LLVM  8.0.0svn
InstCombineAddSub.cpp
Go to the documentation of this file.
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"
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 
65  void set(short C) {
66  assert(!insaneIntVal(C) && "Insane coefficient");
67  IsFp = false; IntVal = C;
68  }
69 
70  void set(const APFloat& C);
71 
72  void negate();
73 
74  bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75  Value *getValue(Type *) const;
76 
77  bool isOne() const { return isInt() && IntVal == 1; }
78  bool isTwo() const { return isInt() && IntVal == 2; }
79  bool isMinusOne() const { return isInt() && IntVal == -1; }
80  bool isMinusTwo() const { return isInt() && IntVal == -2; }
81 
82  private:
83  bool insaneIntVal(int V) { return V > 4 || V < -4; }
84 
85  APFloat *getFpValPtr()
86  { return reinterpret_cast<APFloat *>(&FpValBuf.buffer[0]); }
87 
88  const APFloat *getFpValPtr() const
89  { return reinterpret_cast<const APFloat *>(&FpValBuf.buffer[0]); }
90 
91  const APFloat &getFpVal() const {
92  assert(IsFp && BufHasFpVal && "Incorret state");
93  return *getFpValPtr();
94  }
95 
96  APFloat &getFpVal() {
97  assert(IsFp && BufHasFpVal && "Incorret state");
98  return *getFpValPtr();
99  }
100 
101  bool isInt() const { return !IsFp; }
102 
103  // If the coefficient is represented by an integer, promote it to a
104  // floating point.
105  void convertToFpType(const fltSemantics &Sem);
106 
107  // Construct an APFloat from a signed integer.
108  // TODO: We should get rid of this function when APFloat can be constructed
109  // from an *SIGNED* integer.
110  APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
111 
112  bool IsFp = false;
113 
114  // True iff FpValBuf contains an instance of APFloat.
115  bool BufHasFpVal = false;
116 
117  // The integer coefficient of an individual addend is either 1 or -1,
118  // and we try to simplify at most 4 addends from neighboring at most
119  // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120  // is overkill of this end.
121  short IntVal = 0;
122 
124  };
125 
126  /// FAddend is used to represent floating-point addend. An addend is
127  /// represented as <C, V>, where the V is a symbolic value, and C is a
128  /// constant coefficient. A constant addend is represented as <C, 0>.
129  class FAddend {
130  public:
131  FAddend() = default;
132 
133  void operator+=(const FAddend &T) {
134  assert((Val == T.Val) && "Symbolic-values disagree");
135  Coeff += T.Coeff;
136  }
137 
138  Value *getSymVal() const { return Val; }
139  const FAddendCoef &getCoef() const { return Coeff; }
140 
141  bool isConstant() const { return Val == nullptr; }
142  bool isZero() const { return Coeff.isZero(); }
143 
144  void set(short Coefficient, Value *V) {
145  Coeff.set(Coefficient);
146  Val = V;
147  }
148  void set(const APFloat &Coefficient, Value *V) {
149  Coeff.set(Coefficient);
150  Val = V;
151  }
152  void set(const ConstantFP *Coefficient, Value *V) {
153  Coeff.set(Coefficient->getValueAPF());
154  Val = V;
155  }
156 
157  void negate() { Coeff.negate(); }
158 
159  /// Drill down the U-D chain one step to find the definition of V, and
160  /// try to break the definition into one or two addends.
161  static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
162 
163  /// Similar to FAddend::drillDownOneStep() except that the value being
164  /// splitted is the addend itself.
165  unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
166 
167  private:
168  void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
169 
170  // This addend has the value of "Coeff * Val".
171  Value *Val = nullptr;
172  FAddendCoef Coeff;
173  };
174 
175  /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176  /// with its neighboring at most two instructions.
177  ///
178  class FAddCombine {
179  public:
180  FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
181 
182  Value *simplify(Instruction *FAdd);
183 
184  private:
185  using AddendVect = SmallVector<const FAddend *, 4>;
186 
187  Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
188 
189  /// Convert given addend to a Value
190  Value *createAddendVal(const FAddend &A, bool& NeedNeg);
191 
192  /// Return the number of instructions needed to emit the N-ary addition.
193  unsigned calcInstrNumber(const AddendVect& Vect);
194 
195  Value *createFSub(Value *Opnd0, Value *Opnd1);
196  Value *createFAdd(Value *Opnd0, Value *Opnd1);
197  Value *createFMul(Value *Opnd0, Value *Opnd1);
198  Value *createFNeg(Value *V);
199  Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200  void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
201 
202  // Debugging stuff are clustered here.
203  #ifndef NDEBUG
204  unsigned CreateInstrNum;
205  void initCreateInstNum() { CreateInstrNum = 0; }
206  void incCreateInstNum() { CreateInstrNum++; }
207  #else
208  void initCreateInstNum() {}
209  void incCreateInstNum() {}
210  #endif
211 
212  InstCombiner::BuilderTy &Builder;
213  Instruction *Instr = nullptr;
214  };
215 
216 } // end anonymous namespace
217 
218 //===----------------------------------------------------------------------===//
219 //
220 // Implementation of
221 // {FAddendCoef, FAddend, FAddition, FAddCombine}.
222 //
223 //===----------------------------------------------------------------------===//
224 FAddendCoef::~FAddendCoef() {
225  if (BufHasFpVal)
226  getFpValPtr()->~APFloat();
227 }
228 
229 void FAddendCoef::set(const APFloat& C) {
230  APFloat *P = getFpValPtr();
231 
232  if (isInt()) {
233  // As the buffer is meanless byte stream, we cannot call
234  // APFloat::operator=().
235  new(P) APFloat(C);
236  } else
237  *P = C;
238 
239  IsFp = BufHasFpVal = true;
240 }
241 
242 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
243  if (!isInt())
244  return;
245 
246  APFloat *P = getFpValPtr();
247  if (IntVal > 0)
248  new(P) APFloat(Sem, IntVal);
249  else {
250  new(P) APFloat(Sem, 0 - IntVal);
251  P->changeSign();
252  }
253  IsFp = BufHasFpVal = true;
254 }
255 
256 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
257  if (Val >= 0)
258  return APFloat(Sem, Val);
259 
260  APFloat T(Sem, 0 - Val);
261  T.changeSign();
262 
263  return T;
264 }
265 
266 void FAddendCoef::operator=(const FAddendCoef &That) {
267  if (That.isInt())
268  set(That.IntVal);
269  else
270  set(That.getFpVal());
271 }
272 
273 void FAddendCoef::operator+=(const FAddendCoef &That) {
275  if (isInt() == That.isInt()) {
276  if (isInt())
277  IntVal += That.IntVal;
278  else
279  getFpVal().add(That.getFpVal(), RndMode);
280  return;
281  }
282 
283  if (isInt()) {
284  const APFloat &T = That.getFpVal();
285  convertToFpType(T.getSemantics());
286  getFpVal().add(T, RndMode);
287  return;
288  }
289 
290  APFloat &T = getFpVal();
291  T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292 }
293 
294 void FAddendCoef::operator*=(const FAddendCoef &That) {
295  if (That.isOne())
296  return;
297 
298  if (That.isMinusOne()) {
299  negate();
300  return;
301  }
302 
303  if (isInt() && That.isInt()) {
304  int Res = IntVal * (int)That.IntVal;
305  assert(!insaneIntVal(Res) && "Insane int value");
306  IntVal = Res;
307  return;
308  }
309 
310  const fltSemantics &Semantic =
311  isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312 
313  if (isInt())
314  convertToFpType(Semantic);
315  APFloat &F0 = getFpVal();
316 
317  if (That.isInt())
318  F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
320  else
321  F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322 }
323 
324 void FAddendCoef::negate() {
325  if (isInt())
326  IntVal = 0 - IntVal;
327  else
328  getFpVal().changeSign();
329 }
330 
331 Value *FAddendCoef::getValue(Type *Ty) const {
332  return isInt() ?
333  ConstantFP::get(Ty, float(IntVal)) :
334  ConstantFP::get(Ty->getContext(), getFpVal());
335 }
336 
337 // The definition of <Val> Addends
338 // =========================================
339 // A + B <1, A>, <1,B>
340 // A - B <1, A>, <1,B>
341 // 0 - B <-1, B>
342 // C * A, <C, A>
343 // A + C <1, A> <C, NULL>
344 // 0 +/- 0 <0, NULL> (corner case)
345 //
346 // Legend: A and B are not constant, C is constant
347 unsigned FAddend::drillValueDownOneStep
348  (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349  Instruction *I = nullptr;
350  if (!Val || !(I = dyn_cast<Instruction>(Val)))
351  return 0;
352 
353  unsigned Opcode = I->getOpcode();
354 
355  if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356  ConstantFP *C0, *C1;
357  Value *Opnd0 = I->getOperand(0);
358  Value *Opnd1 = I->getOperand(1);
359  if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360  Opnd0 = nullptr;
361 
362  if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363  Opnd1 = nullptr;
364 
365  if (Opnd0) {
366  if (!C0)
367  Addend0.set(1, Opnd0);
368  else
369  Addend0.set(C0, nullptr);
370  }
371 
372  if (Opnd1) {
373  FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374  if (!C1)
375  Addend.set(1, Opnd1);
376  else
377  Addend.set(C1, nullptr);
378  if (Opcode == Instruction::FSub)
379  Addend.negate();
380  }
381 
382  if (Opnd0 || Opnd1)
383  return Opnd0 && Opnd1 ? 2 : 1;
384 
385  // Both operands are zero. Weird!
