LLVM 19.0.0git
InstCombineAddSub.cpp
Go to the documentation of this file.
1//===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visit functions for add, fadd, sub, and fsub.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APFloat.h"
15#include "llvm/ADT/APInt.h"
16#include "llvm/ADT/STLExtras.h"
20#include "llvm/IR/Constant.h"
21#include "llvm/IR/Constants.h"
22#include "llvm/IR/InstrTypes.h"
23#include "llvm/IR/Instruction.h"
25#include "llvm/IR/Operator.h"
27#include "llvm/IR/Type.h"
28#include "llvm/IR/Value.h"
33#include <cassert>
34#include <utility>
35
36using namespace llvm;
37using namespace PatternMatch;
38
39#define DEBUG_TYPE "instcombine"
40
41namespace {
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() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86
87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
89 }
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
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
213 Instruction *Instr = nullptr;
214 };
215
216} // end anonymous namespace
217
218//===----------------------------------------------------------------------===//
219//
220// Implementation of
221// {FAddendCoef, FAddend, FAddition, FAddCombine}.
222//
223//===----------------------------------------------------------------------===//
224FAddendCoef::~FAddendCoef() {
225 if (BufHasFpVal)
226 getFpValPtr()->~APFloat();
227}
228
229void 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
242void 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
256APFloat 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
266void FAddendCoef::operator=(const FAddendCoef &That) {
267 if (That.isInt())
268 set(That.IntVal);
269 else
270 set(That.getFpVal());
271}
272
273void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
276 if (isInt())
277 IntVal += That.IntVal;
278 else
279 getFpVal().add(That.getFpVal(), RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292}
293
294void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value");
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322}
323
324void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329}
330
331Value *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
347unsigned 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>.
410unsigned 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
427Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429 "Expected 'reassoc'+'nsz' instruction");
430
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
433 return nullptr;
434
435 assert((I->getOpcode() == Instruction::FAdd ||
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
494
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
506
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
508 return R;
509 }
510
511 return nullptr;
512}
513
514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends");
517
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
521
522 // Simplified addends are placed <SimpVect>.
523 AddendVect SimpVect;
524
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529
530 const FAddend *ThisAddend = Addends[SymIdx];
531 if (!ThisAddend) {
532 // This addend was processed before.
533 continue;
534 }
535
536 Value *Val = ThisAddend->getSymVal();
537
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543 // call later.
544
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(ThisAddend);
547
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(T);
561 }
562 }
563
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569 R += *SimpVect[Idx];
570
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(&R);
575 }
576 }
577 }
578
579 assert((NextTmpIdx <= std::size(TmpResult) + 1) && "out-of-bound access");
580
581 Value *Result;
582 if (!SimpVect.empty())
583 Result = createNaryFAdd(SimpVect, InstrQuota);
584 else {
585 // The addition is folded to 0.0.
586 Result = ConstantFP::get(Instr->getType(), 0.0);
587 }
588
589 return Result;
590}
591
592Value *FAddCombine::createNaryFAdd
593 (const AddendVect &Opnds, unsigned InstrQuota) {
594 assert(!Opnds.empty() && "Expect at least one addend");
595
596 // Step 1: Check if the # of instructions needed exceeds the quota.
597
598 unsigned InstrNeeded = calcInstrNumber(Opnds);
599 if (InstrNeeded > InstrQuota)
600 return nullptr;
601
602 initCreateInstNum();
603
604 // step 2: Emit the N-ary addition.
605 // Note that at most three instructions are involved in Fadd-InstCombine: the
606 // addition in question, and at most two neighboring instructions.
607 // The resulting optimized addition should have at least one less instruction
608 // than the original addition expression tree. This implies that the resulting
609 // N-ary addition has at most two instructions, and we don't need to worry
610 // about tree-height when constructing the N-ary addition.
611
612 Value *LastVal = nullptr;
613 bool LastValNeedNeg = false;
614
615 // Iterate the addends, creating fadd/fsub using adjacent two addends.
616 for (const FAddend *Opnd : Opnds) {
617 bool NeedNeg;
618 Value *V = createAddendVal(*Opnd, NeedNeg);
619 if (!LastVal) {
620 LastVal = V;
621 LastValNeedNeg = NeedNeg;
622 continue;
623 }
624
625 if (LastValNeedNeg == NeedNeg) {
626 LastVal = createFAdd(LastVal, V);
627 continue;
628 }
629
630 if (LastValNeedNeg)
631 LastVal = createFSub(V, LastVal);
632 else
633 LastVal = createFSub(LastVal, V);
634
635 LastValNeedNeg = false;
636 }
637
638 if (LastValNeedNeg) {
639 LastVal = createFNeg(LastVal);
640 }
641
642#ifndef NDEBUG
643 assert(CreateInstrNum == InstrNeeded &&
644 "Inconsistent in instruction numbers");
645#endif
646
647 return LastVal;
648}
649
650Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
651 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
652 if (Instruction *I = dyn_cast<Instruction>(V))
653 createInstPostProc(I);
654 return V;
655}
656
657Value *FAddCombine::createFNeg(Value *V) {
658 Value *NewV = Builder.CreateFNeg(V);
659 if (Instruction *I = dyn_cast<Instruction>(NewV))
660 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
661 return NewV;
662}
663
664Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
665 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
666 if (Instruction *I = dyn_cast<Instruction>(V))
667 createInstPostProc(I);
668 return V;
669}
670
671Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
673 if (Instruction *I = dyn_cast<Instruction>(V))
674 createInstPostProc(I);
675 return V;
676}
677
678void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
679 NewInstr->setDebugLoc(Instr->getDebugLoc());
680
681 // Keep track of the number of instruction created.
682 if (!NoNumber)
683 incCreateInstNum();
684
685 // Propagate fast-math flags
686 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
687}
688
689// Return the number of instruction needed to emit the N-ary addition.
690// NOTE: Keep this function in sync with createAddendVal().
691unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
692 unsigned OpndNum = Opnds.size();
693 unsigned InstrNeeded = OpndNum - 1;
694
695 // Adjust the number of instructions needed to emit the N-ary add.
696 for (const FAddend *Opnd : Opnds) {
697 if (Opnd->isConstant())
698 continue;
699
700 // The constant check above is really for a few special constant
701 // coefficients.
702 if (isa<UndefValue>(Opnd->getSymVal()))
703 continue;
704
705 const FAddendCoef &CE = Opnd->getCoef();
706 // Let the addend be "c * x". If "c == +/-1", the value of the addend
707 // is immediately available; otherwise, it needs exactly one instruction
708 // to evaluate the value.
709 if (!CE.isMinusOne() && !CE.isOne())
710 InstrNeeded++;
711 }
712 return InstrNeeded;
713}
714
715// Input Addend Value NeedNeg(output)
716// ================================================================
717// Constant C C false
718// <+/-1, V> V coefficient is -1
719// <2/-2, V> "fadd V, V" coefficient is -2
720// <C, V> "fmul V, C" false
721//
722// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
723Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
724 const FAddendCoef &Coeff = Opnd.getCoef();
725
726 if (Opnd.isConstant()) {
727 NeedNeg = false;
728 return Coeff.getValue(Instr->getType());
729 }
730
731 Value *OpndVal = Opnd.getSymVal();
732
733 if (Coeff.isMinusOne() || Coeff.isOne()) {
734 NeedNeg = Coeff.isMinusOne();
735 return OpndVal;
736 }
737
738 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
739 NeedNeg = Coeff.isMinusTwo();
740 return createFAdd(OpndVal, OpndVal);
741 }
742
743 NeedNeg = false;
744 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
745}
746
747// Checks if any operand is negative and we can convert add to sub.
748// This function checks for following negative patterns
749// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
750// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
751// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
753 InstCombiner::BuilderTy &Builder) {
754 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
755
756 // This function creates 2 instructions to replace ADD, we need at least one
757 // of LHS or RHS to have one use to ensure benefit in transform.
758 if (!LHS->hasOneUse() && !RHS->hasOneUse())
759 return nullptr;
760
761 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
762 const APInt *C1 = nullptr, *C2 = nullptr;
763
764 // if ONE is on other side, swap
765 if (match(RHS, m_Add(m_Value(X), m_One())))
766 std::swap(LHS, RHS);
767
768 if (match(LHS, m_Add(m_Value(X), m_One()))) {
769 // if XOR on other side, swap
770 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
771 std::swap(X, RHS);
772
773 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
774 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
775 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
776 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
777 Value *NewAnd = Builder.CreateAnd(Z, *C1);
778 return Builder.CreateSub(RHS, NewAnd, "sub");
779 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
780 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
781 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
782 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
783 return Builder.CreateSub(RHS, NewOr, "sub");
784 }
785 }
786 }
787
788 // Restore LHS and RHS
789 LHS = I.getOperand(0);
790 RHS = I.getOperand(1);
791
792 // if XOR is on other side, swap
793 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
794 std::swap(LHS, RHS);
795
796 // C2 is ODD
797 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
798 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
799 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
800 if (C1->countr_zero() == 0)
801 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
802 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
803 return Builder.CreateSub(RHS, NewOr, "sub");
804 }
805 return nullptr;
806}
807
808/// Wrapping flags may allow combining constants separated by an extend.
810 InstCombiner::BuilderTy &Builder) {
811 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
812 Type *Ty = Add.getType();
813 Constant *Op1C;
814 if (!match(Op1, m_Constant(Op1C)))
815 return nullptr;
816
817 // Try this match first because it results in an add in the narrow type.
818 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
819 Value *X;
820 const APInt *C1, *C2;
821 if (match(Op1, m_APInt(C1)) &&
823 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824 Constant *NewC =
825 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
826 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
827 }
828
829 // More general combining of constants in the wide type.
830 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
831 // or (zext nneg (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
832 Constant *NarrowC;
833 if (match(Op0, m_OneUse(m_SExtLike(
834 m_NSWAddLike(m_Value(X), m_Constant(NarrowC)))))) {
835 Value *WideC = Builder.CreateSExt(NarrowC, Ty);
836 Value *NewC = Builder.CreateAdd(WideC, Op1C);
837 Value *WideX = Builder.CreateSExt(X, Ty);
838 return BinaryOperator::CreateAdd(WideX, NewC);
839 }
840 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
841 if (match(Op0,
843 Value *WideC = Builder.CreateZExt(NarrowC, Ty);
844 Value *NewC = Builder.CreateAdd(WideC, Op1C);
845 Value *WideX = Builder.CreateZExt(X, Ty);
846 return BinaryOperator::CreateAdd(WideX, NewC);
847 }
848 return nullptr;
849}
850
852 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
853 Type *Ty = Add.getType();
854 Constant *Op1C;
855 if (!match(Op1, m_ImmConstant(Op1C)))
856 return nullptr;
857
859 return NV;
860
861 Value *X;
862 Constant *Op00C;
863
864 // add (sub C1, X), C2 --> sub (add C1, C2), X
865 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
866 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
867
868 Value *Y;
869
870 // add (sub X, Y), -1 --> add (not Y), X
871 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
872 match(Op1, m_AllOnes()))
873 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
874
875 // zext(bool) + C -> bool ? C + 1 : C
876 if (match(Op0, m_ZExt(m_Value(X))) &&
877 X->getType()->getScalarSizeInBits() == 1)
878 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
879 // sext(bool) + C -> bool ? C - 1 : C
880 if (match(Op0, m_SExt(m_Value(X))) &&
881 X->getType()->getScalarSizeInBits() == 1)
882 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
883
884 // ~X + C --> (C-1) - X
885 if (match(Op0, m_Not(m_Value(X)))) {
886 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
887 auto *COne = ConstantInt::get(Op1C->getType(), 1);
888 bool WillNotSOV = willNotOverflowSignedSub(Op1C, COne, Add);
889 BinaryOperator *Res =
890 BinaryOperator::CreateSub(ConstantExpr::getSub(Op1C, COne), X);
891 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
892 return Res;
893 }
894
895 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
896 const APInt *C;
897 unsigned BitWidth = Ty->getScalarSizeInBits();
898 if (match(Op0, m_OneUse(m_AShr(m_Value(X),
900 match(Op1, m_One()))
901 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
902
903 if (!match(Op1, m_APInt(C)))
904 return nullptr;
905
906 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
907 Constant *Op01C;
908 if (match(Op0, m_DisjointOr(m_Value(X), m_ImmConstant(Op01C)))) {
909 BinaryOperator *NewAdd =
910 BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
911 NewAdd->setHasNoSignedWrap(Add.hasNoSignedWrap() &&
912 willNotOverflowSignedAdd(Op01C, Op1C, Add));
913 NewAdd->setHasNoUnsignedWrap(Add.hasNoUnsignedWrap());
914 return NewAdd;
915 }
916
917 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
918 const APInt *C2;
919 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
920 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
921
922 if (C->isSignMask()) {
923 // If wrapping is not allowed, then the addition must set the sign bit:
924 // X + (signmask) --> X | signmask
925 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
926 return BinaryOperator::CreateOr(Op0, Op1);
927
928 // If wrapping is allowed, then the addition flips the sign bit of LHS:
929 // X + (signmask) --> X ^ signmask
930 return BinaryOperator::CreateXor(Op0, Op1);
931 }
932
933 // Is this add the last step in a convoluted sext?
