LLVM  6.0.0svn
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  Value *performFactorization(Instruction *I);
190 
191  /// Convert given addend to a Value
192  Value *createAddendVal(const FAddend &A, bool& NeedNeg);
193 
194  /// Return the number of instructions needed to emit the N-ary addition.
195  unsigned calcInstrNumber(const AddendVect& Vect);
196 
197  Value *createFSub(Value *Opnd0, Value *Opnd1);
198  Value *createFAdd(Value *Opnd0, Value *Opnd1);
199  Value *createFMul(Value *Opnd0, Value *Opnd1);
200  Value *createFDiv(Value *Opnd0, Value *Opnd1);
201  Value *createFNeg(Value *V);
202  Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
203  void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
204 
205  // Debugging stuff are clustered here.
206  #ifndef NDEBUG
207  unsigned CreateInstrNum;
208  void initCreateInstNum() { CreateInstrNum = 0; }
209  void incCreateInstNum() { CreateInstrNum++; }
210  #else
211  void initCreateInstNum() {}
212  void incCreateInstNum() {}
213  #endif
214 
215  InstCombiner::BuilderTy &Builder;
216  Instruction *Instr = nullptr;
217  };
218 
219 } // end anonymous namespace
220 
221 //===----------------------------------------------------------------------===//
222 //
223 // Implementation of
224 // {FAddendCoef, FAddend, FAddition, FAddCombine}.
225 //
226 //===----------------------------------------------------------------------===//
227 FAddendCoef::~FAddendCoef() {
228  if (BufHasFpVal)
229  getFpValPtr()->~APFloat();
230 }
231 
232 void FAddendCoef::set(const APFloat& C) {
233  APFloat *P = getFpValPtr();
234 
235  if (isInt()) {
236  // As the buffer is meanless byte stream, we cannot call
237  // APFloat::operator=().
238  new(P) APFloat(C);
239  } else
240  *P = C;
241 
242  IsFp = BufHasFpVal = true;
243 }
244 
245 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
246  if (!isInt())
247  return;
248 
249  APFloat *P = getFpValPtr();
250  if (IntVal > 0)
251  new(P) APFloat(Sem, IntVal);
252  else {
253  new(P) APFloat(Sem, 0 - IntVal);
254  P->changeSign();
255  }
256  IsFp = BufHasFpVal = true;
257 }
258 
259 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
260  if (Val >= 0)
261  return APFloat(Sem, Val);
262 
263  APFloat T(Sem, 0 - Val);
264  T.changeSign();
265 
266  return T;
267 }
268 
269 void FAddendCoef::operator=(const FAddendCoef &That) {
270  if (That.isInt())
271  set(That.IntVal);
272  else
273  set(That.getFpVal());
274 }
275 
276 void FAddendCoef::operator+=(const FAddendCoef &That) {
278  if (isInt() == That.isInt()) {
279  if (isInt())
280  IntVal += That.IntVal;
281  else
282  getFpVal().add(That.getFpVal(), RndMode);
283  return;
284  }
285 
286  if (isInt()) {
287  const APFloat &T = That.getFpVal();
288  convertToFpType(T.getSemantics());
289  getFpVal().add(T, RndMode);
290  return;
291  }
292 
293  APFloat &T = getFpVal();
294  T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
295 }
296 
297 void FAddendCoef::operator*=(const FAddendCoef &That) {
298  if (That.isOne())
299  return;
300 
301  if (That.isMinusOne()) {
302  negate();
303  return;
304  }
305 
306  if (isInt() && That.isInt()) {
307  int Res = IntVal * (int)That.IntVal;
308  assert(!insaneIntVal(Res) && "Insane int value");
309  IntVal = Res;
310  return;
311  }
312 
313  const fltSemantics &Semantic =
314  isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
315 
316  if (isInt())
317  convertToFpType(Semantic);
318  APFloat &F0 = getFpVal();
319 
320  if (That.isInt())
321  F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
323  else
324  F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
325 }
326 
327 void FAddendCoef::negate() {
328  if (isInt())
329  IntVal = 0 - IntVal;
330  else
331  getFpVal().changeSign();
332 }
333 
334 Value *FAddendCoef::getValue(Type *Ty) const {
335  return isInt() ?
336  ConstantFP::get(Ty, float(IntVal)) :
337  ConstantFP::get(Ty->getContext(), getFpVal());
338 }
339 
340 // The definition of <Val> Addends
341 // =========================================
342 // A + B <1, A>, <1,B>
343 // A - B <1, A>, <1,B>
344 // 0 - B <-1, B>
345 // C * A, <C, A>
346 // A + C <1, A> <C, NULL>
347 // 0 +/- 0 <0, NULL> (corner case)
348 //
349 // Legend: A and B are not constant, C is constant
350 unsigned FAddend::drillValueDownOneStep
351  (Value *Val, FAddend &Addend0, FAddend &Addend1) {
352  Instruction *I = nullptr;
353  if (!Val || !(I = dyn_cast<Instruction>(Val)))
354  return 0;
355 
356  unsigned Opcode = I->getOpcode();
357 
358  if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
359  ConstantFP *C0, *C1;
360  Value *Opnd0 = I->getOperand(0);
361  Value *Opnd1 = I->getOperand(1);
362  if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
363  Opnd0 = nullptr;
364 
365  if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
366  Opnd1 = nullptr;
367 
368  if (Opnd0) {
369  if (!C0)
370  Addend0.set(1, Opnd0);
371  else
372  Addend0.set(C0, nullptr);
373  }
374 
375  if (Opnd1) {
376  FAddend &Addend = Opnd0 ? Addend1 : Addend0;
377  if (!C1)
378  Addend.set(1, Opnd1);
379  else
380  Addend.set(C1, nullptr);
381  if (Opcode == Instruction::FSub)
382  Addend.negate();
383  }
384 
385  if (Opnd0 || Opnd1)
386  return Opnd0 && Opnd1 ? 2 : 1;
387 
388  // Both operands are zero. Weird!
389  Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
390  return 1;
391  }
392 
393  if (I->getOpcode() == Instruction::FMul) {
394  Value *V0 = I->getOperand(0);
395  Value *V1 = I->getOperand(1);
396  if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
397  Addend0.set(C, V1);
398  return 1;
399  }
400 
401  if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
402  Addend0.set(C, V0);
403  return 1;
404  }
405  }
406 
407  return 0;
408 }
409 
410 // Try to break *this* addend into two addends. e.g. Suppose this addend is
411 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
412 // i.e. <2.3, X> and <2.3, Y>.
413 unsigned FAddend::drillAddendDownOneStep
414  (FAddend &Addend0, FAddend &Addend1) const {
415  if (isConstant())
416  return 0;
417 
418  unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
419  if (!BreakNum || Coeff.isOne())
420  return BreakNum;
421 
422  Addend0.Scale(Coeff);
423 
424  if (BreakNum == 2)
425  Addend1.Scale(Coeff);
426 
427  return BreakNum;
428 }
429 
430 // Try to perform following optimization on the input instruction I. Return the
431 // simplified expression if was successful; otherwise, return 0.
