1 //===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for add, fadd, sub, and fsub.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/GetElementPtrTypeIterator.h"
19 #include "llvm/IR/PatternMatch.h"
22 using namespace PatternMatch;
24 #define DEBUG_TYPE "instcombine"
28 /// Class representing coefficient of floating-point addend.
29 /// This class needs to be highly efficient, which is especially true for
30 /// the constructor. As of I write this comment, the cost of the default
31 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
32 /// perform write-merging).
36 // The constructor has to initialize a APFloat, which is unnecessary for
37 // most addends which have coefficient either 1 or -1. So, the constructor
38 // is expensive. In order to avoid the cost of the constructor, we should
39 // reuse some instances whenever possible. The pre-created instances
40 // FAddCombine::Add[0-5] embodies this idea.
42 FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {}
46 assert(!insaneIntVal(C) && "Insane coefficient");
47 IsFp = false; IntVal = C;
50 void set(const APFloat& C);
54 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
55 Value *getValue(Type *) const;
57 // If possible, don't define operator+/operator- etc because these
58 // operators inevitably call FAddendCoef's constructor which is not cheap.
59 void operator=(const FAddendCoef &A);
60 void operator+=(const FAddendCoef &A);
61 void operator-=(const FAddendCoef &A);
62 void operator*=(const FAddendCoef &S);
64 bool isOne() const { return isInt() && IntVal == 1; }
65 bool isTwo() const { return isInt() && IntVal == 2; }
66 bool isMinusOne() const { return isInt() && IntVal == -1; }
67 bool isMinusTwo() const { return isInt() && IntVal == -2; }
70 bool insaneIntVal(int V) { return V > 4 || V < -4; }
71 APFloat *getFpValPtr()
72 { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); }
73 const APFloat *getFpValPtr() const
74 { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); }
76 const APFloat &getFpVal() const {
77 assert(IsFp && BufHasFpVal && "Incorret state");
78 return *getFpValPtr();
82 assert(IsFp && BufHasFpVal && "Incorret state");
83 return *getFpValPtr();
86 bool isInt() const { return !IsFp; }
88 // If the coefficient is represented by an integer, promote it to a
90 void convertToFpType(const fltSemantics &Sem);
92 // Construct an APFloat from a signed integer.
93 // TODO: We should get rid of this function when APFloat can be constructed
94 // from an *SIGNED* integer.
95 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
100 // True iff FpValBuf contains an instance of APFloat.
103 // The integer coefficient of an individual addend is either 1 or -1,
104 // and we try to simplify at most 4 addends from neighboring at most
105 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
106 // is overkill of this end.
109 AlignedCharArrayUnion<APFloat> FpValBuf;
112 /// FAddend is used to represent floating-point addend. An addend is
113 /// represented as <C, V>, where the V is a symbolic value, and C is a
114 /// constant coefficient. A constant addend is represented as <C, 0>.
118 FAddend() : Val(nullptr) {}
120 Value *getSymVal() const { return Val; }
121 const FAddendCoef &getCoef() const { return Coeff; }
123 bool isConstant() const { return Val == nullptr; }
124 bool isZero() const { return Coeff.isZero(); }
126 void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; }
127 void set(const APFloat& Coefficient, Value *V)
128 { Coeff.set(Coefficient); Val = V; }
129 void set(const ConstantFP* Coefficient, Value *V)
130 { Coeff.set(Coefficient->getValueAPF()); Val = V; }
132 void negate() { Coeff.negate(); }
134 /// Drill down the U-D chain one step to find the definition of V, and
135 /// try to break the definition into one or two addends.
136 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
138 /// Similar to FAddend::drillDownOneStep() except that the value being
139 /// splitted is the addend itself.
140 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
142 void operator+=(const FAddend &T) {
143 assert((Val == T.Val) && "Symbolic-values disagree");
148 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
150 // This addend has the value of "Coeff * Val".
155 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
156 /// with its neighboring at most two instructions.
160 FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(nullptr) {}
161 Value *simplify(Instruction *FAdd);
164 typedef SmallVector<const FAddend*, 4> AddendVect;
166 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
168 Value *performFactorization(Instruction *I);
170 /// Convert given addend to a Value
171 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
173 /// Return the number of instructions needed to emit the N-ary addition.
174 unsigned calcInstrNumber(const AddendVect& Vect);
175 Value *createFSub(Value *Opnd0, Value *Opnd1);
176 Value *createFAdd(Value *Opnd0, Value *Opnd1);
177 Value *createFMul(Value *Opnd0, Value *Opnd1);
178 Value *createFDiv(Value *Opnd0, Value *Opnd1);
179 Value *createFNeg(Value *V);
180 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
181 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
183 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() {}
197 } // anonymous namespace
199 //===----------------------------------------------------------------------===//
202 // {FAddendCoef, FAddend, FAddition, FAddCombine}.
