1//===- InstCombineAddSub.cpp ----------------------------------------------===//
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 "InstCombine.h"
15#include "llvm/ADT/STLExtras.h"
16#include "llvm/Analysis/InstructionSimplify.h"
17#include "llvm/IR/DataLayout.h"
18#include "llvm/IR/GetElementPtrTypeIterator.h"
19#include "llvm/IR/PatternMatch.h"
20using namespace llvm;
21using namespace PatternMatch;
22
23#define DEBUG_TYPE "instcombine"
24
25namespace {
26
27  /// Class representing coefficient of floating-point addend.
28  /// This class needs to be highly efficient, which is especially true for
29  /// the constructor. As of I write this comment, the cost of the default
30  /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
31  /// perform write-merging).
32  ///
33  class FAddendCoef {
34  public:
35    // The constructor has to initialize a APFloat, which is uncessary for
36    // most addends which have coefficient either 1 or -1. So, the constructor
37    // is expensive. In order to avoid the cost of the constructor, we should
38    // reuse some instances whenever possible. The pre-created instances
39    // FAddCombine::Add[0-5] embodies this idea.
40    //
41    FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {}
42    ~FAddendCoef();
43
44    void set(short C) {
45      assert(!insaneIntVal(C) && "Insane coefficient");
46      IsFp = false; IntVal = C;
47    }
48
49    void set(const APFloat& C);
50
51    void negate();
52
53    bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
54    Value *getValue(Type *) const;
55
56    // If possible, don't define operator+/operator- etc because these
57    // operators inevitably call FAddendCoef's constructor which is not cheap.
58    void operator=(const FAddendCoef &A);
59    void operator+=(const FAddendCoef &A);
60    void operator-=(const FAddendCoef &A);
61    void operator*=(const FAddendCoef &S);
62
63    bool isOne() const { return isInt() && IntVal == 1; }
64    bool isTwo() const { return isInt() && IntVal == 2; }
65    bool isMinusOne() const { return isInt() && IntVal == -1; }
66    bool isMinusTwo() const { return isInt() && IntVal == -2; }
67
68  private:
69    bool insaneIntVal(int V) { return V > 4 || V < -4; }
70    APFloat *getFpValPtr(void)
71      { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); }
72    const APFloat *getFpValPtr(void) const
73      { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); }
74
75    const APFloat &getFpVal(void) const {
76      assert(IsFp && BufHasFpVal && "Incorret state");
77      return *getFpValPtr();
78    }
79
80    APFloat &getFpVal(void) {
81      assert(IsFp && BufHasFpVal && "Incorret state");
82      return *getFpValPtr();
83    }
84
85    bool isInt() const { return !IsFp; }
86
87    // If the coefficient is represented by an integer, promote it to a
88    // floating point.
89    void convertToFpType(const fltSemantics &Sem);
90
91    // Construct an APFloat from a signed integer.
92    // TODO: We should get rid of this function when APFloat can be constructed
93    //       from an *SIGNED* integer.
94    APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
95  private:
96
97    bool IsFp;
98
99    // True iff FpValBuf contains an instance of APFloat.
100    bool BufHasFpVal;
101
102    // The integer coefficient of an individual addend is either 1 or -1,
103    // and we try to simplify at most 4 addends from neighboring at most
104    // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
105    // is overkill of this end.
106    short IntVal;
107
108    AlignedCharArrayUnion<APFloat> FpValBuf;
109  };
110
111  /// FAddend is used to represent floating-point addend. An addend is
112  /// represented as <C, V>, where the V is a symbolic value, and C is a
113  /// constant coefficient. A constant addend is represented as <C, 0>.
114  ///
115  class FAddend {
116  public:
117    FAddend() { Val = nullptr; }
118
119    Value *getSymVal (void) const { return Val; }
120    const FAddendCoef &getCoef(void) const { return Coeff; }
121
122    bool isConstant() const { return Val == nullptr; }
123    bool isZero() const { return Coeff.isZero(); }
124
125    void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; }
126    void set(const APFloat& Coefficient, Value *V)
127      { Coeff.set(Coefficient); Val = V; }
128    void set(const ConstantFP* Coefficient, Value *V)
129      { Coeff.set(Coefficient->getValueAPF()); Val = V; }
130
131    void negate() { Coeff.negate(); }
132
133    /// Drill down the U-D chain one step to find the definition of V, and
134    /// try to break the definition into one or two addends.
135    static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
136
137    /// Similar to FAddend::drillDownOneStep() except that the value being
138    /// splitted is the addend itself.
139    unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
140
141    void operator+=(const FAddend &T) {
142      assert((Val == T.Val) && "Symbolic-values disagree");
143      Coeff += T.Coeff;
144    }
145
146  private:
147    void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
148
149    // This addend has the value of "Coeff * Val".
150    Value *Val;
151    FAddendCoef Coeff;
152  };
153
154  /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
155  /// with its neighboring at most two instructions.
156  ///
157  class FAddCombine {
158  public:
159    FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(nullptr) {}
160    Value *simplify(Instruction *FAdd);
161
162  private:
163    typedef SmallVector<const FAddend*, 4> AddendVect;
164
165    Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
166
167    Value *performFactorization(Instruction *I);
168
169    /// Convert given addend to a Value
170    Value *createAddendVal(const FAddend &A, bool& NeedNeg);
171
172    /// Return the number of instructions needed to emit the N-ary addition.
173    unsigned calcInstrNumber(const AddendVect& Vect);
174    Value *createFSub(Value *Opnd0, Value *Opnd1);
175    Value *createFAdd(Value *Opnd0, Value *Opnd1);
176    Value *createFMul(Value *Opnd0, Value *Opnd1);
177    Value *createFDiv(Value *Opnd0, Value *Opnd1);
178    Value *createFNeg(Value *V);
179    Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
180    void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
181
182    InstCombiner::BuilderTy *Builder;
183    Instruction *Instr;
184
185  private:
186     // Debugging stuff are clustered here.
187    #ifndef NDEBUG
188      unsigned CreateInstrNum;
189      void initCreateInstNum() { CreateInstrNum = 0; }
190      void incCreateInstNum() { CreateInstrNum++; }
191    #else
192      void initCreateInstNum() {}
193      void incCreateInstNum() {}
194    #endif
195  };
196}
197
198//===----------------------------------------------------------------------===//
199//
200// Implementation of
201//    {FAddendCoef, FAddend, FAddition, FAddCombine}.
202//
203//===----------------------------------------------------------------------===//
204FAddendCoef::~FAddendCoef() {
205  if (BufHasFpVal)
206    getFpValPtr()->~APFloat();
207}
208
209void FAddendCoef::set(const APFloat& C) {
210  APFloat *P = getFpValPtr();
211
212  if (isInt()) {
213    // As the buffer is meanless byte stream, we cannot call
214    // APFloat::operator=().