432 //
433 // Instruction "I" is Simplified into
434 // -------------------------------------------------------
435 // (x * y) +/- (x * z) x * (y +/- z)
436 // (y / x) +/- (z / x) (y +/- z) / x
437 Value *FAddCombine::performFactorization(Instruction *I) {
438  assert((I->getOpcode() == Instruction::FAdd ||
439  I->getOpcode() == Instruction::FSub) && "Expect add/sub");
440 
443 
444  if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
445  return nullptr;
446 
447  bool isMpy = false;
448  if (I0->getOpcode() == Instruction::FMul)
449  isMpy = true;
450  else if (I0->getOpcode() != Instruction::FDiv)
451  return nullptr;
452 
453  Value *Opnd0_0 = I0->getOperand(0);
454  Value *Opnd0_1 = I0->getOperand(1);
455  Value *Opnd1_0 = I1->getOperand(0);
456  Value *Opnd1_1 = I1->getOperand(1);
457 
458  // Input Instr I Factor AddSub0 AddSub1
459  // ----------------------------------------------
460  // (x*y) +/- (x*z) x y z
461  // (y/x) +/- (z/x) x y z
462  Value *Factor = nullptr;
463  Value *AddSub0 = nullptr, *AddSub1 = nullptr;
464 
465  if (isMpy) {
466  if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
467  Factor = Opnd0_0;
468  else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
469  Factor = Opnd0_1;
470 
471  if (Factor) {
472  AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
473  AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
474  }
475  } else if (Opnd0_1 == Opnd1_1) {
476  Factor = Opnd0_1;
477  AddSub0 = Opnd0_0;
478  AddSub1 = Opnd1_0;
479  }
480 
481  if (!Factor)
482  return nullptr;
483 
484  FastMathFlags Flags;
485  Flags.setFast();
486  if (I0) Flags &= I->getFastMathFlags();
487  if (I1) Flags &= I->getFastMathFlags();
488 
489  // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
490  Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
491  createFAdd(AddSub0, AddSub1) :
492  createFSub(AddSub0, AddSub1);
493  if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
494  const APFloat &F = CFP->getValueAPF();
495  if (!F.isNormal())
496  return nullptr;
497  } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
498  II->setFastMathFlags(Flags);
499 
500  if (isMpy) {
501  Value *RI = createFMul(Factor, NewAddSub);
502  if (Instruction *II = dyn_cast<Instruction>(RI))
503  II->setFastMathFlags(Flags);
504  return RI;
505  }
506 
507  Value *RI = createFDiv(NewAddSub, Factor);
508  if (Instruction *II = dyn_cast<Instruction>(RI))
509  II->setFastMathFlags(Flags);
510  return RI;
511 }
512 
514  assert(I->isFast() && "Expected 'fast' instruction");
515 
516  // Currently we are not able to handle vector type.
517  if (I->getType()->isVectorTy())
518  return nullptr;
519 
520  assert((I->getOpcode() == Instruction::FAdd ||
521  I->getOpcode() == Instruction::FSub) && "Expect add/sub");
522 
523  // Save the instruction before calling other member-functions.
524  Instr = I;
525 
526  FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
527 
528  unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
529 
530  // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
531  unsigned Opnd0_ExpNum = 0;
532  unsigned Opnd1_ExpNum = 0;
533 
534  if (!Opnd0.isConstant())
535  Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
536 
537  // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
538  if (OpndNum == 2 && !Opnd1.isConstant())
539  Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
540 
541  // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
542  if (Opnd0_ExpNum && Opnd1_ExpNum) {
543  AddendVect AllOpnds;
544  AllOpnds.push_back(&Opnd0_0);
545  AllOpnds.push_back(&Opnd1_0);
546  if (Opnd0_ExpNum == 2)
547  AllOpnds.push_back(&Opnd0_1);
548  if (Opnd1_ExpNum == 2)
549  AllOpnds.push_back(&Opnd1_1);
550 
551  // Compute instruction quota. We should save at least one instruction.
552  unsigned InstQuota = 0;
553 
554  Value *V0 = I->getOperand(0);
555  Value *V1 = I->getOperand(1);
556  InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
557  (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
558 
559  if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
560  return R;
561  }
562 
563  if (OpndNum != 2) {
564  // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
565  // splitted into two addends, say "V = X - Y", the instruction would have
566  // been optimized into "I = Y - X" in the previous steps.
567  //
568  const FAddendCoef &CE = Opnd0.getCoef();
569  return CE.isOne() ? Opnd0.getSymVal() : nullptr;
570  }
571 
572  // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
573  if (Opnd1_ExpNum) {
574  AddendVect AllOpnds;
575  AllOpnds.push_back(&Opnd0);
576  AllOpnds.push_back(&Opnd1_0);
577  if (Opnd1_ExpNum == 2)
578  AllOpnds.push_back(&Opnd1_1);
579 
580  if (Value *R = simplifyFAdd(AllOpnds, 1))
581  return R;
582  }
583 
584  // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
585  if (Opnd0_ExpNum) {
586  AddendVect AllOpnds;
587  AllOpnds.push_back(&Opnd1);
588  AllOpnds.push_back(&Opnd0_0);
589  if (Opnd0_ExpNum == 2)
590  AllOpnds.push_back(&Opnd0_1);
591 
592  if (Value *R = simplifyFAdd(AllOpnds, 1))
593  return R;
594  }
595 
596  // step 6: Try factorization as the last resort,
597  return performFactorization(I);
598 }
599 
600 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
601  unsigned AddendNum = Addends.size();
602  assert(AddendNum <= 4 && "Too many addends");
603 
604  // For saving intermediate results;
605  unsigned NextTmpIdx = 0;
606  FAddend TmpResult[3];
607 
608  // Points to the constant addend of the resulting simplified expression.
609  // If the resulting expr has constant-addend, this constant-addend is
610  // desirable to reside at the top of the resulting expression tree. Placing
611  // constant close to supper-expr(s) will potentially reveal some optimization
612  // opportunities in super-expr(s).
613  const FAddend *ConstAdd = nullptr;
614 
615  // Simplified addends are placed <SimpVect>.
616  AddendVect SimpVect;
617 
618  // The outer loop works on one symbolic-value at a time. Suppose the input
619  // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
620  // The symbolic-values will be processed in this order: x, y, z.
621  for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
622 
623  const FAddend *ThisAddend = Addends[SymIdx];
624  if (!ThisAddend) {
625  // This addend was processed before.
626  continue;
627  }
628 
629  Value *Val = ThisAddend->getSymVal();
630  unsigned StartIdx = SimpVect.size();
631  SimpVect.push_back(ThisAddend);
632 
633  // The inner loop collects addends sharing same symbolic-value, and these
634  // addends will be later on folded into a single addend. Following above
635  // example, if the symbolic value "y" is being processed, the inner loop
636  // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
637  // be later on folded into "<b1+b2, y>".
638  for (unsigned SameSymIdx = SymIdx + 1;
639  SameSymIdx < AddendNum; SameSymIdx++) {
640  const FAddend *T = Addends[SameSymIdx];
641  if (T && T->getSymVal() == Val) {
642  // Set null such that next iteration of the outer loop will not process
643  // this addend again.
644  Addends[SameSymIdx] = nullptr;
645  SimpVect.push_back(T);
646  }
647  }
648 
649  // If multiple addends share same symbolic value, fold them together.
650  if (StartIdx + 1 != SimpVect.size()) {
651  FAddend &R = TmpResult[NextTmpIdx ++];
652  R = *SimpVect[StartIdx];
653  for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
654  R += *SimpVect[Idx];
655 
656  // Pop all addends being folded and push the resulting folded addend.
