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 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
834 Constant *NewC = ConstantExpr::getAdd(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 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
841 Constant *NewC = ConstantExpr::getAdd(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))) &&
907 haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
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. The canonical pattern for that is an "ugt" compare with SMIN:
1238 // sext (icmp ugt (X & (DivC - 1)), SMIN)
1239 const APInt *MaskC;
1241 if (!match(Add.getOperand(1),
1242 m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1243 m_SignMask()))) ||
1244 Pred != ICmpInst::ICMP_UGT)
1245 return nullptr;
1246
1247 APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1248 if (*MaskC != (SMin | (*DivC - 1)))
1249 return nullptr;
1250
1251 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1252 return BinaryOperator::CreateAShr(
1253 X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1254}
1255
1258 BinaryOperator &I) {
1259 assert((I.getOpcode() == Instruction::Add ||
1260 I.getOpcode() == Instruction::Or ||
1261 I.getOpcode() == Instruction::Sub) &&
1262 "Expecting add/or/sub instruction");
1263
1264 // We have a subtraction/addition between a (potentially truncated) *logical*
1265 // right-shift of X and a "select".
1266 Value *X, *Select;
1267 Instruction *LowBitsToSkip, *Extract;
1269 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1270 m_Instruction(Extract))),
1271 m_Value(Select))))
1272 return nullptr;
1273
1274 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1275 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1276 return nullptr;
1277
1278 Type *XTy = X->getType();
1279 bool HadTrunc = I.getType() != XTy;
1280
1281 // If there was a truncation of extracted value, then we'll need to produce
1282 // one extra instruction, so we need to ensure one instruction will go away.
1283 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1284 return nullptr;
1285
1286 // Extraction should extract high NBits bits, with shift amount calculated as:
1287 // low bits to skip = shift bitwidth - high bits to extract
1288 // The shift amount itself may be extended, and we need to look past zero-ext
1289 // when matching NBits, that will matter for matching later.
1290 Constant *C;
1291 Value *NBits;
1292 if (!match(
1293 LowBitsToSkip,
1296 APInt(C->getType()->getScalarSizeInBits(),
1297 X->getType()->getScalarSizeInBits()))))
1298 return nullptr;
1299
1300 // Sign-extending value can be zero-extended if we `sub`tract it,
1301 // or sign-extended otherwise.
1302 auto SkipExtInMagic = [&I](Value *&V) {
1303 if (I.getOpcode() == Instruction::Sub)
1304 match(V, m_ZExtOrSelf(m_Value(V)));
1305 else
1306 match(V, m_SExtOrSelf(m_Value(V)));
1307 };
1308
1309 // Now, finally validate the sign-extending magic.
1310 // `select` itself may be appropriately extended, look past that.
1311 SkipExtInMagic(Select);
1312
1314 const APInt *Thr;
1315 Value *SignExtendingValue, *Zero;
1316 bool ShouldSignext;
1317 // It must be a select between two values we will later establish to be a
1318 // sign-extending value and a zero constant. The condition guarding the
1319 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1320 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1321 m_Value(SignExtendingValue), m_Value(Zero))) ||
1322 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1323 return nullptr;
1324
1325 // icmp-select pair is commutative.
1326 if (!ShouldSignext)
1327 std::swap(SignExtendingValue, Zero);
1328
1329 // If we should not perform sign-extension then we must add/or/subtract zero.
1330 if (!match(Zero, m_Zero()))
1331 return nullptr;
1332 // Otherwise, it should be some constant, left-shifted by the same NBits we
1333 // had in `lshr`. Said left-shift can also be appropriately extended.
1334 // Again, we must look past zero-ext when looking for NBits.
1335 SkipExtInMagic(SignExtendingValue);
1336 Constant *SignExtendingValueBaseConstant;
1337 if (!match(SignExtendingValue,
1338 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1339 m_ZExtOrSelf(m_Specific(NBits)))))
1340 return nullptr;
1341 // If we `sub`, then the constant should be one, else it should be all-ones.
