.g. Suppose this addend is
408 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409 // i.e. <2.3, X> and <2.3, Y>.
410 unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
419 Addend0.Scale(Coeff);
422 Addend1.Scale(Coeff);
427 Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429 "Expected 'reassoc'+'nsz' instruction");
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
435 assert((I->getOpcode() == Instruction::FAdd ||
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
438 // Save the instruction before calling other member-functions.
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
514 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends");
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
522 // Simplified addends are placed <SimpVect>.
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
530 const FAddend *ThisAddend = Addends[SymIdx];
532 // This addend was processed before.
536 Value *Val = ThisAddend->getSymVal();
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(ThisAddend);
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(T);
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(StartIdx);
574 SimpVect.push_back(&R);
579 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
580 "out-of-bound access");
583 if (!SimpVect.empty())
584 Result = createNaryFAdd(SimpVect, InstrQuota);
586 // The addition is folded to 0.0.
587 Result = ConstantFP::get(Instr->getType(), 0.0);
593 Value *FAddCombine::createNaryFAdd
594 (const AddendVect &Opnds, unsigned InstrQuota) {
595 assert(!Opnds.empty() && "Expect at least one addend");
597 // Step 1: Check if the # of instructions needed exceeds the quota.
599 unsigned InstrNeeded = calcInstrNumber(Opnds);
600 if (InstrNeeded > InstrQuota)
605 // step 2: Emit the N-ary addition.
606 // Note that at most three instructions are involved in Fadd-InstCombine: the
607 // addition in question, and at most two neighboring instructions.
608 // The resulting optimized addition should have at least one less instruction
609 // than the original addition expression tree. This implies that the resulting
610 // N-ary addition has at most two instructions, and we don't need to worry
611 // about tree-height when constructing the N-ary addition.
613 Value *LastVal = nullptr;
614 bool LastValNeedNeg = false;
616 // Iterate the addends, creating fadd/fsub using adjacent two addends.
617 for (const FAddend *Opnd : Opnds) {
619 Value *V = createAddendVal(*Opnd, NeedNeg);
622 LastValNeedNeg = NeedNeg;
626 if (LastValNeedNeg == NeedNeg) {
627 LastVal = createFAdd(LastVal, V);
632 LastVal = createFSub(V, LastVal);
634 LastVal = createFSub(LastVal, V);
636 LastValNeedNeg = false;
639 if (LastValNeedNeg) {
640 LastVal = createFNeg(LastVal);
644 assert(CreateInstrNum == InstrNeeded &&
645 "Inconsistent in instruction numbers");
651 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
652 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
653 if (Instruction *I = dyn_cast<Instruction>(V))
654 createInstPostProc(I);
658 Value *FAddCombine::createFNeg(Value *V) {
659 Value *NewV = Builder.CreateFNeg(V);
660 if (Instruction *I = dyn_cast<Instruction>(NewV))
661 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
665 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
666 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
667 if (Instruction *I = dyn_cast<Instruction>(V))
668 createInstPostProc(I);
672 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
673 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
674 if (Instruction *I = dyn_cast<Instruction>(V))
675 createInstPostProc(I);
679 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
680 NewInstr->setDebugLoc(Instr->getDebugLoc());
682 // Keep track of the number of instruction created.
686 // Propagate fast-math flags
687 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
690 // Return the number of instruction needed to emit the N-ary addition.
691 // NOTE: Keep this function in sync with createAddendVal().
692 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
693 unsigned OpndNum = Opnds.size();
694 unsigned InstrNeeded = OpndNum - 1;
696 // The number of addends in the form of "(-1)*x".
697 unsigned NegOpndNum = 0;
699 // Adjust the number of instructions needed to emit the N-ary add.
700 for (const FAddend *Opnd : Opnds) {
701 if (Opnd->isConstant())
704 // The constant check above is really for a few special constant
706 if (isa<UndefValue>(Opnd->getSymVal()))
709 const FAddendCoef &CE = Opnd->getCoef();
710 if (CE.isMinusOne() || CE.isMinusTwo())
713 // Let the addend be "c * x". If "c == +/-1", the value of the addend
714 // is immediately available; otherwise, it needs exactly one instruction
715 // to evaluate the value.
716 if (!CE.isMinusOne() && !CE.isOne())
722 // Input Addend Value NeedNeg(output)
723 // ================================================================
724 // Constant C C false
725 // <+/-1, V> V coefficient is -1
726 // <2/-2, V> "fadd V, V" coefficient is -2
727 // <C, V> "fmul V, C" false
729 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
730 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
731 const FAddendCoef &Coeff = Opnd.getCoef();
733 if (Opnd.isConstant()) {
735 return Coeff.getValue(Instr->getType());
738 Value *OpndVal = Opnd.getSymVal();
740 if (Coeff.isMinusOne() || Coeff.isOne()) {
741 NeedNeg = Coeff.isMinusOne();
745 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
746 NeedNeg = Coeff.isMinusTwo();
747 return createFAdd(OpndVal, OpndVal);
751 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
754 // Checks if any operand is negative and we can convert add to sub.
755 // This function checks for following negative patterns
756 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
757 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
758 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
759 static Value *checkForNegativeOperand(BinaryOperator &I,
760 InstCombiner::BuilderTy &Builder) {
761 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
763 // This function creates 2 instructions to replace ADD, we need at least one
764 // of LHS or RHS to have one use to ensure benefit in transform.
