• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visit functions for add, fadd, sub, and fsub.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/IR/Constant.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/InstrTypes.h"
24 #include "llvm/IR/Instruction.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/Operator.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/IR/Type.h"
29 #include "llvm/IR/Value.h"
30 #include "llvm/Support/AlignOf.h"
31 #include "llvm/Support/Casting.h"
32 #include "llvm/Support/KnownBits.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()
86       { return reinterpret_cast<APFloat *>(&FpValBuf.buffer[0]); }
87 
getFpValPtr() const88     const APFloat *getFpValPtr() const
89       { return reinterpret_cast<const APFloat *>(&FpValBuf.buffer[0]); }
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     Value *performFactorization(Instruction *I);
190 
191     /// Convert given addend to a Value
192     Value *createAddendVal(const FAddend &A, bool& NeedNeg);
193 
194     /// Return the number of instructions needed to emit the N-ary addition.
195     unsigned calcInstrNumber(const AddendVect& Vect);
196 
197     Value *createFSub(Value *Opnd0, Value *Opnd1);
198     Value *createFAdd(Value *Opnd0, Value *Opnd1);
199     Value *createFMul(Value *Opnd0, Value *Opnd1);
200     Value *createFDiv(Value *Opnd0, Value *Opnd1);
201     Value *createFNeg(Value *V);
202     Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
203     void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
204 
205      // Debugging stuff are clustered here.
206     #ifndef NDEBUG
207       unsigned CreateInstrNum;
initCreateInstNum()208       void initCreateInstNum() { CreateInstrNum = 0; }
incCreateInstNum()209       void incCreateInstNum() { CreateInstrNum++; }
210     #else
initCreateInstNum()211       void initCreateInstNum() {}
incCreateInstNum()212       void incCreateInstNum() {}
213     #endif
214 
215     InstCombiner::BuilderTy &Builder;
216     Instruction *Instr = nullptr;
217   };
218 
219 } // end anonymous namespace
220 
221 //===----------------------------------------------------------------------===//
222 //
223 // Implementation of
224 //    {FAddendCoef, FAddend, FAddition, FAddCombine}.
225 //
226 //===----------------------------------------------------------------------===//
~FAddendCoef()227 FAddendCoef::~FAddendCoef() {
228   if (BufHasFpVal)
229     getFpValPtr()->~APFloat();
230 }
231 
set(const APFloat & C)232 void FAddendCoef::set(const APFloat& C) {
233   APFloat *P = getFpValPtr();
234 
235   if (isInt()) {
236     // As the buffer is meanless byte stream, we cannot call
237     // APFloat::operator=().
238     new(P) APFloat(C);
239   } else
240     *P = C;
241 
242   IsFp = BufHasFpVal = true;
243 }
244 
convertToFpType(const fltSemantics & Sem)245 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
246   if (!isInt())
247     return;
248 
249   APFloat *P = getFpValPtr();
250   if (IntVal > 0)
251     new(P) APFloat(Sem, IntVal);
252   else {
253     new(P) APFloat(Sem, 0 - IntVal);
254     P->changeSign();
255   }
256   IsFp = BufHasFpVal = true;
257 }
258 
createAPFloatFromInt(const fltSemantics & Sem,int Val)259 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
260   if (Val >= 0)
261     return APFloat(Sem, Val);
262 
263   APFloat T(Sem, 0 - Val);
264   T.changeSign();
265 
266   return T;
267 }
268 
operator =(const FAddendCoef & That)269 void FAddendCoef::operator=(const FAddendCoef &That) {
270   if (That.isInt())
271     set(That.IntVal);
272   else
273     set(That.getFpVal());
274 }
275 
operator +=(const FAddendCoef & That)276 void FAddendCoef::operator+=(const FAddendCoef &That) {
277   enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
278   if (isInt() == That.isInt()) {
279     if (isInt())
280       IntVal += That.IntVal;
281     else
282       getFpVal().add(That.getFpVal(), RndMode);
283     return;
284   }
285 
286   if (isInt()) {
287     const APFloat &T = That.getFpVal();
288     convertToFpType(T.getSemantics());
289     getFpVal().add(T, RndMode);
290     return;
291   }
292 
293   APFloat &T = getFpVal();
294   T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
295 }
296 
operator *=(const FAddendCoef & That)297 void FAddendCoef::operator*=(const FAddendCoef &That) {
298   if (That.isOne())
299     return;
300 
301   if (That.isMinusOne()) {
302     negate();
303     return;
304   }
305 
306   if (isInt() && That.isInt()) {
307     int Res = IntVal * (int)That.IntVal;
308     assert(!insaneIntVal(Res) && "Insane int value");
309     IntVal = Res;
310     return;
311   }
312 
313   const fltSemantics &Semantic =
314     isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
315 
316   if (isInt())
317     convertToFpType(Semantic);
318   APFloat &F0 = getFpVal();
319 
320   if (That.isInt())
321     F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
322                 APFloat::rmNearestTiesToEven);
323   else
324     F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
325 }
326 
negate()327 void FAddendCoef::negate() {
328   if (isInt())
329     IntVal = 0 - IntVal;
330   else
331     getFpVal().changeSign();
332 }
333 
getValue(Type * Ty) const334 Value *FAddendCoef::getValue(Type *Ty) const {
335   return isInt() ?
336     ConstantFP::get(Ty, float(IntVal)) :
337     ConstantFP::get(Ty->getContext(), getFpVal());
338 }
339 
340 // The definition of <Val>     Addends
341 // =========================================
342 //  A + B                     <1, A>, <1,B>
343 //  A - B                     <1, A>, <1,B>
344 //  0 - B                     <-1, B>
345 //  C * A,                    <C, A>
346 //  A + C                     <1, A> <C, NULL>
347 //  0 +/- 0                   <0, NULL> (corner case)
348 //
349 // Legend: A and B are not constant, C is constant
drillValueDownOneStep(Value * Val,FAddend & Addend0,FAddend & Addend1)350 unsigned FAddend::drillValueDownOneStep
351   (Value *Val, FAddend &Addend0, FAddend &Addend1) {
352   Instruction *I = nullptr;
353   if (!Val || !(I = dyn_cast<Instruction>(Val)))
354     return 0;
355 
356   unsigned Opcode = I->getOpcode();
357 
358   if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
359     ConstantFP *C0, *C1;
360     Value *Opnd0 = I->getOperand(0);
361     Value *Opnd1 = I->getOperand(1);
362     if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
363       Opnd0 = nullptr;
364 
365     if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
366       Opnd1 = nullptr;
367 
368     if (Opnd0) {
369       if (!C0)
370         Addend0.set(1, Opnd0);
371       else
372         Addend0.set(C0, nullptr);
373     }
374 
375     if (Opnd1) {
376       FAddend &Addend = Opnd0 ? Addend1 : Addend0;
377       if (!C1)
378         Addend.set(1, Opnd1);
379       else
380         Addend.set(C1, nullptr);
381       if (Opcode == Instruction::FSub)
382         Addend.negate();
383     }
384 
385     if (Opnd0 || Opnd1)
386       return Opnd0 && Opnd1 ? 2 : 1;
387 
388     // Both operands are zero. Weird!
