LLVM 18.0.0git
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_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
823 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824 Constant *NewC =
825 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
826 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
827 }
828
829 // More general combining of constants in the wide type.
830 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
831 Constant *NarrowC;
832 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
833 Value *WideC = Builder.CreateSExt(NarrowC, Ty);
834 Value *NewC = Builder.CreateAdd(WideC, Op1C);
835 Value *WideX = Builder.CreateSExt(X, Ty);
836 return BinaryOperator::CreateAdd(WideX, NewC);
837 }
838 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
839 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
840 Value *WideC = Builder.CreateZExt(NarrowC, Ty);
841 Value *NewC = Builder.CreateAdd(WideC, Op1C);
842 Value *WideX = Builder.CreateZExt(X, Ty);
843 return BinaryOperator::CreateAdd(WideX, NewC);
844 }
845
846 return nullptr;
847}
848
850 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
851 Type *Ty = Add.getType();
852 Constant *Op1C;
853 if (!match(Op1, m_ImmConstant(Op1C)))
854 return nullptr;
855
857 return NV;
858
859 Value *X;
860 Constant *Op00C;
861
862 // add (sub C1, X), C2 --> sub (add C1, C2), X
863 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
864 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
865
866 Value *Y;
867
868 // add (sub X, Y), -1 --> add (not Y), X
869 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
870 match(Op1, m_AllOnes()))
871 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
872
873 // zext(bool) + C -> bool ? C + 1 : C
874 if (match(Op0, m_ZExt(m_Value(X))) &&
875 X->getType()->getScalarSizeInBits() == 1)
876 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
877 // sext(bool) + C -> bool ? C - 1 : C
878 if (match(Op0, m_SExt(m_Value(X))) &&
879 X->getType()->getScalarSizeInBits() == 1)
880 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
881
882 // ~X + C --> (C-1) - X
883 if (match(Op0, m_Not(m_Value(X)))) {
884 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
885 auto *COne = ConstantInt::get(Op1C->getType(), 1);
886 bool WillNotSOV = willNotOverflowSignedSub(Op1C, COne, Add);
887 BinaryOperator *Res =
888 BinaryOperator::CreateSub(ConstantExpr::getSub(Op1C, COne), X);
889 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
890 return Res;
891 }
892
893 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
894 const APInt *C;
895 unsigned BitWidth = Ty->getScalarSizeInBits();
896 if (match(Op0, m_OneUse(m_AShr(m_Value(X),
898 match(Op1, m_One()))
899 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
900
901 if (!match(Op1, m_APInt(C)))
902 return nullptr;
903
904 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
905 Constant *Op01C;
906 if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
908 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
909
910 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
911 const APInt *C2;
912 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
913 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
914
915 if (C->isSignMask()) {
916 // If wrapping is not allowed, then the addition must set the sign bit:
917 // X + (signmask) --> X | signmask
918 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
919 return BinaryOperator::CreateOr(Op0, Op1);
920
921 // If wrapping is allowed, then the addition flips the sign bit of LHS:
922 // X + (signmask) --> X ^ signmask
923 return BinaryOperator::CreateXor(Op0, Op1);
924 }
925
926 // Is this add the last step in a convoluted sext?
927 // add(zext(xor i16 X, -32768), -32768) --> sext X
928 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
929 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
930 return CastInst::Create(Instruction::SExt, X, Ty);
931
932 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
933 // (X ^ signmask) + C --> (X + (signmask ^ C))
934 if (C2->isSignMask())
935 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
936
937 // If X has no high-bits set above an xor mask:
938 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
939 if (C2->isMask()) {
940 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
941 if ((*C2 | LHSKnown.Zero).isAllOnes())
942 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
943 }
944
945 // Look for a math+logic pattern that corresponds to sext-in-register of a
946 // value with cleared high bits. Convert that into a pair of shifts:
947 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
948 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
949 if (Op0->hasOneUse() && *C2 == -(*C)) {
950 unsigned BitWidth = Ty->getScalarSizeInBits();
951 unsigned ShAmt = 0;
952 if (C->isPowerOf2())
953 ShAmt = BitWidth - C->logBase2() - 1;
954 else if (C2->isPowerOf2())
955 ShAmt = BitWidth - C2->logBase2() - 1;
957 0, &Add)) {
958 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
959 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
960 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
961 }
962 }
963 }
964
965 if (C->isOne() && Op0->hasOneUse()) {
966 // add (sext i1 X), 1 --> zext (not X)
967 // TODO: The smallest IR representation is (select X, 0, 1), and that would
968 // not require the one-use check. But we need to remove a transform in
969 // visitSelect and make sure that IR value tracking for select is equal or
970 // better than for these ops.
971 if (match(Op0, m_SExt(m_Value(X))) &&
972 X->getType()->getScalarSizeInBits() == 1)
973 return new ZExtInst(Builder.CreateNot(X), Ty);
974
975 // Shifts and add used to flip and mask off the low bit:
976 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
977 const APInt *C3;
978 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
979 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
980 Value *NotX = Builder.CreateNot(X);
981 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
982 }
983 }
984
985 // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
986 // TODO: There's a general form for any constant on the outer add.
987 if (C->isOne()) {
988 if (match(Op0, m_ZExt(m_Add(m_Value(X), m_AllOnes())))) {
990 if (llvm::isKnownNonZero(X, DL, 0, Q.AC, Q.CxtI, Q.DT))
991 return new ZExtInst(X, Ty);
992 }
993 }
994
995 return nullptr;
996}
997
998// match variations of a^2 + 2*a*b + b^2
999//
1000// to reuse the code between the FP and Int versions, the instruction OpCodes
1001// and constant types have been turned into template parameters.
1002//
1003// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1004// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1005// (we're matching `X<<1` instead of `X*2` for Int)
1006template <bool FP, typename Mul2Rhs>
1007static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1008 Value *&B) {
1009 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1010 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1011 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1012
1013 // (a * a) + (((a * 2) + b) * b)
1014 if (match(&I, m_c_BinOp(
1015 AddOp, m_OneUse(m_BinOp(MulOp, m_Value(A), m_Deferred(A))),
1017 MulOp,
1018 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(A), M2Rhs),
1019 m_Value(B)),
1020 m_Deferred(B))))))
1021 return true;
1022
1023 // ((a * b) * 2) or ((a * 2) * b)
1024 // +
1025 // (a * a + b * b) or (b * b + a * a)
1026 return match(
1027 &I,
1028 m_c_BinOp(AddOp,
1031 Mul2Op, m_BinOp(MulOp, m_Value(A), m_Value(B)), M2Rhs)),
1032 m_OneUse(m_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(A), M2Rhs),
1033 m_Value(B)))),
1035 AddOp, m_BinOp(MulOp, m_Deferred(A), m_Deferred(A)),
1036 m_BinOp(MulOp, m_Deferred(B), m_Deferred(B))))));
1037}
1038
1039// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1041 Value *A, *B;
1042 if (matchesSquareSum</*FP*/ false>(I, m_SpecificInt(1), A, B)) {
1043 Value *AB = Builder.CreateAdd(A, B);
1044 return BinaryOperator::CreateMul(AB, AB);
1045 }
1046 return nullptr;
1047}
1048
1049// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1050// Requires `nsz` and `reassoc`.
1052 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1053 Value *A, *B;
1054 if (matchesSquareSum</*FP*/ true>(I, m_SpecificFP(2.0), A, B)) {
1055 Value *AB = Builder.CreateFAddFMF(A, B, &I);
1056 return BinaryOperator::CreateFMulFMF(AB, AB, &I);
1057 }
1058 return nullptr;
1059}
1060
1061// Matches multiplication expression Op * C where C is a constant. Returns the
1062// constant value in C and the other operand in Op. Returns true if such a
1063// match is found.
1064static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1065 const APInt *AI;
1066 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1067 C = *AI;
1068 return true;
1069 }
1070 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1071 C = APInt(AI->getBitWidth(), 1);
1072 C <<= *AI;
1073 return true;
1074 }
1075 return false;
1076}
1077
1078// Matches remainder expression Op % C where C is a constant. Returns the
1079// constant value in C and the other operand in Op. Returns the signedness of
1080// the remainder operation in IsSigned. Returns true if such a match is
1081// found.
1082static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1083 const APInt *AI;
1084 IsSigned = false;
1085 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1086 IsSigned = true;
1087 C = *AI;
1088 return true;
1089 }
1090 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1091 C = *AI;
1092 return true;
1093 }
1094 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1095 C = *AI + 1;
1096 return true;
1097 }
1098 return false;
1099}
1100
1101// Matches division expression Op / C with the given signedness as indicated
1102// by IsSigned, where C is a constant. Returns the constant value in C and the
1103// other operand in Op. Returns true if such a match is found.
