1 //===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visit functions for add, fadd, sub, and fsub.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APFloat.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/Constant.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/InstrTypes.h"
23 #include "llvm/IR/Instruction.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/Operator.h"
26 #include "llvm/IR/PatternMatch.h"
27 #include "llvm/IR/Type.h"
28 #include "llvm/IR/Value.h"
29 #include "llvm/Support/AlignOf.h"
30 #include "llvm/Support/Casting.h"
31 #include "llvm/Support/KnownBits.h"
32 #include "llvm/Transforms/InstCombine/InstCombiner.h"
33 #include <cassert>
34 #include <utility>
35 
36 using namespace llvm;
37 using namespace PatternMatch;
38 
39 #define DEBUG_TYPE "instcombine"
40 
41 namespace {
42 
43   /// Class representing coefficient of floating-point addend.
44   /// This class needs to be highly efficient, which is especially true for
45   /// the constructor. As of I write this comment, the cost of the default
46   /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
47   /// perform write-merging).
48   ///
49   class FAddendCoef {
50   public:
51     // The constructor has to initialize a APFloat, which is unnecessary for
52     // most addends which have coefficient either 1 or -1. So, the constructor
53     // is expensive. In order to avoid the cost of the constructor, we should
54     // reuse some instances whenever possible. The pre-created instances
55     // FAddCombine::Add[0-5] embodies this idea.
56     FAddendCoef() = default;
57     ~FAddendCoef();
58 
59     // If possible, don't define operator+/operator- etc because these
60     // operators inevitably call FAddendCoef's constructor which is not cheap.
61     void operator=(const FAddendCoef &A);
62     void operator+=(const FAddendCoef &A);
63     void operator*=(const FAddendCoef &S);
64 
65     void set(short C) {
66       assert(!insaneIntVal(C) && "Insane coefficient");
67       IsFp = false; IntVal = C;
68     }
69 
70     void set(const APFloat& C);
71 
72     void negate();
73 
74     bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75     Value *getValue(Type *) const;
76 
77     bool isOne() const { return isInt() && IntVal == 1; }
78     bool isTwo() const { return isInt() && IntVal == 2; }
79     bool isMinusOne() const { return isInt() && IntVal == -1; }
80     bool isMinusTwo() const { return isInt() && IntVal == -2; }
81 
82   private:
83     bool insaneIntVal(int V) { return V > 4 || V < -4; }
84 
85     APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86 
87     const APFloat *getFpValPtr() const {
88       return reinterpret_cast<const APFloat *>(&FpValBuf);
89     }
90 
91     const APFloat &getFpVal() const {
92       assert(IsFp && BufHasFpVal && "Incorret state");
93       return *getFpValPtr();
94     }
95 
96     APFloat &getFpVal() {
97       assert(IsFp && BufHasFpVal && "Incorret state");
98       return *getFpValPtr();
99     }
100 
101     bool isInt() const { return !IsFp; }
102 
103     // If the coefficient is represented by an integer, promote it to a
104     // floating point.
105     void convertToFpType(const fltSemantics &Sem);
106 
107     // Construct an APFloat from a signed integer.
108     // TODO: We should get rid of this function when APFloat can be constructed
109     //       from an *SIGNED* integer.
110     APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
111 
112     bool IsFp = false;
113 
114     // True iff FpValBuf contains an instance of APFloat.
115     bool BufHasFpVal = false;
116 
117     // The integer coefficient of an individual addend is either 1 or -1,
118     // and we try to simplify at most 4 addends from neighboring at most
119     // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120     // is overkill of this end.
121     short IntVal = 0;
122 
123     AlignedCharArrayUnion<APFloat> FpValBuf;
124   };
125 
126   /// FAddend is used to represent floating-point addend. An addend is
127   /// represented as <C, V>, where the V is a symbolic value, and C is a
128   /// constant coefficient. A constant addend is represented as <C, 0>.
129   class FAddend {
130   public:
131     FAddend() = default;
132 
133     void operator+=(const FAddend &T) {
134       assert((Val == T.Val) && "Symbolic-values disagree");
135       Coeff += T.Coeff;
136     }
137 
138     Value *getSymVal() const { return Val; }
139     const FAddendCoef &getCoef() const { return Coeff; }
140 
141     bool isConstant() const { return Val == nullptr; }
142     bool isZero() const { return Coeff.isZero(); }
143 
144     void set(short Coefficient, Value *V) {
145       Coeff.set(Coefficient);
146       Val = V;
147     }
148     void set(const APFloat &Coefficient, Value *V) {
149       Coeff.set(Coefficient);
150       Val = V;
151     }
152     void set(const ConstantFP *Coefficient, Value *V) {
153       Coeff.set(Coefficient->getValueAPF());
154       Val = V;
155     }
156 
157     void negate() { Coeff.negate(); }
158 
159     /// Drill down the U-D chain one step to find the definition of V, and
160     /// try to break the definition into one or two addends.
161     static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
162 
163     /// Similar to FAddend::drillDownOneStep() except that the value being
164     /// splitted is the addend itself.
165     unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
166 
167   private:
168     void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
169 
170     // This addend has the value of "Coeff * Val".
171     Value *Val = nullptr;
172     FAddendCoef Coeff;
173   };
174 
175   /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176   /// with its neighboring at most two instructions.
177   ///
178   class FAddCombine {
179   public:
180     FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
181 
182     Value *simplify(Instruction *FAdd);
183 
184   private:
185     using AddendVect = SmallVector<const FAddend *, 4>;
186 
187     Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
188 
189     /// Convert given addend to a Value
190     Value *createAddendVal(const FAddend &A, bool& NeedNeg);
191 
192     /// Return the number of instructions needed to emit the N-ary addition.
193     unsigned calcInstrNumber(const AddendVect& Vect);
194 
195     Value *createFSub(Value *Opnd0, Value *Opnd1);
196     Value *createFAdd(Value *Opnd0, Value *Opnd1);
197     Value *createFMul(Value *Opnd0, Value *Opnd1);
198     Value *createFNeg(Value *V);
199     Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200     void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
201 
202      // Debugging stuff are clustered here.
203     #ifndef NDEBUG
204       unsigned CreateInstrNum;
205       void initCreateInstNum() { CreateInstrNum = 0; }
206       void incCreateInstNum() { CreateInstrNum++; }
207     #else
208       void initCreateInstNum() {}
209       void incCreateInstNum() {}
210     #endif
211 
212     InstCombiner::BuilderTy &Builder;
213     Instruction *Instr = nullptr;
214   };
215 
216 } // end anonymous namespace
217 
218 //===----------------------------------------------------------------------===//
219 //
220 // Implementation of
221 //    {FAddendCoef, FAddend, FAddition, FAddCombine}.
222 //
223 //===----------------------------------------------------------------------===//
224 FAddendCoef::~FAddendCoef() {
225   if (BufHasFpVal)
226     getFpValPtr()->~APFloat();
227 }
228 
229 void FAddendCoef::set(const APFloat& C) {
230   APFloat *P = getFpValPtr();
231 
232   if (isInt()) {
233     // As the buffer is meanless byte stream, we cannot call
234     // APFloat::operator=().
235     new(P) APFloat(C);
236   } else
237     *P = C;
238 
239   IsFp = BufHasFpVal = true;
240 }
241 
242 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
243   if (!isInt())
244     return;
245 
246   APFloat *P = getFpValPtr();
247   if (IntVal > 0)
248     new(P) APFloat(Sem, IntVal);
249   else {
250     new(P) APFloat(Sem, 0 - IntVal);
251     P->changeSign();
252   }
253   IsFp = BufHasFpVal = true;
254 }
255 
256 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
257   if (Val >= 0)
258     return APFloat(Sem, Val);
259 
260   APFloat T(Sem, 0 - Val);
261   T.changeSign();
262 
263   return T;
264 }
265 
266 void FAddendCoef::operator=(const FAddendCoef &That) {
267   if (That.isInt())
268     set(That.IntVal);
269   else
270     set(That.getFpVal());
271 }
272 
273 void FAddendCoef::operator+=(const FAddendCoef &That) {
274   RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275   if (isInt() == That.isInt()) {
276     if (isInt())
277       IntVal += That.IntVal;
278     else
279       getFpVal().add(That.getFpVal(), RndMode);
280     return;
281   }
282 
283   if (isInt()) {
284     const APFloat &T = That.getFpVal();
285     convertToFpType(T.getSemantics());
286     getFpVal().add(T, RndMode);
287     return;
288   }
289 
290   APFloat &T = getFpVal();
291   T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292 }
293 
294 void FAddendCoef::operator*=(const FAddendCoef &That) {
295   if (That.isOne())
296     return;
297 
298   if (That.isMinusOne()) {
299     negate();
300     return;
301   }
302 
303   if (isInt() && That.isInt()) {
304     int Res = IntVal * (int)That.IntVal;
305     assert(!insaneIntVal(Res) && "Insane int value");
306     IntVal = Res;
307     return;
308   }
309 
310   const fltSemantics &Semantic =
311     isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312 
313   if (isInt())
314     convertToFpType(Semantic);
315   APFloat &F0 = getFpVal();
316 
317   if (That.isInt())
318     F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319                 APFloat::rmNearestTiesToEven);
320   else
321     F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322 }
323 
324 void FAddendCoef::negate() {
325   if (isInt())
326     IntVal = 0 - IntVal;
327   else
328     getFpVal().changeSign();
329 }
330 
331 Value *FAddendCoef::getValue(Type *Ty) const {
332   return isInt() ?