389     Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
390     return 1;
391   }
392 
393   if (I->getOpcode() == Instruction::FMul) {
394     Value *V0 = I->getOperand(0);
395     Value *V1 = I->getOperand(1);
396     if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
397       Addend0.set(C, V1);
398       return 1;
399     }
400 
401     if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
402       Addend0.set(C, V0);
403       return 1;
404     }
405   }
406 
407   return 0;
408 }
409 
410 // Try to break *this* addend into two addends. e.g. Suppose this addend is
411 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
412 // i.e. <2.3, X> and <2.3, Y>.
drillAddendDownOneStep(FAddend & Addend0,FAddend & Addend1) const413 unsigned FAddend::drillAddendDownOneStep
414   (FAddend &Addend0, FAddend &Addend1) const {
415   if (isConstant())
416     return 0;
417 
418   unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
419   if (!BreakNum || Coeff.isOne())
420     return BreakNum;
421 
422   Addend0.Scale(Coeff);
423 
424   if (BreakNum == 2)
425     Addend1.Scale(Coeff);
426 
427   return BreakNum;
428 }
429 
430 // Try to perform following optimization on the input instruction I. Return the
431 // simplified expression if was successful; otherwise, return 0.
432 //
433 //   Instruction "I" is                Simplified into
434 // -------------------------------------------------------
435 //   (x * y) +/- (x * z)               x * (y +/- z)
436 //   (y / x) +/- (z / x)               (y +/- z) / x
performFactorization(Instruction * I)437 Value *FAddCombine::performFactorization(Instruction *I) {
438   assert((I->getOpcode() == Instruction::FAdd ||
439           I->getOpcode() == Instruction::FSub) && "Expect add/sub");
440 
441   Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0));
442   Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1));
443 
444   if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
445     return nullptr;
446 
447   bool isMpy = false;
448   if (I0->getOpcode() == Instruction::FMul)
449     isMpy = true;
450   else if (I0->getOpcode() != Instruction::FDiv)
451     return nullptr;
452 
453   Value *Opnd0_0 = I0->getOperand(0);
454   Value *Opnd0_1 = I0->getOperand(1);
455   Value *Opnd1_0 = I1->getOperand(0);
456   Value *Opnd1_1 = I1->getOperand(1);
457 
458   //  Input Instr I       Factor   AddSub0  AddSub1
459   //  ----------------------------------------------
460   // (x*y) +/- (x*z)        x        y         z
461   // (y/x) +/- (z/x)        x        y         z
462   Value *Factor = nullptr;
463   Value *AddSub0 = nullptr, *AddSub1 = nullptr;
464 
465   if (isMpy) {
466     if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
467       Factor = Opnd0_0;
468     else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
469       Factor = Opnd0_1;
470 
471     if (Factor) {
472       AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
473       AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
474     }
475   } else if (Opnd0_1 == Opnd1_1) {
476     Factor = Opnd0_1;
477     AddSub0 = Opnd0_0;
478     AddSub1 = Opnd1_0;
479   }
480 
481   if (!Factor)
482     return nullptr;
483 
484   FastMathFlags Flags;
485   Flags.setFast();
486   if (I0) Flags &= I->getFastMathFlags();
487   if (I1) Flags &= I->getFastMathFlags();
488 
489   // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
490   Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
491                       createFAdd(AddSub0, AddSub1) :
492                       createFSub(AddSub0, AddSub1);
493   if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
494     const APFloat &F = CFP->getValueAPF();
495     if (!F.isNormal())
496       return nullptr;
497   } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
498     II->setFastMathFlags(Flags);
499 
500   if (isMpy) {
501     Value *RI = createFMul(Factor, NewAddSub);
502     if (Instruction *II = dyn_cast<Instruction>(RI))
503       II->setFastMathFlags(Flags);
504     return RI;
505   }
506 
507   Value *RI = createFDiv(NewAddSub, Factor);
508   if (Instruction *II = dyn_cast<Instruction>(RI))
509     II->setFastMathFlags(Flags);
510   return RI;
511 }
512 
simplify(Instruction * I)513 Value *FAddCombine::simplify(Instruction *I) {
514   assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
515          "Expected 'reassoc'+'nsz' instruction");
516 
517   // Currently we are not able to handle vector type.
518   if (I->getType()->isVectorTy())
519     return nullptr;
520 
521   assert((I->getOpcode() == Instruction::FAdd ||
522           I->getOpcode() == Instruction::FSub) && "Expect add/sub");
523 
524   // Save the instruction before calling other member-functions.
525   Instr = I;
526 
527   FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
528 
529   unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
530 
531   // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
532   unsigned Opnd0_ExpNum = 0;
533   unsigned Opnd1_ExpNum = 0;
534 
535   if (!Opnd0.isConstant())
536     Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
537 
538   // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
539   if (OpndNum == 2 && !Opnd1.isConstant())
540     Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
541 
542   // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
543   if (Opnd0_ExpNum && Opnd1_ExpNum) {
544     AddendVect AllOpnds;
545     AllOpnds.push_back(&Opnd0_0);
546     AllOpnds.push_back(&Opnd1_0);
547     if (Opnd0_ExpNum == 2)
548       AllOpnds.push_back(&Opnd0_1);
549     if (Opnd1_ExpNum == 2)
550       AllOpnds.push_back(&Opnd1_1);
551 
552     // Compute instruction quota. We should save at least one instruction.
553     unsigned InstQuota = 0;
554 
555     Value *V0 = I->getOperand(0);
556     Value *V1 = I->getOperand(1);
557     InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
558                  (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
559 
560     if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
561       return R;
562   }
563 
564   if (OpndNum != 2) {
565     // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
566     // splitted into two addends, say "V = X - Y", the instruction would have
567     // been optimized into "I = Y - X" in the previous steps.
568     //
569     const FAddendCoef &CE = Opnd0.getCoef();
570     return CE.isOne() ? Opnd0.getSymVal() : nullptr;
571   }
572 
573   // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
574   if (Opnd1_ExpNum) {
575     AddendVect AllOpnds;
576     AllOpnds.push_back(&Opnd0);
577     AllOpnds.push_back(&Opnd1_0);
578     if (Opnd1_ExpNum == 2)
579       AllOpnds.push_back(&Opnd1_1);
580 
581     if (Value *R = simplifyFAdd(AllOpnds, 1))
582       return R;
583   }
584 
585   // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
586   if (Opnd0_ExpNum) {
587     AddendVect AllOpnds;
588     AllOpnds.push_back(&Opnd1);
589     AllOpnds.push_back(&Opnd0_0);
590     if (Opnd0_ExpNum == 2)
591       AllOpnds.push_back(&Opnd0_1);
592 
593     if (Value *R = simplifyFAdd(AllOpnds, 1))
594       return R;
595   }
596 
597   // step 6: Try factorization as the last resort,
598   return performFactorization(I);
599 }
600 
simplifyFAdd(AddendVect & Addends,unsigned InstrQuota)601 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
602   unsigned AddendNum = Addends.size();
603   assert(AddendNum <= 4 && "Too many addends");
604 
605   // For saving intermediate results;
606   unsigned NextTmpIdx = 0;
607   FAddend TmpResult[3];
608 
609   // Points to the constant addend of the resulting simplified expression.
610   // If the resulting expr has constant-addend, this constant-addend is
611   // desirable to reside at the top of the resulting expression tree. Placing
612   // constant close to supper-expr(s) will potentially reveal some optimization
613   // opportunities in super-expr(s).
614   const FAddend *ConstAdd = nullptr;
615 
616   // Simplified addends are placed <SimpVect>.
617   AddendVect SimpVect;
618 
619   // The outer loop works on one symbolic-value at a time. Suppose the input
620   // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
621   // The symbolic-values will be processed in this order: x, y, z.
622   for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
623 
624     const FAddend *ThisAddend = Addends[SymIdx];
625     if (!ThisAddend) {
626       // This addend was processed before.
627       continue;
628     }
629 
630     Value *Val = ThisAddend->getSymVal();
631     unsigned StartIdx = SimpVect.size();
632     SimpVect.push_back(ThisAddend);
633 
634     // The inner loop collects addends sharing same symbolic-value, and these
635     // addends will be later on folded into a single addend. Following above
636     // example, if the symbolic value "y" is being processed, the inner loop
637     // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
638     // be later on folded into "<b1+b2, y>".