204 //===----------------------------------------------------------------------===//
205 FAddendCoef::~FAddendCoef() {
207 getFpValPtr()->~APFloat();
210 void FAddendCoef::set(const APFloat& C) {
211 APFloat *P = getFpValPtr();
214 // As the buffer is meanless byte stream, we cannot call
215 // APFloat::operator=().
220 IsFp = BufHasFpVal = true;
223 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
227 APFloat *P = getFpValPtr();
229 new(P) APFloat(Sem, IntVal);
231 new(P) APFloat(Sem, 0 - IntVal);
234 IsFp = BufHasFpVal = true;
237 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
239 return APFloat(Sem, Val);
241 APFloat T(Sem, 0 - Val);
247 void FAddendCoef::operator=(const FAddendCoef &That) {
251 set(That.getFpVal());
254 void FAddendCoef::operator+=(const FAddendCoef &That) {
255 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
256 if (isInt() == That.isInt()) {
258 IntVal += That.IntVal;
260 getFpVal().add(That.getFpVal(), RndMode);
265 const APFloat &T = That.getFpVal();
266 convertToFpType(T.getSemantics());
267 getFpVal().add(T, RndMode);
271 APFloat &T = getFpVal();
272 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
275 void FAddendCoef::operator-=(const FAddendCoef &That) {
276 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
277 if (isInt() == That.isInt()) {
279 IntVal -= That.IntVal;
281 getFpVal().subtract(That.getFpVal(), RndMode);
286 const APFloat &T = That.getFpVal();
287 convertToFpType(T.getSemantics());
288 getFpVal().subtract(T, RndMode);
292 APFloat &T = getFpVal();
293 T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode);
296 void FAddendCoef::operator*=(const FAddendCoef &That) {
300 if (That.isMinusOne()) {
305 if (isInt() && That.isInt()) {
306 int Res = IntVal * (int)That.IntVal;
307 assert(!insaneIntVal(Res) && "Insane int value");
312 const fltSemantics &Semantic =
313 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
316 convertToFpType(Semantic);
317 APFloat &F0 = getFpVal();
320 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
321 APFloat::rmNearestTiesToEven);
323 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
328 void FAddendCoef::negate() {
332 getFpVal().changeSign();
335 Value *FAddendCoef::getValue(Type *Ty) const {
337 ConstantFP::get(Ty, float(IntVal)) :
338 ConstantFP::get(Ty->getContext(), getFpVal());
341 // The definition of <Val> Addends
342 // =========================================
343 // A + B <1, A>, <1,B>
344 // A - B <1, A>, <1,B>
347 // A + C <1, A> <C, NULL>
348 // 0 +/- 0 <0, NULL> (corner case)
350 // Legend: A and B are not constant, C is constant
352 unsigned FAddend::drillValueDownOneStep
353 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
354 Instruction *I = nullptr;
355 if (!Val || !(I = dyn_cast<Instruction>(Val)))
358 unsigned Opcode = I->getOpcode();
360 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
362 Value *Opnd0 = I->getOperand(0);
363 Value *Opnd1 = I->getOperand(1);
364 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
367 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
372 Addend0.set(1, Opnd0);
374 Addend0.set(C0, nullptr);
378 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
380 Addend.set(1, Opnd1);
382 Addend.set(C1, nullptr);
383 if (Opcode == Instruction::FSub)
388 return Opnd0 && Opnd1 ? 2 : 1;
390 // Both operands are zero. Weird!
391 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
395 if (I->getOpcode() == Instruction::FMul) {
396 Value *V0 = I->getOperand(0);
397 Value *V1 = I->getOperand(1);
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
403 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
412 // Try to break *this* addend into two addends. e.g. Suppose this addend is
413 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
414 // i.e. <2.3, X> and <2.3, Y>.
416 unsigned FAddend::drillAddendDownOneStep
417 (FAddend &Addend0, FAddend &Addend1) const {
421 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
422 if (!BreakNum || Coeff.isOne())
425 Addend0.Scale(Coeff);
428 Addend1.Scale(Coeff);
433 // Try to perform following optimization on the input instruction I. Return the
434 // simplified expression if was successful; otherwise, return 0.
