1 //===- InstCombineAddSub.cpp ----------------------------------------------===//
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 "InstCombine.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/GetElementPtrTypeIterator.h"
19 #include "llvm/IR/PatternMatch.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
27 /// Class representing coefficient of floating-point addend.
28 /// This class needs to be highly efficient, which is especially true for
29 /// the constructor. As of I write this comment, the cost of the default
30 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
31 /// perform write-merging).
35 // The constructor has to initialize a APFloat, which is unnecessary for
36 // most addends which have coefficient either 1 or -1. So, the constructor
37 // is expensive. In order to avoid the cost of the constructor, we should
38 // reuse some instances whenever possible. The pre-created instances
39 // FAddCombine::Add[0-5] embodies this idea.
41 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {}
45 assert(!insaneIntVal(C) && "Insane coefficient");
46 IsFp = false; IntVal = C;
49 void set(const APFloat& C);
53 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
54 Value *getValue(Type *) const;
56 // If possible, don't define operator+/operator- etc because these
57 // operators inevitably call FAddendCoef's constructor which is not cheap.
58 void operator=(const FAddendCoef &A);
59 void operator+=(const FAddendCoef &A);
60 void operator-=(const FAddendCoef &A);
61 void operator*=(const FAddendCoef &S);
63 bool isOne() const { return isInt() && IntVal == 1; }
64 bool isTwo() const { return isInt() && IntVal == 2; }
65 bool isMinusOne() const { return isInt() && IntVal == -1; }
66 bool isMinusTwo() const { return isInt() && IntVal == -2; }
69 bool insaneIntVal(int V) { return V > 4 || V < -4; }
70 APFloat *getFpValPtr(void)
71 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); }
72 const APFloat *getFpValPtr(void) const
73 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); }
75 const APFloat &getFpVal(void) const {
76 assert(IsFp && BufHasFpVal && "Incorret state");
77 return *getFpValPtr();
80 APFloat &getFpVal(void) {
81 assert(IsFp && BufHasFpVal && "Incorret state");
82 return *getFpValPtr();
85 bool isInt() const { return !IsFp; }
87 // If the coefficient is represented by an integer, promote it to a
89 void convertToFpType(const fltSemantics &Sem);
91 // Construct an APFloat from a signed integer.
92 // TODO: We should get rid of this function when APFloat can be constructed
93 // from an *SIGNED* integer.
94 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
99 // True iff FpValBuf contains an instance of APFloat.
102 // The integer coefficient of an individual addend is either 1 or -1,
103 // and we try to simplify at most 4 addends from neighboring at most
104 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
105 // is overkill of this end.
108 AlignedCharArrayUnion<APFloat> FpValBuf;
111 /// FAddend is used to represent floating-point addend. An addend is
112 /// represented as <C, V>, where the V is a symbolic value, and C is a
113 /// constant coefficient. A constant addend is represented as <C, 0>.
117 FAddend() { Val = nullptr; }
119 Value *getSymVal (void) const { return Val; }
120 const FAddendCoef &getCoef(void) const { return Coeff; }
122 bool isConstant() const { return Val == nullptr; }
123 bool isZero() const { return Coeff.isZero(); }
125 void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; }
126 void set(const APFloat& Coefficient, Value *V)
127 { Coeff.set(Coefficient); Val = V; }
128 void set(const ConstantFP* Coefficient, Value *V)
129 { Coeff.set(Coefficient->getValueAPF()); Val = V; }
131 void negate() { Coeff.negate(); }
133 /// Drill down the U-D chain one step to find the definition of V, and
134 /// try to break the definition into one or two addends.
135 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
137 /// Similar to FAddend::drillDownOneStep() except that the value being
138 /// splitted is the addend itself.
139 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
141 void operator+=(const FAddend &T) {
142 assert((Val == T.Val) && "Symbolic-values disagree");
147 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
149 // This addend has the value of "Coeff * Val".
154 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
155 /// with its neighboring at most two instructions.
159 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(nullptr) {}
160 Value *simplify(Instruction *FAdd);
163 typedef SmallVector<const FAddend*, 4> AddendVect;
165 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
167 Value *performFactorization(Instruction *I);
169 /// Convert given addend to a Value
170 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
172 /// Return the number of instructions needed to emit the N-ary addition.
173 unsigned calcInstrNumber(const AddendVect& Vect);
174 Value *createFSub(Value *Opnd0, Value *Opnd1);
175 Value *createFAdd(Value *Opnd0, Value *Opnd1);
176 Value *createFMul(Value *Opnd0, Value *Opnd1);
177 Value *createFDiv(Value *Opnd0, Value *Opnd1);
178 Value *createFNeg(Value *V);
179 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
180 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
182 InstCombiner::BuilderTy *Builder;
186 // Debugging stuff are clustered here.
188 unsigned CreateInstrNum;
189 void initCreateInstNum() { CreateInstrNum = 0; }
190 void incCreateInstNum() { CreateInstrNum++; }
192 void initCreateInstNum() {}
193 void incCreateInstNum() {}
198 //===----------------------------------------------------------------------===//
201 // {FAddendCoef, FAddend, FAddition, FAddCombine}.
203 //===----------------------------------------------------------------------===//
204 FAddendCoef::~FAddendCoef() {
206 getFpValPtr()->~APFloat();
209 void FAddendCoef::set(const APFloat& C) {
210 APFloat *P = getFpValPtr();
213 // As the buffer is meanless byte stream, we cannot call
214 // APFloat::operator=().