639     for (unsigned SameSymIdx = SymIdx + 1;
640          SameSymIdx < AddendNum; SameSymIdx++) {
641       const FAddend *T = Addends[SameSymIdx];
642       if (T && T->getSymVal() == Val) {
643         // Set null such that next iteration of the outer loop will not process
644         // this addend again.
645         Addends[SameSymIdx] = nullptr;
646         SimpVect.push_back(T);
647       }
648     }
649 
650     // If multiple addends share same symbolic value, fold them together.
651     if (StartIdx + 1 != SimpVect.size()) {
652       FAddend &R = TmpResult[NextTmpIdx ++];
653       R = *SimpVect[StartIdx];
654       for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
655         R += *SimpVect[Idx];
656 
657       // Pop all addends being folded and push the resulting folded addend.
658       SimpVect.resize(StartIdx);
659       if (Val) {
660         if (!R.isZero()) {
661           SimpVect.push_back(&R);
662         }
663       } else {
664         // Don't push constant addend at this time. It will be the last element
665         // of <SimpVect>.
666         ConstAdd = &R;
667       }
668     }
669   }
670 
671   assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
672          "out-of-bound access");
673 
674   if (ConstAdd)
675     SimpVect.push_back(ConstAdd);
676 
677   Value *Result;
678   if (!SimpVect.empty())
679     Result = createNaryFAdd(SimpVect, InstrQuota);
680   else {
681     // The addition is folded to 0.0.
682     Result = ConstantFP::get(Instr->getType(), 0.0);
683   }
684 
685   return Result;
686 }
687 
createNaryFAdd(const AddendVect & Opnds,unsigned InstrQuota)688 Value *FAddCombine::createNaryFAdd
689   (const AddendVect &Opnds, unsigned InstrQuota) {
690   assert(!Opnds.empty() && "Expect at least one addend");
691 
692   // Step 1: Check if the # of instructions needed exceeds the quota.
693 
694   unsigned InstrNeeded = calcInstrNumber(Opnds);
695   if (InstrNeeded > InstrQuota)
696     return nullptr;
697 
698   initCreateInstNum();
699 
700   // step 2: Emit the N-ary addition.
701   // Note that at most three instructions are involved in Fadd-InstCombine: the
702   // addition in question, and at most two neighboring instructions.
703   // The resulting optimized addition should have at least one less instruction
704   // than the original addition expression tree. This implies that the resulting
705   // N-ary addition has at most two instructions, and we don't need to worry
706   // about tree-height when constructing the N-ary addition.
707 
708   Value *LastVal = nullptr;
709   bool LastValNeedNeg = false;
710 
711   // Iterate the addends, creating fadd/fsub using adjacent two addends.
712   for (const FAddend *Opnd : Opnds) {
713     bool NeedNeg;
714     Value *V = createAddendVal(*Opnd, NeedNeg);
715     if (!LastVal) {
716       LastVal = V;
717       LastValNeedNeg = NeedNeg;
718       continue;
719     }
720 
721     if (LastValNeedNeg == NeedNeg) {
722       LastVal = createFAdd(LastVal, V);
723       continue;
724     }
725 
726     if (LastValNeedNeg)
727       LastVal = createFSub(V, LastVal);
728     else
729       LastVal = createFSub(LastVal, V);
730 
731     LastValNeedNeg = false;
732   }
733 
734   if (LastValNeedNeg) {
735     LastVal = createFNeg(LastVal);
736   }
737 
738 #ifndef NDEBUG
739   assert(CreateInstrNum == InstrNeeded &&
740          "Inconsistent in instruction numbers");
741 #endif
742 
743   return LastVal;
744 }
745 
createFSub(Value * Opnd0,Value * Opnd1)746 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
747   Value *V = Builder.CreateFSub(Opnd0, Opnd1);
748   if (Instruction *I = dyn_cast<Instruction>(V))
749     createInstPostProc(I);
750   return V;
751 }
752 
createFNeg(Value * V)753 Value *FAddCombine::createFNeg(Value *V) {
754   Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType()));
755   Value *NewV = createFSub(Zero, V);
756   if (Instruction *I = dyn_cast<Instruction>(NewV))
757     createInstPostProc(I, true); // fneg's don't receive instruction numbers.
758   return NewV;
759 }
760 
createFAdd(Value * Opnd0,Value * Opnd1)761 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
762   Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
763   if (Instruction *I = dyn_cast<Instruction>(V))
764     createInstPostProc(I);
765   return V;
766 }
767 
createFMul(Value * Opnd0,Value * Opnd1)768 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
769   Value *V = Builder.CreateFMul(Opnd0, Opnd1);
770   if (Instruction *I = dyn_cast<Instruction>(V))
771     createInstPostProc(I);
772   return V;
773 }
774 
createFDiv(Value * Opnd0,Value * Opnd1)775 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) {
776   Value *V = Builder.CreateFDiv(Opnd0, Opnd1);
777   if (Instruction *I = dyn_cast<Instruction>(V))
778     createInstPostProc(I);
779   return V;
780 }
781 
createInstPostProc(Instruction * NewInstr,bool NoNumber)782 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
783   NewInstr->setDebugLoc(Instr->getDebugLoc());
784 
785   // Keep track of the number of instruction created.
786   if (!NoNumber)
787     incCreateInstNum();
788 
789   // Propagate fast-math flags
790   NewInstr->setFastMathFlags(Instr->getFastMathFlags());
791 }
792 
793 // Return the number of instruction needed to emit the N-ary addition.
794 // NOTE: Keep this function in sync with createAddendVal().
calcInstrNumber(const AddendVect & Opnds)795 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
796   unsigned OpndNum = Opnds.size();
797   unsigned InstrNeeded = OpndNum - 1;
798 
799   // The number of addends in the form of "(-1)*x".
800   unsigned NegOpndNum = 0;
801 
802   // Adjust the number of instructions needed to emit the N-ary add.
803   for (const FAddend *Opnd : Opnds) {
804     if (Opnd->isConstant())
805       continue;
806 
807     // The constant check above is really for a few special constant
808     // coefficients.
809     if (isa<UndefValue>(Opnd->getSymVal()))
810       continue;
811 
812     const FAddendCoef &CE = Opnd->getCoef();
813     if (CE.isMinusOne() || CE.isMinusTwo())
814       NegOpndNum++;
815 
816     // Let the addend be "c * x". If "c == +/-1", the value of the addend
817     // is immediately available; otherwise, it needs exactly one instruction
818     // to evaluate the value.
819     if (!CE.isMinusOne() && !CE.isOne())
820       InstrNeeded++;
821   }
822   if (NegOpndNum == OpndNum)
823     InstrNeeded++;
824   return InstrNeeded;
825 }
826 
827 // Input Addend        Value           NeedNeg(output)
828 // ================================================================
829 // Constant C          C               false
830 // <+/-1, V>           V               coefficient is -1
831 // <2/-2, V>          "fadd V, V"      coefficient is -2
832 // <C, V>             "fmul V, C"      false
833 //
834 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
createAddendVal(const FAddend & Opnd,bool & NeedNeg)835 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
836   const FAddendCoef &Coeff = Opnd.getCoef();
837 
838   if (Opnd.isConstant()) {
839     NeedNeg = false;
840     return Coeff.getValue(Instr->getType());
841   }
842 
843   Value *OpndVal = Opnd.getSymVal();
844 
845   if (Coeff.isMinusOne() || Coeff.isOne()) {
846     NeedNeg = Coeff.isMinusOne();
847     return OpndVal;
848   }
849 
850   if (Coeff.isTwo() || Coeff.isMinusTwo()) {
851     NeedNeg = Coeff.isMinusTwo();
852     return createFAdd(OpndVal, OpndVal);
853   }
854 
855   NeedNeg = false;
856   return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
857 }
858 
859 // Checks if any operand is negative and we can convert add to sub.