1342 if (I.getOpcode() == Instruction::Sub
1343 ? !match(SignExtendingValueBaseConstant, m_One())
1344 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1345 return nullptr;
1346
1347 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1348 Extract->getName() + ".sext");
1349 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1350 if (!HadTrunc)
1351 return NewAShr;
1352
1353 Builder.Insert(NewAShr);
1354 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1355}
1356
1357/// This is a specialization of a more general transform from
1358/// foldUsingDistributiveLaws. If that code can be made to work optimally
1359/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1361 InstCombiner::BuilderTy &Builder) {
1362 // TODO: Also handle mul by doubling the shift amount?
1363 assert((I.getOpcode() == Instruction::Add ||
1364 I.getOpcode() == Instruction::Sub) &&
1365 "Expected add/sub");
1366 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1367 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1368 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1369 return nullptr;
1370
1371 Value *X, *Y, *ShAmt;
1372 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1373 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1374 return nullptr;
1375
1376 // No-wrap propagates only when all ops have no-wrap.
1377 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1378 Op1->hasNoSignedWrap();
1379 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1380 Op1->hasNoUnsignedWrap();
1381
1382 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1383 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1384 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1385 NewI->setHasNoSignedWrap(HasNSW);
1386 NewI->setHasNoUnsignedWrap(HasNUW);
1387 }
1388 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1389 NewShl->setHasNoSignedWrap(HasNSW);
1390 NewShl->setHasNoUnsignedWrap(HasNUW);
1391 return NewShl;
1392}
1393
1394/// Reduce a sequence of masked half-width multiplies to a single multiply.
1395/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1397 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1398 // Skip the odd bitwidth types.
1399 if ((BitWidth & 0x1))
1400 return nullptr;
1401
1402 unsigned HalfBits = BitWidth >> 1;
1403 APInt HalfMask = APInt::getMaxValue(HalfBits);
1404
1405 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1406 Value *XLo, *YLo;
1407 Value *CrossSum;
1408 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1409 m_Mul(m_Value(YLo), m_Value(XLo)))))
1410 return nullptr;
1411
1412 // XLo = X & HalfMask
1413 // YLo = Y & HalfMask
1414 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1415 // to enhance robustness
1416 Value *X, *Y;
1417 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1418 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1419 return nullptr;
1420
1421 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1422 // X' can be either X or XLo in the pattern (and the same for Y')
1423 if (match(CrossSum,
1428 return BinaryOperator::CreateMul(X, Y);
1429
1430 return nullptr;
1431}
1432
1434 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1435 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1437 return replaceInstUsesWith(I, V);
1438
1440 return &I;
1441
1443 return X;
1444
1446 return Phi;
1447
1448 // (A*B)+(A*C) -> A*(B+C) etc
1450 return replaceInstUsesWith(I, V);
1451
1452 if (Instruction *R = foldBoxMultiply(I))
1453 return R;
1454
1456 return R;
1457
1459 return X;
1460
1462 return X;
1463
1465 return R;
1466
1468 return R;
1469
1470 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1471 Type *Ty = I.getType();
1472 if (Ty->isIntOrIntVectorTy(1))
1473 return BinaryOperator::CreateXor(LHS, RHS);
1474
1475 // X + X --> X << 1
1476 if (LHS == RHS) {
1477 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1478 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1479 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1480 return Shl;
1481 }
1482
1483 Value *A, *B;
1484 if (match(LHS, m_Neg(m_Value(A)))) {
1485 // -A + -B --> -(A + B)
1486 if (match(RHS, m_Neg(m_Value(B))))
1488
1489 // -A + B --> B - A
1490 return BinaryOperator::CreateSub(RHS, A);
1491 }
1492
1493 // A + -B --> A - B
1494 if (match(RHS, m_Neg(m_Value(B))))
1495 return BinaryOperator::CreateSub(LHS, B);
1496
1498 return replaceInstUsesWith(I, V);
1499
1500 // (A + 1) + ~B --> A - B
1501 // ~B + (A + 1) --> A - B
1502 // (~B + A) + 1 --> A - B
1503 // (A + ~B) + 1 --> A - B
1504 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1506 return BinaryOperator::CreateSub(A, B);