436 // Instruction "I" is Simplified into
437 // -------------------------------------------------------
438 // (x * y) +/- (x * z) x * (y +/- z)
439 // (y / x) +/- (z / x) (y +/- z) / x
441 Value *FAddCombine::performFactorization(Instruction *I) {
442 assert((I->getOpcode() == Instruction::FAdd ||
443 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
445 Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0));
446 Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1));
448 if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
452 if (I0->getOpcode() == Instruction::FMul)
454 else if (I0->getOpcode() != Instruction::FDiv)
457 Value *Opnd0_0 = I0->getOperand(0);
458 Value *Opnd0_1 = I0->getOperand(1);
459 Value *Opnd1_0 = I1->getOperand(0);
460 Value *Opnd1_1 = I1->getOperand(1);
462 // Input Instr I Factor AddSub0 AddSub1
463 // ----------------------------------------------
464 // (x*y) +/- (x*z) x y z
465 // (y/x) +/- (z/x) x y z
467 Value *Factor = nullptr;
468 Value *AddSub0 = nullptr, *AddSub1 = nullptr;
471 if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
473 else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
477 AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
478 AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
480 } else if (Opnd0_1 == Opnd1_1) {
490 Flags.setUnsafeAlgebra();
491 if (I0) Flags &= I->getFastMathFlags();
492 if (I1) Flags &= I->getFastMathFlags();
494 // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
495 Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
496 createFAdd(AddSub0, AddSub1) :
497 createFSub(AddSub0, AddSub1);
498 if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
499 const APFloat &F = CFP->getValueAPF();
502 } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
503 II->setFastMathFlags(Flags);
506 Value *RI = createFMul(Factor, NewAddSub);
507 if (Instruction *II = dyn_cast<Instruction>(RI))
508 II->setFastMathFlags(Flags);
512 Value *RI = createFDiv(NewAddSub, Factor);
513 if (Instruction *II = dyn_cast<Instruction>(RI))
514 II->setFastMathFlags(Flags);
518 Value *FAddCombine::simplify(Instruction *I) {
519 assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode");
521 // Currently we are not able to handle vector type.
522 if (I->getType()->isVectorTy())
525 assert((I->getOpcode() == Instruction::FAdd ||
526 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
528 // Save the instruction before calling other member-functions.
531 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
533 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
535 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
536 unsigned Opnd0_ExpNum = 0;
537 unsigned Opnd1_ExpNum = 0;
539 if (!Opnd0.isConstant())
540 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
542 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
543 if (OpndNum == 2 && !Opnd1.isConstant())
544 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
546 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
547 if (Opnd0_ExpNum && Opnd1_ExpNum) {
549 AllOpnds.push_back(&Opnd0_0);
550 AllOpnds.push_back(&Opnd1_0);
551 if (Opnd0_ExpNum == 2)
552 AllOpnds.push_back(&Opnd0_1);
553 if (Opnd1_ExpNum == 2)
554 AllOpnds.push_back(&Opnd1_1);
556 // Compute instruction quota. We should save at least one instruction.
557 unsigned InstQuota = 0;
559 Value *V0 = I->getOperand(0);
560 Value *V1 = I->getOperand(1);
561 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
562 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
564 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
569 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
570 // splitted into two addends, say "V = X - Y", the instruction would have
571 // been optimized into "I = Y - X" in the previous steps.
573 const FAddendCoef &CE = Opnd0.getCoef();
574 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
577 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
580 AllOpnds.push_back(&Opnd0);
581 AllOpnds.push_back(&Opnd1_0);
582 if (Opnd1_ExpNum == 2)
583 AllOpnds.push_back(&Opnd1_1);
585 if (Value *R = simplifyFAdd(AllOpnds, 1))
589 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
592 AllOpnds.push_back(&Opnd1);
593 AllOpnds.push_back(&Opnd0_0);
594 if (Opnd0_ExpNum == 2)
595 AllOpnds.push_back(&Opnd0_1);
597 if (Value *R = simplifyFAdd(AllOpnds, 1))
601 // step 6: Try factorization as the last resort,
602 return performFactorization(I);
605 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 /// 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 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS,
895 // There are different heuristics we can use for this. Here are some simple
898 // If LHS and RHS each have at least two sign bits, the addition will look
904 // If the carry into the most significant position is 0, X and Y can't both
905 // be 1 and therefore the carry out of the addition is also 0.
907 // If the carry into the most significant position is 1, X and Y can't both
908 // be 0 and therefore the carry out of the addition is also 1.
910 // Since the carry into the most significant position is always equal to
911 // the carry out of the addition, there is no signed overflow.
912 if (ComputeNumSignBits(LHS, 0, &CxtI) > 1 &&
913 ComputeNumSignBits(RHS, 0, &CxtI) > 1)
916 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
917 APInt LHSKnownZero(BitWidth, 0);
918 APInt LHSKnownOne(BitWidth, 0);
919 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, &CxtI);
921 APInt RHSKnownZero(BitWidth, 0);
922 APInt RHSKnownOne(BitWidth, 0);
923 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, &CxtI);
925 // Addition of two 2's compliment numbers having opposite signs will never
927 if ((LHSKnownOne[BitWidth - 1] && RHSKnownZero[BitWidth - 1]) ||
928 (LHSKnownZero[BitWidth - 1] && RHSKnownOne[BitWidth - 1]))
931 // Check if carry bit of addition will not cause overflow.
932 if (checkRippleForAdd(LHSKnownZero, RHSKnownZero))
934 if (checkRippleForAdd(RHSKnownZero, LHSKnownZero))
940 /// \brief Return true if we can prove that:
941 /// (sub LHS, RHS) === (sub nsw LHS, RHS)
942 /// This basically requires proving that the add in the original type would not
943 /// overflow to change the sign bit or have a carry out.