1104static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1105 const APInt *AI;
1106 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1107 C = *AI;
1108 return true;
1109 }
1110 if (!IsSigned) {
1111 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1112 C = *AI;
1113 return true;
1114 }
1115 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1116 C = APInt(AI->getBitWidth(), 1);
1117 C <<= *AI;
1118 return true;
1119 }
1120 }
1121 return false;
1122}
1123
1124// Returns whether C0 * C1 with the given signedness overflows.
1125static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1126 bool overflow;
1127 if (IsSigned)
1128 (void)C0.smul_ov(C1, overflow);
1129 else
1130 (void)C0.umul_ov(C1, overflow);
1131 return overflow;
1132}
1133
1134// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1135// does not overflow.
1137 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1138 Value *X, *MulOpV;
1139 APInt C0, MulOpC;
1140 bool IsSigned;
1141 // Match I = X % C0 + MulOpV * C0
1142 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1143 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1144 C0 == MulOpC) {
1145 Value *RemOpV;
1146 APInt C1;
1147 bool Rem2IsSigned;
1148 // Match MulOpC = RemOpV % C1
1149 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1150 IsSigned == Rem2IsSigned) {
1151 Value *DivOpV;
1152 APInt DivOpC;
1153 // Match RemOpV = X / C0
1154 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1155 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1156 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1157 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1158 : Builder.CreateURem(X, NewDivisor, "urem");
1159 }
1160 }
1161 }
1162
1163 return nullptr;
1164}
1165
1166/// Fold
1167/// (1 << NBits) - 1
1168/// Into:
1169/// ~(-(1 << NBits))
1170/// Because a 'not' is better for bit-tracking analysis and other transforms
1171/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1173 InstCombiner::BuilderTy &Builder) {
1174 Value *NBits;
1175 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1176 return nullptr;
1177
1178 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1179 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1180 // Be wary of constant folding.
1181 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1182 // Always NSW. But NUW propagates from `add`.
1183 BOp->setHasNoSignedWrap();
1184 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1185 }
1186
1187 return BinaryOperator::CreateNot(NotMask, I.getName());
1188}
1189
1191 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1192 Type *Ty = I.getType();
1193 auto getUAddSat = [&]() {
1194 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1195 };
1196
1197 // add (umin X, ~Y), Y --> uaddsat X, Y
1198 Value *X, *Y;
1200 m_Deferred(Y))))
1201 return CallInst::Create(getUAddSat(), { X, Y });
1202
1203 // add (umin X, ~C), C --> uaddsat X, C
1204 const APInt *C, *NotC;
1205 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1206 *C == ~*NotC)
1207 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1208
1209 return nullptr;
1210}
1211
1212// Transform:
1213// (add A, (shl (neg B), Y))
1214// -> (sub A, (shl B, Y))
1216 const BinaryOperator &I) {
1217 Value *A, *B, *Cnt;
1218 if (match(&I,
1220 m_Value(A)))) {
1221 Value *NewShl = Builder.CreateShl(B, Cnt);
1222 return BinaryOperator::CreateSub(A, NewShl);
1223 }
1224 return nullptr;
1225}
1226
1227/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1229 // Division must be by power-of-2, but not the minimum signed value.
1230 Value *X;
1231 const APInt *DivC;
1232 if (!match(Add.getOperand(0), m_SDiv(m_Value(X), m_Power2(DivC))) ||
1233 DivC->isNegative())
1234 return nullptr;
1235
1236 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1237 // low bits set. It recognizes two canonical patterns:
1238 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1239 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1240 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1241 // Note that, by the time we end up here, if possible, ugt has been
1242 // canonicalized into eq.
1243 const APInt *MaskC, *MaskCCmp;
1245 if (!match(Add.getOperand(1),
1246 m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1247 m_APInt(MaskCCmp)))))
1248 return nullptr;
1249
1250 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1251 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1252 return nullptr;
1253
1254 APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1255 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1256 ? (*MaskC == (SMin | (*DivC - 1)))
1257 : (*DivC == 2 && *MaskC == SMin + 1);
1258 if (!IsMaskValid)
1259 return nullptr;
1260
1261 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1262 return BinaryOperator::CreateAShr(
1263 X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1264}
1265
1268 BinaryOperator &I) {
1269 assert((I.getOpcode() == Instruction::Add ||
1270 I.getOpcode() == Instruction::Or ||
1271 I.getOpcode() == Instruction::Sub) &&
1272 "Expecting add/or/sub instruction");
1273
1274 // We have a subtraction/addition between a (potentially truncated) *logical*
1275 // right-shift of X and a "select".
1276 Value *X, *Select;
1277 Instruction *LowBitsToSkip, *Extract;
1279 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1280 m_Instruction(Extract))),
1281 m_Value(Select))))
1282 return nullptr;
1283
1284 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1285 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1286 return nullptr;
1287
1288 Type *XTy = X->getType();
1289 bool HadTrunc = I.getType() != XTy;
1290
1291 // If there was a truncation of extracted value, then we'll need to produce
1292 // one extra instruction, so we need to ensure one instruction will go away.
1293 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1294 return nullptr;
1295
1296 // Extraction should extract high NBits bits, with shift amount calculated as:
1297 // low bits to skip = shift bitwidth - high bits to extract
1298 // The shift amount itself may be extended, and we need to look past zero-ext
1299 // when matching NBits, that will matter for matching later.
1300 Constant *C;
1301 Value *NBits;
1302 if (!match(
1303 LowBitsToSkip,
1306 APInt(C->getType()->getScalarSizeInBits(),
1307 X->getType()->getScalarSizeInBits()))))
1308 return nullptr;
1309
1310 // Sign-extending value can be zero-extended if we `sub`tract it,
1311 // or sign-extended otherwise.
1312 auto SkipExtInMagic = [&I](Value *&V) {
1313 if (I.getOpcode() == Instruction::Sub)
1314 match(V, m_ZExtOrSelf(m_Value(V)));
1315 else
1316 match(V, m_SExtOrSelf(m_Value(V)));
1317 };
1318
1319 // Now, finally validate the sign-extending magic.
1320 // `select` itself may be appropriately extended, look past that.
1321 SkipExtInMagic(Select);
1322
1324 const APInt *Thr;
1325 Value *SignExtendingValue, *Zero;
1326 bool ShouldSignext;
1327 // It must be a select between two values we will later establish to be a
1328 // sign-extending value and a zero constant. The condition guarding the
1329 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1330 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1331 m_Value(SignExtendingValue), m_Value(Zero))) ||
1332 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1333 return nullptr;
1334
1335 // icmp-select pair is commutative.
1336 if (!ShouldSignext)
1337 std::swap(SignExtendingValue, Zero);
1338
1339 // If we should not perform sign-extension then we must add/or/subtract zero.
1340 if (!match(Zero, m_Zero()))
1341 return nullptr;
1342 // Otherwise, it should be some constant, left-shifted by the same NBits we
1343 // had in `lshr`. Said left-shift can also be appropriately extended.
1344 // Again, we must look past zero-ext when looking for NBits.
1345 SkipExtInMagic(SignExtendingValue);
1346 Constant *SignExtendingValueBaseConstant;
1347 if (!match(SignExtendingValue,
1348 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1349 m_ZExtOrSelf(m_Specific(NBits)))))
1350 return nullptr;
1351 // If we `sub`, then the constant should be one, else it should be all-ones.
1352 if (I.getOpcode() == Instruction::Sub
1353 ? !match(SignExtendingValueBaseConstant, m_One())
1354 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1355 return nullptr;
1356
1357 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1358 Extract->getName() + ".sext");
1359 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1360 if (!HadTrunc)
1361 return NewAShr;
1362
1363 Builder.Insert(NewAShr);
1364 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1365}
1366
1367/// This is a specialization of a more general transform from
1368/// foldUsingDistributiveLaws. If that code can be made to work optimally
1369/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1371 InstCombiner::BuilderTy &Builder) {
1372 // TODO: Also handle mul by doubling the shift amount?
1373 assert((I.getOpcode() == Instruction::Add ||
1374 I.getOpcode() == Instruction::Sub) &&
1375 "Expected add/sub");
1376 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1377 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1378 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1379 return nullptr;
1380
1381 Value *X, *Y, *ShAmt;
1382 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1383 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1384 return nullptr;
1385
1386 // No-wrap propagates only when all ops have no-wrap.
