LLVM 20.0.0git
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)) &&
822 match(Op0, m_ZExt(m_NUWAddLike(m_Value(X), m_APInt(C2)))) &&
823 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824 APInt NewC = *C2 + C1->trunc(C2->getBitWidth());
825 // If the smaller add will fold to zero, we don't need to check one use.
826 if (NewC.isZero())
827 return new ZExtInst(X, Ty);
828 // Otherwise only do this if the existing zero extend will be removed.
829 if (Op0->hasOneUse())
830 return new ZExtInst(
831 Builder.CreateNUWAdd(X, ConstantInt::get(X->getType(), NewC)), Ty);
832 }
833
834 // More general combining of constants in the wide type.
835 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
836 // or (zext nneg (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
837 Constant *NarrowC;
838 if (match(Op0, m_OneUse(m_SExtLike(
839 m_NSWAddLike(m_Value(X), m_Constant(NarrowC)))))) {
840 Value *WideC = Builder.CreateSExt(NarrowC, Ty);
841 Value *NewC = Builder.CreateAdd(WideC, Op1C);
842 Value *WideX = Builder.CreateSExt(X, Ty);
843 return BinaryOperator::CreateAdd(WideX, NewC);
844 }
845 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
846 if (match(Op0,
848 Value *WideC = Builder.CreateZExt(NarrowC, Ty);
849 Value *NewC = Builder.CreateAdd(WideC, Op1C);
850 Value *WideX = Builder.CreateZExt(X, Ty);
851 return BinaryOperator::CreateAdd(WideX, NewC);
852 }
853 return nullptr;
854}
855
857 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
858 Type *Ty = Add.getType();
859 Constant *Op1C;
860 if (!match(Op1, m_ImmConstant(Op1C)))
861 return nullptr;
862
864 return NV;
865
866 Value *X;
867 Constant *Op00C;
868
869 // add (sub C1, X), C2 --> sub (add C1, C2), X
870 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
871 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
872
873 Value *Y;
874
875 // add (sub X, Y), -1 --> add (not Y), X
876 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
877 match(Op1, m_AllOnes()))
878 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
879
880 // zext(bool) + C -> bool ? C + 1 : C
881 if (match(Op0, m_ZExt(m_Value(X))) &&
882 X->getType()->getScalarSizeInBits() == 1)
883 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
884 // sext(bool) + C -> bool ? C - 1 : C
885 if (match(Op0, m_SExt(m_Value(X))) &&
886 X->getType()->getScalarSizeInBits() == 1)
887 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
888
889 // ~X + C --> (C-1) - X
890 if (match(Op0, m_Not(m_Value(X)))) {
891 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
892 auto *COne = ConstantInt::get(Op1C->getType(), 1);
893 bool WillNotSOV = willNotOverflowSignedSub(Op1C, COne, Add);
894 BinaryOperator *Res =
895 BinaryOperator::CreateSub(ConstantExpr::getSub(Op1C, COne), X);
896 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
897 return Res;
898 }
899
900 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
901 const APInt *C;
902 unsigned BitWidth = Ty->getScalarSizeInBits();
903 if (match(Op0, m_OneUse(m_AShr(m_Value(X),
905 match(Op1, m_One()))
906 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
907
908 if (!match(Op1, m_APInt(C)))
909 return nullptr;
910
911 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
912 Constant *Op01C;
913 if (match(Op0, m_DisjointOr(m_Value(X), m_ImmConstant(Op01C)))) {
914 BinaryOperator *NewAdd =
915 BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
916 NewAdd->setHasNoSignedWrap(Add.hasNoSignedWrap() &&
917 willNotOverflowSignedAdd(Op01C, Op1C, Add));
918 NewAdd->setHasNoUnsignedWrap(Add.hasNoUnsignedWrap());
919 return NewAdd;
920 }
921
922 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
923 const APInt *C2;
924 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
925 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
926
927 if (C->isSignMask()) {
928 // If wrapping is not allowed, then the addition must set the sign bit:
929 // X + (signmask) --> X | signmask
930 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
931 return BinaryOperator::CreateOr(Op0, Op1);
932
933 // If wrapping is allowed, then the addition flips the sign bit of LHS:
934 // X + (signmask) --> X ^ signmask
935 return BinaryOperator::CreateXor(Op0, Op1);
936 }
937
938 // Is this add the last step in a convoluted sext?
939 // add(zext(xor i16 X, -32768), -32768) --> sext X
940 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
941 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
942 return CastInst::Create(Instruction::SExt, X, Ty);
943
944 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
945 // (X ^ signmask) + C --> (X + (signmask ^ C))
946 if (C2->isSignMask())
947 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
948
949 // If X has no high-bits set above an xor mask:
950 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
951 if (C2->isMask()) {
952 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
953 if ((*C2 | LHSKnown.Zero).isAllOnes())
954 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
955 }
956
957 // Look for a math+logic pattern that corresponds to sext-in-register of a
958 // value with cleared high bits. Convert that into a pair of shifts:
959 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
960 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
961 if (Op0->hasOneUse() && *C2 == -(*C)) {
962 unsigned BitWidth = Ty->getScalarSizeInBits();
963 unsigned ShAmt = 0;
964 if (C->isPowerOf2())
965 ShAmt = BitWidth - C->logBase2() - 1;
966 else if (C2->isPowerOf2())
967 ShAmt = BitWidth - C2->logBase2() - 1;
969 0, &Add)) {
970 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
971 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
972 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
973 }
974 }
975 }
976
977 if (C->isOne() && Op0->hasOneUse()) {
978 // add (sext i1 X), 1 --> zext (not X)
979 // TODO: The smallest IR representation is (select X, 0, 1), and that would
980 // not require the one-use check. But we need to remove a transform in
981 // visitSelect and make sure that IR value tracking for select is equal or
982 // better than for these ops.
983 if (match(Op0, m_SExt(m_Value(X))) &&
984 X->getType()->getScalarSizeInBits() == 1)
985 return new ZExtInst(Builder.CreateNot(X), Ty);
986
987 // Shifts and add used to flip and mask off the low bit:
988 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
989 const APInt *C3;
990 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
991 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
992 Value *NotX = Builder.CreateNot(X);
993 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
994 }
995 }
996
997 // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
998 // TODO: There's a general form for any constant on the outer add.
999 if (C->isOne()) {
1000 if (match(Op0, m_ZExt(m_Add(m_Value(X), m_AllOnes())))) {
1002 if (llvm::isKnownNonZero(X, Q))
1003 return new ZExtInst(X, Ty);
1004 }
1005 }
1006
1007 return nullptr;
1008}
1009
1010// match variations of a^2 + 2*a*b + b^2
1011//
1012// to reuse the code between the FP and Int versions, the instruction OpCodes
1013// and constant types have been turned into template parameters.
1014//
1015// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1016// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1017// (we're matching `X<<1` instead of `X*2` for Int)
1018template <bool FP, typename Mul2Rhs>
1019static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1020 Value *&B) {
1021 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1022 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1023 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1024
1025 // (a * a) + (((a * 2) + b) * b)
1026 if (match(&I, m_c_BinOp(
1027 AddOp, m_OneUse(m_BinOp(MulOp, m_Value(A), m_Deferred(A))),
1029 MulOp,
1030 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(A), M2Rhs),
1031 m_Value(B)),
1032 m_Deferred(B))))))
1033 return true;
1034
1035 // ((a * b) * 2) or ((a * 2) * b)
1036 // +
1037 // (a * a + b * b) or (b * b + a * a)
1038 return match(
1039 &I, m_c_BinOp(
1040 AddOp,
1043 Mul2Op, m_BinOp(MulOp, m_Value(A), m_Value(B)), M2Rhs)),
1044 m_OneUse(m_c_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(A), M2Rhs),
1045 m_Value(B)))),
1046 m_OneUse(
1047 m_c_BinOp(AddOp, m_BinOp(MulOp, m_Deferred(A), m_Deferred(A)),
1048 m_BinOp(MulOp, m_Deferred(B), m_Deferred(B))))));
1049}
1050
1051// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1053 Value *A, *B;
1054 if (matchesSquareSum</*FP*/ false>(I, m_SpecificInt(1), A, B)) {
1055 Value *AB = Builder.CreateAdd(A, B);
1056 return BinaryOperator::CreateMul(AB, AB);
1057 }
1058 return nullptr;
1059}
1060
1061// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1062// Requires `nsz` and `reassoc`.