860 // This function checks for following negative patterns
861 //   ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
862 //   ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
863 //   XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
checkForNegativeOperand(BinaryOperator & I,InstCombiner::BuilderTy & Builder)864 static Value *checkForNegativeOperand(BinaryOperator &I,
865                                       InstCombiner::BuilderTy &Builder) {
866   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
867 
868   // This function creates 2 instructions to replace ADD, we need at least one
869   // of LHS or RHS to have one use to ensure benefit in transform.
870   if (!LHS->hasOneUse() && !RHS->hasOneUse())
871     return nullptr;
872 
873   Value *X = nullptr, *Y = nullptr, *Z = nullptr;
874   const APInt *C1 = nullptr, *C2 = nullptr;
875 
876   // if ONE is on other side, swap
877   if (match(RHS, m_Add(m_Value(X), m_One())))
878     std::swap(LHS, RHS);
879 
880   if (match(LHS, m_Add(m_Value(X), m_One()))) {
881     // if XOR on other side, swap
882     if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
883       std::swap(X, RHS);
884 
885     if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
886       // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
887       // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
888       if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
889         Value *NewAnd = Builder.CreateAnd(Z, *C1);
890         return Builder.CreateSub(RHS, NewAnd, "sub");
891       } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
892         // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
893         // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
894         Value *NewOr = Builder.CreateOr(Z, ~(*C1));
895         return Builder.CreateSub(RHS, NewOr, "sub");
896       }
897     }
898   }
899 
900   // Restore LHS and RHS
901   LHS = I.getOperand(0);
902   RHS = I.getOperand(1);
903 
904   // if XOR is on other side, swap
905   if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
906     std::swap(LHS, RHS);
907 
908   // C2 is ODD
909   // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
910   // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
911   if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
912     if (C1->countTrailingZeros() == 0)
913       if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
914         Value *NewOr = Builder.CreateOr(Z, ~(*C2));
915         return Builder.CreateSub(RHS, NewOr, "sub");
916       }
917   return nullptr;
918 }
919 
foldAddWithConstant(BinaryOperator & Add)920 Instruction *InstCombiner::foldAddWithConstant(BinaryOperator &Add) {
921   Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
922   Constant *Op1C;
923   if (!match(Op1, m_Constant(Op1C)))
924     return nullptr;
925 
926   if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
927     return NV;
928 
929   Value *X, *Y;
930 
931   // add (sub X, Y), -1 --> add (not Y), X
932   if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
933       match(Op1, m_AllOnes()))
934     return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
935 
936   // zext(bool) + C -> bool ? C + 1 : C
937   if (match(Op0, m_ZExt(m_Value(X))) &&
938       X->getType()->getScalarSizeInBits() == 1)
939     return SelectInst::Create(X, AddOne(Op1C), Op1);
940 
941   // ~X + C --> (C-1) - X
942   if (match(Op0, m_Not(m_Value(X))))
943     return BinaryOperator::CreateSub(SubOne(Op1C), X);
944 
945   const APInt *C;
946   if (!match(Op1, m_APInt(C)))
947     return nullptr;
948 
949   if (C->isSignMask()) {
950     // If wrapping is not allowed, then the addition must set the sign bit:
951     // X + (signmask) --> X | signmask
952     if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
953       return BinaryOperator::CreateOr(Op0, Op1);
954 
955     // If wrapping is allowed, then the addition flips the sign bit of LHS:
956     // X + (signmask) --> X ^ signmask
957     return BinaryOperator::CreateXor(Op0, Op1);
958   }
959 
960   // Is this add the last step in a convoluted sext?
961   // add(zext(xor i16 X, -32768), -32768) --> sext X
962   Type *Ty = Add.getType();
963   const APInt *C2;
964   if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
965       C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
966     return CastInst::Create(Instruction::SExt, X, Ty);
967 
968   // (add (zext (add nuw X, C2)), C) --> (zext (add nuw X, C2 + C))
969   if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
970       C->isNegative() && C->sge(-C2->sext(C->getBitWidth()))) {
971     Constant *NewC =
972         ConstantInt::get(X->getType(), *C2 + C->trunc(C2->getBitWidth()));
973     return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
974   }
975 
976   if (C->isOneValue() && Op0->hasOneUse()) {
977     // add (sext i1 X), 1 --> zext (not X)
978     // TODO: The smallest IR representation is (select X, 0, 1), and that would
979     // not require the one-use check. But we need to remove a transform in
980     // visitSelect and make sure that IR value tracking for select is equal or
981     // better than for these ops.
982     if (match(Op0, m_SExt(m_Value(X))) &&
983         X->getType()->getScalarSizeInBits() == 1)
984       return new ZExtInst(Builder.CreateNot(X), Ty);
985 
986     // Shifts and add used to flip and mask off the low bit:
987     // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
988     const APInt *C3;
989     if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
990         C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
991       Value *NotX = Builder.CreateNot(X);
992       return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
993     }
994   }
995 
996   return nullptr;
997 }
998 
999 // Matches multiplication expression Op * C where C is a constant. Returns the
1000 // constant value in C and the other operand in Op. Returns true if such a
1001 // match is found.
MatchMul(Value * E,Value * & Op,APInt & C)1002 static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1003   const APInt *AI;
1004   if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1005     C = *AI;
1006     return true;
1007   }
1008   if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1009     C = APInt(AI->getBitWidth(), 1);
1010     C <<= *AI;
1011     return true;
1012   }
1013   return false;
1014 }
1015 
1016 // Matches remainder expression Op % C where C is a constant. Returns the
1017 // constant value in C and the other operand in Op. Returns the signedness of
1018 // the remainder operation in IsSigned. Returns true if such a match is
1019 // found.
MatchRem(Value * E,Value * & Op,APInt & C,bool & IsSigned)1020 static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1021   const APInt *AI;
1022   IsSigned = false;
1023   if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1024     IsSigned = true;
1025     C = *AI;
1026     return true;
1027   }
1028   if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1029     C = *AI;
1030     return true;
1031   }
1032   if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1033     C = *AI + 1;
1034     return true;
1035   }
1036   return false;
1037 }
1038 
1039 // Matches division expression Op / C with the given signedness as indicated
1040 // by IsSigned, where C is a constant. Returns the constant value in C and the
1041 // other operand in Op. Returns true if such a match is found.
MatchDiv(Value * E,Value * & Op,APInt & C,bool IsSigned)1042 static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1043   const APInt *AI;
1044   if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1045     C = *AI;
1046     return true;
1047   }
1048   if (!IsSigned) {
1049     if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1050       C = *AI;
1051       return true;
1052     }
1053     if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1054       C = APInt(AI->getBitWidth(), 1);
1055       C <<= *AI;
1056       return true;
1057     }
1058   }
1059   return false;
1060 }
1061 
1062 // Returns whether C0 * C1 with the given signedness overflows.
MulWillOverflow(APInt & C0,APInt & C1,bool IsSigned)1063 static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1064   bool overflow;
1065   if (IsSigned)
1066     (void)C0.smul_ov(C1, overflow);
1067   else
1068     (void)C0.umul_ov(C1, overflow);
1069   return overflow;
1070 }
1071 
1072 // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1073 // does not overflow.