386  Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387  return 1;
388  }
389 
390  if (I->getOpcode() == Instruction::FMul) {
391  Value *V0 = I->getOperand(0);
392  Value *V1 = I->getOperand(1);
393  if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394  Addend0.set(C, V1);
395  return 1;
396  }
397 
398  if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399  Addend0.set(C, V0);
400  return 1;
401  }
402  }
403 
404  return 0;
405 }
406 
407 // Try to break *this* addend into two addends. e.g. Suppose this addend is
408 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409 // i.e. <2.3, X> and <2.3, Y>.
410 unsigned FAddend::drillAddendDownOneStep
411  (FAddend &Addend0, FAddend &Addend1) const {
412  if (isConstant())
413  return 0;
414 
415  unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416  if (!BreakNum || Coeff.isOne())
417  return BreakNum;
418 
419  Addend0.Scale(Coeff);
420 
421  if (BreakNum == 2)
422  Addend1.Scale(Coeff);
423 
424  return BreakNum;
425 }
426 
428  assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429  "Expected 'reassoc'+'nsz' instruction");
430 
431  // Currently we are not able to handle vector type.
432  if (I->getType()->isVectorTy())
433  return nullptr;
434 
435  assert((I->getOpcode() == Instruction::FAdd ||
436  I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437 
438  // Save the instruction before calling other member-functions.
439  Instr = I;
440 
441  FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442 
443  unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444 
445  // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446  unsigned Opnd0_ExpNum = 0;
447  unsigned Opnd1_ExpNum = 0;
448 
449  if (!Opnd0.isConstant())
450  Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451 
452  // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453  if (OpndNum == 2 && !Opnd1.isConstant())
454  Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455 
456  // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457  if (Opnd0_ExpNum && Opnd1_ExpNum) {
458  AddendVect AllOpnds;
459  AllOpnds.push_back(&Opnd0_0);
460  AllOpnds.push_back(&Opnd1_0);
461  if (Opnd0_ExpNum == 2)
462  AllOpnds.push_back(&Opnd0_1);
463  if (Opnd1_ExpNum == 2)
464  AllOpnds.push_back(&Opnd1_1);
465 
466  // Compute instruction quota. We should save at least one instruction.
467  unsigned InstQuota = 0;
468 
469  Value *V0 = I->getOperand(0);
470  Value *V1 = I->getOperand(1);
471  InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472  (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473 
474  if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475  return R;
476  }
477 
478  if (OpndNum != 2) {
479  // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480  // splitted into two addends, say "V = X - Y", the instruction would have
481  // been optimized into "I = Y - X" in the previous steps.
482  //
483  const FAddendCoef &CE = Opnd0.getCoef();
484  return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485  }
486 
487  // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488  if (Opnd1_ExpNum) {
489  AddendVect AllOpnds;
490  AllOpnds.push_back(&Opnd0);
491  AllOpnds.push_back(&Opnd1_0);
492  if (Opnd1_ExpNum == 2)
493  AllOpnds.push_back(&Opnd1_1);
494 
495  if (Value *R = simplifyFAdd(AllOpnds, 1))
496  return R;
497  }
498 
499  // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500  if (Opnd0_ExpNum) {
501  AddendVect AllOpnds;
502  AllOpnds.push_back(&Opnd1);
503  AllOpnds.push_back(&Opnd0_0);
504  if (Opnd0_ExpNum == 2)
505  AllOpnds.push_back(&Opnd0_1);
506 
507  if (Value *R = simplifyFAdd(AllOpnds, 1))
508  return R;
509  }
510 
511  return nullptr;
512 }
513 
514 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515  unsigned AddendNum = Addends.size();
516  assert(AddendNum <= 4 && "Too many addends");
517 
518  // For saving intermediate results;
519  unsigned NextTmpIdx = 0;
520  FAddend TmpResult[3];
521 
522  // Points to the constant addend of the resulting simplified expression.
523  // If the resulting expr has constant-addend, this constant-addend is
524  // desirable to reside at the top of the resulting expression tree. Placing
525  // constant close to supper-expr(s) will potentially reveal some optimization
526  // opportunities in super-expr(s).
527  const FAddend *ConstAdd = nullptr;
528 
529  // Simplified addends are placed <SimpVect>.
530  AddendVect SimpVect;
531 
532  // The outer loop works on one symbolic-value at a time. Suppose the input
533  // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
534  // The symbolic-values will be processed in this order: x, y, z.
535  for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
536 
537  const FAddend *ThisAddend = Addends[SymIdx];
538  if (!ThisAddend) {
539  // This addend was processed before.
540  continue;
541  }
542 
543  Value *Val = ThisAddend->getSymVal();
544  unsigned StartIdx = SimpVect.size();
545  SimpVect.push_back(ThisAddend);
546 
547  // The inner loop collects addends sharing same symbolic-value, and these
548  // addends will be later on folded into a single addend. Following above
549  // example, if the symbolic value "y" is being processed, the inner loop
550  // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
551  // be later on folded into "<b1+b2, y>".
552  for (unsigned SameSymIdx = SymIdx + 1;
553  SameSymIdx < AddendNum; SameSymIdx++) {
554  const FAddend *T = Addends[SameSymIdx];
555  if (T && T->getSymVal() == Val) {
556  // Set null such that next iteration of the outer loop will not process
557  // this addend again.
558  Addends[SameSymIdx] = nullptr;
559  SimpVect.push_back(T);
560  }
561  }
562 
563  // If multiple addends share same symbolic value, fold them together.
564  if (StartIdx + 1 != SimpVect.size()) {
565  FAddend &R = TmpResult[NextTmpIdx ++];
566  R = *SimpVect[StartIdx];
567  for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
568  R += *SimpVect[Idx];
569 
570  // Pop all addends being folded and push the resulting folded addend.
571  SimpVect.resize(StartIdx);
572  if (Val) {
573  if (!R.isZero()) {
574  SimpVect.push_back(&R);
575  }
576  } else {
577  // Don't push constant addend at this time. It will be the last element
578  // of <SimpVect>.
579  ConstAdd = &R;
580  }
581  }
582  }
583 
584  assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
585  "out-of-bound access");
586 
587  if (ConstAdd)
588  SimpVect.push_back(ConstAdd);
589 
590  Value *Result;
591  if (!SimpVect.empty())
592  Result = createNaryFAdd(SimpVect, InstrQuota);
593  else {
594  // The addition is folded to 0.0.
595  Result = ConstantFP::get(Instr->getType(), 0.0);
596  }
597 
598  return Result;
599 }
600 
601 Value *FAddCombine::createNaryFAdd
602  (const AddendVect &Opnds, unsigned InstrQuota) {
603  assert(!Opnds.empty() && "Expect at least one addend");
604 
605  // Step 1: Check if the # of instructions needed exceeds the quota.
606 
607  unsigned InstrNeeded = calcInstrNumber(Opnds);
608  if (InstrNeeded > InstrQuota)
609  return nullptr;
610 
611  initCreateInstNum();
612 
613  // step 2: Emit the N-ary addition.
614  // Note that at most three instructions are involved in Fadd-InstCombine: the
615  // addition in question, and at most two neighboring instructions.
616  // The resulting optimized addition should have at least one less instruction
617  // than the original addition expression tree. This implies that the resulting
618  // N-ary addition has at most two instructions, and we don't need to worry
619  // about tree-height when constructing the N-ary addition.
620 
621  Value *LastVal = nullptr;
622  bool LastValNeedNeg = false;
623 
624  // Iterate the addends, creating fadd/fsub using adjacent two addends.
625  for (const FAddend *Opnd : Opnds) {
626  bool NeedNeg;
627  Value *V = createAddendVal(*Opnd, NeedNeg);
628  if (!LastVal) {
629  LastVal = V;
630  LastValNeedNeg = NeedNeg;
631  continue;
632  }
633 
634  if (LastValNeedNeg == NeedNeg) {
635  LastVal = createFAdd(LastVal, V);
636  continue;
637  }
638 
639  if (LastValNeedNeg)
640  LastVal = createFSub(V, LastVal);
641  else
642  LastVal = createFSub(LastVal, V);
643 
644  LastValNeedNeg = false;
645  }
646 
647  if (LastValNeedNeg) {
648  LastVal = createFNeg(LastVal);
649  }
650 
651 #ifndef NDEBUG
652  assert(CreateInstrNum == InstrNeeded &&
653  "Inconsistent in instruction numbers");
654 #endif
655 
656  return LastVal;
657 }
658 
659 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
660  Value *V = Builder.CreateFSub(Opnd0, Opnd1);
661  if (Instruction *I = dyn_cast<Instruction>(V))
662  createInstPostProc(I);
663  return V;
664 }
665 
666 Value *FAddCombine::createFNeg(Value *V) {
667  Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType()));
668  Value *NewV = createFSub(Zero, V);
669  if (Instruction *I = dyn_cast<Instruction>(NewV))
670  createInstPostProc(I, true); // fneg's don't receive instruction numbers.