333     ConstantFP::get(Ty, float(IntVal)) :
334     ConstantFP::get(Ty->getContext(), getFpVal());
335 }
336 
337 // The definition of <Val>     Addends
338 // =========================================
339 //  A + B                     <1, A>, <1,B>
340 //  A - B                     <1, A>, <1,B>
341 //  0 - B                     <-1, B>
342 //  C * A,                    <C, A>
343 //  A + C                     <1, A> <C, NULL>
344 //  0 +/- 0                   <0, NULL> (corner case)
345 //
346 // Legend: A and B are not constant, C is constant
347 unsigned FAddend::drillValueDownOneStep
348   (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349   Instruction *I = nullptr;
350   if (!Val || !(I = dyn_cast<Instruction>(Val)))
351     return 0;
352 
353   unsigned Opcode = I->getOpcode();
354 
355   if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356     ConstantFP *C0, *C1;
357     Value *Opnd0 = I->getOperand(0);
358     Value *Opnd1 = I->getOperand(1);
359     if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360       Opnd0 = nullptr;
361 
362     if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363       Opnd1 = nullptr;
364 
365     if (Opnd0) {
366       if (!C0)
367         Addend0.set(1, Opnd0);
368       else
369         Addend0.set(C0, nullptr);
370     }
371 
372     if (Opnd1) {
373       FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374       if (!C1)
375         Addend.set(1, Opnd1);
376       else
377         Addend.set(C1, nullptr);
378       if (Opcode == Instruction::FSub)
379         Addend.negate();
380     }
381 
382     if (Opnd0 || Opnd1)
383       return Opnd0 && Opnd1 ? 2 : 1;
384 
385     // Both operands are zero. Weird!
386     Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387     return 1;
388   }
389 
390   if (I->getOpcode() == Instruction::FMul) {
391     Value *V0 = I->getOperand(0);
392     Value *V1 = I->getOperand(1);
393     if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394       Addend0.set(C, V1);
395       return 1;
396     }
397 
398     if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399       Addend0.set(C, V0);
400       return 1;
401     }
402   }
403 
404   return 0;
405 }
406 
407 // Try to break *this* addend into two addends. e.g. Suppose this addend is
408 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409 // i.e. <2.3, X> and <2.3, Y>.
410 unsigned FAddend::drillAddendDownOneStep
411   (FAddend &Addend0, FAddend &Addend1) const {
412   if (isConstant())
413     return 0;
414 
415   unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416   if (!BreakNum || Coeff.isOne())
417     return BreakNum;
418 
419   Addend0.Scale(Coeff);
420 
421   if (BreakNum == 2)
422     Addend1.Scale(Coeff);
423 
424   return BreakNum;
425 }
426 
427 Value *FAddCombine::simplify(Instruction *I) {
428   assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429          "Expected 'reassoc'+'nsz' instruction");
430 
431   // Currently we are not able to handle vector type.
432   if (I->getType()->isVectorTy())
433     return nullptr;
434 
435   assert((I->getOpcode() == Instruction::FAdd ||
436           I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437 
438   // Save the instruction before calling other member-functions.
439   Instr = I;
440 
441   FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442 
443   unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444 
445   // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446   unsigned Opnd0_ExpNum = 0;
447   unsigned Opnd1_ExpNum = 0;
448 
449   if (!Opnd0.isConstant())
450     Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451 
452   // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453   if (OpndNum == 2 && !Opnd1.isConstant())
454     Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455 
456   // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457   if (Opnd0_ExpNum && Opnd1_ExpNum) {
458     AddendVect AllOpnds;
459     AllOpnds.push_back(&Opnd0_0);
460     AllOpnds.push_back(&Opnd1_0);
461     if (Opnd0_ExpNum == 2)
462       AllOpnds.push_back(&Opnd0_1);
463     if (Opnd1_ExpNum == 2)
464       AllOpnds.push_back(&Opnd1_1);
465 
466     // Compute instruction quota. We should save at least one instruction.
467     unsigned InstQuota = 0;
468 
469     Value *V0 = I->getOperand(0);
470     Value *V1 = I->getOperand(1);
471     InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472                  (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473 
474     if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475       return R;
476   }
477 
478   if (OpndNum != 2) {
479     // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480     // splitted into two addends, say "V = X - Y", the instruction would have
481     // been optimized into "I = Y - X" in the previous steps.
482     //
483     const FAddendCoef &CE = Opnd0.getCoef();
484     return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485   }
486 
487   // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488   if (Opnd1_ExpNum) {
489     AddendVect AllOpnds;
490     AllOpnds.push_back(&Opnd0);
491     AllOpnds.push_back(&Opnd1_0);
492     if (Opnd1_ExpNum == 2)
493       AllOpnds.push_back(&Opnd1_1);
494 
495     if (Value *R = simplifyFAdd(AllOpnds, 1))
496       return R;
497   }
498 
499   // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500   if (Opnd0_ExpNum) {
501     AddendVect AllOpnds;
502     AllOpnds.push_back(&Opnd1);
503     AllOpnds.push_back(&Opnd0_0);
504     if (Opnd0_ExpNum == 2)
505       AllOpnds.push_back(&Opnd0_1);
506 
507     if (Value *R = simplifyFAdd(AllOpnds, 1))
508       return R;
509   }
510 
511   return nullptr;
512 }
513 
514 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515   unsigned AddendNum = Addends.size();
516   assert(AddendNum <= 4 && "Too many addends");
517 
518   // For saving intermediate results;
519   unsigned NextTmpIdx = 0;
520   FAddend TmpResult[3];
521 
522   // Simplified addends are placed <SimpVect>.
523   AddendVect SimpVect;
524 
525   // The outer loop works on one symbolic-value at a time. Suppose the input
526   // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527   // The symbolic-values will be processed in this order: x, y, z.
528   for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529 
530     const FAddend *ThisAddend = Addends[SymIdx];
531     if (!ThisAddend) {
532       // This addend was processed before.
533       continue;
534     }
535 
536     Value *Val = ThisAddend->getSymVal();
537 
538     // If the resulting expr has constant-addend, this constant-addend is
539     // desirable to reside at the top of the resulting expression tree. Placing
540     // constant close to super-expr(s) will potentially reveal some
541     // optimization opportunities in super-expr(s). Here we do not implement
542     // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543     // call later.
544 
545     unsigned StartIdx = SimpVect.size();
546     SimpVect.push_back(ThisAddend);
547 
548     // The inner loop collects addends sharing same symbolic-value, and these
549     // addends will be later on folded into a single addend. Following above
550     // example, if the symbolic value "y" is being processed, the inner loop
551     // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552     // be later on folded into "<b1+b2, y>".
553     for (unsigned SameSymIdx = SymIdx + 1;
554          SameSymIdx < AddendNum; SameSymIdx++) {
555       const FAddend *T = Addends[SameSymIdx];
556       if (T && T->getSymVal() == Val) {
557         // Set null such that next iteration of the outer loop will not process
558         // this addend again.
559         Addends[SameSymIdx] = nullptr;
560         SimpVect.push_back(T);
561       }
562     }
563 
564     // If multiple addends share same symbolic value, fold them together.
565     if (StartIdx + 1 != SimpVect.size()) {
566       FAddend &R = TmpResult[NextTmpIdx ++];
567       R = *SimpVect[StartIdx];
568       for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569         R += *SimpVect[Idx];
570 
571       // Pop all addends being folded and push the resulting folded addend.
572       SimpVect.resize(StartIdx);
573       if (!R.isZero()) {
574         SimpVect.push_back(&R);
575       }
576     }
577   }
578 
579   assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
580          "out-of-bound access");
581 
582   Value *Result;
583   if (!SimpVect.empty())
584     Result = createNaryFAdd(SimpVect, InstrQuota);
585   else {
586     // The addition is folded to 0.0.
587     Result = ConstantFP::get(Instr->getType(), 0.0);
588   }
589 
590   return Result;
591 }
592 
593 Value *FAddCombine::createNaryFAdd
594   (const AddendVect &Opnds, unsigned InstrQuota) {
595   assert(!Opnds.empty() && "Expect at least one addend");
596 
597   // Step 1: Check if the # of instructions needed exceeds the quota.
598 
599   unsigned InstrNeeded = calcInstrNumber(Opnds);
600   if (InstrNeeded > InstrQuota)
601     return nullptr;
602 
603   initCreateInstNum();
604 
605   // step 2: Emit the N-ary addition.
606   // Note that at most three instructions are involved in Fadd-InstCombine: the
607   // addition in question, and at most two neighboring instructions.
608   // The resulting optimized addition should have at least one less instruction
609   // than the original addition expression tree. This implies that the resulting
610   // N-ary addition has at most two instructions, and we don't need to worry
611   // about tree-height when constructing the N-ary addition.
612 
613   Value *LastVal = nullptr;
614   bool LastValNeedNeg = false;
615 
616   // Iterate the addends, creating fadd/fsub using adjacent two addends.
617   for (const FAddend *Opnd : Opnds) {
618     bool NeedNeg;
619     Value *V = createAddendVal(*Opnd, NeedNeg);
620     if (!LastVal) {
621       LastVal = V;
622       LastValNeedNeg = NeedNeg;
623       continue;
624     }
625 
626     if (LastValNeedNeg == NeedNeg) {
627       LastVal = createFAdd(LastVal, V);
628       continue;
629     }
630 
631     if (LastValNeedNeg)
632       LastVal = createFSub(V, LastVal);
633     else
634       LastVal = createFSub(LastVal, V);
635 
636     LastValNeedNeg = false;
637   }
638 
639   if (LastValNeedNeg) {
640     LastVal = createFNeg(LastVal);
641   }
642 
643 #ifndef NDEBUG
644   assert(CreateInstrNum == InstrNeeded &&
645          "Inconsistent in instruction numbers");
646 #endif
647 
648   return LastVal;
649 }
650 
651 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
652   Value *V = Builder.CreateFSub(Opnd0, Opnd1);
653   if (Instruction *I = dyn_cast<Instruction>(V))
654     createInstPostProc(I);
655   return V;
656 }
657 
658 Value *FAddCombine::createFNeg(Value *V) {
659   Value *NewV = Builder.CreateFNeg(V);
660   if (Instruction *I = dyn_cast<Instruction>(NewV))
661     createInstPostProc(I, true); // fneg's don't receive instruction numbers.