SimplifyAddWithRemainder(BinaryOperator & I)1074 Value *InstCombiner::SimplifyAddWithRemainder(BinaryOperator &I) {
1075   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1076   Value *X, *MulOpV;
1077   APInt C0, MulOpC;
1078   bool IsSigned;
1079   // Match I = X % C0 + MulOpV * C0
1080   if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1081        (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1082       C0 == MulOpC) {
1083     Value *RemOpV;
1084     APInt C1;
1085     bool Rem2IsSigned;
1086     // Match MulOpC = RemOpV % C1
1087     if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1088         IsSigned == Rem2IsSigned) {
1089       Value *DivOpV;
1090       APInt DivOpC;
1091       // Match RemOpV = X / C0
1092       if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1093           C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1094         Value *NewDivisor =
1095             ConstantInt::get(X->getType()->getContext(), C0 * C1);
1096         return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1097                         : Builder.CreateURem(X, NewDivisor, "urem");
1098       }
1099     }
1100   }
1101 
1102   return nullptr;
1103 }
1104 
1105 /// Fold
1106 ///   (1 << NBits) - 1
1107 /// Into:
1108 ///   ~(-(1 << NBits))
1109 /// Because a 'not' is better for bit-tracking analysis and other transforms
1110 /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
canonicalizeLowbitMask(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1111 static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1112                                            InstCombiner::BuilderTy &Builder) {
1113   Value *NBits;
1114   if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1115     return nullptr;
1116 
1117   Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1118   Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1119   // Be wary of constant folding.
1120   if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1121     // Always NSW. But NUW propagates from `add`.
1122     BOp->setHasNoSignedWrap();
1123     BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1124   }
1125 
1126   return BinaryOperator::CreateNot(NotMask, I.getName());
1127 }
1128 
visitAdd(BinaryOperator & I)1129 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1130   if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1131                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1132                                  SQ.getWithInstruction(&I)))
1133     return replaceInstUsesWith(I, V);
1134 
1135   if (SimplifyAssociativeOrCommutative(I))
1136     return &I;
1137 
1138   if (Instruction *X = foldShuffledBinop(I))
1139     return X;
1140 
1141   // (A*B)+(A*C) -> A*(B+C) etc
1142   if (Value *V = SimplifyUsingDistributiveLaws(I))
1143     return replaceInstUsesWith(I, V);
1144 
1145   if (Instruction *X = foldAddWithConstant(I))
1146     return X;
1147 
1148   // FIXME: This should be moved into the above helper function to allow these
1149   // transforms for general constant or constant splat vectors.
1150   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1151   Type *Ty = I.getType();
1152   if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1153     Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1154     if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1155       unsigned TySizeBits = Ty->getScalarSizeInBits();
1156       const APInt &RHSVal = CI->getValue();
1157       unsigned ExtendAmt = 0;
1158       // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1159       // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1160       if (XorRHS->getValue() == -RHSVal) {
1161         if (RHSVal.isPowerOf2())
1162           ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1163         else if (XorRHS->getValue().isPowerOf2())
1164           ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1165       }
1166 
1167       if (ExtendAmt) {
1168         APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1169         if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1170           ExtendAmt = 0;
1171       }
1172 
1173       if (ExtendAmt) {
1174         Constant *ShAmt = ConstantInt::get(Ty, ExtendAmt);
1175         Value *NewShl = Builder.CreateShl(XorLHS, ShAmt, "sext");
1176         return BinaryOperator::CreateAShr(NewShl, ShAmt);
1177       }
1178 
1179       // If this is a xor that was canonicalized from a sub, turn it back into
1180       // a sub and fuse this add with it.
1181       if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1182         KnownBits LHSKnown = computeKnownBits(XorLHS, 0, &I);
1183         if ((XorRHS->getValue() | LHSKnown.Zero).isAllOnesValue())
1184           return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1185                                            XorLHS);
1186       }
1187       // (X + signmask) + C could have gotten canonicalized to (X^signmask) + C,
1188       // transform them into (X + (signmask ^ C))
1189       if (XorRHS->getValue().isSignMask())
1190         return BinaryOperator::CreateAdd(XorLHS,
1191                                          ConstantExpr::getXor(XorRHS, CI));
1192     }
1193   }
1194 
1195   if (Ty->isIntOrIntVectorTy(1))
1196     return BinaryOperator::CreateXor(LHS, RHS);
1197 
1198   // X + X --> X << 1
1199   if (LHS == RHS) {
1200     auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1201     Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1202     Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1203     return Shl;
1204   }
1205 
1206   Value *A, *B;
1207   if (match(LHS, m_Neg(m_Value(A)))) {
1208     // -A + -B --> -(A + B)
1209     if (match(RHS, m_Neg(m_Value(B))))
1210       return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1211 
1212     // -A + B --> B - A
1213     return BinaryOperator::CreateSub(RHS, A);
1214   }
1215 
1216   // A + -B  -->  A - B
1217   if (match(RHS, m_Neg(m_Value(B))))
1218     return BinaryOperator::CreateSub(LHS, B);
1219 
1220   if (Value *V = checkForNegativeOperand(I, Builder))
1221     return replaceInstUsesWith(I, V);
1222 
1223   // (A + 1) + ~B --> A - B
1224   // ~B + (A + 1) --> A - B
1225   if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))))
1226     return BinaryOperator::CreateSub(A, B);
1227 
1228   // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1229   if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1230 
1231   // A+B --> A|B iff A and B have no bits set in common.
1232   if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1233     return BinaryOperator::CreateOr(LHS, RHS);
1234 
1235   // FIXME: We already did a check for ConstantInt RHS above this.
1236   // FIXME: Is this pattern covered by another fold? No regression tests fail on
1237   // removal.
1238   if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1239     // (X & FF00) + xx00  -> (X+xx00) & FF00
1240     Value *X;
1241     ConstantInt *C2;
1242     if (LHS->hasOneUse() &&
1243         match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1244         CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1245       // See if all bits from the first bit set in the Add RHS up are included
1246       // in the mask.  First, get the rightmost bit.
1247       const APInt &AddRHSV = CRHS->getValue();
1248 
1249       // Form a mask of all bits from the lowest bit added through the top.
1250       APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1251 
1252       // See if the and mask includes all of these bits.
1253       APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1254 
1255       if (AddRHSHighBits == AddRHSHighBitsAnd) {
1256         // Okay, the xform is safe.  Insert the new add pronto.
1257         Value *NewAdd = Builder.CreateAdd(X, CRHS, LHS->getName());
1258         return BinaryOperator::CreateAnd(NewAdd, C2);
1259       }
1260     }
1261   }
1262 
1263   // add (select X 0 (sub n A)) A  -->  select X A n
1264   {
1265     SelectInst *SI = dyn_cast<SelectInst>(LHS);
1266     Value *A = RHS;
1267     if (!SI) {
1268       SI = dyn_cast<SelectInst>(RHS);
1269       A = LHS;
1270     }
1271     if (SI && SI->hasOneUse()) {
1272       Value *TV = SI->getTrueValue();
1273       Value *FV = SI->getFalseValue();
1274       Value *N;
1275 
1276       // Can we fold the add into the argument of the select?
1277       // We check both true and false select arguments for a matching subtract.
1278       if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1279         // Fold the add into the true select value.
1280         return SelectInst::Create(SI->getCondition(), N, A);
1281 
1282       if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1283         // Fold the add into the false select value.
1284         return SelectInst::Create(SI->getCondition(), A, N);
1285     }
1286   }
1287 
1288   // Check for (add (sext x), y), see if we can merge this into an
1289   // integer add followed by a sext.
1290   if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
1291     // (add (sext x), cst) --> (sext (add x, cst'))
1292     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1293       if (LHSConv->hasOneUse()) {
1294         Constant *CI =
1295             ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1296         if (ConstantExpr::getSExt(CI, Ty) == RHSC &&
1297             willNotOverflowSignedAdd(LHSConv->getOperand(0), CI, I)) {
1298           // Insert the new, smaller add.