671  return NewV;
672 }
673 
674 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
675  Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
676  if (Instruction *I = dyn_cast<Instruction>(V))
677  createInstPostProc(I);
678  return V;
679 }
680 
681 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
682  Value *V = Builder.CreateFMul(Opnd0, Opnd1);
683  if (Instruction *I = dyn_cast<Instruction>(V))
684  createInstPostProc(I);
685  return V;
686 }
687 
688 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
689  NewInstr->setDebugLoc(Instr->getDebugLoc());
690 
691  // Keep track of the number of instruction created.
692  if (!NoNumber)
693  incCreateInstNum();
694 
695  // Propagate fast-math flags
696  NewInstr->setFastMathFlags(Instr->getFastMathFlags());
697 }
698 
699 // Return the number of instruction needed to emit the N-ary addition.
700 // NOTE: Keep this function in sync with createAddendVal().
701 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
702  unsigned OpndNum = Opnds.size();
703  unsigned InstrNeeded = OpndNum - 1;
704 
705  // The number of addends in the form of "(-1)*x".
706  unsigned NegOpndNum = 0;
707 
708  // Adjust the number of instructions needed to emit the N-ary add.
709  for (const FAddend *Opnd : Opnds) {
710  if (Opnd->isConstant())
711  continue;
712 
713  // The constant check above is really for a few special constant
714  // coefficients.
715  if (isa<UndefValue>(Opnd->getSymVal()))
716  continue;
717 
718  const FAddendCoef &CE = Opnd->getCoef();
719  if (CE.isMinusOne() || CE.isMinusTwo())
720  NegOpndNum++;
721 
722  // Let the addend be "c * x". If "c == +/-1", the value of the addend
723  // is immediately available; otherwise, it needs exactly one instruction
724  // to evaluate the value.
725  if (!CE.isMinusOne() && !CE.isOne())
726  InstrNeeded++;
727  }
728  if (NegOpndNum == OpndNum)
729  InstrNeeded++;
730  return InstrNeeded;
731 }
732 
733 // Input Addend Value NeedNeg(output)
734 // ================================================================
735 // Constant C C false
736 // <+/-1, V> V coefficient is -1
737 // <2/-2, V> "fadd V, V" coefficient is -2
738 // <C, V> "fmul V, C" false
739 //
740 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
741 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
742  const FAddendCoef &Coeff = Opnd.getCoef();
743 
744  if (Opnd.isConstant()) {
745  NeedNeg = false;
746  return Coeff.getValue(Instr->getType());
747  }
748 
749  Value *OpndVal = Opnd.getSymVal();
750 
751  if (Coeff.isMinusOne() || Coeff.isOne()) {
752  NeedNeg = Coeff.isMinusOne();
753  return OpndVal;
754  }
755 
756  if (Coeff.isTwo() || Coeff.isMinusTwo()) {
757  NeedNeg = Coeff.isMinusTwo();
758  return createFAdd(OpndVal, OpndVal);
759  }
760 
761  NeedNeg = false;
762  return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
763 }
764 
765 // Checks if any operand is negative and we can convert add to sub.
766 // This function checks for following negative patterns
767 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
768 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
769 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
771  InstCombiner::BuilderTy &Builder) {
772  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
773 
774  // This function creates 2 instructions to replace ADD, we need at least one
775  // of LHS or RHS to have one use to ensure benefit in transform.
776  if (!LHS->hasOneUse() && !RHS->hasOneUse())
777  return nullptr;
778 
779  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
780  const APInt *C1 = nullptr, *C2 = nullptr;
781 
782  // if ONE is on other side, swap
783  if (match(RHS, m_Add(m_Value(X), m_One())))
784  std::swap(LHS, RHS);
785 
786  if (match(LHS, m_Add(m_Value(X), m_One()))) {
787  // if XOR on other side, swap
788  if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
789  std::swap(X, RHS);
790 
791  if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
792  // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
793  // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
794  if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
795  Value *NewAnd = Builder.CreateAnd(Z, *C1);
796  return Builder.CreateSub(RHS, NewAnd, "sub");
797  } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
798  // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
799  // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
800  Value *NewOr = Builder.CreateOr(Z, ~(*C1));
801  return Builder.CreateSub(RHS, NewOr, "sub");
802  }
803  }
804  }
805 
806  // Restore LHS and RHS
807  LHS = I.getOperand(0);
808  RHS = I.getOperand(1);
809 
810  // if XOR is on other side, swap
811  if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
812  std::swap(LHS, RHS);
813 
814  // C2 is ODD
815  // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
816  // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
817  if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
818  if (C1->countTrailingZeros() == 0)
819  if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
820  Value *NewOr = Builder.CreateOr(Z, ~(*C2));
821  return Builder.CreateSub(RHS, NewOr, "sub");
822  }
823  return nullptr;
824 }
825 
826 Instruction *InstCombiner::foldAddWithConstant(BinaryOperator &Add) {
827  Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
828  Constant *Op1C;
829  if (!match(Op1, m_Constant(Op1C)))
830  return nullptr;
831 
832  if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
833  return NV;
834 
835  Value *X, *Y;
836 
837  // add (sub X, Y), -1 --> add (not Y), X
838  if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
839  match(Op1, m_AllOnes()))
840  return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
841 
842  // zext(bool) + C -> bool ? C + 1 : C
843  if (match(Op0, m_ZExt(m_Value(X))) &&
844  X->getType()->getScalarSizeInBits() == 1)
845  return SelectInst::Create(X, AddOne(Op1C), Op1);
846 
847  // ~X + C --> (C-1) - X
848  if (match(Op0, m_Not(m_Value(X))))
849  return BinaryOperator::CreateSub(SubOne(Op1C), X);
850 
851  const APInt *C;
852  if (!match(Op1, m_APInt(C)))
853  return nullptr;
854 
855  if (C->isSignMask()) {
856  // If wrapping is not allowed, then the addition must set the sign bit:
857  // X + (signmask) --> X | signmask
858  if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
859  return BinaryOperator::CreateOr(Op0, Op1);
860 
861  // If wrapping is allowed, then the addition flips the sign bit of LHS:
862  // X + (signmask) --> X ^ signmask
863  return BinaryOperator::CreateXor(Op0, Op1);
864  }
865 
866  // Is this add the last step in a convoluted sext?
867  // add(zext(xor i16 X, -32768), -32768) --> sext X
868  Type *Ty = Add.getType();
869  const APInt *C2;
870  if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
871  C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
872  return CastInst::Create(Instruction::SExt, X, Ty);
873 
874  // (add (zext (add nuw X, C2)), C) --> (zext (add nuw X, C2 + C))
875  if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
876  C->isNegative() && C->sge(-C2->sext(C->getBitWidth()))) {
877  Constant *NewC =
878  ConstantInt::get(X->getType(), *C2 + C->trunc(C2->getBitWidth()));
879  return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
880  }
881 
882  if (C->isOneValue() && Op0->hasOneUse()) {
883  // add (sext i1 X), 1 --> zext (not X)
884  // TODO: The smallest IR representation is (select X, 0, 1), and that would
885  // not require the one-use check. But we need to remove a transform in
886  // visitSelect and make sure that IR value tracking for select is equal or
887  // better than for these ops.
888  if (match(Op0, m_SExt(m_Value(X))) &&
889  X->getType()->getScalarSizeInBits() == 1)
890  return new ZExtInst(Builder.CreateNot(X), Ty);
891 
892  // Shifts and add used to flip and mask off the low bit:
893  // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
894  const APInt *C3;
895  if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
896  C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
897  Value *NotX = Builder.CreateNot(X);
898  return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
899  }
900  }
901 
902  return nullptr;
903 }
904 
905 // Matches multiplication expression Op * C where C is a constant. Returns the
906 // constant value in C and the other operand in Op. Returns true if such a
907 // match is found.
908 static bool MatchMul(Value *E, Value *&Op, APInt &C) {
909  const APInt *AI;
910  if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
911  C = *AI;
912  return true;
913  }
914  if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
915  C = APInt(AI->getBitWidth(), 1);
916  C <<= *AI;
917  return true;
918  }
919  return false;
920 }
921 
922 // Matches remainder expression Op % C where C is a constant. Returns the
923 // constant value in C and the other operand in Op. Returns the signedness of
924 // the remainder operation in IsSigned. Returns true if such a match is
925 // found.
926 static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
927  const APInt *AI;
928  IsSigned = false;
929  if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
930  IsSigned = true;
931  C = *AI;
932  return true;
933  }
934  if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
935  C = *AI;
936  return true;
937  }
938  if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
939  C = *AI + 1;
940  return true;
941  }
942  return false;
943 }
944 
945 // Matches division expression Op / C with the given signedness as indicated
946 // by IsSigned, where C is a constant. Returns the constant value in C and the
947 // other operand in Op. Returns true if such a match is found.