1064 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1065 Value *A, *B;
1066 if (matchesSquareSum</*FP*/ true>(I, m_SpecificFP(2.0), A, B)) {
1067 Value *AB = Builder.CreateFAddFMF(A, B, &I);
1068 return BinaryOperator::CreateFMulFMF(AB, AB, &I);
1069 }
1070 return nullptr;
1071}
1072
1073// Matches multiplication expression Op * C where C is a constant. Returns the
1074// constant value in C and the other operand in Op. Returns true if such a
1075// match is found.
1076static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1077 const APInt *AI;
1078 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1079 C = *AI;
1080 return true;
1081 }
1082 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1083 C = APInt(AI->getBitWidth(), 1);
1084 C <<= *AI;
1085 return true;
1086 }
1087 return false;
1088}
1089
1090// Matches remainder expression Op % C where C is a constant. Returns the
1091// constant value in C and the other operand in Op. Returns the signedness of
1092// the remainder operation in IsSigned. Returns true if such a match is
1093// found.
1094static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1095 const APInt *AI;
1096 IsSigned = false;
1097 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1098 IsSigned = true;
1099 C = *AI;
1100 return true;
1101 }
1102 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1103 C = *AI;
1104 return true;
1105 }
1106 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1107 C = *AI + 1;
1108 return true;
1109 }
1110 return false;
1111}
1112
1113// Matches division expression Op / C with the given signedness as indicated
1114// by IsSigned, where C is a constant. Returns the constant value in C and the
1115// other operand in Op. Returns true if such a match is found.
1116static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1117 const APInt *AI;
1118 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1119 C = *AI;
1120 return true;
1121 }
1122 if (!IsSigned) {
1123 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1124 C = *AI;
1125 return true;
1126 }
1127 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1128 C = APInt(AI->getBitWidth(), 1);
1129 C <<= *AI;
1130 return true;
1131 }
1132 }
1133 return false;
1134}
1135
1136// Returns whether C0 * C1 with the given signedness overflows.
1137static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1138 bool overflow;
1139 if (IsSigned)
1140 (void)C0.smul_ov(C1, overflow);
1141 else
1142 (void)C0.umul_ov(C1, overflow);
1143 return overflow;
1144}
1145
1146// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1147// does not overflow.
1148// Simplifies (X / C0) * C1 + (X % C0) * C2 to
1149// (X / C0) * (C1 - C2 * C0) + X * C2
1151 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1152 Value *X, *MulOpV;
1153 APInt C0, MulOpC;
1154 bool IsSigned;
1155 // Match I = X % C0 + MulOpV * C0
1156 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1157 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1158 C0 == MulOpC) {
1159 Value *RemOpV;
1160 APInt C1;
1161 bool Rem2IsSigned;
1162 // Match MulOpC = RemOpV % C1
1163 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1164 IsSigned == Rem2IsSigned) {
1165 Value *DivOpV;
1166 APInt DivOpC;
1167 // Match RemOpV = X / C0
1168 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1169 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1170 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1171 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1172 : Builder.CreateURem(X, NewDivisor, "urem");
1173 }
1174 }
1175 }
1176
1177 // Match I = (X / C0) * C1 + (X % C0) * C2
1178 Value *Div, *Rem;
1179 APInt C1, C2;
1180 if (!LHS->hasOneUse() || !MatchMul(LHS, Div, C1))
1181 Div = LHS, C1 = APInt(I.getType()->getScalarSizeInBits(), 1);
1182 if (!RHS->hasOneUse() || !MatchMul(RHS, Rem, C2))
1183 Rem = RHS, C2 = APInt(I.getType()->getScalarSizeInBits(), 1);
1184 if (match(Div, m_IRem(m_Value(), m_Value()))) {
1185 std::swap(Div, Rem);
1186 std::swap(C1, C2);
1187 }
1188 Value *DivOpV;
1189 APInt DivOpC;
1190 if (MatchRem(Rem, X, C0, IsSigned) &&
1191 MatchDiv(Div, DivOpV, DivOpC, IsSigned) && X == DivOpV && C0 == DivOpC) {
1192 APInt NewC = C1 - C2 * C0;
1193 if (!NewC.isZero() && !Rem->hasOneUse())
1194 return nullptr;
1195 if (!isGuaranteedNotToBeUndef(X, &AC, &I, &DT))
1196 return nullptr;
1197 Value *MulXC2 = Builder.CreateMul(X, ConstantInt::get(X->getType(), C2));
1198 if (NewC.isZero())
1199 return MulXC2;
1200 return Builder.CreateAdd(
1201 Builder.CreateMul(Div, ConstantInt::get(X->getType(), NewC)), MulXC2);
1202 }
1203
1204 return nullptr;
1205}
1206
1207/// Fold
1208/// (1 << NBits) - 1
1209/// Into:
1210/// ~(-(1 << NBits))
1211/// Because a 'not' is better for bit-tracking analysis and other transforms
1212/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1214 InstCombiner::BuilderTy &Builder) {
1215 Value *NBits;
1216 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1217 return nullptr;
1218
1219 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1220 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1221 // Be wary of constant folding.
1222 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1223 // Always NSW. But NUW propagates from `add`.
1224 BOp->setHasNoSignedWrap();
1225 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1226 }
1227
1228 return BinaryOperator::CreateNot(NotMask, I.getName());
1229}
1230
1232 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1233 Type *Ty = I.getType();
1234 auto getUAddSat = [&]() {
1235 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1236 };
1237
1238 // add (umin X, ~Y), Y --> uaddsat X, Y
1239 Value *X, *Y;
1241 m_Deferred(Y))))
1242 return CallInst::Create(getUAddSat(), { X, Y });
1243
1244 // add (umin X, ~C), C --> uaddsat X, C
1245 const APInt *C, *NotC;
1246 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1247 *C == ~*NotC)
1248 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1249
1250 return nullptr;
1251}
1252
1253// Transform:
1254// (add A, (shl (neg B), Y))
1255// -> (sub A, (shl B, Y))
1257 const BinaryOperator &I) {
1258 Value *A, *B, *Cnt;
1259 if (match(&I,
1261 m_Value(A)))) {
1262 Value *NewShl = Builder.CreateShl(B, Cnt);
1263 return BinaryOperator::CreateSub(A, NewShl);
1264 }
1265 return nullptr;
1266}
1267
1268/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1270 // Division must be by power-of-2, but not the minimum signed value.
1271 Value *X;
1272 const APInt *DivC;
1273 if (!match(Add.getOperand(0), m_SDiv(m_Value(X), m_Power2(DivC))) ||
1274 DivC->isNegative())
1275 return nullptr;
1276
1277 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1278 // low bits set. It recognizes two canonical patterns:
1279 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1280 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1281 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1282 // Note that, by the time we end up here, if possible, ugt has been
1283 // canonicalized into eq.
1284 const APInt *MaskC, *MaskCCmp;
1286 if (!match(Add.getOperand(1),
1287 m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1288 m_APInt(MaskCCmp)))))
1289 return nullptr;
1290
1291 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1292 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1293 return nullptr;
1294
1295 APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1296 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1297 ? (*MaskC == (SMin | (*DivC - 1)))
1298 : (*DivC == 2 && *MaskC == SMin + 1);
1299 if (!IsMaskValid)
1300 return nullptr;
1301
1302 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1303 return BinaryOperator::CreateAShr(
1304 X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1305}
1306
1309 BinaryOperator &I) {
1310 assert((I.getOpcode() == Instruction::Add ||
1311 I.getOpcode() == Instruction::Or ||
1312 I.getOpcode() == Instruction::Sub) &&
1313 "Expecting add/or/sub instruction");
1314
1315 // We have a subtraction/addition between a (potentially truncated) *logical*
1316 // right-shift of X and a "select".
1317 Value *X, *Select;
1318 Instruction *LowBitsToSkip, *Extract;
1320 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1321 m_Instruction(Extract))),
1322 m_Value(Select))))
1323 return nullptr;
1324
1325 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1326 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1327 return nullptr;
1328
1329 Type *XTy = X->getType();
1330 bool HadTrunc = I.getType() != XTy;
1331
1332 // If there was a truncation of extracted value, then we'll need to produce
1333 // one extra instruction, so we need to ensure one instruction will go away.
1334 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1335 return nullptr;
1336
1337 // Extraction should extract high NBits bits, with shift amount calculated as:
1338 // low bits to skip = shift bitwidth - high bits to extract
1339 // The shift amount itself may be extended, and we need to look past zero-ext
1340 // when matching NBits, that will matter for matching later.