657  SimpVect.resize(StartIdx);
658  if (Val) {
659  if (!R.isZero()) {
660  SimpVect.push_back(&R);
661  }
662  } else {
663  // Don't push constant addend at this time. It will be the last element
664  // of <SimpVect>.
665  ConstAdd = &R;
666  }
667  }
668  }
669 
670  assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
671  "out-of-bound access");
672 
673  if (ConstAdd)
674  SimpVect.push_back(ConstAdd);
675 
676  Value *Result;
677  if (!SimpVect.empty())
678  Result = createNaryFAdd(SimpVect, InstrQuota);
679  else {
680  // The addition is folded to 0.0.
681  Result = ConstantFP::get(Instr->getType(), 0.0);
682  }
683 
684  return Result;
685 }
686 
687 Value *FAddCombine::createNaryFAdd
688  (const AddendVect &Opnds, unsigned InstrQuota) {
689  assert(!Opnds.empty() && "Expect at least one addend");
690 
691  // Step 1: Check if the # of instructions needed exceeds the quota.
692 
693  unsigned InstrNeeded = calcInstrNumber(Opnds);
694  if (InstrNeeded > InstrQuota)
695  return nullptr;
696 
697  initCreateInstNum();
698 
699  // step 2: Emit the N-ary addition.
700  // Note that at most three instructions are involved in Fadd-InstCombine: the
701  // addition in question, and at most two neighboring instructions.
702  // The resulting optimized addition should have at least one less instruction
703  // than the original addition expression tree. This implies that the resulting
704  // N-ary addition has at most two instructions, and we don't need to worry
705  // about tree-height when constructing the N-ary addition.
706 
707  Value *LastVal = nullptr;
708  bool LastValNeedNeg = false;
709 
710  // Iterate the addends, creating fadd/fsub using adjacent two addends.
711  for (const FAddend *Opnd : Opnds) {
712  bool NeedNeg;
713  Value *V = createAddendVal(*Opnd, NeedNeg);
714  if (!LastVal) {
715  LastVal = V;
716  LastValNeedNeg = NeedNeg;
717  continue;
718  }
719 
720  if (LastValNeedNeg == NeedNeg) {
721  LastVal = createFAdd(LastVal, V);
722  continue;
723  }
724 
725  if (LastValNeedNeg)
726  LastVal = createFSub(V, LastVal);
727  else
728  LastVal = createFSub(LastVal, V);
729 
730  LastValNeedNeg = false;
731  }
732 
733  if (LastValNeedNeg) {
734  LastVal = createFNeg(LastVal);
735  }
736 
737 #ifndef NDEBUG
738  assert(CreateInstrNum == InstrNeeded &&
739  "Inconsistent in instruction numbers");
740 #endif
741 
742  return LastVal;
743 }
744 
745 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
746  Value *V = Builder.CreateFSub(Opnd0, Opnd1);
747  if (Instruction *I = dyn_cast<Instruction>(V))
748  createInstPostProc(I);
749  return V;
750 }
751 
752 Value *FAddCombine::createFNeg(Value *V) {
753  Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType()));
754  Value *NewV = createFSub(Zero, V);
755  if (Instruction *I = dyn_cast<Instruction>(NewV))
756  createInstPostProc(I, true); // fneg's don't receive instruction numbers.
757  return NewV;
758 }
759 
760 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
761  Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
762  if (Instruction *I = dyn_cast<Instruction>(V))
763  createInstPostProc(I);
764  return V;
765 }
766 
767 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
768  Value *V = Builder.CreateFMul(Opnd0, Opnd1);
769  if (Instruction *I = dyn_cast<Instruction>(V))
770  createInstPostProc(I);
771  return V;
772 }
773 
774 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) {
775  Value *V = Builder.CreateFDiv(Opnd0, Opnd1);
776  if (Instruction *I = dyn_cast<Instruction>(V))
777  createInstPostProc(I);
778  return V;
779 }
780 
781 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
782  NewInstr->setDebugLoc(Instr->getDebugLoc());
783 
784  // Keep track of the number of instruction created.
785  if (!NoNumber)
786  incCreateInstNum();
787 
788  // Propagate fast-math flags
789  NewInstr->setFastMathFlags(Instr->getFastMathFlags());
790 }
791 
792 // Return the number of instruction needed to emit the N-ary addition.
793 // NOTE: Keep this function in sync with createAddendVal().
794 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
795  unsigned OpndNum = Opnds.size();
796  unsigned InstrNeeded = OpndNum - 1;
797 
798  // The number of addends in the form of "(-1)*x".
799  unsigned NegOpndNum = 0;
800 
801  // Adjust the number of instructions needed to emit the N-ary add.
802  for (const FAddend *Opnd : Opnds) {
803  if (Opnd->isConstant())
804  continue;
805 
806  // The constant check above is really for a few special constant
807  // coefficients.
808  if (isa<UndefValue>(Opnd->getSymVal()))
809  continue;
810 
811  const FAddendCoef &CE = Opnd->getCoef();
812  if (CE.isMinusOne() || CE.isMinusTwo())
813  NegOpndNum++;
814 
815  // Let the addend be "c * x". If "c == +/-1", the value of the addend
816  // is immediately available; otherwise, it needs exactly one instruction
817  // to evaluate the value.
818  if (!CE.isMinusOne() && !CE.isOne())
819  InstrNeeded++;
820  }
821  if (NegOpndNum == OpndNum)
822  InstrNeeded++;
823  return InstrNeeded;
824 }
825 
826 // Input Addend Value NeedNeg(output)
827 // ================================================================
828 // Constant C C false
829 // <+/-1, V> V coefficient is -1
830 // <2/-2, V> "fadd V, V" coefficient is -2
831 // <C, V> "fmul V, C" false
832 //
833 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
834 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
835  const FAddendCoef &Coeff = Opnd.getCoef();
836 
837  if (Opnd.isConstant()) {
838  NeedNeg = false;
839  return Coeff.getValue(Instr->getType());
840  }
841 
842  Value *OpndVal = Opnd.getSymVal();
843 
844  if (Coeff.isMinusOne() || Coeff.isOne()) {
845  NeedNeg = Coeff.isMinusOne();
846  return OpndVal;
847  }
848 
849  if (Coeff.isTwo() || Coeff.isMinusTwo()) {
850  NeedNeg = Coeff.isMinusTwo();
851  return createFAdd(OpndVal, OpndVal);
852  }
853 
854  NeedNeg = false;
855  return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
856 }
857 
858 /// \brief Return true if we can prove that:
859 /// (sub LHS, RHS) === (sub nsw LHS, RHS)
860 /// This basically requires proving that the add in the original type would not
861 /// overflow to change the sign bit or have a carry out.
862 /// TODO: Handle this for Vectors.
863 bool InstCombiner::willNotOverflowSignedSub(const Value *LHS,
864  const Value *RHS,
865  const Instruction &CxtI) const {
866  // If LHS and RHS each have at least two sign bits, the subtraction
867  // cannot overflow.
868  if (ComputeNumSignBits(LHS, 0, &CxtI) > 1 &&
869  ComputeNumSignBits(RHS, 0, &CxtI) > 1)
870  return true;
871 
872  KnownBits LHSKnown = computeKnownBits(LHS, 0, &CxtI);
873 
874  KnownBits RHSKnown = computeKnownBits(RHS, 0, &CxtI);
875 
876  // Subtraction of two 2's complement numbers having identical signs will
877  // never overflow.