948 static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
949  const APInt *AI;
950  if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
951  C = *AI;
952  return true;
953  }
954  if (!IsSigned) {
955  if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
956  C = *AI;
957  return true;
958  }
959  if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
960  C = APInt(AI->getBitWidth(), 1);
961  C <<= *AI;
962  return true;
963  }
964  }
965  return false;
966 }
967 
968 // Returns whether C0 * C1 with the given signedness overflows.
969 static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
970  bool overflow;
971  if (IsSigned)
972  (void)C0.smul_ov(C1, overflow);
973  else
974  (void)C0.umul_ov(C1, overflow);
975  return overflow;
976 }
977 
978 // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
979 // does not overflow.
980 Value *InstCombiner::SimplifyAddWithRemainder(BinaryOperator &I) {
981  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
982  Value *X, *MulOpV;
983  APInt C0, MulOpC;
984  bool IsSigned;
985  // Match I = X % C0 + MulOpV * C0
986  if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
987  (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
988  C0 == MulOpC) {
989  Value *RemOpV;
990  APInt C1;
991  bool Rem2IsSigned;
992  // Match MulOpC = RemOpV % C1
993  if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
994  IsSigned == Rem2IsSigned) {
995  Value *DivOpV;
996  APInt DivOpC;
997  // Match RemOpV = X / C0
998  if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
999  C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1000  Value *NewDivisor =
1001  ConstantInt::get(X->getType()->getContext(), C0 * C1);
1002  return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1003  : Builder.CreateURem(X, NewDivisor, "urem");
1004  }
1005  }
1006  }
1007 
1008  return nullptr;
1009 }
1010 
1011 /// Fold
1012 /// (1 << NBits) - 1
1013 /// Into:
1014 /// ~(-(1 << NBits))
1015 /// Because a 'not' is better for bit-tracking analysis and other transforms
1016 /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1018  InstCombiner::BuilderTy &Builder) {
1019  Value *NBits;
1020  if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1021  return nullptr;
1022 
1023  Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1024  Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1025  // Be wary of constant folding.
1026  if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1027  // Always NSW. But NUW propagates from `add`.
1028  BOp->setHasNoSignedWrap();
1029  BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1030  }
1031 
1032  return BinaryOperator::CreateNot(NotMask, I.getName());
1033 }
1034 
1036  if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1038  SQ.getWithInstruction(&I)))
1039  return replaceInstUsesWith(I, V);
1040 
1041  if (SimplifyAssociativeOrCommutative(I))
1042  return &I;
1043 
1044  if (Instruction *X = foldVectorBinop(I))
1045  return X;
1046 
1047  // (A*B)+(A*C) -> A*(B+C) etc
1048  if (Value *V = SimplifyUsingDistributiveLaws(I))
1049  return replaceInstUsesWith(I, V);
1050 
1051  if (Instruction *X = foldAddWithConstant(I))
1052  return X;
1053 
1054  // FIXME: This should be moved into the above helper function to allow these
1055  // transforms for general constant or constant splat vectors.
1056  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1057  Type *Ty = I.getType();
1058  if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1059  Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1060  if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1061  unsigned TySizeBits = Ty->getScalarSizeInBits();
1062  const APInt &RHSVal = CI->getValue();
1063  unsigned ExtendAmt = 0;
1064  // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1065  // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1066  if (XorRHS->getValue() == -RHSVal) {
1067  if (RHSVal.isPowerOf2())
1068  ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1069  else if (XorRHS->getValue().isPowerOf2())
1070  ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1071  }
1072 
1073  if (ExtendAmt) {
1074  APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1075  if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1076  ExtendAmt = 0;
1077  }
1078 
1079  if (ExtendAmt) {
1080  Constant *ShAmt = ConstantInt::get(Ty, ExtendAmt);
1081  Value *NewShl = Builder.CreateShl(XorLHS, ShAmt, "sext");
1082  return BinaryOperator::CreateAShr(NewShl, ShAmt);
1083  }
1084 
1085  // If this is a xor that was canonicalized from a sub, turn it back into
1086  // a sub and fuse this add with it.
1087  if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1088  KnownBits LHSKnown = computeKnownBits(XorLHS, 0, &I);
1089  if ((XorRHS->getValue() | LHSKnown.Zero).isAllOnesValue())
1090  return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1091  XorLHS);
1092  }
1093  // (X + signmask) + C could have gotten canonicalized to (X^signmask) + C,
1094  // transform them into (X + (signmask ^ C))
1095  if (XorRHS->getValue().isSignMask())
1096  return BinaryOperator::CreateAdd(XorLHS,
1097  ConstantExpr::getXor(XorRHS, CI));
1098  }
1099  }
1100 
1101  if (Ty->isIntOrIntVectorTy(1))
1102  return BinaryOperator::CreateXor(LHS, RHS);
1103 
1104  // X + X --> X << 1
1105  if (LHS == RHS) {
1106  auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1107  Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1108  Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1109  return Shl;
1110  }
1111 
1112  Value *A, *B;
1113  if (match(LHS, m_Neg(m_Value(A)))) {
1114  // -A + -B --> -(A + B)
1115  if (match(RHS, m_Neg(m_Value(B))))
1116  return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1117 
1118  // -A + B --> B - A
1119  return BinaryOperator::CreateSub(RHS, A);
1120  }
1121 
1122  // A + -B --> A - B
1123  if (match(RHS, m_Neg(m_Value(B))))
1124  return BinaryOperator::CreateSub(LHS, B);
1125 
1126  if (Value *V = checkForNegativeOperand(I, Builder))
1127  return replaceInstUsesWith(I, V);
1128 
1129  // (A + 1) + ~B --> A - B
1130  // ~B + (A + 1) --> A - B
1131  if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))))
1132  return BinaryOperator::CreateSub(A, B);
1133 
1134  // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1135  if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1136 
1137  // A+B --> A|B iff A and B have no bits set in common.
1138  if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1139  return BinaryOperator::CreateOr(LHS, RHS);
1140 
1141  // FIXME: We already did a check for ConstantInt RHS above this.
1142  // FIXME: Is this pattern covered by another fold? No regression tests fail on
1143  // removal.
1144  if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1145  // (X & FF00) + xx00 -> (X+xx00) & FF00
1146  Value *X;
1147  ConstantInt *C2;
1148  if (LHS->hasOneUse() &&
1149  match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1150  CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1151  // See if all bits from the first bit set in the Add RHS up are included
1152  // in the mask. First, get the rightmost bit.
1153  const APInt &AddRHSV = CRHS->getValue();
1154 
1155  // Form a mask of all bits from the lowest bit added through the top.
1156  APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1157 
1158  // See if the and mask includes all of these bits.
1159  APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1160 
1161  if (AddRHSHighBits == AddRHSHighBitsAnd) {
1162  // Okay, the xform is safe. Insert the new add pronto.
1163  Value *NewAdd = Builder.CreateAdd(X, CRHS, LHS->getName());
1164  return BinaryOperator::CreateAnd(NewAdd, C2);
1165  }
1166  }
1167  }
1168 
1169  // add (select X 0 (sub n A)) A --> select X A n
1170  {
1171  SelectInst *SI = dyn_cast<SelectInst>(LHS);
1172  Value *A = RHS;
1173  if (!SI) {
1174  SI = dyn_cast<SelectInst>(RHS);
1175  A = LHS;
1176  }
1177  if (SI && SI->hasOneUse()) {
1178  Value *TV = SI->getTrueValue();
1179  Value *FV = SI->getFalseValue();
1180  Value *N;
1181 
1182  // Can we fold the add into the argument of the select?
1183  // We check both true and false select arguments for a matching subtract.
1184  if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1185  // Fold the add into the true select value.
1186  return SelectInst::Create(SI->getCondition(), N, A);
1187 
1188  if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1189  // Fold the add into the false select value.
1190  return SelectInst::Create(SI->getCondition(), A, N);
1191  }
1192  }
1193 
1194  if (Instruction *Ext = narrowMathIfNoOverflow(I))
1195  return Ext;
1196 
1197  // (add (xor A, B) (and A, B)) --> (or A, B)
1198  // (add (and A, B) (xor A, B)) --> (or A, B)
1199  if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1200  m_c_And(m_Deferred(A), m_Deferred(B)))))
1201  return BinaryOperator::CreateOr(A, B);
1202 
1203  // (add (or A, B) (and A, B)) --> (add A, B)
1204  // (add (and A, B) (or A, B)) --> (add A, B)
1205  if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1206  m_c_And(m_Deferred(A), m_Deferred(B))))) {
1207  I.setOperand(0, A);
1208  I.setOperand(1, B);
1209  return &I;
1210  }
1211 
1212  // TODO(jingyue): Consider willNotOverflowSignedAdd and
1213  // willNotOverflowUnsignedAdd to reduce the number of invocations of
1214  // computeKnownBits.