219 IsFp = BufHasFpVal = true;
222 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
226 APFloat *P = getFpValPtr();
228 new(P) APFloat(Sem, IntVal);
230 new(P) APFloat(Sem, 0 - IntVal);
233 IsFp = BufHasFpVal = true;
236 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
238 return APFloat(Sem, Val);
240 APFloat T(Sem, 0 - Val);
246 void FAddendCoef::operator=(const FAddendCoef &That) {
250 set(That.getFpVal());
253 void FAddendCoef::operator+=(const FAddendCoef &That) {
254 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
255 if (isInt() == That.isInt()) {
257 IntVal += That.IntVal;
259 getFpVal().add(That.getFpVal(), RndMode);
264 const APFloat &T = That.getFpVal();
265 convertToFpType(T.getSemantics());
266 getFpVal().add(T, RndMode);
270 APFloat &T = getFpVal();
271 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
274 void FAddendCoef::operator-=(const FAddendCoef &That) {
275 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
276 if (isInt() == That.isInt()) {
278 IntVal -= That.IntVal;
280 getFpVal().subtract(That.getFpVal(), RndMode);
285 const APFloat &T = That.getFpVal();
286 convertToFpType(T.getSemantics());
287 getFpVal().subtract(T, RndMode);
291 APFloat &T = getFpVal();
292 T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode);
295 void FAddendCoef::operator*=(const FAddendCoef &That) {
299 if (That.isMinusOne()) {
304 if (isInt() && That.isInt()) {
305 int Res = IntVal * (int)That.IntVal;
306 assert(!insaneIntVal(Res) && "Insane int value");
311 const fltSemantics &Semantic =
312 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
315 convertToFpType(Semantic);
316 APFloat &F0 = getFpVal();
319 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
320 APFloat::rmNearestTiesToEven);
322 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
327 void FAddendCoef::negate() {
331 getFpVal().changeSign();
334 Value *FAddendCoef::getValue(Type *Ty) const {
336 ConstantFP::get(Ty, float(IntVal)) :
337 ConstantFP::get(Ty->getContext(), getFpVal());
340 // The definition of <Val> Addends
341 // =========================================
342 // A + B <1, A>, <1,B>
343 // A - B <1, A>, <1,B>
346 // A + C <1, A> <C, NULL>
347 // 0 +/- 0 <0, NULL> (corner case)
349 // Legend: A and B are not constant, C is constant
351 unsigned FAddend::drillValueDownOneStep
352 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
353 Instruction *I = nullptr;
354 if (!Val || !(I = dyn_cast<Instruction>(Val)))
357 unsigned Opcode = I->getOpcode();
359 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
361 Value *Opnd0 = I->getOperand(0);
362 Value *Opnd1 = I->getOperand(1);
363 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
366 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
371 Addend0.set(1, Opnd0);
373 Addend0.set(C0, nullptr);
377 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
379 Addend.set(1, Opnd1);
381 Addend.set(C1, nullptr);
382 if (Opcode == Instruction::FSub)
387 return Opnd0 && Opnd1 ? 2 : 1;
389 // Both operands are zero. Weird!
390 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
394 if (I->getOpcode() == Instruction::FMul) {
395 Value *V0 = I->getOperand(0);
396 Value *V1 = I->getOperand(1);
397 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
402 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
411 // Try to break *this* addend into two addends. e.g. Suppose this addend is
412 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
413 // i.e. <2.3, X> and <2.3, Y>.
415 unsigned FAddend::drillAddendDownOneStep
416 (FAddend &Addend0, FAddend &Addend1) const {
420 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
421 if (!BreakNum || Coeff.isOne())
424 Addend0.Scale(Coeff);
427 Addend1.Scale(Coeff);
432 // Try to perform following optimization on the input instruction I. Return the
433 // simplified expression if was successful; otherwise, return 0.
435 // Instruction "I" is Simplified into
436 // -------------------------------------------------------
437 // (x * y) +/- (x * z) x * (y +/- z)
438 // (y / x) +/- (z / x) (y +/- z) / x
440 Value *FAddCombine::performFactorization(Instruction *I) {
441 assert((I->getOpcode() == Instruction::FAdd ||
442 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
444 Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0));
445 Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1));
447 if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
451 if (I0->getOpcode() == Instruction::FMul)
453 else if (I0->getOpcode() != Instruction::FDiv)
456 Value *Opnd0_0 = I0->getOperand(0);
457 Value *Opnd0_1 = I0->getOperand(1);
458 Value *Opnd1_0 = I1->getOperand(0);
459 Value *Opnd1_1 = I1->getOperand(1);
461 // Input Instr I Factor AddSub0 AddSub1
462 // ----------------------------------------------
463 // (x*y) +/- (x*z) x y z
464 // (y/x) +/- (z/x) x y z
466 Value *Factor = nullptr;
467 Value *AddSub0 = nullptr, *AddSub1 = nullptr;
470 if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
472 else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
476 AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
477 AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
479 } else if (Opnd0_1 == Opnd1_1) {
489 Flags.setUnsafeAlgebra();
490 if (I0) Flags &= I->getFastMathFlags();
491 if (I1) Flags &= I->getFastMathFlags();
493 // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
494 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
495 createFAdd(AddSub0, AddSub1) :
496 createFSub(AddSub0, AddSub1);
497 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
498 const APFloat &F = CFP->getValueAPF();
501 } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
502 II->setFastMathFlags(Flags);
505 Value *RI = createFMul(Factor, NewAddSub);
506 if (Instruction *II = dyn_cast<Instruction>(RI))
507 II->setFastMathFlags(Flags);
511 Value *RI = createFDiv(NewAddSub, Factor);
512 if (Instruction *II = dyn_cast<Instruction>(RI))
513 II->setFastMathFlags(Flags);
517 Value *FAddCombine::simplify(Instruction *I) {
518 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode");
520 // Currently we are not able to handle vector type.
521 if (I->getType()->isVectorTy())
524 assert((I->getOpcode() == Instruction::FAdd ||
525 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
527 // Save the instruction before calling other member-functions.
530 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
532 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
534 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
535 unsigned Opnd0_ExpNum = 0;
536 unsigned Opnd1_ExpNum = 0;
538 if (!Opnd0.isConstant())
539 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
541 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
542 if (OpndNum == 2 && !Opnd1.isConstant())
543 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
545 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
546 if (Opnd0_ExpNum && Opnd1_ExpNum) {
548 AllOpnds.push_back(&Opnd0_0);
549 AllOpnds.push_back(&Opnd1_0);
550 if (Opnd0_ExpNum == 2)
551 AllOpnds.push_back(&Opnd0_1);
552 if (Opnd1_ExpNum == 2)
553 AllOpnds.push_back(&Opnd1_1);
555 // Compute instruction quota. We should save at least one instruction.