1387 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1388 Op1->hasNoSignedWrap();
1389 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1390 Op1->hasNoUnsignedWrap();
1391
1392 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1393 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1394 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1395 NewI->setHasNoSignedWrap(HasNSW);
1396 NewI->setHasNoUnsignedWrap(HasNUW);
1397 }
1398 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1399 NewShl->setHasNoSignedWrap(HasNSW);
1400 NewShl->setHasNoUnsignedWrap(HasNUW);
1401 return NewShl;
1402}
1403
1404/// Reduce a sequence of masked half-width multiplies to a single multiply.
1405/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1407 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1408 // Skip the odd bitwidth types.
1409 if ((BitWidth & 0x1))
1410 return nullptr;
1411
1412 unsigned HalfBits = BitWidth >> 1;
1413 APInt HalfMask = APInt::getMaxValue(HalfBits);
1414
1415 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1416 Value *XLo, *YLo;
1417 Value *CrossSum;
1418 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1419 m_Mul(m_Value(YLo), m_Value(XLo)))))
1420 return nullptr;
1421
1422 // XLo = X & HalfMask
1423 // YLo = Y & HalfMask
1424 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1425 // to enhance robustness
1426 Value *X, *Y;
1427 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1428 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1429 return nullptr;
1430
1431 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1432 // X' can be either X or XLo in the pattern (and the same for Y')
1433 if (match(CrossSum,
1438 return BinaryOperator::CreateMul(X, Y);
1439
1440 return nullptr;
1441}
1442
1444 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1445 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1447 return replaceInstUsesWith(I, V);
1448
1450 return &I;
1451
1453 return X;
1454
1456 return Phi;
1457
1458 // (A*B)+(A*C) -> A*(B+C) etc
1460 return replaceInstUsesWith(I, V);
1461
1462 if (Instruction *R = foldBoxMultiply(I))
1463 return R;
1464
1466 return R;
1467
1469 return X;
1470
1472 return X;
1473
1475 return R;
1476
1478 return R;
1479
1480 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1481 Type *Ty = I.getType();
1482 if (Ty->isIntOrIntVectorTy(1))
1483 return BinaryOperator::CreateXor(LHS, RHS);
1484
1485 // X + X --> X << 1
1486 if (LHS == RHS) {
1487 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1488 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1489 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1490 return Shl;
1491 }
1492
1493 Value *A, *B;
1494 if (match(LHS, m_Neg(m_Value(A)))) {
1495 // -A + -B --> -(A + B)
1496 if (match(RHS, m_Neg(m_Value(B))))
1498
1499 // -A + B --> B - A
1500 auto *Sub = BinaryOperator::CreateSub(RHS, A);
1501 auto *OB0 = cast<OverflowingBinaryOperator>(LHS);
1502 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1503
1504 return Sub;
1505 }
1506
1507 // A + -B --> A - B
1508 if (match(RHS, m_Neg(m_Value(B))))
1509 return BinaryOperator::CreateSub(LHS, B);
1510
1512 return replaceInstUsesWith(I, V);
1513
1514 // (A + 1) + ~B --> A - B
1515 // ~B + (A + 1) --> A - B
1516 // (~B + A) + 1 --> A - B
1517 // (A + ~B) + 1 --> A - B
1518 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1520 return BinaryOperator::CreateSub(A, B);
1521
1522 // (A + RHS) + RHS --> A + (RHS << 1)
1524 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1525
1526 // LHS + (A + LHS) --> A + (LHS << 1)
1528 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1529
1530 {
1531 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1532 Constant *C1, *C2;
1533 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1534 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1535 (LHS->hasOneUse() || RHS->hasOneUse())) {
1536 Value *Sub = Builder.CreateSub(A, B);
1537 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1538 }
1539
1540 // Canonicalize a constant sub operand as an add operand for better folding:
1541 // (C1 - A) + B --> (B - A) + C1
1543 m_Value(B)))) {
1544 Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1545 return BinaryOperator::CreateAdd(Sub, C1);
1546 }
1547 }
1548
1549 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1551
1552 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1553 const APInt *C1, *C2;
1554 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1555 APInt one(C2->getBitWidth(), 1);
1556 APInt minusC1 = -(*C1);
1557 if (minusC1 == (one << *C2)) {
1558 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1559 return BinaryOperator::CreateSRem(RHS, NewRHS);
1560 }
1561 }
1562
1563 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1564 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1565 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1566 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1567 return BinaryOperator::CreateAnd(A, NewMask);
1568 }
1569
1570 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1571 if ((match(RHS, m_ZExt(m_Value(A))) &&
1573 (match(LHS, m_ZExt(m_Value(A))) &&
1575 return new ZExtInst(B, LHS->getType());
1576
1577 // zext(A) + sext(A) --> 0 if A is i1
1579 A->getType()->isIntOrIntVectorTy(1))
1580 return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1581
1582 // A+B --> A|B iff A and B have no bits set in common.
1583 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1584 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(&I))) {
1585 auto *Or = BinaryOperator::CreateOr(LHS, RHS);
1586 cast<PossiblyDisjointInst>(Or)->setIsDisjoint(true);
1587 return Or;
1588 }
1589
1590 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1591 return Ext;
1592
1593 // (add (xor A, B) (and A, B)) --> (or A, B)
1594 // (add (and A, B) (xor A, B)) --> (or A, B)
1595 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1597 return BinaryOperator::CreateOr(A, B);
1598
1599 // (add (or A, B) (and A, B)) --> (add A, B)
1600 // (add (and A, B) (or A, B)) --> (add A, B)
1601 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1603 // Replacing operands in-place to preserve nuw/nsw flags.
1604 replaceOperand(I, 0, A);
1605 replaceOperand(I, 1, B);
1606 return &I;
1607 }
1608
1609 // (add A (or A, -A)) --> (and (add A, -1) A)
1610 // (add A (or -A, A)) --> (and (add A, -1) A)
1611 // (add (or A, -A) A) --> (and (add A, -1) A)
1612 // (add (or -A, A) A) --> (and (add A, -1) A)
1614 m_Deferred(A)))))) {
1615 Value *Add =
1617 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1618 return BinaryOperator::CreateAnd(Add, A);
1619 }
1620
1621 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1622 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1623 if (match(&I,
1625 m_AllOnes()))) {
1627 Value *Dec = Builder.CreateAdd(A, AllOnes);
1628 Value *Not = Builder.CreateXor(A, AllOnes);
1629 return BinaryOperator::CreateAnd(Dec, Not);
1630 }
1631
1632 // Disguised reassociation/factorization:
1633 // ~(A * C1) + A
1634 // ((A * -C1) - 1) + A
1635 // ((A * -C1) + A) - 1
1636 // (A * (1 - C1)) - 1
1637 if (match(&I,
1639 m_Deferred(A)))) {
1640 Type *Ty = I.getType();
1641 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1642 Value *NewMul = Builder.CreateMul(A, NewMulC);
1643 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1644 }
1645
1646 // (A * -2**C) + B --> B - (A << C)
1647 const APInt *NegPow2C;
1648 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1649 m_Value(B)))) {
1650 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1651 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1652 return BinaryOperator::CreateSub(B, Shl);
1653 }
1654
1655 // Canonicalize signum variant that ends in add:
1656 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1661 m_OneUse(m_ICmp(Pred, m_Specific(A), m_ZeroInt()))))) &&
1662 Pred == CmpInst::ICMP_SGT) {
1663 Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1664 Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1665 return BinaryOperator::CreateOr(LHS, Zext);
1666 }
1667
1668 if (Instruction *Ashr = foldAddToAshr(I))
1669 return Ashr;
1670
1671 // min(A, B) + max(A, B) => A + B.
1676 return BinaryOperator::CreateWithCopiedFlags(Instruction::Add, A, B, &I);
1677
1678 // (~X) + (~Y) --> -2 - (X + Y)
1679 {
1680 // To ensure we can save instructions we need to ensure that we consume both
1681 // LHS/RHS (i.e they have a `not`).
1682 bool ConsumesLHS, ConsumesRHS;
1683 if (isFreeToInvert(LHS, LHS->hasOneUse(), ConsumesLHS) && ConsumesLHS &&
1684 isFreeToInvert(RHS, RHS->hasOneUse(), ConsumesRHS) && ConsumesRHS) {
1687 assert(NotLHS != nullptr && NotRHS != nullptr &&
1688 "isFreeToInvert desynced with getFreelyInverted");
1689 Value *LHSPlusRHS = Builder.CreateAdd(NotLHS, NotRHS);
1690 return BinaryOperator::CreateSub(ConstantInt::get(RHS->getType(), -2),
1691 LHSPlusRHS);
1692 }
1693 }
1694
1695 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1696 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1697 // computeKnownBits.