934 // add(zext(xor i16 X, -32768), -32768) --> sext X
935 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
936 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
937 return CastInst::Create(Instruction::SExt, X, Ty);
938
939 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
940 // (X ^ signmask) + C --> (X + (signmask ^ C))
941 if (C2->isSignMask())
942 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
943
944 // If X has no high-bits set above an xor mask:
945 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
946 if (C2->isMask()) {
947 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
948 if ((*C2 | LHSKnown.Zero).isAllOnes())
949 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
950 }
951
952 // Look for a math+logic pattern that corresponds to sext-in-register of a
953 // value with cleared high bits. Convert that into a pair of shifts:
954 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
955 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
956 if (Op0->hasOneUse() && *C2 == -(*C)) {
957 unsigned BitWidth = Ty->getScalarSizeInBits();
958 unsigned ShAmt = 0;
959 if (C->isPowerOf2())
960 ShAmt = BitWidth - C->logBase2() - 1;
961 else if (C2->isPowerOf2())
962 ShAmt = BitWidth - C2->logBase2() - 1;
964 0, &Add)) {
965 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
966 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
967 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
968 }
969 }
970 }
971
972 if (C->isOne() && Op0->hasOneUse()) {
973 // add (sext i1 X), 1 --> zext (not X)
974 // TODO: The smallest IR representation is (select X, 0, 1), and that would
975 // not require the one-use check. But we need to remove a transform in
976 // visitSelect and make sure that IR value tracking for select is equal or
977 // better than for these ops.
978 if (match(Op0, m_SExt(m_Value(X))) &&
979 X->getType()->getScalarSizeInBits() == 1)
980 return new ZExtInst(Builder.CreateNot(X), Ty);
981
982 // Shifts and add used to flip and mask off the low bit:
983 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
984 const APInt *C3;
985 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
986 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
987 Value *NotX = Builder.CreateNot(X);
988 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
989 }
990 }
991
992 // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
993 // TODO: There's a general form for any constant on the outer add.
994 if (C->isOne()) {
995 if (match(Op0, m_ZExt(m_Add(m_Value(X), m_AllOnes())))) {
997 if (llvm::isKnownNonZero(X, Q))
998 return new ZExtInst(X, Ty);
999 }
1000 }
1001
1002 return nullptr;
1003}
1004
1005// match variations of a^2 + 2*a*b + b^2
1006//
1007// to reuse the code between the FP and Int versions, the instruction OpCodes
1008// and constant types have been turned into template parameters.
1009//
1010// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1011// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1012// (we're matching `X<<1` instead of `X*2` for Int)
1013template <bool FP, typename Mul2Rhs>
1014static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1015 Value *&B) {
1016 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1017 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1018 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1019
1020 // (a * a) + (((a * 2) + b) * b)
1021 if (match(&I, m_c_BinOp(
1022 AddOp, m_OneUse(m_BinOp(MulOp, m_Value(A), m_Deferred(A))),
1024 MulOp,
1025 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(A), M2Rhs),
1026 m_Value(B)),
1027 m_Deferred(B))))))
1028 return true;
1029
1030 // ((a * b) * 2) or ((a * 2) * b)
1031 // +
1032 // (a * a + b * b) or (b * b + a * a)
1033 return match(
1034 &I, m_c_BinOp(
1035 AddOp,
1038 Mul2Op, m_BinOp(MulOp, m_Value(A), m_Value(B)), M2Rhs)),
1039 m_OneUse(m_c_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(A), M2Rhs),
1040 m_Value(B)))),
1041 m_OneUse(
1042 m_c_BinOp(AddOp, m_BinOp(MulOp, m_Deferred(A), m_Deferred(A)),
1043 m_BinOp(MulOp, m_Deferred(B), m_Deferred(B))))));
1044}
1045
1046// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1048 Value *A, *B;
1049 if (matchesSquareSum</*FP*/ false>(I, m_SpecificInt(1), A, B)) {
1050 Value *AB = Builder.CreateAdd(A, B);
1051 return BinaryOperator::CreateMul(AB, AB);
1052 }
1053 return nullptr;
1054}
1055
1056// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1057// Requires `nsz` and `reassoc`.
1059 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1060 Value *A, *B;
1061 if (matchesSquareSum</*FP*/ true>(I, m_SpecificFP(2.0), A, B)) {
1062 Value *AB = Builder.CreateFAddFMF(A, B, &I);
1063 return BinaryOperator::CreateFMulFMF(AB, AB, &I);
1064 }
1065 return nullptr;
1066}
1067
1068// Matches multiplication expression Op * C where C is a constant. Returns the
1069// constant value in C and the other operand in Op. Returns true if such a
1070// match is found.
1071static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1072 const APInt *AI;
1073 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1074 C = *AI;
1075 return true;
1076 }
1077 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1078 C = APInt(AI->getBitWidth(), 1);
1079 C <<= *AI;
1080 return true;
1081 }
1082 return false;
1083}
1084
1085// Matches remainder expression Op % C where C is a constant. Returns the
1086// constant value in C and the other operand in Op. Returns the signedness of
1087// the remainder operation in IsSigned. Returns true if such a match is
1088// found.
1089static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1090 const APInt *AI;
1091 IsSigned = false;
1092 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1093 IsSigned = true;
1094 C = *AI;
1095 return true;
1096 }
1097 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1098 C = *AI;
1099 return true;
1100 }
1101 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1102 C = *AI + 1;
1103 return true;
1104 }
1105 return false;
1106}
1107
1108// Matches division expression Op / C with the given signedness as indicated
1109// by IsSigned, where C is a constant. Returns the constant value in C and the
1110// other operand in Op. Returns true if such a match is found.
1111static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1112 const APInt *AI;
1113 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1114 C = *AI;
1115 return true;
1116 }
1117 if (!IsSigned) {
1118 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1119 C = *AI;
1120 return true;
1121 }
1122 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1123 C = APInt(AI->getBitWidth(), 1);
1124 C <<= *AI;
1125 return true;
1126 }
1127 }
1128 return false;
1129}
1130
1131// Returns whether C0 * C1 with the given signedness overflows.
1132static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1133 bool overflow;
1134 if (IsSigned)
1135 (void)C0.smul_ov(C1, overflow);
1136 else
1137 (void)C0.umul_ov(C1, overflow);
1138 return overflow;
1139}
1140
1141// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1142// does not overflow.
1143// Simplifies (X / C0) * C1 + (X % C0) * C2 to
1144// (X / C0) * (C1 - C2 * C0) + X * C2
1146 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1147 Value *X, *MulOpV;
1148 APInt C0, MulOpC;
1149 bool IsSigned;
1150 // Match I = X % C0 + MulOpV * C0
1151 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1152 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1153 C0 == MulOpC) {
1154 Value *RemOpV;
1155 APInt C1;
1156 bool Rem2IsSigned;
1157 // Match MulOpC = RemOpV % C1
1158 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1159 IsSigned == Rem2IsSigned) {
1160 Value *DivOpV;
1161 APInt DivOpC;
1162 // Match RemOpV = X / C0
1163 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1164 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1165 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1166 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1167 : Builder.CreateURem(X, NewDivisor, "urem");
1168 }
1169 }
1170 }
1171
1172 // Match I = (X / C0) * C1 + (X % C0) * C2
1173 Value *Div, *Rem;
1174 APInt C1, C2;
1175 if (!LHS->hasOneUse() || !MatchMul(LHS, Div, C1))
1176 Div = LHS, C1 = APInt(I.getType()->getScalarSizeInBits(), 1);
1177 if (!RHS->hasOneUse() || !MatchMul(RHS, Rem, C2))
1178 Rem = RHS, C2 = APInt(I.getType()->getScalarSizeInBits(), 1);
1179 if (match(Div, m_IRem(m_Value(), m_Value()))) {
1180 std::swap(Div, Rem);
1181 std::swap(C1, C2);
1182 }
1183 Value *DivOpV;
1184 APInt DivOpC;
1185 if (MatchRem(Rem, X, C0, IsSigned) &&
1186 MatchDiv(Div, DivOpV, DivOpC, IsSigned) && X == DivOpV && C0 == DivOpC) {
1187 APInt NewC = C1 - C2 * C0;
1188 if (!NewC.isZero() && !Rem->hasOneUse())
1189 return nullptr;
1190 if (!isGuaranteedNotToBeUndef(X, &AC, &I, &DT))
1191 return nullptr;
1192 Value *MulXC2 = Builder.CreateMul(X, ConstantInt::get(X->getType(), C2));
1193 if (NewC.isZero())
1194 return MulXC2;
1195 return Builder.CreateAdd(
1196 Builder.CreateMul(Div, ConstantInt::get(X->getType(), NewC)), MulXC2);
1197 }
1198
1199 return nullptr;
1200}
1201
1202/// Fold
1203/// (1 << NBits) - 1
1204/// Into:
1205/// ~(-(1 << NBits))
1206/// Because a 'not' is better for bit-tracking analysis and other transforms
1207/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1209 InstCombiner::BuilderTy &Builder) {
1210 Value *NBits;
1211 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1212 return nullptr;
1213
1214 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1215 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1216 // Be wary of constant folding.
1217 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1218 // Always NSW. But NUW propagates from `add`.
1219 BOp->setHasNoSignedWrap();
1220 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1221 }
1222
1223 return BinaryOperator::CreateNot(NotMask, I.getName());
1224}
1225
1227 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1228 Type *Ty = I.getType();
1229 auto getUAddSat = [&]() {
1230 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1231 };
1232
1233 // add (umin X, ~Y), Y --> uaddsat X, Y
1234 Value *X, *Y;
1236 m_Deferred(Y))))
1237 return CallInst::Create(getUAddSat(), { X, Y });
1238
1239 // add (umin X, ~C), C --> uaddsat X, C
1240 const APInt *C, *NotC;
1241 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1242 *C == ~*NotC)
1243 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1244
1245 return nullptr;
1246}
1247
1248// Transform:
1249// (add A, (shl (neg B), Y))
1250// -> (sub A, (shl B, Y))
1252 const BinaryOperator &I) {
1253 Value *A, *B, *Cnt;
1254 if (match(&I,
1256 m_Value(A)))) {
1257 Value *NewShl = Builder.CreateShl(B, Cnt);
1258 return BinaryOperator::CreateSub(A, NewShl);
1259 }
1260 return nullptr;
1261}
1262
1263/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1265 // Division must be by power-of-2, but not the minimum signed value.