556 unsigned InstQuota = 0;
558 Value *V0 = I->getOperand(0);
559 Value *V1 = I->getOperand(1);
560 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
561 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
563 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
568 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
569 // splitted into two addends, say "V = X - Y", the instruction would have
570 // been optimized into "I = Y - X" in the previous steps.
572 const FAddendCoef &CE = Opnd0.getCoef();
573 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
576 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
579 AllOpnds.push_back(&Opnd0);
580 AllOpnds.push_back(&Opnd1_0);
581 if (Opnd1_ExpNum == 2)
582 AllOpnds.push_back(&Opnd1_1);
584 if (Value *R = simplifyFAdd(AllOpnds, 1))
588 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
591 AllOpnds.push_back(&Opnd1);
592 AllOpnds.push_back(&Opnd0_0);
593 if (Opnd0_ExpNum == 2)
594 AllOpnds.push_back(&Opnd0_1);
596 if (Value *R = simplifyFAdd(AllOpnds, 1))
600 // step 6: Try factorization as the last resort,
601 return performFactorization(I);
604 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
606 unsigned AddendNum = Addends.size();
607 assert(AddendNum <= 4 && "Too many addends");
609 // For saving intermediate results;
610 unsigned NextTmpIdx = 0;
611 FAddend TmpResult[3];
613 // Points to the constant addend of the resulting simplified expression.
614 // If the resulting expr has constant-addend, this constant-addend is
615 // desirable to reside at the top of the resulting expression tree. Placing
616 // constant close to supper-expr(s) will potentially reveal some optimization
617 // opportunities in super-expr(s).
619 const FAddend *ConstAdd = nullptr;
621 // Simplified addends are placed <SimpVect>.
624 // The outer loop works on one symbolic-value at a time. Suppose the input
625 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
626 // The symbolic-values will be processed in this order: x, y, z.
628 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
630 const FAddend *ThisAddend = Addends[SymIdx];
632 // This addend was processed before.
636 Value *Val = ThisAddend->getSymVal();
637 unsigned StartIdx = SimpVect.size();
638 SimpVect.push_back(ThisAddend);
640 // The inner loop collects addends sharing same symbolic-value, and these
641 // addends will be later on folded into a single addend. Following above
642 // example, if the symbolic value "y" is being processed, the inner loop
643 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
644 // be later on folded into "<b1+b2, y>".
646 for (unsigned SameSymIdx = SymIdx + 1;
647 SameSymIdx < AddendNum; SameSymIdx++) {
648 const FAddend *T = Addends[SameSymIdx];
649 if (T && T->getSymVal() == Val) {
650 // Set null such that next iteration of the outer loop will not process
651 // this addend again.
652 Addends[SameSymIdx] = nullptr;
653 SimpVect.push_back(T);
657 // If multiple addends share same symbolic value, fold them together.
658 if (StartIdx + 1 != SimpVect.size()) {
659 FAddend &R = TmpResult[NextTmpIdx ++];
660 R = *SimpVect[StartIdx];
661 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
664 // Pop all addends being folded and push the resulting folded addend.
665 SimpVect.resize(StartIdx);
668 SimpVect.push_back(&R);
671 // Don't push constant addend at this time. It will be the last element
678 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
679 "out-of-bound access");
682 SimpVect.push_back(ConstAdd);
685 if (!SimpVect.empty())
686 Result = createNaryFAdd(SimpVect, InstrQuota);
688 // The addition is folded to 0.0.
689 Result = ConstantFP::get(Instr->getType(), 0.0);
695 Value *FAddCombine::createNaryFAdd
696 (const AddendVect &Opnds, unsigned InstrQuota) {
697 assert(!Opnds.empty() && "Expect at least one addend");
699 // Step 1: Check if the # of instructions needed exceeds the quota.
701 unsigned InstrNeeded = calcInstrNumber(Opnds);
702 if (InstrNeeded > InstrQuota)
707 // step 2: Emit the N-ary addition.
708 // Note that at most three instructions are involved in Fadd-InstCombine: the
709 // addition in question, and at most two neighboring instructions.
710 // The resulting optimized addition should have at least one less instruction
711 // than the original addition expression tree. This implies that the resulting
712 // N-ary addition has at most two instructions, and we don't need to worry
713 // about tree-height when constructing the N-ary addition.
715 Value *LastVal = nullptr;
716 bool LastValNeedNeg = false;
718 // Iterate the addends, creating fadd/fsub using adjacent two addends.
719 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end();
722 Value *V = createAddendVal(**I, NeedNeg);
725 LastValNeedNeg = NeedNeg;
729 if (LastValNeedNeg == NeedNeg) {
730 LastVal = createFAdd(LastVal, V);
735 LastVal = createFSub(V, LastVal);
737 LastVal = createFSub(LastVal, V);
739 LastValNeedNeg = false;
742 if (LastValNeedNeg) {
743 LastVal = createFNeg(LastVal);
747 assert(CreateInstrNum == InstrNeeded &&
748 "Inconsistent in instruction numbers");
754 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
755 Value *V = Builder->CreateFSub(Opnd0, Opnd1);
756 if (Instruction *I = dyn_cast<Instruction>(V))
757 createInstPostProc(I);
761 Value *FAddCombine::createFNeg(Value *V) {
762 Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType()));
763 Value *NewV = createFSub(Zero, V);
764 if (Instruction *I = dyn_cast<Instruction>(NewV))
765 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
769 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
770 Value *V = Builder->CreateFAdd(Opnd0, Opnd1);
771 if (Instruction *I = dyn_cast<Instruction>(V))
772 createInstPostProc(I);
776 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
777 Value *V = Builder->CreateFMul(Opnd0, Opnd1);
778 if (Instruction *I = dyn_cast<Instruction>(V))
779 createInstPostProc(I);
783 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) {
784 Value *V = Builder->CreateFDiv(Opnd0, Opnd1);
785 if (Instruction *I = dyn_cast<Instruction>(V))
786 createInstPostProc(I);
790 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
791 NewInstr->setDebugLoc(Instr->getDebugLoc());
793 // Keep track of the number of instruction created.
797 // Propagate fast-math flags
798 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
801 // Return the number of instruction needed to emit the N-ary addition.