765 if (!LHS->hasOneUse() && !RHS->hasOneUse())
768 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
769 const APInt *C1 = nullptr, *C2 = nullptr;
771 // if ONE is on other side, swap
772 if (match(RHS, m_Add(m_Value(X), m_One())))
775 if (match(LHS, m_Add(m_Value(X), m_One()))) {
776 // if XOR on other side, swap
777 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
780 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
781 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
782 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
783 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
784 Value *NewAnd = Builder.CreateAnd(Z, *C1);
785 return Builder.CreateSub(RHS, NewAnd, "sub");
786 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
787 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
788 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
789 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
790 return Builder.CreateSub(RHS, NewOr, "sub");
795 // Restore LHS and RHS
796 LHS = I.getOperand(0);
797 RHS = I.getOperand(1);
799 // if XOR is on other side, swap
800 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
804 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
805 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
806 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
807 if (C1->countTrailingZeros() == 0)
808 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
809 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
810 return Builder.CreateSub(RHS, NewOr, "sub");
815 /// Wrapping flags may allow combining constants separated by an extend.
816 static Instruction *foldNoWrapAdd(BinaryOperator &Add,
817 InstCombiner::BuilderTy &Builder) {
818 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
819 Type *Ty = Add.getType();
821 if (!match(Op1, m_Constant(Op1C)))
824 // Try this match first because it results in an add in the narrow type.
825 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
827 const APInt *C1, *C2;
828 if (match(Op1, m_APInt(C1)) &&
829 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
830 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
832 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
833 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
836 // More general combining of constants in the wide type.
837 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
839 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
840 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
841 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
842 Value *WideX = Builder.CreateSExt(X, Ty);
843 return BinaryOperator::CreateAdd(WideX, NewC);
845 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
846 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
847 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
848 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
849 Value *WideX = Builder.CreateZExt(X, Ty);
850 return BinaryOperator::CreateAdd(WideX, NewC);
856 Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
857 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
859 if (!match(Op1, m_ImmConstant(Op1C)))
862 if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
868 // add (sub C1, X), C2 --> sub (add C1, C2), X
869 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
870 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
874 // add (sub X, Y), -1 --> add (not Y), X
875 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
876 match(Op1, m_AllOnes()))
877 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
879 // zext(bool) + C -> bool ? C + 1 : C
880 if (match(Op0, m_ZExt(m_Value(X))) &&
881 X->getType()->getScalarSizeInBits() == 1)
882 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
883 // sext(bool) + C -> bool ? C - 1 : C
884 if (match(Op0, m_SExt(m_Value(X))) &&
885 X->getType()->getScalarSizeInBits() == 1)
886 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
888 // ~X + C --> (C-1) - X
889 if (match(Op0, m_Not(m_Value(X))))
890 return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X);
893 if (!match(Op1, m_APInt(C)))
896 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
898 if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
899 haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
900 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
902 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
904 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
905 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
907 if (C->isSignMask()) {
908 // If wrapping is not allowed, then the addition must set the sign bit:
909 // X + (signmask) --> X | signmask
910 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
911 return BinaryOperator::CreateOr(Op0, Op1);
913 // If wrapping is allowed, then the addition flips the sign bit of LHS:
914 // X + (signmask) --> X ^ signmask
915 return BinaryOperator::CreateXor(Op0, Op1);
918 // Is this add the last step in a convoluted sext?
919 // add(zext(xor i16 X, -32768), -32768) --> sext X
920 Type *Ty = Add.getType();
921 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
922 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
923 return CastInst::Create(Instruction::SExt, X, Ty);
925 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
926 // (X ^ signmask) + C --> (X + (signmask ^ C))
927 if (C2->isSignMask())
928 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
930 // If X has no high-bits set above an xor mask:
931 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
933 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
934 if ((*C2 | LHSKnown.Zero).isAllOnes())
935 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
938 // Look for a math+logic pattern that corresponds to sext-in-register of a
939 // value with cleared high bits. Convert that into a pair of shifts:
940 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
941 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
942 if (Op0->hasOneUse() && *C2 == -(*C)) {
943 unsigned BitWidth = Ty->getScalarSizeInBits();
946 ShAmt = BitWidth - C->logBase2() - 1;
947 else if (C2->isPowerOf2())
948 ShAmt = BitWidth - C2->logBase2() - 1;
949 if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
951 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
952 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
953 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
958 if (C->isOne() && Op0->hasOneUse()) {
959 // add (sext i1 X), 1 --> zext (not X)
960 // TODO: The smallest IR representation is (select X, 0, 1), and that would
961 // not require the one-use check. But we need to remove a transform in
962 // visitSelect and make sure that IR value tracking for select is equal or
963 // better than for these ops.
964 if (match(Op0, m_SExt(m_Value(X))) &&
965 X->getType()->getScalarSizeInBits() == 1)
966 return new ZExtInst(Builder.CreateNot(X), Ty);
968 // Shifts and add used to flip and mask off the low bit:
969 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
971 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
972 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
973 Value *NotX = Builder.CreateNot(X);
974 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
978 // If all bits affected by the add are included in a high-bit-mask, do the
979 // add before the mask op:
980 // (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00
981 if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) &&
982 C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) {
983 Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C));
984 return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2));
990 // Matches multiplication expression Op * C where C is a constant. Returns the
991 // constant value in C and the other operand in Op. Returns true if such a
993 static bool MatchMul(Value *E, Value *&Op, APInt &C) {
995 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
999 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1000 C = APInt(AI->getBitWidth(), 1);
1007 // Matches remainder expression Op % C where C is a constant. Returns the
1008 // constant value in C and the other operand in Op. Returns the signedness of
1009 // the remainder operation in IsSigned. Returns true if such a match is
1011 static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1014 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1019 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1023 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1030 // Matches division expression Op / C with the given signedness as indicated
1031 // by IsSigned, where C is a constant. Returns the constant value in C and the
1032 // other operand in Op. Returns true if such a match is found.
1033 static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1035 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1040 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1044 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1045 C = APInt(AI->getBitWidth(), 1);
1053 // Returns whether C0 * C1 with the given signedness overflows.