1266 Value *X;
1267 const APInt *DivC;
1268 if (!match(Add.getOperand(0), m_SDiv(m_Value(X), m_Power2(DivC))) ||
1269 DivC->isNegative())
1270 return nullptr;
1271
1272 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1273 // low bits set. It recognizes two canonical patterns:
1274 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1275 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1276 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1277 // Note that, by the time we end up here, if possible, ugt has been
1278 // canonicalized into eq.
1279 const APInt *MaskC, *MaskCCmp;
1281 if (!match(Add.getOperand(1),
1282 m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1283 m_APInt(MaskCCmp)))))
1284 return nullptr;
1285
1286 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1287 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1288 return nullptr;
1289
1290 APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1291 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1292 ? (*MaskC == (SMin | (*DivC - 1)))
1293 : (*DivC == 2 && *MaskC == SMin + 1);
1294 if (!IsMaskValid)
1295 return nullptr;
1296
1297 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1298 return BinaryOperator::CreateAShr(
1299 X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1300}
1301
1304 BinaryOperator &I) {
1305 assert((I.getOpcode() == Instruction::Add ||
1306 I.getOpcode() == Instruction::Or ||
1307 I.getOpcode() == Instruction::Sub) &&
1308 "Expecting add/or/sub instruction");
1309
1310 // We have a subtraction/addition between a (potentially truncated) *logical*
1311 // right-shift of X and a "select".
1312 Value *X, *Select;
1313 Instruction *LowBitsToSkip, *Extract;
1315 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1316 m_Instruction(Extract))),
1317 m_Value(Select))))
1318 return nullptr;
1319
1320 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1321 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1322 return nullptr;
1323
1324 Type *XTy = X->getType();
1325 bool HadTrunc = I.getType() != XTy;
1326
1327 // If there was a truncation of extracted value, then we'll need to produce
1328 // one extra instruction, so we need to ensure one instruction will go away.
1329 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1330 return nullptr;
1331
1332 // Extraction should extract high NBits bits, with shift amount calculated as:
1333 // low bits to skip = shift bitwidth - high bits to extract
1334 // The shift amount itself may be extended, and we need to look past zero-ext
1335 // when matching NBits, that will matter for matching later.
1336 Constant *C;
1337 Value *NBits;
1338 if (!match(
1339 LowBitsToSkip,
1342 APInt(C->getType()->getScalarSizeInBits(),
1343 X->getType()->getScalarSizeInBits()))))
1344 return nullptr;
1345
1346 // Sign-extending value can be zero-extended if we `sub`tract it,
1347 // or sign-extended otherwise.
1348 auto SkipExtInMagic = [&I](Value *&V) {
1349 if (I.getOpcode() == Instruction::Sub)
1350 match(V, m_ZExtOrSelf(m_Value(V)));
1351 else
1352 match(V, m_SExtOrSelf(m_Value(V)));
1353 };
1354
1355 // Now, finally validate the sign-extending magic.
1356 // `select` itself may be appropriately extended, look past that.
1357 SkipExtInMagic(Select);
1358
1360 const APInt *Thr;
1361 Value *SignExtendingValue, *Zero;
1362 bool ShouldSignext;
1363 // It must be a select between two values we will later establish to be a
1364 // sign-extending value and a zero constant. The condition guarding the
1365 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1366 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1367 m_Value(SignExtendingValue), m_Value(Zero))) ||
1368 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1369 return nullptr;
1370
1371 // icmp-select pair is commutative.
1372 if (!ShouldSignext)
1373 std::swap(SignExtendingValue, Zero);
1374
1375 // If we should not perform sign-extension then we must add/or/subtract zero.
1376 if (!match(Zero, m_Zero()))
1377 return nullptr;
1378 // Otherwise, it should be some constant, left-shifted by the same NBits we
1379 // had in `lshr`. Said left-shift can also be appropriately extended.
1380 // Again, we must look past zero-ext when looking for NBits.
1381 SkipExtInMagic(SignExtendingValue);
1382 Constant *SignExtendingValueBaseConstant;
1383 if (!match(SignExtendingValue,
1384 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1385 m_ZExtOrSelf(m_Specific(NBits)))))
1386 return nullptr;
1387 // If we `sub`, then the constant should be one, else it should be all-ones.
1388 if (I.getOpcode() == Instruction::Sub
1389 ? !match(SignExtendingValueBaseConstant, m_One())
1390 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1391 return nullptr;
1392
1393 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1394 Extract->getName() + ".sext");
1395 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1396 if (!HadTrunc)
1397 return NewAShr;
1398
1399 Builder.Insert(NewAShr);
1400 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1401}
1402
1403/// This is a specialization of a more general transform from
1404/// foldUsingDistributiveLaws. If that code can be made to work optimally
1405/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1407 InstCombiner::BuilderTy &Builder) {
1408 // TODO: Also handle mul by doubling the shift amount?
1409 assert((I.getOpcode() == Instruction::Add ||
1410 I.getOpcode() == Instruction::Sub) &&
1411 "Expected add/sub");
1412 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1413 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1414 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1415 return nullptr;
1416
1417 Value *X, *Y, *ShAmt;
1418 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1419 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1420 return nullptr;
1421
1422 // No-wrap propagates only when all ops have no-wrap.
1423 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1424 Op1->hasNoSignedWrap();
1425 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1426 Op1->hasNoUnsignedWrap();
1427
1428 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1429 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1430 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1431 NewI->setHasNoSignedWrap(HasNSW);
1432 NewI->setHasNoUnsignedWrap(HasNUW);
1433 }
1434 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1435 NewShl->setHasNoSignedWrap(HasNSW);
1436 NewShl->setHasNoUnsignedWrap(HasNUW);
1437 return NewShl;
1438}
1439
1440/// Reduce a sequence of masked half-width multiplies to a single multiply.
1441/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1443 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1444 // Skip the odd bitwidth types.
1445 if ((BitWidth & 0x1))
1446 return nullptr;
1447
1448 unsigned HalfBits = BitWidth >> 1;
1449 APInt HalfMask = APInt::getMaxValue(HalfBits);
1450
1451 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1452 Value *XLo, *YLo;
1453 Value *CrossSum;
1454 // Require one-use on the multiply to avoid increasing the number of
1455 // multiplications.
1456 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1457 m_OneUse(m_Mul(m_Value(YLo), m_Value(XLo))))))
1458 return nullptr;
1459
1460 // XLo = X & HalfMask
1461 // YLo = Y & HalfMask
1462 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1463 // to enhance robustness
1464 Value *X, *Y;
1465 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1466 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1467 return nullptr;
1468
1469 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1470 // X' can be either X or XLo in the pattern (and the same for Y')
1471 if (match(CrossSum,
1476 return BinaryOperator::CreateMul(X, Y);
1477
1478 return nullptr;
1479}
1480
1482 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1483 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1485 return replaceInstUsesWith(I, V);
1486
1488 return &I;
1489
1491 return X;
1492
1494 return Phi;
1495
1496 // (A*B)+(A*C) -> A*(B+C) etc
1498 return replaceInstUsesWith(I, V);
1499
1500 if (Instruction *R = foldBoxMultiply(I))
1501 return R;
1502
1504 return R;
1505
1507 return X;
1508
1510 return X;
1511
1513 return R;
1514
1516 return R;
1517
1518 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1519 Type *Ty = I.getType();
1520 if (Ty->isIntOrIntVectorTy(1))
1521 return BinaryOperator::CreateXor(LHS, RHS);
1522
1523 // X + X --> X << 1
1524 if (LHS == RHS) {
1525 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1526 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1527 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1528 return Shl;
1529 }
1530
1531 Value *A, *B;
1532 if (match(LHS, m_Neg(m_Value(A)))) {
1533 // -A + -B --> -(A + B)
1534 if (match(RHS, m_Neg(m_Value(B))))
1536
1537 // -A + B --> B - A
1538 auto *Sub = BinaryOperator::CreateSub(RHS, A);
1539 auto *OB0 = cast<OverflowingBinaryOperator>(LHS);
1540 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1541
1542 return Sub;
1543 }
1544
1545 // A + -B --> A - B
1546 if (match(RHS, m_Neg(m_Value(B))))
1547 return BinaryOperator::CreateSub(LHS, B);
1548
1550 return replaceInstUsesWith(I, V);
1551
1552 // (A + 1) + ~B --> A - B
1553 // ~B + (A + 1) --> A - B
1554 // (~B + A) + 1 --> A - B
1555 // (A + ~B) + 1 --> A - B
1556 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1558 return BinaryOperator::CreateSub(A, B);
1559
1560 // (A + RHS) + RHS --> A + (RHS << 1)
1562 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1563
1564 // LHS + (A + LHS) --> A + (LHS << 1)
1566 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1567
1568 {
1569 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1570 Constant *C1, *C2;
1571 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1572 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1573 (LHS->hasOneUse() || RHS->hasOneUse())) {
1574 Value *Sub = Builder.CreateSub(A, B);
1575 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1576 }
1577
1578 // Canonicalize a constant sub operand as an add operand for better folding:
1579 // (C1 - A) + B --> (B - A) + C1
1581 m_Value(B)))) {
1582 Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1583 return BinaryOperator::CreateAdd(Sub, C1);
1584 }
1585 }
1586
1587 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1589
1590 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1591 const APInt *C1, *C2;
1592 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1593 APInt one(C2->getBitWidth(), 1);
1594 APInt minusC1 = -(*C1);
1595 if (minusC1 == (one << *C2)) {
1596 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1597 return BinaryOperator::CreateSRem(RHS, NewRHS);
1598 }
1599 }
1600
1601 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1602 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1603 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1604 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1605 return BinaryOperator::CreateAnd(A, NewMask);
1606 }
1607
1608 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1609 if ((match(RHS, m_ZExt(m_Value(A))) &&
1611 (match(LHS, m_ZExt(m_Value(A))) &&
1613 return new ZExtInst(B, LHS->getType());
1614
1615 // zext(A) + sext(A) --> 0 if A is i1
1617 A->getType()->isIntOrIntVectorTy(1))
1618 return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1619
1620 // A+B --> A|B iff A and B have no bits set in common.
1621 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1622 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(&I)))
1623 return BinaryOperator::CreateDisjointOr(LHS, RHS);
1624
1625 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1626 return Ext;
1627
1628 // (add (xor A, B) (and A, B)) --> (or A, B)
1629 // (add (and A, B) (xor A, B)) --> (or A, B)
1630 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1632 return BinaryOperator::CreateOr(A, B);
1633
1634 // (add (or A, B) (and A, B)) --> (add A, B)
1635 // (add (and A, B) (or A, B)) --> (add A, B)
1636 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1638 // Replacing operands in-place to preserve nuw/nsw flags.