662   return NewV;
663 }
664 
665 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
666   Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
667   if (Instruction *I = dyn_cast<Instruction>(V))
668     createInstPostProc(I);
669   return V;
670 }
671 
672 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
673   Value *V = Builder.CreateFMul(Opnd0, Opnd1);
674   if (Instruction *I = dyn_cast<Instruction>(V))
675     createInstPostProc(I);
676   return V;
677 }
678 
679 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
680   NewInstr->setDebugLoc(Instr->getDebugLoc());
681 
682   // Keep track of the number of instruction created.
683   if (!NoNumber)
684     incCreateInstNum();
685 
686   // Propagate fast-math flags
687   NewInstr->setFastMathFlags(Instr->getFastMathFlags());
688 }
689 
690 // Return the number of instruction needed to emit the N-ary addition.
691 // NOTE: Keep this function in sync with createAddendVal().
692 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
693   unsigned OpndNum = Opnds.size();
694   unsigned InstrNeeded = OpndNum - 1;
695 
696   // The number of addends in the form of "(-1)*x".
697   unsigned NegOpndNum = 0;
698 
699   // Adjust the number of instructions needed to emit the N-ary add.
700   for (const FAddend *Opnd : Opnds) {
701     if (Opnd->isConstant())
702       continue;
703 
704     // The constant check above is really for a few special constant
705     // coefficients.
706     if (isa<UndefValue>(Opnd->getSymVal()))
707       continue;
708 
709     const FAddendCoef &CE = Opnd->getCoef();
710     if (CE.isMinusOne() || CE.isMinusTwo())
711       NegOpndNum++;
712 
713     // Let the addend be "c * x". If "c == +/-1", the value of the addend
714     // is immediately available; otherwise, it needs exactly one instruction
715     // to evaluate the value.
716     if (!CE.isMinusOne() && !CE.isOne())
717       InstrNeeded++;
718   }
719   return InstrNeeded;
720 }
721 
722 // Input Addend        Value           NeedNeg(output)
723 // ================================================================
724 // Constant C          C               false
725 // <+/-1, V>           V               coefficient is -1
726 // <2/-2, V>          "fadd V, V"      coefficient is -2
727 // <C, V>             "fmul V, C"      false
728 //
729 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
730 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
731   const FAddendCoef &Coeff = Opnd.getCoef();
732 
733   if (Opnd.isConstant()) {
734     NeedNeg = false;
735     return Coeff.getValue(Instr->getType());
736   }
737 
738   Value *OpndVal = Opnd.getSymVal();
739 
740   if (Coeff.isMinusOne() || Coeff.isOne()) {
741     NeedNeg = Coeff.isMinusOne();
742     return OpndVal;
743   }
744 
745   if (Coeff.isTwo() || Coeff.isMinusTwo()) {
746     NeedNeg = Coeff.isMinusTwo();
747     return createFAdd(OpndVal, OpndVal);
748   }
749 
750   NeedNeg = false;
751   return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
752 }
753 
754 // Checks if any operand is negative and we can convert add to sub.
755 // This function checks for following negative patterns
756 //   ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
757 //   ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
758 //   XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
759 static Value *checkForNegativeOperand(BinaryOperator &I,
760                                       InstCombiner::BuilderTy &Builder) {
761   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
762 
763   // This function creates 2 instructions to replace ADD, we need at least one
764   // of LHS or RHS to have one use to ensure benefit in transform.
765   if (!LHS->hasOneUse() && !RHS->hasOneUse())
766     return nullptr;
767 
768   Value *X = nullptr, *Y = nullptr, *Z = nullptr;
769   const APInt *C1 = nullptr, *C2 = nullptr;
770 
771   // if ONE is on other side, swap
772   if (match(RHS, m_Add(m_Value(X), m_One())))
773     std::swap(LHS, RHS);
774 
775   if (match(LHS, m_Add(m_Value(X), m_One()))) {
776     // if XOR on other side, swap
777     if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
778       std::swap(X, RHS);
779 
780     if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
781       // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
782       // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
783       if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
784         Value *NewAnd = Builder.CreateAnd(Z, *C1);
785         return Builder.CreateSub(RHS, NewAnd, "sub");
786       } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
787         // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
788         // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
789         Value *NewOr = Builder.CreateOr(Z, ~(*C1));
790         return Builder.CreateSub(RHS, NewOr, "sub");
791       }
792     }
793   }
794 
795   // Restore LHS and RHS
796   LHS = I.getOperand(0);
797   RHS = I.getOperand(1);
798 
799   // if XOR is on other side, swap
800   if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
801     std::swap(LHS, RHS);
802 
803   // C2 is ODD
804   // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
805   // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
806   if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
807     if (C1->countTrailingZeros() == 0)
808       if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
809         Value *NewOr = Builder.CreateOr(Z, ~(*C2));
810         return Builder.CreateSub(RHS, NewOr, "sub");
811       }
812   return nullptr;
813 }
814 
815 /// Wrapping flags may allow combining constants separated by an extend.
816 static Instruction *foldNoWrapAdd(BinaryOperator &Add,
817                                   InstCombiner::BuilderTy &Builder) {
818   Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
819   Type *Ty = Add.getType();
820   Constant *Op1C;
821   if (!match(Op1, m_Constant(Op1C)))
822     return nullptr;
823 
824   // Try this match first because it results in an add in the narrow type.
825   // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
826   Value *X;
827   const APInt *C1, *C2;
828   if (match(Op1, m_APInt(C1)) &&
829       match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
830       C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
831     Constant *NewC =
832         ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
833     return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
834   }
835 
836   // More general combining of constants in the wide type.
837   // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
838   Constant *NarrowC;
839   if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
840     Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
841     Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
842     Value *WideX = Builder.CreateSExt(X, Ty);
843     return BinaryOperator::CreateAdd(WideX, NewC);
844   }
845   // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
846   if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
847     Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
848     Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
849     Value *WideX = Builder.CreateZExt(X, Ty);
850     return BinaryOperator::CreateAdd(WideX, NewC);
851   }
852 
853   return nullptr;
854 }
855 
856 Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
857   Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
858   Constant *Op1C;
859   if (!match(Op1, m_ImmConstant(Op1C)))
860     return nullptr;
861 
862   if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
863     return NV;
864 
865   Value *X;
866   Constant *Op00C;
867 
868   // add (sub C1, X), C2 --> sub (add C1, C2), X
869   if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
870     return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
871 
872   Value *Y;
873 
874   // add (sub X, Y), -1 --> add (not Y), X
875   if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
876       match(Op1, m_AllOnes()))
877     return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
878 
879   // zext(bool) + C -> bool ? C + 1 : C
880   if (match(Op0, m_ZExt(m_Value(X))) &&
881       X->getType()->getScalarSizeInBits() == 1)
882     return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
883   // sext(bool) + C -> bool ? C - 1 : C
884   if (match(Op0, m_SExt(m_Value(X))) &&
885       X->getType()->getScalarSizeInBits() == 1)
886     return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
887 
888   // ~X + C --> (C-1) - X
889   if (match(Op0, m_Not(m_Value(X))))
890     return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X);
891 
892   const APInt *C;
893   if (!match(Op1, m_APInt(C)))
894     return nullptr;
895 
896   // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
897   Constant *Op01C;
898   if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
899       haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
900     return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
901 
902   // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
903   const APInt *C2;
904   if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
905     return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
906 
907   if (C->isSignMask()) {
908     // If wrapping is not allowed, then the addition must set the sign bit:
909     // X + (signmask) --> X | signmask
910     if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
911       return BinaryOperator::CreateOr(Op0, Op1);
912 
913     // If wrapping is allowed, then the addition flips the sign bit of LHS:
914     // X + (signmask) --> X ^ signmask
915     return BinaryOperator::CreateXor(Op0, Op1);
916   }
917 
918   // Is this add the last step in a convoluted sext?
919   // add(zext(xor i16 X, -32768), -32768) --> sext X
920   Type *Ty = Add.getType();
921   if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
922       C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
923     return CastInst::Create(Instruction::SExt, X, Ty);
924 
925   if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
926     // (X ^ signmask) + C --> (X + (signmask ^ C))
927     if (C2->isSignMask())
928       return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
929 
930     // If X has no high-bits set above an xor mask:
931     // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
932     if (C2->isMask()) {
933       KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
934       if ((*C2 | LHSKnown.Zero).isAllOnes())
935         return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
936     }
937 
938     // Look for a math+logic pattern that corresponds to sext-in-register of a
939     // value with cleared high bits. Convert that into a pair of shifts:
940     // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
941     // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
942     if (Op0->hasOneUse() && *C2 == -(*C)) {
943       unsigned BitWidth = Ty->getScalarSizeInBits();
944       unsigned ShAmt = 0;
945       if (C->isPowerOf2())
946         ShAmt = BitWidth - C->logBase2() - 1;
947       else if (C2->isPowerOf2())
948         ShAmt = BitWidth - C2->logBase2() - 1;
949       if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
950                                      0, &Add)) {
951         Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
952         Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
953         return BinaryOperator::CreateAShr(NewShl, ShAmtC);
954       }
955     }
956   }
957 
958   if (C->isOne() && Op0->hasOneUse()) {
959     // add (sext i1 X), 1 --> zext (not X)
960     // TODO: The smallest IR representation is (select X, 0, 1), and that would
961     // not require the one-use check. But we need to remove a transform in
962     // visitSelect and make sure that IR value tracking for select is equal or
963     // better than for these ops.