878  if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
879  (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
880  return true;
881 
882  // TODO: implement logic similar to checkRippleForAdd
883  return false;
884 }
885 
886 /// \brief Return true if we can prove that:
887 /// (sub LHS, RHS) === (sub nuw LHS, RHS)
888 bool InstCombiner::willNotOverflowUnsignedSub(const Value *LHS,
889  const Value *RHS,
890  const Instruction &CxtI) const {
891  // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
892  KnownBits LHSKnown = computeKnownBits(LHS, /*Depth=*/0, &CxtI);
893  KnownBits RHSKnown = computeKnownBits(RHS, /*Depth=*/0, &CxtI);
894  if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
895  return true;
896 
897  return false;
898 }
899 
900 // Checks if any operand is negative and we can convert add to sub.
901 // This function checks for following negative patterns
902 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
903 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
904 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
906  InstCombiner::BuilderTy &Builder) {
907  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
908 
909  // This function creates 2 instructions to replace ADD, we need at least one
910  // of LHS or RHS to have one use to ensure benefit in transform.
911  if (!LHS->hasOneUse() && !RHS->hasOneUse())
912  return nullptr;
913 
914  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
915  const APInt *C1 = nullptr, *C2 = nullptr;
916 
917  // if ONE is on other side, swap
918  if (match(RHS, m_Add(m_Value(X), m_One())))
919  std::swap(LHS, RHS);
920 
921  if (match(LHS, m_Add(m_Value(X), m_One()))) {
922  // if XOR on other side, swap
923  if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
924  std::swap(X, RHS);
925 
926  if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
927  // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
928  // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
929  if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
930  Value *NewAnd = Builder.CreateAnd(Z, *C1);
931  return Builder.CreateSub(RHS, NewAnd, "sub");
932  } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
933  // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
934  // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
935  Value *NewOr = Builder.CreateOr(Z, ~(*C1));
936  return Builder.CreateSub(RHS, NewOr, "sub");
937  }
938  }
939  }
940 
941  // Restore LHS and RHS
942  LHS = I.getOperand(0);
943  RHS = I.getOperand(1);
944 
945  // if XOR is on other side, swap
946  if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
947  std::swap(LHS, RHS);
948 
949  // C2 is ODD
950  // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
951  // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
952  if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
953  if (C1->countTrailingZeros() == 0)
954  if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
955  Value *NewOr = Builder.CreateOr(Z, ~(*C2));
956  return Builder.CreateSub(RHS, NewOr, "sub");
957  }
958  return nullptr;
959 }
960 
961 Instruction *InstCombiner::foldAddWithConstant(BinaryOperator &Add) {
962  Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
963  Constant *Op1C;
964  if (!match(Op1, m_Constant(Op1C)))
965  return nullptr;
966 
967  if (Instruction *NV = foldOpWithConstantIntoOperand(Add))
968  return NV;
969 
970  Value *X;
971  // zext(bool) + C -> bool ? C + 1 : C
972  if (match(Op0, m_ZExt(m_Value(X))) &&
973  X->getType()->getScalarSizeInBits() == 1)
974  return SelectInst::Create(X, AddOne(Op1C), Op1);
975 
976  // ~X + C --> (C-1) - X
977  if (match(Op0, m_Not(m_Value(X))))
978  return BinaryOperator::CreateSub(SubOne(Op1C), X);
979 
980  const APInt *C;
981  if (!match(Op1, m_APInt(C)))
982  return nullptr;
983 
984  if (C->isSignMask()) {
985  // If wrapping is not allowed, then the addition must set the sign bit:
986  // X + (signmask) --> X | signmask
987  if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
988  return BinaryOperator::CreateOr(Op0, Op1);
989 
990  // If wrapping is allowed, then the addition flips the sign bit of LHS:
991  // X + (signmask) --> X ^ signmask
992  return BinaryOperator::CreateXor(Op0, Op1);
993  }
994 
995  // Is this add the last step in a convoluted sext?
996  // add(zext(xor i16 X, -32768), -32768) --> sext X
997  Type *Ty = Add.getType();
998  const APInt *C2;
999  if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
1000  C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
1001  return CastInst::Create(Instruction::SExt, X, Ty);
1002 
1003  // (add (zext (add nuw X, C2)), C) --> (zext (add nuw X, C2 + C))
1004  if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
1005  C->isNegative() && C->sge(-C2->sext(C->getBitWidth()))) {
1006  Constant *NewC =
1007  ConstantInt::get(X->getType(), *C2 + C->trunc(C2->getBitWidth()));
1008  return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
1009  }
1010 
1011  if (C->isOneValue() && Op0->hasOneUse()) {
1012  // add (sext i1 X), 1 --> zext (not X)
1013  // TODO: The smallest IR representation is (select X, 0, 1), and that would
1014  // not require the one-use check. But we need to remove a transform in
1015  // visitSelect and make sure that IR value tracking for select is equal or
1016  // better than for these ops.
1017  if (match(Op0, m_SExt(m_Value(X))) &&
1018  X->getType()->getScalarSizeInBits() == 1)
1019  return new ZExtInst(Builder.CreateNot(X), Ty);
1020 
1021  // Shifts and add used to flip and mask off the low bit:
1022  // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
1023  const APInt *C3;
1024  if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
1025  C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
1026  Value *NotX = Builder.CreateNot(X);
1027  return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
1028  }
1029  }
1030 
1031  return nullptr;
1032 }
1033 
1035  bool Changed = SimplifyAssociativeOrCommutative(I);
1036  if (Value *V = SimplifyVectorOp(I))
1037  return replaceInstUsesWith(I, V);
1038 
1039  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1040  if (Value *V =
1042  SQ.getWithInstruction(&I)))
1043  return replaceInstUsesWith(I, V);
1044 
1045  // (A*B)+(A*C) -> A*(B+C) etc
1046  if (Value *V = SimplifyUsingDistributiveLaws(I))
1047  return replaceInstUsesWith(I, V);
1048 
1049  if (Instruction *X = foldAddWithConstant(I))
1050  return X;
1051 
1052  // FIXME: This should be moved into the above helper function to allow these
1053  // transforms for general constant or constant splat vectors.
1054  Type *Ty = I.getType();
1055  if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1056  Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1057  if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1058  unsigned TySizeBits = Ty->getScalarSizeInBits();
1059  const APInt &RHSVal = CI->getValue();
1060  unsigned ExtendAmt = 0;
1061  // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1062  // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1063  if (XorRHS->getValue() == -RHSVal) {
1064  if (RHSVal.isPowerOf2())
1065  ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1066  else if (XorRHS->getValue().isPowerOf2())
1067  ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1068  }
1069 
1070  if (ExtendAmt) {
1071  APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1072  if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1073  ExtendAmt = 0;
1074  }
1075 
1076  if (ExtendAmt) {
1077  Constant *ShAmt = ConstantInt::get(Ty, ExtendAmt);
1078  Value *NewShl = Builder.CreateShl(XorLHS, ShAmt, "sext");
1079  return BinaryOperator::CreateAShr(NewShl, ShAmt);
1080  }
1081 
1082  // If this is a xor that was canonicalized from a sub, turn it back into
1083  // a sub and fuse this add with it.