1215  bool Changed = false;
1216  if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1217  Changed = true;
1218  I.setHasNoSignedWrap(true);
1219  }
1220  if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1221  Changed = true;
1222  I.setHasNoUnsignedWrap(true);
1223  }
1224 
1225  if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1226  return V;
1227 
1228  return Changed ? &I : nullptr;
1229 }
1230 
1231 /// Factor a common operand out of fadd/fsub of fmul/fdiv.
1233  InstCombiner::BuilderTy &Builder) {
1234  assert((I.getOpcode() == Instruction::FAdd ||
1235  I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1237  "FP factorization requires FMF");
1238  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1239  Value *X, *Y, *Z;
1240  bool IsFMul;
1241  if ((match(Op0, m_OneUse(m_FMul(m_Value(X), m_Value(Z)))) &&
1242  match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))) ||
1243  (match(Op0, m_OneUse(m_FMul(m_Value(Z), m_Value(X)))) &&
1244  match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))))
1245  IsFMul = true;
1246  else if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Z)))) &&
1247  match(Op1, m_OneUse(m_FDiv(m_Value(Y), m_Specific(Z)))))
1248  IsFMul = false;
1249  else
1250  return nullptr;
1251 
1252  // (X * Z) + (Y * Z) --> (X + Y) * Z
1253  // (X * Z) - (Y * Z) --> (X - Y) * Z
1254  // (X / Z) + (Y / Z) --> (X + Y) / Z
1255  // (X / Z) - (Y / Z) --> (X - Y) / Z
1256  bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1257  Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1258  : Builder.CreateFSubFMF(X, Y, &I);
1259 
1260  // Bail out if we just created a denormal constant.
1261  // TODO: This is copied from a previous implementation. Is it necessary?
1262  const APFloat *C;
1263  if (match(XY, m_APFloat(C)) && !C->isNormal())
1264  return nullptr;
1265 
1266  return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1267  : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1268 }
1269 
1271  if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1272  I.getFastMathFlags(),
1273  SQ.getWithInstruction(&I)))
1274  return replaceInstUsesWith(I, V);
1275 
1276  if (SimplifyAssociativeOrCommutative(I))
1277  return &I;
1278 
1279  if (Instruction *X = foldVectorBinop(I))
1280  return X;
1281 
1282  if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1283  return FoldedFAdd;
1284 
1285  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1286  Value *X;
1287  // (-X) + Y --> Y - X
1288  if (match(LHS, m_FNeg(m_Value(X))))
1289  return BinaryOperator::CreateFSubFMF(RHS, X, &I);
1290  // Y + (-X) --> Y - X
1291  if (match(RHS, m_FNeg(m_Value(X))))
1292  return BinaryOperator::CreateFSubFMF(LHS, X, &I);
1293 
1294  // Check for (fadd double (sitofp x), y), see if we can merge this into an
1295  // integer add followed by a promotion.
1296  if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1297  Value *LHSIntVal = LHSConv->getOperand(0);
1298  Type *FPType = LHSConv->getType();
1299 
1300  // TODO: This check is overly conservative. In many cases known bits
1301  // analysis can tell us that the result of the addition has less significant
1302  // bits than the integer type can hold.
1303  auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1304  Type *FScalarTy = FTy->getScalarType();
1305  Type *IScalarTy = ITy->getScalarType();
1306 
1307  // Do we have enough bits in the significand to represent the result of
1308  // the integer addition?
1309  unsigned MaxRepresentableBits =
1311  return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1312  };
1313 
1314  // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1315  // ... if the constant fits in the integer value. This is useful for things
1316  // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1317  // requires a constant pool load, and generally allows the add to be better
1318  // instcombined.
1319  if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1320  if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1321  Constant *CI =
1322  ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1323  if (LHSConv->hasOneUse() &&
1324  ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1325  willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1326  // Insert the new integer add.
1327  Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1328  return new SIToFPInst(NewAdd, I.getType());
1329  }
1330  }
1331 
1332  // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1333  if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1334  Value *RHSIntVal = RHSConv->getOperand(0);
1335  // It's enough to check LHS types only because we require int types to
1336  // be the same for this transform.
1337  if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1338  // Only do this if x/y have the same type, if at least one of them has a
1339  // single use (so we don't increase the number of int->fp conversions),
1340  // and if the integer add will not overflow.
1341  if (LHSIntVal->getType() == RHSIntVal->getType() &&
1342  (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1343  willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1344  // Insert the new integer add.
1345  Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1346  return new SIToFPInst(NewAdd, I.getType());
1347  }
1348  }
1349  }
1350  }
1351 
1352  // Handle specials cases for FAdd with selects feeding the operation
1353  if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1354  return replaceInstUsesWith(I, V);
1355 
1356  if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1357  if (Instruction *F = factorizeFAddFSub(I, Builder))
1358  return F;
1359  if (Value *V = FAddCombine(Builder).simplify(&I))
1360  return replaceInstUsesWith(I, V);
1361  }
1362 
1363  return nullptr;
1364 }
1365 
1366 /// Optimize pointer differences into the same array into a size. Consider:
1367 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1368 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1370  Type *Ty) {
1371  // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1372  // this.
1373  bool Swapped = false;
1374  GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1375 
1376  // For now we require one side to be the base pointer "A" or a constant
1377  // GEP derived from it.
1378  if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1379  // (gep X, ...) - X
1380  if (LHSGEP->getOperand(0) == RHS) {
1381  GEP1 = LHSGEP;
1382  Swapped = false;
1383  } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1384  // (gep X, ...) - (gep X, ...)
1385  if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1386  RHSGEP->getOperand(0)->stripPointerCasts()) {
1387  GEP2 = RHSGEP;
1388  GEP1 = LHSGEP;
1389  Swapped = false;
1390  }
1391  }
1392  }
1393 
1394  if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1395  // X - (gep X, ...)
1396  if (RHSGEP->getOperand(0) == LHS) {
1397  GEP1 = RHSGEP;
1398  Swapped = true;
1399  } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1400  // (gep X, ...) - (gep X, ...)
1401  if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1402  LHSGEP->getOperand(0)->stripPointerCasts()) {
1403  GEP2 = LHSGEP;
1404  GEP1 = RHSGEP;
1405  Swapped = true;
1406  }
1407  }
1408  }
1409 
1410  if (!GEP1)
1411  // No GEP found.
1412  return nullptr;
1413 
1414  if (GEP2) {
1415  // (gep X, ...) - (gep X, ...)
1416  //
1417  // Avoid duplicating the arithmetic if there are more than one non-constant
1418  // indices between the two GEPs and either GEP has a non-constant index and
1419  // multiple users. If zero non-constant index, the result is a constant and
1420  // there is no duplication. If one non-constant index, the result is an add
1421  // or sub with a constant, which is no larger than the original code, and
1422  // there's no duplicated arithmetic, even if either GEP has multiple
1423  // users. If more than one non-constant indices combined, as long as the GEP
1424  // with at least one non-constant index doesn't have multiple users, there
1425  // is no duplication.
1426  unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1427  unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1428  if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1429  ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1430  (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1431  return nullptr;
1432  }
1433  }
1434 
1435  // Emit the offset of the GEP and an intptr_t.
1436  Value *Result = EmitGEPOffset(GEP1);
1437 
1438  // If we had a constant expression GEP on the other side offsetting the
1439  // pointer, subtract it from the offset we have.
1440  if (GEP2) {
1441  Value *Offset = EmitGEPOffset(GEP2);
1442  Result = Builder.CreateSub(Result, Offset);
1443  }
1444 
1445  // If we have p - gep(p, ...) then we have to negate the result.
1446  if (Swapped)
1447  Result = Builder.CreateNeg(Result, "diff.neg");
1448 
1449  return Builder.CreateIntCast(Result, Ty, true);
1450 }
1451 
1453  if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1455  SQ.getWithInstruction(&I)))
1456  return replaceInstUsesWith(I, V);
1457 
1458  if (Instruction *X = foldVectorBinop(I))
1459  return X;
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.
1466  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1467  if (Value *V = dyn_castNegVal(Op1)) {
1469 
1470  if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1471  assert(BO->getOpcode() == Instruction::Sub &&
1472  "Expected a subtraction operator!");
1473  if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1474  Res->setHasNoSignedWrap(true);
1475  } else {
1476  if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1477  Res->setHasNoSignedWrap(true);
1478  }
1479 
1480  return Res;
1481  }
1482 
1483  if (I.getType()->isIntOrIntVectorTy(1))
1484  return BinaryOperator::CreateXor(Op0, Op1);
1485 
1486  // Replace (-1 - A) with (~A).