802 // NOTE: Keep this function in sync with createAddendVal().
803 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
804 unsigned OpndNum = Opnds.size();
805 unsigned InstrNeeded = OpndNum - 1;
807 // The number of addends in the form of "(-1)*x".
808 unsigned NegOpndNum = 0;
810 // Adjust the number of instructions needed to emit the N-ary add.
811 for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end();
813 const FAddend *Opnd = *I;
814 if (Opnd->isConstant())
817 const FAddendCoef &CE = Opnd->getCoef();
818 if (CE.isMinusOne() || CE.isMinusTwo())
821 // Let the addend be "c * x". If "c == +/-1", the value of the addend
822 // is immediately available; otherwise, it needs exactly one instruction
823 // to evaluate the value.
824 if (!CE.isMinusOne() && !CE.isOne())
827 if (NegOpndNum == OpndNum)
832 // Input Addend Value NeedNeg(output)
833 // ================================================================
834 // Constant C C false
835 // <+/-1, V> V coefficient is -1
836 // <2/-2, V> "fadd V, V" coefficient is -2
837 // <C, V> "fmul V, C" false
839 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
840 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
841 const FAddendCoef &Coeff = Opnd.getCoef();
843 if (Opnd.isConstant()) {
845 return Coeff.getValue(Instr->getType());
848 Value *OpndVal = Opnd.getSymVal();
850 if (Coeff.isMinusOne() || Coeff.isOne()) {
851 NeedNeg = Coeff.isMinusOne();
855 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
856 NeedNeg = Coeff.isMinusTwo();
857 return createFAdd(OpndVal, OpndVal);
861 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
864 // If one of the operands only has one non-zero bit, and if the other
865 // operand has a known-zero bit in a more significant place than it (not
866 // including the sign bit) the ripple may go up to and fill the zero, but
867 // won't change the sign. For example, (X & ~4) + 1.
868 static bool checkRippleForAdd(const APInt &Op0KnownZero,
869 const APInt &Op1KnownZero) {
870 APInt Op1MaybeOne = ~Op1KnownZero;
871 // Make sure that one of the operand has at most one bit set to 1.
872 if (Op1MaybeOne.countPopulation() != 1)
875 // Find the most significant known 0 other than the sign bit.
876 int BitWidth = Op0KnownZero.getBitWidth();
877 APInt Op0KnownZeroTemp(Op0KnownZero);
878 Op0KnownZeroTemp.clearBit(BitWidth - 1);
879 int Op0ZeroPosition = BitWidth - Op0KnownZeroTemp.countLeadingZeros() - 1;
881 int Op1OnePosition = BitWidth - Op1MaybeOne.countLeadingZeros() - 1;
882 assert(Op1OnePosition >= 0);
884 // This also covers the case of no known zero, since in that case
885 // Op0ZeroPosition is -1.
886 return Op0ZeroPosition >= Op1OnePosition;
889 /// WillNotOverflowSignedAdd - Return true if we can prove that:
890 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
891 /// This basically requires proving that the add in the original type would not
892 /// overflow to change the sign bit or have a carry out.
893 /// TODO: Handle this for Vectors.
894 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS,
896 // There are different heuristics we can use for this. Here are some simple
899 // If LHS and RHS each have at least two sign bits, the addition will look
905 // If the carry into the most significant position is 0, X and Y can't both
906 // be 1 and therefore the carry out of the addition is also 0.
908 // If the carry into the most significant position is 1, X and Y can't both
909 // be 0 and therefore the carry out of the addition is also 1.
911 // Since the carry into the most significant position is always equal to
912 // the carry out of the addition, there is no signed overflow.
913 if (ComputeNumSignBits(LHS, 0, CxtI) > 1 &&
914 ComputeNumSignBits(RHS, 0, CxtI) > 1)
917 if (IntegerType *IT = dyn_cast<IntegerType>(LHS->getType())) {
918 int BitWidth = IT->getBitWidth();
919 APInt LHSKnownZero(BitWidth, 0);
920 APInt LHSKnownOne(BitWidth, 0);
921 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, CxtI);
923 APInt RHSKnownZero(BitWidth, 0);
924 APInt RHSKnownOne(BitWidth, 0);
925 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, CxtI);
927 // Addition of two 2's compliment numbers having opposite signs will never
929 if ((LHSKnownOne[BitWidth - 1] && RHSKnownZero[BitWidth - 1]) ||
930 (LHSKnownZero[BitWidth - 1] && RHSKnownOne[BitWidth - 1]))
933 // Check if carry bit of addition will not cause overflow.
934 if (checkRippleForAdd(LHSKnownZero, RHSKnownZero))
936 if (checkRippleForAdd(RHSKnownZero, LHSKnownZero))
942 /// WillNotOverflowUnsignedAdd - Return true if we can prove that:
943 /// (zext (add LHS, RHS)) === (add (zext LHS), (zext RHS))
944 bool InstCombiner::WillNotOverflowUnsignedAdd(Value *LHS, Value *RHS,
946 // There are different heuristics we can use for this. Here is a simple one.
947 // If the sign bit of LHS and that of RHS are both zero, no unsigned wrap.
948 bool LHSKnownNonNegative, LHSKnownNegative;
949 bool RHSKnownNonNegative, RHSKnownNegative;
950 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, 0, AT, CxtI, DT);
951 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, 0, AT, CxtI, DT);
952 if (LHSKnownNonNegative && RHSKnownNonNegative)
958 /// \brief Return true if we can prove that:
959 /// (sub LHS, RHS) === (sub nsw LHS, RHS)
960 /// This basically requires proving that the add in the original type would not
961 /// overflow to change the sign bit or have a carry out.
962 /// TODO: Handle this for Vectors.