1639 replaceOperand(I, 0, A);
1640 replaceOperand(I, 1, B);
1641 return &I;
1642 }
1643
1644 // (add A (or A, -A)) --> (and (add A, -1) A)
1645 // (add A (or -A, A)) --> (and (add A, -1) A)
1646 // (add (or A, -A) A) --> (and (add A, -1) A)
1647 // (add (or -A, A) A) --> (and (add A, -1) A)
1649 m_Deferred(A)))))) {
1650 Value *Add =
1652 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1653 return BinaryOperator::CreateAnd(Add, A);
1654 }
1655
1656 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1657 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1658 if (match(&I,
1660 m_AllOnes()))) {
1662 Value *Dec = Builder.CreateAdd(A, AllOnes);
1663 Value *Not = Builder.CreateXor(A, AllOnes);
1664 return BinaryOperator::CreateAnd(Dec, Not);
1665 }
1666
1667 // Disguised reassociation/factorization:
1668 // ~(A * C1) + A
1669 // ((A * -C1) - 1) + A
1670 // ((A * -C1) + A) - 1
1671 // (A * (1 - C1)) - 1
1672 if (match(&I,
1674 m_Deferred(A)))) {
1675 Type *Ty = I.getType();
1676 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1677 Value *NewMul = Builder.CreateMul(A, NewMulC);
1678 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1679 }
1680
1681 // (A * -2**C) + B --> B - (A << C)
1682 const APInt *NegPow2C;
1683 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1684 m_Value(B)))) {
1685 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1686 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1687 return BinaryOperator::CreateSub(B, Shl);
1688 }
1689
1690 // Canonicalize signum variant that ends in add:
1691 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1696 m_OneUse(m_ICmp(Pred, m_Specific(A), m_ZeroInt()))))) &&
1697 Pred == CmpInst::ICMP_SGT) {
1698 Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1699 Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1700 return BinaryOperator::CreateOr(LHS, Zext);
1701 }
1702
1703 {
1704 Value *Cond, *Ext;
1705 Constant *C;
1706 // (add X, (sext/zext (icmp eq X, C)))
1707 // -> (select (icmp eq X, C), (add C, (sext/zext 1)), X)
1708 auto CondMatcher = m_CombineAnd(
1710
1711 if (match(&I,
1712 m_c_Add(m_Value(A),
1713 m_CombineAnd(m_Value(Ext), m_ZExtOrSExt(CondMatcher)))) &&
1714 Pred == ICmpInst::ICMP_EQ && Ext->hasOneUse()) {
1715 Value *Add = isa<ZExtInst>(Ext) ? InstCombiner::AddOne(C)
1718 }
1719 }
1720
1721 if (Instruction *Ashr = foldAddToAshr(I))
1722 return Ashr;
1723
1724 // (~X) + (~Y) --> -2 - (X + Y)
1725 {
1726 // To ensure we can save instructions we need to ensure that we consume both
1727 // LHS/RHS (i.e they have a `not`).
1728 bool ConsumesLHS, ConsumesRHS;
1729 if (isFreeToInvert(LHS, LHS->hasOneUse(), ConsumesLHS) && ConsumesLHS &&
1730 isFreeToInvert(RHS, RHS->hasOneUse(), ConsumesRHS) && ConsumesRHS) {
1733 assert(NotLHS != nullptr && NotRHS != nullptr &&
1734 "isFreeToInvert desynced with getFreelyInverted");
1735 Value *LHSPlusRHS = Builder.CreateAdd(NotLHS, NotRHS);
1736 return BinaryOperator::CreateSub(
1737 ConstantInt::getSigned(RHS->getType(), -2), LHSPlusRHS);
1738 }
1739 }
1740
1742 return R;
1743
1744 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1745 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1746 // computeKnownBits.
1747 bool Changed = false;
1748 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHSCache, RHSCache, I)) {
1749 Changed = true;
1750 I.setHasNoSignedWrap(true);
1751 }
1752 if (!I.hasNoUnsignedWrap() &&
1753 willNotOverflowUnsignedAdd(LHSCache, RHSCache, I)) {
1754 Changed = true;
1755 I.setHasNoUnsignedWrap(true);
1756 }
1757
1759 return V;
1760
1761 if (Instruction *V =
1763 return V;
1764
1766 return SatAdd;
1767
1768 // usub.sat(A, B) + B => umax(A, B)
1769 if (match(&I, m_c_BinOp(
1770 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1771 m_Deferred(B)))) {
1772 return replaceInstUsesWith(I,
1773 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1774 }
1775
1776 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1777 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1778 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1780 return replaceInstUsesWith(
1781 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1782 {Builder.CreateOr(A, B)}));
1783
1784 // Fold the log2_ceil idiom:
1785 // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
1786 // -->
1787 // BW - ctlz(A - 1, false)
1788 const APInt *XorC;
1789 if (match(&I,
1790 m_c_Add(
1791 m_ZExt(m_ICmp(Pred, m_Intrinsic<Intrinsic::ctpop>(m_Value(A)),
1792 m_One())),
1795 m_Intrinsic<Intrinsic::ctlz>(m_Deferred(A), m_One())))),
1796 m_APInt(XorC))))))) &&
1797 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
1798 *XorC == A->getType()->getScalarSizeInBits() - 1) {
1799 Value *Sub = Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()));
1800 Value *Ctlz = Builder.CreateIntrinsic(Intrinsic::ctlz, {A->getType()},
1801 {Sub, Builder.getFalse()});
1802 Value *Ret = Builder.CreateSub(
1803 ConstantInt::get(A->getType(), A->getType()->getScalarSizeInBits()),
1804 Ctlz, "", /*HasNUW*/ true, /*HasNSW*/ true);
1805 return replaceInstUsesWith(I, Builder.CreateZExtOrTrunc(Ret, I.getType()));
1806 }
1807
1808 if (Instruction *Res = foldSquareSumInt(I))
1809 return Res;
1810
1811 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1812 return Res;
1813
1815 return Res;
1816
1817 return Changed ? &I : nullptr;
1818}
1819
1820/// Eliminate an op from a linear interpolation (lerp) pattern.
1822 InstCombiner::BuilderTy &Builder) {
1823 Value *X, *Y, *Z;
1826 m_Value(Z))))),
1828 return nullptr;
1829
1830 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1831 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1832 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1833 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1834}
1835
1836/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1838 InstCombiner::BuilderTy &Builder) {
1839 assert((I.getOpcode() == Instruction::FAdd ||
1840 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1841 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1842 "FP factorization requires FMF");
1843
1844 if (Instruction *Lerp = factorizeLerp(I, Builder))
1845 return Lerp;
1846
1847 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1848 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1849 return nullptr;
1850
1851 Value *X, *Y, *Z;
1852 bool IsFMul;
1853 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1854 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1855 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1856 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1857 IsFMul = true;
1858 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1859 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1860 IsFMul = false;
1861 else
1862 return nullptr;
1863
1864 // (X * Z) + (Y * Z) --> (X + Y) * Z
1865 // (X * Z) - (Y * Z) --> (X - Y) * Z
1866 // (X / Z) + (Y / Z) --> (X + Y) / Z
1867 // (X / Z) - (Y / Z) --> (X - Y) / Z
1868 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1869 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1870 : Builder.CreateFSubFMF(X, Y, &I);
1871
1872 // Bail out if we just created a denormal constant.
1873 // TODO: This is copied from a previous implementation. Is it necessary?
1874 const APFloat *C;
1875 if (match(XY, m_APFloat(C)) && !C->isNormal())
1876 return nullptr;
1877
1878 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1880}
1881
1883 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1884 I.getFastMathFlags(),
1886 return replaceInstUsesWith(I, V);
1887
1889 return &I;
1890
1892 return X;
1893
1895 return Phi;
1896
1897 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1898 return FoldedFAdd;
1899
1900 // (-X) + Y --> Y - X
1901 Value *X, *Y;
1902 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1904
1905 // Similar to above, but look through fmul/fdiv for the negated term.
1906 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1907 Value *Z;
1909 m_Value(Z)))) {
1910 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1911 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1912 }
1913 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1914 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1916 m_Value(Z))) ||
1918 m_Value(Z)))) {
1919 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1920 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1921 }
1922
1923 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1924 // integer add followed by a promotion.
1925 if (Instruction *R = foldFBinOpOfIntCasts(I))
1926 return R;
1927
1928 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1929 // Handle specials cases for FAdd with selects feeding the operation
1931 return replaceInstUsesWith(I, V);
1932
1933 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1935 return F;
1936
1938 return F;
1939
1940 // Try to fold fadd into start value of reduction intrinsic.
1941 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1942 m_AnyZeroFP(), m_Value(X))),
1943 m_Value(Y)))) {
1944 // fadd (rdx 0.0, X), Y --> rdx Y, X
1945 return replaceInstUsesWith(
1946 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1947 {X->getType()}, {Y, X}, &I));
1948 }
1949 const APFloat *StartC, *C;
1950 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1951 m_APFloat(StartC), m_Value(X)))) &&
1952 match(RHS, m_APFloat(C))) {
1953 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1954 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1955 return replaceInstUsesWith(
1956 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1957 {X->getType()}, {NewStartC, X}, &I));
1958 }
1959
1960 // (X * MulC) + X --> X * (MulC + 1.0)
1961 Constant *MulC;
1962 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1963 m_Deferred(X)))) {
1965 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1966 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1967 }
1968
1969 // (-X - Y) + (X + Z) --> Z - Y
1971 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1972 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1973
1974 if (Value *V = FAddCombine(Builder).simplify(&I))
1975 return replaceInstUsesWith(I, V);
1976 }
1977
1978 // minumum(X, Y) + maximum(X, Y) => X + Y.
1979 if (match(&I,
1980 m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1981 m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1982 m_Deferred(Y))))) {
1984 // We cannot preserve ninf if nnan flag is not set.
1985 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1986 // while in optimized version NaN + Inf and this is a poison with ninf flag.
1987 if (!Result->hasNoNaNs())
1988 Result->setHasNoInfs(false);
1989 return Result;
1990 }
1991
1992 return nullptr;
1993}
1994
1995/// Optimize pointer differences into the same array into a size. Consider:
1996/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1997/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1999 Type *Ty, bool IsNUW) {
2000 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
2001 // this.
2002 bool Swapped = false;
2003 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
2004 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
2005 std::swap(LHS, RHS);
2006 Swapped = true;
2007 }
2008
2009 // Require at least one GEP with a common base pointer on both sides.
2010 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
2011 // (gep X, ...) - X
2012 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
2014 GEP1 = LHSGEP;
2015 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
2016 // (gep X, ...) - (gep X, ...)
2017 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
2018 RHSGEP->getOperand(0)->stripPointerCasts()) {
2019 GEP1 = LHSGEP;
2020 GEP2 = RHSGEP;
2021 }
2022 }
2023 }
2024
2025 if (!GEP1)
2026 return nullptr;
2027
2028 // To avoid duplicating the offset arithmetic, rewrite the GEP to use the
2029 // computed offset. This may erase the original GEP, so be sure to cache the
2030 // inbounds flag before emitting the offset.
2031 // TODO: We should probably do this even if there is only one GEP.
2032 bool RewriteGEPs = GEP2 != nullptr;
2033
2034 // Emit the offset of the GEP and an intptr_t.
2035 bool GEP1IsInBounds = GEP1->isInBounds();
2036 Value *Result = EmitGEPOffset(GEP1, RewriteGEPs);
2037
2038 // If this is a single inbounds GEP and the original sub was nuw,
2039 // then the final multiplication is also nuw.
2040 if (auto *I = dyn_cast<Instruction>(Result))
2041 if (IsNUW && !GEP2 && !Swapped && GEP1IsInBounds &&
2042 I->getOpcode() == Instruction::Mul)
2043 I->setHasNoUnsignedWrap();
2044
2045 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2046 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2047 if (GEP2) {
2048 bool GEP2IsInBounds = GEP2->isInBounds();
2049 Value *Offset = EmitGEPOffset(GEP2, RewriteGEPs);
2050 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2051 GEP1IsInBounds && GEP2IsInBounds);
2052 }
2053
2054 // If we have p - gep(p, ...) then we have to negate the result.