1507
1508 // (A + RHS) + RHS --> A + (RHS << 1)
1510 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1511
1512 // LHS + (A + LHS) --> A + (LHS << 1)
1514 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1515
1516 {
1517 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1518 Constant *C1, *C2;
1519 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1520 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1521 (LHS->hasOneUse() || RHS->hasOneUse())) {
1522 Value *Sub = Builder.CreateSub(A, B);
1523 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1524 }
1525
1526 // Canonicalize a constant sub operand as an add operand for better folding:
1527 // (C1 - A) + B --> (B - A) + C1
1529 m_Value(B)))) {
1530 Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1531 return BinaryOperator::CreateAdd(Sub, C1);
1532 }
1533 }
1534
1535 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1537
1538 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1539 const APInt *C1, *C2;
1540 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1541 APInt one(C2->getBitWidth(), 1);
1542 APInt minusC1 = -(*C1);
1543 if (minusC1 == (one << *C2)) {
1544 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1545 return BinaryOperator::CreateSRem(RHS, NewRHS);
1546 }
1547 }
1548
1549 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1550 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1551 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1552 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1553 return BinaryOperator::CreateAnd(A, NewMask);
1554 }
1555
1556 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1557 if ((match(RHS, m_ZExt(m_Value(A))) &&
1559 (match(LHS, m_ZExt(m_Value(A))) &&
1561 return new ZExtInst(B, LHS->getType());
1562
1563 // zext(A) + sext(A) --> 0 if A is i1
1565 A->getType()->isIntOrIntVectorTy(1))
1566 return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1567
1568 // A+B --> A|B iff A and B have no bits set in common.
1569 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1570 return BinaryOperator::CreateOr(LHS, RHS);
1571
1572 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1573 return Ext;
1574
1575 // (add (xor A, B) (and A, B)) --> (or A, B)
1576 // (add (and A, B) (xor A, B)) --> (or A, B)
1577 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1579 return BinaryOperator::CreateOr(A, B);
1580
1581 // (add (or A, B) (and A, B)) --> (add A, B)
1582 // (add (and A, B) (or A, B)) --> (add A, B)
1583 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1585 // Replacing operands in-place to preserve nuw/nsw flags.
1586 replaceOperand(I, 0, A);
1587 replaceOperand(I, 1, B);
1588 return &I;
1589 }
1590
1591 // (add A (or A, -A)) --> (and (add A, -1) A)
1592 // (add A (or -A, A)) --> (and (add A, -1) A)
1593 // (add (or A, -A) A) --> (and (add A, -1) A)
1594 // (add (or -A, A) A) --> (and (add A, -1) A)
1596 m_Deferred(A)))))) {
1597 Value *Add =
1599 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1600 return BinaryOperator::CreateAnd(Add, A);
1601 }
1602
1603 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1604 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1605 if (match(&I,
1607 m_AllOnes()))) {
1609 Value *Dec = Builder.CreateAdd(A, AllOnes);
1610 Value *Not = Builder.CreateXor(A, AllOnes);
1611 return BinaryOperator::CreateAnd(Dec, Not);
1612 }
1613
1614 // Disguised reassociation/factorization:
1615 // ~(A * C1) + A
1616 // ((A * -C1) - 1) + A
1617 // ((A * -C1) + A) - 1
1618 // (A * (1 - C1)) - 1
1619 if (match(&I,
1621 m_Deferred(A)))) {
1622 Type *Ty = I.getType();
1623 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1624 Value *NewMul = Builder.CreateMul(A, NewMulC);
1625 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1626 }
1627
1628 // (A * -2**C) + B --> B - (A << C)
1629 const APInt *NegPow2C;
1630 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1631 m_Value(B)))) {
1632 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1633 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1634 return BinaryOperator::CreateSub(B, Shl);
1635 }
1636
1637 // Canonicalize signum variant that ends in add:
1638 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1643 m_OneUse(m_ICmp(Pred, m_Specific(A), m_ZeroInt()))))) &&
1644 Pred == CmpInst::ICMP_SGT) {
1645 Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1646 Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1647 return BinaryOperator::CreateOr(LHS, Zext);
1648 }
1649
1650 if (Instruction *Ashr = foldAddToAshr(I))
1651 return Ashr;
1652
1653 // min(A, B) + max(A, B) => A + B.