944 /// TODO: Handle this for Vectors.
945 bool InstCombiner::WillNotOverflowSignedSub(Value *LHS, Value *RHS,
947 // If LHS and RHS each have at least two sign bits, the subtraction
949 if (ComputeNumSignBits(LHS, 0, &CxtI) > 1 &&
950 ComputeNumSignBits(RHS, 0, &CxtI) > 1)
953 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
954 APInt LHSKnownZero(BitWidth, 0);
955 APInt LHSKnownOne(BitWidth, 0);
956 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, &CxtI);
958 APInt RHSKnownZero(BitWidth, 0);
959 APInt RHSKnownOne(BitWidth, 0);
960 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, &CxtI);
962 // Subtraction of two 2's compliment numbers having identical signs will
964 if ((LHSKnownOne[BitWidth - 1] && RHSKnownOne[BitWidth - 1]) ||
965 (LHSKnownZero[BitWidth - 1] && RHSKnownZero[BitWidth - 1]))
968 // TODO: implement logic similar to checkRippleForAdd
972 /// \brief Return true if we can prove that:
973 /// (sub LHS, RHS) === (sub nuw LHS, RHS)
974 bool InstCombiner::WillNotOverflowUnsignedSub(Value *LHS, Value *RHS,
976 // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
977 bool LHSKnownNonNegative, LHSKnownNegative;
978 bool RHSKnownNonNegative, RHSKnownNegative;
979 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, /*Depth=*/0,
981 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, /*Depth=*/0,
983 if (LHSKnownNegative && RHSKnownNonNegative)
989 // Checks if any operand is negative and we can convert add to sub.
990 // This function checks for following negative patterns
991 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
992 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
993 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
994 static Value *checkForNegativeOperand(BinaryOperator &I,
995 InstCombiner::BuilderTy *Builder) {
996 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
998 // This function creates 2 instructions to replace ADD, we need at least one
999 // of LHS or RHS to have one use to ensure benefit in transform.
1000 if (!LHS->hasOneUse() && !RHS->hasOneUse())
1003 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1004 const APInt *C1 = nullptr, *C2 = nullptr;
1006 // if ONE is on other side, swap
1007 if (match(RHS, m_Add(m_Value(X), m_One())))
1008 std::swap(LHS, RHS);
1010 if (match(LHS, m_Add(m_Value(X), m_One()))) {
1011 // if XOR on other side, swap
1012 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
1015 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
1016 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
1017 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
1018 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
1019 Value *NewAnd = Builder->CreateAnd(Z, *C1);
1020 return Builder->CreateSub(RHS, NewAnd, "sub");
1021 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
1022 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
1023 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
1024 Value *NewOr = Builder->CreateOr(Z, ~(*C1));
1025 return Builder->CreateSub(RHS, NewOr, "sub");
1030 // Restore LHS and RHS
1031 LHS = I.getOperand(0);
1032 RHS = I.getOperand(1);
1034 // if XOR is on other side, swap
1035 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
1036 std::swap(LHS, RHS);
1039 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
1040 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
1041 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
1042 if (C1->countTrailingZeros() == 0)
1043 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
1044 Value *NewOr = Builder->CreateOr(Z, ~(*C2));
1045 return Builder->CreateSub(RHS, NewOr, "sub");
1050 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1051 bool Changed = SimplifyAssociativeOrCommutative(I);
1052 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1054 if (Value *V = SimplifyVectorOp(I))
1055 return ReplaceInstUsesWith(I, V);
1057 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
1058 I.hasNoUnsignedWrap(), DL, TLI, DT, AC))
1059 return ReplaceInstUsesWith(I, V);
1061 // (A*B)+(A*C) -> A*(B+C) etc
1062 if (Value *V = SimplifyUsingDistributiveLaws(I))
1063 return ReplaceInstUsesWith(I, V);
1065 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1066 // X + (signbit) --> X ^ signbit
1067 const APInt &Val = CI->getValue();
1068 if (Val.isSignBit())
1069 return BinaryOperator::CreateXor(LHS, RHS);
1071 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1072 // (X & 254)+1 -> (X&254)|1
1073 if (SimplifyDemandedInstructionBits(I))
1076 // zext(bool) + C -> bool ? C + 1 : C
1077 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
1078 if (ZI->getSrcTy()->isIntegerTy(1))
1079 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
1081 Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1082 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1083 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
1084 const APInt &RHSVal = CI->getValue();
1085 unsigned ExtendAmt = 0;
1086 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1087 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1088 if (XorRHS->getValue() == -RHSVal) {
1089 if (RHSVal.isPowerOf2())
1090 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1091 else if (XorRHS->getValue().isPowerOf2())
1092 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1096 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1097 if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1102 Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt);
1103 Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext");
1104 return BinaryOperator::CreateAShr(NewShl, ShAmt);
1107 // If this is a xor that was canonicalized from a sub, turn it back into
1108 // a sub and fuse this add with it.