2055 if (Swapped)
2056 Result = Builder.CreateNeg(Result, "diff.neg");
2057
2058 return Builder.CreateIntCast(Result, Ty, true);
2059}
2060
2062 InstCombiner::BuilderTy &Builder) {
2063 Value *Op0 = I.getOperand(0);
2064 Value *Op1 = I.getOperand(1);
2065 Type *Ty = I.getType();
2066 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2067 if (!MinMax)
2068 return nullptr;
2069
2070 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2071 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2072 Value *X = MinMax->getLHS();
2073 Value *Y = MinMax->getRHS();
2074 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2075 (Op0->hasOneUse() || Op1->hasOneUse())) {
2076 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2077 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2078 return CallInst::Create(F, {X, Y});
2079 }
2080
2081 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2082 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2083 Value *Z;
2084 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2085 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2086 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2087 return BinaryOperator::CreateAdd(X, USub);
2088 }
2089 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2090 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2091 return BinaryOperator::CreateAdd(X, USub);
2092 }
2093 }
2094
2095 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2096 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2097 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2098 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2099 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2100 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2101 return CallInst::Create(F, {Op0, Z});
2102 }
2103
2104 return nullptr;
2105}
2106
2108 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2109 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2111 return replaceInstUsesWith(I, V);
2112
2114 return X;
2115
2117 return Phi;
2118
2119 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2120
2121 // If this is a 'B = x-(-A)', change to B = x+A.
2122 // We deal with this without involving Negator to preserve NSW flag.
2123 if (Value *V = dyn_castNegVal(Op1)) {
2124 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2125
2126 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2127 assert(BO->getOpcode() == Instruction::Sub &&
2128 "Expected a subtraction operator!");
2129 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2130 Res->setHasNoSignedWrap(true);
2131 } else {
2132 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2133 Res->setHasNoSignedWrap(true);
2134 }
2135
2136 return Res;
2137 }
2138
2139 // Try this before Negator to preserve NSW flag.
2141 return R;
2142
2143 Constant *C;
2144 if (match(Op0, m_ImmConstant(C))) {
2145 Value *X;
2146 Constant *C2;
2147
2148 // C-(X+C2) --> (C-C2)-X
2149 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2150 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2151 // => (C-C2)-X can have NSW/NUW
2152 bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2153 BinaryOperator *Res =
2154 BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2155 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2156 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2157 WillNotSOV);
2158 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2159 OBO1->hasNoUnsignedWrap());
2160 return Res;
2161 }
2162 }
2163
2164 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2165 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2166 return Ext;
2167
2168 bool Changed = false;
2169 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2170 Changed = true;
2171 I.setHasNoSignedWrap(true);
2172 }
2173 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2174 Changed = true;
2175 I.setHasNoUnsignedWrap(true);
2176 }
2177
2178 return Changed ? &I : nullptr;
2179 };
2180
2181 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2182 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2183 // a pure negation used by a select that looks like abs/nabs.
2184 bool IsNegation = match(Op0, m_ZeroInt());
2185 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2186 const Instruction *UI = dyn_cast<Instruction>(U);
2187 if (!UI)
2188 return false;
2189 return match(UI,
2190 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2191 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2192 })) {
2193 if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2194 I.hasNoSignedWrap(),
2195 Op1, *this))
2196 return BinaryOperator::CreateAdd(NegOp1, Op0);
2197 }
2198 if (IsNegation)
2199 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2200
2201 // (A*B)-(A*C) -> A*(B-C) etc
2203 return replaceInstUsesWith(I, V);
2204
2205 if (I.getType()->isIntOrIntVectorTy(1))
2206 return BinaryOperator::CreateXor(Op0, Op1);
2207
2208 // Replace (-1 - A) with (~A).
2209 if (match(Op0, m_AllOnes()))
2210 return BinaryOperator::CreateNot(Op1);
2211
2212 // (X + -1) - Y --> ~Y + X
2213 Value *X, *Y;
2214 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2215 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2216
2217 // Reassociate sub/add sequences to create more add instructions and
2218 // reduce dependency chains:
2219 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2220 Value *Z;
2222 m_Value(Z))))) {
2223 Value *XZ = Builder.CreateAdd(X, Z);
2224 Value *YW = Builder.CreateAdd(Y, Op1);
2225 return BinaryOperator::CreateSub(XZ, YW);
2226 }
2227
2228 // ((X - Y) - Op1) --> X - (Y + Op1)
2229 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2230 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Op0);
2231 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2232 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2233 Value *Add = Builder.CreateAdd(Y, Op1, "", /* HasNUW */ HasNUW,
2234 /* HasNSW */ HasNSW);
2235 BinaryOperator *Sub = BinaryOperator::CreateSub(X, Add);
2236 Sub->setHasNoUnsignedWrap(HasNUW);
2237 Sub->setHasNoSignedWrap(HasNSW);
2238 return Sub;
2239 }
2240
2241 {
2242 // (X + Z) - (Y + Z) --> (X - Y)
2243 // This is done in other passes, but we want to be able to consume this
2244 // pattern in InstCombine so we can generate it without creating infinite
2245 // loops.
2246 if (match(Op0, m_Add(m_Value(X), m_Value(Z))) &&
2247 match(Op1, m_c_Add(m_Value(Y), m_Specific(Z))))
2248 return BinaryOperator::CreateSub(X, Y);
2249
2250 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2251 Constant *CX, *CY;
2252 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_ImmConstant(CX)))) &&
2253 match(Op1, m_OneUse(m_Add(m_Value(Y), m_ImmConstant(CY))))) {
2254 Value *OpsSub = Builder.CreateSub(X, Y);
2255 Constant *ConstsSub = ConstantExpr::getSub(CX, CY);
2256 return BinaryOperator::CreateAdd(OpsSub, ConstsSub);
2257 }
2258 }
2259
2260 // (~X) - (~Y) --> Y - X
2261 {
2262 // Need to ensure we can consume at least one of the `not` instructions,
2263 // otherwise this can inf loop.
2264 bool ConsumesOp0, ConsumesOp1;
2265 if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
2266 isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
2267 (ConsumesOp0 || ConsumesOp1)) {
2268 Value *NotOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
2269 Value *NotOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
2270 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2271 "isFreeToInvert desynced with getFreelyInverted");
2272 return BinaryOperator::CreateSub(NotOp1, NotOp0);
2273 }
2274 }
2275
2276 auto m_AddRdx = [](Value *&Vec) {
2277 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2278 };
2279 Value *V0, *V1;
2280 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2281 V0->getType() == V1->getType()) {
2282 // Difference of sums is sum of differences:
2283 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2284 Value *Sub = Builder.CreateSub(V0, V1);
2285 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2286 {Sub->getType()}, {Sub});
2287 return replaceInstUsesWith(I, Rdx);
2288 }
2289
2290 if (Constant *C = dyn_cast<Constant>(Op0)) {
2291 Value *X;
2292 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2293 // C - (zext bool) --> bool ? C - 1 : C
2295 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2296 // C - (sext bool) --> bool ? C + 1 : C
2298
2299 // C - ~X == X + (1+C)
2300 if (match(Op1, m_Not(m_Value(X))))
2301 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2302
2303 // Try to fold constant sub into select arguments.
2304 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2305 if (Instruction *R = FoldOpIntoSelect(I, SI))
2306 return R;
2307
2308 // Try to fold constant sub into PHI values.
2309 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2310 if (Instruction *R = foldOpIntoPhi(I, PN))
2311 return R;
2312
2313 Constant *C2;
2314
2315 // C-(C2-X) --> X+(C-C2)
2316 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2317 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2318 }
2319
2320 const APInt *Op0C;
2321 if (match(Op0, m_APInt(Op0C))) {
2322 if (Op0C->isMask()) {
2323 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2324 // zero. We don't use information from dominating conditions so this
2325 // transform is easier to reverse if necessary.
2328 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2329 return BinaryOperator::CreateXor(Op1, Op0);
2330 }
2331
2332 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2333 // (C3 - ((C2 & C3) - 1)) is pow2
2334 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2335 // C2 is negative pow2 || sub nuw
2336 const APInt *C2, *C3;
2337 BinaryOperator *InnerSub;
2338 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2339 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2340 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2341 APInt C2AndC3 = *C2 & *C3;
2342 APInt C2AndC3Minus1 = C2AndC3 - 1;
2343 APInt C2AddC3 = *C2 + *C3;
2344 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2345 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2346 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2347 return BinaryOperator::CreateAdd(
2348 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2349 }
2350 }
2351 }
2352
2353 {
2354 Value *Y;
2355 // X-(X+Y) == -Y X-(Y+X) == -Y
2356 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2358
2359 // (X-Y)-X == -Y
2360 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2362 }
2363
2364 // (sub (or A, B) (and A, B)) --> (xor A, B)
2365 {
2366 Value *A, *B;
2367 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2368 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2369 return BinaryOperator::CreateXor(A, B);
2370 }
2371
2372 // (sub (add A, B) (or A, B)) --> (and A, B)
2373 {
2374 Value *A, *B;
2375 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2376 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2377 return BinaryOperator::CreateAnd(A, B);
2378 }
2379
2380 // (sub (add A, B) (and A, B)) --> (or A, B)
2381 {
2382 Value *A, *B;
2383 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2385 return BinaryOperator::CreateOr(A, B);
2386 }
2387
2388 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2389 {
2390 Value *A, *B;
2391 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2392 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2393 (Op0->hasOneUse() || Op1->hasOneUse()))
2395 }
2396
2397 // (sub (or A, B), (xor A, B)) --> (and A, B)
2398 {
2399 Value *A, *B;
2400 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2401 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2402 return BinaryOperator::CreateAnd(A, B);
2403 }
2404
2405 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2406 {
2407 Value *A, *B;
2408 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2409 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2410 (Op0->hasOneUse() || Op1->hasOneUse()))
2412 }
2413
2414 {
2415 Value *Y;
2416 // ((X | Y) - X) --> (~X & Y)
2417 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2418 return BinaryOperator::CreateAnd(
2419 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2420 }
2421
2422 {
2423 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2424 Value *X;
2425 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2426 m_OneUse(m_Neg(m_Value(X))))))) {
2428 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2429 }
2430 }
2431
2432 {
2433 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2434 Constant *C;
2435 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2438 }
2439 }
2440
2441 {
2442 // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2443 // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2444 Value *C, *X;
2445 auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2446 return match(LHS, m_OneUse(m_c_Xor(m_Value(X), m_Specific(RHS)))) &&
2447 match(RHS, m_SExt(m_Value(C))) &&
2448 (C->getType()->getScalarSizeInBits() == 1);
2449 };
2450 if (m_SubXorCmp(Op0, Op1))
2452 if (m_SubXorCmp(Op1, Op0))
2454 }
2455
2457 return R;
2458
2460 return R;
2461
2462 {
2463 // If we have a subtraction between some value and a select between
2464 // said value and something else, sink subtraction into select hands, i.e.:
2465 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2466 // ->
2467 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2468 // or
2469 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2470 // ->
2471 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2472 // This will result in select between new subtraction and 0.
2473 auto SinkSubIntoSelect =
2474 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2475 auto SubBuilder) -> Instruction * {
2476 Value *Cond, *TrueVal, *FalseVal;
2477 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2478 m_Value(FalseVal)))))
2479 return nullptr;
2480 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2481 return nullptr;
2482 // While it is really tempting to just create two subtractions and let
2483 // InstCombine fold one of those to 0, it isn't possible to do so
2484 // because of worklist visitation order. So ugly it is.
2485 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2486 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2487 Constant *Zero = Constant::getNullValue(Ty);
2488 SelectInst *NewSel =
2489 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2490 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2491 // Preserve prof metadata if any.
2492 NewSel->copyMetadata(cast<Instruction>(*Select));
2493 return NewSel;
2494 };
2495 if (Instruction *NewSel = SinkSubIntoSelect(
2496 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2497 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2498 return Builder->CreateSub(OtherHandOfSelect,
2499 /*OtherHandOfSub=*/Op1);
2500 }))
2501 return NewSel;
2502 if (Instruction *NewSel = SinkSubIntoSelect(
2503 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2504 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2505 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2506 OtherHandOfSelect);
2507 }))
2508 return NewSel;
2509 }
2510
2511 // (X - (X & Y)) --> (X & ~Y)
2512 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2513 (Op1->hasOneUse() || isa<Constant>(Y)))
2514 return BinaryOperator::CreateAnd(
2515 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2516
2517 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2518 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2519 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2520 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2521 // As long as Y is freely invertible, this will be neutral or a win.