1341 Constant *C;
1342 Value *NBits;
1343 if (!match(
1344 LowBitsToSkip,
1347 APInt(C->getType()->getScalarSizeInBits(),
1348 X->getType()->getScalarSizeInBits()))))
1349 return nullptr;
1350
1351 // Sign-extending value can be zero-extended if we `sub`tract it,
1352 // or sign-extended otherwise.
1353 auto SkipExtInMagic = [&I](Value *&V) {
1354 if (I.getOpcode() == Instruction::Sub)
1355 match(V, m_ZExtOrSelf(m_Value(V)));
1356 else
1357 match(V, m_SExtOrSelf(m_Value(V)));
1358 };
1359
1360 // Now, finally validate the sign-extending magic.
1361 // `select` itself may be appropriately extended, look past that.
1362 SkipExtInMagic(Select);
1363
1365 const APInt *Thr;
1366 Value *SignExtendingValue, *Zero;
1367 bool ShouldSignext;
1368 // It must be a select between two values we will later establish to be a
1369 // sign-extending value and a zero constant. The condition guarding the
1370 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1371 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1372 m_Value(SignExtendingValue), m_Value(Zero))) ||
1373 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1374 return nullptr;
1375
1376 // icmp-select pair is commutative.
1377 if (!ShouldSignext)
1378 std::swap(SignExtendingValue, Zero);
1379
1380 // If we should not perform sign-extension then we must add/or/subtract zero.
1381 if (!match(Zero, m_Zero()))
1382 return nullptr;
1383 // Otherwise, it should be some constant, left-shifted by the same NBits we
1384 // had in `lshr`. Said left-shift can also be appropriately extended.
1385 // Again, we must look past zero-ext when looking for NBits.
1386 SkipExtInMagic(SignExtendingValue);
1387 Constant *SignExtendingValueBaseConstant;
1388 if (!match(SignExtendingValue,
1389 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1390 m_ZExtOrSelf(m_Specific(NBits)))))
1391 return nullptr;
1392 // If we `sub`, then the constant should be one, else it should be all-ones.
1393 if (I.getOpcode() == Instruction::Sub
1394 ? !match(SignExtendingValueBaseConstant, m_One())
1395 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1396 return nullptr;
1397
1398 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1399 Extract->getName() + ".sext");
1400 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1401 if (!HadTrunc)
1402 return NewAShr;
1403
1404 Builder.Insert(NewAShr);
1405 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1406}
1407
1408/// This is a specialization of a more general transform from
1409/// foldUsingDistributiveLaws. If that code can be made to work optimally
1410/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1412 InstCombiner::BuilderTy &Builder) {
1413 // TODO: Also handle mul by doubling the shift amount?
1414 assert((I.getOpcode() == Instruction::Add ||
1415 I.getOpcode() == Instruction::Sub) &&
1416 "Expected add/sub");
1417 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1418 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1419 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1420 return nullptr;
1421
1422 Value *X, *Y, *ShAmt;
1423 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1424 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1425 return nullptr;
1426
1427 // No-wrap propagates only when all ops have no-wrap.
1428 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1429 Op1->hasNoSignedWrap();
1430 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1431 Op1->hasNoUnsignedWrap();
1432
1433 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1434 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1435 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1436 NewI->setHasNoSignedWrap(HasNSW);
1437 NewI->setHasNoUnsignedWrap(HasNUW);
1438 }
1439 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1440 NewShl->setHasNoSignedWrap(HasNSW);
1441 NewShl->setHasNoUnsignedWrap(HasNUW);
1442 return NewShl;
1443}
1444
1445/// Reduce a sequence of masked half-width multiplies to a single multiply.
1446/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1448 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1449 // Skip the odd bitwidth types.
1450 if ((BitWidth & 0x1))
1451 return nullptr;
1452
1453 unsigned HalfBits = BitWidth >> 1;
1454 APInt HalfMask = APInt::getMaxValue(HalfBits);
1455
1456 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1457 Value *XLo, *YLo;
1458 Value *CrossSum;
1459 // Require one-use on the multiply to avoid increasing the number of
1460 // multiplications.
1461 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1462 m_OneUse(m_Mul(m_Value(YLo), m_Value(XLo))))))
1463 return nullptr;
1464
1465 // XLo = X & HalfMask
1466 // YLo = Y & HalfMask
1467 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1468 // to enhance robustness
1469 Value *X, *Y;
1470 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1471 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1472 return nullptr;
1473
1474 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1475 // X' can be either X or XLo in the pattern (and the same for Y')
1476 if (match(CrossSum,
1481 return BinaryOperator::CreateMul(X, Y);
1482
1483 return nullptr;
1484}
1485
1487 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1488 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1490 return replaceInstUsesWith(I, V);
1491
1493 return &I;
1494
1496 return X;
1497
1499 return Phi;
1500
1501 // (A*B)+(A*C) -> A*(B+C) etc
1503 return replaceInstUsesWith(I, V);
1504
1505 if (Instruction *R = foldBoxMultiply(I))
1506 return R;
1507
1509 return R;
1510
1512 return X;
1513
1515 return X;
1516
1518 return R;
1519
1521 return R;
1522
1523 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1524 Type *Ty = I.getType();
1525 if (Ty->isIntOrIntVectorTy(1))
1526 return BinaryOperator::CreateXor(LHS, RHS);
1527
1528 // X + X --> X << 1
1529 if (LHS == RHS) {
1530 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1531 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1532 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1533 return Shl;
1534 }
1535
1536 Value *A, *B;
1537 if (match(LHS, m_Neg(m_Value(A)))) {
1538 // -A + -B --> -(A + B)
1539 if (match(RHS, m_Neg(m_Value(B))))
1541
1542 // -A + B --> B - A
1543 auto *Sub = BinaryOperator::CreateSub(RHS, A);
1544 auto *OB0 = cast<OverflowingBinaryOperator>(LHS);
1545 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1546
1547 return Sub;
1548 }
1549
1550 // A + -B --> A - B
1551 if (match(RHS, m_Neg(m_Value(B)))) {
1552 auto *Sub = BinaryOperator::CreateSub(LHS, B);
1553 auto *OBO = cast<OverflowingBinaryOperator>(RHS);
1554 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO->hasNoSignedWrap());
1555 return Sub;
1556 }
1557
1559 return replaceInstUsesWith(I, V);
1560
1561 // (A + 1) + ~B --> A - B
1562 // ~B + (A + 1) --> A - B
1563 // (~B + A) + 1 --> A - B
1564 // (A + ~B) + 1 --> A - B
1565 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1567 return BinaryOperator::CreateSub(A, B);
1568
1569 // (A + RHS) + RHS --> A + (RHS << 1)
1571 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1572
1573 // LHS + (A + LHS) --> A + (LHS << 1)
1575 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1576
1577 {
1578 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1579 Constant *C1, *C2;
1580 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1581 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1582 (LHS->hasOneUse() || RHS->hasOneUse())) {
1583 Value *Sub = Builder.CreateSub(A, B);
1584 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1585 }
1586
1587 // Canonicalize a constant sub operand as an add operand for better folding:
1588 // (C1 - A) + B --> (B - A) + C1
1590 m_Value(B)))) {
1591 Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1592 return BinaryOperator::CreateAdd(Sub, C1);
1593 }
1594 }
1595
1596 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1598
1599 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1600 const APInt *C1, *C2;
1601 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1602 APInt one(C2->getBitWidth(), 1);
1603 APInt minusC1 = -(*C1);
1604 if (minusC1 == (one << *C2)) {
1605 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1606 return BinaryOperator::CreateSRem(RHS, NewRHS);
1607 }
1608 }
1609
1610 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1611 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1612 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1613 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1614 return BinaryOperator::CreateAnd(A, NewMask);
1615 }
1616
1617 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1618 if ((match(RHS, m_ZExt(m_Value(A))) &&
1620 (match(LHS, m_ZExt(m_Value(A))) &&
1622 return new ZExtInst(B, LHS->getType());
1623
1624 // zext(A) + sext(A) --> 0 if A is i1
1626 A->getType()->isIntOrIntVectorTy(1))
1627 return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1628
1629 // A+B --> A|B iff A and B have no bits set in common.