1299           Value *NewAdd =
1300               Builder.CreateNSWAdd(LHSConv->getOperand(0), CI, "addconv");
1301           return new SExtInst(NewAdd, Ty);
1302         }
1303       }
1304     }
1305 
1306     // (add (sext x), (sext y)) --> (sext (add int x, y))
1307     if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
1308       // Only do this if x/y have the same type, if at least one of them has a
1309       // single use (so we don't increase the number of sexts), and if the
1310       // integer add will not overflow.
1311       if (LHSConv->getOperand(0)->getType() ==
1312               RHSConv->getOperand(0)->getType() &&
1313           (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1314           willNotOverflowSignedAdd(LHSConv->getOperand(0),
1315                                    RHSConv->getOperand(0), I)) {
1316         // Insert the new integer add.
1317         Value *NewAdd = Builder.CreateNSWAdd(LHSConv->getOperand(0),
1318                                              RHSConv->getOperand(0), "addconv");
1319         return new SExtInst(NewAdd, Ty);
1320       }
1321     }
1322   }
1323 
1324   // Check for (add (zext x), y), see if we can merge this into an
1325   // integer add followed by a zext.
1326   if (auto *LHSConv = dyn_cast<ZExtInst>(LHS)) {
1327     // (add (zext x), cst) --> (zext (add x, cst'))
1328     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1329       if (LHSConv->hasOneUse()) {
1330         Constant *CI =
1331             ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1332         if (ConstantExpr::getZExt(CI, Ty) == RHSC &&
1333             willNotOverflowUnsignedAdd(LHSConv->getOperand(0), CI, I)) {
1334           // Insert the new, smaller add.
1335           Value *NewAdd =
1336               Builder.CreateNUWAdd(LHSConv->getOperand(0), CI, "addconv");
1337           return new ZExtInst(NewAdd, Ty);
1338         }
1339       }
1340     }
1341 
1342     // (add (zext x), (zext y)) --> (zext (add int x, y))
1343     if (auto *RHSConv = dyn_cast<ZExtInst>(RHS)) {
1344       // Only do this if x/y have the same type, if at least one of them has a
1345       // single use (so we don't increase the number of zexts), and if the
1346       // integer add will not overflow.
1347       if (LHSConv->getOperand(0)->getType() ==
1348               RHSConv->getOperand(0)->getType() &&
1349           (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1350           willNotOverflowUnsignedAdd(LHSConv->getOperand(0),
1351                                      RHSConv->getOperand(0), I)) {
1352         // Insert the new integer add.
1353         Value *NewAdd = Builder.CreateNUWAdd(
1354             LHSConv->getOperand(0), RHSConv->getOperand(0), "addconv");
1355         return new ZExtInst(NewAdd, Ty);
1356       }
1357     }
1358   }
1359 
1360   // (add (xor A, B) (and A, B)) --> (or A, B)
1361   // (add (and A, B) (xor A, B)) --> (or A, B)
1362   if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1363                           m_c_And(m_Deferred(A), m_Deferred(B)))))
1364     return BinaryOperator::CreateOr(A, B);
1365 
1366   // (add (or A, B) (and A, B)) --> (add A, B)
1367   // (add (and A, B) (or A, B)) --> (add A, B)
1368   if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1369                           m_c_And(m_Deferred(A), m_Deferred(B))))) {
1370     I.setOperand(0, A);
1371     I.setOperand(1, B);
1372     return &I;
1373   }
1374 
1375   // TODO(jingyue): Consider willNotOverflowSignedAdd and
1376   // willNotOverflowUnsignedAdd to reduce the number of invocations of
1377   // computeKnownBits.
1378   bool Changed = false;
1379   if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1380     Changed = true;
1381     I.setHasNoSignedWrap(true);
1382   }
1383   if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1384     Changed = true;
1385     I.setHasNoUnsignedWrap(true);
1386   }
1387 
1388   if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1389     return V;
1390 
1391   return Changed ? &I : nullptr;
1392 }
1393 
visitFAdd(BinaryOperator & I)1394 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
1395   if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1396                                   I.getFastMathFlags(),
1397                                   SQ.getWithInstruction(&I)))
1398     return replaceInstUsesWith(I, V);
1399 
1400   if (SimplifyAssociativeOrCommutative(I))
1401     return &I;
1402 
1403   if (Instruction *X = foldShuffledBinop(I))
1404     return X;
1405 
1406   if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1407     return FoldedFAdd;
1408 
1409   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1410   Value *X;
1411   // (-X) + Y --> Y - X
1412   if (match(LHS, m_FNeg(m_Value(X))))
1413     return BinaryOperator::CreateFSubFMF(RHS, X, &I);
1414   // Y + (-X) --> Y - X
1415   if (match(RHS, m_FNeg(m_Value(X))))
1416     return BinaryOperator::CreateFSubFMF(LHS, X, &I);
1417 
1418   // Check for (fadd double (sitofp x), y), see if we can merge this into an
1419   // integer add followed by a promotion.
1420   if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1421     Value *LHSIntVal = LHSConv->getOperand(0);
1422     Type *FPType = LHSConv->getType();
1423 
1424     // TODO: This check is overly conservative. In many cases known bits
1425     // analysis can tell us that the result of the addition has less significant
1426     // bits than the integer type can hold.
1427     auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1428       Type *FScalarTy = FTy->getScalarType();
1429       Type *IScalarTy = ITy->getScalarType();
1430 
1431       // Do we have enough bits in the significand to represent the result of
1432       // the integer addition?
1433       unsigned MaxRepresentableBits =
1434           APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1435       return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1436     };
1437 
1438     // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1439     // ... if the constant fits in the integer value.  This is useful for things
1440     // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1441     // requires a constant pool load, and generally allows the add to be better
1442     // instcombined.
1443     if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1444       if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1445         Constant *CI =
1446           ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1447         if (LHSConv->hasOneUse() &&
1448             ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1449             willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1450           // Insert the new integer add.
1451           Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1452           return new SIToFPInst(NewAdd, I.getType());
1453         }
1454       }
1455 
1456     // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1457     if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1458       Value *RHSIntVal = RHSConv->getOperand(0);
1459       // It's enough to check LHS types only because we require int types to
1460       // be the same for this transform.
1461       if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1462         // Only do this if x/y have the same type, if at least one of them has a
1463         // single use (so we don't increase the number of int->fp conversions),
1464         // and if the integer add will not overflow.
1465         if (LHSIntVal->getType() == RHSIntVal->getType() &&
1466             (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1467             willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1468           // Insert the new integer add.
1469           Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1470           return new SIToFPInst(NewAdd, I.getType());
1471         }
1472       }
1473     }
1474   }
1475 
1476   // Handle specials cases for FAdd with selects feeding the operation
1477   if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1478     return replaceInstUsesWith(I, V);
1479 
1480   if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1481     if (Value *V = FAddCombine(Builder).simplify(&I))
1482       return replaceInstUsesWith(I, V);
1483   }
1484 
1485   return nullptr;
1486 }
1487 
1488 /// Optimize pointer differences into the same array into a size.  Consider:
1489 ///  &A[10] - &A[0]: we should compile this to "10".  LHS/RHS are the pointer
1490 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
OptimizePointerDifference(Value * LHS,Value * RHS,Type * Ty)1491 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1492                                                Type *Ty) {
1493   // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1494   // this.
1495   bool Swapped = false;
1496   GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1497 
1498   // For now we require one side to be the base pointer "A" or a constant
1499   // GEP derived from it.