1084  if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1085  KnownBits LHSKnown = computeKnownBits(XorLHS, 0, &I);
1086  if ((XorRHS->getValue() | LHSKnown.Zero).isAllOnesValue())
1087  return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1088  XorLHS);
1089  }
1090  // (X + signmask) + C could have gotten canonicalized to (X^signmask) + C,
1091  // transform them into (X + (signmask ^ C))
1092  if (XorRHS->getValue().isSignMask())
1093  return BinaryOperator::CreateAdd(XorLHS,
1094  ConstantExpr::getXor(XorRHS, CI));
1095  }
1096  }
1097 
1098  if (Ty->isIntOrIntVectorTy(1))
1099  return BinaryOperator::CreateXor(LHS, RHS);
1100 
1101  // X + X --> X << 1
1102  if (LHS == RHS) {
1103  auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1104  Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1105  Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1106  return Shl;
1107  }
1108 
1109  Value *A, *B;
1110  if (match(LHS, m_Neg(m_Value(A)))) {
1111  // -A + -B --> -(A + B)
1112  if (match(RHS, m_Neg(m_Value(B))))
1113  return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1114 
1115  // -A + B --> B - A
1116  return BinaryOperator::CreateSub(RHS, A);
1117  }
1118 
1119  // A + -B --> A - B
1120  if (match(RHS, m_Neg(m_Value(B))))
1121  return BinaryOperator::CreateSub(LHS, B);
1122 
1123  if (Value *V = checkForNegativeOperand(I, Builder))
1124  return replaceInstUsesWith(I, V);
1125 
1126  // A+B --> A|B iff A and B have no bits set in common.
1127  if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1128  return BinaryOperator::CreateOr(LHS, RHS);
1129 
1130  // FIXME: We already did a check for ConstantInt RHS above this.
1131  // FIXME: Is this pattern covered by another fold? No regression tests fail on
1132  // removal.
1133  if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1134  // (X & FF00) + xx00 -> (X+xx00) & FF00
1135  Value *X;
1136  ConstantInt *C2;
1137  if (LHS->hasOneUse() &&
1138  match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1139  CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1140  // See if all bits from the first bit set in the Add RHS up are included
1141  // in the mask. First, get the rightmost bit.
1142  const APInt &AddRHSV = CRHS->getValue();
1143 
1144  // Form a mask of all bits from the lowest bit added through the top.
1145  APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1146 
1147  // See if the and mask includes all of these bits.
1148  APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1149 
1150  if (AddRHSHighBits == AddRHSHighBitsAnd) {
1151  // Okay, the xform is safe. Insert the new add pronto.
1152  Value *NewAdd = Builder.CreateAdd(X, CRHS, LHS->getName());
1153  return BinaryOperator::CreateAnd(NewAdd, C2);
1154  }
1155  }
1156  }
1157 
1158  // add (select X 0 (sub n A)) A --> select X A n
1159  {
1160  SelectInst *SI = dyn_cast<SelectInst>(LHS);
1161  Value *A = RHS;
1162  if (!SI) {
1163  SI = dyn_cast<SelectInst>(RHS);
1164  A = LHS;
1165  }
1166  if (SI && SI->hasOneUse()) {
1167  Value *TV = SI->getTrueValue();
1168  Value *FV = SI->getFalseValue();
1169  Value *N;
1170 
1171  // Can we fold the add into the argument of the select?
1172  // We check both true and false select arguments for a matching subtract.
1173  if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1174  // Fold the add into the true select value.
1175  return SelectInst::Create(SI->getCondition(), N, A);
1176 
1177  if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1178  // Fold the add into the false select value.
1179  return SelectInst::Create(SI->getCondition(), A, N);
1180  }
1181  }
1182 
1183  // Check for (add (sext x), y), see if we can merge this into an
1184  // integer add followed by a sext.
1185  if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
1186  // (add (sext x), cst) --> (sext (add x, cst'))
1187  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1188  if (LHSConv->hasOneUse()) {
1189  Constant *CI =
1190  ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1191  if (ConstantExpr::getSExt(CI, Ty) == RHSC &&
1192  willNotOverflowSignedAdd(LHSConv->getOperand(0), CI, I)) {
1193  // Insert the new, smaller add.
1194  Value *NewAdd =
1195  Builder.CreateNSWAdd(LHSConv->getOperand(0), CI, "addconv");
1196  return new SExtInst(NewAdd, Ty);
1197  }
1198  }
1199  }
1200 
1201  // (add (sext x), (sext y)) --> (sext (add int x, y))
1202  if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
1203  // Only do this if x/y have the same type, if at least one of them has a
1204  // single use (so we don't increase the number of sexts), and if the
1205  // integer add will not overflow.
1206  if (LHSConv->getOperand(0)->getType() ==
1207  RHSConv->getOperand(0)->getType() &&
1208  (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1209  willNotOverflowSignedAdd(LHSConv->getOperand(0),
1210  RHSConv->getOperand(0), I)) {
1211  // Insert the new integer add.
1212  Value *NewAdd = Builder.CreateNSWAdd(LHSConv->getOperand(0),
1213  RHSConv->getOperand(0), "addconv");
1214  return new SExtInst(NewAdd, Ty);
1215  }
1216  }
1217  }
1218 
1219  // Check for (add (zext x), y), see if we can merge this into an
1220  // integer add followed by a zext.
1221  if (auto *LHSConv = dyn_cast<ZExtInst>(LHS)) {
1222  // (add (zext x), cst) --> (zext (add x, cst'))
1223  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
1224  if (LHSConv->hasOneUse()) {
1225  Constant *CI =
1226  ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
1227  if (ConstantExpr::getZExt(CI, Ty) == RHSC &&
1228  willNotOverflowUnsignedAdd(LHSConv->getOperand(0), CI, I)) {
1229  // Insert the new, smaller add.
1230  Value *NewAdd =
1231  Builder.CreateNUWAdd(LHSConv->getOperand(0), CI, "addconv");
1232  return new ZExtInst(NewAdd, Ty);
1233  }
1234  }
1235  }
1236 
1237  // (add (zext x), (zext y)) --> (zext (add int x, y))
1238  if (auto *RHSConv = dyn_cast<ZExtInst>(RHS)) {
1239  // Only do this if x/y have the same type, if at least one of them has a
1240  // single use (so we don't increase the number of zexts), and if the
1241  // integer add will not overflow.
1242  if (LHSConv->getOperand(0)->getType() ==
1243  RHSConv->getOperand(0)->getType() &&
1244  (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1245  willNotOverflowUnsignedAdd(LHSConv->getOperand(0),
1246  RHSConv->getOperand(0), I)) {
1247  // Insert the new integer add.
1248  Value *NewAdd = Builder.CreateNUWAdd(
1249  LHSConv->getOperand(0), RHSConv->getOperand(0), "addconv");
1250  return new ZExtInst(NewAdd, Ty);
1251  }
1252  }
1253  }
1254 
1255  // (add (xor A, B) (and A, B)) --> (or A, B)
1256  if (match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
1257  match(RHS, m_c_And(m_Specific(A), m_Specific(B))))
1258  return BinaryOperator::CreateOr(A, B);
1259 
1260  // (add (and A, B) (xor A, B)) --> (or A, B)
1261  if (match(RHS, m_Xor(m_Value(A), m_Value(B))) &&
1262  match(LHS, m_c_And(m_Specific(A), m_Specific(B))))
1263  return BinaryOperator::CreateOr(A, B);
1264 
1265  // (add (or A, B) (and A, B)) --> (add A, B)
1266  if (match(LHS, m_Or(m_Value(A), m_Value(B))) &&
1267  match(RHS, m_c_And(m_Specific(A), m_Specific(B)))) {
1268  I.setOperand(0, A);
1269  I.setOperand(1, B);
1270  return &I;
1271  }
1272 
1273  // (add (and A, B) (or A, B)) --> (add A, B)
1274  if (match(RHS, m_Or(m_Value(A), m_Value(B))) &&
1275  match(LHS, m_c_And(m_Specific(A), m_Specific(B)))) {
1276  I.setOperand(0, A);
1277  I.setOperand(1, B);
1278  return &I;
1279  }
1280 
1281  // TODO(jingyue): Consider willNotOverflowSignedAdd and
1282  // willNotOverflowUnsignedAdd to reduce the number of invocations of
1283  // computeKnownBits.