1630 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1631 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(&I)))
1632 return BinaryOperator::CreateDisjointOr(LHS, RHS);
1633
1634 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1635 return Ext;
1636
1637 // (add (xor A, B) (and A, B)) --> (or A, B)
1638 // (add (and A, B) (xor A, B)) --> (or A, B)
1639 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1641 return BinaryOperator::CreateOr(A, B);
1642
1643 // (add (or A, B) (and A, B)) --> (add A, B)
1644 // (add (and A, B) (or A, B)) --> (add A, B)
1645 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1647 // Replacing operands in-place to preserve nuw/nsw flags.
1648 replaceOperand(I, 0, A);
1649 replaceOperand(I, 1, B);
1650 return &I;
1651 }
1652
1653 // (add A (or A, -A)) --> (and (add A, -1) A)
1654 // (add A (or -A, A)) --> (and (add A, -1) A)
1655 // (add (or A, -A) A) --> (and (add A, -1) A)
1656 // (add (or -A, A) A) --> (and (add A, -1) A)
1658 m_Deferred(A)))))) {
1659 Value *Add =
1661 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1662 return BinaryOperator::CreateAnd(Add, A);
1663 }
1664
1665 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1666 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1667 if (match(&I,
1669 m_AllOnes()))) {
1671 Value *Dec = Builder.CreateAdd(A, AllOnes);
1672 Value *Not = Builder.CreateXor(A, AllOnes);
1673 return BinaryOperator::CreateAnd(Dec, Not);
1674 }
1675
1676 // Disguised reassociation/factorization:
1677 // ~(A * C1) + A
1678 // ((A * -C1) - 1) + A
1679 // ((A * -C1) + A) - 1
1680 // (A * (1 - C1)) - 1
1681 if (match(&I,
1683 m_Deferred(A)))) {
1684 Type *Ty = I.getType();
1685 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1686 Value *NewMul = Builder.CreateMul(A, NewMulC);
1687 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1688 }
1689
1690 // (A * -2**C) + B --> B - (A << C)
1691 const APInt *NegPow2C;
1692 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1693 m_Value(B)))) {
1694 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1695 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1696 return BinaryOperator::CreateSub(B, Shl);
1697 }
1698
1699 // Canonicalize signum variant that ends in add:
1700 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1705 Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1706 Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1707 return BinaryOperator::CreateOr(LHS, Zext);
1708 }
1709
1710 {
1711 Value *Cond, *Ext;
1712 Constant *C;
1713 // (add X, (sext/zext (icmp eq X, C)))
1714 // -> (select (icmp eq X, C), (add C, (sext/zext 1)), X)
1715 auto CondMatcher = m_CombineAnd(
1716 m_Value(Cond),
1718
1719 if (match(&I,
1720 m_c_Add(m_Value(A),
1721 m_CombineAnd(m_Value(Ext), m_ZExtOrSExt(CondMatcher)))) &&
1722 Ext->hasOneUse()) {
1723 Value *Add = isa<ZExtInst>(Ext) ? InstCombiner::AddOne(C)
1726 }
1727 }
1728
1729 if (Instruction *Ashr = foldAddToAshr(I))
1730 return Ashr;
1731
1732 // (~X) + (~Y) --> -2 - (X + Y)
1733 {
1734 // To ensure we can save instructions we need to ensure that we consume both
1735 // LHS/RHS (i.e they have a `not`).
1736 bool ConsumesLHS, ConsumesRHS;
1737 if (isFreeToInvert(LHS, LHS->hasOneUse(), ConsumesLHS) && ConsumesLHS &&
1738 isFreeToInvert(RHS, RHS->hasOneUse(), ConsumesRHS) && ConsumesRHS) {
1741 assert(NotLHS != nullptr && NotRHS != nullptr &&
1742 "isFreeToInvert desynced with getFreelyInverted");
1743 Value *LHSPlusRHS = Builder.CreateAdd(NotLHS, NotRHS);
1744 return BinaryOperator::CreateSub(
1745 ConstantInt::getSigned(RHS->getType(), -2), LHSPlusRHS);
1746 }
1747 }
1748
1750 return R;
1751
1752 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1753 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1754 // computeKnownBits.
1755 bool Changed = false;
1756 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHSCache, RHSCache, I)) {
1757 Changed = true;
1758 I.setHasNoSignedWrap(true);
1759 }
1760 if (!I.hasNoUnsignedWrap() &&
1761 willNotOverflowUnsignedAdd(LHSCache, RHSCache, I)) {
1762 Changed = true;
1763 I.setHasNoUnsignedWrap(true);
1764 }
1765
1767 return V;
1768
1769 if (Instruction *V =
1771 return V;
1772
1774 return SatAdd;
1775
1776 // usub.sat(A, B) + B => umax(A, B)
1777 if (match(&I, m_c_BinOp(
1778 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1779 m_Deferred(B)))) {
1780 return replaceInstUsesWith(I,
1781 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1782 }
1783
1784 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1785 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1786 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1788 return replaceInstUsesWith(
1789 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1790 {Builder.CreateOr(A, B)}));
1791
1792 // Fold the log2_ceil idiom:
1793 // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
1794 // -->
1795 // BW - ctlz(A - 1, false)
1796 const APInt *XorC;
1798 if (match(&I,
1799 m_c_Add(
1800 m_ZExt(m_ICmp(Pred, m_Intrinsic<Intrinsic::ctpop>(m_Value(A)),
1801 m_One())),
1804 m_Intrinsic<Intrinsic::ctlz>(m_Deferred(A), m_One())))),
1805 m_APInt(XorC))))))) &&
1806 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
1807 *XorC == A->getType()->getScalarSizeInBits() - 1) {
1808 Value *Sub = Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()));
1809 Value *Ctlz = Builder.CreateIntrinsic(Intrinsic::ctlz, {A->getType()},
1810 {Sub, Builder.getFalse()});
1811 Value *Ret = Builder.CreateSub(
1812 ConstantInt::get(A->getType(), A->getType()->getScalarSizeInBits()),
1813 Ctlz, "", /*HasNUW*/ true, /*HasNSW*/ true);
1814 return replaceInstUsesWith(I, Builder.CreateZExtOrTrunc(Ret, I.getType()));
1815 }
1816
1817 if (Instruction *Res = foldSquareSumInt(I))
1818 return Res;
1819
1820 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1821 return Res;
1822
1824 return Res;
1825
1826 return Changed ? &I : nullptr;
1827}
1828
1829/// Eliminate an op from a linear interpolation (lerp) pattern.
1831 InstCombiner::BuilderTy &Builder) {
1832 Value *X, *Y, *Z;
1835 m_Value(Z))))),
1837 return nullptr;
1838
1839 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1840 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1841 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1842 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1843}
1844
1845/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1847 InstCombiner::BuilderTy &Builder) {
1848 assert((I.getOpcode() == Instruction::FAdd ||
1849 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1850 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1851 "FP factorization requires FMF");
1852
1853 if (Instruction *Lerp = factorizeLerp(I, Builder))
1854 return Lerp;
1855
1856 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1857 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1858 return nullptr;
1859
1860 Value *X, *Y, *Z;
1861 bool IsFMul;
1862 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1863 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1864 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1865 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1866 IsFMul = true;
1867 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1868 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1869 IsFMul = false;
1870 else
1871 return nullptr;
1872
1873 // (X * Z) + (Y * Z) --> (X + Y) * Z
1874 // (X * Z) - (Y * Z) --> (X - Y) * Z
1875 // (X / Z) + (Y / Z) --> (X + Y) / Z
1876 // (X / Z) - (Y / Z) --> (X - Y) / Z
1877 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1878 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1879 : Builder.CreateFSubFMF(X, Y, &I);
1880
1881 // Bail out if we just created a denormal constant.
1882 // TODO: This is copied from a previous implementation. Is it necessary?
1883 const APFloat *C;
1884 if (match(XY, m_APFloat(C)) && !C->isNormal())
1885 return nullptr;
1886
1887 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1889}
1890
1892 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1893 I.getFastMathFlags(),
1895 return replaceInstUsesWith(I, V);
1896
1898 return &I;
1899
1901 return X;
1902
1904 return Phi;
1905
1906 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1907 return FoldedFAdd;
1908
1909 // (-X) + Y --> Y - X
1910 Value *X, *Y;
1911 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1913
1914 // Similar to above, but look through fmul/fdiv for the negated term.
1915 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1916 Value *Z;
1918 m_Value(Z)))) {
1919 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1920 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1921 }
1922 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1923 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1925 m_Value(Z))) ||
1927 m_Value(Z)))) {
1928 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1929 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1930 }
1931
1932 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1933 // integer add followed by a promotion.