2522 // Note: We don't generate the inverse max/min, just create the 'not' of
2523 // it and let other folds do the rest.
2524 if (match(Op0, m_Not(m_Value(X))) &&
2525 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2526 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2527 Value *Not = Builder.CreateNot(Op1);
2528 return BinaryOperator::CreateSub(Not, X);
2529 }
2530 if (match(Op1, m_Not(m_Value(X))) &&
2531 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2532 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2533 Value *Not = Builder.CreateNot(Op0);
2534 return BinaryOperator::CreateSub(X, Not);
2535 }
2536
2537 // Optimize pointer differences into the same array into a size. Consider:
2538 // &A[10] - &A[0]: we should compile this to "10".
2539 Value *LHSOp, *RHSOp;
2540 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2541 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2542 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2543 I.hasNoUnsignedWrap()))
2544 return replaceInstUsesWith(I, Res);
2545
2546 // trunc(p)-trunc(q) -> trunc(p-q)
2547 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2548 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2549 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2550 /* IsNUW */ false))
2551 return replaceInstUsesWith(I, Res);
2552
2553 // Canonicalize a shifty way to code absolute value to the common pattern.
2554 // There are 2 potential commuted variants.
2555 // We're relying on the fact that we only do this transform when the shift has
2556 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2557 // instructions).
2558 Value *A;
2559 const APInt *ShAmt;
2560 Type *Ty = I.getType();
2561 unsigned BitWidth = Ty->getScalarSizeInBits();
2562 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2563 Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2564 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2565 // B = ashr i32 A, 31 ; smear the sign bit
2566 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2567 // --> (A < 0) ? -A : A
2568 Value *IsNeg = Builder.CreateIsNeg(A);
2569 // Copy the nsw flags from the sub to the negate.
2570 Value *NegA = I.hasNoUnsignedWrap()
2571 ? Constant::getNullValue(A->getType())
2572 : Builder.CreateNeg(A, "", I.hasNoSignedWrap());
2573 return SelectInst::Create(IsNeg, NegA, A);
2574 }
2575
2576 // If we are subtracting a low-bit masked subset of some value from an add
2577 // of that same value with no low bits changed, that is clearing some low bits
2578 // of the sum:
2579 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2580 const APInt *AddC, *AndC;
2581 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2582 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2583 unsigned Cttz = AddC->countr_zero();
2584 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2585 if ((HighMask & *AndC).isZero())
2586 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2587 }
2588
2589 if (Instruction *V =
2591 return V;
2592
2593 // X - usub.sat(X, Y) => umin(X, Y)
2594 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2595 m_Value(Y)))))
2596 return replaceInstUsesWith(
2597 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2598
2599 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2600 // TODO: The one-use restriction is not strictly necessary, but it may
2601 // require improving other pattern matching and/or codegen.
2602 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2603 return replaceInstUsesWith(
2604 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2605
2606 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2607 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2608 return replaceInstUsesWith(
2609 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2610
2611 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2612 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2613 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2614 return BinaryOperator::CreateNeg(USub);
2615 }
2616
2617 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2618 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2619 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2620 return BinaryOperator::CreateNeg(USub);
2621 }
2622
2623 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2624 if (match(Op0, m_SpecificInt(BitWidth)) &&
2625 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2626 return replaceInstUsesWith(
2627 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2628 {Builder.CreateNot(X)}));
2629
2630 // Reduce multiplies for difference-of-squares by factoring:
2631 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2632 if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2633 match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2634 auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2635 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2636 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2637 OBO1->hasNoSignedWrap() && BitWidth > 2;
2638 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2639 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2640 Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2641 Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2642 Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2643 return replaceInstUsesWith(I, Mul);
2644 }
2645
2646 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2647 if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2649 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2650 Value *Sub =
2651 Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2652 Value *Call =
2653 Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2654 return replaceInstUsesWith(I, Call);
2655 }
2656 }
2657
2659 return Res;
2660
2661 return TryToNarrowDeduceFlags();
2662}
2663
2664/// This eliminates floating-point negation in either 'fneg(X)' or
2665/// 'fsub(-0.0, X)' form by combining into a constant operand.
2667 // This is limited with one-use because fneg is assumed better for
2668 // reassociation and cheaper in codegen than fmul/fdiv.
2669 // TODO: Should the m_OneUse restriction be removed?
2670 Instruction *FNegOp;
2671 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2672 return nullptr;
2673
2674 Value *X;
2675 Constant *C;
2676
2677 // Fold negation into constant operand.
2678 // -(X * C) --> X * (-C)
2679 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2680 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2681 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2682 // -(X / C) --> X / (-C)
2683 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2684 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2685 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2686 // -(C / X) --> (-C) / X
2687 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2688 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2690
2691 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2692 // not apply to the fdiv. Everything else propagates from the fneg.
2693 // TODO: We could propagate nsz/ninf from fdiv alone?
2694 FastMathFlags FMF = I.getFastMathFlags();
2695 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2696 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2697 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2698 return FDiv;
2699 }
2700 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2701 // -(X + C) --> -X + -C --> -C - X
2702 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2703 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2704 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2705
2706 return nullptr;
2707}
2708
2709Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2710 Instruction &FMFSource) {
2711 Value *X, *Y;
2712 if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2713 return cast<Instruction>(Builder.CreateFMulFMF(
2714 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2715 }
2716
2717 if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2718 return cast<Instruction>(Builder.CreateFDivFMF(
2719 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2720 }
2721
2722 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2723 // Make sure to preserve flags and metadata on the call.
2724 if (II->getIntrinsicID() == Intrinsic::ldexp) {
2725 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2728
2730 II->getCalledFunction(),
2731 {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2732 New->copyMetadata(*II);
2733 return New;
2734 }
2735 }
2736
2737 return nullptr;
2738}
2739
2741 Value *Op = I.getOperand(0);
2742
2743 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2744 getSimplifyQuery().getWithInstruction(&I)))
2745 return replaceInstUsesWith(I, V);
2746
2748 return X;
2749
2750 Value *X, *Y;
2751
2752 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2753 if (I.hasNoSignedZeros() &&
2756
2757 Value *OneUse;
2758 if (!match(Op, m_OneUse(m_Value(OneUse))))
2759 return nullptr;
2760
2761 if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2762 return replaceInstUsesWith(I, R);
2763
2764 // Try to eliminate fneg if at least 1 arm of the select is negated.
2765 Value *Cond;
2766 if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2767 // Unlike most transforms, this one is not safe to propagate nsz unless
2768 // it is present on the original select. We union the flags from the select
2769 // and fneg and then remove nsz if needed.
2770 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2771 S->copyFastMathFlags(&I);
2772 if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2773 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2774 S->setFastMathFlags(FMF);
2775 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2776 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2777 S->setHasNoSignedZeros(false);
2778 }
2779 };
2780 // -(Cond ? -P : Y) --> Cond ? P : -Y
2781 Value *P;
2782 if (match(X, m_FNeg(m_Value(P)))) {
2783 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2784 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2785 propagateSelectFMF(NewSel, P == Y);
2786 return NewSel;
2787 }
2788 // -(Cond ? X : -P) --> Cond ? -X : P
2789 if (match(Y, m_FNeg(m_Value(P)))) {
2790 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2791 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2792 propagateSelectFMF(NewSel, P == X);
2793 return NewSel;
2794 }
2795
2796 // -(Cond ? X : C) --> Cond ? -X : -C
2797 // -(Cond ? C : Y) --> Cond ? -C : -Y
2798 if (match(X, m_ImmConstant()) || match(Y, m_ImmConstant())) {
2799 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2800 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2801 SelectInst *NewSel = SelectInst::Create(Cond, NegX, NegY);
2802 propagateSelectFMF(NewSel, /*CommonOperand=*/true);
2803 return NewSel;
2804 }
2805 }
2806
2807 // fneg (copysign x, y) -> copysign x, (fneg y)
2808 if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2809 // The source copysign has an additional value input, so we can't propagate
2810 // flags the copysign doesn't also have.
2811 FastMathFlags FMF = I.getFastMathFlags();
2812 FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2813
2816
2817 Value *NegY = Builder.CreateFNeg(Y);
2818 Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2819 return replaceInstUsesWith(I, NewCopySign);
2820 }
2821
2822 return nullptr;
2823}
2824
2826 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2827 I.getFastMathFlags(),
2828 getSimplifyQuery().getWithInstruction(&I)))
2829 return replaceInstUsesWith(I, V);
2830
2832 return X;
2833
2835 return Phi;
2836
2837 // Subtraction from -0.0 is the canonical form of fneg.
2838 // fsub -0.0, X ==> fneg X
2839 // fsub nsz 0.0, X ==> fneg nsz X
2840 //
2841 // FIXME This matcher does not respect FTZ or DAZ yet:
2842 // fsub -0.0, Denorm ==> +-0
2843 // fneg Denorm ==> -Denorm
2844 Value *Op;
2845 if (match(&I, m_FNeg(m_Value(Op))))
2847
2849 return X;
2850
2851 if (Instruction *R = foldFBinOpOfIntCasts(I))
2852 return R;
2853
2854 Value *X, *Y;
2855 Constant *C;
2856
2857 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2858 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2859 // Canonicalize to fadd to make analysis easier.
2860 // This can also help codegen because fadd is commutative.
2861 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2862 // killed later. We still limit that particular transform with 'hasOneUse'
2863 // because an fneg is assumed better/cheaper than a generic fsub.
2864 if (I.hasNoSignedZeros() ||
2865 cannotBeNegativeZero(Op0, 0, getSimplifyQuery().getWithInstruction(&I))) {
2866 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2867 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2868 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2869 }
2870 }
2871
2872 // (-X) - Op1 --> -(X + Op1)
2873 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2874 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2875 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2877 }
2878
2879 if (isa<Constant>(Op0))
2880 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2881 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2882 return NV;
2883
2884 // X - C --> X + (-C)
2885 // But don't transform constant expressions because there's an inverse fold
2886 // for X + (-Y) --> X - Y.
2887 if (match(Op1, m_ImmConstant(C)))
2888 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2889 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2890
2891 // X - (-Y) --> X + Y
2892 if (match(Op1, m_FNeg(m_Value(Y))))
2893 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2894
2895 // Similar to above, but look through a cast of the negated value:
2896 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2897 Type *Ty = I.getType();
2898 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2900
2901 // X - (fpext(-Y)) --> X + fpext(Y)
2902 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2904
2905 // Similar to above, but look through fmul/fdiv of the negated value:
2906 // Op0 - (-X * Y) --> Op0 + (X * Y)
2907 // Op0 - (Y * -X) --> Op0 + (X * Y)
2908 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2910 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2911 }
2912 // Op0 - (-X / Y) --> Op0 + (X / Y)
2913 // Op0 - (X / -Y) --> Op0 + (X / Y)
2914 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2915 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2916 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2917 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2918 }
2919
2920 // Handle special cases for FSub with selects feeding the operation
2921 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2922 return replaceInstUsesWith(I, V);
2923
2924 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2925 // (Y - X) - Y --> -X
2926 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2928
2929 // Y - (X + Y) --> -X
2930 // Y - (Y + X) --> -X
2931 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2933
2934 // (X * C) - X --> X * (C - 1.0)
2935 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2937 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2938 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2939 }
2940 // X - (X * C) --> X * (1.0 - C)
2941 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2943 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2944 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2945 }
2946
2947 // Reassociate fsub/fadd sequences to create more fadd instructions and
2948 // reduce dependency chains:
2949 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2950 Value *Z;
2952 m_Value(Z))))) {
2953 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2954 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2955 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2956 }
2957
2958 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2959 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2960 m_Value(Vec)));
2961 };
2962 Value *A0, *A1, *V0, *V1;
2963 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2964 V0->getType() == V1->getType()) {
2965 // Difference of sums is sum of differences:
2966 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2967 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2968 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2969 {Sub->getType()}, {A0, Sub}, &I);
2970 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2971 }
2972
2974 return F;
2975
2976 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2977 // functionality has been subsumed by simple pattern matching here and in
2978 // InstSimplify. We should let a dedicated reassociation pass handle more
2979 // complex pattern matching and remove this from InstCombine.