215    new(P) APFloat(C);
216  } else
217    *P = C;
218
219  IsFp = BufHasFpVal = true;
220}
221
222void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
223  if (!isInt())
224    return;
225
226  APFloat *P = getFpValPtr();
227  if (IntVal > 0)
228    new(P) APFloat(Sem, IntVal);
229  else {
230    new(P) APFloat(Sem, 0 - IntVal);
231    P->changeSign();
232  }
233  IsFp = BufHasFpVal = true;
234}
235
236APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
237  if (Val >= 0)
238    return APFloat(Sem, Val);
239
240  APFloat T(Sem, 0 - Val);
241  T.changeSign();
242
243  return T;
244}
245
246void FAddendCoef::operator=(const FAddendCoef &That) {
247  if (That.isInt())
248    set(That.IntVal);
249  else
250    set(That.getFpVal());
251}
252
253void FAddendCoef::operator+=(const FAddendCoef &That) {
254  enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
255  if (isInt() == That.isInt()) {
256    if (isInt())
257      IntVal += That.IntVal;
258    else
259      getFpVal().add(That.getFpVal(), RndMode);
260    return;
261  }
262
263  if (isInt()) {
264    const APFloat &T = That.getFpVal();
265    convertToFpType(T.getSemantics());
266    getFpVal().add(T, RndMode);
267    return;
268  }
269
270  APFloat &T = getFpVal();
271  T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
272}
273
274void FAddendCoef::operator-=(const FAddendCoef &That) {
275  enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
276  if (isInt() == That.isInt()) {
277    if (isInt())
278      IntVal -= That.IntVal;
279    else
280      getFpVal().subtract(That.getFpVal(), RndMode);
281    return;
282  }
283
284  if (isInt()) {
285    const APFloat &T = That.getFpVal();
286    convertToFpType(T.getSemantics());
287    getFpVal().subtract(T, RndMode);
288    return;
289  }
290
291  APFloat &T = getFpVal();
292  T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode);
293}
294
295void FAddendCoef::operator*=(const FAddendCoef &That) {
296  if (That.isOne())
297    return;
298
299  if (That.isMinusOne()) {
300    negate();
301    return;
302  }
303
304  if (isInt() && That.isInt()) {
305    int Res = IntVal * (int)That.IntVal;
306    assert(!insaneIntVal(Res) && "Insane int value");
307    IntVal = Res;
308    return;
309  }
310
311  const fltSemantics &Semantic =
312    isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
313
314  if (isInt())
315    convertToFpType(Semantic);
316  APFloat &F0 = getFpVal();
317
318  if (That.isInt())
319    F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
320                APFloat::rmNearestTiesToEven);
321  else
322    F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
323
324  return;
325}
326
327void FAddendCoef::negate() {
328  if (isInt())
329    IntVal = 0 - IntVal;
330  else
331    getFpVal().changeSign();
332}
333
334Value *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
350//
351unsigned FAddend::drillValueDownOneStep
352  (Value *Val, FAddend &Addend0, FAddend &Addend1) {
353  Instruction *I = nullptr;
354  if (!Val || !(I = dyn_cast<Instruction>(Val)))
355    return 0;
356
357  unsigned Opcode = I->getOpcode();
358
359  if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
360    ConstantFP *C0, *C1;
361    Value *Opnd0 = I->getOperand(0);
362    Value *Opnd1 = I->getOperand(1);
363    if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
364      Opnd0 = nullptr;
365
366    if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
367      Opnd1 = nullptr;
368
369    if (Opnd0) {
370      if (!C0)
371        Addend0.set(1, Opnd0);
372      else
373        Addend0.set(C0, nullptr);
374    }
375
376    if (Opnd1) {
377      FAddend &Addend = Opnd0 ? Addend1 : Addend0;
378      if (!C1)
379        Addend.set(1, Opnd1);
380      else
381        Addend.set(C1, nullptr);
382      if (Opcode == Instruction::FSub)
383        Addend.negate();
384    }
385
386    if (Opnd0 || Opnd1)
387      return Opnd0 && Opnd1 ? 2 : 1;
388
389    // Both operands are zero. Weird!
390    Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
391    return 1;
392  }
393
394  if (I->getOpcode() == Instruction::FMul) {
395    Value *V0 = I->getOperand(0);
396    Value *V1 = I->getOperand(1);
397    if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
398      Addend0.set(C, V1);
399      return 1;
400    }
401
402    if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
403      Addend0.set(C, V0);
404      return 1;
405    }
406  }
407
408  return 0;
409}
410
411// Try to break *this* addend into two addends. e.g. Suppose this addend is
412// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
413// i.e. <2.3, X> and <2.3, Y>.
414//
415unsigned FAddend::drillAddendDownOneStep
416  (FAddend &Addend0, FAddend &Addend1) const {
417  if (isConstant())
418    return 0;
419
420  unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
421  if (!BreakNum || Coeff.isOne())
422    return BreakNum;
423
424  Addend0.Scale(Coeff);
425
426  if (BreakNum == 2)
427    Addend1.Scale(Coeff);
428
429  return BreakNum;
430}
431
432// Try to perform following optimization on the input instruction I. Return the
433// simplified expression if was successful; otherwise, return 0.
434//
435//   Instruction "I" is                Simplified into
436// -------------------------------------------------------
437//   (x * y) +/- (x * z)               x * (y +/- z)
438//   (y / x) +/- (z / x)               (y +/- z) / x
439//
440Value *FAddCombine::performFactorization(Instruction *I) {
441  assert((I->getOpcode() == Instruction::FAdd ||
442          I->getOpcode() == Instruction::FSub) && "Expect add/sub");
443
444  Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0));
445  Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1));
446
447  if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
448    return nullptr;
449
450  bool isMpy = false;
451  if (I0->getOpcode() == Instruction::FMul)
452    isMpy = true;
453  else if (I0->getOpcode() != Instruction::FDiv)
454    return nullptr;
455
456  Value *Opnd0_0 = I0->getOperand(0);
457  Value *Opnd0_1 = I0->getOperand(1);
458  Value *Opnd1_0 = I1->getOperand(0);
459  Value *Opnd1_1 = I1->getOperand(1);
460
461  //  Input Instr I       Factor   AddSub0  AddSub1
462  //  ----------------------------------------------
463  // (x*y) +/- (x*z)        x        y         z
464  // (y/x) +/- (z/x)        x        y         z
465  //
466  Value *Factor = nullptr;
467  Value *AddSub0 = nullptr, *AddSub1 = nullptr;
468
469  if (isMpy) {
470    if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
471      Factor = Opnd0_0;
472    else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
473      Factor = Opnd0_1;
474
475    if (Factor) {
476      AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
477      AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
478    }
479  } else if (Opnd0_1 == Opnd1_1) {
480    Factor = Opnd0_1;
481    AddSub0 = Opnd0_0;
482    AddSub1 = Opnd1_0;
483  }
484
485  if (!Factor)
486    return nullptr;
487
488  FastMathFlags Flags;
489  Flags.setUnsafeAlgebra();
490  if (I0) Flags &= I->getFastMathFlags();
491  if (I1) Flags &= I->getFastMathFlags();
492
493  // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
494  Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
495                      createFAdd(AddSub0, AddSub1) :
496                      createFSub(AddSub0, AddSub1);
497  if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
498    const APFloat &F = CFP->getValueAPF();
499    if (!F.isNormal())
500      return nullptr;
501  } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
502    II->setFastMathFlags(Flags);
503
504  if (isMpy) {
505    Value *RI = createFMul(Factor, NewAddSub);
506    if (Instruction *II = dyn_cast<Instruction>(RI))
507      II->setFastMathFlags(Flags);
508    return RI;
509  }
510
511  Value *RI = createFDiv(NewAddSub, Factor);
512  if (Instruction *II = dyn_cast<Instruction>(RI))
513    II->setFastMathFlags(Flags);
514  return RI;
515}
516
517Value *FAddCombine::simplify(Instruction *I) {
518  assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode");
519
520  // Currently we are not able to handle vector type.