1658 return BinaryOperator::CreateWithCopiedFlags(Instruction::Add, A, B, &I);
1659
1660 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1661 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1662 // computeKnownBits.
1663 bool Changed = false;
1664 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1665 Changed = true;
1666 I.setHasNoSignedWrap(true);
1667 }
1668 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1669 Changed = true;
1670 I.setHasNoUnsignedWrap(true);
1671 }
1672
1674 return V;
1675
1676 if (Instruction *V =
1678 return V;
1679
1681 return SatAdd;
1682
1683 // usub.sat(A, B) + B => umax(A, B)
1684 if (match(&I, m_c_BinOp(
1685 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1686 m_Deferred(B)))) {
1687 return replaceInstUsesWith(I,
1688 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1689 }
1690
1691 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1692 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1693 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1694 haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1695 return replaceInstUsesWith(
1696 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1697 {Builder.CreateOr(A, B)}));
1698
1699 if (Instruction *Res = foldSquareSumInt(I))
1700 return Res;
1701
1702 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1703 return Res;
1704
1706 return Res;
1707
1708 return Changed ? &I : nullptr;
1709}
1710
1711/// Eliminate an op from a linear interpolation (lerp) pattern.
1713 InstCombiner::BuilderTy &Builder) {
1714 Value *X, *Y, *Z;
1717 m_Value(Z))))),
1719 return nullptr;
1720
1721 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1722 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1723 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1724 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1725}
1726
1727/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1729 InstCombiner::BuilderTy &Builder) {
1730 assert((I.getOpcode() == Instruction::FAdd ||
1731 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1732 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1733 "FP factorization requires FMF");
1734
1735 if (Instruction *Lerp = factorizeLerp(I, Builder))
1736 return Lerp;
1737
1738 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1739 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1740 return nullptr;
1741
1742 Value *X, *Y, *Z;
1743 bool IsFMul;
1744 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1745 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1746 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1747 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1748 IsFMul = true;
1749 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1750 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1751 IsFMul = false;
1752 else
1753 return nullptr;
1754
1755 // (X * Z) + (Y * Z) --> (X + Y) * Z
1756 // (X * Z) - (Y * Z) --> (X - Y) * Z
1757 // (X / Z) + (Y / Z) --> (X + Y) / Z
1758 // (X / Z) - (Y / Z) --> (X - Y) / Z
1759 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1760 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1761 : Builder.CreateFSubFMF(X, Y, &I);
1762
1763 // Bail out if we just created a denormal constant.
1764 // TODO: This is copied from a previous implementation. Is it necessary?
1765 const APFloat *C;
1766 if (match(XY, m_APFloat(C)) && !C->isNormal())
1767 return nullptr;
1768
1769 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1771}
1772
1774 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1775 I.getFastMathFlags(),
1777 return replaceInstUsesWith(I, V);
1778
1780 return &I;
1781
1783 return X;
1784
1786 return Phi;
1787
1788 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1789 return FoldedFAdd;
1790
1791 // (-X) + Y --> Y - X
1792 Value *X, *Y;
1793 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1795
1796 // Similar to above, but look through fmul/fdiv for the negated term.