1487  if (match(Op0, m_AllOnes()))
1488  return BinaryOperator::CreateNot(Op1);
1489 
1490  // (~X) - (~Y) --> Y - X
1491  Value *X, *Y;
1492  if (match(Op0, m_Not(m_Value(X))) && match(Op1, m_Not(m_Value(Y))))
1493  return BinaryOperator::CreateSub(Y, X);
1494 
1495  // (X + -1) - Y --> ~Y + X
1496  if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1497  return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1498 
1499  // Y - (X + 1) --> ~X + Y
1500  if (match(Op1, m_OneUse(m_Add(m_Value(X), m_One()))))
1501  return BinaryOperator::CreateAdd(Builder.CreateNot(X), Op0);
1502 
1503  if (Constant *C = dyn_cast<Constant>(Op0)) {
1504  bool IsNegate = match(C, m_ZeroInt());
1505  Value *X;
1506  if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1507  // 0 - (zext bool) --> sext bool
1508  // C - (zext bool) --> bool ? C - 1 : C
1509  if (IsNegate)
1510  return CastInst::CreateSExtOrBitCast(X, I.getType());
1511  return SelectInst::Create(X, SubOne(C), C);
1512  }
1513  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1514  // 0 - (sext bool) --> zext bool
1515  // C - (sext bool) --> bool ? C + 1 : C
1516  if (IsNegate)
1517  return CastInst::CreateZExtOrBitCast(X, I.getType());
1518  return SelectInst::Create(X, AddOne(C), C);
1519  }
1520 
1521  // C - ~X == X + (1+C)
1522  if (match(Op1, m_Not(m_Value(X))))
1523  return BinaryOperator::CreateAdd(X, AddOne(C));
1524 
1525  // Try to fold constant sub into select arguments.
1526  if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1527  if (Instruction *R = FoldOpIntoSelect(I, SI))
1528  return R;
1529 
1530  // Try to fold constant sub into PHI values.
1531  if (PHINode *PN = dyn_cast<PHINode>(Op1))
1532  if (Instruction *R = foldOpIntoPhi(I, PN))
1533  return R;
1534 
1535  // C-(X+C2) --> (C-C2)-X
1536  Constant *C2;
1537  if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1538  return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1539  }
1540 
1541  const APInt *Op0C;
1542  if (match(Op0, m_APInt(Op0C))) {
1543  unsigned BitWidth = I.getType()->getScalarSizeInBits();
1544 
1545  // -(X >>u 31) -> (X >>s 31)
1546  // -(X >>s 31) -> (X >>u 31)
1547  if (Op0C->isNullValue()) {
1548  Value *X;
1549  const APInt *ShAmt;
1550  if (match(Op1, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1551  *ShAmt == BitWidth - 1) {
1552  Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1553  return BinaryOperator::CreateAShr(X, ShAmtOp);
1554  }
1555  if (match(Op1, m_AShr(m_Value(X), m_APInt(ShAmt))) &&
1556  *ShAmt == BitWidth - 1) {
1557  Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1558  return BinaryOperator::CreateLShr(X, ShAmtOp);
1559  }
1560 
1561  if (Op1->hasOneUse()) {
1562  Value *LHS, *RHS;
1563  SelectPatternFlavor SPF = matchSelectPattern(Op1, LHS, RHS).Flavor;
1564  if (SPF == SPF_ABS || SPF == SPF_NABS) {
1565  // This is a negate of an ABS/NABS pattern. Just swap the operands
1566  // of the select.
1567  SelectInst *SI = cast<SelectInst>(Op1);
1568  Value *TrueVal = SI->getTrueValue();
1569  Value *FalseVal = SI->getFalseValue();
1570  SI->setTrueValue(FalseVal);
1571  SI->setFalseValue(TrueVal);
1572  // Don't swap prof metadata, we didn't change the branch behavior.
1573  return replaceInstUsesWith(I, SI);
1574  }
1575  }
1576  }
1577 
1578  // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1579  // zero.
1580  if (Op0C->isMask()) {
1581  KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1582  if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
1583  return BinaryOperator::CreateXor(Op1, Op0);
1584  }
1585  }
1586 
1587  {
1588  Value *Y;
1589  // X-(X+Y) == -Y X-(Y+X) == -Y
1590  if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1591  return BinaryOperator::CreateNeg(Y);
1592 
1593  // (X-Y)-X == -Y
1594  if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1595  return BinaryOperator::CreateNeg(Y);
1596  }
1597 
1598  // (sub (or A, B), (xor A, B)) --> (and A, B)
1599  {
1600  Value *A, *B;
1601  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1602  match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1603  return BinaryOperator::CreateAnd(A, B);
1604  }
1605 
1606  {
1607  Value *Y;
1608  // ((X | Y) - X) --> (~X & Y)
1609  if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1610  return BinaryOperator::CreateAnd(
1611  Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1612  }
1613 
1614  if (Op1->hasOneUse()) {
1615  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1616  Constant *C = nullptr;
1617 
1618  // (X - (Y - Z)) --> (X + (Z - Y)).
1619  if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1620  return BinaryOperator::CreateAdd(Op0,
1621  Builder.CreateSub(Z, Y, Op1->getName()));
1622 
1623  // (X - (X & Y)) --> (X & ~Y)
1624  if (match(Op1, m_c_And(m_Value(Y), m_Specific(Op0))))
1625  return BinaryOperator::CreateAnd(Op0,
1626  Builder.CreateNot(Y, Y->getName() + ".not"));
1627 
1628  // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow.
1629  if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1630  C->isNotMinSignedValue() && !C->isOneValue())
1631  return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1632 
1633  // 0 - (X << Y) -> (-X << Y) when X is freely negatable.
1634  if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1635  if (Value *XNeg = dyn_castNegVal(X))
1636  return BinaryOperator::CreateShl(XNeg, Y);
1637 
1638  // Subtracting -1/0 is the same as adding 1/0:
1639  // sub [nsw] Op0, sext(bool Y) -> add [nsw] Op0, zext(bool Y)
1640  // 'nuw' is dropped in favor of the canonical form.
1641  if (match(Op1, m_SExt(m_Value(Y))) &&
1642  Y->getType()->getScalarSizeInBits() == 1) {
1643  Value *Zext = Builder.CreateZExt(Y, I.getType());
1644  BinaryOperator *Add = BinaryOperator::CreateAdd(Op0, Zext);
1646  return Add;
1647  }
1648 
1649  // X - A*-B -> X + A*B
1650  // X - -A*B -> X + A*B
1651  Value *A, *B;
1652  Constant *CI;
1653  if (match(Op1, m_c_Mul(m_Value(A), m_Neg(m_Value(B)))))
1654  return BinaryOperator::CreateAdd(Op0, Builder.CreateMul(A, B));
1655 
1656  // X - A*CI -> X + A*-CI
1657  // No need to handle commuted multiply because multiply handling will
1658  // ensure constant will be move to the right hand side.
1659  if (match(Op1, m_Mul(m_Value(A), m_Constant(CI)))) {
1660  Value *NewMul = Builder.CreateMul(A, ConstantExpr::getNeg(CI));
1661  return BinaryOperator::CreateAdd(Op0, NewMul);
1662  }
1663  }
1664 
1665  {
1666  // ~A - Min/Max(~A, O) -> Max/Min(A, ~O) - A
1667  // ~A - Min/Max(O, ~A) -> Max/Min(A, ~O) - A
1668  // Min/Max(~A, O) - ~A -> A - Max/Min(A, ~O)
1669  // Min/Max(O, ~A) - ~A -> A - Max/Min(A, ~O)
1670  // So long as O here is freely invertible, this will be neutral or a win.
1671  Value *LHS, *RHS, *A;
1672  Value *NotA = Op0, *MinMax = Op1;
1673  SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
1674  if (!SelectPatternResult::isMinOrMax(SPF)) {
1675  NotA = Op1;
1676  MinMax = Op0;
1677  SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
1678  }
1680  match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
1681  if (NotA == LHS)
1682  std::swap(LHS, RHS);
1683  // LHS is now O above and expected to have at least 2 uses (the min/max)
1684  // NotA is epected to have 2 uses from the min/max and 1 from the sub.
1685  if (IsFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
1686  !NotA->hasNUsesOrMore(4)) {
1687  // Note: We don't generate the inverse max/min, just create the not of
1688  // it and let other folds do the rest.
1689  Value *Not = Builder.CreateNot(MinMax);
1690  if (NotA == Op0)
1691  return BinaryOperator::CreateSub(Not, A);
1692  else
1693  return BinaryOperator::CreateSub(A, Not);
1694  }
1695  }
1696  }
1697 
1698  // Optimize pointer differences into the same array into a size. Consider:
1699  // &A[10] - &A[0]: we should compile this to "10".
1700  Value *LHSOp, *RHSOp;
1701  if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1702  match(Op1, m_PtrToInt(m_Value(RHSOp))))
1703  if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1704  return replaceInstUsesWith(I, Res);
1705 
1706  // trunc(p)-trunc(q) -> trunc(p-q)
1707  if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1708  match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1709  if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1710  return replaceInstUsesWith(I, Res);
1711 
1712  // Canonicalize a shifty way to code absolute value to the common pattern.
1713  // There are 2 potential commuted variants.
1714  // We're relying on the fact that we only do this transform when the shift has
1715  // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
1716  // instructions).
1717  Value *A;
1718  const APInt *ShAmt;
1719  Type *Ty = I.getType();
1720  if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
1721  Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
1722  match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
1723  // B = ashr i32 A, 31 ; smear the sign bit
1724  // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
1725  // --> (A < 0) ? -A : A
1726  Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
1727  // Copy the nuw/nsw flags from the sub to the negate.
