xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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 
set(short C)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 
isZero() const74     bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75     Value *getValue(Type *) const;
76 
isOne() const77     bool isOne() const { return isInt() && IntVal == 1; }
isTwo() const78     bool isTwo() const { return isInt() && IntVal == 2; }
isMinusOne() const79     bool isMinusOne() const { return isInt() && IntVal == -1; }
isMinusTwo() const80     bool isMinusTwo() const { return isInt() && IntVal == -2; }
81 
82   private:
insaneIntVal(int V)83     bool insaneIntVal(int V) { return V > 4 || V < -4; }
84 
getFpValPtr()85     APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86 
getFpValPtr() const87     const APFloat *getFpValPtr() const {
88       return reinterpret_cast<const APFloat *>(&FpValBuf);
89     }
90 
getFpVal() const91     const APFloat &getFpVal() const {
92       assert(IsFp && BufHasFpVal && "Incorret state");
93       return *getFpValPtr();
94     }
95 
getFpVal()96     APFloat &getFpVal() {
97       assert(IsFp && BufHasFpVal && "Incorret state");
98       return *getFpValPtr();
99     }
100 
isInt() const101     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 
operator +=(const FAddend & T)133     void operator+=(const FAddend &T) {
134       assert((Val == T.Val) && "Symbolic-values disagree");
135       Coeff += T.Coeff;
136     }
137 
getSymVal() const138     Value *getSymVal() const { return Val; }
getCoef() const139     const FAddendCoef &getCoef() const { return Coeff; }
140 
isConstant() const141     bool isConstant() const { return Val == nullptr; }
isZero() const142     bool isZero() const { return Coeff.isZero(); }
143 
set(short Coefficient,Value * V)144     void set(short Coefficient, Value *V) {
145       Coeff.set(Coefficient);
146       Val = V;
147     }
set(const APFloat & Coefficient,Value * V)148     void set(const APFloat &Coefficient, Value *V) {
149       Coeff.set(Coefficient);
150       Val = V;
151     }
set(const ConstantFP * Coefficient,Value * V)152     void set(const ConstantFP *Coefficient, Value *V) {
153       Coeff.set(Coefficient->getValueAPF());
154       Val = V;
155     }
156 
negate()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:
Scale(const FAddendCoef & ScaleAmt)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:
FAddCombine(InstCombiner::BuilderTy & B)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;
initCreateInstNum()205       void initCreateInstNum() { CreateInstrNum = 0; }
incCreateInstNum()206       void incCreateInstNum() { CreateInstrNum++; }
207     #else
initCreateInstNum()208       void initCreateInstNum() {}
incCreateInstNum()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 //===----------------------------------------------------------------------===//
~FAddendCoef()224 FAddendCoef::~FAddendCoef() {
225   if (BufHasFpVal)
226     getFpValPtr()->~APFloat();
227 }
228 
set(const APFloat & C)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 
convertToFpType(const fltSemantics & Sem)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 
createAPFloatFromInt(const fltSemantics & Sem,int Val)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 
operator =(const FAddendCoef & That)266 void FAddendCoef::operator=(const FAddendCoef &That) {
267   if (That.isInt())
268     set(That.IntVal);
269   else
270     set(That.getFpVal());
271 }
272 
operator +=(const FAddendCoef & That)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 
operator *=(const FAddendCoef & That)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 
negate()324 void FAddendCoef::negate() {
325   if (isInt())
326     IntVal = 0 - IntVal;
327   else
328     getFpVal().changeSign();
329 }
330 
getValue(Type * Ty) const331 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
drillValueDownOneStep(Value * Val,FAddend & Addend0,FAddend & Addend1)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>.
drillAddendDownOneStep(FAddend & Addend0,FAddend & Addend1) const410 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 
simplify(Instruction * I)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 
simplifyFAdd(AddendVect & Addends,unsigned InstrQuota)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 <= std::size(TmpResult) + 1) && "out-of-bound access");
580 
581   Value *Result;
582   if (!SimpVect.empty())
583     Result = createNaryFAdd(SimpVect, InstrQuota);
584   else {
585     // The addition is folded to 0.0.
586     Result = ConstantFP::get(Instr->getType(), 0.0);
587   }
588 
589   return Result;
590 }
591 
createNaryFAdd(const AddendVect & Opnds,unsigned InstrQuota)592 Value *FAddCombine::createNaryFAdd
593   (const AddendVect &Opnds, unsigned InstrQuota) {
594   assert(!Opnds.empty() && "Expect at least one addend");
595 
596   // Step 1: Check if the # of instructions needed exceeds the quota.
597 
598   unsigned InstrNeeded = calcInstrNumber(Opnds);
599   if (InstrNeeded > InstrQuota)
600     return nullptr;
601 
602   initCreateInstNum();
603 
604   // step 2: Emit the N-ary addition.
605   // Note that at most three instructions are involved in Fadd-InstCombine: the
606   // addition in question, and at most two neighboring instructions.
607   // The resulting optimized addition should have at least one less instruction
608   // than the original addition expression tree. This implies that the resulting
609   // N-ary addition has at most two instructions, and we don't need to worry
610   // about tree-height when constructing the N-ary addition.
611 
612   Value *LastVal = nullptr;
613   bool LastValNeedNeg = false;
614 
615   // Iterate the addends, creating fadd/fsub using adjacent two addends.
616   for (const FAddend *Opnd : Opnds) {
617     bool NeedNeg;
618     Value *V = createAddendVal(*Opnd, NeedNeg);
619     if (!LastVal) {
620       LastVal = V;
621       LastValNeedNeg = NeedNeg;
622       continue;
623     }
624 
625     if (LastValNeedNeg == NeedNeg) {
626       LastVal = createFAdd(LastVal, V);
627       continue;
628     }
629 
630     if (LastValNeedNeg)
631       LastVal = createFSub(V, LastVal);
632     else
633       LastVal = createFSub(LastVal, V);
634 
635     LastValNeedNeg = false;
636   }
637 
638   if (LastValNeedNeg) {
639     LastVal = createFNeg(LastVal);
640   }
641 
642 #ifndef NDEBUG
643   assert(CreateInstrNum == InstrNeeded &&
644          "Inconsistent in instruction numbers");
645 #endif
646 
647   return LastVal;
648 }
649 
createFSub(Value * Opnd0,Value * Opnd1)650 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
651   Value *V = Builder.CreateFSub(Opnd0, Opnd1);
652   if (Instruction *I = dyn_cast<Instruction>(V))
653     createInstPostProc(I);
654   return V;
655 }
656 
createFNeg(Value * V)657 Value *FAddCombine::createFNeg(Value *V) {
658   Value *NewV = Builder.CreateFNeg(V);
659   if (Instruction *I = dyn_cast<Instruction>(NewV))
660     createInstPostProc(I, true); // fneg's don't receive instruction numbers.
661   return NewV;
662 }
663 
createFAdd(Value * Opnd0,Value * Opnd1)664 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
665   Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
666   if (Instruction *I = dyn_cast<Instruction>(V))
667     createInstPostProc(I);
668   return V;
669 }
670 
createFMul(Value * Opnd0,Value * Opnd1)671 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672   Value *V = Builder.CreateFMul(Opnd0, Opnd1);
673   if (Instruction *I = dyn_cast<Instruction>(V))
674     createInstPostProc(I);
675   return V;
676 }
677 
createInstPostProc(Instruction * NewInstr,bool NoNumber)678 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
679   NewInstr->setDebugLoc(Instr->getDebugLoc());
680 
681   // Keep track of the number of instruction created.
682   if (!NoNumber)
683     incCreateInstNum();
684 
685   // Propagate fast-math flags
686   NewInstr->setFastMathFlags(Instr->getFastMathFlags());
687 }
688 
689 // Return the number of instruction needed to emit the N-ary addition.
690 // NOTE: Keep this function in sync with createAddendVal().
calcInstrNumber(const AddendVect & Opnds)691 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
692   unsigned OpndNum = Opnds.size();
693   unsigned InstrNeeded = OpndNum - 1;
694 
695   // Adjust the number of instructions needed to emit the N-ary add.
696   for (const FAddend *Opnd : Opnds) {
697     if (Opnd->isConstant())
698       continue;
699 
700     // The constant check above is really for a few special constant
701     // coefficients.
702     if (isa<UndefValue>(Opnd->getSymVal()))
703       continue;
704 
705     const FAddendCoef &CE = Opnd->getCoef();
706     // Let the addend be "c * x". If "c == +/-1", the value of the addend
707     // is immediately available; otherwise, it needs exactly one instruction
708     // to evaluate the value.
709     if (!CE.isMinusOne() && !CE.isOne())
710       InstrNeeded++;
711   }
712   return InstrNeeded;
713 }
714 
715 // Input Addend        Value           NeedNeg(output)
716 // ================================================================
717 // Constant C          C               false
718 // <+/-1, V>           V               coefficient is -1
719 // <2/-2, V>          "fadd V, V"      coefficient is -2
720 // <C, V>             "fmul V, C"      false
721 //
722 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
createAddendVal(const FAddend & Opnd,bool & NeedNeg)723 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
724   const FAddendCoef &Coeff = Opnd.getCoef();
725 
726   if (Opnd.isConstant()) {
727     NeedNeg = false;
728     return Coeff.getValue(Instr->getType());
729   }
730 
731   Value *OpndVal = Opnd.getSymVal();
732 
733   if (Coeff.isMinusOne() || Coeff.isOne()) {
734     NeedNeg = Coeff.isMinusOne();
735     return OpndVal;
736   }
737 
738   if (Coeff.isTwo() || Coeff.isMinusTwo()) {
739     NeedNeg = Coeff.isMinusTwo();
740     return createFAdd(OpndVal, OpndVal);
741   }
742 
743   NeedNeg = false;
744   return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
745 }
746 
747 // Checks if any operand is negative and we can convert add to sub.
748 // This function checks for following negative patterns
749 //   ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
750 //   ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
751 //   XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
checkForNegativeOperand(BinaryOperator & I,InstCombiner::BuilderTy & Builder)752 static Value *checkForNegativeOperand(BinaryOperator &I,
753                                       InstCombiner::BuilderTy &Builder) {
754   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
755 
756   // This function creates 2 instructions to replace ADD, we need at least one
757   // of LHS or RHS to have one use to ensure benefit in transform.