1797 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1798 Value *Z;
1800 m_Value(Z)))) {
1801 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1802 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1803 }
1804 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1805 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1807 m_Value(Z))) ||
1809 m_Value(Z)))) {
1810 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1811 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1812 }
1813
1814 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1815 // integer add followed by a promotion.
1816 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1817 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1818 Value *LHSIntVal = LHSConv->getOperand(0);
1819 Type *FPType = LHSConv->getType();
1820
1821 // TODO: This check is overly conservative. In many cases known bits
1822 // analysis can tell us that the result of the addition has less significant
1823 // bits than the integer type can hold.
1824 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1825 Type *FScalarTy = FTy->getScalarType();
1826 Type *IScalarTy = ITy->getScalarType();
1827
1828 // Do we have enough bits in the significand to represent the result of
1829 // the integer addition?
1830 unsigned MaxRepresentableBits =
1832 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1833 };
1834
1835 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1836 // ... if the constant fits in the integer value. This is useful for things
1837 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1838 // requires a constant pool load, and generally allows the add to be better
1839 // instcombined.
1840 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1841 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1842 Constant *CI =
1843 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1844 if (LHSConv->hasOneUse() &&
1845 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1846 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1847 // Insert the new integer add.
1848 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1849 return new SIToFPInst(NewAdd, I.getType());
1850 }
1851 }
1852
1853 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1854 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1855 Value *RHSIntVal = RHSConv->getOperand(0);
1856 // It's enough to check LHS types only because we require int types to
1857 // be the same for this transform.
1858 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1859 // Only do this if x/y have the same type, if at least one of them has a
1860 // single use (so we don't increase the number of int->fp conversions),
1861 // and if the integer add will not overflow.
1862 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1863 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1864 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1865 // Insert the new integer add.
1866 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1867 return new SIToFPInst(NewAdd, I.getType());
1868 }
1869 }
1870 }
1871 }
1872
1873 // Handle specials cases for FAdd with selects feeding the operation
1875 return replaceInstUsesWith(I, V);
1876
1877 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1879 return F;
1880
1882 return F;
1883
1884 // Try to fold fadd into start value of reduction intrinsic.
1885 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1886 m_AnyZeroFP(), m_Value(X))),
1887 m_Value(Y)))) {
1888 // fadd (rdx 0.0, X), Y --> rdx Y, X
1889 return replaceInstUsesWith(
1890 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1891 {X->getType()}, {Y, X}, &I));
1892 }
1893 const APFloat *StartC, *C;
1894 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1895 m_APFloat(StartC), m_Value(X)))) &&
1896 match(RHS, m_APFloat(C))) {
1897 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1898 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1899 return replaceInstUsesWith(
1900 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1901 {X->getType()}, {NewStartC, X}, &I));
1902 }
1903
1904 // (X * MulC) + X --> X * (MulC + 1.0)
1905 Constant *MulC;
1906 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1907 m_Deferred(X)))) {
1909 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1910 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1911 }
1912
1913 // (-X - Y) + (X + Z) --> Z - Y
1915 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1916 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1917
1918 if (Value *V = FAddCombine(Builder).simplify(&I))
1919 return replaceInstUsesWith(I, V);
1920 }
1921
1922 // minumum(X, Y) + maximum(X, Y) => X + Y.
1923 if (match(&I,
1924 m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1925 m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1926 m_Deferred(Y))))) {
1928 // We cannot preserve ninf if nnan flag is not set.
1929 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1930 // while in optimized version NaN + Inf and this is a poison with ninf flag.
1931 if (!Result->hasNoNaNs())
1932 Result->setHasNoInfs(false);
1933 return Result;
1934 }
1935
1936 return nullptr;
1937}
1938
1939/// Optimize pointer differences into the same array into a size. Consider:
1940/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1941/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1943 Type *Ty, bool IsNUW) {
1944 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1945 // this.