964     if (match(Op0, m_SExt(m_Value(X))) &&
965         X->getType()->getScalarSizeInBits() == 1)
966       return new ZExtInst(Builder.CreateNot(X), Ty);
967 
968     // Shifts and add used to flip and mask off the low bit:
969     // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
970     const APInt *C3;
971     if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
972         C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
973       Value *NotX = Builder.CreateNot(X);
974       return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
975     }
976   }
977 
978   // If all bits affected by the add are included in a high-bit-mask, do the
979   // add before the mask op:
980   // (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00
981   if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) &&
982       C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) {
983     Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C));
984     return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2));
985   }
986 
987   return nullptr;
988 }
989 
990 // Matches multiplication expression Op * C where C is a constant. Returns the
991 // constant value in C and the other operand in Op. Returns true if such a
992 // match is found.
993 static bool MatchMul(Value *E, Value *&Op, APInt &C) {
994   const APInt *AI;
995   if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
996     C = *AI;
997     return true;
998   }
999   if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1000     C = APInt(AI->getBitWidth(), 1);
1001     C <<= *AI;
1002     return true;
1003   }
1004   return false;
1005 }
1006 
1007 // Matches remainder expression Op % C where C is a constant. Returns the
1008 // constant value in C and the other operand in Op. Returns the signedness of
1009 // the remainder operation in IsSigned. Returns true if such a match is
1010 // found.
1011 static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1012   const APInt *AI;
1013   IsSigned = false;
1014   if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1015     IsSigned = true;
1016     C = *AI;
1017     return true;
1018   }
1019   if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1020     C = *AI;
1021     return true;
1022   }
1023   if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1024     C = *AI + 1;
1025     return true;
1026   }
1027   return false;
1028 }
1029 
1030 // Matches division expression Op / C with the given signedness as indicated
1031 // by IsSigned, where C is a constant. Returns the constant value in C and the
1032 // other operand in Op. Returns true if such a match is found.
1033 static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1034   const APInt *AI;
1035   if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1036     C = *AI;
1037     return true;
1038   }
1039   if (!IsSigned) {
1040     if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1041       C = *AI;
1042       return true;
1043     }
1044     if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1045       C = APInt(AI->getBitWidth(), 1);
1046       C <<= *AI;
1047       return true;
1048     }
1049   }
1050   return false;
1051 }
1052 
1053 // Returns whether C0 * C1 with the given signedness overflows.
1054 static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1055   bool overflow;
1056   if (IsSigned)
1057     (void)C0.smul_ov(C1, overflow);
1058   else
1059     (void)C0.umul_ov(C1, overflow);
1060   return overflow;
1061 }
1062 
1063 // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1064 // does not overflow.
1065 Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1066   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1067   Value *X, *MulOpV;
1068   APInt C0, MulOpC;
1069   bool IsSigned;
1070   // Match I = X % C0 + MulOpV * C0
1071   if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1072        (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1073       C0 == MulOpC) {
1074     Value *RemOpV;
1075     APInt C1;
1076     bool Rem2IsSigned;
1077     // Match MulOpC = RemOpV % C1
1078     if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1079         IsSigned == Rem2IsSigned) {
1080       Value *DivOpV;
1081       APInt DivOpC;
1082       // Match RemOpV = X / C0
1083       if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1084           C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1085         Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1086         return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1087                         : Builder.CreateURem(X, NewDivisor, "urem");
1088       }
1089     }
1090   }
1091 
1092   return nullptr;
1093 }
1094 
1095 /// Fold
1096 ///   (1 << NBits) - 1
1097 /// Into:
1098 ///   ~(-(1 << NBits))
1099 /// Because a 'not' is better for bit-tracking analysis and other transforms
1100 /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1101 static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1102                                            InstCombiner::BuilderTy &Builder) {
1103   Value *NBits;
1104   if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1105     return nullptr;
1106 
1107   Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1108   Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1109   // Be wary of constant folding.
1110   if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1111     // Always NSW. But NUW propagates from `add`.
1112     BOp->setHasNoSignedWrap();
1113     BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1114   }
1115 
1116   return BinaryOperator::CreateNot(NotMask, I.getName());
1117 }
1118 
1119 static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1120   assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1121   Type *Ty = I.getType();
1122   auto getUAddSat = [&]() {
1123     return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1124   };
1125 
1126   // add (umin X, ~Y), Y --> uaddsat X, Y
1127   Value *X, *Y;
1128   if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1129                         m_Deferred(Y))))
1130     return CallInst::Create(getUAddSat(), { X, Y });
1131 
1132   // add (umin X, ~C), C --> uaddsat X, C
1133   const APInt *C, *NotC;
1134   if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1135       *C == ~*NotC)
1136     return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1137 
1138   return nullptr;
1139 }
1140 
1141 Instruction *InstCombinerImpl::
1142     canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1143         BinaryOperator &I) {
1144   assert((I.getOpcode() == Instruction::Add ||
1145           I.getOpcode() == Instruction::Or ||
1146           I.getOpcode() == Instruction::Sub) &&
1147          "Expecting add/or/sub instruction");
1148 
1149   // We have a subtraction/addition between a (potentially truncated) *logical*
1150   // right-shift of X and a "select".
1151   Value *X, *Select;
1152   Instruction *LowBitsToSkip, *Extract;
1153   if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd(
1154                                m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1155                                m_Instruction(Extract))),
1156                            m_Value(Select))))
1157     return nullptr;
1158 
1159   // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1160   if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1161     return nullptr;
1162 
1163   Type *XTy = X->getType();
1164   bool HadTrunc = I.getType() != XTy;
1165 
1166   // If there was a truncation of extracted value, then we'll need to produce
1167   // one extra instruction, so we need to ensure one instruction will go away.
1168   if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1169     return nullptr;
1170 
1171   // Extraction should extract high NBits bits, with shift amount calculated as:
1172   //   low bits to skip = shift bitwidth - high bits to extract
1173   // The shift amount itself may be extended, and we need to look past zero-ext
1174   // when matching NBits, that will matter for matching later.
1175   Constant *C;
1176   Value *NBits;
1177   if (!match(
1178           LowBitsToSkip,
1179           m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1180       !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1181                                    APInt(C->getType()->getScalarSizeInBits(),
1182                                          X->getType()->getScalarSizeInBits()))))
1183     return nullptr;
1184 
1185   // Sign-extending value can be zero-extended if we `sub`tract it,
1186   // or sign-extended otherwise.
1187   auto SkipExtInMagic = [&I](Value *&V) {
1188     if (I.getOpcode() == Instruction::Sub)
1189       match(V, m_ZExtOrSelf(m_Value(V)));
1190     else
1191       match(V, m_SExtOrSelf(m_Value(V)));
1192   };
1193 
1194   // Now, finally validate the sign-extending magic.
1195   // `select` itself may be appropriately extended, look past that.
1196   SkipExtInMagic(Select);
1197 
1198   ICmpInst::Predicate Pred;
1199   const APInt *Thr;
1200   Value *SignExtendingValue, *Zero;
1201   bool ShouldSignext;
1202   // It must be a select between two values we will later establish to be a
1203   // sign-extending value and a zero constant. The condition guarding the
1204   // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1205   if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1206                               m_Value(SignExtendingValue), m_Value(Zero))) ||
1207       !isSignBitCheck(Pred, *Thr, ShouldSignext))
1208     return nullptr;
1209 
1210   // icmp-select pair is commutative.
1211   if (!ShouldSignext)
1212     std::swap(SignExtendingValue, Zero);
1213 
1214   // If we should not perform sign-extension then we must add/or/subtract zero.
1215   if (!match(Zero, m_Zero()))
1216     return nullptr;
1217   // Otherwise, it should be some constant, left-shifted by the same NBits we
1218   // had in `lshr`. Said left-shift can also be appropriately extended.
1219   // Again, we must look past zero-ext when looking for NBits.
1220   SkipExtInMagic(SignExtendingValue);
1221   Constant *SignExtendingValueBaseConstant;
1222   if (!match(SignExtendingValue,
1223              m_Shl(m_Constant(SignExtendingValueBaseConstant),
1224                    m_ZExtOrSelf(m_Specific(NBits)))))
1225     return nullptr;
1226   // If we `sub`, then the constant should be one, else it should be all-ones.
1227   if (I.getOpcode() == Instruction::Sub
1228           ? !match(SignExtendingValueBaseConstant, m_One())
1229           : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1230     return nullptr;
1231 
1232   auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1233                                              Extract->getName() + ".sext");
1234   NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1235   if (!HadTrunc)
1236     return NewAShr;
1237 
1238   Builder.Insert(NewAShr);
1239   return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1240 }
1241 
1242 /// This is a specialization of a more general transform from
1243 /// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
1244 /// for multi-use cases or propagating nsw/nuw, then we would not need this.
1245 static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1246                                             InstCombiner::BuilderTy &Builder) {
1247   // TODO: Also handle mul by doubling the shift amount?
1248   assert((I.getOpcode() == Instruction::Add ||
1249           I.getOpcode() == Instruction::Sub) &&
1250          "Expected add/sub");
1251   auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1252   auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1253   if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1254     return nullptr;
1255 
1256   Value *X, *Y, *ShAmt;
1257   if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1258       !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1259     return nullptr;
1260 
1261   // No-wrap propagates only when all ops have no-wrap.