758   if (!LHS->hasOneUse() && !RHS->hasOneUse())
759     return nullptr;
760 
761   Value *X = nullptr, *Y = nullptr, *Z = nullptr;
762   const APInt *C1 = nullptr, *C2 = nullptr;
763 
764   // if ONE is on other side, swap
765   if (match(RHS, m_Add(m_Value(X), m_One())))
766     std::swap(LHS, RHS);
767 
768   if (match(LHS, m_Add(m_Value(X), m_One()))) {
769     // if XOR on other side, swap
770     if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
771       std::swap(X, RHS);
772 
773     if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
774       // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
775       // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
776       if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
777         Value *NewAnd = Builder.CreateAnd(Z, *C1);
778         return Builder.CreateSub(RHS, NewAnd, "sub");
779       } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
780         // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
781         // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
782         Value *NewOr = Builder.CreateOr(Z, ~(*C1));
783         return Builder.CreateSub(RHS, NewOr, "sub");
784       }
785     }
786   }
787 
788   // Restore LHS and RHS
789   LHS = I.getOperand(0);
790   RHS = I.getOperand(1);
791 
792   // if XOR is on other side, swap
793   if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
794     std::swap(LHS, RHS);
795 
796   // C2 is ODD
797   // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
798   // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
799   if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
800     if (C1->countr_zero() == 0)
801       if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
802         Value *NewOr = Builder.CreateOr(Z, ~(*C2));
803         return Builder.CreateSub(RHS, NewOr, "sub");
804       }
805   return nullptr;
806 }
807 
808 /// Wrapping flags may allow combining constants separated by an extend.
foldNoWrapAdd(BinaryOperator & Add,InstCombiner::BuilderTy & Builder)809 static Instruction *foldNoWrapAdd(BinaryOperator &Add,
810                                   InstCombiner::BuilderTy &Builder) {
811   Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
812   Type *Ty = Add.getType();
813   Constant *Op1C;
814   if (!match(Op1, m_Constant(Op1C)))
815     return nullptr;
816 
817   // Try this match first because it results in an add in the narrow type.
818   // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
819   Value *X;
820   const APInt *C1, *C2;
821   if (match(Op1, m_APInt(C1)) &&
822       match(Op0, m_ZExt(m_NUWAddLike(m_Value(X), m_APInt(C2)))) &&
823       C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824     APInt NewC = *C2 + C1->trunc(C2->getBitWidth());
825     // If the smaller add will fold to zero, we don't need to check one use.
826     if (NewC.isZero())
827       return new ZExtInst(X, Ty);
828     // Otherwise only do this if the existing zero extend will be removed.
829     if (Op0->hasOneUse())
830       return new ZExtInst(
831           Builder.CreateNUWAdd(X, ConstantInt::get(X->getType(), NewC)), Ty);
832   }
833 
834   // More general combining of constants in the wide type.
835   // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
836   // or (zext nneg (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
837   Constant *NarrowC;
838   if (match(Op0, m_OneUse(m_SExtLike(
839                      m_NSWAddLike(m_Value(X), m_Constant(NarrowC)))))) {
840     Value *WideC = Builder.CreateSExt(NarrowC, Ty);
841     Value *NewC = Builder.CreateAdd(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,
847             m_OneUse(m_ZExt(m_NUWAddLike(m_Value(X), m_Constant(NarrowC)))))) {
848     Value *WideC = Builder.CreateZExt(NarrowC, Ty);
849     Value *NewC = Builder.CreateAdd(WideC, Op1C);
850     Value *WideX = Builder.CreateZExt(X, Ty);
851     return BinaryOperator::CreateAdd(WideX, NewC);
852   }
853   return nullptr;
854 }
855 
foldAddWithConstant(BinaryOperator & Add)856 Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
857   Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
858   Type *Ty = Add.getType();
859   Constant *Op1C;
860   if (!match(Op1, m_ImmConstant(Op1C)))
861     return nullptr;
862 
863   if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
864     return NV;
865 
866   Value *X;
867   Constant *Op00C;
868 
869   // add (sub C1, X), C2 --> sub (add C1, C2), X
870   if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
871     return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
872 
873   Value *Y;
874 
875   // add (sub X, Y), -1 --> add (not Y), X
876   if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
877       match(Op1, m_AllOnes()))
878     return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
879 
880   // zext(bool) + C -> bool ? C + 1 : C
881   if (match(Op0, m_ZExt(m_Value(X))) &&
882       X->getType()->getScalarSizeInBits() == 1)
883     return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
884   // sext(bool) + C -> bool ? C - 1 : C
885   if (match(Op0, m_SExt(m_Value(X))) &&
886       X->getType()->getScalarSizeInBits() == 1)
887     return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
888 
889   // ~X + C --> (C-1) - X
890   if (match(Op0, m_Not(m_Value(X)))) {
891     // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
892     auto *COne = ConstantInt::get(Op1C->getType(), 1);
893     bool WillNotSOV = willNotOverflowSignedSub(Op1C, COne, Add);
894     BinaryOperator *Res =
895         BinaryOperator::CreateSub(ConstantExpr::getSub(Op1C, COne), X);
896     Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
897     return Res;
898   }
899 
900   // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
901   const APInt *C;
902   unsigned BitWidth = Ty->getScalarSizeInBits();
903   if (match(Op0, m_OneUse(m_AShr(m_Value(X),
904                                  m_SpecificIntAllowPoison(BitWidth - 1)))) &&
905       match(Op1, m_One()))
906     return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
907 
908   if (!match(Op1, m_APInt(C)))
909     return nullptr;
910 
911   // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
912   Constant *Op01C;
913   if (match(Op0, m_DisjointOr(m_Value(X), m_ImmConstant(Op01C)))) {
914     BinaryOperator *NewAdd =
915         BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
916     NewAdd->setHasNoSignedWrap(Add.hasNoSignedWrap() &&
917                                willNotOverflowSignedAdd(Op01C, Op1C, Add));
918     NewAdd->setHasNoUnsignedWrap(Add.hasNoUnsignedWrap());
919     return NewAdd;
920   }
921 
922   // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
923   const APInt *C2;
924   if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
925     return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
926 
927   if (C->isSignMask()) {
928     // If wrapping is not allowed, then the addition must set the sign bit:
929     // X + (signmask) --> X | signmask
930     if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
931       return BinaryOperator::CreateOr(Op0, Op1);
932 
933     // If wrapping is allowed, then the addition flips the sign bit of LHS:
934     // X + (signmask) --> X ^ signmask
935     return BinaryOperator::CreateXor(Op0, Op1);
936   }
937 
938   // Is this add the last step in a convoluted sext?
939   // add(zext(xor i16 X, -32768), -32768) --> sext X
940   if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
941       C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
942     return CastInst::Create(Instruction::SExt, X, Ty);
943 
944   if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
945     // (X ^ signmask) + C --> (X + (signmask ^ C))
946     if (C2->isSignMask())
947       return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
948 
949     // If X has no high-bits set above an xor mask:
950     // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
951     if (C2->isMask()) {
952       KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
953       if ((*C2 | LHSKnown.Zero).isAllOnes())
954         return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
955     }
956 
957     // Look for a math+logic pattern that corresponds to sext-in-register of a
958     // value with cleared high bits. Convert that into a pair of shifts:
959     // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
960     // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
961     if (Op0->hasOneUse() && *C2 == -(*C)) {
962       unsigned BitWidth = Ty->getScalarSizeInBits();
963       unsigned ShAmt = 0;
964       if (C->isPowerOf2())
965         ShAmt = BitWidth - C->logBase2() - 1;
966       else if (C2->isPowerOf2())
967         ShAmt = BitWidth - C2->logBase2() - 1;
968       if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
969                                      0, &Add)) {
970         Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
971         Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
972         return BinaryOperator::CreateAShr(NewShl, ShAmtC);
973       }
974     }
975   }
976 
977   if (C->isOne() && Op0->hasOneUse()) {
978     // add (sext i1 X), 1 --> zext (not X)
979     // TODO: The smallest IR representation is (select X, 0, 1), and that would
980     // not require the one-use check. But we need to remove a transform in
981     // visitSelect and make sure that IR value tracking for select is equal or
982     // better than for these ops.
983     if (match(Op0, m_SExt(m_Value(X))) &&
984         X->getType()->getScalarSizeInBits() == 1)
985       return new ZExtInst(Builder.CreateNot(X), Ty);
986 
987     // Shifts and add used to flip and mask off the low bit:
988     // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
989     const APInt *C3;
990     if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
991         C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
992       Value *NotX = Builder.CreateNot(X);
993       return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
994     }
995   }
996 
997   // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
998   // TODO: There's a general form for any constant on the outer add.
999   if (C->isOne()) {
1000     if (match(Op0, m_ZExt(m_Add(m_Value(X), m_AllOnes())))) {
1001       const SimplifyQuery Q = SQ.getWithInstruction(&Add);
1002       if (llvm::isKnownNonZero(X, Q))
1003         return new ZExtInst(X, Ty);
1004     }
1005   }
1006 
1007   return nullptr;
1008 }
1009 
1010 // match variations of a^2 + 2*a*b + b^2
1011 //
1012 // to reuse the code between the FP and Int versions, the instruction OpCodes
1013 //  and constant types have been turned into template parameters.
1014 //
1015 // Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1016 //  should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1017 //  (we're matching `X<<1` instead of `X*2` for Int)
1018 template <bool FP, typename Mul2Rhs>
matchesSquareSum(BinaryOperator & I,Mul2Rhs M2Rhs,Value * & A,Value * & B)1019 static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1020                              Value *&B) {
1021   constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1022   constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1023   constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1024 
1025   // (a * a) + (((a * 2) + b) * b)
1026   if (match(&I, m_c_BinOp(
1027                     AddOp, m_OneUse(m_BinOp(MulOp, m_Value(A), m_Deferred(A))),
1028                     m_OneUse(m_c_BinOp(
1029                         MulOp,
1030                         m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(A), M2Rhs),
1031                                   m_Value(B)),
1032                         m_Deferred(B))))))
1033     return true;
1034 
1035   // ((a * b) * 2)  or ((a * 2) * b)
1036   // +
1037   // (a * a + b * b) or (b * b + a * a)
1038   return match(
1039       &I, m_c_BinOp(
1040               AddOp,
1041               m_CombineOr(
1042                   m_OneUse(m_BinOp(
1043                       Mul2Op, m_BinOp(MulOp, m_Value(A), m_Value(B)), M2Rhs)),
1044                   m_OneUse(m_c_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(A), M2Rhs),
1045                                      m_Value(B)))),
1046               m_OneUse(
1047                   m_c_BinOp(AddOp, m_BinOp(MulOp, m_Deferred(A), m_Deferred(A)),
1048                             m_BinOp(MulOp, m_Deferred(B), m_Deferred(B))))));
1049 }
1050 
1051 // Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
foldSquareSumInt(BinaryOperator & I)1052 Instruction *InstCombinerImpl::foldSquareSumInt(BinaryOperator &I) {
1053   Value *A, *B;
1054   if (matchesSquareSum</*FP*/ false>(I, m_SpecificInt(1), A, B)) {
1055     Value *AB = Builder.CreateAdd(A, B);
1056     return BinaryOperator::CreateMul(AB, AB);
1057   }
1058   return nullptr;
1059 }
1060 
1061 // Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1062 // Requires `nsz` and `reassoc`.
foldSquareSumFP(BinaryOperator & I)1063 Instruction *InstCombinerImpl::foldSquareSumFP(BinaryOperator &I) {
1064   assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1065   Value *A, *B;
1066   if (matchesSquareSum</*FP*/ true>(I, m_SpecificFP(2.0), A, B)) {
1067     Value *AB = Builder.CreateFAddFMF(A, B, &I);
1068     return BinaryOperator::CreateFMulFMF(AB, AB, &I);
1069   }
1070   return nullptr;
1071 }
1072 
1073 // Matches multiplication expression Op * C where C is a constant. Returns the
1074 // constant value in C and the other operand in Op. Returns true if such a
1075 // match is found.