1054 static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1057 (void)C0.smul_ov(C1, overflow);
1059 (void)C0.umul_ov(C1, overflow);
1063 // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1064 // does not overflow.
1065 Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1066 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1070 // Match I = X % C0 + MulOpV * C0
1071 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1072 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1077 // Match MulOpC = RemOpV % C1
1078 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1079 IsSigned == Rem2IsSigned) {
1082 // Match RemOpV = X / C0
1083 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1084 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1085 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1086 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1087 : Builder.CreateURem(X, NewDivisor, "urem");
1096 /// (1 << NBits) - 1
1098 /// ~(-(1 << NBits))
1099 /// Because a 'not' is better for bit-tracking analysis and other transforms
1100 /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1101 static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1102 InstCombiner::BuilderTy &Builder) {
1104 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1107 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1108 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1109 // Be wary of constant folding.
1110 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1111 // Always NSW. But NUW propagates from `add`.
1112 BOp->setHasNoSignedWrap();
1113 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1116 return BinaryOperator::CreateNot(NotMask, I.getName());
1119 static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1120 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1121 Type *Ty = I.getType();
1122 auto getUAddSat = [&]() {
1123 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1126 // add (umin X, ~Y), Y --> uaddsat X, Y
1128 if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1130 return CallInst::Create(getUAddSat(), { X, Y });
1132 // add (umin X, ~C), C --> uaddsat X, C
1133 const APInt *C, *NotC;
1134 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1136 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1141 Instruction *InstCombinerImpl::
1142 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1143 BinaryOperator &I) {
1144 assert((I.getOpcode() == Instruction::Add ||
1145 I.getOpcode() == Instruction::Or ||
1146 I.getOpcode() == Instruction::Sub) &&
1147 "Expecting add/or/sub instruction");
1149 // We have a subtraction/addition between a (potentially truncated) *logical*
1150 // right-shift of X and a "select".
1152 Instruction *LowBitsToSkip, *Extract;
1153 if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd(
1154 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1155 m_Instruction(Extract))),
1159 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1160 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1163 Type *XTy = X->getType();
1164 bool HadTrunc = I.getType() != XTy;
1166 // If there was a truncation of extracted value, then we'll need to produce
1167 // one extra instruction, so we need to ensure one instruction will go away.
1168 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1171 // Extraction should extract high NBits bits, with shift amount calculated as:
1172 // low bits to skip = shift bitwidth - high bits to extract
1173 // The shift amount itself may be extended, and we need to look past zero-ext
1174 // when matching NBits, that will matter for matching later.
1179 m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1180 !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1181 APInt(C->getType()->getScalarSizeInBits(),
1182 X->getType()->getScalarSizeInBits()))))
1185 // Sign-extending value can be zero-extended if we `sub`tract it,
1186 // or sign-extended otherwise.
1187 auto SkipExtInMagic = [&I](Value *&V) {
1188 if (I.getOpcode() == Instruction::Sub)
1189 match(V, m_ZExtOrSelf(m_Value(V)));
1191 match(V, m_SExtOrSelf(m_Value(V)));
1194 // Now, finally validate the sign-extending magic.
1195 // `select` itself may be appropriately extended, look past that.
1196 SkipExtInMagic(Select);
1198 ICmpInst::Predicate Pred;
1200 Value *SignExtendingValue, *Zero;
1202 // It must be a select between two values we will later establish to be a
1203 // sign-extending value and a zero constant. The condition guarding the
1204 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1205 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1206 m_Value(SignExtendingValue), m_Value(Zero))) ||
1207 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1210 // icmp-select pair is commutative.
1212 std::swap(SignExtendingValue, Zero);
1214 // If we should not perform sign-extension then we must add/or/subtract zero.
1215 if (!match(Zero, m_Zero()))
1217 // Otherwise, it should be some constant, left-shifted by the same NBits we
1218 // had in `lshr`. Said left-shift can also be appropriately extended.
1219 // Again, we must look past zero-ext when looking for NBits.
1220 SkipExtInMagic(SignExtendingValue);
1221 Constant *SignExtendingValueBaseConstant;
1222 if (!match(SignExtendingValue,
1223 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1224 m_ZExtOrSelf(m_Specific(NBits)))))
1226 // If we `sub`, then the constant should be one, else it should be all-ones.
1227 if (I.getOpcode() == Instruction::Sub
1228 ? !match(SignExtendingValueBaseConstant, m_One())
1229 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1232 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1233 Extract->getName() + ".sext");
1234 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1238 Builder.Insert(NewAShr);
1239 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1242 /// This is a specialization of a more general transform from
1243 /// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
1244 /// for multi-use cases or propagating nsw/nuw, then we would not need this.
1245 static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1246 InstCombiner::BuilderTy &Builder) {
1247 // TODO: Also handle mul by doubling the shift amount?
1248 assert((I.getOpcode() == Instruction::Add ||
1249 I.getOpcode() == Instruction::Sub) &&
1250 "Expected add/sub");
1251 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1252 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1253 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1256 Value *X, *Y, *ShAmt;
1257 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1258 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1261 // No-wrap propagates only when all ops have no-wrap.