963 bool InstCombiner::WillNotOverflowSignedSub(Value *LHS, Value *RHS,
965 // If LHS and RHS each have at least two sign bits, the subtraction
967 if (ComputeNumSignBits(LHS, 0, CxtI) > 1 &&
968 ComputeNumSignBits(RHS, 0, CxtI) > 1)
971 if (IntegerType *IT = dyn_cast<IntegerType>(LHS->getType())) {
972 unsigned BitWidth = IT->getBitWidth();
973 APInt LHSKnownZero(BitWidth, 0);
974 APInt LHSKnownOne(BitWidth, 0);
975 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, CxtI);
977 APInt RHSKnownZero(BitWidth, 0);
978 APInt RHSKnownOne(BitWidth, 0);
979 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, CxtI);
981 // Subtraction of two 2's compliment numbers having identical signs will
983 if ((LHSKnownOne[BitWidth - 1] && RHSKnownOne[BitWidth - 1]) ||
984 (LHSKnownZero[BitWidth - 1] && RHSKnownZero[BitWidth - 1]))
987 // TODO: implement logic similar to checkRippleForAdd
992 /// \brief Return true if we can prove that:
993 /// (sub LHS, RHS) === (sub nuw LHS, RHS)
994 bool InstCombiner::WillNotOverflowUnsignedSub(Value *LHS, Value *RHS,
996 // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
997 bool LHSKnownNonNegative, LHSKnownNegative;
998 bool RHSKnownNonNegative, RHSKnownNegative;
999 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, 0, AT, CxtI, DT);
1000 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, 0, AT, CxtI, DT);
1001 if (LHSKnownNegative && RHSKnownNonNegative)
1007 /// \brief Return true if we can prove that:
1008 /// (mul LHS, RHS) === (mul nsw LHS, RHS)
1009 bool InstCombiner::WillNotOverflowSignedMul(Value *LHS, Value *RHS,
1010 Instruction *CxtI) {
1011 if (IntegerType *IT = dyn_cast<IntegerType>(LHS->getType())) {
1013 // Multiplying n * m significant bits yields a result of n + m significant
1014 // bits. If the total number of significant bits does not exceed the
1015 // result bit width (minus 1), there is no overflow.
1016 // This means if we have enough leading sign bits in the operands
1017 // we can guarantee that the result does not overflow.
1018 // Ref: "Hacker's Delight" by Henry Warren
1019 unsigned BitWidth = IT->getBitWidth();
1021 // Note that underestimating the number of sign bits gives a more
1022 // conservative answer.
1023 unsigned SignBits = ComputeNumSignBits(LHS, 0, CxtI) +
1024 ComputeNumSignBits(RHS, 0, CxtI);
1026 // First handle the easy case: if we have enough sign bits there's
1027 // definitely no overflow.
1028 if (SignBits > BitWidth + 1)
1031 // There are two ambiguous cases where there can be no overflow:
1032 // SignBits == BitWidth + 1 and
1033 // SignBits == BitWidth
1034 // The second case is difficult to check, therefore we only handle the
1036 if (SignBits == BitWidth + 1) {
1037 // It overflows only when both arguments are negative and the true
1038 // product is exactly the minimum negative number.
1039 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
1040 // For simplicity we just check if at least one side is not negative.
1041 bool LHSNonNegative, LHSNegative;
1042 bool RHSNonNegative, RHSNegative;
1043 ComputeSignBit(LHS, LHSNonNegative, LHSNegative, DL, 0, AT, CxtI, DT);
1044 ComputeSignBit(RHS, RHSNonNegative, RHSNegative, DL, 0, AT, CxtI, DT);
1045 if (LHSNonNegative || RHSNonNegative)
1052 // Checks if any operand is negative and we can convert add to sub.
1053 // This function checks for following negative patterns
1054 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
1055 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
1056 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
1057 static Value *checkForNegativeOperand(BinaryOperator &I,
1058 InstCombiner::BuilderTy *Builder) {
1059 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1061 // This function creates 2 instructions to replace ADD, we need at least one
1062 // of LHS or RHS to have one use to ensure benefit in transform.
1063 if (!LHS->hasOneUse() && !RHS->hasOneUse())
1066 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1067 const APInt *C1 = nullptr, *C2 = nullptr;
1069 // if ONE is on other side, swap
1070 if (match(RHS, m_Add(m_Value(X), m_One())))
1071 std::swap(LHS, RHS);
1073 if (match(LHS, m_Add(m_Value(X), m_One()))) {
1074 // if XOR on other side, swap
1075 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
1078 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
1079 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
1080 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
1081 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
1082 Value *NewAnd = Builder->CreateAnd(Z, *C1);
1083 return Builder->CreateSub(RHS, NewAnd, "sub");
1084 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
1085 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
1086 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
1087 Value *NewOr = Builder->CreateOr(Z, ~(*C1));
1088 return Builder->CreateSub(RHS, NewOr, "sub");
1093 // Restore LHS and RHS
1094 LHS = I.getOperand(0);
1095 RHS = I.getOperand(1);
1097 // if XOR is on other side, swap
1098 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
1099 std::swap(LHS, RHS);
1102 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
1103 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
1104 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
1105 if (C1->countTrailingZeros() == 0)
1106 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
1107 Value *NewOr = Builder->CreateOr(Z, ~(*C2));
1108 return Builder->CreateSub(RHS, NewOr, "sub");
1113 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1114 bool Changed = SimplifyAssociativeOrCommutative(I);
1115 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1117 if (Value *V = SimplifyVectorOp(I))
1118 return ReplaceInstUsesWith(I, V);
1120 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
1121 I.hasNoUnsignedWrap(), DL, TLI, DT, AT))
1122 return ReplaceInstUsesWith(I, V);
1124 // (A*B)+(A*C) -> A*(B+C) etc
1125 if (Value *V = SimplifyUsingDistributiveLaws(I))
1126 return ReplaceInstUsesWith(I, V);
1128 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1129 // X + (signbit) --> X ^ signbit
1130 const APInt &Val = CI->getValue();
1131 if (Val.isSignBit())
1132 return BinaryOperator::CreateXor(LHS, RHS);
1134 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1135 // (X & 254)+1 -> (X&254)|1
1136 if (SimplifyDemandedInstructionBits(I))
1139 // zext(bool) + C -> bool ? C + 1 : C
1140 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
1141 if (ZI->getSrcTy()->isIntegerTy(1))
1142 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
1144 Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1145 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1146 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
1147 const APInt &RHSVal = CI->getValue();
1148 unsigned ExtendAmt = 0;
1149 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1150 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1151 if (XorRHS->getValue() == -RHSVal) {
1152 if (RHSVal.isPowerOf2())
1153 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1154 else if (XorRHS->getValue().isPowerOf2())
1155 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1159 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1160 if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1165 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt);
1166 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext");
1167 return BinaryOperator::CreateAShr(NewShl, ShAmt);
1170 // If this is a xor that was canonicalized from a sub, turn it back into
1171 // a sub and fuse this add with it.