1109 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1110 IntegerType *IT = cast<IntegerType>(I.getType());
1111 APInt LHSKnownOne(IT->getBitWidth(), 0);
1112 APInt LHSKnownZero(IT->getBitWidth(), 0);
1113 computeKnownBits(XorLHS, LHSKnownZero, LHSKnownOne, 0, &I);
1114 if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue())
1115 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1118 // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C,
1119 // transform them into (X + (signbit ^ C))
1120 if (XorRHS->getValue().isSignBit())
1121 return BinaryOperator::CreateAdd(XorLHS,
1122 ConstantExpr::getXor(XorRHS, CI));
1126 if (isa<Constant>(RHS) && isa<PHINode>(LHS))
1127 if (Instruction *NV = FoldOpIntoPhi(I))
1130 if (I.getType()->getScalarType()->isIntegerTy(1))
1131 return BinaryOperator::CreateXor(LHS, RHS);
1135 BinaryOperator *New =
1136 BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1));
1137 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1138 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1143 // -A + -B --> -(A + B)
1144 if (Value *LHSV = dyn_castNegVal(LHS)) {
1145 if (!isa<Constant>(RHS))
1146 if (Value *RHSV = dyn_castNegVal(RHS)) {
1147 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
1148 return BinaryOperator::CreateNeg(NewAdd);
1151 return BinaryOperator::CreateSub(RHS, LHSV);
1155 if (!isa<Constant>(RHS))
1156 if (Value *V = dyn_castNegVal(RHS))
1157 return BinaryOperator::CreateSub(LHS, V);
1159 if (Value *V = checkForNegativeOperand(I, Builder))
1160 return ReplaceInstUsesWith(I, V);
1162 // A+B --> A|B iff A and B have no bits set in common.
1163 if (haveNoCommonBitsSet(LHS, RHS, DL, AC, &I, DT))
1164 return BinaryOperator::CreateOr(LHS, RHS);
1166 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
1168 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
1169 return BinaryOperator::CreateSub(SubOne(CRHS), X);
1172 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1173 // (X & FF00) + xx00 -> (X+xx00) & FF00
1176 if (LHS->hasOneUse() &&
1177 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1178 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1179 // See if all bits from the first bit set in the Add RHS up are included
1180 // in the mask. First, get the rightmost bit.
1181 const APInt &AddRHSV = CRHS->getValue();
1183 // Form a mask of all bits from the lowest bit added through the top.
1184 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1186 // See if the and mask includes all of these bits.
1187 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1189 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1190 // Okay, the xform is safe. Insert the new add pronto.
1191 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
1192 return BinaryOperator::CreateAnd(NewAdd, C2);
1196 // Try to fold constant add into select arguments.
1197 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1198 if (Instruction *R = FoldOpIntoSelect(I, SI))
1202 // add (select X 0 (sub n A)) A --> select X A n
1204 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1207 SI = dyn_cast<SelectInst>(RHS);
1210 if (SI && SI->hasOneUse()) {
1211 Value *TV = SI->getTrueValue();
1212 Value *FV = SI->getFalseValue();
1215 // Can we fold the add into the argument of the select?
1216 // We check both true and false select arguments for a matching subtract.
1217 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1218 // Fold the add into the true select value.
1219 return SelectInst::Create(SI->getCondition(), N, A);
1221 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1222 // Fold the add into the false select value.
1223 return SelectInst::Create(SI->getCondition(), A, N);
1227 // Check for (add (sext x), y), see if we can merge this into an
1228 // integer add followed by a sext.
1229 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
1230 // (add (sext x), cst) --> (sext (add x, cst'))
1231 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1233 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1234 if (LHSConv->hasOneUse() &&
1235 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
1236 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI, I)) {
1237 // Insert the new, smaller add.
1238 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1240 return new SExtInst(NewAdd, I.getType());
1244 // (add (sext x), (sext y)) --> (sext (add int x, y))
1245 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
1246 // Only do this if x/y have the same type, if at last one of them has a
1247 // single use (so we don't increase the number of sexts), and if the
1248 // integer add will not overflow.
1249 if (LHSConv->getOperand(0)->getType() ==
1250 RHSConv->getOperand(0)->getType() &&
1251 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1252 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1253 RHSConv->getOperand(0), I)) {
1254 // Insert the new integer add.