1262 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1263 Op1->hasNoSignedWrap();
1264 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1265 Op1->hasNoUnsignedWrap();
1267 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1268 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1269 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1270 NewI->setHasNoSignedWrap(HasNSW);
1271 NewI->setHasNoUnsignedWrap(HasNUW);
1273 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1274 NewShl->setHasNoSignedWrap(HasNSW);
1275 NewShl->setHasNoUnsignedWrap(HasNUW);
1279 Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1280 if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1281 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1282 SQ.getWithInstruction(&I)))
1283 return replaceInstUsesWith(I, V);
1285 if (SimplifyAssociativeOrCommutative(I))
1288 if (Instruction *X = foldVectorBinop(I))
1291 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1294 // (A*B)+(A*C) -> A*(B+C) etc
1295 if (Value *V = SimplifyUsingDistributiveLaws(I))
1296 return replaceInstUsesWith(I, V);
1298 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1301 if (Instruction *X = foldAddWithConstant(I))
1304 if (Instruction *X = foldNoWrapAdd(I, Builder))
1307 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1308 Type *Ty = I.getType();
1309 if (Ty->isIntOrIntVectorTy(1))
1310 return BinaryOperator::CreateXor(LHS, RHS);
1314 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1315 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1316 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1321 if (match(LHS, m_Neg(m_Value(A)))) {
1322 // -A + -B --> -(A + B)
1323 if (match(RHS, m_Neg(m_Value(B))))
1324 return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1327 return BinaryOperator::CreateSub(RHS, A);
1331 if (match(RHS, m_Neg(m_Value(B))))
1332 return BinaryOperator::CreateSub(LHS, B);
1334 if (Value *V = checkForNegativeOperand(I, Builder))
1335 return replaceInstUsesWith(I, V);
1337 // (A + 1) + ~B --> A - B
1338 // ~B + (A + 1) --> A - B
1339 // (~B + A) + 1 --> A - B
1340 // (A + ~B) + 1 --> A - B
1341 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1342 match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1343 return BinaryOperator::CreateSub(A, B);
1345 // (A + RHS) + RHS --> A + (RHS << 1)
1346 if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1347 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1349 // LHS + (A + LHS) --> A + (LHS << 1)
1350 if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1351 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1354 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1356 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1357 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1358 (LHS->hasOneUse() || RHS->hasOneUse())) {
1359 Value *Sub = Builder.CreateSub(A, B);
1360 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1364 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1365 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1367 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1368 const APInt *C1, *C2;
1369 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1370 APInt one(C2->getBitWidth(), 1);
1371 APInt minusC1 = -(*C1);
1372 if (minusC1 == (one << *C2)) {
1373 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1374 return BinaryOperator::CreateSRem(RHS, NewRHS);
1378 // A+B --> A|B iff A and B have no bits set in common.
1379 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1380 return BinaryOperator::CreateOr(LHS, RHS);
1382 // add (select X 0 (sub n A)) A --> select X A n
1384 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1387 SI = dyn_cast<SelectInst>(RHS);
1390 if (SI && SI->hasOneUse()) {
1391 Value *TV = SI->getTrueValue();
1392 Value *FV = SI->getFalseValue();
1395 // Can we fold the add into the argument of the select?
1396 // We check both true and false select arguments for a matching subtract.
1397 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1398 // Fold the add into the true select value.
1399 return SelectInst::Create(SI->getCondition(), N, A);
1401 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1402 // Fold the add into the false select value.
1403 return SelectInst::Create(SI->getCondition(), A, N);
1407 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1410 // (add (xor A, B) (and A, B)) --> (or A, B)
1411 // (add (and A, B) (xor A, B)) --> (or A, B)
1412 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1413 m_c_And(m_Deferred(A), m_Deferred(B)))))
1414 return BinaryOperator::CreateOr(A, B);
1416 // (add (or A, B) (and A, B)) --> (add A, B)
1417 // (add (and A, B) (or A, B)) --> (add A, B)
1418 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1419 m_c_And(m_Deferred(A), m_Deferred(B))))) {
1420 // Replacing operands in-place to preserve nuw/nsw flags.
1421 replaceOperand(I, 0, A);
1422 replaceOperand(I, 1, B);
1426 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1427 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1428 // computeKnownBits.
1429 bool Changed = false;
1430 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1432 I.setHasNoSignedWrap(true);
1434 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1436 I.setHasNoUnsignedWrap(true);
1439 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1442 if (Instruction *V =
1443 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1446 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1449 // usub.sat(A, B) + B => umax(A, B)
1450 if (match(&I, m_c_BinOp(
1451 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1453 return replaceInstUsesWith(I,
1454 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1457 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1458 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1459 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1460 haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1461 return replaceInstUsesWith(
1462 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1463 {Builder.CreateOr(A, B)}));
1465 return Changed ? &I : nullptr;
1468 /// Eliminate an op from a linear interpolation (lerp) pattern.
1469 static Instruction *factorizeLerp(BinaryOperator &I,
1470 InstCombiner::BuilderTy &Builder) {
1472 if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1473 m_OneUse(m_FSub(m_FPOne(),
1475 m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1478 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1479 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1480 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1481 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1484 /// Factor a common operand out of fadd/fsub of fmul/fdiv.
1485 static Instruction *factorizeFAddFSub(BinaryOperator &I,
1486 InstCombiner::BuilderTy &Builder) {
1487 assert((I.getOpcode() == Instruction::FAdd ||
1488 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1489 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1490 "FP factorization requires FMF");
1492 if (Instruction *Lerp = factorizeLerp(I, Builder))
1495 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1496 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1501 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1502 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1503 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1504 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1506 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1507 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1512 // (X * Z) + (Y * Z) --> (X + Y) * Z
1513 // (X * Z) - (Y * Z) --> (X - Y) * Z
1514 // (X / Z) + (Y / Z) --> (X + Y) / Z
1515 // (X / Z) - (Y / Z) --> (X - Y) / Z
1516 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1517 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1518 : Builder.CreateFSubFMF(X, Y, &I);
1520 // Bail out if we just created a denormal constant.
1521 // TODO: This is copied from a previous implementation. Is it necessary?
1523 if (match(XY, m_APFloat(C)) && !C->isNormal())
1526 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1527 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1530 Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1531 if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1532 I.getFastMathFlags(),
1533 SQ.getWithInstruction(&I)))
1534 return replaceInstUsesWith(I, V);
1536 if (SimplifyAssociativeOrCommutative(I))
1539 if (Instruction *X = foldVectorBinop(I))
1542 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1545 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1548 // (-X) + Y --> Y - X
1550 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1551 return BinaryOperator::CreateFSubFMF(Y, X, &I);
1553 // Similar to above, but look through fmul/fdiv for the negated term.