MatchMul(Value * E,Value * & Op,APInt & C)1076 static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1077   const APInt *AI;
1078   if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1079     C = *AI;
1080     return true;
1081   }
1082   if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1083     C = APInt(AI->getBitWidth(), 1);
1084     C <<= *AI;
1085     return true;
1086   }
1087   return false;
1088 }
1089 
1090 // Matches remainder expression Op % C where C is a constant. Returns the
1091 // constant value in C and the other operand in Op. Returns the signedness of
1092 // the remainder operation in IsSigned. Returns true if such a match is
1093 // found.
MatchRem(Value * E,Value * & Op,APInt & C,bool & IsSigned)1094 static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1095   const APInt *AI;
1096   IsSigned = false;
1097   if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1098     IsSigned = true;
1099     C = *AI;
1100     return true;
1101   }
1102   if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1103     C = *AI;
1104     return true;
1105   }
1106   if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1107     C = *AI + 1;
1108     return true;
1109   }
1110   return false;
1111 }
1112 
1113 // Matches division expression Op / C with the given signedness as indicated
1114 // by IsSigned, where C is a constant. Returns the constant value in C and the
1115 // other operand in Op. Returns true if such a match is found.
MatchDiv(Value * E,Value * & Op,APInt & C,bool IsSigned)1116 static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1117   const APInt *AI;
1118   if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1119     C = *AI;
1120     return true;
1121   }
1122   if (!IsSigned) {
1123     if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1124       C = *AI;
1125       return true;
1126     }
1127     if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1128       C = APInt(AI->getBitWidth(), 1);
1129       C <<= *AI;
1130       return true;
1131     }
1132   }
1133   return false;
1134 }
1135 
1136 // Returns whether C0 * C1 with the given signedness overflows.
MulWillOverflow(APInt & C0,APInt & C1,bool IsSigned)1137 static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1138   bool overflow;
1139   if (IsSigned)
1140     (void)C0.smul_ov(C1, overflow);
1141   else
1142     (void)C0.umul_ov(C1, overflow);
1143   return overflow;
1144 }
1145 
1146 // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1147 // does not overflow.
1148 // Simplifies (X / C0) * C1 + (X % C0) * C2 to
1149 // (X / C0) * (C1 - C2 * C0) + X * C2
SimplifyAddWithRemainder(BinaryOperator & I)1150 Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1151   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1152   Value *X, *MulOpV;
1153   APInt C0, MulOpC;
1154   bool IsSigned;
1155   // Match I = X % C0 + MulOpV * C0
1156   if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1157        (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1158       C0 == MulOpC) {
1159     Value *RemOpV;
1160     APInt C1;
1161     bool Rem2IsSigned;
1162     // Match MulOpC = RemOpV % C1
1163     if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1164         IsSigned == Rem2IsSigned) {
1165       Value *DivOpV;
1166       APInt DivOpC;
1167       // Match RemOpV = X / C0
1168       if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1169           C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1170         Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1171         return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1172                         : Builder.CreateURem(X, NewDivisor, "urem");
1173       }
1174     }
1175   }
1176 
1177   // Match I = (X / C0) * C1 + (X % C0) * C2
1178   Value *Div, *Rem;
1179   APInt C1, C2;
1180   if (!LHS->hasOneUse() || !MatchMul(LHS, Div, C1))
1181     Div = LHS, C1 = APInt(I.getType()->getScalarSizeInBits(), 1);
1182   if (!RHS->hasOneUse() || !MatchMul(RHS, Rem, C2))
1183     Rem = RHS, C2 = APInt(I.getType()->getScalarSizeInBits(), 1);
1184   if (match(Div, m_IRem(m_Value(), m_Value()))) {
1185     std::swap(Div, Rem);
1186     std::swap(C1, C2);
1187   }
1188   Value *DivOpV;
1189   APInt DivOpC;
1190   if (MatchRem(Rem, X, C0, IsSigned) &&
1191       MatchDiv(Div, DivOpV, DivOpC, IsSigned) && X == DivOpV && C0 == DivOpC) {
1192     APInt NewC = C1 - C2 * C0;
1193     if (!NewC.isZero() && !Rem->hasOneUse())
1194       return nullptr;
1195     if (!isGuaranteedNotToBeUndef(X, &AC, &I, &DT))
1196       return nullptr;
1197     Value *MulXC2 = Builder.CreateMul(X, ConstantInt::get(X->getType(), C2));
1198     if (NewC.isZero())
1199       return MulXC2;
1200     return Builder.CreateAdd(
1201         Builder.CreateMul(Div, ConstantInt::get(X->getType(), NewC)), MulXC2);
1202   }
1203 
1204   return nullptr;
1205 }
1206 
1207 /// Fold
1208 ///   (1 << NBits) - 1
1209 /// Into:
1210 ///   ~(-(1 << NBits))
1211 /// Because a 'not' is better for bit-tracking analysis and other transforms
1212 /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
canonicalizeLowbitMask(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1213 static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1214                                            InstCombiner::BuilderTy &Builder) {
1215   Value *NBits;
1216   if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1217     return nullptr;
1218 
1219   Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1220   Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1221   // Be wary of constant folding.
1222   if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1223     // Always NSW. But NUW propagates from `add`.
1224     BOp->setHasNoSignedWrap();
1225     BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1226   }
1227 
1228   return BinaryOperator::CreateNot(NotMask, I.getName());
1229 }
1230 
foldToUnsignedSaturatedAdd(BinaryOperator & I)1231 static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1232   assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1233   Type *Ty = I.getType();
1234   auto getUAddSat = [&]() {
1235     return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1236   };
1237 
1238   // add (umin X, ~Y), Y --> uaddsat X, Y
1239   Value *X, *Y;
1240   if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1241                         m_Deferred(Y))))
1242     return CallInst::Create(getUAddSat(), { X, Y });
1243 
1244   // add (umin X, ~C), C --> uaddsat X, C
1245   const APInt *C, *NotC;
1246   if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1247       *C == ~*NotC)
1248     return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1249 
1250   return nullptr;
1251 }
1252 
1253 // Transform:
1254 //  (add A, (shl (neg B), Y))
1255 //      -> (sub A, (shl B, Y))
combineAddSubWithShlAddSub(InstCombiner::BuilderTy & Builder,const BinaryOperator & I)1256 static Instruction *combineAddSubWithShlAddSub(InstCombiner::BuilderTy &Builder,
1257                                                const BinaryOperator &I) {
1258   Value *A, *B, *Cnt;
1259   if (match(&I,
1260             m_c_Add(m_OneUse(m_Shl(m_OneUse(m_Neg(m_Value(B))), m_Value(Cnt))),
1261                     m_Value(A)))) {
1262     Value *NewShl = Builder.CreateShl(B, Cnt);
1263     return BinaryOperator::CreateSub(A, NewShl);
1264   }
1265   return nullptr;
1266 }
1267 
1268 /// Try to reduce signed division by power-of-2 to an arithmetic shift right.
foldAddToAshr(BinaryOperator & Add)1269 static Instruction *foldAddToAshr(BinaryOperator &Add) {
1270   // Division must be by power-of-2, but not the minimum signed value.
1271   Value *X;
1272   const APInt *DivC;
1273   if (!match(Add.getOperand(0), m_SDiv(m_Value(X), m_Power2(DivC))) ||
1274       DivC->isNegative())
1275     return nullptr;
1276 
1277   // Rounding is done by adding -1 if the dividend (X) is negative and has any
1278   // low bits set. It recognizes two canonical patterns:
1279   // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1280   //    pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1281   // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1282   // Note that, by the time we end up here, if possible, ugt has been
1283   // canonicalized into eq.
1284   const APInt *MaskC, *MaskCCmp;
1285   ICmpInst::Predicate Pred;
1286   if (!match(Add.getOperand(1),
1287              m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1288                            m_APInt(MaskCCmp)))))
1289     return nullptr;
1290 
1291   if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1292       (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1293     return nullptr;
1294 
1295   APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1296   bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1297                          ? (*MaskC == (SMin | (*DivC - 1)))
1298                          : (*DivC == 2 && *MaskC == SMin + 1);
1299   if (!IsMaskValid)
1300     return nullptr;
1301 
1302   // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1303   return BinaryOperator::CreateAShr(
1304       X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1305 }
1306 
1307 Instruction *InstCombinerImpl::
canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(BinaryOperator & I)1308     canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1309         BinaryOperator &I) {
1310   assert((I.getOpcode() == Instruction::Add ||
1311           I.getOpcode() == Instruction::Or ||
1312           I.getOpcode() == Instruction::Sub) &&
1313          "Expecting add/or/sub instruction");
1314 
1315   // We have a subtraction/addition between a (potentially truncated) *logical*
1316   // right-shift of X and a "select".
1317   Value *X, *Select;
1318   Instruction *LowBitsToSkip, *Extract;
1319   if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd(
1320                                m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1321                                m_Instruction(Extract))),
1322                            m_Value(Select))))
1323     return nullptr;
1324 
1325   // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1326   if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1327     return nullptr;
1328 
1329   Type *XTy = X->getType();
1330   bool HadTrunc = I.getType() != XTy;
1331 
1332   // If there was a truncation of extracted value, then we'll need to produce
1333   // one extra instruction, so we need to ensure one instruction will go away.
1334   if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1335     return nullptr;
1336 
1337   // Extraction should extract high NBits bits, with shift amount calculated as:
1338   //   low bits to skip = shift bitwidth - high bits to extract
1339   // The shift amount itself may be extended, and we need to look past zero-ext
1340   // when matching NBits, that will matter for matching later.
1341   Constant *C;
1342   Value *NBits;
1343   if (!match(
1344           LowBitsToSkip,
1345           m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1346       !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1347                                    APInt(C->getType()->getScalarSizeInBits(),
1348                                          X->getType()->getScalarSizeInBits()))))
1349     return nullptr;
1350 
1351   // Sign-extending value can be zero-extended if we `sub`tract it,
1352   // or sign-extended otherwise.
1353   auto SkipExtInMagic = [&I](Value *&V) {
1354     if (I.getOpcode() == Instruction::Sub)
1355       match(V, m_ZExtOrSelf(m_Value(V)));
1356     else
1357       match(V, m_SExtOrSelf(m_Value(V)));
1358   };
1359 
1360   // Now, finally validate the sign-extending magic.
1361   // `select` itself may be appropriately extended, look past that.
1362   SkipExtInMagic(Select);
1363 
1364   ICmpInst::Predicate Pred;
1365   const APInt *Thr;
1366   Value *SignExtendingValue, *Zero;
1367   bool ShouldSignext;
1368   // It must be a select between two values we will later establish to be a
1369   // sign-extending value and a zero constant. The condition guarding the
1370   // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1371   if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1372                               m_Value(SignExtendingValue), m_Value(Zero))) ||
1373       !isSignBitCheck(Pred, *Thr, ShouldSignext))
1374     return nullptr;
1375 
1376   // icmp-select pair is commutative.
1377   if (!ShouldSignext)
1378     std::swap(SignExtendingValue, Zero);
1379 
1380   // If we should not perform sign-extension then we must add/or/subtract zero.