1172 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1173 IntegerType *IT = cast<IntegerType>(I.getType());
1174 APInt LHSKnownOne(IT->getBitWidth(), 0);
1175 APInt LHSKnownZero(IT->getBitWidth(), 0);
1176 computeKnownBits(XorLHS, LHSKnownZero, LHSKnownOne, 0, &I);
1177 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue())
1178 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1181 // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C,
1182 // transform them into (X + (signbit ^ C))
1183 if (XorRHS->getValue().isSignBit())
1184 return BinaryOperator::CreateAdd(XorLHS,
1185 ConstantExpr::getXor(XorRHS, CI));
1189 if (isa<Constant>(RHS) && isa<PHINode>(LHS))
1190 if (Instruction *NV = FoldOpIntoPhi(I))
1193 if (I.getType()->getScalarType()->isIntegerTy(1))
1194 return BinaryOperator::CreateXor(LHS, RHS);
1198 BinaryOperator *New =
1199 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1));
1200 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1201 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1206 // -A + -B --> -(A + B)
1207 if (Value *LHSV = dyn_castNegVal(LHS)) {
1208 if (!isa<Constant>(RHS))
1209 if (Value *RHSV = dyn_castNegVal(RHS)) {
1210 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
1211 return BinaryOperator::CreateNeg(NewAdd);
1214 return BinaryOperator::CreateSub(RHS, LHSV);
1218 if (!isa<Constant>(RHS))
1219 if (Value *V = dyn_castNegVal(RHS))
1220 return BinaryOperator::CreateSub(LHS, V);
1222 if (Value *V = checkForNegativeOperand(I, Builder))
1223 return ReplaceInstUsesWith(I, V);
1225 // A+B --> A|B iff A and B have no bits set in common.
1226 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
1227 APInt LHSKnownOne(IT->getBitWidth(), 0);
1228 APInt LHSKnownZero(IT->getBitWidth(), 0);
1229 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, &I);
1230 if (LHSKnownZero != 0) {
1231 APInt RHSKnownOne(IT->getBitWidth(), 0);
1232 APInt RHSKnownZero(IT->getBitWidth(), 0);
1233 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, &I);
1235 // No bits in common -> bitwise or.
1236 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
1237 return BinaryOperator::CreateOr(LHS, RHS);
1241 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
1243 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
1244 return BinaryOperator::CreateSub(SubOne(CRHS), X);
1247 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1248 // (X & FF00) + xx00 -> (X+xx00) & FF00
1251 if (LHS->hasOneUse() &&
1252 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1253 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1254 // See if all bits from the first bit set in the Add RHS up are included
1255 // in the mask. First, get the rightmost bit.
1256 const APInt &AddRHSV = CRHS->getValue();
1258 // Form a mask of all bits from the lowest bit added through the top.
1259 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1261 // See if the and mask includes all of these bits.
1262 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1264 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1265 // Okay, the xform is safe. Insert the new add pronto.
1266 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
1267 return BinaryOperator::CreateAnd(NewAdd, C2);
1271 // Try to fold constant add into select arguments.
1272 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1273 if (Instruction *R = FoldOpIntoSelect(I, SI))
1277 // add (select X 0 (sub n A)) A --> select X A n
1279 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1282 SI = dyn_cast<SelectInst>(RHS);
1285 if (SI && SI->hasOneUse()) {
1286 Value *TV = SI->getTrueValue();
1287 Value *FV = SI->getFalseValue();
1290 // Can we fold the add into the argument of the select?
1291 // We check both true and false select arguments for a matching subtract.
1292 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1293 // Fold the add into the true select value.
1294 return SelectInst::Create(SI->getCondition(), N, A);
1296 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1297 // Fold the add into the false select value.
1298 return SelectInst::Create(SI->getCondition(), A, N);
1302 // Check for (add (sext x), y), see if we can merge this into an
1303 // integer add followed by a sext.
1304 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
1305 // (add (sext x), cst) --> (sext (add x, cst'))
1306 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1308 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1309 if (LHSConv->hasOneUse() &&
1310 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
1311 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI, &I)) {
1312 // Insert the new, smaller add.
1313 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1315 return new SExtInst(NewAdd, I.getType());
1319 // (add (sext x), (sext y)) --> (sext (add int x, y))
1320 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
1321 // Only do this if x/y have the same type, if at last one of them has a
1322 // single use (so we don't increase the number of sexts), and if the
1323 // integer add will not overflow.
1324 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
1325 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1326 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1327 RHSConv->getOperand(0), &I)) {
1328 // Insert the new integer add.
1329 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1330 RHSConv->getOperand(0), "addconv");
1331 return new SExtInst(NewAdd, I.getType());
1336 // (add (xor A, B) (and A, B)) --> (or A, B)
1338 Value *A = nullptr, *B = nullptr;
1339 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) &&
1340 (match(LHS, m_And(m_Specific(A), m_Specific(B))) ||
1341 match(LHS, m_And(m_Specific(B), m_Specific(A)))))
1342 return BinaryOperator::CreateOr(A, B);
1344 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
1345 (match(RHS, m_And(m_Specific(A), m_Specific(B))) ||
1346 match(RHS, m_And(m_Specific(B), m_Specific(A)))))
1347 return BinaryOperator::CreateOr(A, B);
1350 // (add (or A, B) (and A, B)) --> (add A, B)
1352 Value *A = nullptr, *B = nullptr;
1353 if (match(RHS, m_Or(m_Value(A), m_Value(B))) &&
1354 (match(LHS, m_And(m_Specific(A), m_Specific(B))) ||
1355 match(LHS, m_And(m_Specific(B), m_Specific(A))))) {
1356 auto *New = BinaryOperator::CreateAdd(A, B);
1357 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1358 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1362 if (match(LHS, m_Or(m_Value(A), m_Value(B))) &&
1363 (match(RHS, m_And(m_Specific(A), m_Specific(B))) ||
1364 match(RHS, m_And(m_Specific(B), m_Specific(A))))) {
1365 auto *New = BinaryOperator::CreateAdd(A, B);
1366 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1367 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1372 // TODO(jingyue): Consider WillNotOverflowSignedAdd and
1373 // WillNotOverflowUnsignedAdd to reduce the number of invocations of
1374 // computeKnownBits.