1554 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1556 if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1558 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1559 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1561 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1562 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1563 if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1565 match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1567 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1568 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1571 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1572 // integer add followed by a promotion.
1573 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1574 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1575 Value *LHSIntVal = LHSConv->getOperand(0);
1576 Type *FPType = LHSConv->getType();
1578 // TODO: This check is overly conservative. In many cases known bits
1579 // analysis can tell us that the result of the addition has less significant
1580 // bits than the integer type can hold.
1581 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1582 Type *FScalarTy = FTy->getScalarType();
1583 Type *IScalarTy = ITy->getScalarType();
1585 // Do we have enough bits in the significand to represent the result of
1586 // the integer addition?
1587 unsigned MaxRepresentableBits =
1588 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1589 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1592 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1593 // ... if the constant fits in the integer value. This is useful for things
1594 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1595 // requires a constant pool load, and generally allows the add to be better
1597 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1598 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1600 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1601 if (LHSConv->hasOneUse() &&
1602 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1603 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1604 // Insert the new integer add.
1605 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1606 return new SIToFPInst(NewAdd, I.getType());
1610 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1611 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1612 Value *RHSIntVal = RHSConv->getOperand(0);
1613 // It's enough to check LHS types only because we require int types to
1614 // be the same for this transform.
1615 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1616 // Only do this if x/y have the same type, if at least one of them has a
1617 // single use (so we don't increase the number of int->fp conversions),
1618 // and if the integer add will not overflow.
1619 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1620 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1621 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1622 // Insert the new integer add.
1623 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1624 return new SIToFPInst(NewAdd, I.getType());
1630 // Handle specials cases for FAdd with selects feeding the operation
1631 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1632 return replaceInstUsesWith(I, V);
1634 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1635 if (Instruction *F = factorizeFAddFSub(I, Builder))
1638 // Try to fold fadd into start value of reduction intrinsic.
1639 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1640 m_AnyZeroFP(), m_Value(X))),
1642 // fadd (rdx 0.0, X), Y --> rdx Y, X
1643 return replaceInstUsesWith(
1644 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1645 {X->getType()}, {Y, X}, &I));
1647 const APFloat *StartC, *C;
1648 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1649 m_APFloat(StartC), m_Value(X)))) &&
1650 match(RHS, m_APFloat(C))) {
1651 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1652 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1653 return replaceInstUsesWith(
1654 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1655 {X->getType()}, {NewStartC, X}, &I));
1658 // (X * MulC) + X --> X * (MulC + 1.0)
1660 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1662 MulC = ConstantExpr::getFAdd(MulC, ConstantFP::get(I.getType(), 1.0));
1663 return BinaryOperator::CreateFMulFMF(X, MulC, &I);
1666 if (Value *V = FAddCombine(Builder).simplify(&I))
1667 return replaceInstUsesWith(I, V);
1673 /// Optimize pointer differences into the same array into a size. Consider:
1674 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1675 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1676 Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
1677 Type *Ty, bool IsNUW) {
1678 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1680 bool Swapped = false;
1681 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1682 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1683 std::swap(LHS, RHS);
1687 // Require at least one GEP with a common base pointer on both sides.
1688 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1690 if (LHSGEP->getOperand(0) == RHS) {
1692 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1693 // (gep X, ...) - (gep X, ...)
1694 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1695 RHSGEP->getOperand(0)->stripPointerCasts()) {
1706 // (gep X, ...) - (gep X, ...)
1708 // Avoid duplicating the arithmetic if there are more than one non-constant
1709 // indices between the two GEPs and either GEP has a non-constant index and
1710 // multiple users. If zero non-constant index, the result is a constant and
1711 // there is no duplication. If one non-constant index, the result is an add
1712 // or sub with a constant, which is no larger than the original code, and
1713 // there's no duplicated arithmetic, even if either GEP has multiple
1714 // users. If more than one non-constant indices combined, as long as the GEP
1715 // with at least one non-constant index doesn't have multiple users, there
1716 // is no duplication.
1717 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1718 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1719 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1720 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1721 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1726 // Emit the offset of the GEP and an intptr_t.
1727 Value *Result = EmitGEPOffset(GEP1);
1729 // If this is a single inbounds GEP and the original sub was nuw,
1730 // then the final multiplication is also nuw.
1731 if (auto *I = dyn_cast<Instruction>(Result))
1732 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1733 I->getOpcode() == Instruction::Mul)
1734 I->setHasNoUnsignedWrap();
1736 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
1737 // If both GEPs are inbounds, then the subtract does not have signed overflow.
1739 Value *Offset = EmitGEPOffset(GEP2);
1740 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
1741 GEP1->isInBounds() && GEP2->isInBounds());
1744 // If we have p - gep(p, ...) then we have to negate the result.
1746 Result = Builder.CreateNeg(Result, "diff.neg");
1748 return Builder.CreateIntCast(Result, Ty, true);
1751 Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
1752 if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1753 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1754 SQ.getWithInstruction(&I)))
1755 return replaceInstUsesWith(I, V);
1757 if (Instruction *X = foldVectorBinop(I))
1760 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1763 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1765 // If this is a 'B = x-(-A)', change to B = x+A.
1766 // We deal with this without involving Negator to preserve NSW flag.
1767 if (Value *V = dyn_castNegVal(Op1)) {
1768 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1770 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1771 assert(BO->getOpcode() == Instruction::Sub &&
1772 "Expected a subtraction operator!");
1773 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1774 Res->setHasNoSignedWrap(true);
1776 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1777 Res->setHasNoSignedWrap(true);
1783 // Try this before Negator to preserve NSW flag.