521  if (I->getType()->isVectorTy())
522    return nullptr;
523
524  assert((I->getOpcode() == Instruction::FAdd ||
525          I->getOpcode() == Instruction::FSub) && "Expect add/sub");
526
527  // Save the instruction before calling other member-functions.
528  Instr = I;
529
530  FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
531
532  unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
533
534  // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
535  unsigned Opnd0_ExpNum = 0;
536  unsigned Opnd1_ExpNum = 0;
537
538  if (!Opnd0.isConstant())
539    Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
540
541  // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
542  if (OpndNum == 2 && !Opnd1.isConstant())
543    Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
544
545  // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
546  if (Opnd0_ExpNum && Opnd1_ExpNum) {
547    AddendVect AllOpnds;
548    AllOpnds.push_back(&Opnd0_0);
549    AllOpnds.push_back(&Opnd1_0);
550    if (Opnd0_ExpNum == 2)
551      AllOpnds.push_back(&Opnd0_1);
552    if (Opnd1_ExpNum == 2)
553      AllOpnds.push_back(&Opnd1_1);
554
555    // Compute instruction quota. We should save at least one instruction.
556    unsigned InstQuota = 0;
557
558    Value *V0 = I->getOperand(0);
559    Value *V1 = I->getOperand(1);
560    InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
561                 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
562
563    if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
564      return R;
565  }
566
567  if (OpndNum != 2) {
568    // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
569    // splitted into two addends, say "V = X - Y", the instruction would have
570    // been optimized into "I = Y - X" in the previous steps.
571    //
572    const FAddendCoef &CE = Opnd0.getCoef();
573    return CE.isOne() ? Opnd0.getSymVal() : nullptr;
574  }
575
576  // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
577  if (Opnd1_ExpNum) {
578    AddendVect AllOpnds;
579    AllOpnds.push_back(&Opnd0);
580    AllOpnds.push_back(&Opnd1_0);
581    if (Opnd1_ExpNum == 2)
582      AllOpnds.push_back(&Opnd1_1);
583
584    if (Value *R = simplifyFAdd(AllOpnds, 1))
585      return R;
586  }
587
588  // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
589  if (Opnd0_ExpNum) {
590    AddendVect AllOpnds;
591    AllOpnds.push_back(&Opnd1);
592    AllOpnds.push_back(&Opnd0_0);
593    if (Opnd0_ExpNum == 2)
594      AllOpnds.push_back(&Opnd0_1);
595
596    if (Value *R = simplifyFAdd(AllOpnds, 1))
597      return R;
598  }
599
600  // step 6: Try factorization as the last resort,
601  return performFactorization(I);
602}
603
604Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
605
606  unsigned AddendNum = Addends.size();
607  assert(AddendNum <= 4 && "Too many addends");
608
609  // For saving intermediate results;
610  unsigned NextTmpIdx = 0;
611  FAddend TmpResult[3];
612
613  // Points to the constant addend of the resulting simplified expression.
614  // If the resulting expr has constant-addend, this constant-addend is
615  // desirable to reside at the top of the resulting expression tree. Placing
616  // constant close to supper-expr(s) will potentially reveal some optimization
617  // opportunities in super-expr(s).
618  //
619  const FAddend *ConstAdd = nullptr;
620
621  // Simplified addends are placed <SimpVect>.
622  AddendVect SimpVect;
623
624  // The outer loop works on one symbolic-value at a time. Suppose the input
625  // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
626  // The symbolic-values will be processed in this order: x, y, z.
627  //
628  for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
629
630    const FAddend *ThisAddend = Addends[SymIdx];
631    if (!ThisAddend) {
632      // This addend was processed before.
633      continue;
634    }
635
636    Value *Val = ThisAddend->getSymVal();
637    unsigned StartIdx = SimpVect.size();
638    SimpVect.push_back(ThisAddend);
639
640    // The inner loop collects addends sharing same symbolic-value, and these
641    // addends will be later on folded into a single addend. Following above
642    // example, if the symbolic value "y" is being processed, the inner loop
643    // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
644    // be later on folded into "<b1+b2, y>".
645    //
646    for (unsigned SameSymIdx = SymIdx + 1;
647         SameSymIdx < AddendNum; SameSymIdx++) {
648      const FAddend *T = Addends[SameSymIdx];
649      if (T && T->getSymVal() == Val) {
650        // Set null such that next iteration of the outer loop will not process
651        // this addend again.
652        Addends[SameSymIdx] = nullptr;
653        SimpVect.push_back(T);
654      }
655    }
656
657    // If multiple addends share same symbolic value, fold them together.
658    if (StartIdx + 1 != SimpVect.size()) {
659      FAddend &R = TmpResult[NextTmpIdx ++];
660      R = *SimpVect[StartIdx];
661      for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
662        R += *SimpVect[Idx];
663
664      // Pop all addends being folded and push the resulting folded addend.
665      SimpVect.resize(StartIdx);
666      if (Val) {
667        if (!R.isZero()) {
668          SimpVect.push_back(&R);
669        }
670      } else {
671        // Don't push constant addend at this time. It will be the last element
672        // of <SimpVect>.
673        ConstAdd = &R;
674      }
675    }
676  }
677
678  assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
679         "out-of-bound access");
680
681  if (ConstAdd)
682    SimpVect.push_back(ConstAdd);
683
684  Value *Result;
685  if (!SimpVect.empty())
686    Result = createNaryFAdd(SimpVect, InstrQuota);
687  else {
688    // The addition is folded to 0.0.