1381   if (!match(Zero, m_Zero()))
1382     return nullptr;
1383   // Otherwise, it should be some constant, left-shifted by the same NBits we
1384   // had in `lshr`. Said left-shift can also be appropriately extended.
1385   // Again, we must look past zero-ext when looking for NBits.
1386   SkipExtInMagic(SignExtendingValue);
1387   Constant *SignExtendingValueBaseConstant;
1388   if (!match(SignExtendingValue,
1389              m_Shl(m_Constant(SignExtendingValueBaseConstant),
1390                    m_ZExtOrSelf(m_Specific(NBits)))))
1391     return nullptr;
1392   // If we `sub`, then the constant should be one, else it should be all-ones.
1393   if (I.getOpcode() == Instruction::Sub
1394           ? !match(SignExtendingValueBaseConstant, m_One())
1395           : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1396     return nullptr;
1397 
1398   auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1399                                              Extract->getName() + ".sext");
1400   NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1401   if (!HadTrunc)
1402     return NewAShr;
1403 
1404   Builder.Insert(NewAShr);
1405   return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1406 }
1407 
1408 /// This is a specialization of a more general transform from
1409 /// foldUsingDistributiveLaws. If that code can be made to work optimally
1410 /// for multi-use cases or propagating nsw/nuw, then we would not need this.
factorizeMathWithShlOps(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1411 static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1412                                             InstCombiner::BuilderTy &Builder) {
1413   // TODO: Also handle mul by doubling the shift amount?
1414   assert((I.getOpcode() == Instruction::Add ||
1415           I.getOpcode() == Instruction::Sub) &&
1416          "Expected add/sub");
1417   auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1418   auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1419   if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1420     return nullptr;
1421 
1422   Value *X, *Y, *ShAmt;
1423   if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1424       !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1425     return nullptr;
1426 
1427   // No-wrap propagates only when all ops have no-wrap.
1428   bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1429                 Op1->hasNoSignedWrap();
1430   bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1431                 Op1->hasNoUnsignedWrap();
1432 
1433   // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1434   Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1435   if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1436     NewI->setHasNoSignedWrap(HasNSW);
1437     NewI->setHasNoUnsignedWrap(HasNUW);
1438   }
1439   auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1440   NewShl->setHasNoSignedWrap(HasNSW);
1441   NewShl->setHasNoUnsignedWrap(HasNUW);
1442   return NewShl;
1443 }
1444 
1445 /// Reduce a sequence of masked half-width multiplies to a single multiply.
1446 /// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
foldBoxMultiply(BinaryOperator & I)1447 static Instruction *foldBoxMultiply(BinaryOperator &I) {
1448   unsigned BitWidth = I.getType()->getScalarSizeInBits();
1449   // Skip the odd bitwidth types.
1450   if ((BitWidth & 0x1))
1451     return nullptr;
1452 
1453   unsigned HalfBits = BitWidth >> 1;
1454   APInt HalfMask = APInt::getMaxValue(HalfBits);
1455 
1456   // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1457   Value *XLo, *YLo;
1458   Value *CrossSum;
1459   // Require one-use on the multiply to avoid increasing the number of
1460   // multiplications.
1461   if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1462                          m_OneUse(m_Mul(m_Value(YLo), m_Value(XLo))))))
1463     return nullptr;
1464 
1465   // XLo = X & HalfMask
1466   // YLo = Y & HalfMask
1467   // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1468   // to enhance robustness
1469   Value *X, *Y;
1470   if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1471       !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1472     return nullptr;
1473 
1474   // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1475   // X' can be either X or XLo in the pattern (and the same for Y')
1476   if (match(CrossSum,
1477             m_c_Add(m_c_Mul(m_LShr(m_Specific(Y), m_SpecificInt(HalfBits)),
1478                             m_CombineOr(m_Specific(X), m_Specific(XLo))),
1479                     m_c_Mul(m_LShr(m_Specific(X), m_SpecificInt(HalfBits)),
1480                             m_CombineOr(m_Specific(Y), m_Specific(YLo))))))
1481     return BinaryOperator::CreateMul(X, Y);
1482 
1483   return nullptr;
1484 }
1485 
visitAdd(BinaryOperator & I)1486 Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1487   if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1488                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1489                                  SQ.getWithInstruction(&I)))
1490     return replaceInstUsesWith(I, V);
1491 
1492   if (SimplifyAssociativeOrCommutative(I))
1493     return &I;
1494 
1495   if (Instruction *X = foldVectorBinop(I))
1496     return X;
1497 
1498   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1499     return Phi;
1500 
1501   // (A*B)+(A*C) -> A*(B+C) etc
1502   if (Value *V = foldUsingDistributiveLaws(I))
1503     return replaceInstUsesWith(I, V);
1504 
1505   if (Instruction *R = foldBoxMultiply(I))
1506     return R;
1507 
1508   if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1509     return R;
1510 
1511   if (Instruction *X = foldAddWithConstant(I))
1512     return X;
1513 
1514   if (Instruction *X = foldNoWrapAdd(I, Builder))
1515     return X;
1516 
1517   if (Instruction *R = foldBinOpShiftWithShift(I))
1518     return R;
1519 
1520   if (Instruction *R = combineAddSubWithShlAddSub(Builder, I))
1521     return R;
1522 
1523   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1524   Type *Ty = I.getType();
1525   if (Ty->isIntOrIntVectorTy(1))
1526     return BinaryOperator::CreateXor(LHS, RHS);
1527 
1528   // X + X --> X << 1
1529   if (LHS == RHS) {
1530     auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1531     Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1532     Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1533     return Shl;
1534   }
1535 
1536   Value *A, *B;
1537   if (match(LHS, m_Neg(m_Value(A)))) {
1538     // -A + -B --> -(A + B)
1539     if (match(RHS, m_Neg(m_Value(B))))
1540       return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1541 
1542     // -A + B --> B - A
1543     auto *Sub = BinaryOperator::CreateSub(RHS, A);
1544     auto *OB0 = cast<OverflowingBinaryOperator>(LHS);
1545     Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1546 
1547     return Sub;
1548   }
1549 
1550   // A + -B  -->  A - B
1551   if (match(RHS, m_Neg(m_Value(B))))
1552     return BinaryOperator::CreateSub(LHS, B);
1553 
1554   if (Value *V = checkForNegativeOperand(I, Builder))
1555     return replaceInstUsesWith(I, V);
1556 
1557   // (A + 1) + ~B --> A - B
1558   // ~B + (A + 1) --> A - B
1559   // (~B + A) + 1 --> A - B
1560   // (A + ~B) + 1 --> A - B
1561   if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1562       match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1563     return BinaryOperator::CreateSub(A, B);
1564 
1565   // (A + RHS) + RHS --> A + (RHS << 1)
1566   if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1567     return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1568 
1569   // LHS + (A + LHS) --> A + (LHS << 1)
1570   if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1571     return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1572 
1573   {
1574     // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1575     Constant *C1, *C2;
1576     if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1577                           m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1578         (LHS->hasOneUse() || RHS->hasOneUse())) {
1579       Value *Sub = Builder.CreateSub(A, B);
1580       return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1581     }
1582 
1583     // Canonicalize a constant sub operand as an add operand for better folding:
1584     // (C1 - A) + B --> (B - A) + C1
1585     if (match(&I, m_c_Add(m_OneUse(m_Sub(m_ImmConstant(C1), m_Value(A))),
1586                           m_Value(B)))) {
1587       Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1588       return BinaryOperator::CreateAdd(Sub, C1);
1589     }
1590   }
1591 
1592   // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1593   if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1594 
1595   // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1596   const APInt *C1, *C2;
1597   if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1598     APInt one(C2->getBitWidth(), 1);
1599     APInt minusC1 = -(*C1);
1600     if (minusC1 == (one << *C2)) {
1601       Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1602       return BinaryOperator::CreateSRem(RHS, NewRHS);
1603     }
1604   }
1605 
1606   // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1607   if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1608       C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1609     Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1610     return BinaryOperator::CreateAnd(A, NewMask);
1611   }
1612 
1613   // ZExt (B - A) + ZExt(A) --> ZExt(B)
1614   if ((match(RHS, m_ZExt(m_Value(A))) &&
1615        match(LHS, m_ZExt(m_NUWSub(m_Value(B), m_Specific(A))))) ||
1616       (match(LHS, m_ZExt(m_Value(A))) &&
1617        match(RHS, m_ZExt(m_NUWSub(m_Value(B), m_Specific(A))))))
1618     return new ZExtInst(B, LHS->getType());
1619 
1620   // zext(A) + sext(A) --> 0 if A is i1
1621   if (match(&I, m_c_BinOp(m_ZExt(m_Value(A)), m_SExt(m_Deferred(A)))) &&
1622       A->getType()->isIntOrIntVectorTy(1))
1623     return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1624 
1625   // A+B --> A|B iff A and B have no bits set in common.
1626   WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1627   if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(&I)))
1628     return BinaryOperator::CreateDisjointOr(LHS, RHS);
1629 
1630   if (Instruction *Ext = narrowMathIfNoOverflow(I))
1631     return Ext;
1632 
1633   // (add (xor A, B) (and A, B)) --> (or A, B)
1634   // (add (and A, B) (xor A, B)) --> (or A, B)
1635   if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1636                           m_c_And(m_Deferred(A), m_Deferred(B)))))
1637     return BinaryOperator::CreateOr(A, B);
1638 
1639   // (add (or A, B) (and A, B)) --> (add A, B)
1640   // (add (and A, B) (or A, B)) --> (add A, B)
1641   if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1642                           m_c_And(m_Deferred(A), m_Deferred(B))))) {
1643     // Replacing operands in-place to preserve nuw/nsw flags.