1946 bool Swapped = false;
1947 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1948 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1949 std::swap(LHS, RHS);
1950 Swapped = true;
1951 }
1952
1953 // Require at least one GEP with a common base pointer on both sides.
1954 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1955 // (gep X, ...) - X
1956 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1958 GEP1 = LHSGEP;
1959 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1960 // (gep X, ...) - (gep X, ...)
1961 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1962 RHSGEP->getOperand(0)->stripPointerCasts()) {
1963 GEP1 = LHSGEP;
1964 GEP2 = RHSGEP;
1965 }
1966 }
1967 }
1968
1969 if (!GEP1)
1970 return nullptr;
1971
1972 if (GEP2) {
1973 // (gep X, ...) - (gep X, ...)
1974 //
1975 // Avoid duplicating the arithmetic if there are more than one non-constant
1976 // indices between the two GEPs and either GEP has a non-constant index and
1977 // multiple users. If zero non-constant index, the result is a constant and
1978 // there is no duplication. If one non-constant index, the result is an add
1979 // or sub with a constant, which is no larger than the original code, and
1980 // there's no duplicated arithmetic, even if either GEP has multiple
1981 // users. If more than one non-constant indices combined, as long as the GEP
1982 // with at least one non-constant index doesn't have multiple users, there
1983 // is no duplication.
1984 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1985 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1986 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1987 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1988 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1989 return nullptr;
1990 }
1991 }
1992
1993 // Emit the offset of the GEP and an intptr_t.
1994 Value *Result = EmitGEPOffset(GEP1);
1995
1996 // If this is a single inbounds GEP and the original sub was nuw,
1997 // then the final multiplication is also nuw.
1998 if (auto *I = dyn_cast<Instruction>(Result))
1999 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
2000 I->getOpcode() == Instruction::Mul)
2001 I->setHasNoUnsignedWrap();
2002
2003 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2004 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2005 if (GEP2) {
2006 Value *Offset = EmitGEPOffset(GEP2);
2007 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2008 GEP1->isInBounds() && GEP2->isInBounds());
2009 }
2010
2011 // If we have p - gep(p, ...) then we have to negate the result.
2012 if (Swapped)
2013 Result = Builder.CreateNeg(Result, "diff.neg");
2014
2015 return Builder.CreateIntCast(Result, Ty, true);
2016}
2017
2019 InstCombiner::BuilderTy &Builder) {
2020 Value *Op0 = I.getOperand(0);
2021 Value *Op1 = I.getOperand(1);
2022 Type *Ty = I.getType();
2023 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2024 if (!MinMax)
2025 return nullptr;
2026
2027 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2028 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2029 Value *X = MinMax->getLHS();
2030 Value *Y = MinMax->getRHS();
2031 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2032 (Op0->hasOneUse() || Op1->hasOneUse())) {
2033 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2034 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2035 return CallInst::Create(F, {X, Y});
2036 }
2037
2038 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2039 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2040 Value *Z;
2041 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2042 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2043 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2044 return BinaryOperator::CreateAdd(X, USub);
2045 }
2046 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2047 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2048 return BinaryOperator::CreateAdd(X, USub);
2049 }
2050 }
2051
2052 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2053 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2054 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2055 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2056 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2057 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2058 return CallInst::Create(F, {Op0, Z});
2059 }
2060
2061 return nullptr;
2062}
2063
2065 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2066 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2068 return replaceInstUsesWith(I, V);
2069
2071 return X;
2072
2074 return Phi;
2075
2076 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2077
2078 // If this is a 'B = x-(-A)', change to B = x+A.
2079 // We deal with this without involving Negator to preserve NSW flag.
2080 if (Value *V = dyn_castNegVal(Op1)) {
2081 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2082
2083 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2084 assert(BO->getOpcode() == Instruction::Sub &&
2085 "Expected a subtraction operator!");
2086 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2087 Res->setHasNoSignedWrap(true);
2088 } else {
2089 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2090 Res->setHasNoSignedWrap(true);
2091 }
2092
2093 return Res;
2094 }
2095
2096 // Try this before Negator to preserve NSW flag.