1255 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1256 RHSConv->getOperand(0), "addconv");
1257 return new SExtInst(NewAdd, I.getType());
1262 // (add (xor A, B) (and A, B)) --> (or A, B)
1264 Value *A = nullptr, *B = nullptr;
1265 if (match(RHS, m_Xor(m_Value(A), m_Value(B))) &&
1266 (match(LHS, m_And(m_Specific(A), m_Specific(B))) ||
1267 match(LHS, m_And(m_Specific(B), m_Specific(A)))))
1268 return BinaryOperator::CreateOr(A, B);
1270 if (match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
1271 (match(RHS, m_And(m_Specific(A), m_Specific(B))) ||
1272 match(RHS, m_And(m_Specific(B), m_Specific(A)))))
1273 return BinaryOperator::CreateOr(A, B);
1276 // (add (or A, B) (and A, B)) --> (add A, B)
1278 Value *A = nullptr, *B = nullptr;
1279 if (match(RHS, m_Or(m_Value(A), m_Value(B))) &&
1280 (match(LHS, m_And(m_Specific(A), m_Specific(B))) ||
1281 match(LHS, m_And(m_Specific(B), m_Specific(A))))) {
1282 auto *New = BinaryOperator::CreateAdd(A, B);
1283 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1284 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1288 if (match(LHS, m_Or(m_Value(A), m_Value(B))) &&
1289 (match(RHS, m_And(m_Specific(A), m_Specific(B))) ||
1290 match(RHS, m_And(m_Specific(B), m_Specific(A))))) {
1291 auto *New = BinaryOperator::CreateAdd(A, B);
1292 New->setHasNoSignedWrap(I.hasNoSignedWrap());
1293 New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1298 // TODO(jingyue): Consider WillNotOverflowSignedAdd and
1299 // WillNotOverflowUnsignedAdd to reduce the number of invocations of
1300 // computeKnownBits.
1301 if (!I.hasNoSignedWrap() && WillNotOverflowSignedAdd(LHS, RHS, I)) {
1303 I.setHasNoSignedWrap(true);
1305 if (!I.hasNoUnsignedWrap() &&
1306 computeOverflowForUnsignedAdd(LHS, RHS, &I) ==
1307 OverflowResult::NeverOverflows) {
1309 I.setHasNoUnsignedWrap(true);
1312 return Changed ? &I : nullptr;
1315 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
1316 bool Changed = SimplifyAssociativeOrCommutative(I);
1317 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1319 if (Value *V = SimplifyVectorOp(I))
1320 return ReplaceInstUsesWith(I, V);
1323 SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL, TLI, DT, AC))
1324 return ReplaceInstUsesWith(I, V);
1326 if (isa<Constant>(RHS)) {
1327 if (isa<PHINode>(LHS))
1328 if (Instruction *NV = FoldOpIntoPhi(I))
1331 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1332 if (Instruction *NV = FoldOpIntoSelect(I, SI))
1337 // -A + -B --> -(A + B)
1338 if (Value *LHSV = dyn_castFNegVal(LHS)) {
1339 Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV);
1340 RI->copyFastMathFlags(&I);
1345 if (!isa<Constant>(RHS))
1346 if (Value *V = dyn_castFNegVal(RHS)) {
1347 Instruction *RI = BinaryOperator::CreateFSub(LHS, V);
1348 RI->copyFastMathFlags(&I);
1352 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1353 // integer add followed by a promotion.
1354 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1355 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1356 // ... if the constant fits in the integer value. This is useful for things
1357 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1358 // requires a constant pool load, and generally allows the add to be better
1360 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
1362 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
1363 if (LHSConv->hasOneUse() &&
1364 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1365 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI, I)) {
1366 // Insert the new integer add.
1367 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1369 return new SIToFPInst(NewAdd, I.getType());
1373 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1374 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1375 // Only do this if x/y have the same type, if at last one of them has a
1376 // single use (so we don't increase the number of int->fp conversions),
1377 // and if the integer add will not overflow.
1378 if (LHSConv->getOperand(0)->getType() ==
1379 RHSConv->getOperand(0)->getType() &&
1380 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1381 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1382 RHSConv->getOperand(0), I)) {
1383 // Insert the new integer add.
1384 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1385 RHSConv->getOperand(0),"addconv");
1386 return new SIToFPInst(NewAdd, I.getType());
1391 // select C, 0, B + select C, A, 0 -> select C, A, B
1393 Value *A1, *B1, *C1, *A2, *B2, *C2;
1394 if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) &&
1395 match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) {
1397 Constant *Z1=nullptr, *Z2=nullptr;
1398 Value *A, *B, *C=C1;
1399 if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) {
1400 Z1 = dyn_cast<Constant>(A1); A = A2;
1401 Z2 = dyn_cast<Constant>(B2); B = B1;
1402 } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) {
1403 Z1 = dyn_cast<Constant>(B1); B = B2;
1404 Z2 = dyn_cast<Constant>(A2); A = A1;
1408 (I.hasNoSignedZeros() ||
1409 (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) {
1410 return SelectInst::Create(C, A, B);
1416 if (I.hasUnsafeAlgebra()) {
1417 if (Value *V = FAddCombine(Builder).simplify(&I))
1418 return ReplaceInstUsesWith(I, V);
1421 return Changed ? &I : nullptr;
1424 /// Optimize pointer differences into the same array into a size. Consider:
1425 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1426 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1428 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1430 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1432 bool Swapped = false;
1433 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1435 // For now we require one side to be the base pointer "A" or a constant
1436 // GEP derived from it.