689    Result = ConstantFP::get(Instr->getType(), 0.0);
690  }
691
692  return Result;
693}
694
695Value *FAddCombine::createNaryFAdd
696  (const AddendVect &Opnds, unsigned InstrQuota) {
697  assert(!Opnds.empty() && "Expect at least one addend");
698
699  // Step 1: Check if the # of instructions needed exceeds the quota.
700  //
701  unsigned InstrNeeded = calcInstrNumber(Opnds);
702  if (InstrNeeded > InstrQuota)
703    return nullptr;
704
705  initCreateInstNum();
706
707  // step 2: Emit the N-ary addition.
708  // Note that at most three instructions are involved in Fadd-InstCombine: the
709  // addition in question, and at most two neighboring instructions.
710  // The resulting optimized addition should have at least one less instruction
711  // than the original addition expression tree. This implies that the resulting
712  // N-ary addition has at most two instructions, and we don't need to worry
713  // about tree-height when constructing the N-ary addition.
714
715  Value *LastVal = nullptr;
716  bool LastValNeedNeg = false;
717
718  // Iterate the addends, creating fadd/fsub using adjacent two addends.
719  for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end();
720       I != E; I++) {
721    bool NeedNeg;
722    Value *V = createAddendVal(**I, NeedNeg);
723    if (!LastVal) {
724      LastVal = V;
725      LastValNeedNeg = NeedNeg;
726      continue;
727    }
728
729    if (LastValNeedNeg == NeedNeg) {
730      LastVal = createFAdd(LastVal, V);
731      continue;
732    }
733
734    if (LastValNeedNeg)
735      LastVal = createFSub(V, LastVal);
736    else
737      LastVal = createFSub(LastVal, V);
738
739    LastValNeedNeg = false;
740  }
741
742  if (LastValNeedNeg) {
743    LastVal = createFNeg(LastVal);
744  }
745
746  #ifndef NDEBUG
747    assert(CreateInstrNum == InstrNeeded &&
748           "Inconsistent in instruction numbers");
749  #endif
750
751  return LastVal;
752}
753
754Value *FAddCombine::createFSub
755  (Value *Opnd0, Value *Opnd1) {
756  Value *V = Builder->CreateFSub(Opnd0, Opnd1);
757  if (Instruction *I = dyn_cast<Instruction>(V))
758    createInstPostProc(I);
759  return V;
760}
761
762Value *FAddCombine::createFNeg(Value *V) {
763  Value *Zero = cast<Value>(ConstantFP::get(V->getType(), 0.0));
764  Value *NewV = createFSub(Zero, V);
765  if (Instruction *I = dyn_cast<Instruction>(NewV))
766    createInstPostProc(I, true); // fneg's don't receive instruction numbers.
767  return NewV;
768}
769
770Value *FAddCombine::createFAdd
771  (Value *Opnd0, Value *Opnd1) {
772  Value *V = Builder->CreateFAdd(Opnd0, Opnd1);
773  if (Instruction *I = dyn_cast<Instruction>(V))
774    createInstPostProc(I);
775  return V;
776}
777
778Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
779  Value *V = Builder->CreateFMul(Opnd0, Opnd1);
780  if (Instruction *I = dyn_cast<Instruction>(V))
781    createInstPostProc(I);
782  return V;
783}
784
785Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) {
786  Value *V = Builder->CreateFDiv(Opnd0, Opnd1);
787  if (Instruction *I = dyn_cast<Instruction>(V))
788    createInstPostProc(I);
789  return V;
790}
791
792void FAddCombine::createInstPostProc(Instruction *NewInstr,
793                                     bool NoNumber) {
794  NewInstr->setDebugLoc(Instr->getDebugLoc());
795
796  // Keep track of the number of instruction created.
797  if (!NoNumber)
798    incCreateInstNum();
799
800  // Propagate fast-math flags
801  NewInstr->setFastMathFlags(Instr->getFastMathFlags());
802}
803
804// Return the number of instruction needed to emit the N-ary addition.
805// NOTE: Keep this function in sync with createAddendVal().
806unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
807  unsigned OpndNum = Opnds.size();
808  unsigned InstrNeeded = OpndNum - 1;
809
810  // The number of addends in the form of "(-1)*x".
811  unsigned NegOpndNum = 0;
812
813  // Adjust the number of instructions needed to emit the N-ary add.
814  for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end();
815       I != E; I++) {
816    const FAddend *Opnd = *I;
817    if (Opnd->isConstant())
818      continue;
819
820    const FAddendCoef &CE = Opnd->getCoef();
821    if (CE.isMinusOne() || CE.isMinusTwo())
822      NegOpndNum++;
823
824    // Let the addend be "c * x". If "c == +/-1", the value of the addend
825    // is immediately available; otherwise, it needs exactly one instruction
826    // to evaluate the value.
827    if (!CE.isMinusOne() && !CE.isOne())
828      InstrNeeded++;
829  }
830  if (NegOpndNum == OpndNum)
831    InstrNeeded++;
832  return InstrNeeded;
833}
834
835// Input Addend        Value           NeedNeg(output)
836// ================================================================
837// Constant C          C               false
838// <+/-1, V>           V               coefficient is -1
839// <2/-2, V>          "fadd V, V"      coefficient is -2
840// <C, V>             "fmul V, C"      false
841//
842// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
843Value *FAddCombine::createAddendVal
844  (const FAddend &Opnd, bool &NeedNeg) {
845  const FAddendCoef &Coeff = Opnd.getCoef();
846
847  if (Opnd.isConstant()) {
848    NeedNeg = false;
849    return Coeff.getValue(Instr->getType());
850  }
851
852  Value *OpndVal = Opnd.getSymVal();
853
854  if (Coeff.isMinusOne() || Coeff.isOne()) {
855    NeedNeg = Coeff.isMinusOne();
856    return OpndVal;
857  }
858
859  if (Coeff.isTwo() || Coeff.isMinusTwo()) {
860    NeedNeg = Coeff.isMinusTwo();
861    return createFAdd(OpndVal, OpndVal);
862  }
863
864  NeedNeg = false;
865  return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
866}
867
868// If one of the operands only has one non-zero bit, and if the other
869// operand has a known-zero bit in a more significant place than it (not
870// including the sign bit) the ripple may go up to and fill the zero, but
871// won't change the sign. For example, (X & ~4) + 1.
872static bool checkRippleForAdd(const APInt &Op0KnownZero,
873                              const APInt &Op1KnownZero) {
874  APInt Op1MaybeOne = ~Op1KnownZero;
875  // Make sure that one of the operand has at most one bit set to 1.
876  if (Op1MaybeOne.countPopulation() != 1)
877    return false;
878
879  // Find the most significant known 0 other than the sign bit.