2098 return R;
2099
2100 Constant *C;
2101 if (match(Op0, m_ImmConstant(C))) {
2102 Value *X;
2103 Constant *C2;
2104
2105 // C-(X+C2) --> (C-C2)-X
2106 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2107 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW
2108 // => (C-C2)-X can have NSW
2109 bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2110 BinaryOperator *Res =
2111 BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2112 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2113 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2114 WillNotSOV);
2115 return Res;
2116 }
2117 }
2118
2119 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2120 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2121 return Ext;
2122
2123 bool Changed = false;
2124 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2125 Changed = true;
2126 I.setHasNoSignedWrap(true);
2127 }
2128 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2129 Changed = true;
2130 I.setHasNoUnsignedWrap(true);
2131 }
2132
2133 return Changed ? &I : nullptr;
2134 };
2135
2136 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2137 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2138 // a pure negation used by a select that looks like abs/nabs.
2139 bool IsNegation = match(Op0, m_ZeroInt());
2140 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2141 const Instruction *UI = dyn_cast<Instruction>(U);
2142 if (!UI)
2143 return false;
2144 return match(UI,
2145 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2146 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2147 })) {
2148 if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2149 I.hasNoSignedWrap(),
2150 Op1, *this))
2151 return BinaryOperator::CreateAdd(NegOp1, Op0);
2152 }
2153 if (IsNegation)
2154 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2155
2156 // (A*B)-(A*C) -> A*(B-C) etc
2158 return replaceInstUsesWith(I, V);
2159
2160 if (I.getType()->isIntOrIntVectorTy(1))
2161 return BinaryOperator::CreateXor(Op0, Op1);
2162
2163 // Replace (-1 - A) with (~A).
2164 if (match(Op0, m_AllOnes()))
2165 return BinaryOperator::CreateNot(Op1);
2166
2167 // (X + -1) - Y --> ~Y + X
2168 Value *X, *Y;
2169 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2170 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2171
2172 // Reassociate sub/add sequences to create more add instructions and
2173 // reduce dependency chains:
2174 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2175 Value *Z;
2177 m_Value(Z))))) {
2178 Value *XZ = Builder.CreateAdd(X, Z);
2179 Value *YW = Builder.CreateAdd(Y, Op1);
2180 return BinaryOperator::CreateSub(XZ, YW);
2181 }
2182
2183 // ((X - Y) - Op1) --> X - (Y + Op1)
2184 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2185 Value *Add = Builder.CreateAdd(Y, Op1);
2186 return BinaryOperator::CreateSub(X, Add);
2187 }
2188
2189 // (~X) - (~Y) --> Y - X
2190 // This is placed after the other reassociations and explicitly excludes a
2191 // sub-of-sub pattern to avoid infinite looping.
2192 if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
2193 isFreeToInvert(Op1, Op1->hasOneUse()) &&
2194 !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
2195 Value *NotOp0 = Builder.CreateNot(Op0);
2196 Value *NotOp1 = Builder.CreateNot(Op1);
2197 return BinaryOperator::CreateSub(NotOp1, NotOp0);
2198 }
2199
2200 auto m_AddRdx = [](Value *&Vec) {
2201 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2202 };
2203 Value *V0, *V1;
2204 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2205 V0->getType() == V1->getType()) {
2206 // Difference of sums is sum of differences:
2207 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2208 Value *Sub = Builder.CreateSub(V0, V1);
2209 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2210 {Sub->getType()}, {Sub});
2211 return replaceInstUsesWith(I, Rdx);
2212 }
2213
2214 if (Constant *C = dyn_cast<Constant>(Op0)) {
2215 Value *X;
2216 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2217 // C - (zext bool) --> bool ? C - 1 : C
2219 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2220 // C - (sext bool) --> bool ? C + 1 : C
2222
2223 // C - ~X == X + (1+C)
2224 if (match(Op1, m_Not(m_Value(X))))
2225 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2226
2227 // Try to fold constant sub into select arguments.