1784 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1788 if (match(Op0, m_ImmConstant(C))) {
1792 // C-(X+C2) --> (C-C2)-X
1793 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
1794 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1797 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
1798 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1801 bool Changed = false;
1802 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1804 I.setHasNoSignedWrap(true);
1806 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1808 I.setHasNoUnsignedWrap(true);
1811 return Changed ? &I : nullptr;
1814 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
1815 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
1816 // a pure negation used by a select that looks like abs/nabs.
1817 bool IsNegation = match(Op0, m_ZeroInt());
1818 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
1819 const Instruction *UI = dyn_cast<Instruction>(U);
1823 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
1824 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
1826 if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
1827 return BinaryOperator::CreateAdd(NegOp1, Op0);
1830 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
1832 // (A*B)-(A*C) -> A*(B-C) etc
1833 if (Value *V = SimplifyUsingDistributiveLaws(I))
1834 return replaceInstUsesWith(I, V);
1836 if (I.getType()->isIntOrIntVectorTy(1))
1837 return BinaryOperator::CreateXor(Op0, Op1);
1839 // Replace (-1 - A) with (~A).
1840 if (match(Op0, m_AllOnes()))
1841 return BinaryOperator::CreateNot(Op1);
1843 // (X + -1) - Y --> ~Y + X
1845 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1846 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1848 // Reassociate sub/add sequences to create more add instructions and
1849 // reduce dependency chains:
1850 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
1852 if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
1854 Value *XZ = Builder.CreateAdd(X, Z);
1855 Value *YW = Builder.CreateAdd(Y, Op1);
1856 return BinaryOperator::CreateSub(XZ, YW);
1859 // ((X - Y) - Op1) --> X - (Y + Op1)
1860 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
1861 Value *Add = Builder.CreateAdd(Y, Op1);
1862 return BinaryOperator::CreateSub(X, Add);
1865 // (~X) - (~Y) --> Y - X
1866 // This is placed after the other reassociations and explicitly excludes a
1867 // sub-of-sub pattern to avoid infinite looping.
1868 if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
1869 isFreeToInvert(Op1, Op1->hasOneUse()) &&
1870 !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
1871 Value *NotOp0 = Builder.CreateNot(Op0);
1872 Value *NotOp1 = Builder.CreateNot(Op1);
1873 return BinaryOperator::CreateSub(NotOp1, NotOp0);
1876 auto m_AddRdx = [](Value *&Vec) {
1877 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
1880 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
1881 V0->getType() == V1->getType()) {
1882 // Difference of sums is sum of differences:
1883 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
1884 Value *Sub = Builder.CreateSub(V0, V1);
1885 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
1886 {Sub->getType()}, {Sub});
1887 return replaceInstUsesWith(I, Rdx);
1890 if (Constant *C = dyn_cast<Constant>(Op0)) {
1892 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1893 // C - (zext bool) --> bool ? C - 1 : C
1894 return SelectInst::Create(X, InstCombiner::SubOne(C), C);
1895 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1896 // C - (sext bool) --> bool ? C + 1 : C
1897 return SelectInst::Create(X, InstCombiner::AddOne(C), C);
1899 // C - ~X == X + (1+C)
1900 if (match(Op1, m_Not(m_Value(X))))
1901 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
1903 // Try to fold constant sub into select arguments.
1904 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1905 if (Instruction *R = FoldOpIntoSelect(I, SI))
1908 // Try to fold constant sub into PHI values.
1909 if (PHINode *PN = dyn_cast<PHINode>(Op1))
1910 if (Instruction *R = foldOpIntoPhi(I, PN))
1915 // C-(C2-X) --> X+(C-C2)
1916 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
1917 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
1921 if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) {
1922 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1924 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1925 if ((*Op0C | RHSKnown.Zero).isAllOnes())
1926 return BinaryOperator::CreateXor(Op1, Op0);
1931 // X-(X+Y) == -Y X-(Y+X) == -Y
1932 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1933 return BinaryOperator::CreateNeg(Y);
1936 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1937 return BinaryOperator::CreateNeg(Y);
1940 // (sub (or A, B) (and A, B)) --> (xor A, B)
1943 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1944 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1945 return BinaryOperator::CreateXor(A, B);
1948 // (sub (add A, B) (or A, B)) --> (and A, B)
1951 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1952 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1953 return BinaryOperator::CreateAnd(A, B);
1956 // (sub (add A, B) (and A, B)) --> (or A, B)
1959 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1960 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
1961 return BinaryOperator::CreateOr(A, B);
1964 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
1967 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1968 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1969 (Op0->hasOneUse() || Op1->hasOneUse()))
1970 return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
1973 // (sub (or A, B), (xor A, B)) --> (and A, B)
1976 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1977 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1978 return BinaryOperator::CreateAnd(A, B);
1981 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
1984 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1985 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1986 (Op0->hasOneUse() || Op1->hasOneUse()))
1987 return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
1992 // ((X | Y) - X) --> (~X & Y)
1993 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1994 return BinaryOperator::CreateAnd(
1995 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1999 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2001 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2002 m_OneUse(m_Neg(m_Value(X))))))) {
2003 return BinaryOperator::CreateNeg(Builder.CreateAnd(
2004 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2009 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2011 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2012 return BinaryOperator::CreateNeg(
2013 Builder.CreateAnd(Op1, Builder.CreateNot(C)));
2018 // If we have a subtraction between some value and a select between
2019 // said value and something else, sink subtraction into select hands, i.e.:
2020 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2022 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2024 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2026 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2027 // This will result in select between new subtraction and 0.