1262   bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1263                 Op1->hasNoSignedWrap();
1264   bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1265                 Op1->hasNoUnsignedWrap();
1266 
1267   // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1268   Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1269   if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1270     NewI->setHasNoSignedWrap(HasNSW);
1271     NewI->setHasNoUnsignedWrap(HasNUW);
1272   }
1273   auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1274   NewShl->setHasNoSignedWrap(HasNSW);
1275   NewShl->setHasNoUnsignedWrap(HasNUW);
1276   return NewShl;
1277 }
1278 
1279 Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1280   if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1281                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1282                                  SQ.getWithInstruction(&I)))
1283     return replaceInstUsesWith(I, V);
1284 
1285   if (SimplifyAssociativeOrCommutative(I))
1286     return &I;
1287 
1288   if (Instruction *X = foldVectorBinop(I))
1289     return X;
1290 
1291   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1292     return Phi;
1293 
1294   // (A*B)+(A*C) -> A*(B+C) etc
1295   if (Value *V = SimplifyUsingDistributiveLaws(I))
1296     return replaceInstUsesWith(I, V);
1297 
1298   if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1299     return R;
1300 
1301   if (Instruction *X = foldAddWithConstant(I))
1302     return X;
1303 
1304   if (Instruction *X = foldNoWrapAdd(I, Builder))
1305     return X;
1306 
1307   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1308   Type *Ty = I.getType();
1309   if (Ty->isIntOrIntVectorTy(1))
1310     return BinaryOperator::CreateXor(LHS, RHS);
1311 
1312   // X + X --> X << 1
1313   if (LHS == RHS) {
1314     auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1315     Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1316     Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1317     return Shl;
1318   }
1319 
1320   Value *A, *B;
1321   if (match(LHS, m_Neg(m_Value(A)))) {
1322     // -A + -B --> -(A + B)
1323     if (match(RHS, m_Neg(m_Value(B))))
1324       return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1325 
1326     // -A + B --> B - A
1327     return BinaryOperator::CreateSub(RHS, A);
1328   }
1329 
1330   // A + -B  -->  A - B
1331   if (match(RHS, m_Neg(m_Value(B))))
1332     return BinaryOperator::CreateSub(LHS, B);
1333 
1334   if (Value *V = checkForNegativeOperand(I, Builder))
1335     return replaceInstUsesWith(I, V);
1336 
1337   // (A + 1) + ~B --> A - B
1338   // ~B + (A + 1) --> A - B
1339   // (~B + A) + 1 --> A - B
1340   // (A + ~B) + 1 --> A - B
1341   if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1342       match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1343     return BinaryOperator::CreateSub(A, B);
1344 
1345   // (A + RHS) + RHS --> A + (RHS << 1)
1346   if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1347     return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1348 
1349   // LHS + (A + LHS) --> A + (LHS << 1)
1350   if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1351     return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1352 
1353   {
1354     // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1355     Constant *C1, *C2;
1356     if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1357                           m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1358         (LHS->hasOneUse() || RHS->hasOneUse())) {
1359       Value *Sub = Builder.CreateSub(A, B);
1360       return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1361     }
1362   }
1363 
1364   // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1365   if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1366 
1367   // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1368   const APInt *C1, *C2;
1369   if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1370     APInt one(C2->getBitWidth(), 1);
1371     APInt minusC1 = -(*C1);
1372     if (minusC1 == (one << *C2)) {
1373       Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1374       return BinaryOperator::CreateSRem(RHS, NewRHS);
1375     }
1376   }
1377 
1378   // A+B --> A|B iff A and B have no bits set in common.
1379   if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1380     return BinaryOperator::CreateOr(LHS, RHS);
1381 
1382   // add (select X 0 (sub n A)) A  -->  select X A n
1383   {
1384     SelectInst *SI = dyn_cast<SelectInst>(LHS);
1385     Value *A = RHS;
1386     if (!SI) {
1387       SI = dyn_cast<SelectInst>(RHS);
1388       A = LHS;
1389     }
1390     if (SI && SI->hasOneUse()) {
1391       Value *TV = SI->getTrueValue();
1392       Value *FV = SI->getFalseValue();
1393       Value *N;
1394 
1395       // Can we fold the add into the argument of the select?
1396       // We check both true and false select arguments for a matching subtract.
1397       if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1398         // Fold the add into the true select value.
1399         return SelectInst::Create(SI->getCondition(), N, A);
1400 
1401       if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1402         // Fold the add into the false select value.
1403         return SelectInst::Create(SI->getCondition(), A, N);
1404     }
1405   }
1406 
1407   if (Instruction *Ext = narrowMathIfNoOverflow(I))
1408     return Ext;
1409 
1410   // (add (xor A, B) (and A, B)) --> (or A, B)
1411   // (add (and A, B) (xor A, B)) --> (or A, B)
1412   if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1413                           m_c_And(m_Deferred(A), m_Deferred(B)))))
1414     return BinaryOperator::CreateOr(A, B);
1415 
1416   // (add (or A, B) (and A, B)) --> (add A, B)
1417   // (add (and A, B) (or A, B)) --> (add A, B)
1418   if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1419                           m_c_And(m_Deferred(A), m_Deferred(B))))) {
1420     // Replacing operands in-place to preserve nuw/nsw flags.
1421     replaceOperand(I, 0, A);
1422     replaceOperand(I, 1, B);
1423     return &I;
1424   }
1425 
1426   // TODO(jingyue): Consider willNotOverflowSignedAdd and
1427   // willNotOverflowUnsignedAdd to reduce the number of invocations of
1428   // computeKnownBits.
1429   bool Changed = false;
1430   if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1431     Changed = true;
1432     I.setHasNoSignedWrap(true);
1433   }
1434   if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1435     Changed = true;
1436     I.setHasNoUnsignedWrap(true);
1437   }
1438 
1439   if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1440     return V;
1441 
1442   if (Instruction *V =
1443           canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1444     return V;
1445 
1446   if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1447     return SatAdd;
1448 
1449   // usub.sat(A, B) + B => umax(A, B)
1450   if (match(&I, m_c_BinOp(
1451           m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1452           m_Deferred(B)))) {
1453     return replaceInstUsesWith(I,
1454         Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1455   }
1456 
1457   // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1458   if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1459       match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1460       haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1461     return replaceInstUsesWith(
1462         I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1463                                    {Builder.CreateOr(A, B)}));
1464 
1465   return Changed ? &I : nullptr;
1466 }
1467 
1468 /// Eliminate an op from a linear interpolation (lerp) pattern.
1469 static Instruction *factorizeLerp(BinaryOperator &I,
1470                                   InstCombiner::BuilderTy &Builder) {
1471   Value *X, *Y, *Z;
1472   if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1473                                             m_OneUse(m_FSub(m_FPOne(),
1474                                                             m_Value(Z))))),
1475                           m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1476     return nullptr;
1477 
1478   // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1479   Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1480   Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1481   return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1482 }
1483 
1484 /// Factor a common operand out of fadd/fsub of fmul/fdiv.
1485 static Instruction *factorizeFAddFSub(BinaryOperator &I,
1486                                       InstCombiner::BuilderTy &Builder) {
1487   assert((I.getOpcode() == Instruction::FAdd ||
1488           I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1489   assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1490          "FP factorization requires FMF");
1491 
1492   if (Instruction *Lerp = factorizeLerp(I, Builder))
1493     return Lerp;
1494 
1495   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1496   if (!Op0->hasOneUse() || !Op1->hasOneUse())
1497     return nullptr;
1498 
1499   Value *X, *Y, *Z;
1500   bool IsFMul;
1501   if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1502        match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1503       (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1504        match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1505     IsFMul = true;
1506   else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1507            match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1508     IsFMul = false;
1509   else
1510     return nullptr;
1511 
1512   // (X * Z) + (Y * Z) --> (X + Y) * Z
1513   // (X * Z) - (Y * Z) --> (X - Y) * Z
1514   // (X / Z) + (Y / Z) --> (X + Y) / Z
1515   // (X / Z) - (Y / Z) --> (X - Y) / Z
1516   bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1517   Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1518                      : Builder.CreateFSubFMF(X, Y, &I);
1519 
1520   // Bail out if we just created a denormal constant.
1521   // TODO: This is copied from a previous implementation. Is it necessary?
1522   const APFloat *C;
1523   if (match(XY, m_APFloat(C)) && !C->isNormal())
1524     return nullptr;
1525 
1526   return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1527                 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1528 }
1529 
1530 Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1531   if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1532                                   I.getFastMathFlags(),
1533                                   SQ.getWithInstruction(&I)))
1534     return replaceInstUsesWith(I, V);
1535 
1536   if (SimplifyAssociativeOrCommutative(I))
1537     return &I;
1538 
1539   if (Instruction *X = foldVectorBinop(I))
1540     return X;
1541 
1542   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1543     return Phi;
1544 
1545   if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1546     return FoldedFAdd;
1547 
1548   // (-X) + Y --> Y - X
1549   Value *X, *Y;
1550   if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1551     return BinaryOperator::CreateFSubFMF(Y, X, &I);
1552 
1553   // Similar to above, but look through fmul/fdiv for the negated term.
1554   // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1555   Value *Z;
1556   if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1557                          m_Value(Z)))) {
1558     Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1559     return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1560   }
1561   // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1562   // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1563   if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1564                          m_Value(Z))) ||
1565       match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1566                          m_Value(Z)))) {
1567     Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1568     return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1569   }
1570 
1571   // Check for (fadd double (sitofp x), y), see if we can merge this into an
1572   // integer add followed by a promotion.