2980 if (Value *V = FAddCombine(Builder).simplify(&I))
2981 return replaceInstUsesWith(I, V);
2982
2983 // (X - Y) - Op1 --> X - (Y + Op1)
2984 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2985 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2987 }
2988 }
2989
2990 return nullptr;
2991}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static bool isConstant(const MachineInstr &MI)
amdgpu AMDGPU Register Bank Select
This file declares a class to represent arbitrary precision floating point values and provide a varie...
This file implements a class to represent arbitrary precision integral constant values and operations...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
hexagon bit simplify
static Instruction * factorizeFAddFSub(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Factor a common operand out of fadd/fsub of fmul/fdiv.
static Instruction * foldAddToAshr(BinaryOperator &Add)
Try to reduce signed division by power-of-2 to an arithmetic shift right.
static bool MatchMul(Value *E, Value *&Op, APInt &C)
static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned)
static Instruction * foldFNegIntoConstant(Instruction &I, const DataLayout &DL)
This eliminates floating-point negation in either 'fneg(X)' or 'fsub(-0.0, X)' form by combining into...
static Instruction * combineAddSubWithShlAddSub(InstCombiner::BuilderTy &Builder, const BinaryOperator &I)
static Instruction * factorizeLerp(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Eliminate an op from a linear interpolation (lerp) pattern.
static Instruction * foldSubOfMinMax(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static Instruction * foldBoxMultiply(BinaryOperator &I)
Reduce a sequence of masked half-width multiplies to a single multiply.
static Value * checkForNegativeOperand(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned)
static Instruction * foldNoWrapAdd(BinaryOperator &Add, InstCombiner::BuilderTy &Builder)
Wrapping flags may allow combining constants separated by an extend.
static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A, Value *&B)
static Instruction * factorizeMathWithShlOps(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
This is a specialization of a more general transform from foldUsingDistributiveLaws.
static Instruction * canonicalizeLowbitMask(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Fold (1 << NBits) - 1 Into: ~(-(1 << NBits)) Because a 'not' is better for bit-tracking analysis and ...
static Instruction * foldToUnsignedSaturatedAdd(BinaryOperator &I)
static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned)
This file provides internal interfaces used to implement the InstCombine.
This file provides the interface for the instcombine pass implementation.
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:528
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
uint64_t IntrinsicInst * II
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
#define P(N)
const SmallVectorImpl< MachineOperand > & Cond
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file contains some templates that are useful if you are working with the STL at all.
This file defines the SmallVector class.
Value * RHS
Value * LHS
const fltSemantics & getSemantics() const
Definition: APFloat.h:1356
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1104
Class for arbitrary precision integers.
Definition: APInt.h:77
APInt umul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1941
bool isNegatedPowerOf2() const
Check if this APInt's negated value is a power of two greater than zero.
Definition: APInt.h:428
bool isMinSignedValue() const
Determine if this is the smallest signed value.
Definition: APInt.h:402
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:906
static APInt getMaxValue(unsigned numBits)
Gets maximum unsigned value of APInt for specific bit width.
Definition: APInt.h:185
bool isZero() const
Determine if this value is zero, i.e. all bits are clear.
Definition: APInt.h:359
bool isSignMask() const
Check if the APInt's value is returned by getSignMask.
Definition: APInt.h:445
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1447
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:308
int32_t exactLogBase2() const
Definition: APInt.h:1740
unsigned countr_zero() const
Count the number of trailing zero bits.
Definition: APInt.h:1597
unsigned countl_zero() const
The APInt version of std::countl_zero.
Definition: APInt.h:1556
static APInt getSignedMinValue(unsigned numBits)
Gets minimum signed value of APInt for a specific bit width.
Definition: APInt.h:198
unsigned logBase2() const
Definition: APInt.h:1718
APInt smul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1930
bool isMask(unsigned numBits) const
Definition: APInt.h:467
APInt sext(unsigned width) const
Sign extend to a new width.
Definition: APInt.cpp:954
bool isSubsetOf(const APInt &RHS) const
This operation checks that all bits set in this APInt are also set in RHS.
Definition: APInt.h:1236
bool isPowerOf2() const
Check if this APInt's value is a power of two greater than zero.
Definition: APInt.h:419
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Constructs an APInt value that has the top hiBitsSet bits set.
Definition: APInt.h:275
bool sge(const APInt &RHS) const
Signed greater or equal comparison.
Definition: APInt.h:1216
static BinaryOperator * CreateFAddFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:324
static BinaryOperator * CreateNeg(Value *Op, const Twine &Name, BasicBlock::iterator InsertBefore)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static BinaryOperator * CreateFMulFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:332
static BinaryOperator * CreateNot(Value *Op, const Twine &Name, BasicBlock::iterator InsertBefore)
static BinaryOperator * CreateFDivFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:336
static BinaryOperator * CreateFSubFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:328
This class represents a function call, abstracting a target machine's calling convention.
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr, BasicBlock::iterator InsertBefore)
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name, BasicBlock::iterator InsertBefore)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass's ...
static CastInst * CreateTruncOrBitCast(Value *S, Type *Ty, const Twine &Name, BasicBlock::iterator InsertBefore)
Create a Trunc or BitCast cast instruction.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:993
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:1016
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:1020
@ ICMP_EQ
equal
Definition: InstrTypes.h:1014
@ ICMP_NE
not equal
Definition: InstrTypes.h:1015
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2574
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2567
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:269
const APFloat & getValueAPF() const
Definition: Constants.h:312
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:316
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
Definition: Constants.h:124
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:417
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:370
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
Convenience struct for specifying and reasoning about fast-math flags.
Definition: FMF.h:20
bool noSignedZeros() const
Definition: FMF.h:68
bool noInfs() const
Definition: FMF.h:67
bool isInBounds() const
Test whether this is an inbounds GEP, as defined by LangRef.html.
Definition: Operator.h:411
Value * CreateFAddFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1547
Value * CreateSRem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1410
Value * CreateBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 2 operands which is mangled on the first type.
Definition: IRBuilder.cpp:921
Value * CreateFMulFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1601
Value * CreateFSubFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1574
Value * CreateFDivFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1628
Value * CreateFPTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2101
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:2039
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:466
CallInst * CreateIntrinsic(Intrinsic::ID ID, ArrayRef< Type * > Types, ArrayRef< Value * > Args, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with Args, mangled using Types.
Definition: IRBuilder.cpp:932
Value * CreateFNegFMF(Value *V, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1740
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.cpp:1090
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2033
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2559
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:311
Value * CreateNUWAdd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1340
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNSW=false)
Definition: IRBuilder.h:1721
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1749
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:145
Value * CreateIsNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg < 0.
Definition: IRBuilder.h:2554
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1344
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1416
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="", bool IsNonNeg=false)
Definition: IRBuilder.h:2021
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1475
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1327
ConstantInt * getFalse()
Get the constant value for i1 false.
Definition: IRBuilder.h:471
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2549
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1497
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1666
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:2196
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2412
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2110
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1519
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1730
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1404
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1361
Value * CreateCopySign(Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create call to the copysign intrinsic.
Definition: IRBuilder.h:1022
Instruction * FoldOpIntoSelect(Instruction &Op, SelectInst *SI, bool FoldWithMultiUse=false)
Given an instruction with a select as one operand and a constant as the other operand,...
Instruction * foldBinOpOfSelectAndCastOfSelectCondition(BinaryOperator &I)
Tries to simplify binops of select and cast of the select condition.
Instruction * visitAdd(BinaryOperator &I)
Instruction * canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(BinaryOperator &I)
Instruction * foldBinOpIntoSelectOrPhi(BinaryOperator &I)
This is a convenience wrapper function for the above two functions.
bool SimplifyAssociativeOrCommutative(BinaryOperator &I)
Performs a few simplifications for operators which are associative or commutative.
Value * foldUsingDistributiveLaws(BinaryOperator &I)
Tries to simplify binary operations which some other binary operation distributes over.
Instruction * foldBinOpShiftWithShift(BinaryOperator &I)
Instruction * foldSquareSumInt(BinaryOperator &I)
Instruction * foldOpIntoPhi(Instruction &I, PHINode *PN)
Given a binary operator, cast instruction, or select which has a PHI node as operand #0,...
Instruction * foldSquareSumFP(BinaryOperator &I)
Instruction * visitSub(BinaryOperator &I)
Value * OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty, bool isNUW)
Optimize pointer differences into the same array into a size.
Instruction * visitFAdd(BinaryOperator &I)
Instruction * foldBinopWithPhiOperands(BinaryOperator &BO)
For a binary operator with 2 phi operands, try to hoist the binary operation before the phi.
Instruction * tryFoldInstWithCtpopWithNot(Instruction *I)
Value * SimplifyAddWithRemainder(BinaryOperator &I)
Tries to simplify add operations using the definition of remainder.
Instruction * foldAddWithConstant(BinaryOperator &Add)
Instruction * foldVectorBinop(BinaryOperator &Inst)
Canonicalize the position of binops relative to shufflevector.
Value * SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS, Value *RHS)
Instruction * visitFNeg(UnaryOperator &I)
Instruction * visitFSub(BinaryOperator &I)
SimplifyQuery SQ
Definition: InstCombiner.h:76
bool isFreeToInvert(Value *V, bool WillInvertAllUses, bool &DoesConsume)
Return true if the specified value is free to invert (apply ~ to).
Definition: InstCombiner.h:232
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:386
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
Definition: InstCombiner.h:180
const DataLayout & DL
Definition: InstCombiner.h:75
unsigned ComputeNumSignBits(const Value *Op, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:452
AssumptionCache & AC
Definition: InstCombiner.h:72
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:410
DominatorTree & DT
Definition: InstCombiner.h:74
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
Definition: InstCombiner.h:431
BuilderTy & Builder
Definition: InstCombiner.h:60
bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:447
Value * getFreelyInverted(Value *V, bool WillInvertAllUses, BuilderTy *Builder, bool &DoesConsume)
Definition: InstCombiner.h:213
const SimplifyQuery & getSimplifyQuery() const
Definition: InstCombiner.h:342
static Constant * AddOne(Constant *C)
Add one to a Constant.
Definition: InstCombiner.h:175
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nuw flag on this instruction, which must be an operator which supports this flag.
bool hasNoUnsignedWrap() const LLVM_READONLY
Determine whether the no unsigned wrap flag is set.
void copyFastMathFlags(FastMathFlags FMF)
Convenience function for transferring all fast-math flag values to this instruction,...
void setHasNoSignedZeros(bool B)
Set or clear the no-signed-zeros flag on this instruction, which must be an operator which supports t...
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag.
void setFastMathFlags(FastMathFlags FMF)
Convenience function for setting multiple fast-math flags on this instruction, which must be an opera...
void setHasNoInfs(bool B)
Set or clear the no-infs flag on this instruction, which must be an operator which supports this flag...
FastMathFlags getFastMathFlags() const LLVM_READONLY
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:451
void copyMetadata(const Instruction &SrcInst, ArrayRef< unsigned > WL=ArrayRef< unsigned >())
Copy metadata from SrcInst to this instruction.
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:48
static Value * Negate(bool LHSIsZero, bool IsNSW, Value *Root, InstCombinerImpl &IC)
Attempt to negate Root.