1437 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1439 if (LHSGEP->getOperand(0) == RHS) {
1442 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1443 // (gep X, ...) - (gep X, ...)
1444 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1445 RHSGEP->getOperand(0)->stripPointerCasts()) {
1453 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1455 if (RHSGEP->getOperand(0) == LHS) {
1458 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1459 // (gep X, ...) - (gep X, ...)
1460 if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1461 LHSGEP->getOperand(0)->stripPointerCasts()) {
1469 // Avoid duplicating the arithmetic if GEP2 has non-constant indices and
1472 (GEP2 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse()))
1475 // Emit the offset of the GEP and an intptr_t.
1476 Value *Result = EmitGEPOffset(GEP1);
1478 // If we had a constant expression GEP on the other side offsetting the
1479 // pointer, subtract it from the offset we have.
1481 Value *Offset = EmitGEPOffset(GEP2);
1482 Result = Builder->CreateSub(Result, Offset);
1485 // If we have p - gep(p, ...) then we have to negate the result.
1487 Result = Builder->CreateNeg(Result, "diff.neg");
1489 return Builder->CreateIntCast(Result, Ty, true);
1492 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1493 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1495 if (Value *V = SimplifyVectorOp(I))
1496 return ReplaceInstUsesWith(I, V);
1498 if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(),
1499 I.hasNoUnsignedWrap(), DL, TLI, DT, AC))
1500 return ReplaceInstUsesWith(I, V);
1502 // (A*B)-(A*C) -> A*(B-C) etc
1503 if (Value *V = SimplifyUsingDistributiveLaws(I))
1504 return ReplaceInstUsesWith(I, V);
1506 // If this is a 'B = x-(-A)', change to B = x+A.
1507 if (Value *V = dyn_castNegVal(Op1)) {
1508 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1510 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1511 assert(BO->getOpcode() == Instruction::Sub &&
1512 "Expected a subtraction operator!");
1513 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1514 Res->setHasNoSignedWrap(true);
1516 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1517 Res->setHasNoSignedWrap(true);
1523 if (I.getType()->isIntegerTy(1))
1524 return BinaryOperator::CreateXor(Op0, Op1);
1526 // Replace (-1 - A) with (~A).
1527 if (match(Op0, m_AllOnes()))
1528 return BinaryOperator::CreateNot(Op1);
1530 if (Constant *C = dyn_cast<Constant>(Op0)) {
1531 // C - ~X == X + (1+C)
1533 if (match(Op1, m_Not(m_Value(X))))
1534 return BinaryOperator::CreateAdd(X, AddOne(C));
1536 // Try to fold constant sub into select arguments.
1537 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1538 if (Instruction *R = FoldOpIntoSelect(I, SI))
1541 // C-(X+C2) --> (C-C2)-X
1543 if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1544 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1546 if (SimplifyDemandedInstructionBits(I))
1549 // Fold (sub 0, (zext bool to B)) --> (sext bool to B)
1550 if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X))))
1551 if (X->getType()->getScalarType()->isIntegerTy(1))
1552 return CastInst::CreateSExtOrBitCast(X, Op1->getType());
1554 // Fold (sub 0, (sext bool to B)) --> (zext bool to B)
1555 if (C->isNullValue() && match(Op1, m_SExt(m_Value(X))))
1556 if (X->getType()->getScalarType()->isIntegerTy(1))
1557 return CastInst::CreateZExtOrBitCast(X, Op1->getType());
1560 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1561 // -(X >>u 31) -> (X >>s 31)
1562 // -(X >>s 31) -> (X >>u 31)
1566 if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) &&
1567 // Verify we are shifting out everything but the sign bit.
1568 CI->getValue() == I.getType()->getPrimitiveSizeInBits() - 1)
1569 return BinaryOperator::CreateAShr(X, CI);
1571 if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) &&
1572 // Verify we are shifting out everything but the sign bit.