1500   if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1501     // (gep X, ...) - X
1502     if (LHSGEP->getOperand(0) == RHS) {
1503       GEP1 = LHSGEP;
1504       Swapped = false;
1505     } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1506       // (gep X, ...) - (gep X, ...)
1507       if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1508             RHSGEP->getOperand(0)->stripPointerCasts()) {
1509         GEP2 = RHSGEP;
1510         GEP1 = LHSGEP;
1511         Swapped = false;
1512       }
1513     }
1514   }
1515 
1516   if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1517     // X - (gep X, ...)
1518     if (RHSGEP->getOperand(0) == LHS) {
1519       GEP1 = RHSGEP;
1520       Swapped = true;
1521     } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1522       // (gep X, ...) - (gep X, ...)
1523       if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1524             LHSGEP->getOperand(0)->stripPointerCasts()) {
1525         GEP2 = LHSGEP;
1526         GEP1 = RHSGEP;
1527         Swapped = true;
1528       }
1529     }
1530   }
1531 
1532   if (!GEP1)
1533     // No GEP found.
1534     return nullptr;
1535 
1536   if (GEP2) {
1537     // (gep X, ...) - (gep X, ...)
1538     //
1539     // Avoid duplicating the arithmetic if there are more than one non-constant
1540     // indices between the two GEPs and either GEP has a non-constant index and
1541     // multiple users. If zero non-constant index, the result is a constant and
1542     // there is no duplication. If one non-constant index, the result is an add
1543     // or sub with a constant, which is no larger than the original code, and
1544     // there's no duplicated arithmetic, even if either GEP has multiple
1545     // users. If more than one non-constant indices combined, as long as the GEP
1546     // with at least one non-constant index doesn't have multiple users, there
1547     // is no duplication.
1548     unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1549     unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1550     if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1551         ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1552          (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1553       return nullptr;
1554     }
1555   }
1556 
1557   // Emit the offset of the GEP and an intptr_t.
1558   Value *Result = EmitGEPOffset(GEP1);
1559 
1560   // If we had a constant expression GEP on the other side offsetting the
1561   // pointer, subtract it from the offset we have.
1562   if (GEP2) {
1563     Value *Offset = EmitGEPOffset(GEP2);
1564     Result = Builder.CreateSub(Result, Offset);
1565   }
1566 
1567   // If we have p - gep(p, ...)  then we have to negate the result.
1568   if (Swapped)
1569     Result = Builder.CreateNeg(Result, "diff.neg");
1570 
1571   return Builder.CreateIntCast(Result, Ty, true);
1572 }
1573 
visitSub(BinaryOperator & I)1574 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1575   if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1576                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1577                                  SQ.getWithInstruction(&I)))
1578     return replaceInstUsesWith(I, V);
1579 
1580   if (Instruction *X = foldShuffledBinop(I))
1581     return X;
1582 
1583   // (A*B)-(A*C) -> A*(B-C) etc
1584   if (Value *V = SimplifyUsingDistributiveLaws(I))
1585     return replaceInstUsesWith(I, V);
1586 
1587   // If this is a 'B = x-(-A)', change to B = x+A.
1588   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1589   if (Value *V = dyn_castNegVal(Op1)) {
1590     BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1591 
1592     if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1593       assert(BO->getOpcode() == Instruction::Sub &&
1594              "Expected a subtraction operator!");
1595       if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1596         Res->setHasNoSignedWrap(true);
1597     } else {
1598       if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1599         Res->setHasNoSignedWrap(true);
1600     }
1601 
1602     return Res;
1603   }
1604 
1605   if (I.getType()->isIntOrIntVectorTy(1))
1606     return BinaryOperator::CreateXor(Op0, Op1);
1607 
1608   // Replace (-1 - A) with (~A).
1609   if (match(Op0, m_AllOnes()))
1610     return BinaryOperator::CreateNot(Op1);
1611 
1612   // (~X) - (~Y) --> Y - X
1613   Value *X, *Y;
1614   if (match(Op0, m_Not(m_Value(X))) && match(Op1, m_Not(m_Value(Y))))
1615     return BinaryOperator::CreateSub(Y, X);
1616 
1617   // (X + -1) - Y --> ~Y + X
1618   if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1619     return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1620 
1621   // Y - (X + 1) --> ~X + Y
1622   if (match(Op1, m_OneUse(m_Add(m_Value(X), m_One()))))
1623     return BinaryOperator::CreateAdd(Builder.CreateNot(X), Op0);
1624 
1625   if (Constant *C = dyn_cast<Constant>(Op0)) {
1626     bool IsNegate = match(C, m_ZeroInt());
1627     Value *X;
1628     if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1629       // 0 - (zext bool) --> sext bool
1630       // C - (zext bool) --> bool ? C - 1 : C
1631       if (IsNegate)
1632         return CastInst::CreateSExtOrBitCast(X, I.getType());
1633       return SelectInst::Create(X, SubOne(C), C);
1634     }
1635     if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1636       // 0 - (sext bool) --> zext bool
1637       // C - (sext bool) --> bool ? C + 1 : C
1638       if (IsNegate)
1639         return CastInst::CreateZExtOrBitCast(X, I.getType());
1640       return SelectInst::Create(X, AddOne(C), C);
1641     }
1642 
1643     // C - ~X == X + (1+C)
1644     if (match(Op1, m_Not(m_Value(X))))
1645       return BinaryOperator::CreateAdd(X, AddOne(C));
1646 
1647     // Try to fold constant sub into select arguments.
1648     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1649       if (Instruction *R = FoldOpIntoSelect(I, SI))
1650         return R;
1651 
1652     // Try to fold constant sub into PHI values.
1653     if (PHINode *PN = dyn_cast<PHINode>(Op1))
1654       if (Instruction *R = foldOpIntoPhi(I, PN))
1655         return R;
1656 
1657     // C-(X+C2) --> (C-C2)-X
1658     Constant *C2;
1659     if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1660       return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1661   }
1662 
1663   const APInt *Op0C;
1664   if (match(Op0, m_APInt(Op0C))) {
1665     unsigned BitWidth = I.getType()->getScalarSizeInBits();
1666 
1667     // -(X >>u 31) -> (X >>s 31)
1668     // -(X >>s 31) -> (X >>u 31)
1669     if (Op0C->isNullValue()) {
1670       Value *X;
1671       const APInt *ShAmt;
1672       if (match(Op1, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1673           *ShAmt == BitWidth - 1) {
1674         Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1675         return BinaryOperator::CreateAShr(X, ShAmtOp);
1676       }
1677       if (match(Op1, m_AShr(m_Value(X), m_APInt(ShAmt))) &&
1678           *ShAmt == BitWidth - 1) {
1679         Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1680         return BinaryOperator::CreateLShr(X, ShAmtOp);
1681       }
1682 
1683       if (Op1->hasOneUse()) {
1684         Value *LHS, *RHS;
1685         SelectPatternFlavor SPF = matchSelectPattern(Op1, LHS, RHS).Flavor;
1686         if (SPF == SPF_ABS || SPF == SPF_NABS) {
1687           // This is a negate of an ABS/NABS pattern. Just swap the operands
1688           // of the select.
1689           SelectInst *SI = cast<SelectInst>(Op1);
1690           Value *TrueVal = SI->getTrueValue();
1691           Value *FalseVal = SI->getFalseValue();
1692           SI->setTrueValue(FalseVal);
1693           SI->setFalseValue(TrueVal);
1694           // Don't swap prof metadata, we didn't change the branch behavior.
1695           return replaceInstUsesWith(I, SI);
1696         }
1697       }
1698     }
1699 
1700     // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1701     // zero.