1698 bool Changed = false;
1699 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHSCache, RHSCache, I)) {
1700 Changed = true;
1701 I.setHasNoSignedWrap(true);
1702 }
1703 if (!I.hasNoUnsignedWrap() &&
1704 willNotOverflowUnsignedAdd(LHSCache, RHSCache, I)) {
1705 Changed = true;
1706 I.setHasNoUnsignedWrap(true);
1707 }
1708
1710 return V;
1711
1712 if (Instruction *V =
1714 return V;
1715
1717 return SatAdd;
1718
1719 // usub.sat(A, B) + B => umax(A, B)
1720 if (match(&I, m_c_BinOp(
1721 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1722 m_Deferred(B)))) {
1723 return replaceInstUsesWith(I,
1724 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1725 }
1726
1727 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1728 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1729 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1731 return replaceInstUsesWith(
1732 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1733 {Builder.CreateOr(A, B)}));
1734
1735 if (Instruction *Res = foldSquareSumInt(I))
1736 return Res;
1737
1738 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1739 return Res;
1740
1742 return Res;
1743
1744 return Changed ? &I : nullptr;
1745}
1746
1747/// Eliminate an op from a linear interpolation (lerp) pattern.
1749 InstCombiner::BuilderTy &Builder) {
1750 Value *X, *Y, *Z;
1753 m_Value(Z))))),
1755 return nullptr;
1756
1757 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1758 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1759 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1760 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1761}
1762
1763/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1765 InstCombiner::BuilderTy &Builder) {
1766 assert((I.getOpcode() == Instruction::FAdd ||
1767 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1768 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1769 "FP factorization requires FMF");
1770
1771 if (Instruction *Lerp = factorizeLerp(I, Builder))
1772 return Lerp;
1773
1774 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1775 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1776 return nullptr;
1777
1778 Value *X, *Y, *Z;
1779 bool IsFMul;
1780 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1781 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1782 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1783 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1784 IsFMul = true;
1785 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1786 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1787 IsFMul = false;
1788 else
1789 return nullptr;
1790
1791 // (X * Z) + (Y * Z) --> (X + Y) * Z
1792 // (X * Z) - (Y * Z) --> (X - Y) * Z
1793 // (X / Z) + (Y / Z) --> (X + Y) / Z
1794 // (X / Z) - (Y / Z) --> (X - Y) / Z
1795 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1796 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1797 : Builder.CreateFSubFMF(X, Y, &I);
1798
1799 // Bail out if we just created a denormal constant.
1800 // TODO: This is copied from a previous implementation. Is it necessary?
1801 const APFloat *C;
1802 if (match(XY, m_APFloat(C)) && !C->isNormal())
1803 return nullptr;
1804
1805 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1807}
1808
1810 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1811 I.getFastMathFlags(),
1813 return replaceInstUsesWith(I, V);
1814
1816 return &I;
1817
1819 return X;
1820
1822 return Phi;
1823
1824 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1825 return FoldedFAdd;
1826
1827 // (-X) + Y --> Y - X
1828 Value *X, *Y;
1829 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1831
1832 // Similar to above, but look through fmul/fdiv for the negated term.
1833 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1834 Value *Z;
1836 m_Value(Z)))) {
1837 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1838 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1839 }
1840 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1841 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1843 m_Value(Z))) ||
1845 m_Value(Z)))) {
1846 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1847 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1848 }
1849
1850 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1851 // integer add followed by a promotion.
1852 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1853 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1854 Value *LHSIntVal = LHSConv->getOperand(0);
1855 Type *FPType = LHSConv->getType();
1856
1857 // TODO: This check is overly conservative. In many cases known bits
1858 // analysis can tell us that the result of the addition has less significant
1859 // bits than the integer type can hold.
1860 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1861 Type *FScalarTy = FTy->getScalarType();
1862 Type *IScalarTy = ITy->getScalarType();
1863
1864 // Do we have enough bits in the significand to represent the result of
1865 // the integer addition?
1866 unsigned MaxRepresentableBits =
1868 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1869 };
1870
1871 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1872 // ... if the constant fits in the integer value. This is useful for things
1873 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1874 // requires a constant pool load, and generally allows the add to be better
1875 // instcombined.
1876 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1877 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1878 Constant *CI = ConstantFoldCastOperand(Instruction::FPToSI, CFP,
1879 LHSIntVal->getType(), DL);
1880 if (LHSConv->hasOneUse() &&
1881 ConstantFoldCastOperand(Instruction::SIToFP, CI, I.getType(), DL) ==
1882 CFP &&
1883 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1884 // Insert the new integer add.
1885 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1886 return new SIToFPInst(NewAdd, I.getType());
1887 }
1888 }
1889
1890 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1891 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1892 Value *RHSIntVal = RHSConv->getOperand(0);
1893 // It's enough to check LHS types only because we require int types to
1894 // be the same for this transform.
1895 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1896 // Only do this if x/y have the same type, if at least one of them has a
1897 // single use (so we don't increase the number of int->fp conversions),
1898 // and if the integer add will not overflow.
1899 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1900 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1901 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1902 // Insert the new integer add.
1903 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1904 return new SIToFPInst(NewAdd, I.getType());
1905 }
1906 }
1907 }
1908 }
1909
1910 // Handle specials cases for FAdd with selects feeding the operation
1912 return replaceInstUsesWith(I, V);
1913
1914 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1916 return F;
1917
1919 return F;
1920
1921 // Try to fold fadd into start value of reduction intrinsic.
1922 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1923 m_AnyZeroFP(), m_Value(X))),
1924 m_Value(Y)))) {
1925 // fadd (rdx 0.0, X), Y --> rdx Y, X
1926 return replaceInstUsesWith(
1927 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1928 {X->getType()}, {Y, X}, &I));
1929 }
1930 const APFloat *StartC, *C;
1931 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1932 m_APFloat(StartC), m_Value(X)))) &&
1933 match(RHS, m_APFloat(C))) {
1934 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1935 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1936 return replaceInstUsesWith(
1937 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1938 {X->getType()}, {NewStartC, X}, &I));
1939 }
1940
1941 // (X * MulC) + X --> X * (MulC + 1.0)
1942 Constant *MulC;
1943 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1944 m_Deferred(X)))) {
1946 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1947 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1948 }
1949
1950 // (-X - Y) + (X + Z) --> Z - Y
1952 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1953 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1954
1955 if (Value *V = FAddCombine(Builder).simplify(&I))
1956 return replaceInstUsesWith(I, V);
1957 }
1958
1959 // minumum(X, Y) + maximum(X, Y) => X + Y.
1960 if (match(&I,
1961 m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1962 m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1963 m_Deferred(Y))))) {
1965 // We cannot preserve ninf if nnan flag is not set.
1966 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1967 // while in optimized version NaN + Inf and this is a poison with ninf flag.
1968 if (!Result->hasNoNaNs())
1969 Result->setHasNoInfs(false);
1970 return Result;
1971 }
1972
1973 return nullptr;
1974}
1975
1976/// Optimize pointer differences into the same array into a size. Consider:
1977/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1978/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1980 Type *Ty, bool IsNUW) {
1981 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1982 // this.
1983 bool Swapped = false;
1984 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1985 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1986 std::swap(LHS, RHS);
1987 Swapped = true;
1988 }
1989
1990 // Require at least one GEP with a common base pointer on both sides.
1991 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1992 // (gep X, ...) - X
1993 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1995 GEP1 = LHSGEP;
1996 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1997 // (gep X, ...) - (gep X, ...)
1998 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1999 RHSGEP->getOperand(0)->stripPointerCasts()) {
2000 GEP1 = LHSGEP;
2001 GEP2 = RHSGEP;
2002 }
2003 }
2004 }
2005
2006 if (!GEP1)
2007 return nullptr;
2008
2009 if (GEP2) {
2010 // (gep X, ...) - (gep X, ...)
2011 //
2012 // Avoid duplicating the arithmetic if there are more than one non-constant
2013 // indices between the two GEPs and either GEP has a non-constant index and
2014 // multiple users. If zero non-constant index, the result is a constant and
2015 // there is no duplication. If one non-constant index, the result is an add
2016 // or sub with a constant, which is no larger than the original code, and
2017 // there's no duplicated arithmetic, even if either GEP has multiple
2018 // users. If more than one non-constant indices combined, as long as the GEP
2019 // with at least one non-constant index doesn't have multiple users, there
2020 // is no duplication.
2021 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
2022 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
2023 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
2024 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
2025 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
2026 return nullptr;
2027 }
2028 }
2029
2030 // Emit the offset of the GEP and an intptr_t.
2031 Value *Result = EmitGEPOffset(GEP1);
2032
2033 // If this is a single inbounds GEP and the original sub was nuw,
2034 // then the final multiplication is also nuw.