880  int BitWidth = Op0KnownZero.getBitWidth();
881  APInt Op0KnownZeroTemp(Op0KnownZero);
882  Op0KnownZeroTemp.clearBit(BitWidth - 1);
883  int Op0ZeroPosition = BitWidth - Op0KnownZeroTemp.countLeadingZeros() - 1;
884
885  int Op1OnePosition = BitWidth - Op1MaybeOne.countLeadingZeros() - 1;
886  assert(Op1OnePosition >= 0);
887
888  // This also covers the case of no known zero, since in that case
889  // Op0ZeroPosition is -1.
890  return Op0ZeroPosition >= Op1OnePosition;
891}
892
893/// WillNotOverflowSignedAdd - Return true if we can prove that:
894///    (sext (add LHS, RHS))  === (add (sext LHS), (sext RHS))
895/// This basically requires proving that the add in the original type would not
896/// overflow to change the sign bit or have a carry out.
897/// TODO: Handle this for Vectors.
898bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
899  // There are different heuristics we can use for this.  Here are some simple
900  // ones.
901
902  // If LHS and RHS each have at least two sign bits, the addition will look
903  // like
904  //
905  // XX..... +
906  // YY.....
907  //
908  // If the carry into the most significant position is 0, X and Y can't both
909  // be 1 and therefore the carry out of the addition is also 0.
910  //
911  // If the carry into the most significant position is 1, X and Y can't both
912  // be 0 and therefore the carry out of the addition is also 1.
913  //
914  // Since the carry into the most significant position is always equal to
915  // the carry out of the addition, there is no signed overflow.
916  if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
917    return true;
918
919  if (IntegerType *IT = dyn_cast<IntegerType>(LHS->getType())) {
920    int BitWidth = IT->getBitWidth();
921    APInt LHSKnownZero(BitWidth, 0);
922    APInt LHSKnownOne(BitWidth, 0);
923    computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
924
925    APInt RHSKnownZero(BitWidth, 0);
926    APInt RHSKnownOne(BitWidth, 0);
927    computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
928
929    // Addition of two 2's compliment numbers having opposite signs will never
930    // overflow.
931    if ((LHSKnownOne[BitWidth - 1] && RHSKnownZero[BitWidth - 1]) ||
932        (LHSKnownZero[BitWidth - 1] && RHSKnownOne[BitWidth - 1]))
933      return true;
934
935    // Check if carry bit of addition will not cause overflow.
936    if (checkRippleForAdd(LHSKnownZero, RHSKnownZero))
937      return true;
938    if (checkRippleForAdd(RHSKnownZero, LHSKnownZero))
939      return true;
940  }
941  return false;
942}
943
944/// WillNotOverflowUnsignedAdd - Return true if we can prove that:
945///    (zext (add LHS, RHS))  === (add (zext LHS), (zext RHS))
946bool InstCombiner::WillNotOverflowUnsignedAdd(Value *LHS, Value *RHS) {
947  // There are different heuristics we can use for this. Here is a simple one.
948  // If the sign bit of LHS and that of RHS are both zero, no unsigned wrap.
949  bool LHSKnownNonNegative, LHSKnownNegative;
950  bool RHSKnownNonNegative, RHSKnownNegative;
951  ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, 0);
952  ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, 0);
953  if (LHSKnownNonNegative && RHSKnownNonNegative)
954    return true;
955
956  return false;
957}
958
959// Checks if any operand is negative and we can convert add to sub.
960// This function checks for following negative patterns
961//   ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
962//   ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
963//   XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
964static Value *checkForNegativeOperand(BinaryOperator &I,
965                                      InstCombiner::BuilderTy *Builder) {
966  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
967
968  // This function creates 2 instructions to replace ADD, we need at least one
969  // of LHS or RHS to have one use to ensure benefit in transform.
970  if (!LHS->hasOneUse() && !RHS->hasOneUse())
971    return nullptr;
972
973  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
974  const APInt *C1 = nullptr, *C2 = nullptr;
975
976  // if ONE is on other side, swap
977  if (match(RHS, m_Add(m_Value(X), m_One())))
978    std::swap(LHS, RHS);
979
980  if (match(LHS, m_Add(m_Value(X), m_One()))) {
981    // if XOR on other side, swap
982    if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
983      std::swap(X, RHS);
984
985    if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
986      // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
987      // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
988      if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
989        Value *NewAnd = Builder->CreateAnd(Z, *C1);
990        return Builder->CreateSub(RHS, NewAnd, "sub");
991      } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
992        // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
993        // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
994        Value *NewOr = Builder->CreateOr(Z, ~(*C1));
995        return Builder->CreateSub(RHS, NewOr, "sub");
996      }
997    }
998  }
999
1000  // Restore LHS and RHS
1001  LHS = I.getOperand(0);
1002  RHS = I.getOperand(1);
1003
1004  // if XOR is on other side, swap
1005  if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
1006    std::swap(LHS, RHS);
1007
1008  // C2 is ODD
1009  // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
1010  // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
1011  if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
1012    if (C1->countTrailingZeros() == 0)
1013      if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
1014        Value *NewOr = Builder->CreateOr(Z, ~(*C2));
1015        return Builder->CreateSub(RHS, NewOr, "sub");
1016      }
1017  return nullptr;
1018}
1019
1020Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1021   bool Changed = SimplifyAssociativeOrCommutative(I);
1022   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1023
1024   if (Value *V = SimplifyVectorOp(I))
1025     return ReplaceInstUsesWith(I, V);
1026
1027   if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
1028                                  I.hasNoUnsignedWrap(), DL))
1029     return ReplaceInstUsesWith(I, V);
1030
1031   // (A*B)+(A*C) -> A*(B+C) etc
1032  if (Value *V = SimplifyUsingDistributiveLaws(I))
1033    return ReplaceInstUsesWith(I, V);
1034
1035  if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1036    // X + (signbit) --> X ^ signbit
1037    const APInt &Val = CI->getValue();
1038    if (Val.isSignBit())
1039      return BinaryOperator::CreateXor(LHS, RHS);
1040
1041    // See if SimplifyDemandedBits can simplify this.  This handles stuff like
1042    // (X & 254)+1 -> (X&254)|1
1043    if (SimplifyDemandedInstructionBits(I))
1044      return &I;
1045
1046    // zext(bool) + C -> bool ? C + 1 : C
1047    if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
1048      if (ZI->getSrcTy()->isIntegerTy(1))
1049        return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
1050
1051    Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1052    if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1053      uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
1054      const APInt &RHSVal = CI->getValue();
1055      unsigned ExtendAmt = 0;
1056      // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1057      // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1058      if (XorRHS->getValue() == -RHSVal) {
1059        if (RHSVal.isPowerOf2())
1060          ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1061        else if (XorRHS->getValue().isPowerOf2())
1062          ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1063      }
1064
1065      if (ExtendAmt) {
1066        APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1067        if (!MaskedValueIsZero(XorLHS, Mask))
1068          ExtendAmt = 0;
1069      }
1070
1071      if (ExtendAmt) {
1072        Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt);
1073        Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext");
1074        return BinaryOperator::CreateAShr(NewShl, ShAmt);
1075      }
1076
1077      // If this is a xor that was canonicalized from a sub, turn it back into
1078      // a sub and fuse this add with it.