1728  Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
1729  I.hasNoSignedWrap());
1730  return SelectInst::Create(Cmp, Neg, A);
1731  }
1732 
1733  if (Instruction *Ext = narrowMathIfNoOverflow(I))
1734  return Ext;
1735 
1736  bool Changed = false;
1737  if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1738  Changed = true;
1739  I.setHasNoSignedWrap(true);
1740  }
1741  if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1742  Changed = true;
1743  I.setHasNoUnsignedWrap(true);
1744  }
1745 
1746  return Changed ? &I : nullptr;
1747 }
1748 
1750  if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
1751  I.getFastMathFlags(),
1752  SQ.getWithInstruction(&I)))
1753  return replaceInstUsesWith(I, V);
1754 
1755  if (Instruction *X = foldVectorBinop(I))
1756  return X;
1757 
1758  // Subtraction from -0.0 is the canonical form of fneg.
1759  // fsub nsz 0, X ==> fsub nsz -0.0, X
1760  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1761  if (I.hasNoSignedZeros() && match(Op0, m_PosZeroFP()))
1762  return BinaryOperator::CreateFNegFMF(Op1, &I);
1763 
1764  Value *X, *Y;
1765  Constant *C;
1766 
1767  // Fold negation into constant operand. This is limited with one-use because
1768  // fneg is assumed better for analysis and cheaper in codegen than fmul/fdiv.
1769  // -(X * C) --> X * (-C)
1770  if (match(&I, m_FNeg(m_OneUse(m_FMul(m_Value(X), m_Constant(C))))))
1772  // -(X / C) --> X / (-C)
1773  if (match(&I, m_FNeg(m_OneUse(m_FDiv(m_Value(X), m_Constant(C))))))
1775  // -(C / X) --> (-C) / X
1776  if (match(&I, m_FNeg(m_OneUse(m_FDiv(m_Constant(C), m_Value(X))))))
1778 
1779  // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
1780  // Canonicalize to fadd to make analysis easier.
1781  // This can also help codegen because fadd is commutative.
1782  // Note that if this fsub was really an fneg, the fadd with -0.0 will get
1783  // killed later. We still limit that particular transform with 'hasOneUse'
1784  // because an fneg is assumed better/cheaper than a generic fsub.
1785  if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
1786  if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
1787  Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
1788  return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
1789  }
1790  }
1791 
1792  if (isa<Constant>(Op0))
1793  if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1794  if (Instruction *NV = FoldOpIntoSelect(I, SI))
1795  return NV;
1796 
1797  // X - C --> X + (-C)
1798  // But don't transform constant expressions because there's an inverse fold
1799  // for X + (-Y) --> X - Y.
1800  if (match(Op1, m_Constant(C)) && !isa<ConstantExpr>(Op1))
1802 
1803  // X - (-Y) --> X + Y
1804  if (match(Op1, m_FNeg(m_Value(Y))))
1805  return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
1806 
1807  // Similar to above, but look through a cast of the negated value:
1808  // X - (fptrunc(-Y)) --> X + fptrunc(Y)
1809  Type *Ty = I.getType();
1810  if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
1811  return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
1812 
1813  // X - (fpext(-Y)) --> X + fpext(Y)
1814  if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
1815  return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
1816 
1817  // Handle special cases for FSub with selects feeding the operation
1818  if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
1819  return replaceInstUsesWith(I, V);
1820 
1821  if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1822  // (Y - X) - Y --> -X
1823  if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
1824  return BinaryOperator::CreateFNegFMF(X, &I);
1825 
1826  // Y - (X + Y) --> -X
1827  // Y - (Y + X) --> -X
1828  if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
1829  return BinaryOperator::CreateFNegFMF(X, &I);
1830 
1831  // (X * C) - X --> X * (C - 1.0)
1832  if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
1833  Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
1834  return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
1835  }
1836  // X - (X * C) --> X * (1.0 - C)
1837  if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
1838  Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
1839  return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
1840  }
1841 
1842  if (Instruction *F = factorizeFAddFSub(I, Builder))
1843  return F;
1844 
1845  // TODO: This performs reassociative folds for FP ops. Some fraction of the
1846  // functionality has been subsumed by simple pattern matching here and in
1847  // InstSimplify. We should let a dedicated reassociation pass handle more
1848  // complex pattern matching and remove this from InstCombine.
1849  if (Value *V = FAddCombine(Builder).simplify(&I))
1850  return replaceInstUsesWith(I, V);
1851  }
1852 
1853  return nullptr;
1854 }
static BinaryOperator * CreateFMulFMF(Value *V1, Value *V2, BinaryOperator *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:177
Value * EmitGEPOffset(IRBuilderTy *Builder, const DataLayout &DL, User *GEP, bool NoAssumptions=false)
Given a getelementptr instruction/constantexpr, emit the code necessary to compute the offset from th...
Definition: Local.h:29
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:750
uint64_t CallInst * C
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, OptimizationRemarkEmitter *ORE=nullptr, bool UseInstrInfo=true)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
std::string & operator+=(std::string &buffer, StringRef string)
Definition: StringRef.h:921
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
void setFastMathFlags(FastMathFlags FMF)
Convenience function for setting multiple fast-math flags on this instruction, which must be an opera...
bool isSignMask() const
Check if the APInt&#39;s value is returned by getSignMask.
Definition: APInt.h:473
static bool isConstant(const MachineInstr &MI)
static bool IsFreeToInvert(Value *V, bool WillInvertAllUses)
Return true if the specified value is free to invert (apply ~ to).
bool hasNoSignedZeros() const
Determine whether the no-signed-zeros flag is set.
APInt sext(unsigned width) const
Sign extend to a new width.
Definition: APInt.cpp:834
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:655
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:377
DiagnosticInfoOptimizationBase::Argument NV
static BinaryOperator * CreateNot(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:311
BinaryOp_match< LHS, RHS, Instruction::Xor, true > m_c_Xor(const LHS &L, const RHS &R)
Matches an Xor with LHS and RHS in either order.
static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned)
BinaryOp_match< LHS, RHS, Instruction::FDiv > m_FDiv(const LHS &L, const RHS &R)
Definition: PatternMatch.h:726
BinaryOp_match< LHS, RHS, Instruction::SRem > m_SRem(const LHS &L, const RHS &R)
Definition: PatternMatch.h:738
BinaryOp_match< LHS, RHS, Instruction::FAdd, true > m_c_FAdd(const LHS &L, const RHS &R)
Matches FAdd with LHS and RHS in either order.
This class represents zero extension of integer types.
BinaryOp_match< LHS, RHS, Instruction::Mul > m_Mul(const LHS &L, const RHS &R)
Definition: PatternMatch.h:702
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:91
const Value * getTrueValue() const
BinaryOp_match< LHS, RHS, Instruction::AShr > m_AShr(const LHS &L, const RHS &R)
Definition: PatternMatch.h:780
static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned)
Instruction * visitFSub(BinaryOperator &I)
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr="", Instruction *InsertBefore=nullptr, Instruction *MDFrom=nullptr)
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:811
F(f)
const fltSemantics & getSemantics() const
Definition: APFloat.h:1155
BinaryOp_match< LHS, RHS, Instruction::FSub > m_FSub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:661
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2227
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:230
void changeSign()
Definition: APFloat.h:1050
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
bool hasNoSignedWrap() const
Determine whether the no signed wrap flag is set.
static BinaryOperator * CreateFSubFMF(Value *V1, Value *V2, BinaryOperator *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:172
cst_pred_ty< is_zero_int > m_ZeroInt()
Match an integer 0 or a vector with all elements equal to 0.
Definition: PatternMatch.h:365
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1509
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:130
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:268
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2216
unsigned countTrailingZeros() const
Count the number of trailing zero bits.
Definition: APInt.h:1632
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
BinaryOp_match< LHS, RHS, Instruction::Xor > m_Xor(const LHS &L, const RHS &R)
Definition: PatternMatch.h:762
This class represents the LLVM &#39;select&#39; instruction.
Absolute value.
roundingMode
IEEE-754R 4.3: Rounding-direction attributes.
Definition: APFloat.h:174
CastClass_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:743
static Constant * AddOne(Constant *C)
Add one to a Constant.
BinaryOp_match< LHS, RHS, Instruction::FMul, true > m_c_FMul(const LHS &L, const RHS &R)
Matches FMul with LHS and RHS in either order.
Value * CreateFAddFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder&#39;s default FMF.
Definition: IRBuilder.h:1182
bool MaskedValueIsZero(const Value *V, const APInt &Mask, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Return true if &#39;V & Mask&#39; is known to be zero.
This file implements a class to represent arbitrary precision integral constant values and operations...