1573   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1574   if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1575     Value *LHSIntVal = LHSConv->getOperand(0);
1576     Type *FPType = LHSConv->getType();
1577 
1578     // TODO: This check is overly conservative. In many cases known bits
1579     // analysis can tell us that the result of the addition has less significant
1580     // bits than the integer type can hold.
1581     auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1582       Type *FScalarTy = FTy->getScalarType();
1583       Type *IScalarTy = ITy->getScalarType();
1584 
1585       // Do we have enough bits in the significand to represent the result of
1586       // the integer addition?
1587       unsigned MaxRepresentableBits =
1588           APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1589       return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1590     };
1591 
1592     // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1593     // ... if the constant fits in the integer value.  This is useful for things
1594     // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1595     // requires a constant pool load, and generally allows the add to be better
1596     // instcombined.
1597     if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1598       if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1599         Constant *CI =
1600           ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1601         if (LHSConv->hasOneUse() &&
1602             ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1603             willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1604           // Insert the new integer add.
1605           Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1606           return new SIToFPInst(NewAdd, I.getType());
1607         }
1608       }
1609 
1610     // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1611     if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1612       Value *RHSIntVal = RHSConv->getOperand(0);
1613       // It's enough to check LHS types only because we require int types to
1614       // be the same for this transform.
1615       if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1616         // Only do this if x/y have the same type, if at least one of them has a
1617         // single use (so we don't increase the number of int->fp conversions),
1618         // and if the integer add will not overflow.
1619         if (LHSIntVal->getType() == RHSIntVal->getType() &&
1620             (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1621             willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1622           // Insert the new integer add.
1623           Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1624           return new SIToFPInst(NewAdd, I.getType());
1625         }
1626       }
1627     }
1628   }
1629 
1630   // Handle specials cases for FAdd with selects feeding the operation
1631   if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1632     return replaceInstUsesWith(I, V);
1633 
1634   if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1635     if (Instruction *F = factorizeFAddFSub(I, Builder))
1636       return F;
1637 
1638     // Try to fold fadd into start value of reduction intrinsic.
1639     if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1640                                m_AnyZeroFP(), m_Value(X))),
1641                            m_Value(Y)))) {
1642       // fadd (rdx 0.0, X), Y --> rdx Y, X
1643       return replaceInstUsesWith(
1644           I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1645                                      {X->getType()}, {Y, X}, &I));
1646     }
1647     const APFloat *StartC, *C;
1648     if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1649                        m_APFloat(StartC), m_Value(X)))) &&
1650         match(RHS, m_APFloat(C))) {
1651       // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1652       Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1653       return replaceInstUsesWith(
1654           I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1655                                      {X->getType()}, {NewStartC, X}, &I));
1656     }
1657 
1658     // (X * MulC) + X --> X * (MulC + 1.0)
1659     Constant *MulC;
1660     if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1661                            m_Deferred(X)))) {
1662       MulC = ConstantExpr::getFAdd(MulC, ConstantFP::get(I.getType(), 1.0));
1663       return BinaryOperator::CreateFMulFMF(X, MulC, &I);
1664     }
1665 
1666     if (Value *V = FAddCombine(Builder).simplify(&I))
1667       return replaceInstUsesWith(I, V);
1668   }
1669 
1670   return nullptr;
1671 }
1672 
1673 /// Optimize pointer differences into the same array into a size.  Consider:
1674 ///  &A[10] - &A[0]: we should compile this to "10".  LHS/RHS are the pointer
1675 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1676 Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
1677                                                    Type *Ty, bool IsNUW) {
1678   // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1679   // this.
1680   bool Swapped = false;
1681   GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1682   if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1683     std::swap(LHS, RHS);
1684     Swapped = true;
1685   }
1686 
1687   // Require at least one GEP with a common base pointer on both sides.
1688   if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1689     // (gep X, ...) - X
1690     if (LHSGEP->getOperand(0) == RHS) {
1691       GEP1 = LHSGEP;
1692     } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1693       // (gep X, ...) - (gep X, ...)
1694       if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1695           RHSGEP->getOperand(0)->stripPointerCasts()) {
1696         GEP1 = LHSGEP;
1697         GEP2 = RHSGEP;
1698       }
1699     }
1700   }
1701 
1702   if (!GEP1)
1703     return nullptr;
1704 
1705   if (GEP2) {
1706     // (gep X, ...) - (gep X, ...)
1707     //
1708     // Avoid duplicating the arithmetic if there are more than one non-constant
1709     // indices between the two GEPs and either GEP has a non-constant index and
1710     // multiple users. If zero non-constant index, the result is a constant and
1711     // there is no duplication. If one non-constant index, the result is an add
1712     // or sub with a constant, which is no larger than the original code, and
1713     // there's no duplicated arithmetic, even if either GEP has multiple
1714     // users. If more than one non-constant indices combined, as long as the GEP
1715     // with at least one non-constant index doesn't have multiple users, there
1716     // is no duplication.
1717     unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1718     unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1719     if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1720         ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1721          (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1722       return nullptr;
1723     }
1724   }
1725 
1726   // Emit the offset of the GEP and an intptr_t.
1727   Value *Result = EmitGEPOffset(GEP1);
1728 
1729   // If this is a single inbounds GEP and the original sub was nuw,
1730   // then the final multiplication is also nuw.
1731   if (auto *I = dyn_cast<Instruction>(Result))
1732     if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1733         I->getOpcode() == Instruction::Mul)
1734       I->setHasNoUnsignedWrap();
1735 
1736   // If we have a 2nd GEP of the same base pointer, subtract the offsets.
1737   // If both GEPs are inbounds, then the subtract does not have signed overflow.
1738   if (GEP2) {
1739     Value *Offset = EmitGEPOffset(GEP2);
1740     Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
1741                                GEP1->isInBounds() && GEP2->isInBounds());
1742   }
1743 
1744   // If we have p - gep(p, ...)  then we have to negate the result.
1745   if (Swapped)
1746     Result = Builder.CreateNeg(Result, "diff.neg");
1747 
1748   return Builder.CreateIntCast(Result, Ty, true);
1749 }
1750 
1751 Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
1752   if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1753                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1754                                  SQ.getWithInstruction(&I)))
1755     return replaceInstUsesWith(I, V);
1756 
1757   if (Instruction *X = foldVectorBinop(I))
1758     return X;
1759 
1760   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1761     return Phi;
1762 
1763   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1764 
1765   // If this is a 'B = x-(-A)', change to B = x+A.
1766   // We deal with this without involving Negator to preserve NSW flag.
1767   if (Value *V = dyn_castNegVal(Op1)) {
1768     BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1769 
1770     if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1771       assert(BO->getOpcode() == Instruction::Sub &&
1772              "Expected a subtraction operator!");
1773       if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1774         Res->setHasNoSignedWrap(true);
1775     } else {
1776       if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1777         Res->setHasNoSignedWrap(true);
1778     }
1779 
1780     return Res;
1781   }
1782 
1783   // Try this before Negator to preserve NSW flag.
1784   if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1785     return R;
1786 
1787   Constant *C;
1788   if (match(Op0, m_ImmConstant(C))) {
1789     Value *X;
1790     Constant *C2;
1791 
1792     // C-(X+C2) --> (C-C2)-X
1793     if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
1794       return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1795   }
1796 
1797   auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
1798     if (Instruction *Ext = narrowMathIfNoOverflow(I))
1799       return Ext;
1800 
1801     bool Changed = false;
1802     if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1803       Changed = true;
1804       I.setHasNoSignedWrap(true);
1805     }
1806     if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1807       Changed = true;
1808       I.setHasNoUnsignedWrap(true);
1809     }
1810 
1811     return Changed ? &I : nullptr;
1812   };
1813 
1814   // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
1815   // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
1816   // a pure negation used by a select that looks like abs/nabs.
1817   bool IsNegation = match(Op0, m_ZeroInt());
1818   if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
1819         const Instruction *UI = dyn_cast<Instruction>(U);
1820         if (!UI)
1821           return false;
1822         return match(UI,
1823                      m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
1824                match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
1825       })) {
1826     if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
1827       return BinaryOperator::CreateAdd(NegOp1, Op0);
1828   }
1829   if (IsNegation)
1830     return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
1831 
1832   // (A*B)-(A*C) -> A*(B-C) etc
1833   if (Value *V = SimplifyUsingDistributiveLaws(I))
1834     return replaceInstUsesWith(I, V);
1835 
1836   if (I.getType()->isIntOrIntVectorTy(1))
1837     return BinaryOperator::CreateXor(Op0, Op1);
1838 
1839   // Replace (-1 - A) with (~A).
1840   if (match(Op0, m_AllOnes()))
1841     return BinaryOperator::CreateNot(Op1);
1842 
1843   // (X + -1) - Y --> ~Y + X
1844   Value *X, *Y;
1845   if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1846     return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1847 
1848   // Reassociate sub/add sequences to create more add instructions and
1849   // reduce dependency chains:
1850   // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
1851   Value *Z;
1852   if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
1853                                   m_Value(Z))))) {
1854     Value *XZ = Builder.CreateAdd(X, Z);
1855     Value *YW = Builder.CreateAdd(Y, Op1);
1856     return BinaryOperator::CreateSub(XZ, YW);
1857   }
1858 
1859   // ((X - Y) - Op1)  -->  X - (Y + Op1)
1860   if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
1861     Value *Add = Builder.CreateAdd(Y, Op1);
1862     return BinaryOperator::CreateSub(X, Add);
1863   }
1864 
1865   // (~X) - (~Y) --> Y - X
1866   // This is placed after the other reassociations and explicitly excludes a
1867   // sub-of-sub pattern to avoid infinite looping.