1375 if (!I.hasNoSignedWrap() && WillNotOverflowSignedAdd(LHS, RHS, &I)) {
1377 I.setHasNoSignedWrap(true);
1379 if (!I.hasNoUnsignedWrap() && WillNotOverflowUnsignedAdd(LHS, RHS, &I)) {
1381 I.setHasNoUnsignedWrap(true);
1384 return Changed ? &I : nullptr;
1387 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
1388 bool Changed = SimplifyAssociativeOrCommutative(I);
1389 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1391 if (Value *V = SimplifyVectorOp(I))
1392 return ReplaceInstUsesWith(I, V);
1394 if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL,
1396 return ReplaceInstUsesWith(I, V);
1398 if (isa<Constant>(RHS)) {
1399 if (isa<PHINode>(LHS))
1400 if (Instruction *NV = FoldOpIntoPhi(I))
1403 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1404 if (Instruction *NV = FoldOpIntoSelect(I, SI))
1409 // -A + -B --> -(A + B)
1410 if (Value *LHSV = dyn_castFNegVal(LHS)) {
1411 Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV);
1412 RI->copyFastMathFlags(&I);
1417 if (!isa<Constant>(RHS))
1418 if (Value *V = dyn_castFNegVal(RHS)) {
1419 Instruction *RI = BinaryOperator::CreateFSub(LHS, V);
1420 RI->copyFastMathFlags(&I);
1424 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1425 // integer add followed by a promotion.
1426 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1427 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1428 // ... if the constant fits in the integer value. This is useful for things
1429 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1430 // requires a constant pool load, and generally allows the add to be better
1432 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
1434 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
1435 if (LHSConv->hasOneUse() &&
1436 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1437 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI, &I)) {
1438 // Insert the new integer add.
1439 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1441 return new SIToFPInst(NewAdd, I.getType());
1445 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1446 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1447 // Only do this if x/y have the same type, if at last one of them has a
1448 // single use (so we don't increase the number of int->fp conversions),
1449 // and if the integer add will not overflow.
1450 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
1451 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1452 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1453 RHSConv->getOperand(0), &I)) {
1454 // Insert the new integer add.
1455 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1456 RHSConv->getOperand(0),"addconv");
1457 return new SIToFPInst(NewAdd, I.getType());
1462 // select C, 0, B + select C, A, 0 -> select C, A, B
1464 Value *A1, *B1, *C1, *A2, *B2, *C2;
1465 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) &&
1466 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) {
1468 Constant *Z1=nullptr, *Z2=nullptr;
1469 Value *A, *B, *C=C1;
1470 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) {
1471 Z1 = dyn_cast<Constant>(A1); A = A2;
1472 Z2 = dyn_cast<Constant>(B2); B = B1;
1473 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) {
1474 Z1 = dyn_cast<Constant>(B1); B = B2;
1475 Z2 = dyn_cast<Constant>(A2); A = A1;
1479 (I.hasNoSignedZeros() ||
1480 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) {
1481 return SelectInst::Create(C, A, B);
1487 if (I.hasUnsafeAlgebra()) {
1488 if (Value *V = FAddCombine(Builder).simplify(&I))
1489 return ReplaceInstUsesWith(I, V);
1492 return Changed ? &I : nullptr;
1496 /// Optimize pointer differences into the same array into a size. Consider:
1497 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1498 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1500 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1502 assert(DL && "Must have target data info for this");
1504 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1506 bool Swapped = false;
1507 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1509 // For now we require one side to be the base pointer "A" or a constant
1510 // GEP derived from it.
1511 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1513 if (LHSGEP->getOperand(0) == RHS) {
1516 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1517 // (gep X, ...) - (gep X, ...)
1518 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1519 RHSGEP->getOperand(0)->stripPointerCasts()) {
1527 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1529 if (RHSGEP->getOperand(0) == LHS) {
1532 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1533 // (gep X, ...) - (gep X, ...)
1534 if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1535 LHSGEP->getOperand(0)->stripPointerCasts()) {
1543 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and
1546 (GEP2 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse()))
1549 // Emit the offset of the GEP and an intptr_t.
1550 Value *Result = EmitGEPOffset(GEP1);
1552 // If we had a constant expression GEP on the other side offsetting the
1553 // pointer, subtract it from the offset we have.
1555 Value *Offset = EmitGEPOffset(GEP2);
1556 Result = Builder->CreateSub(Result, Offset);
1559 // If we have p - gep(p, ...) then we have to negate the result.
1561 Result = Builder->CreateNeg(Result, "diff.neg");
1563 return Builder->CreateIntCast(Result, Ty, true);
1566 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1567 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1569 if (Value *V = SimplifyVectorOp(I))
1570 return ReplaceInstUsesWith(I, V);
1572 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(),
1573 I.hasNoUnsignedWrap(), DL, TLI, DT, AT))
1574 return ReplaceInstUsesWith(I, V);
1576 // (A*B)-(A*C) -> A*(B-C) etc
1577 if (Value *V = SimplifyUsingDistributiveLaws(I))
1578 return ReplaceInstUsesWith(I, V);
1580 // If this is a 'B = x-(-A)', change to B = x+A.
1581 if (Value *V = dyn_castNegVal(Op1)) {
1582 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1584 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1585 assert(BO->getOpcode() == Instruction::Sub &&
1586 "Expected a subtraction operator!");
1587 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1588 Res->setHasNoSignedWrap(true);
1590 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1591 Res->setHasNoSignedWrap(true);
1597 if (I.getType()->isIntegerTy(1))
1598 return BinaryOperator::CreateXor(Op0, Op1);
1600 // Replace (-1 - A) with (~A).