1934 if (Instruction *R = foldFBinOpOfIntCasts(I))
1935 return R;
1936
1937 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1938 // Handle specials cases for FAdd with selects feeding the operation
1940 return replaceInstUsesWith(I, V);
1941
1942 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1944 return F;
1945
1947 return F;
1948
1949 // Try to fold fadd into start value of reduction intrinsic.
1950 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1951 m_AnyZeroFP(), m_Value(X))),
1952 m_Value(Y)))) {
1953 // fadd (rdx 0.0, X), Y --> rdx Y, X
1954 return replaceInstUsesWith(
1955 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1956 {X->getType()}, {Y, X}, &I));
1957 }
1958 const APFloat *StartC, *C;
1959 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1960 m_APFloat(StartC), m_Value(X)))) &&
1961 match(RHS, m_APFloat(C))) {
1962 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1963 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1964 return replaceInstUsesWith(
1965 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1966 {X->getType()}, {NewStartC, X}, &I));
1967 }
1968
1969 // (X * MulC) + X --> X * (MulC + 1.0)
1970 Constant *MulC;
1971 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1972 m_Deferred(X)))) {
1974 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1975 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1976 }
1977
1978 // (-X - Y) + (X + Z) --> Z - Y
1980 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1981 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1982
1983 if (Value *V = FAddCombine(Builder).simplify(&I))
1984 return replaceInstUsesWith(I, V);
1985 }
1986
1987 // minumum(X, Y) + maximum(X, Y) => X + Y.
1988 if (match(&I,
1989 m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1990 m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1991 m_Deferred(Y))))) {
1993 // We cannot preserve ninf if nnan flag is not set.
1994 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1995 // while in optimized version NaN + Inf and this is a poison with ninf flag.
1996 if (!Result->hasNoNaNs())
1997 Result->setHasNoInfs(false);
1998 return Result;
1999 }
2000
2001 return nullptr;
2002}
2003
2004/// Optimize pointer differences into the same array into a size. Consider:
2005/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
2006/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
2008 Type *Ty, bool IsNUW) {
2009 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
2010 // this.
2011 bool Swapped = false;
2012 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
2013 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
2014 std::swap(LHS, RHS);
2015 Swapped = true;
2016 }
2017
2018 // Require at least one GEP with a common base pointer on both sides.
2019 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
2020 // (gep X, ...) - X
2021 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
2023 GEP1 = LHSGEP;
2024 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
2025 // (gep X, ...) - (gep X, ...)
2026 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
2027 RHSGEP->getOperand(0)->stripPointerCasts()) {
2028 GEP1 = LHSGEP;
2029 GEP2 = RHSGEP;
2030 }
2031 }
2032 }
2033
2034 if (!GEP1)
2035 return nullptr;
2036
2037 // To avoid duplicating the offset arithmetic, rewrite the GEP to use the
2038 // computed offset. This may erase the original GEP, so be sure to cache the
2039 // inbounds flag before emitting the offset.
2040 // TODO: We should probably do this even if there is only one GEP.
2041 bool RewriteGEPs = GEP2 != nullptr;
2042
2043 // Emit the offset of the GEP and an intptr_t.
2044 bool GEP1IsInBounds = GEP1->isInBounds();
2045 Value *Result = EmitGEPOffset(GEP1, RewriteGEPs);
2046
2047 // If this is a single inbounds GEP and the original sub was nuw,
2048 // then the final multiplication is also nuw.
2049 if (auto *I = dyn_cast<Instruction>(Result))
2050 if (IsNUW && !GEP2 && !Swapped && GEP1IsInBounds &&
2051 I->getOpcode() == Instruction::Mul)
2052 I->setHasNoUnsignedWrap();
2053
2054 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2055 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2056 if (GEP2) {
2057 bool GEP2IsInBounds = GEP2->isInBounds();
2058 Value *Offset = EmitGEPOffset(GEP2, RewriteGEPs);
2059 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2060 GEP1IsInBounds && GEP2IsInBounds);
2061 }
2062
2063 // If we have p - gep(p, ...) then we have to negate the result.
2064 if (Swapped)
2065 Result = Builder.CreateNeg(Result, "diff.neg");
2066
2067 return Builder.CreateIntCast(Result, Ty, true);
2068}
2069
2071 InstCombiner::BuilderTy &Builder) {
2072 Value *Op0 = I.getOperand(0);
2073 Value *Op1 = I.getOperand(1);
2074 Type *Ty = I.getType();
2075 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2076 if (!MinMax)
2077 return nullptr;
2078
2079 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2080 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2081 Value *X = MinMax->getLHS();
2082 Value *Y = MinMax->getRHS();
2083 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2084 (Op0->hasOneUse() || Op1->hasOneUse())) {
2085 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2086 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2087 return CallInst::Create(F, {X, Y});
2088 }
2089
2090 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2091 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2092 Value *Z;
2093 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2094 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2095 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2096 return BinaryOperator::CreateAdd(X, USub);
2097 }
2098 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2099 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2100 return BinaryOperator::CreateAdd(X, USub);
2101 }
2102 }
2103
2104 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2105 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2106 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2107 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2108 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2109 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2110 return CallInst::Create(F, {Op0, Z});
2111 }
2112
2113 return nullptr;
2114}
2115
2117 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2118 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2120 return replaceInstUsesWith(I, V);
2121
2123 return X;
2124
2126 return Phi;
2127
2128 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2129
2130 // If this is a 'B = x-(-A)', change to B = x+A.
2131 // We deal with this without involving Negator to preserve NSW flag.
2132 if (Value *V = dyn_castNegVal(Op1)) {
2133 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2134
2135 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2136 assert(BO->getOpcode() == Instruction::Sub &&
2137 "Expected a subtraction operator!");
2138 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2139 Res->setHasNoSignedWrap(true);
2140 } else {
2141 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2142 Res->setHasNoSignedWrap(true);
2143 }
2144
2145 return Res;
2146 }
2147
2148 // Try this before Negator to preserve NSW flag.
2150 return R;
2151
2152 Constant *C;
2153 if (match(Op0, m_ImmConstant(C))) {
2154 Value *X;
2155 Constant *C2;
2156
2157 // C-(X+C2) --> (C-C2)-X
2158 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2159 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2160 // => (C-C2)-X can have NSW/NUW
2161 bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2162 BinaryOperator *Res =
2163 BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2164 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2165 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2166 WillNotSOV);
2167 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2168 OBO1->hasNoUnsignedWrap());
2169 return Res;
2170 }
2171 }
2172
2173 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2174 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2175 return Ext;
2176
2177 bool Changed = false;
2178 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2179 Changed = true;
2180 I.setHasNoSignedWrap(true);
2181 }
2182 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2183 Changed = true;
2184 I.setHasNoUnsignedWrap(true);
2185 }
2186
2187 return Changed ? &I : nullptr;
2188 };
2189
2190 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2191 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2192 // a pure negation used by a select that looks like abs/nabs.
2193 bool IsNegation = match(Op0, m_ZeroInt());
2194 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2195 const Instruction *UI = dyn_cast<Instruction>(U);
2196 if (!UI)
2197 return false;
2198 return match(UI,
2199 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2200 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2201 })) {
2202 if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2203 I.hasNoSignedWrap(),
2204 Op1, *this))
2205 return BinaryOperator::CreateAdd(NegOp1, Op0);
2206 }
2207 if (IsNegation)
2208 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2209
2210 // (A*B)-(A*C) -> A*(B-C) etc
2212 return replaceInstUsesWith(I, V);
2213
2214 if (I.getType()->isIntOrIntVectorTy(1))
2215 return BinaryOperator::CreateXor(Op0, Op1);
2216
2217 // Replace (-1 - A) with (~A).
2218 if (match(Op0, m_AllOnes()))
2219 return BinaryOperator::CreateNot(Op1);
2220
2221 // (X + -1) - Y --> ~Y + X
2222 Value *X, *Y;
2223 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2224 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2225
2226 // Reassociate sub/add sequences to create more add instructions and
2227 // reduce dependency chains:
2228 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2229 Value *Z;
2231 m_Value(Z))))) {
2232 Value *XZ = Builder.CreateAdd(X, Z);
2233 Value *YW = Builder.CreateAdd(Y, Op1);
2234 return BinaryOperator::CreateSub(XZ, YW);
2235 }
2236
2237 // ((X - Y) - Op1) --> X - (Y + Op1)
2238 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2239 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Op0);
2240 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2241 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2242 Value *Add = Builder.CreateAdd(Y, Op1, "", /* HasNUW */ HasNUW,
2243 /* HasNSW */ HasNSW);
2244 BinaryOperator *Sub = BinaryOperator::CreateSub(X, Add);
2245 Sub->setHasNoUnsignedWrap(HasNUW);
2246 Sub->setHasNoSignedWrap(HasNSW);
2247 return Sub;
2248 }
2249
2250 {
2251 // (X + Z) - (Y + Z) --> (X - Y)
2252 // This is done in other passes, but we want to be able to consume this
2253 // pattern in InstCombine so we can generate it without creating infinite
2254 // loops.