2035 if (auto *I = dyn_cast<Instruction>(Result))
2036 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
2037 I->getOpcode() == Instruction::Mul)
2038 I->setHasNoUnsignedWrap();
2039
2040 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2041 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2042 if (GEP2) {
2043 Value *Offset = EmitGEPOffset(GEP2);
2044 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2045 GEP1->isInBounds() && GEP2->isInBounds());
2046 }
2047
2048 // If we have p - gep(p, ...) then we have to negate the result.
2049 if (Swapped)
2050 Result = Builder.CreateNeg(Result, "diff.neg");
2051
2052 return Builder.CreateIntCast(Result, Ty, true);
2053}
2054
2056 InstCombiner::BuilderTy &Builder) {
2057 Value *Op0 = I.getOperand(0);
2058 Value *Op1 = I.getOperand(1);
2059 Type *Ty = I.getType();
2060 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2061 if (!MinMax)
2062 return nullptr;
2063
2064 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2065 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2066 Value *X = MinMax->getLHS();
2067 Value *Y = MinMax->getRHS();
2068 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2069 (Op0->hasOneUse() || Op1->hasOneUse())) {
2070 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2071 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2072 return CallInst::Create(F, {X, Y});
2073 }
2074
2075 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2076 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2077 Value *Z;
2078 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2079 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2080 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2081 return BinaryOperator::CreateAdd(X, USub);
2082 }
2083 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2084 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2085 return BinaryOperator::CreateAdd(X, USub);
2086 }
2087 }
2088
2089 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2090 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2091 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2092 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2093 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2094 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2095 return CallInst::Create(F, {Op0, Z});
2096 }
2097
2098 return nullptr;
2099}
2100
2102 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2103 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2105 return replaceInstUsesWith(I, V);
2106
2108 return X;
2109
2111 return Phi;
2112
2113 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2114
2115 // If this is a 'B = x-(-A)', change to B = x+A.
2116 // We deal with this without involving Negator to preserve NSW flag.
2117 if (Value *V = dyn_castNegVal(Op1)) {
2118 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2119
2120 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2121 assert(BO->getOpcode() == Instruction::Sub &&
2122 "Expected a subtraction operator!");
2123 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2124 Res->setHasNoSignedWrap(true);
2125 } else {
2126 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2127 Res->setHasNoSignedWrap(true);
2128 }
2129
2130 return Res;
2131 }
2132
2133 // Try this before Negator to preserve NSW flag.
2135 return R;
2136
2137 Constant *C;
2138 if (match(Op0, m_ImmConstant(C))) {
2139 Value *X;
2140 Constant *C2;
2141
2142 // C-(X+C2) --> (C-C2)-X
2143 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2144 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2145 // => (C-C2)-X can have NSW/NUW
2146 bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2147 BinaryOperator *Res =
2148 BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2149 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2150 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2151 WillNotSOV);
2152 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2153 OBO1->hasNoUnsignedWrap());
2154 return Res;
2155 }
2156 }
2157
2158 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2159 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2160 return Ext;
2161
2162 bool Changed = false;
2163 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2164 Changed = true;
2165 I.setHasNoSignedWrap(true);
2166 }
2167 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2168 Changed = true;
2169 I.setHasNoUnsignedWrap(true);
2170 }
2171
2172 return Changed ? &I : nullptr;
2173 };
2174
2175 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2176 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2177 // a pure negation used by a select that looks like abs/nabs.
2178 bool IsNegation = match(Op0, m_ZeroInt());
2179 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2180 const Instruction *UI = dyn_cast<Instruction>(U);
2181 if (!UI)
2182 return false;
2183 return match(UI,
2184 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2185 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2186 })) {
2187 if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2188 I.hasNoSignedWrap(),
2189 Op1, *this))
2190 return BinaryOperator::CreateAdd(NegOp1, Op0);
2191 }
2192 if (IsNegation)
2193 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2194
2195 // (A*B)-(A*C) -> A*(B-C) etc
2197 return replaceInstUsesWith(I, V);
2198
2199 if (I.getType()->isIntOrIntVectorTy(1))
2200 return BinaryOperator::CreateXor(Op0, Op1);
2201
2202 // Replace (-1 - A) with (~A).
2203 if (match(Op0, m_AllOnes()))
2204 return BinaryOperator::CreateNot(Op1);
2205
2206 // (X + -1) - Y --> ~Y + X
2207 Value *X, *Y;
2208 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2209 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2210
2211 // Reassociate sub/add sequences to create more add instructions and
2212 // reduce dependency chains:
2213 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2214 Value *Z;
2216 m_Value(Z))))) {
2217 Value *XZ = Builder.CreateAdd(X, Z);
2218 Value *YW = Builder.CreateAdd(Y, Op1);
2219 return BinaryOperator::CreateSub(XZ, YW);
2220 }
2221
2222 // ((X - Y) - Op1) --> X - (Y + Op1)
2223 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2224 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Op0);
2225 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2226 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2227 Value *Add = Builder.CreateAdd(Y, Op1, "", /* HasNUW */ HasNUW,
2228 /* HasNSW */ HasNSW);
2229 BinaryOperator *Sub = BinaryOperator::CreateSub(X, Add);
2230 Sub->setHasNoUnsignedWrap(HasNUW);
2231 Sub->setHasNoSignedWrap(HasNSW);
2232 return Sub;
2233 }
2234
2235 {
2236 // (X + Z) - (Y + Z) --> (X - Y)
2237 // This is done in other passes, but we want to be able to consume this
2238 // pattern in InstCombine so we can generate it without creating infinite
2239 // loops.
2240 if (match(Op0, m_Add(m_Value(X), m_Value(Z))) &&
2241 match(Op1, m_c_Add(m_Value(Y), m_Specific(Z))))
2242 return BinaryOperator::CreateSub(X, Y);
2243
2244 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2245 Constant *CX, *CY;
2246 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_ImmConstant(CX)))) &&
2247 match(Op1, m_OneUse(m_Add(m_Value(Y), m_ImmConstant(CY))))) {
2248 Value *OpsSub = Builder.CreateSub(X, Y);
2249 Constant *ConstsSub = ConstantExpr::getSub(CX, CY);
2250 return BinaryOperator::CreateAdd(OpsSub, ConstsSub);
2251 }
2252 }
2253
2254 // (~X) - (~Y) --> Y - X
2255 {
2256 // Need to ensure we can consume at least one of the `not` instructions,
2257 // otherwise this can inf loop.
2258 bool ConsumesOp0, ConsumesOp1;
2259 if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
2260 isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
2261 (ConsumesOp0 || ConsumesOp1)) {
2262 Value *NotOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
2263 Value *NotOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
2264 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2265 "isFreeToInvert desynced with getFreelyInverted");
2266 return BinaryOperator::CreateSub(NotOp1, NotOp0);
2267 }
2268 }
2269
2270 auto m_AddRdx = [](Value *&Vec) {
2271 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2272 };
2273 Value *V0, *V1;
2274 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2275 V0->getType() == V1->getType()) {
2276 // Difference of sums is sum of differences:
2277 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2278 Value *Sub = Builder.CreateSub(V0, V1);
2279 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2280 {Sub->getType()}, {Sub});
2281 return replaceInstUsesWith(I, Rdx);
2282 }
2283
2284 if (Constant *C = dyn_cast<Constant>(Op0)) {
2285 Value *X;
2286 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2287 // C - (zext bool) --> bool ? C - 1 : C
2289 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2290 // C - (sext bool) --> bool ? C + 1 : C
2292
2293 // C - ~X == X + (1+C)
2294 if (match(Op1, m_Not(m_Value(X))))
2295 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2296
2297 // Try to fold constant sub into select arguments.
2298 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2299 if (Instruction *R = FoldOpIntoSelect(I, SI))
2300 return R;
2301
2302 // Try to fold constant sub into PHI values.
2303 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2304 if (Instruction *R = foldOpIntoPhi(I, PN))
2305 return R;
2306
2307 Constant *C2;
2308
2309 // C-(C2-X) --> X+(C-C2)
2310 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2311 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2312 }
2313
2314 const APInt *Op0C;
2315 if (match(Op0, m_APInt(Op0C))) {
2316 if (Op0C->isMask()) {
2317 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2318 // zero.