1601 if (match(Op0, m_AllOnes()))
1602 return BinaryOperator::CreateNot(Op1);
1604 if (Constant *C = dyn_cast<Constant>(Op0)) {
1605 // C - ~X == X + (1+C)
1607 if (match(Op1, m_Not(m_Value(X))))
1608 return BinaryOperator::CreateAdd(X, AddOne(C));
1610 // Try to fold constant sub into select arguments.
1611 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1612 if (Instruction *R = FoldOpIntoSelect(I, SI))
1615 // C-(X+C2) --> (C-C2)-X
1617 if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1618 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1620 if (SimplifyDemandedInstructionBits(I))
1623 // Fold (sub 0, (zext bool to B)) --> (sext bool to B)
1624 if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X))))
1625 if (X->getType()->getScalarType()->isIntegerTy(1))
1626 return CastInst::CreateSExtOrBitCast(X, Op1->getType());
1628 // Fold (sub 0, (sext bool to B)) --> (zext bool to B)
1629 if (C->isNullValue() && match(Op1, m_SExt(m_Value(X))))
1630 if (X->getType()->getScalarType()->isIntegerTy(1))
1631 return CastInst::CreateZExtOrBitCast(X, Op1->getType());
1634 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1635 // -(X >>u 31) -> (X >>s 31)
1636 // -(X >>s 31) -> (X >>u 31)
1640 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) &&
1641 // Verify we are shifting out everything but the sign bit.
1642 CI->getValue() == I.getType()->getPrimitiveSizeInBits() - 1)
1643 return BinaryOperator::CreateAShr(X, CI);
1645 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) &&
1646 // Verify we are shifting out everything but the sign bit.
1647 CI->getValue() == I.getType()->getPrimitiveSizeInBits() - 1)
1648 return BinaryOperator::CreateLShr(X, CI);
1655 // X-(X+Y) == -Y X-(Y+X) == -Y
1656 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) ||
1657 match(Op1, m_Add(m_Value(Y), m_Specific(Op0))))
1658 return BinaryOperator::CreateNeg(Y);
1661 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1662 return BinaryOperator::CreateNeg(Y);
1665 // (sub (or A, B) (xor A, B)) --> (and A, B)
1667 Value *A = nullptr, *B = nullptr;
1668 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1669 (match(Op0, m_Or(m_Specific(A), m_Specific(B))) ||
1670 match(Op0, m_Or(m_Specific(B), m_Specific(A)))))
1671 return BinaryOperator::CreateAnd(A, B);
1674 if (Op0->hasOneUse()) {
1676 // ((X | Y) - X) --> (~X & Y)
1677 if (match(Op0, m_Or(m_Value(Y), m_Specific(Op1))) ||
1678 match(Op0, m_Or(m_Specific(Op1), m_Value(Y))))
1679 return BinaryOperator::CreateAnd(
1680 Y, Builder->CreateNot(Op1, Op1->getName() + ".not"));
1683 if (Op1->hasOneUse()) {
1684 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1685 Constant *C = nullptr;
1686 Constant *CI = nullptr;
1688 // (X - (Y - Z)) --> (X + (Z - Y)).
1689 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1690 return BinaryOperator::CreateAdd(Op0,
1691 Builder->CreateSub(Z, Y, Op1->getName()));
1693 // (X - (X & Y)) --> (X & ~Y)
1695 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) ||
1696 match(Op1, m_And(m_Specific(Op0), m_Value(Y))))
1697 return BinaryOperator::CreateAnd(Op0,
1698 Builder->CreateNot(Y, Y->getName() + ".not"));
1700 // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow.
1701 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1702 C->isNotMinSignedValue() && !C->isOneValue())
1703 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1705 // 0 - (X << Y) -> (-X << Y) when X is freely negatable.
1706 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1707 if (Value *XNeg = dyn_castNegVal(X))
1708 return BinaryOperator::CreateShl(XNeg, Y);
1710 // X - A*-B -> X + A*B
1711 // X - -A*B -> X + A*B
1713 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) ||
1714 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B))))
1715 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B));
1717 // X - A*CI -> X + A*-CI
1718 // X - CI*A -> X + A*-CI
1719 if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) ||
1720 match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) {
1721 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI));
1722 return BinaryOperator::CreateAdd(Op0, NewMul);
1726 // Optimize pointer differences into the same array into a size. Consider:
1727 // &A[10] - &A[0]: we should compile this to "10".
1729 Value *LHSOp, *RHSOp;
1730 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1731 match(Op1, m_PtrToInt(m_Value(RHSOp))))
1732 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1733 return ReplaceInstUsesWith(I, Res);
1735 // trunc(p)-trunc(q) -> trunc(p-q)
1736 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1737 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1738 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1739 return ReplaceInstUsesWith(I, Res);
1742 bool Changed = false;
1743 if (!I.hasNoSignedWrap() && WillNotOverflowSignedSub(Op0, Op1, &I)) {
1745 I.setHasNoSignedWrap(true);
1747 if (!I.hasNoUnsignedWrap() && WillNotOverflowUnsignedSub(Op0, Op1, &I)) {
1749 I.setHasNoUnsignedWrap(true);
1752 return Changed ? &I : nullptr;
1755 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
1756 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1758 if (Value *V = SimplifyVectorOp(I))
1759 return ReplaceInstUsesWith(I, V);
1761 if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL,
1763 return ReplaceInstUsesWith(I, V);
1765 if (isa<Constant>(Op0))
1766 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1767 if (Instruction *NV = FoldOpIntoSelect(I, SI))
1770 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking
1771 // through FP extensions/truncations along the way.
1772 if (Value *V = dyn_castFNegVal(Op1)) {
1773 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V);
1774 NewI->copyFastMathFlags(&I);
1777 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) {
1778 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) {
1779 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType());
1780 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc);
1781 NewI->copyFastMathFlags(&I);
1784 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) {
1785 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
1786 Value *NewExt = Builder->CreateFPExt(V, I.getType());
1787 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt);
1788 NewI->copyFastMathFlags(&I);
1793 if (I.hasUnsafeAlgebra()) {
1794 if (Value *V = FAddCombine(Builder).simplify(&I))
1795 return ReplaceInstUsesWith(I, V);