2255 if (match(Op0, m_Add(m_Value(X), m_Value(Z))) &&
2256 match(Op1, m_c_Add(m_Value(Y), m_Specific(Z))))
2257 return BinaryOperator::CreateSub(X, Y);
2258
2259 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2260 Constant *CX, *CY;
2261 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_ImmConstant(CX)))) &&
2262 match(Op1, m_OneUse(m_Add(m_Value(Y), m_ImmConstant(CY))))) {
2263 Value *OpsSub = Builder.CreateSub(X, Y);
2264 Constant *ConstsSub = ConstantExpr::getSub(CX, CY);
2265 return BinaryOperator::CreateAdd(OpsSub, ConstsSub);
2266 }
2267 }
2268
2269 // (~X) - (~Y) --> Y - X
2270 {
2271 // Need to ensure we can consume at least one of the `not` instructions,
2272 // otherwise this can inf loop.
2273 bool ConsumesOp0, ConsumesOp1;
2274 if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
2275 isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
2276 (ConsumesOp0 || ConsumesOp1)) {
2277 Value *NotOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
2278 Value *NotOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
2279 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2280 "isFreeToInvert desynced with getFreelyInverted");
2281 return BinaryOperator::CreateSub(NotOp1, NotOp0);
2282 }
2283 }
2284
2285 auto m_AddRdx = [](Value *&Vec) {
2286 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2287 };
2288 Value *V0, *V1;
2289 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2290 V0->getType() == V1->getType()) {
2291 // Difference of sums is sum of differences:
2292 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2293 Value *Sub = Builder.CreateSub(V0, V1);
2294 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2295 {Sub->getType()}, {Sub});
2296 return replaceInstUsesWith(I, Rdx);
2297 }
2298
2299 if (Constant *C = dyn_cast<Constant>(Op0)) {
2300 Value *X;
2301 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2302 // C - (zext bool) --> bool ? C - 1 : C
2304 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2305 // C - (sext bool) --> bool ? C + 1 : C
2307
2308 // C - ~X == X + (1+C)
2309 if (match(Op1, m_Not(m_Value(X))))
2310 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2311
2312 // Try to fold constant sub into select arguments.
2313 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2314 if (Instruction *R = FoldOpIntoSelect(I, SI))
2315 return R;
2316
2317 // Try to fold constant sub into PHI values.
2318 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2319 if (Instruction *R = foldOpIntoPhi(I, PN))
2320 return R;
2321
2322 Constant *C2;
2323
2324 // C-(C2-X) --> X+(C-C2)
2325 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2326 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2327 }
2328
2329 const APInt *Op0C;
2330 if (match(Op0, m_APInt(Op0C))) {
2331 if (Op0C->isMask()) {
2332 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2333 // zero. We don't use information from dominating conditions so this
2334 // transform is easier to reverse if necessary.
2337 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2338 return BinaryOperator::CreateXor(Op1, Op0);
2339 }
2340
2341 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2342 // (C3 - ((C2 & C3) - 1)) is pow2
2343 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2344 // C2 is negative pow2 || sub nuw
2345 const APInt *C2, *C3;
2346 BinaryOperator *InnerSub;
2347 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2348 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2349 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2350 APInt C2AndC3 = *C2 & *C3;
2351 APInt C2AndC3Minus1 = C2AndC3 - 1;
2352 APInt C2AddC3 = *C2 + *C3;
2353 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2354 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2355 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2356 return BinaryOperator::CreateAdd(
2357 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2358 }
2359 }
2360 }
2361
2362 {
2363 Value *Y;
2364 // X-(X+Y) == -Y X-(Y+X) == -Y
2365 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2367
2368 // (X-Y)-X == -Y
2369 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2371 }
2372
2373 // (sub (or A, B) (and A, B)) --> (xor A, B)
2374 {
2375 Value *A, *B;
2376 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2377 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2378 return BinaryOperator::CreateXor(A, B);
2379 }
2380
2381 // (sub (add A, B) (or A, B)) --> (and A, B)
2382 {
2383 Value *A, *B;
2384 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2385 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2386 return BinaryOperator::CreateAnd(A, B);
2387 }
2388
2389 // (sub (add A, B) (and A, B)) --> (or A, B)
2390 {
2391 Value *A, *B;
2392 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2394 return BinaryOperator::CreateOr(A, B);
2395 }
2396
2397 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2398 {
2399 Value *A, *B;
2400 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2401 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2402 (Op0->hasOneUse() || Op1->hasOneUse()))
2404 }
2405
2406 // (sub (or A, B), (xor A, B)) --> (and A, B)
2407 {
2408 Value *A, *B;
2409 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2410 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2411 return BinaryOperator::CreateAnd(A, B);
2412 }
2413
2414 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2415 {
2416 Value *A, *B;
2417 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2418 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2419 (Op0->hasOneUse() || Op1->hasOneUse()))
2421 }
2422
2423 {
2424 Value *Y;
2425 // ((X | Y) - X) --> (~X & Y)
2426 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2427 return BinaryOperator::CreateAnd(
2428 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2429 }
2430
2431 {
2432 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2433 Value *X;
2434 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2435 m_OneUse(m_Neg(m_Value(X))))))) {
2437 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2438 }
2439 }
2440
2441 {
2442 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2443 Constant *C;
2444 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2447 }
2448 }
2449
2450 {
2451 // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2452 // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2453 Value *C, *X;
2454 auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2455 return match(LHS, m_OneUse(m_c_Xor(m_Value(X), m_Specific(RHS)))) &&
2456 match(RHS, m_SExt(m_Value(C))) &&
2457 (C->getType()->getScalarSizeInBits() == 1);
2458 };
2459 if (m_SubXorCmp(Op0, Op1))
2461 if (m_SubXorCmp(Op1, Op0))
2463 }
2464
2466 return R;
2467
2469 return R;
2470
2471 {
2472 // If we have a subtraction between some value and a select between
2473 // said value and something else, sink subtraction into select hands, i.e.:
2474 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2475 // ->
2476 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2477 // or
2478 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2479 // ->
2480 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2481 // This will result in select between new subtraction and 0.
2482 auto SinkSubIntoSelect =
2483 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2484 auto SubBuilder) -> Instruction * {
2485 Value *Cond, *TrueVal, *FalseVal;
2486 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2487 m_Value(FalseVal)))))
2488 return nullptr;
2489 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2490 return nullptr;
2491 // While it is really tempting to just create two subtractions and let
2492 // InstCombine fold one of those to 0, it isn't possible to do so
2493 // because of worklist visitation order. So ugly it is.
2494 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2495 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2496 Constant *Zero = Constant::getNullValue(Ty);
2497 SelectInst *NewSel =
2498 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2499 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2500 // Preserve prof metadata if any.
2501 NewSel->copyMetadata(cast<Instruction>(*Select));
2502 return NewSel;
2503 };
2504 if (Instruction *NewSel = SinkSubIntoSelect(
2505 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2506 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2507 return Builder->CreateSub(OtherHandOfSelect,
2508 /*OtherHandOfSub=*/Op1);
2509 }))
2510 return NewSel;
2511 if (Instruction *NewSel = SinkSubIntoSelect(
2512 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2513 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2514 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2515 OtherHandOfSelect);
2516 }))
2517 return NewSel;
2518 }
2519
2520 // (X - (X & Y)) --> (X & ~Y)
2521 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2522 (Op1->hasOneUse() || isa<Constant>(Y)))
2523 return BinaryOperator::CreateAnd(
2524 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2525
2526 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2527 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2528 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2529 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2530 // As long as Y is freely invertible, this will be neutral or a win.
2531 // Note: We don't generate the inverse max/min, just create the 'not' of
2532 // it and let other folds do the rest.