1644     replaceOperand(I, 0, A);
1645     replaceOperand(I, 1, B);
1646     return &I;
1647   }
1648 
1649   // (add A (or A, -A)) --> (and (add A, -1) A)
1650   // (add A (or -A, A)) --> (and (add A, -1) A)
1651   // (add (or A, -A) A) --> (and (add A, -1) A)
1652   // (add (or -A, A) A) --> (and (add A, -1) A)
1653   if (match(&I, m_c_BinOp(m_Value(A), m_OneUse(m_c_Or(m_Neg(m_Deferred(A)),
1654                                                       m_Deferred(A)))))) {
1655     Value *Add =
1656         Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()), "",
1657                           I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1658     return BinaryOperator::CreateAnd(Add, A);
1659   }
1660 
1661   // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1662   // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1663   if (match(&I,
1664             m_Add(m_OneUse(m_c_And(m_Value(A), m_OneUse(m_Neg(m_Deferred(A))))),
1665                   m_AllOnes()))) {
1666     Constant *AllOnes = ConstantInt::getAllOnesValue(RHS->getType());
1667     Value *Dec = Builder.CreateAdd(A, AllOnes);
1668     Value *Not = Builder.CreateXor(A, AllOnes);
1669     return BinaryOperator::CreateAnd(Dec, Not);
1670   }
1671 
1672   // Disguised reassociation/factorization:
1673   // ~(A * C1) + A
1674   // ((A * -C1) - 1) + A
1675   // ((A * -C1) + A) - 1
1676   // (A * (1 - C1)) - 1
1677   if (match(&I,
1678             m_c_Add(m_OneUse(m_Not(m_OneUse(m_Mul(m_Value(A), m_APInt(C1))))),
1679                     m_Deferred(A)))) {
1680     Type *Ty = I.getType();
1681     Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1682     Value *NewMul = Builder.CreateMul(A, NewMulC);
1683     return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1684   }
1685 
1686   // (A * -2**C) + B --> B - (A << C)
1687   const APInt *NegPow2C;
1688   if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1689                         m_Value(B)))) {
1690     Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1691     Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1692     return BinaryOperator::CreateSub(B, Shl);
1693   }
1694 
1695   // Canonicalize signum variant that ends in add:
1696   // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1697   ICmpInst::Predicate Pred;
1698   uint64_t BitWidth = Ty->getScalarSizeInBits();
1699   if (match(LHS, m_AShr(m_Value(A), m_SpecificIntAllowPoison(BitWidth - 1))) &&
1700       match(RHS, m_OneUse(m_ZExt(
1701                      m_OneUse(m_ICmp(Pred, m_Specific(A), m_ZeroInt()))))) &&
1702       Pred == CmpInst::ICMP_SGT) {
1703     Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1704     Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1705     return BinaryOperator::CreateOr(LHS, Zext);
1706   }
1707 
1708   {
1709     Value *Cond, *Ext;
1710     Constant *C;
1711     // (add X, (sext/zext (icmp eq X, C)))
1712     //    -> (select (icmp eq X, C), (add C, (sext/zext 1)), X)
1713     auto CondMatcher = m_CombineAnd(
1714         m_Value(Cond), m_ICmp(Pred, m_Deferred(A), m_ImmConstant(C)));
1715 
1716     if (match(&I,
1717               m_c_Add(m_Value(A),
1718                       m_CombineAnd(m_Value(Ext), m_ZExtOrSExt(CondMatcher)))) &&
1719         Pred == ICmpInst::ICMP_EQ && Ext->hasOneUse()) {
1720       Value *Add = isa<ZExtInst>(Ext) ? InstCombiner::AddOne(C)
1721                                       : InstCombiner::SubOne(C);
1722       return replaceInstUsesWith(I, Builder.CreateSelect(Cond, Add, A));
1723     }
1724   }
1725 
1726   if (Instruction *Ashr = foldAddToAshr(I))
1727     return Ashr;
1728 
1729   // (~X) + (~Y) --> -2 - (X + Y)
1730   {
1731     // To ensure we can save instructions we need to ensure that we consume both
1732     // LHS/RHS (i.e they have a `not`).
1733     bool ConsumesLHS, ConsumesRHS;
1734     if (isFreeToInvert(LHS, LHS->hasOneUse(), ConsumesLHS) && ConsumesLHS &&
1735         isFreeToInvert(RHS, RHS->hasOneUse(), ConsumesRHS) && ConsumesRHS) {
1736       Value *NotLHS = getFreelyInverted(LHS, LHS->hasOneUse(), &Builder);
1737       Value *NotRHS = getFreelyInverted(RHS, RHS->hasOneUse(), &Builder);
1738       assert(NotLHS != nullptr && NotRHS != nullptr &&
1739              "isFreeToInvert desynced with getFreelyInverted");
1740       Value *LHSPlusRHS = Builder.CreateAdd(NotLHS, NotRHS);
1741       return BinaryOperator::CreateSub(
1742           ConstantInt::getSigned(RHS->getType(), -2), LHSPlusRHS);
1743     }
1744   }
1745 
1746   if (Instruction *R = tryFoldInstWithCtpopWithNot(&I))
1747     return R;
1748 
1749   // TODO(jingyue): Consider willNotOverflowSignedAdd and
1750   // willNotOverflowUnsignedAdd to reduce the number of invocations of
1751   // computeKnownBits.
1752   bool Changed = false;
1753   if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHSCache, RHSCache, I)) {
1754     Changed = true;
1755     I.setHasNoSignedWrap(true);
1756   }
1757   if (!I.hasNoUnsignedWrap() &&
1758       willNotOverflowUnsignedAdd(LHSCache, RHSCache, I)) {
1759     Changed = true;
1760     I.setHasNoUnsignedWrap(true);
1761   }
1762 
1763   if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1764     return V;
1765 
1766   if (Instruction *V =
1767           canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1768     return V;
1769 
1770   if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1771     return SatAdd;
1772 
1773   // usub.sat(A, B) + B => umax(A, B)
1774   if (match(&I, m_c_BinOp(
1775           m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1776           m_Deferred(B)))) {
1777     return replaceInstUsesWith(I,
1778         Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1779   }
1780 
1781   // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1782   if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1783       match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1784       haveNoCommonBitsSet(A, B, SQ.getWithInstruction(&I)))
1785     return replaceInstUsesWith(
1786         I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1787                                    {Builder.CreateOr(A, B)}));
1788 
1789   // Fold the log2_ceil idiom:
1790   // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
1791   // -->
1792   // BW - ctlz(A - 1, false)
1793   const APInt *XorC;
1794   if (match(&I,
1795             m_c_Add(
1796                 m_ZExt(m_ICmp(Pred, m_Intrinsic<Intrinsic::ctpop>(m_Value(A)),
1797                               m_One())),
1798                 m_OneUse(m_ZExtOrSelf(m_OneUse(m_Xor(
1799                     m_OneUse(m_TruncOrSelf(m_OneUse(
1800                         m_Intrinsic<Intrinsic::ctlz>(m_Deferred(A), m_One())))),
1801                     m_APInt(XorC))))))) &&
1802       (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
1803       *XorC == A->getType()->getScalarSizeInBits() - 1) {
1804     Value *Sub = Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()));
1805     Value *Ctlz = Builder.CreateIntrinsic(Intrinsic::ctlz, {A->getType()},
1806                                           {Sub, Builder.getFalse()});
1807     Value *Ret = Builder.CreateSub(
1808         ConstantInt::get(A->getType(), A->getType()->getScalarSizeInBits()),
1809         Ctlz, "", /*HasNUW*/ true, /*HasNSW*/ true);
1810     return replaceInstUsesWith(I, Builder.CreateZExtOrTrunc(Ret, I.getType()));
1811   }
1812 
1813   if (Instruction *Res = foldSquareSumInt(I))
1814     return Res;
1815 
1816   if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1817     return Res;
1818 
1819   if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
1820     return Res;
1821 
1822   return Changed ? &I : nullptr;
1823 }
1824 
1825 /// Eliminate an op from a linear interpolation (lerp) pattern.
factorizeLerp(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1826 static Instruction *factorizeLerp(BinaryOperator &I,
1827                                   InstCombiner::BuilderTy &Builder) {
1828   Value *X, *Y, *Z;
1829   if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1830                                             m_OneUse(m_FSub(m_FPOne(),
1831                                                             m_Value(Z))))),
1832                           m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1833     return nullptr;
1834 
1835   // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1836   Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1837   Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1838   return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1839 }
1840 
1841 /// Factor a common operand out of fadd/fsub of fmul/fdiv.
factorizeFAddFSub(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1842 static Instruction *factorizeFAddFSub(BinaryOperator &I,
1843                                       InstCombiner::BuilderTy &Builder) {
1844   assert((I.getOpcode() == Instruction::FAdd ||
1845           I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1846   assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1847          "FP factorization requires FMF");
1848 
1849   if (Instruction *Lerp = factorizeLerp(I, Builder))
1850     return Lerp;
1851 
1852   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1853   if (!Op0->hasOneUse() || !Op1->hasOneUse())
1854     return nullptr;
1855 
1856   Value *X, *Y, *Z;
1857   bool IsFMul;
1858   if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1859        match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1860       (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1861        match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1862     IsFMul = true;
1863   else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1864            match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1865     IsFMul = false;
1866   else
1867     return nullptr;
1868 
1869   // (X * Z) + (Y * Z) --> (X + Y) * Z
1870   // (X * Z) - (Y * Z) --> (X - Y) * Z
1871   // (X / Z) + (Y / Z) --> (X + Y) / Z
1872   // (X / Z) - (Y / Z) --> (X - Y) / Z
1873   bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1874   Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1875                      : Builder.CreateFSubFMF(X, Y, &I);
1876 
1877   // Bail out if we just created a denormal constant.
1878   // TODO: This is copied from a previous implementation. Is it necessary?
1879   const APFloat *C;
1880   if (match(XY, m_APFloat(C)) && !C->isNormal())
1881     return nullptr;
1882 
1883   return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1884                 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1885 }
1886 
visitFAdd(BinaryOperator & I)1887 Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1888   if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1889                                   I.getFastMathFlags(),
1890                                   SQ.getWithInstruction(&I)))
1891     return replaceInstUsesWith(I, V);
1892 
1893   if (SimplifyAssociativeOrCommutative(I))
1894     return &I;
1895 
1896   if (Instruction *X = foldVectorBinop(I))
1897     return X;
1898 
1899   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1900     return Phi;
1901 
1902   if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1903     return FoldedFAdd;
1904 
1905   // (-X) + Y --> Y - X
1906   Value *X, *Y;
1907   if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1908     return BinaryOperator::CreateFSubFMF(Y, X, &I);
1909 
1910   // Similar to above, but look through fmul/fdiv for the negated term.
1911   // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1912   Value *Z;
1913   if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1914                          m_Value(Z)))) {
1915     Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1916     return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1917   }
1918   // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1919   // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1920   if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1921                          m_Value(Z))) ||
1922       match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1923                          m_Value(Z)))) {
1924     Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1925     return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1926   }
1927 
1928   // Check for (fadd double (sitofp x), y), see if we can merge this into an
1929   // integer add followed by a promotion.
1930   if (Instruction *R = foldFBinOpOfIntCasts(I))
1931     return R;
1932 
1933   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1934   // Handle specials cases for FAdd with selects feeding the operation
1935   if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1936     return replaceInstUsesWith(I, V);
1937 
1938   if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1939     if (Instruction *F = factorizeFAddFSub(I, Builder))
1940       return F;
1941 
1942     if (Instruction *F = foldSquareSumFP(I))
1943       return F;
1944 
1945     // Try to fold fadd into start value of reduction intrinsic.
1946     if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1947                                m_AnyZeroFP(), m_Value(X))),
1948                            m_Value(Y)))) {
1949       // fadd (rdx 0.0, X), Y --> rdx Y, X
1950       return replaceInstUsesWith(
1951           I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1952                                      {X->getType()}, {Y, X}, &I));
1953     }
1954     const APFloat *StartC, *C;
1955     if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1956                        m_APFloat(StartC), m_Value(X)))) &&
1957         match(RHS, m_APFloat(C))) {
1958       // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1959       Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1960       return replaceInstUsesWith(
1961           I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1962                                      {X->getType()}, {NewStartC, X}, &I));
1963     }
1964 
1965     // (X * MulC) + X --> X * (MulC + 1.0)
1966     Constant *MulC;
1967     if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1968                            m_Deferred(X)))) {
1969       if (Constant *NewMulC = ConstantFoldBinaryOpOperands(
1970               Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1971         return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1972     }
1973 
1974     // (-X - Y) + (X + Z) --> Z - Y
1975     if (match(&I, m_c_FAdd(m_FSub(m_FNeg(m_Value(X)), m_Value(Y)),
1976                            m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1977       return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1978 
1979     if (Value *V = FAddCombine(Builder).simplify(&I))
1980       return replaceInstUsesWith(I, V);
1981   }
1982 
1983   // minumum(X, Y) + maximum(X, Y) => X + Y.