1079      if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1080        IntegerType *IT = cast<IntegerType>(I.getType());
1081        APInt LHSKnownOne(IT->getBitWidth(), 0);
1082        APInt LHSKnownZero(IT->getBitWidth(), 0);
1083        computeKnownBits(XorLHS, LHSKnownZero, LHSKnownOne);
1084        if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue())
1085          return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1086                                           XorLHS);
1087      }
1088      // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C,
1089      // transform them into (X + (signbit ^ C))
1090      if (XorRHS->getValue().isSignBit())
1091          return BinaryOperator::CreateAdd(XorLHS,
1092                                           ConstantExpr::getXor(XorRHS, CI));
1093    }
1094  }
1095
1096  if (isa<Constant>(RHS) && isa<PHINode>(LHS))
1097    if (Instruction *NV = FoldOpIntoPhi(I))
1098      return NV;
1099
1100  if (I.getType()->getScalarType()->isIntegerTy(1))
1101    return BinaryOperator::CreateXor(LHS, RHS);
1102
1103  // X + X --> X << 1
1104  if (LHS == RHS) {
1105    BinaryOperator *New =
1106      BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1));
1107    New->setHasNoSignedWrap(I.hasNoSignedWrap());
1108    New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1109    return New;
1110  }
1111
1112  // -A + B  -->  B - A
1113  // -A + -B  -->  -(A + B)
1114  if (Value *LHSV = dyn_castNegVal(LHS)) {
1115    if (!isa<Constant>(RHS))
1116      if (Value *RHSV = dyn_castNegVal(RHS)) {
1117        Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
1118        return BinaryOperator::CreateNeg(NewAdd);
1119      }
1120
1121    return BinaryOperator::CreateSub(RHS, LHSV);
1122  }
1123
1124  // A + -B  -->  A - B
1125  if (!isa<Constant>(RHS))
1126    if (Value *V = dyn_castNegVal(RHS))
1127      return BinaryOperator::CreateSub(LHS, V);
1128
1129  if (Value *V = checkForNegativeOperand(I, Builder))
1130    return ReplaceInstUsesWith(I, V);
1131
1132  // A+B --> A|B iff A and B have no bits set in common.
1133  if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
1134    APInt LHSKnownOne(IT->getBitWidth(), 0);
1135    APInt LHSKnownZero(IT->getBitWidth(), 0);
1136    computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
1137    if (LHSKnownZero != 0) {
1138      APInt RHSKnownOne(IT->getBitWidth(), 0);
1139      APInt RHSKnownZero(IT->getBitWidth(), 0);
1140      computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
1141
1142      // No bits in common -> bitwise or.
1143      if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
1144        return BinaryOperator::CreateOr(LHS, RHS);
1145    }
1146  }
1147
1148  if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
1149    Value *X;
1150    if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
1151      return BinaryOperator::CreateSub(SubOne(CRHS), X);
1152  }
1153
1154  if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1155    // (X & FF00) + xx00  -> (X+xx00) & FF00
1156    Value *X;
1157    ConstantInt *C2;
1158    if (LHS->hasOneUse() &&
1159        match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1160        CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1161      // See if all bits from the first bit set in the Add RHS up are included
1162      // in the mask.  First, get the rightmost bit.
1163      const APInt &AddRHSV = CRHS->getValue();
1164
1165      // Form a mask of all bits from the lowest bit added through the top.
1166      APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1167
1168      // See if the and mask includes all of these bits.
1169      APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1170
1171      if (AddRHSHighBits == AddRHSHighBitsAnd) {
1172        // Okay, the xform is safe.  Insert the new add pronto.
1173        Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
1174        return BinaryOperator::CreateAnd(NewAdd, C2);
1175      }
1176    }
1177
1178    // Try to fold constant add into select arguments.
1179    if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1180      if (Instruction *R = FoldOpIntoSelect(I, SI))
1181        return R;
1182  }
1183
1184  // add (select X 0 (sub n A)) A  -->  select X A n
1185  {
1186    SelectInst *SI = dyn_cast<SelectInst>(LHS);
1187    Value *A = RHS;
1188    if (!SI) {
1189      SI = dyn_cast<SelectInst>(RHS);
1190      A = LHS;
1191    }
1192    if (SI && SI->hasOneUse()) {
1193      Value *TV = SI->getTrueValue();
1194      Value *FV = SI->getFalseValue();
1195      Value *N;
1196
1197      // Can we fold the add into the argument of the select?
1198      // We check both true and false select arguments for a matching subtract.
1199      if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1200        // Fold the add into the true select value.
1201        return SelectInst::Create(SI->getCondition(), N, A);
1202
1203      if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1204        // Fold the add into the false select value.
1205        return SelectInst::Create(SI->getCondition(), A, N);
1206    }
1207  }
1208
1209  // Check for (add (sext x), y), see if we can merge this into an
1210  // integer add followed by a sext.
1211  if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
1212    // (add (sext x), cst) --> (sext (add x, cst'))
1213    if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1214      Constant *CI =
1215        ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1216      if (LHSConv->hasOneUse() &&
1217          ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
1218          WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
1219        // Insert the new, smaller add.
1220        Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1221                                              CI, "addconv");
1222        return new SExtInst(NewAdd, I.getType());
1223      }
1224    }
1225
1226    // (add (sext x), (sext y)) --> (sext (add int x, y))
1227    if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
1228      // Only do this if x/y have the same type, if at last one of them has a
1229      // single use (so we don't increase the number of sexts), and if the
1230      // integer add will not overflow.
1231      if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
1232          (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1233          WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1234                                   RHSConv->getOperand(0))) {
1235        // Insert the new integer add.
1236        Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1237                                             RHSConv->getOperand(0), "addconv");
1238        return new SExtInst(NewAdd, I.getType());
1239      }
1240    }
1241  }
1242
1243  // Check for (x & y) + (x ^ y)
1244  {
1245    Value *A = nullptr, *B = nullptr;
1246    if (match(RHS, m_Xor(m_Value(A), m_Value(B))) &&
1247        (match(LHS, m_And(m_Specific(A), m_Specific(B))) ||
1248         match(LHS, m_And(m_Specific(B), m_Specific(A)))))
1249      return BinaryOperator::CreateOr(A, B);
1250
1251    if (match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
1252        (match(RHS, m_And(m_Specific(A), m_Specific(B))) ||
1253         match(RHS, m_And(m_Specific(B), m_Specific(A)))))
1254      return BinaryOperator::CreateOr(A, B);
1255  }
1256
1257  // TODO(jingyue): Consider WillNotOverflowSignedAdd and
1258  // WillNotOverflowUnsignedAdd to reduce the number of invocations of
1259  // computeKnownBits.