1284  if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1285  Changed = true;
1286  I.setHasNoSignedWrap(true);
1287  }
1288  if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1289  Changed = true;
1290  I.setHasNoUnsignedWrap(true);
1291  }
1292 
1293  return Changed ? &I : nullptr;
1294 }
1295 
1297  bool Changed = SimplifyAssociativeOrCommutative(I);
1298  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1299 
1300  if (Value *V = SimplifyVectorOp(I))
1301  return replaceInstUsesWith(I, V);
1302 
1303  if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(),
1304  SQ.getWithInstruction(&I)))
1305  return replaceInstUsesWith(I, V);
1306 
1307  if (isa<Constant>(RHS))
1308  if (Instruction *FoldedFAdd = foldOpWithConstantIntoOperand(I))
1309  return FoldedFAdd;
1310 
1311  // -A + B --> B - A
1312  // -A + -B --> -(A + B)
1313  if (Value *LHSV = dyn_castFNegVal(LHS)) {
1314  Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV);
1315  RI->copyFastMathFlags(&I);
1316  return RI;
1317  }
1318 
1319  // A + -B --> A - B
1320  if (!isa<Constant>(RHS))
1321  if (Value *V = dyn_castFNegVal(RHS)) {
1322  Instruction *RI = BinaryOperator::CreateFSub(LHS, V);
1323  RI->copyFastMathFlags(&I);
1324  return RI;
1325  }
1326 
1327  // Check for (fadd double (sitofp x), y), see if we can merge this into an
1328  // integer add followed by a promotion.
1329  if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1330  Value *LHSIntVal = LHSConv->getOperand(0);
1331  Type *FPType = LHSConv->getType();
1332 
1333  // TODO: This check is overly conservative. In many cases known bits
1334  // analysis can tell us that the result of the addition has less significant
1335  // bits than the integer type can hold.
1336  auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1337  Type *FScalarTy = FTy->getScalarType();
1338  Type *IScalarTy = ITy->getScalarType();
1339 
1340  // Do we have enough bits in the significand to represent the result of
1341  // the integer addition?
1342  unsigned MaxRepresentableBits =
1344  return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1345  };
1346 
1347  // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1348  // ... if the constant fits in the integer value. This is useful for things
1349  // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1350  // requires a constant pool load, and generally allows the add to be better
1351  // instcombined.
1352  if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1353  if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1354  Constant *CI =
1355  ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1356  if (LHSConv->hasOneUse() &&
1357  ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1358  willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1359  // Insert the new integer add.
1360  Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1361  return new SIToFPInst(NewAdd, I.getType());
1362  }
1363  }
1364 
1365  // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1366  if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1367  Value *RHSIntVal = RHSConv->getOperand(0);
1368  // It's enough to check LHS types only because we require int types to
1369  // be the same for this transform.
1370  if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1371  // Only do this if x/y have the same type, if at least one of them has a
1372  // single use (so we don't increase the number of int->fp conversions),
1373  // and if the integer add will not overflow.
1374  if (LHSIntVal->getType() == RHSIntVal->getType() &&
1375  (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1376  willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1377  // Insert the new integer add.
1378  Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1379  return new SIToFPInst(NewAdd, I.getType());
1380  }
1381  }
1382  }
1383  }
1384 
1385  // Handle specials cases for FAdd with selects feeding the operation
1386  if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1387  return replaceInstUsesWith(I, V);
1388 
1389  if (I.isFast()) {
1390  if (Value *V = FAddCombine(Builder).simplify(&I))
1391  return replaceInstUsesWith(I, V);
1392  }
1393 
1394  return Changed ? &I : nullptr;
1395 }
1396 
1397 /// Optimize pointer differences into the same array into a size. Consider:
1398 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1399 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1401  Type *Ty) {
1402  // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1403  // this.
1404  bool Swapped = false;
1405  GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1406 
1407  // For now we require one side to be the base pointer "A" or a constant
1408  // GEP derived from it.
1409  if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1410  // (gep X, ...) - X
1411  if (LHSGEP->getOperand(0) == RHS) {
1412  GEP1 = LHSGEP;
1413  Swapped = false;
1414  } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1415  // (gep X, ...) - (gep X, ...)
1416  if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1417  RHSGEP->getOperand(0)->stripPointerCasts()) {
1418  GEP2 = RHSGEP;
1419  GEP1 = LHSGEP;
1420  Swapped = false;
1421  }
1422  }
1423  }
1424 
1425  if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1426  // X - (gep X, ...)
1427  if (RHSGEP->getOperand(0) == LHS) {
1428  GEP1 = RHSGEP;
1429  Swapped = true;
1430  } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1431  // (gep X, ...) - (gep X, ...)
1432  if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1433  LHSGEP->getOperand(0)->stripPointerCasts()) {
1434  GEP2 = LHSGEP;
1435  GEP1 = RHSGEP;
1436  Swapped = true;
1437  }
1438  }
1439  }
1440 
1441  if (!GEP1)
1442  // No GEP found.
1443  return nullptr;
1444 
1445  if (GEP2) {
1446  // (gep X, ...) - (gep X, ...)
1447  //
1448  // Avoid duplicating the arithmetic if there are more than one non-constant
1449  // indices between the two GEPs and either GEP has a non-constant index and
1450  // multiple users. If zero non-constant index, the result is a constant and
1451  // there is no duplication. If one non-constant index, the result is an add
1452  // or sub with a constant, which is no larger than the original code, and
1453  // there's no duplicated arithmetic, even if either GEP has multiple
1454  // users. If more than one non-constant indices combined, as long as the GEP
1455  // with at least one non-constant index doesn't have multiple users, there
1456  // is no duplication.
1457  unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1458  unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1459  if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1460  ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1461  (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1462  return nullptr;
1463  }
1464  }
1465 
1466  // Emit the offset of the GEP and an intptr_t.
1467  Value *Result = EmitGEPOffset(GEP1);
1468 
1469  // If we had a constant expression GEP on the other side offsetting the
1470  // pointer, subtract it from the offset we have.
1471  if (GEP2) {
1472  Value *Offset = EmitGEPOffset(GEP2);
1473  Result = Builder.CreateSub(Result, Offset);
1474  }
1475 
1476  // If we have p - gep(p, ...) then we have to negate the result.
1477  if (Swapped)
1478  Result = Builder.CreateNeg(Result, "diff.neg");
1479 
1480  return Builder.CreateIntCast(Result, Ty, true);
1481 }
1482 
1484  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1485 
1486  if (Value *V = SimplifyVectorOp(I))
1487  return replaceInstUsesWith(I, V);
1488 
1489  if (Value *V =
1491  SQ.getWithInstruction(&I)))
1492  return replaceInstUsesWith(I, V);
1493 
1494  // (A*B)-(A*C) -> A*(B-C) etc
1495  if (Value *V = SimplifyUsingDistributiveLaws(I))
1496  return replaceInstUsesWith(I, V);
1497 
1498  // If this is a 'B = x-(-A)', change to B = x+A.