2028 auto SinkSubIntoSelect =
2029 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2030 auto SubBuilder) -> Instruction * {
2031 Value *Cond, *TrueVal, *FalseVal;
2032 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2033 m_Value(FalseVal)))))
2035 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2037 // While it is really tempting to just create two subtractions and let
2038 // InstCombine fold one of those to 0, it isn't possible to do so
2039 // because of worklist visitation order. So ugly it is.
2040 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2041 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2042 Constant *Zero = Constant::getNullValue(Ty);
2043 SelectInst *NewSel =
2044 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2045 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2046 // Preserve prof metadata if any.
2047 NewSel->copyMetadata(cast<Instruction>(*Select));
2050 if (Instruction *NewSel = SinkSubIntoSelect(
2051 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2052 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2053 return Builder->CreateSub(OtherHandOfSelect,
2054 /*OtherHandOfSub=*/Op1);
2057 if (Instruction *NewSel = SinkSubIntoSelect(
2058 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2059 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2060 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2066 // (X - (X & Y)) --> (X & ~Y)
2067 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2068 (Op1->hasOneUse() || isa<Constant>(Y)))
2069 return BinaryOperator::CreateAnd(
2070 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2072 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2073 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2074 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2075 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2076 // As long as Y is freely invertible, this will be neutral or a win.
2077 // Note: We don't generate the inverse max/min, just create the 'not' of
2078 // it and let other folds do the rest.
2079 if (match(Op0, m_Not(m_Value(X))) &&
2080 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2081 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2082 Value *Not = Builder.CreateNot(Op1);
2083 return BinaryOperator::CreateSub(Not, X);
2085 if (match(Op1, m_Not(m_Value(X))) &&
2086 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2087 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2088 Value *Not = Builder.CreateNot(Op0);
2089 return BinaryOperator::CreateSub(X, Not);
2092 // TODO: This is the same logic as above but handles the cmp-select idioms
2093 // for min/max, so the use checks are increased to account for the
2094 // extra instructions. If we canonicalize to intrinsics, this block
2095 // can likely be removed.
2097 Value *LHS, *RHS, *A;
2098 Value *NotA = Op0, *MinMax = Op1;
2099 SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2100 if (!SelectPatternResult::isMinOrMax(SPF)) {
2103 SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2105 if (SelectPatternResult::isMinOrMax(SPF) &&
2106 match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
2108 std::swap(LHS, RHS);
2109 // LHS is now Y above and expected to have at least 2 uses (the min/max)
2110 // NotA is expected to have 2 uses from the min/max and 1 from the sub.
2111 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
2112 !NotA->hasNUsesOrMore(4)) {
2113 Value *Not = Builder.CreateNot(MinMax);
2115 return BinaryOperator::CreateSub(Not, A);
2117 return BinaryOperator::CreateSub(A, Not);
2122 // Optimize pointer differences into the same array into a size. Consider:
2123 // &A[10] - &A[0]: we should compile this to "10".
2124 Value *LHSOp, *RHSOp;
2125 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2126 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2127 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2128 I.hasNoUnsignedWrap()))
2129 return replaceInstUsesWith(I, Res);
2131 // trunc(p)-trunc(q) -> trunc(p-q)
2132 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2133 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2134 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2136 return replaceInstUsesWith(I, Res);
2138 // Canonicalize a shifty way to code absolute value to the common pattern.
2139 // There are 2 potential commuted variants.
2140 // We're relying on the fact that we only do this transform when the shift has
2141 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2145 Type *Ty = I.getType();
2146 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2147 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2148 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2149 // B = ashr i32 A, 31 ; smear the sign bit
2150 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2151 // --> (A < 0) ? -A : A
2152 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2153 // Copy the nuw/nsw flags from the sub to the negate.
2154 Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2155 I.hasNoSignedWrap());
2156 return SelectInst::Create(Cmp, Neg, A);
2159 // If we are subtracting a low-bit masked subset of some value from an add
2160 // of that same value with no low bits changed, that is clearing some low bits
2162 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2163 const APInt *AddC, *AndC;
2164 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2165 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2166 unsigned BitWidth = Ty->getScalarSizeInBits();
2167 unsigned Cttz = AddC->countTrailingZeros();
2168 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2169 if ((HighMask & *AndC).isZero())
2170 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2173 if (Instruction *V =
2174 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2177 // X - usub.sat(X, Y) => umin(X, Y)
2178 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2180 return replaceInstUsesWith(
2181 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2183 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2184 // TODO: The one-use restriction is not strictly necessary, but it may
2185 // require improving other pattern matching and/or codegen.
2186 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2187 return replaceInstUsesWith(
2188 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2190 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2191 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2192 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2193 return BinaryOperator::CreateNeg(USub);
2196 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2197 if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
2198 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2199 return replaceInstUsesWith(
2200 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2201 {Builder.CreateNot(X)}));
2203 return TryToNarrowDeduceFlags();
2206 /// This eliminates floating-point negation in either 'fneg(X)' or
2207 /// 'fsub(-0.0, X)' form by combining into a constant operand.
2208 static Instruction *foldFNegIntoConstant(Instruction &I) {
2209 // This is limited with one-use because fneg is assumed better for
2210 // reassociation and cheaper in codegen than fmul/fdiv.
2211 // TODO: Should the m_OneUse restriction be removed?
2212 Instruction *FNegOp;
2213 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2219 // Fold negation into constant operand.
2220 // -(X * C) --> X * (-C)
2221 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2222 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
2223 // -(X / C) --> X / (-C)
2224 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2225 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
2226 // -(C / X) --> (-C) / X
2227 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) {
2229 BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
2231 // Intersect 'nsz' and 'ninf' because those special value exceptions may not
2232 // apply to the fdiv. Everything else propagates from the fneg.
2233 // TODO: We could propagate nsz/ninf from fdiv alone?