1984   if (match(&I,
1985             m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1986                      m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1987                                                        m_Deferred(Y))))) {
1988     BinaryOperator *Result = BinaryOperator::CreateFAddFMF(X, Y, &I);
1989     // We cannot preserve ninf if nnan flag is not set.
1990     // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1991     // while in optimized version NaN + Inf and this is a poison with ninf flag.
1992     if (!Result->hasNoNaNs())
1993       Result->setHasNoInfs(false);
1994     return Result;
1995   }
1996 
1997   return nullptr;
1998 }
1999 
2000 /// Optimize pointer differences into the same array into a size.  Consider:
2001 ///  &A[10] - &A[0]: we should compile this to "10".  LHS/RHS are the pointer
2002 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
OptimizePointerDifference(Value * LHS,Value * RHS,Type * Ty,bool IsNUW)2003 Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
2004                                                    Type *Ty, bool IsNUW) {
2005   // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
2006   // this.
2007   bool Swapped = false;
2008   GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
2009   if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
2010     std::swap(LHS, RHS);
2011     Swapped = true;
2012   }
2013 
2014   // Require at least one GEP with a common base pointer on both sides.
2015   if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
2016     // (gep X, ...) - X
2017     if (LHSGEP->getOperand(0)->stripPointerCasts() ==
2018         RHS->stripPointerCasts()) {
2019       GEP1 = LHSGEP;
2020     } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
2021       // (gep X, ...) - (gep X, ...)
2022       if (LHSGEP->getOperand(0)->stripPointerCasts() ==
2023           RHSGEP->getOperand(0)->stripPointerCasts()) {
2024         GEP1 = LHSGEP;
2025         GEP2 = RHSGEP;
2026       }
2027     }
2028   }
2029 
2030   if (!GEP1)
2031     return nullptr;
2032 
2033   // To avoid duplicating the offset arithmetic, rewrite the GEP to use the
2034   // computed offset. This may erase the original GEP, so be sure to cache the
2035   // inbounds flag before emitting the offset.
2036   // TODO: We should probably do this even if there is only one GEP.
2037   bool RewriteGEPs = GEP2 != nullptr;
2038 
2039   // Emit the offset of the GEP and an intptr_t.
2040   bool GEP1IsInBounds = GEP1->isInBounds();
2041   Value *Result = EmitGEPOffset(GEP1, RewriteGEPs);
2042 
2043   // If this is a single inbounds GEP and the original sub was nuw,
2044   // then the final multiplication is also nuw.
2045   if (auto *I = dyn_cast<Instruction>(Result))
2046     if (IsNUW && !GEP2 && !Swapped && GEP1IsInBounds &&
2047         I->getOpcode() == Instruction::Mul)
2048       I->setHasNoUnsignedWrap();
2049 
2050   // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2051   // If both GEPs are inbounds, then the subtract does not have signed overflow.
2052   if (GEP2) {
2053     bool GEP2IsInBounds = GEP2->isInBounds();
2054     Value *Offset = EmitGEPOffset(GEP2, RewriteGEPs);
2055     Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2056                                GEP1IsInBounds && GEP2IsInBounds);
2057   }
2058 
2059   // If we have p - gep(p, ...)  then we have to negate the result.
2060   if (Swapped)
2061     Result = Builder.CreateNeg(Result, "diff.neg");
2062 
2063   return Builder.CreateIntCast(Result, Ty, true);
2064 }
2065 
foldSubOfMinMax(BinaryOperator & I,InstCombiner::BuilderTy & Builder)2066 static Instruction *foldSubOfMinMax(BinaryOperator &I,
2067                                     InstCombiner::BuilderTy &Builder) {
2068   Value *Op0 = I.getOperand(0);
2069   Value *Op1 = I.getOperand(1);
2070   Type *Ty = I.getType();
2071   auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2072   if (!MinMax)
2073     return nullptr;
2074 
2075   // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2076   // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2077   Value *X = MinMax->getLHS();
2078   Value *Y = MinMax->getRHS();
2079   if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2080       (Op0->hasOneUse() || Op1->hasOneUse())) {
2081     Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2082     Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2083     return CallInst::Create(F, {X, Y});
2084   }
2085 
2086   // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2087   // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2088   Value *Z;
2089   if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2090     if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2091       Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2092       return BinaryOperator::CreateAdd(X, USub);
2093     }
2094     if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2095       Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2096       return BinaryOperator::CreateAdd(X, USub);
2097     }
2098   }
2099 
2100   // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2101   // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2102   if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2103       match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2104     Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2105     Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2106     return CallInst::Create(F, {Op0, Z});
2107   }
2108 
2109   return nullptr;
2110 }
2111 
visitSub(BinaryOperator & I)2112 Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
2113   if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2114                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2115                                  SQ.getWithInstruction(&I)))
2116     return replaceInstUsesWith(I, V);
2117 
2118   if (Instruction *X = foldVectorBinop(I))
2119     return X;
2120 
2121   if (Instruction *Phi = foldBinopWithPhiOperands(I))
2122     return Phi;
2123 
2124   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2125 
2126   // If this is a 'B = x-(-A)', change to B = x+A.
2127   // We deal with this without involving Negator to preserve NSW flag.
2128   if (Value *V = dyn_castNegVal(Op1)) {
2129     BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2130 
2131     if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2132       assert(BO->getOpcode() == Instruction::Sub &&
2133              "Expected a subtraction operator!");
2134       if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2135         Res->setHasNoSignedWrap(true);
2136     } else {
2137       if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2138         Res->setHasNoSignedWrap(true);
2139     }
2140 
2141     return Res;
2142   }
2143 
2144   // Try this before Negator to preserve NSW flag.
2145   if (Instruction *R = factorizeMathWithShlOps(I, Builder))
2146     return R;
2147 
2148   Constant *C;
2149   if (match(Op0, m_ImmConstant(C))) {
2150     Value *X;
2151     Constant *C2;
2152 
2153     // C-(X+C2) --> (C-C2)-X
2154     if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2155       // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2156       // => (C-C2)-X can have NSW/NUW
2157       bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2158       BinaryOperator *Res =
2159           BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2160       auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2161       Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2162                               WillNotSOV);
2163       Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2164                                 OBO1->hasNoUnsignedWrap());
2165       return Res;
2166     }
2167   }
2168 
2169   auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2170     if (Instruction *Ext = narrowMathIfNoOverflow(I))
2171       return Ext;
2172 
2173     bool Changed = false;
2174     if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2175       Changed = true;
2176       I.setHasNoSignedWrap(true);
2177     }
2178     if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2179       Changed = true;
2180       I.setHasNoUnsignedWrap(true);
2181     }
2182 
2183     return Changed ? &I : nullptr;
2184   };
2185 
2186   // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2187   // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2188   // a pure negation used by a select that looks like abs/nabs.
2189   bool IsNegation = match(Op0, m_ZeroInt());
2190   if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2191         const Instruction *UI = dyn_cast<Instruction>(U);
2192         if (!UI)
2193           return false;
2194         return match(UI,
2195                      m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2196                match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2197       })) {
2198     if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2199                                                         I.hasNoSignedWrap(),
2200                                         Op1, *this))
2201       return BinaryOperator::CreateAdd(NegOp1, Op0);
2202   }
2203   if (IsNegation)
2204     return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2205 
2206   // (A*B)-(A*C) -> A*(B-C) etc
2207   if (Value *V = foldUsingDistributiveLaws(I))
2208     return replaceInstUsesWith(I, V);
2209 
2210   if (I.getType()->isIntOrIntVectorTy(1))
2211     return BinaryOperator::CreateXor(Op0, Op1);
2212 
2213   // Replace (-1 - A) with (~A).
2214   if (match(Op0, m_AllOnes()))
2215     return BinaryOperator::CreateNot(Op1);
2216 
2217   // (X + -1) - Y --> ~Y + X
2218   Value *X, *Y;
2219   if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2220     return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2221 
2222   // Reassociate sub/add sequences to create more add instructions and
2223   // reduce dependency chains:
2224   // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2225   Value *Z;
2226   if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
2227                                   m_Value(Z))))) {
2228     Value *XZ = Builder.CreateAdd(X, Z);
2229     Value *YW = Builder.CreateAdd(Y, Op1);
2230     return BinaryOperator::CreateSub(XZ, YW);
2231   }
2232 
2233   // ((X - Y) - Op1)  -->  X - (Y + Op1)
2234   if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2235     OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Op0);
2236     bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2237     bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2238     Value *Add = Builder.CreateAdd(Y, Op1, "", /* HasNUW */ HasNUW,
2239                                    /* HasNSW */ HasNSW);
2240     BinaryOperator *Sub = BinaryOperator::CreateSub(X, Add);
2241     Sub->setHasNoUnsignedWrap(HasNUW);
2242     Sub->setHasNoSignedWrap(HasNSW);
2243     return Sub;
2244   }
2245 
2246   {
2247     // (X + Z) - (Y + Z) --> (X - Y)
2248     // This is done in other passes, but we want to be able to consume this
2249     // pattern in InstCombine so we can generate it without creating infinite
2250     // loops.
2251     if (match(Op0, m_Add(m_Value(X), m_Value(Z))) &&
2252         match(Op1, m_c_Add(m_Value(Y), m_Specific(Z))))
2253       return BinaryOperator::CreateSub(X, Y);
2254 
2255     // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2256     Constant *CX, *CY;
2257     if (match(Op0, m_OneUse(m_Add(m_Value(X), m_ImmConstant(CX)))) &&
2258         match(Op1, m_OneUse(m_Add(m_Value(Y), m_ImmConstant(CY))))) {
2259       Value *OpsSub = Builder.CreateSub(X, Y);
2260       Constant *ConstsSub = ConstantExpr::getSub(CX, CY);
2261       return BinaryOperator::CreateAdd(OpsSub, ConstsSub);
2262     }
2263   }
2264 
2265   // (~X) - (~Y) --> Y - X
2266   {
2267     // Need to ensure we can consume at least one of the `not` instructions,
2268     // otherwise this can inf loop.
2269     bool ConsumesOp0, ConsumesOp1;
2270     if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
2271         isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
2272         (ConsumesOp0 || ConsumesOp1)) {
2273       Value *NotOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
2274       Value *NotOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
2275       assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2276              "isFreeToInvert desynced with getFreelyInverted");
2277       return BinaryOperator::CreateSub(NotOp1, NotOp0);
2278     }
2279   }
2280 
2281   auto m_AddRdx = [](Value *&Vec) {
2282     return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2283   };
2284   Value *V0, *V1;
2285   if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2286       V0->getType() == V1->getType()) {
2287     // Difference of sums is sum of differences:
2288     // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2289     Value *Sub = Builder.CreateSub(V0, V1);
2290     Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2291                                          {Sub->getType()}, {Sub});
2292     return replaceInstUsesWith(I, Rdx);
2293   }
2294 
2295   if (Constant *C = dyn_cast<Constant>(Op0)) {
2296     Value *X;
2297     if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2298       // C - (zext bool) --> bool ? C - 1 : C
2299       return SelectInst::Create(X, InstCombiner::SubOne(C), C);
2300     if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2301       // C - (sext bool) --> bool ? C + 1 : C
2302       return SelectInst::Create(X, InstCombiner::AddOne(C), C);
2303 
2304     // C - ~X == X + (1+C)
2305     if (match(Op1, m_Not(m_Value(X))))
2306       return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2307 
2308     // Try to fold constant sub into select arguments.