1499  if (Value *V = dyn_castNegVal(Op1)) {
1501 
1502  if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1503  assert(BO->getOpcode() == Instruction::Sub &&
1504  "Expected a subtraction operator!");
1505  if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1506  Res->setHasNoSignedWrap(true);
1507  } else {
1508  if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1509  Res->setHasNoSignedWrap(true);
1510  }
1511 
1512  return Res;
1513  }
1514 
1515  if (I.getType()->isIntOrIntVectorTy(1))
1516  return BinaryOperator::CreateXor(Op0, Op1);
1517 
1518  // Replace (-1 - A) with (~A).
1519  if (match(Op0, m_AllOnes()))
1520  return BinaryOperator::CreateNot(Op1);
1521 
1522  if (Constant *C = dyn_cast<Constant>(Op0)) {
1523  // C - ~X == X + (1+C)
1524  Value *X = nullptr;
1525  if (match(Op1, m_Not(m_Value(X))))
1526  return BinaryOperator::CreateAdd(X, AddOne(C));
1527 
1528  // Try to fold constant sub into select arguments.
1529  if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1530  if (Instruction *R = FoldOpIntoSelect(I, SI))
1531  return R;
1532 
1533  // Try to fold constant sub into PHI values.
1534  if (PHINode *PN = dyn_cast<PHINode>(Op1))
1535  if (Instruction *R = foldOpIntoPhi(I, PN))
1536  return R;
1537 
1538  // C-(X+C2) --> (C-C2)-X
1539  Constant *C2;
1540  if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1541  return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1542 
1543  // Fold (sub 0, (zext bool to B)) --> (sext bool to B)
1544  if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X))))
1545  if (X->getType()->isIntOrIntVectorTy(1))
1546  return CastInst::CreateSExtOrBitCast(X, Op1->getType());
1547 
1548  // Fold (sub 0, (sext bool to B)) --> (zext bool to B)
1549  if (C->isNullValue() && match(Op1, m_SExt(m_Value(X))))
1550  if (X->getType()->isIntOrIntVectorTy(1))
1551  return CastInst::CreateZExtOrBitCast(X, Op1->getType());
1552  }
1553 
1554  const APInt *Op0C;
1555  if (match(Op0, m_APInt(Op0C))) {
1556  unsigned BitWidth = I.getType()->getScalarSizeInBits();
1557 
1558  // -(X >>u 31) -> (X >>s 31)
1559  // -(X >>s 31) -> (X >>u 31)
1560  if (Op0C->isNullValue()) {
1561  Value *X;
1562  const APInt *ShAmt;
1563  if (match(Op1, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1564  *ShAmt == BitWidth - 1) {
1565  Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1566  return BinaryOperator::CreateAShr(X, ShAmtOp);
1567  }
1568  if (match(Op1, m_AShr(m_Value(X), m_APInt(ShAmt))) &&
1569  *ShAmt == BitWidth - 1) {
1570  Value *ShAmtOp = cast<Instruction>(Op1)->getOperand(1);
1571  return BinaryOperator::CreateLShr(X, ShAmtOp);
1572  }
1573  }
1574 
1575  // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1576  // zero.
1577  if (Op0C->isMask()) {
1578  KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1579  if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
1580  return BinaryOperator::CreateXor(Op1, Op0);
1581  }
1582  }
1583 
1584  {
1585  Value *Y;
1586  // X-(X+Y) == -Y X-(Y+X) == -Y
1587  if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1588  return BinaryOperator::CreateNeg(Y);
1589 
1590  // (X-Y)-X == -Y
1591  if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1592  return BinaryOperator::CreateNeg(Y);
1593  }
1594 
1595  // (sub (or A, B), (xor A, B)) --> (and A, B)
1596  {
1597  Value *A, *B;
1598  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1599  match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1600  return BinaryOperator::CreateAnd(A, B);
1601  }
1602 
1603  {
1604  Value *Y;
1605  // ((X | Y) - X) --> (~X & Y)
1606  if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1607  return BinaryOperator::CreateAnd(
1608  Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1609  }
1610 
1611  if (Op1->hasOneUse()) {
1612  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1613  Constant *C = nullptr;
1614 
1615  // (X - (Y - Z)) --> (X + (Z - Y)).
1616  if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1617  return BinaryOperator::CreateAdd(Op0,
1618  Builder.CreateSub(Z, Y, Op1->getName()));
1619 
1620  // (X - (X & Y)) --> (X & ~Y)
1621  if (match(Op1, m_c_And(m_Value(Y), m_Specific(Op0))))
1622  return BinaryOperator::CreateAnd(Op0,
1623  Builder.CreateNot(Y, Y->getName() + ".not"));
1624 
1625  // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow.
1626  if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) && match(Op0, m_Zero()) &&
1627  C->isNotMinSignedValue() && !C->isOneValue())
1628  return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
1629 
1630  // 0 - (X << Y) -> (-X << Y) when X is freely negatable.
1631  if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1632  if (Value *XNeg = dyn_castNegVal(X))
1633  return BinaryOperator::CreateShl(XNeg, Y);
1634 
1635  // Subtracting -1/0 is the same as adding 1/0:
1636  // sub [nsw] Op0, sext(bool Y) -> add [nsw] Op0, zext(bool Y)
1637  // 'nuw' is dropped in favor of the canonical form.
1638  if (match(Op1, m_SExt(m_Value(Y))) &&
1639  Y->getType()->getScalarSizeInBits() == 1) {
1640  Value *Zext = Builder.CreateZExt(Y, I.getType());
1641  BinaryOperator *Add = BinaryOperator::CreateAdd(Op0, Zext);
1643  return Add;
1644  }
1645 
1646  // X - A*-B -> X + A*B
1647  // X - -A*B -> X + A*B
1648  Value *A, *B;
1649  Constant *CI;
1650  if (match(Op1, m_c_Mul(m_Value(A), m_Neg(m_Value(B)))))
1651  return BinaryOperator::CreateAdd(Op0, Builder.CreateMul(A, B));
1652 
1653  // X - A*CI -> X + A*-CI
1654  // No need to handle commuted multiply because multiply handling will
1655  // ensure constant will be move to the right hand side.
1656  if (match(Op1, m_Mul(m_Value(A), m_Constant(CI)))) {
1657  Value *NewMul = Builder.CreateMul(A, ConstantExpr::getNeg(CI));
1658  return BinaryOperator::CreateAdd(Op0, NewMul);
1659  }
1660  }
1661 
1662  // Optimize pointer differences into the same array into a size. Consider:
1663  // &A[10] - &A[0]: we should compile this to "10".
1664  Value *LHSOp, *RHSOp;
1665  if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1666  match(Op1, m_PtrToInt(m_Value(RHSOp))))
1667  if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1668  return replaceInstUsesWith(I, Res);
1669 
1670  // trunc(p)-trunc(q) -> trunc(p-q)
1671  if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1672  match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1673  if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1674  return replaceInstUsesWith(I, Res);
1675 
1676  bool Changed = false;
1677  if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1678  Changed = true;
1679  I.setHasNoSignedWrap(true);
1680  }
1681  if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1682  Changed = true;
1683  I.setHasNoUnsignedWrap(true);
1684  }
1685 
1686  return Changed ? &I : nullptr;
1687 }
1688 
1690  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1691 
1692  if (Value *V = SimplifyVectorOp(I))
1693  return replaceInstUsesWith(I, V);
1694 
1695  if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(),
1696  SQ.getWithInstruction(&I)))
1697  return replaceInstUsesWith(I, V);
1698 
1699  // fsub nsz 0, X ==> fsub nsz -0.0, X
1700  if (I.getFastMathFlags().noSignedZeros() && match(Op0, m_Zero())) {
1701  // Subtraction from -0.0 is the canonical form of fneg.