2319 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
2320 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2321 return BinaryOperator::CreateXor(Op1, Op0);
2322 }
2323
2324 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2325 // (C3 - ((C2 & C3) - 1)) is pow2
2326 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2327 // C2 is negative pow2 || sub nuw
2328 const APInt *C2, *C3;
2329 BinaryOperator *InnerSub;
2330 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2331 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2332 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2333 APInt C2AndC3 = *C2 & *C3;
2334 APInt C2AndC3Minus1 = C2AndC3 - 1;
2335 APInt C2AddC3 = *C2 + *C3;
2336 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2337 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2338 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2339 return BinaryOperator::CreateAdd(
2340 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2341 }
2342 }
2343 }
2344
2345 {
2346 Value *Y;
2347 // X-(X+Y) == -Y X-(Y+X) == -Y
2348 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2350
2351 // (X-Y)-X == -Y
2352 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2354 }
2355
2356 // (sub (or A, B) (and A, B)) --> (xor A, B)
2357 {
2358 Value *A, *B;
2359 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2360 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2361 return BinaryOperator::CreateXor(A, B);
2362 }
2363
2364 // (sub (add A, B) (or A, B)) --> (and A, B)
2365 {
2366 Value *A, *B;
2367 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2368 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2369 return BinaryOperator::CreateAnd(A, B);
2370 }
2371
2372 // (sub (add A, B) (and A, B)) --> (or A, B)
2373 {
2374 Value *A, *B;
2375 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2377 return BinaryOperator::CreateOr(A, B);
2378 }
2379
2380 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2381 {
2382 Value *A, *B;
2383 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2384 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2385 (Op0->hasOneUse() || Op1->hasOneUse()))
2387 }
2388
2389 // (sub (or A, B), (xor A, B)) --> (and A, B)
2390 {
2391 Value *A, *B;
2392 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2393 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2394 return BinaryOperator::CreateAnd(A, B);
2395 }
2396
2397 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2398 {
2399 Value *A, *B;
2400 if (match(Op0, m_Xor(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 {
2407 Value *Y;
2408 // ((X | Y) - X) --> (~X & Y)
2409 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2410 return BinaryOperator::CreateAnd(
2411 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2412 }
2413
2414 {
2415 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2416 Value *X;
2417 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2418 m_OneUse(m_Neg(m_Value(X))))))) {
2420 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2421 }
2422 }
2423
2424 {
2425 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2426 Constant *C;
2427 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2430 }
2431 }
2432
2434 return R;
2435
2436 {
2437 // If we have a subtraction between some value and a select between
2438 // said value and something else, sink subtraction into select hands, i.e.:
2439 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2440 // ->
2441 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2442 // or
2443 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2444 // ->
2445 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2446 // This will result in select between new subtraction and 0.
2447 auto SinkSubIntoSelect =
2448 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2449 auto SubBuilder) -> Instruction * {
2450 Value *Cond, *TrueVal, *FalseVal;
2451 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2452 m_Value(FalseVal)))))
2453 return nullptr;
2454 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2455 return nullptr;
2456 // While it is really tempting to just create two subtractions and let
2457 // InstCombine fold one of those to 0, it isn't possible to do so
2458 // because of worklist visitation order. So ugly it is.
2459 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2460 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2461 Constant *Zero = Constant::getNullValue(Ty);
2462 SelectInst *NewSel =
2463 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2464 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2465 // Preserve prof metadata if any.
2466 NewSel->copyMetadata(cast<Instruction>(*Select));
2467 return NewSel;
2468 };
2469 if (Instruction *NewSel = SinkSubIntoSelect(
2470 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2471 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2472 return Builder->CreateSub(OtherHandOfSelect,
2473 /*OtherHandOfSub=*/Op1);
2474 }))
2475 return NewSel;
2476 if (Instruction *NewSel = SinkSubIntoSelect(
2477 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2478 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2479 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2480 OtherHandOfSelect);
2481 }))
2482 return NewSel;
2483 }
2484
2485 // (X - (X & Y)) --> (X & ~Y)
2486 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2487 (Op1->hasOneUse() || isa<Constant>(Y)))
2488 return BinaryOperator::CreateAnd(
2489 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2490
2491 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2492 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2493 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2494 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2495 // As long as Y is freely invertible, this will be neutral or a win.
2496 // Note: We don't generate the inverse max/min, just create the 'not' of
2497 // it and let other folds do the rest.
2498 if (match(Op0, m_Not(m_Value(X))) &&
2499 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2500 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2501 Value *Not = Builder.CreateNot(Op1);
2502 return BinaryOperator::CreateSub(Not, X);
2503 }
2504 if (match(Op1, m_Not(m_Value(X))) &&
2505 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2506 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2507 Value *Not = Builder.CreateNot(Op0);
2508 return BinaryOperator::CreateSub(X, Not);
2509 }
2510
2511 // Optimize pointer differences into the same array into a size. Consider:
2512 // &A[10] - &A[0]: we should compile this to "10".
2513 Value *LHSOp, *RHSOp;
2514 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2515 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2516 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2517 I.hasNoUnsignedWrap()))
2518 return replaceInstUsesWith(I, Res);
2519
2520 // trunc(p)-trunc(q) -> trunc(p-q)
2521 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2522 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2523 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2524 /* IsNUW */ false))
2525 return replaceInstUsesWith(I, Res);
2526
2527 // Canonicalize a shifty way to code absolute value to the common pattern.
2528 // There are 2 potential commuted variants.
2529 // We're relying on the fact that we only do this transform when the shift has
2530 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2531 // instructions).
2532 Value *A;
2533 const APInt *ShAmt;
2534 Type *Ty = I.getType();
2535 unsigned BitWidth = Ty->getScalarSizeInBits();
2536 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2537 Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2538 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2539 // B = ashr i32 A, 31 ; smear the sign bit
2540 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2541 // --> (A < 0) ? -A : A
2542 Value *IsNeg = Builder.CreateIsNeg(A);
2543 // Copy the nuw/nsw flags from the sub to the negate.
2544 Value *NegA = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2545 I.hasNoSignedWrap());
2546 return SelectInst::Create(IsNeg, NegA, A);
2547 }
2548
2549 // If we are subtracting a low-bit masked subset of some value from an add
2550 // of that same value with no low bits changed, that is clearing some low bits
2551 // of the sum:
2552 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2553 const APInt *AddC, *AndC;
2554 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2555 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2556 unsigned Cttz = AddC->countr_zero();
2557 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2558 if ((HighMask & *AndC).isZero())
2559 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2560 }
2561
2562 if (Instruction *V =
2564 return V;
2565
2566 // X - usub.sat(X, Y) => umin(X, Y)
2567 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2568 m_Value(Y)))))
2569 return replaceInstUsesWith(
2570 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2571
2572 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2573 // TODO: The one-use restriction is not strictly necessary, but it may
2574 // require improving other pattern matching and/or codegen.
2575 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2576 return replaceInstUsesWith(
2577 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2578
2579 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2580 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2581 return replaceInstUsesWith(
2582 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2583
2584 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2585 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2586 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2587 return BinaryOperator::CreateNeg(USub);
2588 }
2589
2590 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2591 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2592 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2593 return BinaryOperator::CreateNeg(USub);
2594 }
2595
2596 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2597 if (match(Op0, m_SpecificInt(BitWidth)) &&
2598 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2599 return replaceInstUsesWith(
2600 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2601 {Builder.CreateNot(X)}));
2602
2603 // Reduce multiplies for difference-of-squares by factoring:
2604 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2605 if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2606 match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2607 auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2608 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2609 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2610 OBO1->hasNoSignedWrap() && BitWidth > 2;
2611 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2612 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2613 Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2614 Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2615 Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2616 return replaceInstUsesWith(I, Mul);
2617 }
2618
2619 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2620 if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2622 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2623 Value *Sub =
2624 Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2625 Value *Call =
2626 Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2627 return replaceInstUsesWith(I, Call);
2628 }
2629 }
2630
2632 return Res;
2633
2634 return TryToNarrowDeduceFlags();
2635}
2636
2637/// This eliminates floating-point negation in either 'fneg(X)' or
2638/// 'fsub(-0.0, X)' form by combining into a constant operand.
2640 // This is limited with one-use because fneg is assumed better for
2641 // reassociation and cheaper in codegen than fmul/fdiv.
2642 // TODO: Should the m_OneUse restriction be removed?
2643 Instruction *FNegOp;
2644 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2645 return nullptr;
2646
2647 Value *X;
2648 Constant *C;
2649
2650 // Fold negation into constant operand.
2651 // -(X * C) --> X * (-C)
2652 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2653 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2654 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2655 // -(X / C) --> X / (-C)
2656 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2657 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2658 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2659 // -(C / X) --> (-C) / X
2660 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2661 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2663
2664 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2665 // not apply to the fdiv. Everything else propagates from the fneg.
2666 // TODO: We could propagate nsz/ninf from fdiv alone?