2533 if (match(Op0, m_Not(m_Value(X))) &&
2534 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2535 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2536 Value *Not = Builder.CreateNot(Op1);
2537 return BinaryOperator::CreateSub(Not, X);
2538 }
2539 if (match(Op1, m_Not(m_Value(X))) &&
2540 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2541 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2542 Value *Not = Builder.CreateNot(Op0);
2543 return BinaryOperator::CreateSub(X, Not);
2544 }
2545
2546 // Optimize pointer differences into the same array into a size. Consider:
2547 // &A[10] - &A[0]: we should compile this to "10".
2548 Value *LHSOp, *RHSOp;
2549 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2550 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2551 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2552 I.hasNoUnsignedWrap()))
2553 return replaceInstUsesWith(I, Res);
2554
2555 // trunc(p)-trunc(q) -> trunc(p-q)
2556 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2557 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2558 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2559 /* IsNUW */ false))
2560 return replaceInstUsesWith(I, Res);
2561
2562 // Canonicalize a shifty way to code absolute value to the common pattern.
2563 // There are 2 potential commuted variants.
2564 // We're relying on the fact that we only do this transform when the shift has
2565 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2566 // instructions).
2567 Value *A;
2568 const APInt *ShAmt;
2569 Type *Ty = I.getType();
2570 unsigned BitWidth = Ty->getScalarSizeInBits();
2571 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2572 Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2573 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2574 // B = ashr i32 A, 31 ; smear the sign bit
2575 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2576 // --> (A < 0) ? -A : A
2577 Value *IsNeg = Builder.CreateIsNeg(A);
2578 // Copy the nsw flags from the sub to the negate.
2579 Value *NegA = I.hasNoUnsignedWrap()
2580 ? Constant::getNullValue(A->getType())
2581 : Builder.CreateNeg(A, "", I.hasNoSignedWrap());
2582 return SelectInst::Create(IsNeg, NegA, A);
2583 }
2584
2585 // If we are subtracting a low-bit masked subset of some value from an add
2586 // of that same value with no low bits changed, that is clearing some low bits
2587 // of the sum:
2588 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2589 const APInt *AddC, *AndC;
2590 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2591 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2592 unsigned Cttz = AddC->countr_zero();
2593 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2594 if ((HighMask & *AndC).isZero())
2595 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2596 }
2597
2598 if (Instruction *V =
2600 return V;
2601
2602 // X - usub.sat(X, Y) => umin(X, Y)
2603 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2604 m_Value(Y)))))
2605 return replaceInstUsesWith(
2606 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2607
2608 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2609 // TODO: The one-use restriction is not strictly necessary, but it may
2610 // require improving other pattern matching and/or codegen.
2611 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2612 return replaceInstUsesWith(
2613 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2614
2615 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2616 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2617 return replaceInstUsesWith(
2618 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2619
2620 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2621 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2622 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2623 return BinaryOperator::CreateNeg(USub);
2624 }
2625
2626 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2627 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2628 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2629 return BinaryOperator::CreateNeg(USub);
2630 }
2631
2632 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2633 if (match(Op0, m_SpecificInt(BitWidth)) &&
2634 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2635 return replaceInstUsesWith(
2636 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2637 {Builder.CreateNot(X)}));
2638
2639 // Reduce multiplies for difference-of-squares by factoring:
2640 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2641 if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2642 match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2643 auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2644 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2645 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2646 OBO1->hasNoSignedWrap() && BitWidth > 2;
2647 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2648 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2649 Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2650 Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2651 Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2652 return replaceInstUsesWith(I, Mul);
2653 }
2654
2655 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2656 if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2658 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2659 Value *Sub =
2660 Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2661 Value *Call =
2662 Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2663 return replaceInstUsesWith(I, Call);
2664 }
2665 }
2666
2668 return Res;
2669
2670 return TryToNarrowDeduceFlags();
2671}
2672
2673/// This eliminates floating-point negation in either 'fneg(X)' or
2674/// 'fsub(-0.0, X)' form by combining into a constant operand.
2676 // This is limited with one-use because fneg is assumed better for
2677 // reassociation and cheaper in codegen than fmul/fdiv.
2678 // TODO: Should the m_OneUse restriction be removed?
2679 Instruction *FNegOp;
2680 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2681 return nullptr;
2682
2683 Value *X;
2684 Constant *C;
2685
2686 // Fold negation into constant operand.
2687 // -(X * C) --> X * (-C)
2688 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2689 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2690 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2691 // -(X / C) --> X / (-C)
2692 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2693 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2694 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2695 // -(C / X) --> (-C) / X
2696 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2697 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2699
2700 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2701 // not apply to the fdiv. Everything else propagates from the fneg.
2702 // TODO: We could propagate nsz/ninf from fdiv alone?
2703 FastMathFlags FMF = I.getFastMathFlags();
2704 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2705 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2706 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2707 return FDiv;
2708 }
2709 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2710 // -(X + C) --> -X + -C --> -C - X
2711 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2712 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2713 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2714
2715 return nullptr;
2716}
2717
2718Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2719 Instruction &FMFSource) {
2720 Value *X, *Y;
2721 if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2722 // Push into RHS which is more likely to simplify (const or another fneg).
2723 // FIXME: It would be better to invert the transform.
2724 return cast<Instruction>(Builder.CreateFMulFMF(
2725 X, Builder.CreateFNegFMF(Y, &FMFSource), &FMFSource));
2726 }
2727
2728 if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2729 return cast<Instruction>(Builder.CreateFDivFMF(
2730 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2731 }
2732
2733 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2734 // Make sure to preserve flags and metadata on the call.
2735 if (II->getIntrinsicID() == Intrinsic::ldexp) {
2736 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2739
2741 II->getCalledFunction(),
2742 {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2743 New->copyMetadata(*II);
2744 return New;
2745 }
2746 }
2747
2748 return nullptr;
2749}
2750
2752 Value *Op = I.getOperand(0);
2753
2754 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2755 getSimplifyQuery().getWithInstruction(&I)))
2756 return replaceInstUsesWith(I, V);
2757
2759 return X;
2760
2761 Value *X, *Y;
2762
2763 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2764 if (I.hasNoSignedZeros() &&
2767
2768 Value *OneUse;
2769 if (!match(Op, m_OneUse(m_Value(OneUse))))
2770 return nullptr;
2771
2772 if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2773 return replaceInstUsesWith(I, R);
2774
2775 // Try to eliminate fneg if at least 1 arm of the select is negated.
2776 Value *Cond;
2777 if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2778 // Unlike most transforms, this one is not safe to propagate nsz unless
2779 // it is present on the original select. We union the flags from the select
2780 // and fneg and then remove nsz if needed.
2781 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2782 S->copyFastMathFlags(&I);
2783 if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2784 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2785 S->setFastMathFlags(FMF);
2786 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2787 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2788 S->setHasNoSignedZeros(false);
2789 }
2790 };
2791 // -(Cond ? -P : Y) --> Cond ? P : -Y
2792 Value *P;
2793 if (match(X, m_FNeg(m_Value(P)))) {
2794 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2795 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2796 propagateSelectFMF(NewSel, P == Y);
2797 return NewSel;
2798 }
2799 // -(Cond ? X : -P) --> Cond ? -X : P
2800 if (match(Y, m_FNeg(m_Value(P)))) {
2801 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2802 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2803 propagateSelectFMF(NewSel, P == X);
2804 return NewSel;
2805 }
2806
2807 // -(Cond ? X : C) --> Cond ? -X : -C
2808 // -(Cond ? C : Y) --> Cond ? -C : -Y
2809 if (match(X, m_ImmConstant()) || match(Y, m_ImmConstant())) {
2810 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2811 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2812 SelectInst *NewSel = SelectInst::Create(Cond, NegX, NegY);
2813 propagateSelectFMF(NewSel, /*CommonOperand=*/true);
2814 return NewSel;
2815 }
2816 }
2817
2818 // fneg (copysign x, y) -> copysign x, (fneg y)
2819 if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2820 // The source copysign has an additional value input, so we can't propagate
2821 // flags the copysign doesn't also have.
2822 FastMathFlags FMF = I.getFastMathFlags();
2823 FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2824
2827
2828 Value *NegY = Builder.CreateFNeg(Y);
2829 Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2830 return replaceInstUsesWith(I, NewCopySign);
2831 }
2832
2833 return nullptr;
2834}
2835
2837 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2838 I.getFastMathFlags(),
2839 getSimplifyQuery().getWithInstruction(&I)))
2840 return replaceInstUsesWith(I, V);
2841
2843 return X;
2844
2846 return Phi;
2847
2848 // Subtraction from -0.0 is the canonical form of fneg.