Utility class for integer operators which may exhibit overflow - Add, Sub, Mul, and Shl.
Definition: Operator.h:77
bool hasNoSignedWrap() const
Test whether this operation is known to never undergo signed overflow, aka the nsw property.
Definition: Operator.h:110
bool hasNoUnsignedWrap() const
Test whether this operation is known to never undergo unsigned overflow, aka the nuw property.
Definition: Operator.h:104
This class represents the LLVM 'select' instruction.
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr, BasicBlock::iterator InsertBefore, Instruction *MDFrom=nullptr)
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1209
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:234
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
static UnaryOperator * CreateFNegFMF(Value *Op, Instruction *FMFSource, const Twine &Name, BasicBlock::iterator InsertBefore)
Definition: InstrTypes.h:191
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
bool hasOneUse() const
Return true if there is exactly one use of this value.
Definition: Value.h:434
bool hasNUsesOrMore(unsigned N) const
Return true if this value has N uses or more.
Definition: Value.cpp:153
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs and address space casts.
Definition: Value.cpp:693
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
This class represents zero extension of integer types.
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
Function * getDeclaration(Module *M, ID id, ArrayRef< Type * > Tys=std::nullopt)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1474
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:524
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:100
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.
BinaryOp_match< LHS, RHS, Instruction::AShr > m_AShr(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::FSub > m_FSub(const LHS &L, const RHS &R)
cst_pred_ty< is_power2 > m_Power2()
Match an integer or vector power-of-2.
Definition: PatternMatch.h:619
BinaryOp_match< LHS, RHS, Instruction::URem > m_URem(const LHS &L, const RHS &R)
match_combine_or< CastInst_match< OpTy, TruncInst >, OpTy > m_TruncOrSelf(const OpTy &Op)
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:165
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.
CastInst_match< OpTy, TruncInst > m_Trunc(const OpTy &Op)
Matches Trunc.
BinaryOp_match< LHS, RHS, Instruction::Xor > m_Xor(const LHS &L, const RHS &R)
OverflowingBinaryOp_match< LHS, RHS, Instruction::Sub, OverflowingBinaryOperator::NoSignedWrap > m_NSWSub(const LHS &L, const RHS &R)
specific_intval< false > m_SpecificInt(const APInt &V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:972
BinaryOp_match< LHS, RHS, Instruction::FMul > m_FMul(const LHS &L, const RHS &R)
match_combine_or< CastInst_match< OpTy, ZExtInst >, OpTy > m_ZExtOrSelf(const OpTy &Op)
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
bind_ty< Instruction > m_Instruction(Instruction *&I)
Match an instruction, capturing it if we match.
Definition: PatternMatch.h:816
cstfp_pred_ty< is_any_zero_fp > m_AnyZeroFP()
Match a floating-point negative zero or positive zero.
Definition: PatternMatch.h:764
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:875
DisjointOr_match< LHS, RHS > m_DisjointOr(const LHS &L, const RHS &R)
specific_intval< true > m_SpecificIntAllowPoison(const APInt &V)
Definition: PatternMatch.h:980
cst_pred_ty< is_one > m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:592
ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)
Matches SelectInst.
match_combine_or< CastInst_match< OpTy, SExtInst >, OpTy > m_SExtOrSelf(const OpTy &Op)
specific_fpval m_SpecificFP(double V)
Match a specific floating point value or vector with all elements equal to the value.
Definition: PatternMatch.h:918
match_combine_and< LTy, RTy > m_CombineAnd(const LTy &L, const RTy &R)
Combine two pattern matchers matching L && R.
Definition: PatternMatch.h:245
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.
BinaryOp_match< LHS, RHS, Instruction::FAdd > m_FAdd(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::Mul > m_Mul(const LHS &L, const RHS &R)
deferredval_ty< Value > m_Deferred(Value *const &V)
Like m_Specific(), but works if the specific value to match is determined as part of the same match()...
Definition: PatternMatch.h:893
cst_pred_ty< is_zero_int > m_ZeroInt()
Match an integer 0 or a vector with all elements equal to 0.
Definition: PatternMatch.h:599
CmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R)
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:67
MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty, true > m_c_SMin(const LHS &L, const RHS &R)
Matches an SMin with LHS and RHS in either order.
BinaryOp_match< cst_pred_ty< is_zero_int >, ValTy, Instruction::Sub > m_Neg(const ValTy &V)
Matches a 'Neg' as 'sub 0, V'.
match_combine_and< class_match< Constant >, match_unless< constantexpr_match > > m_ImmConstant()
Match an arbitrary immediate Constant and ignore it.
Definition: PatternMatch.h:854
MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty, true > m_c_UMax(const LHS &L, const RHS &R)
Matches a UMax with LHS and RHS in either order.
CastInst_match< OpTy, FPExtInst > m_FPExt(const OpTy &Op)
CastInst_match< OpTy, ZExtInst > m_ZExt(const OpTy &Op)
Matches ZExt.
BinaryOp_match< LHS, RHS, Instruction::UDiv > m_UDiv(const LHS &L, const RHS &R)
cst_pred_ty< is_negated_power2 > m_NegatedPower2()
Match a integer or vector negated power-of-2.
Definition: PatternMatch.h:627
specific_fpval m_FPOne()
Match a float 1.0 or vector with all elements equal to 1.0.
Definition: PatternMatch.h:921
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty, true > m_c_UMin(const LHS &L, const RHS &R)
Matches a UMin with LHS and RHS in either order.
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.
MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty, true > m_c_SMax(const LHS &L, const RHS &R)
Matches an SMax with LHS and RHS in either order.
match_combine_or< match_combine_or< MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty, true >, MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty, true > >, match_combine_or< MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty, true >, MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty, true > > > m_c_MaxOrMin(const LHS &L, const RHS &R)
match_combine_or< CastInst_match< OpTy, SExtInst >, NNegZExt_match< OpTy > > m_SExtLike(const OpTy &Op)
Match either "sext" or "zext nneg".
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
OverflowingBinaryOp_match< LHS, RHS, Instruction::Sub, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWSub(const LHS &L, const RHS &R)
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt.
Definition: PatternMatch.h:299
match_combine_or< OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoSignedWrap >, DisjointOr_match< LHS, RHS > > m_NSWAddLike(const LHS &L, const RHS &R)
Match either "add nsw" or "or disjoint".
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:92
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
match_combine_or< CastInst_match< OpTy, ZExtInst >, CastInst_match< OpTy, SExtInst > > m_ZExtOrSExt(const OpTy &Op)
FNeg_match< OpTy > m_FNeg(const OpTy &X)
Match 'fneg X' as 'fsub -0.0, X'.
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.
BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::FDiv > m_FDiv(const LHS &L, const RHS &R)
BinOpPred_match< LHS, RHS, is_irem_op > m_IRem(const LHS &L, const RHS &R)
Matches integer remainder operations.
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:316
CastInst_match< OpTy, FPTruncInst > m_FPTrunc(const OpTy &Op)
BinaryOp_match< LHS, RHS, Instruction::SRem > m_SRem(const LHS &L, const RHS &R)
BinaryOp_match< cst_pred_ty< is_all_ones >, ValTy, Instruction::Xor, true > m_Not(const ValTy &V)
Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
CastInst_match< OpTy, SExtInst > m_SExt(const OpTy &Op)
Matches SExt.
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:612
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.
match_combine_or< OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap >, DisjointOr_match< LHS, RHS > > m_NUWAddLike(const LHS &L, const RHS &R)
Match either "add nuw" or "or disjoint".
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.
m_Intrinsic_Ty< Opnd0, Opnd1 >::Ty m_CopySign(const Opnd0 &Op0, const Opnd1 &Op1)
CastOperator_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty > m_UMin(const LHS &L, const RHS &R)
match_combine_or< LTy, RTy > m_CombineOr(const LTy &L, const RTy &R)
Combine two pattern matchers matching L || R.
Definition: PatternMatch.h:239
cst_pred_ty< icmp_pred_with_threshold > m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold)
Match an integer or vector with every element comparing 'pred' (eg/ne/...) to Threshold.
Definition: PatternMatch.h:698
@ CE
Windows NT (Windows on ARM)
NodeAddr< InstrNode * > Instr
Definition: RDFGraph.h:389
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
bool haveNoCommonBitsSet(const WithCache< const Value * > &LHSCache, const WithCache< const Value * > &RHSCache, const SimplifyQuery &SQ)
Return true if LHS and RHS have no common bits set.
Intrinsic::ID getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID)
@ Offset
Definition: DWP.cpp:456
constexpr bool isInt(int64_t x)
Checks if an integer fits into the given bit width.
Definition: MathExtras.h:153
bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS, bool &TrueIfSigned)
Given an exploded icmp instruction, return true if the comparison only checks the sign bit.
LLVM_ATTRIBUTE_ALWAYS_INLINE DynamicAPInt & operator+=(DynamicAPInt &A, int64_t B)
Definition: DynamicAPInt.h:511
bool isGuaranteedNotToBeUndef(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Returns true if V cannot be undef, but may be poison.
Value * simplifySubInst(Value *LHS, Value *RHS, bool IsNSW, bool IsNUW, const SimplifyQuery &Q)
Given operands for a Sub, fold the result or return null.
Value * simplifyAddInst(Value *LHS, Value *RHS, bool IsNSW, bool IsNUW, const SimplifyQuery &Q)
Given operands for an Add, fold the result or return null.
LLVM_ATTRIBUTE_ALWAYS_INLINE DynamicAPInt & operator*=(DynamicAPInt &A, int64_t B)
Definition: DynamicAPInt.h:519
Constant * ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op, const DataLayout &DL)
Attempt to constant fold a unary operation with the specified operand.
Value * simplifyFNegInst(Value *Op, FastMathFlags FMF, const SimplifyQuery &Q)
Given operand for an FNeg, fold the result or return null.
Value * simplifyFSubInst(Value *LHS, Value *RHS, FastMathFlags FMF, const SimplifyQuery &Q, fp::ExceptionBehavior ExBehavior=fp::ebIgnore, RoundingMode Rounding=RoundingMode::NearestTiesToEven)
Given operands for an FSub, fold the result or return null.
decltype(auto) get(const PointerIntPair< PointerTy, IntBits, IntType, PtrTraits, Info > &Pair)
Value * simplifyFAddInst(Value *LHS, Value *RHS, FastMathFlags FMF, const SimplifyQuery &Q, fp::ExceptionBehavior ExBehavior=fp::ebIgnore, RoundingMode Rounding=RoundingMode::NearestTiesToEven)
Given operands for an FAdd, fold the result or return null.
bool none_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::none_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1736
Constant * ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, Constant *RHS, const DataLayout &DL)
Attempt to constant fold a binary operation with the specified operands.
bool isKnownNonZero(const Value *V, const SimplifyQuery &Q, unsigned Depth=0)
Return true if the given value is known to be non-zero when defined.
@ Mul
Product of integers.
@ FMul
Product of floats.
@ And
Bitwise or logical AND of integers.
@ SMin
Signed integer min implemented in terms of select(cmp()).
@ Add
Sum of integers.
@ FAdd
Sum of floats.
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
DWARFExpression::Operation Op
RoundingMode
Rounding mode.
bool isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Return true if this function can prove that V does not have undef bits and is never poison.
constexpr unsigned BitWidth
Definition: BitmaskEnum.h:191
bool cannotBeNegativeZero(const Value *V, unsigned Depth, const SimplifyQuery &SQ)
Return true if we can prove that the specified FP value is never equal to -0.0.
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:860
A suitably aligned and sized character array member which can hold elements of any type.
Definition: AlignOf.h:27
SimplifyQuery getWithInstruction(const Instruction *I) const
Definition: SimplifyQuery.h:96
SimplifyQuery getWithoutDomCondCache() const