1868   if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
1869       isFreeToInvert(Op1, Op1->hasOneUse()) &&
1870       !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
1871     Value *NotOp0 = Builder.CreateNot(Op0);
1872     Value *NotOp1 = Builder.CreateNot(Op1);
1873     return BinaryOperator::CreateSub(NotOp1, NotOp0);
1874   }
1875 
1876   auto m_AddRdx = [](Value *&Vec) {
1877     return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
1878   };
1879   Value *V0, *V1;
1880   if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
1881       V0->getType() == V1->getType()) {
1882     // Difference of sums is sum of differences:
1883     // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
1884     Value *Sub = Builder.CreateSub(V0, V1);
1885     Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
1886                                          {Sub->getType()}, {Sub});
1887     return replaceInstUsesWith(I, Rdx);
1888   }
1889 
1890   if (Constant *C = dyn_cast<Constant>(Op0)) {
1891     Value *X;
1892     if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1893       // C - (zext bool) --> bool ? C - 1 : C
1894       return SelectInst::Create(X, InstCombiner::SubOne(C), C);
1895     if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1896       // C - (sext bool) --> bool ? C + 1 : C
1897       return SelectInst::Create(X, InstCombiner::AddOne(C), C);
1898 
1899     // C - ~X == X + (1+C)
1900     if (match(Op1, m_Not(m_Value(X))))
1901       return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
1902 
1903     // Try to fold constant sub into select arguments.
1904     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1905       if (Instruction *R = FoldOpIntoSelect(I, SI))
1906         return R;
1907 
1908     // Try to fold constant sub into PHI values.
1909     if (PHINode *PN = dyn_cast<PHINode>(Op1))
1910       if (Instruction *R = foldOpIntoPhi(I, PN))
1911         return R;
1912 
1913     Constant *C2;
1914 
1915     // C-(C2-X) --> X+(C-C2)
1916     if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
1917       return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
1918   }
1919 
1920   const APInt *Op0C;
1921   if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) {
1922     // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1923     // zero.
1924     KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1925     if ((*Op0C | RHSKnown.Zero).isAllOnes())
1926       return BinaryOperator::CreateXor(Op1, Op0);
1927   }
1928 
1929   {
1930     Value *Y;
1931     // X-(X+Y) == -Y    X-(Y+X) == -Y
1932     if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1933       return BinaryOperator::CreateNeg(Y);
1934 
1935     // (X-Y)-X == -Y
1936     if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1937       return BinaryOperator::CreateNeg(Y);
1938   }
1939 
1940   // (sub (or A, B) (and A, B)) --> (xor A, B)
1941   {
1942     Value *A, *B;
1943     if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1944         match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1945       return BinaryOperator::CreateXor(A, B);
1946   }
1947 
1948   // (sub (add A, B) (or A, B)) --> (and A, B)
1949   {
1950     Value *A, *B;
1951     if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1952         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1953       return BinaryOperator::CreateAnd(A, B);
1954   }
1955 
1956   // (sub (add A, B) (and A, B)) --> (or A, B)
1957   {
1958     Value *A, *B;
1959     if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1960         match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
1961       return BinaryOperator::CreateOr(A, B);
1962   }
1963 
1964   // (sub (and A, B) (or A, B)) --> neg (xor A, B)
1965   {
1966     Value *A, *B;
1967     if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1968         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1969         (Op0->hasOneUse() || Op1->hasOneUse()))
1970       return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
1971   }
1972 
1973   // (sub (or A, B), (xor A, B)) --> (and A, B)
1974   {
1975     Value *A, *B;
1976     if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1977         match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1978       return BinaryOperator::CreateAnd(A, B);
1979   }
1980 
1981   // (sub (xor A, B) (or A, B)) --> neg (and A, B)
1982   {
1983     Value *A, *B;
1984     if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1985         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1986         (Op0->hasOneUse() || Op1->hasOneUse()))
1987       return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
1988   }
1989 
1990   {
1991     Value *Y;
1992     // ((X | Y) - X) --> (~X & Y)
1993     if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1994       return BinaryOperator::CreateAnd(
1995           Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1996   }
1997 
1998   {
1999     // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2000     Value *X;
2001     if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2002                                     m_OneUse(m_Neg(m_Value(X))))))) {
2003       return BinaryOperator::CreateNeg(Builder.CreateAnd(
2004           Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2005     }
2006   }
2007 
2008   {
2009     // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2010     Constant *C;
2011     if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2012       return BinaryOperator::CreateNeg(
2013           Builder.CreateAnd(Op1, Builder.CreateNot(C)));
2014     }
2015   }
2016 
2017   {
2018     // If we have a subtraction between some value and a select between
2019     // said value and something else, sink subtraction into select hands, i.e.:
2020     //   sub (select %Cond, %TrueVal, %FalseVal), %Op1
2021     //     ->
2022     //   select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2023     //  or
2024     //   sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2025     //     ->
2026     //   select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2027     // This will result in select between new subtraction and 0.
2028     auto SinkSubIntoSelect =
2029         [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2030                            auto SubBuilder) -> Instruction * {
2031       Value *Cond, *TrueVal, *FalseVal;
2032       if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2033                                            m_Value(FalseVal)))))
2034         return nullptr;
2035       if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2036         return nullptr;
2037       // While it is really tempting to just create two subtractions and let
2038       // InstCombine fold one of those to 0, it isn't possible to do so
2039       // because of worklist visitation order. So ugly it is.
2040       bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2041       Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2042       Constant *Zero = Constant::getNullValue(Ty);
2043       SelectInst *NewSel =
2044           SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2045                              OtherHandOfSubIsTrueVal ? NewSub : Zero);
2046       // Preserve prof metadata if any.
2047       NewSel->copyMetadata(cast<Instruction>(*Select));
2048       return NewSel;
2049     };
2050     if (Instruction *NewSel = SinkSubIntoSelect(
2051             /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2052             [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2053               return Builder->CreateSub(OtherHandOfSelect,
2054                                         /*OtherHandOfSub=*/Op1);
2055             }))
2056       return NewSel;
2057     if (Instruction *NewSel = SinkSubIntoSelect(
2058             /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2059             [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2060               return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2061                                         OtherHandOfSelect);
2062             }))
2063       return NewSel;
2064   }
2065 
2066   // (X - (X & Y))   -->   (X & ~Y)
2067   if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2068       (Op1->hasOneUse() || isa<Constant>(Y)))
2069     return BinaryOperator::CreateAnd(
2070         Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2071 
2072   // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2073   // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2074   // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2075   // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2076   // As long as Y is freely invertible, this will be neutral or a win.
2077   // Note: We don't generate the inverse max/min, just create the 'not' of
2078   // it and let other folds do the rest.
2079   if (match(Op0, m_Not(m_Value(X))) &&
2080       match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2081       !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2082     Value *Not = Builder.CreateNot(Op1);
2083     return BinaryOperator::CreateSub(Not, X);
2084   }
2085   if (match(Op1, m_Not(m_Value(X))) &&
2086       match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2087       !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2088     Value *Not = Builder.CreateNot(Op0);
2089     return BinaryOperator::CreateSub(X, Not);
2090   }
2091 
2092   // TODO: This is the same logic as above but handles the cmp-select idioms
2093   //       for min/max, so the use checks are increased to account for the
2094   //       extra instructions. If we canonicalize to intrinsics, this block
2095   //       can likely be removed.
2096   {
2097     Value *LHS, *RHS, *A;
2098     Value *NotA = Op0, *MinMax = Op1;
2099     SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2100     if (!SelectPatternResult::isMinOrMax(SPF)) {
2101       NotA = Op1;
2102       MinMax = Op0;
2103       SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2104     }
2105     if (SelectPatternResult::isMinOrMax(SPF) &&
2106         match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
2107       if (NotA == LHS)
2108         std::swap(LHS, RHS);
2109       // LHS is now Y above and expected to have at least 2 uses (the min/max)
2110       // NotA is expected to have 2 uses from the min/max and 1 from the sub.
2111       if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
2112           !NotA->hasNUsesOrMore(4)) {
2113         Value *Not = Builder.CreateNot(MinMax);
2114         if (NotA == Op0)
2115           return BinaryOperator::CreateSub(Not, A);
2116         else
2117           return BinaryOperator::CreateSub(A, Not);
2118       }
2119     }
2120   }
2121 
2122   // Optimize pointer differences into the same array into a size.  Consider:
2123   //  &A[10] - &A[0]: we should compile this to "10".
2124   Value *LHSOp, *RHSOp;
2125   if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2126       match(Op1, m_PtrToInt(m_Value(RHSOp))))
2127     if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2128                                                I.hasNoUnsignedWrap()))
2129       return replaceInstUsesWith(I, Res);
2130 
2131   // trunc(p)-trunc(q) -> trunc(p-q)
2132   if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2133       match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2134     if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2135                                                /* IsNUW */ false))
2136       return replaceInstUsesWith(I, Res);
2137 
2138   // Canonicalize a shifty way to code absolute value to the common pattern.
2139   // There are 2 potential commuted variants.
2140   // We're relying on the fact that we only do this transform when the shift has
2141   // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2142   // instructions).
2143   Value *A;
2144   const APInt *ShAmt;
2145   Type *Ty = I.getType();
2146   if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2147       Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2148       match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2149     // B = ashr i32 A, 31 ; smear the sign bit
2150     // sub (xor A, B), B  ; flip bits if negative and subtract -1 (add 1)
2151     // --> (A < 0) ? -A : A
2152     Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2153     // Copy the nuw/nsw flags from the sub to the negate.