2849 // fsub -0.0, X ==> fneg X
2850 // fsub nsz 0.0, X ==> fneg nsz X
2851 //
2852 // FIXME This matcher does not respect FTZ or DAZ yet:
2853 // fsub -0.0, Denorm ==> +-0
2854 // fneg Denorm ==> -Denorm
2855 Value *Op;
2856 if (match(&I, m_FNeg(m_Value(Op))))
2858
2860 return X;
2861
2862 if (Instruction *R = foldFBinOpOfIntCasts(I))
2863 return R;
2864
2865 Value *X, *Y;
2866 Constant *C;
2867
2868 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2869 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2870 // Canonicalize to fadd to make analysis easier.
2871 // This can also help codegen because fadd is commutative.
2872 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2873 // killed later. We still limit that particular transform with 'hasOneUse'
2874 // because an fneg is assumed better/cheaper than a generic fsub.
2875 if (I.hasNoSignedZeros() ||
2876 cannotBeNegativeZero(Op0, 0, getSimplifyQuery().getWithInstruction(&I))) {
2877 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2878 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2879 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2880 }
2881 }
2882
2883 // (-X) - Op1 --> -(X + Op1)
2884 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2885 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2886 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2888 }
2889
2890 if (isa<Constant>(Op0))
2891 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2892 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2893 return NV;
2894
2895 // X - C --> X + (-C)
2896 // But don't transform constant expressions because there's an inverse fold
2897 // for X + (-Y) --> X - Y.
2898 if (match(Op1, m_ImmConstant(C)))
2899 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2900 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2901
2902 // X - (-Y) --> X + Y
2903 if (match(Op1, m_FNeg(m_Value(Y))))
2904 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2905
2906 // Similar to above, but look through a cast of the negated value:
2907 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2908 Type *Ty = I.getType();
2909 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2911
2912 // X - (fpext(-Y)) --> X + fpext(Y)
2913 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2915
2916 // Similar to above, but look through fmul/fdiv of the negated value:
2917 // Op0 - (-X * Y) --> Op0 + (X * Y)
2918 // Op0 - (Y * -X) --> Op0 + (X * Y)
2919 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2921 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2922 }
2923 // Op0 - (-X / Y) --> Op0 + (X / Y)
2924 // Op0 - (X / -Y) --> Op0 + (X / Y)
2925 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2926 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2927 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2928 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2929 }
2930
2931 // Handle special cases for FSub with selects feeding the operation
2932 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2933 return replaceInstUsesWith(I, V);
2934
2935 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2936 // (Y - X) - Y --> -X
2937 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2939
2940 // Y - (X + Y) --> -X
2941 // Y - (Y + X) --> -X
2942 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2944
2945 // (X * C) - X --> X * (C - 1.0)
2946 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2948 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2949 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2950 }
2951 // X - (X * C) --> X * (1.0 - C)
2952 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2954 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2955 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2956 }
2957
2958 // Reassociate fsub/fadd sequences to create more fadd instructions and
2959 // reduce dependency chains:
2960 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2961 Value *Z;
2963 m_Value(Z))))) {
2964 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2965 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2966 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2967 }
2968
2969 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2970 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2971 m_Value(Vec)));
2972 };
2973 Value *A0, *A1, *V0, *V1;
2974 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2975 V0->getType() == V1->getType()) {
2976 // Difference of sums is sum of differences:
2977 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2978 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2979 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2980 {Sub->getType()}, {A0, Sub}, &I);
2981 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2982 }
2983
2985 return F;
2986
2987 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2988 // functionality has been subsumed by simple pattern matching here and in
2989 // InstSimplify. We should let a dedicated reassociation pass handle more
2990 // complex pattern matching and remove this from InstCombine.
2991 if (Value *V = FAddCombine(Builder).simplify(&I))
2992 return replaceInstUsesWith(I, V);
2993
2994 // (X - Y) - Op1 --> X - (Y + Op1)
2995 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2996 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2998 }
2999 }
3000
3001 return nullptr;
3002}
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...
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
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:512
#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:1368
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1116
Class for arbitrary precision integers.
Definition: APInt.h:78
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:427
bool isMinSignedValue() const
Determine if this is the smallest signed value.
Definition: APInt.h:401
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:184
bool isZero() const
Determine if this value is zero, i.e. all bits are clear.
Definition: APInt.h:358
bool isSignMask() const
Check if the APInt's value is returned by getSignMask.
Definition: APInt.h:444
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1446
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:307
int32_t exactLogBase2() const
Definition: APInt.h:1739
unsigned countr_zero() const
Count the number of trailing zero bits.
Definition: APInt.h:1596
unsigned countl_zero() const
The APInt version of std::countl_zero.
Definition: APInt.h:1555
static APInt getSignedMinValue(unsigned numBits)
Gets minimum signed value of APInt for a specific bit width.
Definition: APInt.h:197
unsigned logBase2() const
Definition: APInt.h:1717
APInt smul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1930
bool isMask(unsigned numBits) const
Definition: APInt.h:466
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:1235
bool isPowerOf2() const
Check if this APInt's value is a power of two greater than zero.
Definition: APInt.h:418
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Constructs an APInt value that has the top hiBitsSet bits set.
Definition: APInt.h:274
bool sge(const APInt &RHS) const
Signed greater or equal comparison.
Definition: APInt.h:1215
static BinaryOperator * CreateFAddFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:263
static BinaryOperator * CreateNeg(Value *Op, const Twine &Name="", InsertPosition InsertBefore=nullptr)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static BinaryOperator * CreateNot(Value *Op, const Twine &Name="", InsertPosition InsertBefore=nullptr)
static BinaryOperator * CreateFMulFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:271
static BinaryOperator * CreateFDivFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:275
static BinaryOperator * CreateFSubFMF(Value *V1, Value *V2, FastMathFlags FMF, const Twine &Name="")
Definition: InstrTypes.h:267
This class represents a function call, abstracting a target machine's calling convention.
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr="", InsertPosition InsertBefore=nullptr)
static CastInst * CreateTruncOrBitCast(Value *S, Type *Ty, const Twine &Name="", InsertPosition InsertBefore=nullptr)
Create a Trunc or BitCast cast instruction.
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", InsertPosition InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass's ...
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:757
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:780
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:784
@ ICMP_EQ
equal
Definition: InstrTypes.h:778
@ ICMP_NE
not equal
Definition: InstrTypes.h:779
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2618
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2611
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:42
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:63
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:1564
Value * CreateSRem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1427
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:922
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:1618
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:1591
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:1645
Value * CreateFPTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2121
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:2059
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:463
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:933
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:1757
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.cpp:1091
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2053
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2579
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:308
Value * CreateNUWAdd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1357
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNSW=false)
Definition: IRBuilder.h:1738
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1766
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:142
Value * CreateIsNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg < 0.
Definition: IRBuilder.h:2574
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1361
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1433
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="", bool IsNonNeg=false)
Definition: IRBuilder.h:2041
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1492
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1344
ConstantInt * getFalse()
Get the constant value for i1 false.
Definition: IRBuilder.h:468
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2569
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1514
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1683
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:2216
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2432
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2130
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1536
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1747
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1421
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1378
Value * CreateCopySign(Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create call to the copysign intrinsic.
Definition: IRBuilder.h:1031
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:229
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:383
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
Definition: InstCombiner.h:177
const DataLayout & DL
Definition: InstCombiner.h:75
unsigned ComputeNumSignBits(const Value *Op, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:449
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:407
DominatorTree & DT
Definition: InstCombiner.h:74
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
Definition: InstCombiner.h:428
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:444
Value * getFreelyInverted(Value *V, bool WillInvertAllUses, BuilderTy *Builder, bool &DoesConsume)
Definition: InstCombiner.h:210
const SimplifyQuery & getSimplifyQuery() const
Definition: InstCombiner.h:339
static Constant * AddOne(Constant *C)
Add one to a Constant.
Definition: InstCombiner.h:172
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:463
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="", InsertPosition InsertBefore=nullptr, 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:230
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="", InsertPosition InsertBefore=nullptr)
Definition: InstrTypes.h:164
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:694
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:1539
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.
SpecificCmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_SpecificICmp(ICmpInst::Predicate MatchPred, const LHS &L, const RHS &R)
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:480
constexpr bool isInt(int64_t x)
Checks if an integer fits into the given bit width.
Definition: MathExtras.h:169
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:518
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:526
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
SimplifyQuery getWithoutDomCondCache() const