2234 FastMathFlags FMF = I.getFastMathFlags();
2235 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2236 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2237 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2240 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2241 // -(X + C) --> -X + -C --> -C - X
2242 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2243 return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I);
2248 static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2249 InstCombiner::BuilderTy &Builder) {
2251 if (!match(&I, m_FNeg(m_Value(FNeg))))
2255 if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2256 return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2258 if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2259 return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2264 Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2265 Value *Op = I.getOperand(0);
2267 if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
2268 getSimplifyQuery().getWithInstruction(&I)))
2269 return replaceInstUsesWith(I, V);
2271 if (Instruction *X = foldFNegIntoConstant(I))
2276 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2277 if (I.hasNoSignedZeros() &&
2278 match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2279 return BinaryOperator::CreateFSubFMF(Y, X, &I);
2281 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2284 // Try to eliminate fneg if at least 1 arm of the select is negated.
2286 if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
2287 // Unlike most transforms, this one is not safe to propagate nsz unless
2288 // it is present on the original select. (We are conservatively intersecting
2289 // the nsz flags from the select and root fneg instruction.)
2290 auto propagateSelectFMF = [&](SelectInst *S) {
2291 S->copyFastMathFlags(&I);
2292 if (auto *OldSel = dyn_cast<SelectInst>(Op))
2293 if (!OldSel->hasNoSignedZeros())
2294 S->setHasNoSignedZeros(false);
2296 // -(Cond ? -P : Y) --> Cond ? P : -Y
2298 if (match(X, m_FNeg(m_Value(P)))) {
2299 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2300 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2301 propagateSelectFMF(NewSel);
2304 // -(Cond ? X : -P) --> Cond ? -X : P
2305 if (match(Y, m_FNeg(m_Value(P)))) {
2306 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2307 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2308 propagateSelectFMF(NewSel);
2316 Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2317 if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
2318 I.getFastMathFlags(),
2319 getSimplifyQuery().getWithInstruction(&I)))
2320 return replaceInstUsesWith(I, V);
2322 if (Instruction *X = foldVectorBinop(I))
2325 if (Instruction *Phi = foldBinopWithPhiOperands(I))
2328 // Subtraction from -0.0 is the canonical form of fneg.
2329 // fsub -0.0, X ==> fneg X
2330 // fsub nsz 0.0, X ==> fneg nsz X
2332 // FIXME This matcher does not respect FTZ or DAZ yet:
2333 // fsub -0.0, Denorm ==> +-0
2334 // fneg Denorm ==> -Denorm
2336 if (match(&I, m_FNeg(m_Value(Op))))
2337 return UnaryOperator::CreateFNegFMF(Op, &I);
2339 if (Instruction *X = foldFNegIntoConstant(I))
2342 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2348 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2349 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2350 // Canonicalize to fadd to make analysis easier.
2351 // This can also help codegen because fadd is commutative.
2352 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2353 // killed later. We still limit that particular transform with 'hasOneUse'
2354 // because an fneg is assumed better/cheaper than a generic fsub.
2355 if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2356 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2357 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2358 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2362 // (-X) - Op1 --> -(X + Op1)
2363 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2364 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2365 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2366 return UnaryOperator::CreateFNegFMF(FAdd, &I);
2369 if (isa<Constant>(Op0))
2370 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2371 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2374 // X - C --> X + (-C)
2375 // But don't transform constant expressions because there's an inverse fold
2376 // for X + (-Y) --> X - Y.
2377 if (match(Op1, m_ImmConstant(C)))
2378 return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
2380 // X - (-Y) --> X + Y
2381 if (match(Op1, m_FNeg(m_Value(Y))))
2382 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2384 // Similar to above, but look through a cast of the negated value:
2385 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2386 Type *Ty = I.getType();
2387 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2388 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2390 // X - (fpext(-Y)) --> X + fpext(Y)
2391 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2392 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2394 // Similar to above, but look through fmul/fdiv of the negated value:
2395 // Op0 - (-X * Y) --> Op0 + (X * Y)
2396 // Op0 - (Y * -X) --> Op0 + (X * Y)
2397 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2398 Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2399 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2401 // Op0 - (-X / Y) --> Op0 + (X / Y)
2402 // Op0 - (X / -Y) --> Op0 + (X / Y)
2403 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2404 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2405 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2406 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2409 // Handle special cases for FSub with selects feeding the operation
2410 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2411 return replaceInstUsesWith(I, V);
2413 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2414 // (Y - X) - Y --> -X
2415 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2416 return UnaryOperator::CreateFNegFMF(X, &I);
2418 // Y - (X + Y) --> -X
2419 // Y - (Y + X) --> -X
2420 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2421 return UnaryOperator::CreateFNegFMF(X, &I);
2423 // (X * C) - X --> X * (C - 1.0)
2424 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2425 Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
2426 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2428 // X - (X * C) --> X * (1.0 - C)
2429 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2430 Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
2431 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2434 // Reassociate fsub/fadd sequences to create more fadd instructions and
2435 // reduce dependency chains:
2436 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2438 if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2440 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2441 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2442 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2445 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2446 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2449 Value *A0, *A1, *V0, *V1;
2450 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2451 V0->getType() == V1->getType()) {
2452 // Difference of sums is sum of differences:
2453 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2454 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2455 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2456 {Sub->getType()}, {A0, Sub}, &I);
2457 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2460 if (Instruction *F = factorizeFAddFSub(I, Builder))
2463 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2464 // functionality has been subsumed by simple pattern matching here and in
2465 // InstSimplify. We should let a dedicated reassociation pass handle more
2466 // complex pattern matching and remove this from InstCombine.
2467 if (Value *V = FAddCombine(Builder).simplify(&I))
2468 return replaceInstUsesWith(I, V);
2470 // (X - Y) - Op1 --> X - (Y + Op1)
2471 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2472 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2473 return BinaryOperator::CreateFSubFMF(X, FAdd, &I);