2667 FastMathFlags FMF = I.getFastMathFlags();
2668 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2669 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2670 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2671 return FDiv;
2672 }
2673 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2674 // -(X + C) --> -X + -C --> -C - X
2675 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2676 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2677 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2678
2679 return nullptr;
2680}
2681
2682Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2683 Instruction &FMFSource) {
2684 Value *X, *Y;
2685 if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2686 return cast<Instruction>(Builder.CreateFMulFMF(
2687 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2688 }
2689
2690 if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2691 return cast<Instruction>(Builder.CreateFDivFMF(
2692 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2693 }
2694
2695 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2696 // Make sure to preserve flags and metadata on the call.
2697 if (II->getIntrinsicID() == Intrinsic::ldexp) {
2698 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2701
2703 II->getCalledFunction(),
2704 {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2705 New->copyMetadata(*II);
2706 return New;
2707 }
2708 }
2709
2710 return nullptr;
2711}
2712
2714 Value *Op = I.getOperand(0);
2715
2716 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2717 getSimplifyQuery().getWithInstruction(&I)))
2718 return replaceInstUsesWith(I, V);
2719
2721 return X;
2722
2723 Value *X, *Y;
2724
2725 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2726 if (I.hasNoSignedZeros() &&
2729
2730 Value *OneUse;
2731 if (!match(Op, m_OneUse(m_Value(OneUse))))
2732 return nullptr;
2733
2734 if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2735 return replaceInstUsesWith(I, R);
2736
2737 // Try to eliminate fneg if at least 1 arm of the select is negated.
2738 Value *Cond;
2739 if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2740 // Unlike most transforms, this one is not safe to propagate nsz unless
2741 // it is present on the original select. We union the flags from the select
2742 // and fneg and then remove nsz if needed.
2743 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2744 S->copyFastMathFlags(&I);
2745 if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2746 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2747 S->setFastMathFlags(FMF);
2748 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2749 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2750 S->setHasNoSignedZeros(false);
2751 }
2752 };
2753 // -(Cond ? -P : Y) --> Cond ? P : -Y
2754 Value *P;
2755 if (match(X, m_FNeg(m_Value(P)))) {
2756 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2757 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2758 propagateSelectFMF(NewSel, P == Y);
2759 return NewSel;
2760 }
2761 // -(Cond ? X : -P) --> Cond ? -X : P
2762 if (match(Y, m_FNeg(m_Value(P)))) {
2763 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2764 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2765 propagateSelectFMF(NewSel, P == X);
2766 return NewSel;
2767 }
2768 }
2769
2770 // fneg (copysign x, y) -> copysign x, (fneg y)
2771 if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2772 // The source copysign has an additional value input, so we can't propagate
2773 // flags the copysign doesn't also have.
2774 FastMathFlags FMF = I.getFastMathFlags();
2775 FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2776
2779
2780 Value *NegY = Builder.CreateFNeg(Y);
2781 Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2782 return replaceInstUsesWith(I, NewCopySign);
2783 }
2784
2785 return nullptr;
2786}
2787
2789 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2790 I.getFastMathFlags(),
2791 getSimplifyQuery().getWithInstruction(&I)))
2792 return replaceInstUsesWith(I, V);
2793
2795 return X;
2796
2798 return Phi;
2799
2800 // Subtraction from -0.0 is the canonical form of fneg.
2801 // fsub -0.0, X ==> fneg X
2802 // fsub nsz 0.0, X ==> fneg nsz X
2803 //
2804 // FIXME This matcher does not respect FTZ or DAZ yet:
2805 // fsub -0.0, Denorm ==> +-0
2806 // fneg Denorm ==> -Denorm
2807 Value *Op;
2808 if (match(&I, m_FNeg(m_Value(Op))))
2810
2812 return X;
2813
2814 Value *X, *Y;
2815 Constant *C;
2816
2817 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2818 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2819 // Canonicalize to fadd to make analysis easier.
2820 // This can also help codegen because fadd is commutative.
2821 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2822 // killed later. We still limit that particular transform with 'hasOneUse'
2823 // because an fneg is assumed better/cheaper than a generic fsub.
2824 if (I.hasNoSignedZeros() || cannotBeNegativeZero(Op0, SQ.DL, SQ.TLI)) {
2825 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2826 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2827 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2828 }
2829 }
2830
2831 // (-X) - Op1 --> -(X + Op1)
2832 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2833 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2834 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2836 }
2837
2838 if (isa<Constant>(Op0))
2839 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2840 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2841 return NV;
2842
2843 // X - C --> X + (-C)
2844 // But don't transform constant expressions because there's an inverse fold
2845 // for X + (-Y) --> X - Y.
2846 if (match(Op1, m_ImmConstant(C)))
2847 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2848 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2849
2850 // X - (-Y) --> X + Y
2851 if (match(Op1, m_FNeg(m_Value(Y))))
2852 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2853
2854 // Similar to above, but look through a cast of the negated value:
2855 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2856 Type *Ty = I.getType();
2857 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2859
2860 // X - (fpext(-Y)) --> X + fpext(Y)
2861 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2863
2864 // Similar to above, but look through fmul/fdiv of the negated value:
2865 // Op0 - (-X * Y) --> Op0 + (X * Y)
2866 // Op0 - (Y * -X) --> Op0 + (X * Y)
2867 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2869 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2870 }
2871 // Op0 - (-X / Y) --> Op0 + (X / Y)
2872 // Op0 - (X / -Y) --> Op0 + (X / Y)
2873 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2874 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2875 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2876 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2877 }
2878
2879 // Handle special cases for FSub with selects feeding the operation
2880 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2881 return replaceInstUsesWith(I, V);
2882
2883 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2884 // (Y - X) - Y --> -X
2885 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2887
2888 // Y - (X + Y) --> -X
2889 // Y - (Y + X) --> -X
2890 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2892
2893 // (X * C) - X --> X * (C - 1.0)
2894 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2896 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2897 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2898 }
2899 // X - (X * C) --> X * (1.0 - C)
2900 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2902 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2903 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2904 }
2905
2906 // Reassociate fsub/fadd sequences to create more fadd instructions and
2907 // reduce dependency chains:
2908 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2909 Value *Z;
2911 m_Value(Z))))) {
2912 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2913 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2914 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2915 }
2916
2917 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2918 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2919 m_Value(Vec)));
2920 };
2921 Value *A0, *A1, *V0, *V1;
2922 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2923 V0->getType() == V1->getType()) {
2924 // Difference of sums is sum of differences:
2925 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2926 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2927 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2928 {Sub->getType()}, {A0, Sub}, &I);
2929 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2930 }
2931
2933 return F;
2934
2935 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2936 // functionality has been subsumed by simple pattern matching here and in
2937 // InstSimplify. We should let a dedicated reassociation pass handle more
2938 // complex pattern matching and remove this from InstCombine.
2939 if (Value *V = FAddCombine(Builder).simplify(&I))
2940 return replaceInstUsesWith(I, V);
2941
2942 // (X - Y) - Op1 --> X - (Y + Op1)
2943 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2944 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2946 }
2947 }
2948
2949 return nullptr;
2950}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static bool isConstant(const MachineInstr &MI)
amdgpu AMDGPU Register Bank Select
This file declares a class to represent arbitrary precision floating point values and provide a varie...
This file implements a class to represent arbitrary precision integral constant values and operations...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
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:526
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
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
static constexpr uint32_t Opcode
Definition: aarch32.h:200
const fltSemantics & getSemantics() const
Definition: APFloat.h:1303
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1060
Class for arbitrary precision integers.
Definition: APInt.h:76
APInt umul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1977
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 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:1433
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:307
int32_t exactLogBase2() const
Definition: APInt.h:1718
unsigned countr_zero() const
Count the number of trailing zero bits.
Definition: APInt.h:1583
unsigned countl_zero() const
The APInt version of std::countl_zero.
Definition: APInt.h:1542
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:1696
APInt smul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1966
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:1229
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:1209
static BinaryOperator * CreateFDivFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:271
static BinaryOperator * CreateNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static BinaryOperator * CreateNot(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
static BinaryOperator * CreateWithCopiedFlags(BinaryOps Opc, Value *V1, Value *V2, Value *CopyO, const Twine &Name="", Instruction *InsertBefore=nullptr)
Definition: InstrTypes.h:248
static BinaryOperator * CreateFMulFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:266
static BinaryOperator * CreateFSubFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:261
static BinaryOperator * CreateFAddFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:256
This class represents a function call, abstracting a target machine's calling convention.
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass's ...
static CastInst * CreateTruncOrBitCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a Trunc or BitCast cast instruction.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:748
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:771
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:775
@ ICMP_EQ
equal
Definition: InstrTypes.h:769
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2460
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2453
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:260
const APFloat & getValueAPF() const
Definition: Constants.h:296
static Constant * get(Type *Ty, double V)
This returns a ConstantFP, or a vector containing a splat of a ConstantFP, for the specified value in...