2309     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2310       if (Instruction *R = FoldOpIntoSelect(I, SI))
2311         return R;
2312 
2313     // Try to fold constant sub into PHI values.
2314     if (PHINode *PN = dyn_cast<PHINode>(Op1))
2315       if (Instruction *R = foldOpIntoPhi(I, PN))
2316         return R;
2317 
2318     Constant *C2;
2319 
2320     // C-(C2-X) --> X+(C-C2)
2321     if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2322       return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2323   }
2324 
2325   const APInt *Op0C;
2326   if (match(Op0, m_APInt(Op0C))) {
2327     if (Op0C->isMask()) {
2328       // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2329       // zero. We don't use information from dominating conditions so this
2330       // transform is easier to reverse if necessary.
2331       KnownBits RHSKnown = llvm::computeKnownBits(
2332           Op1, 0, SQ.getWithInstruction(&I).getWithoutDomCondCache());
2333       if ((*Op0C | RHSKnown.Zero).isAllOnes())
2334         return BinaryOperator::CreateXor(Op1, Op0);
2335     }
2336 
2337     // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2338     // (C3 - ((C2 & C3) - 1)) is pow2
2339     // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2340     // C2 is negative pow2 || sub nuw
2341     const APInt *C2, *C3;
2342     BinaryOperator *InnerSub;
2343     if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2344         match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2345         (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2346       APInt C2AndC3 = *C2 & *C3;
2347       APInt C2AndC3Minus1 = C2AndC3 - 1;
2348       APInt C2AddC3 = *C2 + *C3;
2349       if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2350           C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2351         Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2352         return BinaryOperator::CreateAdd(
2353             And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2354       }
2355     }
2356   }
2357 
2358   {
2359     Value *Y;
2360     // X-(X+Y) == -Y    X-(Y+X) == -Y
2361     if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2362       return BinaryOperator::CreateNeg(Y);
2363 
2364     // (X-Y)-X == -Y
2365     if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2366       return BinaryOperator::CreateNeg(Y);
2367   }
2368 
2369   // (sub (or A, B) (and A, B)) --> (xor A, B)
2370   {
2371     Value *A, *B;
2372     if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2373         match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2374       return BinaryOperator::CreateXor(A, B);
2375   }
2376 
2377   // (sub (add A, B) (or A, B)) --> (and A, B)
2378   {
2379     Value *A, *B;
2380     if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2381         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2382       return BinaryOperator::CreateAnd(A, B);
2383   }
2384 
2385   // (sub (add A, B) (and A, B)) --> (or A, B)
2386   {
2387     Value *A, *B;
2388     if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2389         match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
2390       return BinaryOperator::CreateOr(A, B);
2391   }
2392 
2393   // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2394   {
2395     Value *A, *B;
2396     if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2397         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2398         (Op0->hasOneUse() || Op1->hasOneUse()))
2399       return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
2400   }
2401 
2402   // (sub (or A, B), (xor A, B)) --> (and A, B)
2403   {
2404     Value *A, *B;
2405     if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2406         match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2407       return BinaryOperator::CreateAnd(A, B);
2408   }
2409 
2410   // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2411   {
2412     Value *A, *B;
2413     if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2414         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2415         (Op0->hasOneUse() || Op1->hasOneUse()))
2416       return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
2417   }
2418 
2419   {
2420     Value *Y;
2421     // ((X | Y) - X) --> (~X & Y)
2422     if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2423       return BinaryOperator::CreateAnd(
2424           Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2425   }
2426 
2427   {
2428     // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2429     Value *X;
2430     if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2431                                     m_OneUse(m_Neg(m_Value(X))))))) {
2432       return BinaryOperator::CreateNeg(Builder.CreateAnd(
2433           Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2434     }
2435   }
2436 
2437   {
2438     // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2439     Constant *C;
2440     if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2441       return BinaryOperator::CreateNeg(
2442           Builder.CreateAnd(Op1, Builder.CreateNot(C)));
2443     }
2444   }
2445 
2446   {
2447     // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2448     // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2449     Value *C, *X;
2450     auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2451       return match(LHS, m_OneUse(m_c_Xor(m_Value(X), m_Specific(RHS)))) &&
2452              match(RHS, m_SExt(m_Value(C))) &&
2453              (C->getType()->getScalarSizeInBits() == 1);
2454     };
2455     if (m_SubXorCmp(Op0, Op1))
2456       return SelectInst::Create(C, Builder.CreateNeg(X), X);
2457     if (m_SubXorCmp(Op1, Op0))
2458       return SelectInst::Create(C, X, Builder.CreateNeg(X));
2459   }
2460 
2461   if (Instruction *R = tryFoldInstWithCtpopWithNot(&I))
2462     return R;
2463 
2464   if (Instruction *R = foldSubOfMinMax(I, Builder))
2465     return R;
2466 
2467   {
2468     // If we have a subtraction between some value and a select between
2469     // said value and something else, sink subtraction into select hands, i.e.:
2470     //   sub (select %Cond, %TrueVal, %FalseVal), %Op1
2471     //     ->
2472     //   select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2473     //  or
2474     //   sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2475     //     ->
2476     //   select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2477     // This will result in select between new subtraction and 0.
2478     auto SinkSubIntoSelect =
2479         [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2480                            auto SubBuilder) -> Instruction * {
2481       Value *Cond, *TrueVal, *FalseVal;
2482       if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2483                                            m_Value(FalseVal)))))
2484         return nullptr;
2485       if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2486         return nullptr;
2487       // While it is really tempting to just create two subtractions and let
2488       // InstCombine fold one of those to 0, it isn't possible to do so
2489       // because of worklist visitation order. So ugly it is.
2490       bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2491       Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2492       Constant *Zero = Constant::getNullValue(Ty);
2493       SelectInst *NewSel =
2494           SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2495                              OtherHandOfSubIsTrueVal ? NewSub : Zero);
2496       // Preserve prof metadata if any.
2497       NewSel->copyMetadata(cast<Instruction>(*Select));
2498       return NewSel;
2499     };
2500     if (Instruction *NewSel = SinkSubIntoSelect(
2501             /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2502             [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2503               return Builder->CreateSub(OtherHandOfSelect,
2504                                         /*OtherHandOfSub=*/Op1);
2505             }))
2506       return NewSel;
2507     if (Instruction *NewSel = SinkSubIntoSelect(
2508             /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2509             [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2510               return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2511                                         OtherHandOfSelect);
2512             }))
2513       return NewSel;
2514   }
2515 
2516   // (X - (X & Y))   -->   (X & ~Y)
2517   if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2518       (Op1->hasOneUse() || isa<Constant>(Y)))
2519     return BinaryOperator::CreateAnd(
2520         Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2521 
2522   // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2523   // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2524   // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2525   // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2526   // As long as Y is freely invertible, this will be neutral or a win.
2527   // Note: We don't generate the inverse max/min, just create the 'not' of
2528   // it and let other folds do the rest.
2529   if (match(Op0, m_Not(m_Value(X))) &&
2530       match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2531       !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2532     Value *Not = Builder.CreateNot(Op1);
2533     return BinaryOperator::CreateSub(Not, X);
2534   }
2535   if (match(Op1, m_Not(m_Value(X))) &&
2536       match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2537       !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2538     Value *Not = Builder.CreateNot(Op0);
2539     return BinaryOperator::CreateSub(X, Not);
2540   }
2541 
2542   // Optimize pointer differences into the same array into a size.  Consider:
2543   //  &A[10] - &A[0]: we should compile this to "10".
2544   Value *LHSOp, *RHSOp;
2545   if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2546       match(Op1, m_PtrToInt(m_Value(RHSOp))))
2547     if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2548                                                I.hasNoUnsignedWrap()))
2549       return replaceInstUsesWith(I, Res);
2550 
2551   // trunc(p)-trunc(q) -> trunc(p-q)
2552   if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2553       match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2554     if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2555                                                /* IsNUW */ false))
2556       return replaceInstUsesWith(I, Res);
2557 
2558   // Canonicalize a shifty way to code absolute value to the common pattern.
2559   // There are 2 potential commuted variants.
2560   // We're relying on the fact that we only do this transform when the shift has
2561   // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2562   // instructions).
2563   Value *A;
2564   const APInt *ShAmt;
2565   Type *Ty = I.getType();
2566   unsigned BitWidth = Ty->getScalarSizeInBits();
2567   if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2568       Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2569       match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2570     // B = ashr i32 A, 31 ; smear the sign bit
2571     // sub (xor A, B), B  ; flip bits if negative and subtract -1 (add 1)
2572     // --> (A < 0) ? -A : A
2573     Value *IsNeg = Builder.CreateIsNeg(A);
2574     // Copy the nsw flags from the sub to the negate.
2575     Value *NegA = I.hasNoUnsignedWrap()
2576                       ? Constant::getNullValue(A->getType())
2577                       : Builder.CreateNeg(A, "", I.hasNoSignedWrap());
2578     return SelectInst::Create(IsNeg, NegA, A);
2579   }
2580 
2581   // If we are subtracting a low-bit masked subset of some value from an add
2582   // of that same value with no low bits changed, that is clearing some low bits
2583   // of the sum:
2584   // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2585   const APInt *AddC, *AndC;
2586   if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2587       match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2588     unsigned Cttz = AddC->countr_zero();
2589     APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2590     if ((HighMask & *AndC).isZero())
2591       return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2592   }
2593 
2594   if (Instruction *V =
2595           canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2596     return V;
2597 
2598   // X - usub.sat(X, Y) => umin(X, Y)
2599   if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2600                                                            m_Value(Y)))))
2601     return replaceInstUsesWith(
2602         I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2603 
2604   // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2605   // TODO: The one-use restriction is not strictly necessary, but it may
2606   //       require improving other pattern matching and/or codegen.