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:643
OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWAdd(const LHS &L, const RHS &R)
Definition: PatternMatch.h:846
Value * SimplifyFAddInst(Value *LHS, Value *RHS, FastMathFlags FMF, const SimplifyQuery &Q)
Given operands for an FAdd, fold the result or return null.
static BinaryOperator * CreateFAddFMF(Value *V1, Value *V2, BinaryOperator *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:167
FastMathFlags getFastMathFlags() const
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:181
CastClass_match< OpTy, Instruction::FPExt > m_FPExt(const OpTy &Op)
Matches FPExt.
CastClass_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
#define T
CastClass_match< OpTy, Instruction::FPTrunc > m_FPTrunc(const OpTy &Op)
Matches FPTrunc.
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:83
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:138
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
cstfp_pred_ty< is_pos_zero_fp > m_PosZeroFP()
Match a floating-point positive zero.
Definition: PatternMatch.h:446
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:991
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
Value * getOperand(unsigned i) const
Definition: User.h:170
bool CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, unsigned Depth=0)
Return true if we can prove that the specified FP value is never equal to -0.0.
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1142
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:304
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Get a value with high bits set.
Definition: APInt.h:636
Value * OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty)
Optimize pointer differences into the same array into a size.
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:63
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:364
#define P(N)
static Constant * getFNeg(Constant *C)
Definition: Constants.cpp:2204
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
Definition: PatternMatch.h:774
bool hasNUsesOrMore(unsigned N) const
Return true if this value has N users or more.
Definition: Value.cpp:136
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:177
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:308
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
Definition: PatternMatch.h:720
BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)
Matches a Add with LHS and RHS in either order.
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:756
constexpr bool isInt(int64_t x)
Checks if an integer fits into the given bit width.
Definition: MathExtras.h:299
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
This file contains the declarations for the subclasses of Constant, which represent the different fla...
BinaryOp_match< LHS, RHS, Instruction::And, true > m_c_And(const LHS &L, const RHS &R)
Matches an And with LHS and RHS in either order.
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:264
bool isMask(unsigned numBits) const
Definition: APInt.h:495
bool isOneValue() const
Determine if this is a value of 1.
Definition: APInt.h:411
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:310
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:503
Value * SimplifyAddInst(Value *LHS, Value *RHS, bool isNSW, bool isNUW, const SimplifyQuery &Q)
Given operands for an Add, fold the result or return null.
bool isMinSignedValue() const
Determine if this is the smallest signed value.
Definition: APInt.h:443
This file declares a class to represent arbitrary precision floating point values and provide a varie...
BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)
Definition: PatternMatch.h:768
Value * CreateFSubFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder&#39;s default FMF.
Definition: IRBuilder.h:1199
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:959
static Instruction * factorizeFAddFSub(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Factor a common operand out of fadd/fsub of fmul/fdiv.
Instruction * visitFAdd(BinaryOperator &I)
const Value * getCondition() const
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:322
NUW NUW NUW NUW Exact static Exact BinaryOperator * CreateNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static CastInst * CreateZExtOrBitCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a ZExt or BitCast cast instruction.
deferredval_ty< Value > m_Deferred(Value *const &V)
A commutative-friendly version of m_Specific().
Definition: PatternMatch.h:516
const APFloat & getValueAPF() const
Definition: Constants.h:299
CastClass_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
Floating point maxnum.
static Constant * getSIToFP(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1691
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
hexagon bit simplify
BinaryOp_match< LHS, RHS, Instruction::Or, true > m_c_Or(const LHS &L, const RHS &R)
Matches an Or with LHS and RHS in either order.
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
SelectPatternFlavor Flavor
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type...
Definition: Type.cpp:130
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:847
SelectPatternFlavor
Specific patterns of select instructions we can match.
BinaryOp_match< LHS, RHS, Instruction::URem > m_URem(const LHS &L, const RHS &R)
Definition: PatternMatch.h:732
constexpr size_t array_lengthof(T(&)[N])
Find the length of an array.
Definition: STLExtras.h:906
BinaryOp_match< LHS, RHS, Instruction::UDiv > m_UDiv(const LHS &L, const RHS &R)
Definition: PatternMatch.h:714
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:621
BinaryOp_match< LHS, RHS, Instruction::Mul, true > m_c_Mul(const LHS &L, const RHS &R)
Matches a Mul with LHS and RHS in either order.
static Constant * get(Type *Ty, double V)
This returns a ConstantFP, or a vector containing a splat of a ConstantFP, for the specified value in...
Definition: Constants.cpp:684
BinaryOp_match< LHS, RHS, Instruction::FMul > m_FMul(const LHS &L, const RHS &R)
Definition: PatternMatch.h:708
static unsigned int semanticsPrecision(const fltSemantics &)
Definition: APFloat.cpp:155
static BinaryOperator * CreateFDivFMF(Value *V1, Value *V2, BinaryOperator *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:182
void setOperand(unsigned i, Value *Val)
Definition: User.h:175
unsigned logBase2() const
Definition: APInt.h:1748
BinaryOp_match< cst_pred_ty< is_zero_int >, ValTy, Instruction::Sub > m_Neg(const ValTy &V)
Matches a &#39;Neg&#39; as &#39;sub 0, V&#39;.
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:941
Class for arbitrary precision integers.
Definition: APInt.h:70
Value * SimplifyFSubInst(Value *LHS, Value *RHS, FastMathFlags FMF, const SimplifyQuery &Q)
Given operands for an FSub, fold the result or return null.
bool isPowerOf2() const
Check if this APInt&#39;s value is a power of two greater than zero.
Definition: APInt.h:464
bool sge(const APInt &RHS) const
Signed greater or equal comparison.
Definition: APInt.h:1309
CastClass_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
This union template exposes a suitably aligned and sized character array member which can hold elemen...
Definition: AlignOf.h:138
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1063
const Value * getFalseValue() const
static Constant * getFSub(Constant *C1, Constant *C2)
Definition: Constants.cpp:2234
static Instruction * canonicalizeLowbitMask(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Fold (1 << NBits) - 1 Into: ~(-(1 << NBits)) Because a &#39;not&#39; is better for bit-tracking analysis and ...
static Constant * getNeg(Constant *C, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2197
static Value * checkForNegativeOperand(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass&#39;s ...
opStatus add(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:941
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:546
FNeg_match< OpTy > m_FNeg(const OpTy &X)
Match &#39;fneg X&#39; as &#39;fsub -0.0, X&#39;.
Definition: PatternMatch.h:690
static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned)
void setTrueValue(Value *V)
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
Instruction * visitAdd(BinaryOperator &I)
static bool isMinOrMax(SelectPatternFlavor SPF)
When implementing this min/max pattern as fcmp; select, does the fcmp have to be ordered?
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:302
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:215
APInt smul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1907
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
bool haveNoCommonBitsSet(const Value *LHS, const Value *RHS, const DataLayout &DL, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Return true if LHS and RHS have no common bits set.
bool isNormal() const
Definition: APFloat.h:1151
static Constant * getZeroValueForNegation(Type *Ty)
Floating point negation must be implemented with f(x) = -0.0 - x.
Definition: Constants.cpp:748
static BinaryOperator * CreateFNegFMF(Value *Op, BinaryOperator *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:192
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
void setFalseValue(Value *V)
bool hasNoUnsignedWrap() const
Determine whether the no unsigned wrap flag is set.
static bool MatchMul(Value *E, Value *&Op, APInt &C)
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1124
APInt umul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1917
bool isOneValue() const
Returns true if the value is one.
Definition: Constants.cpp:126
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nuw flag on this instruction, which must be an operator which supports this flag...
static CastInst * CreateSExtOrBitCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a SExt or BitCast cast instruction.
This class represents a cast from signed integer to floating point.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
LLVM Value Representation.
Definition: Value.h:73
This file provides internal interfaces used to implement the InstCombine.
SelectPatternResult matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, Instruction::CastOps *CastOp=nullptr, unsigned Depth=0)
Pattern match integer [SU]MIN, [SU]MAX and ABS idioms, returning the kind and providing the out param...
cst_pred_ty< is_one > m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:356
bool hasAllowReassoc() const
Determine whether the allow-reassociation flag is set.
std::underlying_type< E >::type Mask()
Get a bitmask with 1s in all places up to the high-order bit of E&#39;s largest value.
Definition: BitmaskEnum.h:81
static Constant * getFPToSI(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1713
Value * SimplifySubInst(Value *LHS, Value *RHS, bool isNSW, bool isNUW, const SimplifyQuery &Q)
Given operands for a Sub, fold the result or return null.
bool isNotMinSignedValue() const
Return true if the value is not the smallest signed value.
Definition: Constants.cpp:178
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:413
unsigned countNonConstantIndices() const
Definition: Operator.h:522
Instruction * visitSub(BinaryOperator &I)
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
BinaryOp_match< ValTy, cst_pred_ty< is_all_ones >, Instruction::Xor, true > m_Not(const ValTy &V)
Matches a &#39;Not&#39; as &#39;xor V, -1&#39; or &#39;xor -1, V&#39;.
bool isNullValue() const
Determine if all bits are clear.
Definition: APInt.h:406
const fltSemantics & getFltSemantics() const
Definition: Type.h:169
static Constant * getXor(Constant *C1, Constant *C2)
Definition: Constants.cpp:2283