1702     if (Op0C->isMask()) {
1703       KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1704       if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
1705         return BinaryOperator::CreateXor(Op1, Op0);
1706     }
1707   }
1708 
1709   {
1710     Value *Y;
1711     // X-(X+Y) == -Y    X-(Y+X) == -Y
1712     if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1713       return BinaryOperator::CreateNeg(Y);
1714 
1715     // (X-Y)-X == -Y
1716     if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1717       return BinaryOperator::CreateNeg(Y);
1718   }
1719 
1720   // (sub (or A, B), (xor A, B)) --> (and A, B)
1721   {
1722     Value *A, *B;
1723     if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1724         match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1725       return BinaryOperator::CreateAnd(A, B);
1726   }
1727 
1728   {
1729     Value *Y;
1730     // ((X | Y) - X) --> (~X & Y)
1731     if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1732       return BinaryOperator::CreateAnd(
1733           Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1734   }
1735 
1736   if (Op1->hasOneUse()) {
1737     Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1738     Constant *C = nullptr;
1739 
1740     // (X - (Y - Z))  -->  (X + (Z - Y)).
1741     if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1742       return BinaryOperator::CreateAdd(Op0,
1743                                       Builder.CreateSub(Z, Y, Op1->getName()));
1744 
1745     // (X - (X & Y))   -->   (X & ~Y)
1746     if (match(Op1, m_c_And(m_Value(Y), m_Specific(Op0))))
1747       return BinaryOperator::CreateAnd(Op0,
1748                                   Builder.CreateNot(Y, Y->getName() + ".not"));
1749 
1750     // 0 - (X sdiv C)  -> (X sdiv -C)  provided the negation doesn't overflow.
1751     if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1752         C->isNotMinSignedValue() && !C->isOneValue())
1753       return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1754 
1755     // 0 - (X << Y)  -> (-X << Y)   when X is freely negatable.
1756     if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1757       if (Value *XNeg = dyn_castNegVal(X))
1758         return BinaryOperator::CreateShl(XNeg, Y);
1759 
1760     // Subtracting -1/0 is the same as adding 1/0:
1761     // sub [nsw] Op0, sext(bool Y) -> add [nsw] Op0, zext(bool Y)
1762     // 'nuw' is dropped in favor of the canonical form.
1763     if (match(Op1, m_SExt(m_Value(Y))) &&
1764         Y->getType()->getScalarSizeInBits() == 1) {
1765       Value *Zext = Builder.CreateZExt(Y, I.getType());
1766       BinaryOperator *Add = BinaryOperator::CreateAdd(Op0, Zext);
1767       Add->setHasNoSignedWrap(I.hasNoSignedWrap());
1768       return Add;
1769     }
1770 
1771     // X - A*-B -> X + A*B
1772     // X - -A*B -> X + A*B
1773     Value *A, *B;
1774     Constant *CI;
1775     if (match(Op1, m_c_Mul(m_Value(A), m_Neg(m_Value(B)))))
1776       return BinaryOperator::CreateAdd(Op0, Builder.CreateMul(A, B));
1777 
1778     // X - A*CI -> X + A*-CI
1779     // No need to handle commuted multiply because multiply handling will
1780     // ensure constant will be move to the right hand side.
1781     if (match(Op1, m_Mul(m_Value(A), m_Constant(CI)))) {
1782       Value *NewMul = Builder.CreateMul(A, ConstantExpr::getNeg(CI));
1783       return BinaryOperator::CreateAdd(Op0, NewMul);
1784     }
1785   }
1786 
1787   // Optimize pointer differences into the same array into a size.  Consider:
1788   //  &A[10] - &A[0]: we should compile this to "10".
1789   Value *LHSOp, *RHSOp;
1790   if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1791       match(Op1, m_PtrToInt(m_Value(RHSOp))))
1792     if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1793       return replaceInstUsesWith(I, Res);
1794 
1795   // trunc(p)-trunc(q) -> trunc(p-q)
1796   if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1797       match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1798     if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1799       return replaceInstUsesWith(I, Res);
1800 
1801   // Canonicalize a shifty way to code absolute value to the common pattern.
1802   // There are 2 potential commuted variants.
1803   // We're relying on the fact that we only do this transform when the shift has
1804   // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
1805   // instructions).
1806   Value *A;
1807   const APInt *ShAmt;
1808   Type *Ty = I.getType();
1809   if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
1810       Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
1811       match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
1812     // B = ashr i32 A, 31 ; smear the sign bit
1813     // sub (xor A, B), B  ; flip bits if negative and subtract -1 (add 1)
1814     // --> (A < 0) ? -A : A
1815     Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
1816     // Copy the nuw/nsw flags from the sub to the negate.
1817     Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
1818                                    I.hasNoSignedWrap());
1819     return SelectInst::Create(Cmp, Neg, A);
1820   }
1821 
1822   bool Changed = false;
1823   if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1824     Changed = true;
1825     I.setHasNoSignedWrap(true);
1826   }
1827   if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1828     Changed = true;
1829     I.setHasNoUnsignedWrap(true);
1830   }
1831 
1832   return Changed ? &I : nullptr;
1833 }
1834 
visitFSub(BinaryOperator & I)1835 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
1836   if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
1837                                   I.getFastMathFlags(),
1838                                   SQ.getWithInstruction(&I)))
1839     return replaceInstUsesWith(I, V);
1840 
1841   if (Instruction *X = foldShuffledBinop(I))
1842     return X;
1843 
1844   // Subtraction from -0.0 is the canonical form of fneg.
1845   // fsub nsz 0, X ==> fsub nsz -0.0, X
1846   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1847   if (I.hasNoSignedZeros() && match(Op0, m_PosZeroFP()))
1848     return BinaryOperator::CreateFNegFMF(Op1, &I);
1849 
1850   // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
1851   // Canonicalize to fadd to make analysis easier.
1852   // This can also help codegen because fadd is commutative.
1853   // Note that if this fsub was really an fneg, the fadd with -0.0 will get
1854   // killed later. We still limit that particular transform with 'hasOneUse'
1855   // because an fneg is assumed better/cheaper than a generic fsub.
1856   Value *X, *Y;
1857   if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
1858     if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
1859       Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
1860       return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
1861     }
1862   }
1863 
1864   if (isa<Constant>(Op0))
1865     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1866       if (Instruction *NV = FoldOpIntoSelect(I, SI))
1867         return NV;
1868 
1869   // X - C --> X + (-C)
1870   // But don't transform constant expressions because there's an inverse fold
1871   // for X + (-Y) --> X - Y.
1872   Constant *C;
1873   if (match(Op1, m_Constant(C)) && !isa<ConstantExpr>(Op1))
1874     return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
1875 
1876   // X - (-Y) --> X + Y
1877   if (match(Op1, m_FNeg(m_Value(Y))))
1878     return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
1879 
1880   // Similar to above, but look through a cast of the negated value:
1881   // X - (fptrunc(-Y)) --> X + fptrunc(Y)
1882   if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y)))))) {
1883     Value *TruncY = Builder.CreateFPTrunc(Y, I.getType());
1884     return BinaryOperator::CreateFAddFMF(Op0, TruncY, &I);
1885   }
1886   // X - (fpext(-Y)) --> X + fpext(Y)
1887   if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y)))))) {
1888     Value *ExtY = Builder.CreateFPExt(Y, I.getType());
1889     return BinaryOperator::CreateFAddFMF(Op0, ExtY, &I);
1890   }
1891 
1892   // Handle specials cases for FSub with selects feeding the operation
1893   if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
1894     return replaceInstUsesWith(I, V);
1895 
1896   if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1897     if (Value *V = FAddCombine(Builder).simplify(&I))
1898       return replaceInstUsesWith(I, V);
1899   }
1900 
1901   return nullptr;
1902 }
1903