1260  if (!I.hasNoSignedWrap() && WillNotOverflowSignedAdd(LHS, RHS)) {
1261    Changed = true;
1262    I.setHasNoSignedWrap(true);
1263  }
1264  if (!I.hasNoUnsignedWrap() && WillNotOverflowUnsignedAdd(LHS, RHS)) {
1265    Changed = true;
1266    I.setHasNoUnsignedWrap(true);
1267  }
1268
1269  return Changed ? &I : nullptr;
1270}
1271
1272Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
1273  bool Changed = SimplifyAssociativeOrCommutative(I);
1274  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1275
1276  if (Value *V = SimplifyVectorOp(I))
1277    return ReplaceInstUsesWith(I, V);
1278
1279  if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL))
1280    return ReplaceInstUsesWith(I, V);
1281
1282  if (isa<Constant>(RHS)) {
1283    if (isa<PHINode>(LHS))
1284      if (Instruction *NV = FoldOpIntoPhi(I))
1285        return NV;
1286
1287    if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1288      if (Instruction *NV = FoldOpIntoSelect(I, SI))
1289        return NV;
1290  }
1291
1292  // -A + B  -->  B - A
1293  // -A + -B  -->  -(A + B)
1294  if (Value *LHSV = dyn_castFNegVal(LHS)) {
1295    Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV);
1296    RI->copyFastMathFlags(&I);
1297    return RI;
1298  }
1299
1300  // A + -B  -->  A - B
1301  if (!isa<Constant>(RHS))
1302    if (Value *V = dyn_castFNegVal(RHS)) {
1303      Instruction *RI = BinaryOperator::CreateFSub(LHS, V);
1304      RI->copyFastMathFlags(&I);
1305      return RI;
1306    }
1307
1308  // Check for (fadd double (sitofp x), y), see if we can merge this into an
1309  // integer add followed by a promotion.
1310  if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1311    // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1312    // ... if the constant fits in the integer value.  This is useful for things
1313    // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1314    // requires a constant pool load, and generally allows the add to be better
1315    // instcombined.
1316    if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
1317      Constant *CI =
1318      ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
1319      if (LHSConv->hasOneUse() &&
1320          ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1321          WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
1322        // Insert the new integer add.
1323        Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1324                                              CI, "addconv");
1325        return new SIToFPInst(NewAdd, I.getType());
1326      }
1327    }
1328
1329    // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1330    if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1331      // Only do this if x/y have the same type, if at last one of them has a
1332      // single use (so we don't increase the number of int->fp conversions),
1333      // and if the integer add will not overflow.
1334      if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
1335          (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1336          WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1337                                   RHSConv->getOperand(0))) {
1338        // Insert the new integer add.
1339        Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1340                                              RHSConv->getOperand(0),"addconv");
1341        return new SIToFPInst(NewAdd, I.getType());
1342      }
1343    }
1344  }
1345
1346  // select C, 0, B + select C, A, 0 -> select C, A, B
1347  {
1348    Value *A1, *B1, *C1, *A2, *B2, *C2;
1349    if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) &&
1350        match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) {
1351      if (C1 == C2) {
1352        Constant *Z1=nullptr, *Z2=nullptr;
1353        Value *A, *B, *C=C1;
1354        if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) {
1355            Z1 = dyn_cast<Constant>(A1); A = A2;
1356            Z2 = dyn_cast<Constant>(B2); B = B1;
1357        } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) {
1358            Z1 = dyn_cast<Constant>(B1); B = B2;
1359            Z2 = dyn_cast<Constant>(A2); A = A1;
1360        }
1361
1362        if (Z1 && Z2 &&
1363            (I.hasNoSignedZeros() ||
1364             (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) {
1365          return SelectInst::Create(C, A, B);
1366        }
1367      }
1368    }
1369  }
1370
1371  if (I.hasUnsafeAlgebra()) {
1372    if (Value *V = FAddCombine(Builder).simplify(&I))
1373      return ReplaceInstUsesWith(I, V);
1374  }
1375
1376  return Changed ? &I : nullptr;
1377}
1378
1379
1380/// Optimize pointer differences into the same array into a size.  Consider:
1381///  &A[10] - &A[0]: we should compile this to "10".  LHS/RHS are the pointer
1382/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1383///
1384Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1385                                               Type *Ty) {
1386  assert(DL && "Must have target data info for this");
1387
1388  // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1389  // this.
1390  bool Swapped = false;
1391  GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1392
1393  // For now we require one side to be the base pointer "A" or a constant
1394  // GEP derived from it.
1395  if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1396    // (gep X, ...) - X
1397    if (LHSGEP->getOperand(0) == RHS) {
1398      GEP1 = LHSGEP;
1399      Swapped = false;
1400    } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1401      // (gep X, ...) - (gep X, ...)
1402      if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1403            RHSGEP->getOperand(0)->stripPointerCasts()) {
1404        GEP2 = RHSGEP;
1405        GEP1 = LHSGEP;
1406        Swapped = false;
1407      }
1408    }
1409  }
1410
1411  if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1412    // X - (gep X, ...)
1413    if (RHSGEP->getOperand(0) == LHS) {
1414      GEP1 = RHSGEP;
1415      Swapped = true;
1416    } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1417      // (gep X, ...) - (gep X, ...)
1418      if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1419            LHSGEP->getOperand(0)->stripPointerCasts()) {
1420        GEP2 = LHSGEP;
1421        GEP1 = RHSGEP;
1422        Swapped = true;
1423      }
1424    }
1425  }
1426
1427  // Avoid duplicating the arithmetic if GEP2 has non-constant indices and
1428  // multiple users.
1429  if (!GEP1 ||
1430      (GEP2 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse()))
1431    return nullptr;
1432
1433  // Emit the offset of the GEP and an intptr_t.
1434  Value *Result = EmitGEPOffset(GEP1);
1435
1436  // If we had a constant expression GEP on the other side offsetting the
1437  // pointer, subtract it from the offset we have.
1438  if (GEP2) {
1439    Value *Offset = EmitGEPOffset(GEP2);
1440    Result = Builder->CreateSub(Result, Offset);
1441  }
1442
1443  // If we have p - gep(p, ...)  then we have to negate the result.