2607   if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2608     return replaceInstUsesWith(
2609         I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2610 
2611   // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2612   if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2613     return replaceInstUsesWith(
2614         I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2615 
2616   // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2617   if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2618     Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2619     return BinaryOperator::CreateNeg(USub);
2620   }
2621 
2622   // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2623   if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2624     Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2625     return BinaryOperator::CreateNeg(USub);
2626   }
2627 
2628   // C - ctpop(X) => ctpop(~X) if C is bitwidth
2629   if (match(Op0, m_SpecificInt(BitWidth)) &&
2630       match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2631     return replaceInstUsesWith(
2632         I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2633                                    {Builder.CreateNot(X)}));
2634 
2635   // Reduce multiplies for difference-of-squares by factoring:
2636   // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2637   if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2638       match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2639     auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2640     auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2641     bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2642                         OBO1->hasNoSignedWrap() && BitWidth > 2;
2643     bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2644                         OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2645     Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2646     Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2647     Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2648     return replaceInstUsesWith(I, Mul);
2649   }
2650 
2651   // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2652   if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2653       match(Op1, m_OneUse(m_c_SMin(m_Specific(X), m_Specific(Y))))) {
2654     if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2655       Value *Sub =
2656           Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2657       Value *Call =
2658           Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2659       return replaceInstUsesWith(I, Call);
2660     }
2661   }
2662 
2663   if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
2664     return Res;
2665 
2666   return TryToNarrowDeduceFlags();
2667 }
2668 
2669 /// This eliminates floating-point negation in either 'fneg(X)' or
2670 /// 'fsub(-0.0, X)' form by combining into a constant operand.
foldFNegIntoConstant(Instruction & I,const DataLayout & DL)2671 static Instruction *foldFNegIntoConstant(Instruction &I, const DataLayout &DL) {
2672   // This is limited with one-use because fneg is assumed better for
2673   // reassociation and cheaper in codegen than fmul/fdiv.
2674   // TODO: Should the m_OneUse restriction be removed?
2675   Instruction *FNegOp;
2676   if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2677     return nullptr;
2678 
2679   Value *X;
2680   Constant *C;
2681 
2682   // Fold negation into constant operand.
2683   // -(X * C) --> X * (-C)
2684   if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2685     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2686       return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2687   // -(X / C) --> X / (-C)
2688   if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2689     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2690       return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2691   // -(C / X) --> (-C) / X
2692   if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2693     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2694       Instruction *FDiv = BinaryOperator::CreateFDivFMF(NegC, X, &I);
2695 
2696       // Intersect 'nsz' and 'ninf' because those special value exceptions may
2697       // not apply to the fdiv. Everything else propagates from the fneg.
2698       // TODO: We could propagate nsz/ninf from fdiv alone?
2699       FastMathFlags FMF = I.getFastMathFlags();
2700       FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2701       FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2702       FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2703       return FDiv;
2704     }
2705   // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2706   // -(X + C) --> -X + -C --> -C - X
2707   if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2708     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2709       return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2710 
2711   return nullptr;
2712 }
2713 
hoistFNegAboveFMulFDiv(Value * FNegOp,Instruction & FMFSource)2714 Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2715                                                       Instruction &FMFSource) {
2716   Value *X, *Y;
2717   if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2718     return cast<Instruction>(Builder.CreateFMulFMF(
2719         Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2720   }
2721 
2722   if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2723     return cast<Instruction>(Builder.CreateFDivFMF(
2724         Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2725   }
2726 
2727   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2728     // Make sure to preserve flags and metadata on the call.
2729     if (II->getIntrinsicID() == Intrinsic::ldexp) {
2730       FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2731       IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
2732       Builder.setFastMathFlags(FMF);
2733 
2734       CallInst *New = Builder.CreateCall(
2735           II->getCalledFunction(),
2736           {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2737       New->copyMetadata(*II);
2738       return New;
2739     }
2740   }
2741 
2742   return nullptr;
2743 }
2744 
visitFNeg(UnaryOperator & I)2745 Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2746   Value *Op = I.getOperand(0);
2747 
2748   if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2749                                   getSimplifyQuery().getWithInstruction(&I)))
2750     return replaceInstUsesWith(I, V);
2751 
2752   if (Instruction *X = foldFNegIntoConstant(I, DL))
2753     return X;
2754 
2755   Value *X, *Y;
2756 
2757   // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2758   if (I.hasNoSignedZeros() &&
2759       match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2760     return BinaryOperator::CreateFSubFMF(Y, X, &I);
2761 
2762   Value *OneUse;
2763   if (!match(Op, m_OneUse(m_Value(OneUse))))
2764     return nullptr;
2765 
2766   if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2767     return replaceInstUsesWith(I, R);
2768 
2769   // Try to eliminate fneg if at least 1 arm of the select is negated.
2770   Value *Cond;
2771   if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2772     // Unlike most transforms, this one is not safe to propagate nsz unless
2773     // it is present on the original select. We union the flags from the select
2774     // and fneg and then remove nsz if needed.
2775     auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2776       S->copyFastMathFlags(&I);
2777       if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2778         FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2779         S->setFastMathFlags(FMF);
2780         if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2781             !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2782           S->setHasNoSignedZeros(false);
2783       }
2784     };
2785     // -(Cond ? -P : Y) --> Cond ? P : -Y
2786     Value *P;
2787     if (match(X, m_FNeg(m_Value(P)))) {
2788       Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2789       SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2790       propagateSelectFMF(NewSel, P == Y);
2791       return NewSel;
2792     }
2793     // -(Cond ? X : -P) --> Cond ? -X : P
2794     if (match(Y, m_FNeg(m_Value(P)))) {
2795       Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2796       SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2797       propagateSelectFMF(NewSel, P == X);
2798       return NewSel;
2799     }
2800 
2801     // -(Cond ? X : C) --> Cond ? -X : -C
2802     // -(Cond ? C : Y) --> Cond ? -C : -Y
2803     if (match(X, m_ImmConstant()) || match(Y, m_ImmConstant())) {
2804       Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2805       Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2806       SelectInst *NewSel = SelectInst::Create(Cond, NegX, NegY);
2807       propagateSelectFMF(NewSel, /*CommonOperand=*/true);
2808       return NewSel;
2809     }
2810   }
2811 
2812   // fneg (copysign x, y) -> copysign x, (fneg y)
2813   if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2814     // The source copysign has an additional value input, so we can't propagate
2815     // flags the copysign doesn't also have.
2816     FastMathFlags FMF = I.getFastMathFlags();
2817     FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2818 
2819     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
2820     Builder.setFastMathFlags(FMF);
2821 
2822     Value *NegY = Builder.CreateFNeg(Y);
2823     Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2824     return replaceInstUsesWith(I, NewCopySign);
2825   }
2826 
2827   return nullptr;
2828 }
2829 
visitFSub(BinaryOperator & I)2830 Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2831   if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2832                                   I.getFastMathFlags(),
2833                                   getSimplifyQuery().getWithInstruction(&I)))
2834     return replaceInstUsesWith(I, V);
2835 
2836   if (Instruction *X = foldVectorBinop(I))
2837     return X;
2838 
2839   if (Instruction *Phi = foldBinopWithPhiOperands(I))
2840     return Phi;
2841 
2842   // Subtraction from -0.0 is the canonical form of fneg.
2843   // fsub -0.0, X ==> fneg X
2844   // fsub nsz 0.0, X ==> fneg nsz X
2845   //
2846   // FIXME This matcher does not respect FTZ or DAZ yet:
2847   // fsub -0.0, Denorm ==> +-0
2848   // fneg Denorm ==> -Denorm
2849   Value *Op;
2850   if (match(&I, m_FNeg(m_Value(Op))))
2851     return UnaryOperator::CreateFNegFMF(Op, &I);
2852 
2853   if (Instruction *X = foldFNegIntoConstant(I, DL))
2854     return X;
2855 
2856   if (Instruction *R = foldFBinOpOfIntCasts(I))
2857     return R;
2858 
2859   Value *X, *Y;
2860   Constant *C;
2861 
2862   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2863   // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2864   // Canonicalize to fadd to make analysis easier.
2865   // This can also help codegen because fadd is commutative.
2866   // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2867   // killed later. We still limit that particular transform with 'hasOneUse'
2868   // because an fneg is assumed better/cheaper than a generic fsub.
2869   if (I.hasNoSignedZeros() ||
2870       cannotBeNegativeZero(Op0, 0, getSimplifyQuery().getWithInstruction(&I))) {
2871     if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2872       Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2873       return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2874     }
2875   }
2876 
2877   // (-X) - Op1 --> -(X + Op1)
2878   if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2879       match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2880     Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2881     return UnaryOperator::CreateFNegFMF(FAdd, &I);
2882   }
2883 
2884   if (isa<Constant>(Op0))
2885     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2886       if (Instruction *NV = FoldOpIntoSelect(I, SI))
2887         return NV;
2888 
2889   // X - C --> X + (-C)
2890   // But don't transform constant expressions because there's an inverse fold
2891   // for X + (-Y) --> X - Y.
2892   if (match(Op1, m_ImmConstant(C)))
2893     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2894       return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2895 
2896   // X - (-Y) --> X + Y
2897   if (match(Op1, m_FNeg(m_Value(Y))))
2898     return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2899 
2900   // Similar to above, but look through a cast of the negated value:
2901   // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2902   Type *Ty = I.getType();
2903   if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2904     return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2905 
2906   // X - (fpext(-Y)) --> X + fpext(Y)
2907   if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2908     return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2909 
2910   // Similar to above, but look through fmul/fdiv of the negated value:
2911   // Op0 - (-X * Y) --> Op0 + (X * Y)
2912   // Op0 - (Y * -X) --> Op0 + (X * Y)
2913   if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2914     Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2915     return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2916   }
2917   // Op0 - (-X / Y) --> Op0 + (X / Y)
2918   // Op0 - (X / -Y) --> Op0 + (X / Y)
2919   if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2920       match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2921     Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2922     return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2923   }
2924 
2925   // Handle special cases for FSub with selects feeding the operation
2926   if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2927     return replaceInstUsesWith(I, V);
2928 
2929   if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2930     // (Y - X) - Y --> -X
2931     if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2932       return UnaryOperator::CreateFNegFMF(X, &I);
2933 
2934     // Y - (X + Y) --> -X
2935     // Y - (Y + X) --> -X
2936     if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2937       return UnaryOperator::CreateFNegFMF(X, &I);
2938 
2939     // (X * C) - X --> X * (C - 1.0)
2940     if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2941       if (Constant *CSubOne = ConstantFoldBinaryOpOperands(
2942               Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2943         return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2944     }
2945     // X - (X * C) --> X * (1.0 - C)
2946     if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2947       if (Constant *OneSubC = ConstantFoldBinaryOpOperands(
2948               Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2949         return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2950     }
2951 
2952     // Reassociate fsub/fadd sequences to create more fadd instructions and
2953     // reduce dependency chains:
2954     // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2955     Value *Z;
2956     if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2957                                      m_Value(Z))))) {
2958       Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2959       Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2960       return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2961     }
2962 
2963     auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2964       return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2965                                                                  m_Value(Vec)));
2966     };
2967     Value *A0, *A1, *V0, *V1;
2968     if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2969         V0->getType() == V1->getType()) {
2970       // Difference of sums is sum of differences:
2971       // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2972       Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2973       Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2974                                            {Sub->getType()}, {A0, Sub}, &I);
2975       return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2976     }
2977 
2978     if (Instruction *F = factorizeFAddFSub(I, Builder))
2979       return F;
2980 
2981     // TODO: This performs reassociative folds for FP ops. Some fraction of the
2982     // functionality has been subsumed by simple pattern matching here and in
2983     // InstSimplify. We should let a dedicated reassociation pass handle more
2984     // complex pattern matching and remove this from InstCombine.
2985     if (Value *V = FAddCombine(Builder).simplify(&I))
2986       return replaceInstUsesWith(I, V);
2987 
2988     // (X - Y) - Op1 --> X - (Y + Op1)
2989     if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2990       Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2991       return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
2992     }
2993   }
2994 
2995   return nullptr;
2996 }
2997