1703  NewI->copyFastMathFlags(&I);
1704  return NewI;
1705  }
1706 
1707  if (isa<Constant>(Op0))
1708  if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1709  if (Instruction *NV = FoldOpIntoSelect(I, SI))
1710  return NV;
1711 
1712  // If this is a 'B = x-(-A)', change to B = x+A, potentially looking
1713  // through FP extensions/truncations along the way.
1714  if (Value *V = dyn_castFNegVal(Op1)) {
1715  Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V);
1716  NewI->copyFastMathFlags(&I);
1717  return NewI;
1718  }
1719  if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) {
1720  if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) {
1721  Value *NewTrunc = Builder.CreateFPTrunc(V, I.getType());
1722  Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc);
1723  NewI->copyFastMathFlags(&I);
1724  return NewI;
1725  }
1726  } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) {
1727  if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
1728  Value *NewExt = Builder.CreateFPExt(V, I.getType());
1729  Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt);
1730  NewI->copyFastMathFlags(&I);
1731  return NewI;
1732  }
1733  }
1734 
1735  // Handle specials cases for FSub with selects feeding the operation
1736  if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
1737  return replaceInstUsesWith(I, V);
1738 
1739  if (I.isFast()) {
1740  if (Value *V = FAddCombine(Builder).simplify(&I))
1741  return replaceInstUsesWith(I, V);
1742  }
1743 
1744  return nullptr;
1745 }
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:247
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:574
uint64_t CallInst * C
std::string & operator+=(std::string &buffer, StringRef string)
Definition: StringRef.h:899
void copyFastMathFlags(FastMathFlags FMF)
Convenience function for transferring all fast-math flag values to this instruction, which must be an operator which supports these flags.
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:466
static bool isConstant(const MachineInstr &MI)
APInt sext(unsigned width) const
Sign extend to a new width.
Definition: APInt.cpp:841
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:514
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
match_zero m_Zero()
Match an arbitrary zero/null constant.
Definition: PatternMatch.h:145
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:526
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:604
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:818
F(f)
const fltSemantics & getSemantics() const
Definition: APFloat.h:1155
This class represents a sign extension of integer types.
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2126
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
void changeSign()
Definition: APFloat.h:1050
bool hasNoSignedWrap() const
Determine whether the no signed wrap flag is set.
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1488
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:130
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2115
unsigned countTrailingZeros() const
Count the number of trailing zero bits.
Definition: APInt.h:1611
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:586
This class represents the LLVM &#39;select&#39; instruction.
roundingMode
IEEE-754R 4.3: Rounding-direction attributes.
Definition: APFloat.h:174
CastClass_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
Definition: PatternMatch.h:912
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:668
static Constant * AddOne(Constant *C)
Add one to a Constant.
static Constant * getSExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1556
This file implements a class to represent arbitrary precision integral constant values and operations...
not_match< LHS > m_Not(const LHS &L)
Definition: PatternMatch.h:985
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:502
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1570
OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWAdd(const LHS &L, const RHS &R)
Definition: PatternMatch.h:670
Value * SimplifyFAddInst(Value *LHS, Value *RHS, FastMathFlags FMF, const SimplifyQuery &Q)
Given operands for an FAdd, fold the result or return null.
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
CastClass_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
Definition: PatternMatch.h:924
#define T
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:125
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:915
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
bool haveNoCommonBitsSet(const Value *LHS, const Value *RHS, const DataLayout &DL, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr)
Return true if LHS and RHS have no common bits set.
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
Value * getOperand(unsigned i) const
Definition: User.h:154
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1083
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:301
bool isNegative() const
Returns true if this value is known to be negative.
Definition: KnownBits.h:96
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Get a value with high bits set.
Definition: APInt.h:629
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:357
#define P(N)
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
Definition: PatternMatch.h:598
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:260
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:281
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
Definition: PatternMatch.h:544
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:580
constexpr bool isInt(int64_t x)
Checks if an integer fits into the given bit width.
Definition: MathExtras.h:291
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...
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)
Return true if &#39;V & Mask&#39; is known to be zero.
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:488
bool isOneValue() const
Determine if this is a value of 1.
Definition: APInt.h:404
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:382
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:436
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:592
bool isFast() const
Determine whether all fast-math-flags are set.
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:959
Instruction * visitFAdd(BinaryOperator &I)
const Value * getCondition() const
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.
neg_match< LHS > m_Neg(const LHS &L)
Match an integer negate.
const APFloat & getValueAPF() const
Definition: Constants.h:294
CastClass_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
Definition: PatternMatch.h:918
static Constant * getSIToFP(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1619
static BinaryOperator * CreateFNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
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
unsigned ComputeNumSignBits(const Value *Op, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr)
Return the number of times the sign bit of the register is replicated into the other bits...
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:864
constexpr size_t array_lengthof(T(&)[N])
Find the length of an array.
Definition: STLExtras.h:674
static Constant * getTrunc(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1542
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:560
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:623
static unsigned int semanticsPrecision(const fltSemantics &)
Definition: APFloat.cpp:154
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
unsigned logBase2() const
Definition: APInt.h:1727
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
Class for arbitrary precision integers.
Definition: APInt.h:69
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:457
bool sge(const APInt &RHS) const
Signed greather or equal comparison.
Definition: APInt.h:1288
CastClass_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
Definition: PatternMatch.h:906
This union template exposes a suitably aligned and sized character array member which can hold elemen...
Definition: AlignOf.h:138
const Value * getFalseValue() const
static Constant * getNeg(Constant *C, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2096
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:538
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
Instruction * visitAdd(BinaryOperator &I)
match_one m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:194
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:297
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:220
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
static Constant * getZeroValueForNegation(Type *Ty)
Floating point negation must be implemented with f(x) = -0.0 - x.
Definition: Constants.cpp:676
bool isNormal() const
Definition: APFloat.h:1151
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
bool hasNoUnsignedWrap() const
Determine whether the no unsigned wrap flag is set.
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1067
bool isOneValue() const
Returns true if the value is one.
Definition: Constants.cpp:127
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nsw 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())
match_all_ones m_AllOnes()
Match an integer or vector with all bits set to true.
Definition: PatternMatch.h:205
This class represents a truncation of floating point types.
LLVM Value Representation.
Definition: Value.h:73
This file provides internal interfaces used to implement the InstCombine.
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:1641
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:179
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:414
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
bool isNonNegative() const
Returns true if this value is known to be non-negative.
Definition: KnownBits.h:99
loop simplify
This class represents an extension of floating point types.
unsigned countNonConstantIndices() const
Definition: Operator.h:504
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)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
Instruction * visitSub(BinaryOperator &I)
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
bool isNullValue() const
Determine if all bits are clear.
Definition: APInt.h:399
bool noSignedZeros() const
Definition: Operator.h:202
const fltSemantics & getFltSemantics() const
Definition: Type.h:169
static Constant * getXor(Constant *C1, Constant *C2)
Definition: Constants.cpp:2182