Definition: Constants.cpp:927
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:300
static Constant * get(Type *Ty, uint64_t V, bool IsSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:888
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:403
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:356
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
Convenience struct for specifying and reasoning about fast-math flags.
Definition: FMF.h:20
bool noSignedZeros() const
Definition: FMF.h:68
bool noInfs() const
Definition: FMF.h:67
bool isInBounds() const
Test whether this is an inbounds GEP, as defined by LangRef.html.
Definition: Operator.h:387
unsigned countNonConstantIndices() const
Definition: Operator.h:467
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:1546
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1720
Value * CreateSRem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1409
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:1600
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:1573
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:1627
Value * CreateFPTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2079
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:460
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:930
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:1744
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2017
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2537
Value * CreateNSWAdd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1335
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:305
CallInst * CreateCopySign(Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create call to the copysign intrinsic.
Definition: IRBuilder.h:1021
Value * CreateNUWAdd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1339
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1753
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:145
Value * CreateIsNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg < 0.
Definition: IRBuilder.h:2532
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1343
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1415
CallInst * CreateBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 2 operands which is mangled on the first type.
Definition: IRBuilder.cpp:921
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="", bool IsNonNeg=false)
Definition: IRBuilder.h:2005
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1474
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1326
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2527
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1496
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1665
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:2174
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2390
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2088
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1518
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1734
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1403
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1360
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.
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)
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:420
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
Definition: InstCombiner.h:214
const SimplifyQuery SQ
Definition: InstCombiner.h:75
const DataLayout & DL
Definition: InstCombiner.h:74
unsigned ComputeNumSignBits(const Value *Op, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:486
static bool isFreeToInvert(Value *V, bool WillInvertAllUses, bool &DoesConsume)
Return true if the specified value is free to invert (apply ~ to).
Definition: InstCombiner.h:266
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:444
static 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.
Definition: InstCombiner.h:172
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
Definition: InstCombiner.h:465
BuilderTy & Builder
Definition: InstCombiner.h:59
bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:481
const SimplifyQuery & getSimplifyQuery() const
Definition: InstCombiner.h:376
static Value * getFreelyInverted(Value *V, bool WillInvertAllUses, BuilderTy *Builder, bool &DoesConsume)
Definition: InstCombiner.h:247
static Constant * AddOne(Constant *C)
Add one to a Constant.
Definition: InstCombiner.h:209
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:435
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:47
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:75
bool hasNoSignedWrap() const
Test whether this operation is known to never undergo signed overflow, aka the nsw property.
Definition: Operator.h:105
bool hasNoUnsignedWrap() const
Test whether this operation is known to never undergo unsigned overflow, aka the nuw property.
Definition: Operator.h:99
This class represents a cast from signed integer to floating point.
This class represents the LLVM 'select' instruction.
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr="", Instruction *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:1200
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
unsigned getIntegerBitWidth() const
const fltSemantics & getFltSemantics() const
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:234
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return 'this'.
Definition: Type.h:348
static UnaryOperator * CreateFNegFMF(Value *Op, Instruction *FMFSource, const Twine &Name="", Instruction *InsertBefore=nullptr)
Definition: InstrTypes.h:163
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
bool hasOneUse() const
Return true if there is exactly one use of this value.
Definition: Value.h:434
bool hasNUsesOrMore(unsigned N) const
Return true if this value has N uses or more.
Definition: Value.cpp:153
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs and address space casts.
Definition: Value.cpp:693
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
This class represents zero extension of integer types.
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
Function * getDeclaration(Module *M, ID id, ArrayRef< Type * > Tys=std::nullopt)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1444
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:461
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
specific_intval< false > m_SpecificInt(APInt V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:862
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:982
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:84
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.
OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWAdd(const LHS &L, const RHS &R)
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:552
BinaryOp_match< LHS, RHS, Instruction::URem > m_URem(const LHS &L, const RHS &R)
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:144
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.
BinaryOp_match< LHS, RHS, Instruction::Xor > m_Xor(const LHS &L, const RHS &R)
CastInst_match< OpTy, Instruction::FPTrunc > m_FPTrunc(const OpTy &Op)
OverflowingBinaryOp_match< LHS, RHS, Instruction::Sub, OverflowingBinaryOperator::NoSignedWrap > m_NSWSub(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::FMul > m_FMul(const LHS &L, const RHS &R)
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:724
cstfp_pred_ty< is_any_zero_fp > m_AnyZeroFP()
Match a floating-point negative zero or positive zero.
Definition: PatternMatch.h:672
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:780
cst_pred_ty< is_one > m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:525
ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)
Matches SelectInst.
specific_fpval m_SpecificFP(double V)
Match a specific floating point value or vector with all elements equal to the value.
Definition: PatternMatch.h:823
match_combine_and< LTy, RTy > m_CombineAnd(const LTy &L, const RTy &R)
Combine two pattern matchers matching L && R.
Definition: PatternMatch.h:224
CastInst_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
CastOperator_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
CastInst_match< OpTy, Instruction::FPExt > m_FPExt(const OpTy &Op)
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)
Definition: PatternMatch.h:988
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:798
cst_pred_ty< is_zero_int > m_ZeroInt()
Match an integer 0 or a vector with all elements equal to 0.
Definition: PatternMatch.h:532
CastInst_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
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:759
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.
BinaryOp_match< LHS, RHS, Instruction::UDiv > m_UDiv(const LHS &L, const RHS &R)
MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty > m_UMax(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:560
specific_fpval m_FPOne()
Match a float 1.0 or vector with all elements equal to 1.0.
Definition: PatternMatch.h:826
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty, true > m_c_UMin(const LHS &L, const RHS &R)
Matches a UMin with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)
Matches a Add with LHS and RHS in either order.
MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty, true > m_c_SMax(const LHS &L, const RHS &R)
Matches an SMax with LHS and RHS in either order.
match_combine_or< match_combine_or< MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty, true >, MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty, true > >, match_combine_or< MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty, true >, MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty, true > > > m_c_MaxOrMin(const LHS &L, const RHS &R)
match_combine_or< CastOperator_match< OpTy, Instruction::Trunc >, OpTy > m_TruncOrSelf(const OpTy &Op)
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
specific_intval< true > m_SpecificIntAllowUndef(APInt V)
Definition: PatternMatch.h:870
OverflowingBinaryOp_match< LHS, RHS, Instruction::Sub, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWSub(const LHS &L, const RHS &R)
MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty > m_SMax(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:278
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:76
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoSignedWrap > m_NSWAdd(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
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)
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:295
match_combine_or< CastInst_match< OpTy, Instruction::SExt >, OpTy > m_SExtOrSelf(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)
match_combine_or< CastInst_match< OpTy, Instruction::ZExt >, OpTy > m_ZExtOrSelf(const OpTy &Op)
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:545
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.
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)
Definition: PatternMatch.h:994
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:218
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:606
@ 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:440
bool isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Return true if the given value is known to be non-zero when defined.
constexpr bool isInt(int64_t x)
Checks if an integer fits into the given bit width.
Definition: MathExtras.h:151
std::string & operator+=(std::string &buffer, StringRef string)
Definition: StringRef.h:907
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.
Constant * ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op, const DataLayout &DL)
Attempt to constant fold a unary operation with the specified operand.
bool cannotBeNegativeZero(const Value *V, const DataLayout &DL, const TargetLibraryInfo *TLI=nullptr, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Return true if we can prove that the specified FP value is never equal to -0.0.
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:1740
Constant * ConstantFoldCastOperand(unsigned Opcode, Constant *C, Type *DestTy, const DataLayout &DL)
Attempt to constant fold a cast with the specified operand.
Constant * ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, Constant *RHS, const DataLayout &DL)
Attempt to constant fold a binary operation with the specified operands.
@ Or
Bitwise or logical OR of integers.
@ 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.
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
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:860
static unsigned int semanticsPrecision(const fltSemantics &)
Definition: APFloat.cpp:292
A suitably aligned and sized character array member which can hold elements of any type.
Definition: AlignOf.h:27
const DataLayout & DL
Definition: SimplifyQuery.h:60
const Instruction * CxtI
Definition: SimplifyQuery.h:64
const DominatorTree * DT
Definition: SimplifyQuery.h:62
SimplifyQuery getWithInstruction(const Instruction *I) const
Definition: SimplifyQuery.h:94
AssumptionCache * AC
Definition: SimplifyQuery.h:63
const TargetLibraryInfo * TLI
Definition: SimplifyQuery.h:61