1444  if (Swapped)
1445    Result = Builder->CreateNeg(Result, "diff.neg");
1446
1447  return Builder->CreateIntCast(Result, Ty, true);
1448}
1449
1450
1451Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1452  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1453
1454  if (Value *V = SimplifyVectorOp(I))
1455    return ReplaceInstUsesWith(I, V);
1456
1457  if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(),
1458                                 I.hasNoUnsignedWrap(), DL))
1459    return ReplaceInstUsesWith(I, V);
1460
1461  // (A*B)-(A*C) -> A*(B-C) etc
1462  if (Value *V = SimplifyUsingDistributiveLaws(I))
1463    return ReplaceInstUsesWith(I, V);
1464
1465  // If this is a 'B = x-(-A)', change to B = x+A.  This preserves NSW/NUW.
1466  if (Value *V = dyn_castNegVal(Op1)) {
1467    BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1468    Res->setHasNoSignedWrap(I.hasNoSignedWrap());
1469    Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1470    return Res;
1471  }
1472
1473  if (I.getType()->isIntegerTy(1))
1474    return BinaryOperator::CreateXor(Op0, Op1);
1475
1476  // Replace (-1 - A) with (~A).
1477  if (match(Op0, m_AllOnes()))
1478    return BinaryOperator::CreateNot(Op1);
1479
1480  if (Constant *C = dyn_cast<Constant>(Op0)) {
1481    // C - ~X == X + (1+C)
1482    Value *X = nullptr;
1483    if (match(Op1, m_Not(m_Value(X))))
1484      return BinaryOperator::CreateAdd(X, AddOne(C));
1485
1486    // Try to fold constant sub into select arguments.
1487    if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1488      if (Instruction *R = FoldOpIntoSelect(I, SI))
1489        return R;
1490
1491    // C-(X+C2) --> (C-C2)-X
1492    Constant *C2;
1493    if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1494      return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1495
1496    if (SimplifyDemandedInstructionBits(I))
1497      return &I;
1498
1499    // Fold (sub 0, (zext bool to B)) --> (sext bool to B)
1500    if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X))))
1501      if (X->getType()->getScalarType()->isIntegerTy(1))
1502        return CastInst::CreateSExtOrBitCast(X, Op1->getType());
1503
1504    // Fold (sub 0, (sext bool to B)) --> (zext bool to B)
1505    if (C->isNullValue() && match(Op1, m_SExt(m_Value(X))))
1506      if (X->getType()->getScalarType()->isIntegerTy(1))
1507        return CastInst::CreateZExtOrBitCast(X, Op1->getType());
1508  }
1509
1510  if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1511    // -(X >>u 31) -> (X >>s 31)
1512    // -(X >>s 31) -> (X >>u 31)
1513    if (C->isZero()) {
1514      Value *X; ConstantInt *CI;
1515      if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) &&
1516          // Verify we are shifting out everything but the sign bit.
1517          CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1)
1518        return BinaryOperator::CreateAShr(X, CI);
1519
1520      if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) &&
1521          // Verify we are shifting out everything but the sign bit.
1522          CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1)
1523        return BinaryOperator::CreateLShr(X, CI);
1524    }
1525  }
1526
1527
1528  { Value *Y;
1529    // X-(X+Y) == -Y    X-(Y+X) == -Y
1530    if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) ||
1531        match(Op1, m_Add(m_Value(Y), m_Specific(Op0))))
1532      return BinaryOperator::CreateNeg(Y);
1533
1534    // (X-Y)-X == -Y
1535    if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1536      return BinaryOperator::CreateNeg(Y);
1537  }
1538
1539  if (Op1->hasOneUse()) {
1540    Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1541    Constant *C = nullptr;
1542    Constant *CI = nullptr;
1543
1544    // (X - (Y - Z))  -->  (X + (Z - Y)).
1545    if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1546      return BinaryOperator::CreateAdd(Op0,
1547                                      Builder->CreateSub(Z, Y, Op1->getName()));
1548
1549    // (X - (X & Y))   -->   (X & ~Y)
1550    //
1551    if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) ||
1552        match(Op1, m_And(m_Specific(Op0), m_Value(Y))))
1553      return BinaryOperator::CreateAnd(Op0,
1554                                  Builder->CreateNot(Y, Y->getName() + ".not"));
1555
1556    // 0 - (X sdiv C)  -> (X sdiv -C)  provided the negation doesn't overflow.
1557    if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1558        !C->isMinSignedValue())
1559      return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1560
1561    // 0 - (X << Y)  -> (-X << Y)   when X is freely negatable.
1562    if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1563      if (Value *XNeg = dyn_castNegVal(X))
1564        return BinaryOperator::CreateShl(XNeg, Y);
1565
1566    // X - A*-B -> X + A*B
1567    // X - -A*B -> X + A*B
1568    Value *A, *B;
1569    if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) ||
1570        match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B))))
1571      return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B));
1572
1573    // X - A*CI -> X + A*-CI
1574    // X - CI*A -> X + A*-CI
1575    if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) ||
1576        match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) {
1577      Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI));
1578      return BinaryOperator::CreateAdd(Op0, NewMul);
1579    }
1580  }
1581
1582  // Optimize pointer differences into the same array into a size.  Consider:
1583  //  &A[10] - &A[0]: we should compile this to "10".
1584  if (DL) {
1585    Value *LHSOp, *RHSOp;
1586    if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1587        match(Op1, m_PtrToInt(m_Value(RHSOp))))
1588      if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1589        return ReplaceInstUsesWith(I, Res);
1590
1591    // trunc(p)-trunc(q) -> trunc(p-q)
1592    if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1593        match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1594      if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1595        return ReplaceInstUsesWith(I, Res);
1596      }
1597
1598  return nullptr;
1599}
1600
1601Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
1602  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1603
1604  if (Value *V = SimplifyVectorOp(I))
1605    return ReplaceInstUsesWith(I, V);
1606
1607  if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL))
1608    return ReplaceInstUsesWith(I, V);
1609
1610  if (isa<Constant>(Op0))
1611    if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1612      if (Instruction *NV = FoldOpIntoSelect(I, SI))
1613        return NV;
1614
1615  // If this is a 'B = x-(-A)', change to B = x+A, potentially looking
1616  // through FP extensions/truncations along the way.
1617  if (Value *V = dyn_castFNegVal(Op1)) {
1618    Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V);
1619    NewI->copyFastMathFlags(&I);
1620    return NewI;
1621  }
1622  if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) {
1623    if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) {
1624      Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType());
1625      Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc);
1626      NewI->copyFastMathFlags(&I);
1627      return NewI;
1628    }
1629  } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) {
1630    if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
1631      Value *NewExt = Builder->CreateFPExt(V, I.getType());
1632      Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt);
1633      NewI->copyFastMathFlags(&I);
1634      return NewI;
1635    }
1636  }
1637
1638  if (I.hasUnsafeAlgebra()) {
1639    if (Value *V = FAddCombine(Builder).simplify(&I))
1640      return ReplaceInstUsesWith(I, V);
1641  }
1642
1643  return nullptr;
1644}
1645