1573 CI->getValue() == I.getType()->getPrimitiveSizeInBits() - 1)
1574 return BinaryOperator::CreateLShr(X, CI);
1577 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1579 APInt IntVal = C->getValue();
1580 if ((IntVal + 1).isPowerOf2()) {
1581 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1582 APInt KnownZero(BitWidth, 0);
1583 APInt KnownOne(BitWidth, 0);
1584 computeKnownBits(&I, KnownZero, KnownOne, 0, &I);
1585 if ((IntVal | KnownZero).isAllOnesValue()) {
1586 return BinaryOperator::CreateXor(Op1, C);
1593 // X-(X+Y) == -Y X-(Y+X) == -Y
1594 if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) ||
1595 match(Op1, m_Add(m_Value(Y), m_Specific(Op0))))
1596 return BinaryOperator::CreateNeg(Y);
1599 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1600 return BinaryOperator::CreateNeg(Y);
1603 // (sub (or A, B) (xor A, B)) --> (and A, B)
1605 Value *A = nullptr, *B = nullptr;
1606 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1607 (match(Op0, m_Or(m_Specific(A), m_Specific(B))) ||
1608 match(Op0, m_Or(m_Specific(B), m_Specific(A)))))
1609 return BinaryOperator::CreateAnd(A, B);
1612 if (Op0->hasOneUse()) {
1614 // ((X | Y) - X) --> (~X & Y)
1615 if (match(Op0, m_Or(m_Value(Y), m_Specific(Op1))) ||
1616 match(Op0, m_Or(m_Specific(Op1), m_Value(Y))))
1617 return BinaryOperator::CreateAnd(
1618 Y, Builder->CreateNot(Op1, Op1->getName() + ".not"));
1621 if (Op1->hasOneUse()) {
1622 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1623 Constant *C = nullptr;
1624 Constant *CI = nullptr;
1626 // (X - (Y - Z)) --> (X + (Z - Y)).
1627 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1628 return BinaryOperator::CreateAdd(Op0,
1629 Builder->CreateSub(Z, Y, Op1->getName()));
1631 // (X - (X & Y)) --> (X & ~Y)
1633 if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) ||
1634 match(Op1, m_And(m_Specific(Op0), m_Value(Y))))
1635 return BinaryOperator::CreateAnd(Op0,
1636 Builder->CreateNot(Y, Y->getName() + ".not"));
1638 // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow.
1639 if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1640 C->isNotMinSignedValue() && !C->isOneValue())
1641 return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1643 // 0 - (X << Y) -> (-X << Y) when X is freely negatable.
1644 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1645 if (Value *XNeg = dyn_castNegVal(X))
1646 return BinaryOperator::CreateShl(XNeg, Y);
1648 // X - A*-B -> X + A*B
1649 // X - -A*B -> X + A*B
1651 if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) ||
1652 match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B))))
1653 return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B));
1655 // X - A*CI -> X + A*-CI
1656 // X - CI*A -> X + A*-CI
1657 if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) ||
1658 match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) {
1659 Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI));
1660 return BinaryOperator::CreateAdd(Op0, NewMul);
1664 // Optimize pointer differences into the same array into a size. Consider:
1665 // &A[10] - &A[0]: we should compile this to "10".
1666 Value *LHSOp, *RHSOp;
1667 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1668 match(Op1, m_PtrToInt(m_Value(RHSOp))))
1669 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1670 return ReplaceInstUsesWith(I, Res);
1672 // trunc(p)-trunc(q) -> trunc(p-q)
1673 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1674 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1675 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1676 return ReplaceInstUsesWith(I, Res);
1678 bool Changed = false;
1679 if (!I.hasNoSignedWrap() && WillNotOverflowSignedSub(Op0, Op1, I)) {
1681 I.setHasNoSignedWrap(true);
1683 if (!I.hasNoUnsignedWrap() && WillNotOverflowUnsignedSub(Op0, Op1, I)) {
1685 I.setHasNoUnsignedWrap(true);
1688 return Changed ? &I : nullptr;
1691 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
1692 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1694 if (Value *V = SimplifyVectorOp(I))
1695 return ReplaceInstUsesWith(I, V);
1698 SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL, TLI, DT, AC))
1699 return ReplaceInstUsesWith(I, V);
1701 // fsub nsz 0, X ==> fsub nsz -0.0, X
1702 if (I.getFastMathFlags().noSignedZeros() && match(Op0, m_Zero())) {
1703 // Subtraction from -0.0 is the canonical form of fneg.
1704 Instruction *NewI = BinaryOperator::CreateFNeg(Op1);
1705 NewI->copyFastMathFlags(&I);
1709 if (isa<Constant>(Op0))
1710 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1711 if (Instruction *NV = FoldOpIntoSelect(I, SI))
1714 // If this is a 'B = x-(-A)', change to B = x+A, potentially looking
1715 // through FP extensions/truncations along the way.
1716 if (Value *V = dyn_castFNegVal(Op1)) {
1717 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V);
1718 NewI->copyFastMathFlags(&I);
1721 if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) {
1722 if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) {
1723 Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType());
1724 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc);
1725 NewI->copyFastMathFlags(&I);
1728 } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) {
1729 if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
1730 Value *NewExt = Builder->CreateFPExt(V, I.getType());
1731 Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt);
1732 NewI->copyFastMathFlags(&I);
1737 if (I.hasUnsafeAlgebra()) {
1738 if (Value *V = FAddCombine(Builder).simplify(&I))
1739 return ReplaceInstUsesWith(I, V);