2154     Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2155                                    I.hasNoSignedWrap());
2156     return SelectInst::Create(Cmp, Neg, A);
2157   }
2158 
2159   // If we are subtracting a low-bit masked subset of some value from an add
2160   // of that same value with no low bits changed, that is clearing some low bits
2161   // of the sum:
2162   // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2163   const APInt *AddC, *AndC;
2164   if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2165       match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2166     unsigned BitWidth = Ty->getScalarSizeInBits();
2167     unsigned Cttz = AddC->countTrailingZeros();
2168     APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2169     if ((HighMask & *AndC).isZero())
2170       return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2171   }
2172 
2173   if (Instruction *V =
2174           canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2175     return V;
2176 
2177   // X - usub.sat(X, Y) => umin(X, Y)
2178   if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2179                                                            m_Value(Y)))))
2180     return replaceInstUsesWith(
2181         I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2182 
2183   // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2184   // TODO: The one-use restriction is not strictly necessary, but it may
2185   //       require improving other pattern matching and/or codegen.
2186   if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2187     return replaceInstUsesWith(
2188         I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2189 
2190   // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2191   if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2192     Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2193     return BinaryOperator::CreateNeg(USub);
2194   }
2195 
2196   // C - ctpop(X) => ctpop(~X) if C is bitwidth
2197   if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
2198       match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2199     return replaceInstUsesWith(
2200         I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2201                                    {Builder.CreateNot(X)}));
2202 
2203   return TryToNarrowDeduceFlags();
2204 }
2205 
2206 /// This eliminates floating-point negation in either 'fneg(X)' or
2207 /// 'fsub(-0.0, X)' form by combining into a constant operand.
2208 static Instruction *foldFNegIntoConstant(Instruction &I) {
2209   // This is limited with one-use because fneg is assumed better for
2210   // reassociation and cheaper in codegen than fmul/fdiv.
2211   // TODO: Should the m_OneUse restriction be removed?
2212   Instruction *FNegOp;
2213   if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2214     return nullptr;
2215 
2216   Value *X;
2217   Constant *C;
2218 
2219   // Fold negation into constant operand.
2220   // -(X * C) --> X * (-C)
2221   if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2222     return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
2223   // -(X / C) --> X / (-C)
2224   if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2225     return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
2226   // -(C / X) --> (-C) / X
2227   if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) {
2228     Instruction *FDiv =
2229         BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
2230 
2231     // Intersect 'nsz' and 'ninf' because those special value exceptions may not
2232     // apply to the fdiv. Everything else propagates from the fneg.
2233     // TODO: We could propagate nsz/ninf from fdiv alone?
2234     FastMathFlags FMF = I.getFastMathFlags();
2235     FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2236     FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2237     FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2238     return FDiv;
2239   }
2240   // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2241   // -(X + C) --> -X + -C --> -C - X
2242   if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2243     return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I);
2244 
2245   return nullptr;
2246 }
2247 
2248 static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2249                                            InstCombiner::BuilderTy &Builder) {
2250   Value *FNeg;
2251   if (!match(&I, m_FNeg(m_Value(FNeg))))
2252     return nullptr;
2253 
2254   Value *X, *Y;
2255   if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2256     return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2257 
2258   if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2259     return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2260 
2261   return nullptr;
2262 }
2263 
2264 Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2265   Value *Op = I.getOperand(0);
2266 
2267   if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
2268                                   getSimplifyQuery().getWithInstruction(&I)))
2269     return replaceInstUsesWith(I, V);
2270 
2271   if (Instruction *X = foldFNegIntoConstant(I))
2272     return X;
2273 
2274   Value *X, *Y;
2275 
2276   // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2277   if (I.hasNoSignedZeros() &&
2278       match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2279     return BinaryOperator::CreateFSubFMF(Y, X, &I);
2280 
2281   if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2282     return R;
2283 
2284   // Try to eliminate fneg if at least 1 arm of the select is negated.
2285   Value *Cond;
2286   if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
2287     // Unlike most transforms, this one is not safe to propagate nsz unless
2288     // it is present on the original select. (We are conservatively intersecting
2289     // the nsz flags from the select and root fneg instruction.)
2290     auto propagateSelectFMF = [&](SelectInst *S) {
2291       S->copyFastMathFlags(&I);
2292       if (auto *OldSel = dyn_cast<SelectInst>(Op))
2293         if (!OldSel->hasNoSignedZeros())
2294           S->setHasNoSignedZeros(false);
2295     };
2296     // -(Cond ? -P : Y) --> Cond ? P : -Y
2297     Value *P;
2298     if (match(X, m_FNeg(m_Value(P)))) {
2299       Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2300       SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2301       propagateSelectFMF(NewSel);
2302       return NewSel;
2303     }
2304     // -(Cond ? X : -P) --> Cond ? -X : P
2305     if (match(Y, m_FNeg(m_Value(P)))) {
2306       Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2307       SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2308       propagateSelectFMF(NewSel);
2309       return NewSel;
2310     }
2311   }
2312 
2313   return nullptr;
2314 }
2315 
2316 Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2317   if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
2318                                   I.getFastMathFlags(),
2319                                   getSimplifyQuery().getWithInstruction(&I)))
2320     return replaceInstUsesWith(I, V);
2321 
2322   if (Instruction *X = foldVectorBinop(I))
2323     return X;
2324 
2325   if (Instruction *Phi = foldBinopWithPhiOperands(I))
2326     return Phi;
2327 
2328   // Subtraction from -0.0 is the canonical form of fneg.
2329   // fsub -0.0, X ==> fneg X
2330   // fsub nsz 0.0, X ==> fneg nsz X
2331   //
2332   // FIXME This matcher does not respect FTZ or DAZ yet:
2333   // fsub -0.0, Denorm ==> +-0
2334   // fneg Denorm ==> -Denorm
2335   Value *Op;
2336   if (match(&I, m_FNeg(m_Value(Op))))
2337     return UnaryOperator::CreateFNegFMF(Op, &I);
2338 
2339   if (Instruction *X = foldFNegIntoConstant(I))
2340     return X;
2341 
2342   if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2343     return R;
2344 
2345   Value *X, *Y;
2346   Constant *C;
2347 
2348   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2349   // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2350   // Canonicalize to fadd to make analysis easier.
2351   // This can also help codegen because fadd is commutative.
2352   // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2353   // killed later. We still limit that particular transform with 'hasOneUse'
2354   // because an fneg is assumed better/cheaper than a generic fsub.
2355   if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2356     if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2357       Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2358       return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2359     }
2360   }
2361 
2362   // (-X) - Op1 --> -(X + Op1)
2363   if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2364       match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2365     Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2366     return UnaryOperator::CreateFNegFMF(FAdd, &I);
2367   }
2368 
2369   if (isa<Constant>(Op0))
2370     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2371       if (Instruction *NV = FoldOpIntoSelect(I, SI))
2372         return NV;
2373 
2374   // X - C --> X + (-C)
2375   // But don't transform constant expressions because there's an inverse fold
2376   // for X + (-Y) --> X - Y.
2377   if (match(Op1, m_ImmConstant(C)))
2378     return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
2379 
2380   // X - (-Y) --> X + Y
2381   if (match(Op1, m_FNeg(m_Value(Y))))
2382     return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2383 
2384   // Similar to above, but look through a cast of the negated value:
2385   // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2386   Type *Ty = I.getType();
2387   if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2388     return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2389 
2390   // X - (fpext(-Y)) --> X + fpext(Y)
2391   if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2392     return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2393 
2394   // Similar to above, but look through fmul/fdiv of the negated value:
2395   // Op0 - (-X * Y) --> Op0 + (X * Y)
2396   // Op0 - (Y * -X) --> Op0 + (X * Y)
2397   if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2398     Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2399     return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2400   }
2401   // Op0 - (-X / Y) --> Op0 + (X / Y)
2402   // Op0 - (X / -Y) --> Op0 + (X / Y)
2403   if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2404       match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2405     Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2406     return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2407   }
2408 
2409   // Handle special cases for FSub with selects feeding the operation
2410   if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2411     return replaceInstUsesWith(I, V);
2412 
2413   if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2414     // (Y - X) - Y --> -X
2415     if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2416       return UnaryOperator::CreateFNegFMF(X, &I);
2417 
2418     // Y - (X + Y) --> -X
2419     // Y - (Y + X) --> -X
2420     if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2421       return UnaryOperator::CreateFNegFMF(X, &I);
2422 
2423     // (X * C) - X --> X * (C - 1.0)
2424     if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2425       Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
2426       return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2427     }
2428     // X - (X * C) --> X * (1.0 - C)
2429     if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2430       Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
2431       return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2432     }
2433 
2434     // Reassociate fsub/fadd sequences to create more fadd instructions and
2435     // reduce dependency chains:
2436     // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2437     Value *Z;
2438     if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2439                                      m_Value(Z))))) {
2440       Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2441       Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2442       return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2443     }
2444 
2445     auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2446       return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2447                                                                  m_Value(Vec)));
2448     };
2449     Value *A0, *A1, *V0, *V1;
2450     if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2451         V0->getType() == V1->getType()) {
2452       // Difference of sums is sum of differences:
2453       // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2454       Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2455       Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2456                                            {Sub->getType()}, {A0, Sub}, &I);
2457       return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2458     }
2459 
2460     if (Instruction *F = factorizeFAddFSub(I, Builder))
2461       return F;
2462 
2463     // TODO: This performs reassociative folds for FP ops. Some fraction of the
2464     // functionality has been subsumed by simple pattern matching here and in
2465     // InstSimplify. We should let a dedicated reassociation pass handle more
2466     // complex pattern matching and remove this from InstCombine.
2467     if (Value *V = FAddCombine(Builder).simplify(&I))
2468       return replaceInstUsesWith(I, V);
2469 
2470     // (X - Y) - Op1 --> X - (Y + Op1)
2471     if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2472       Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2473       return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
2474     }
2475   }
2476 
2477   return nullptr;
2478 }
2479