2228 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2229 if (Instruction *R = FoldOpIntoSelect(I, SI))
2230 return R;
2231
2232 // Try to fold constant sub into PHI values.
2233 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2234 if (Instruction *R = foldOpIntoPhi(I, PN))
2235 return R;
2236
2237 Constant *C2;
2238
2239 // C-(C2-X) --> X+(C-C2)
2240 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2241 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2242 }
2243
2244 const APInt *Op0C;
2245 if (match(Op0, m_APInt(Op0C))) {
2246 if (Op0C->isMask()) {
2247 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2248 // zero.
2249 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
2250 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2251 return BinaryOperator::CreateXor(Op1, Op0);
2252 }
2253
2254 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2255 // (C3 - ((C2 & C3) - 1)) is pow2
2256 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2257 // C2 is negative pow2 || sub nuw
2258 const APInt *C2, *C3;
2259 BinaryOperator *InnerSub;
2260 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2261 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2262 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2263 APInt C2AndC3 = *C2 & *C3;
2264 APInt C2AndC3Minus1 = C2AndC3 - 1;
2265 APInt C2AddC3 = *C2 + *C3;
2266 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2267 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2268 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2269 return BinaryOperator::CreateAdd(
2270 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2271 }
2272 }
2273 }
2274
2275 {
2276 Value *Y;
2277 // X-(X+Y) == -Y X-(Y+X) == -Y
2278 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2280
2281 // (X-Y)-X == -Y
2282 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2284 }
2285
2286 // (sub (or A, B) (and A, B)) --> (xor A, B)
2287 {
2288 Value *A, *B;
2289 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2290 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2291 return BinaryOperator::CreateXor(A, B);
2292 }
2293
2294 // (sub (add A, B) (or A, B)) --> (and A, B)
2295 {
2296 Value *A, *B;
2297 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2298 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2299 return BinaryOperator::CreateAnd(A, B);
2300 }
2301
2302 // (sub (add A, B) (and A, B)) --> (or A, B)
2303 {
2304 Value *A, *B;
2305 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2307 return BinaryOperator::CreateOr(A, B);
2308 }
2309
2310 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2311 {
2312 Value *A, *B;
2313 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2314 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2315 (Op0->hasOneUse() || Op1->hasOneUse()))
2317 }
2318
2319 // (sub (or A, B), (xor A, B)) --> (and A, B)
2320 {
2321 Value *A, *B;
2322 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2323 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2324 return BinaryOperator::CreateAnd(A, B);
2325 }
2326
2327 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2328 {
2329 Value *A, *B;
2330 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2331 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2332 (Op0->hasOneUse() || Op1->hasOneUse()))
2334 }
2335
2336 {
2337 Value *Y;
2338 // ((X | Y) - X) --> (~X & Y)
2339 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2340 return BinaryOperator::CreateAnd(
2341 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2342 }
2343
2344 {
2345 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2346 Value *X;
2347 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2348 m_OneUse(m_Neg(m_Value(X))))))) {
2350 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2351 }
2352 }
2353
2354 {
2355 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2356 Constant *C;
2357 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2360 }
2361 }
2362
2364 return R;
2365
2366 {
2367 // If we have a subtraction between some value and a select between
2368 // said value and something else, sink subtraction into select hands, i.e.:
2369 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2370 // ->
2371 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2372 // or
2373 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2374 // ->
2375 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2376 // This will result in select between new subtraction and 0.
2377 auto SinkSubIntoSelect =
2378 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2379 auto SubBuilder) -> Instruction * {
2380 Value *Cond, *TrueVal, *FalseVal;
2381 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2382 m_Value(FalseVal)))))
2383 return nullptr;
2384 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2385 return nullptr;
2386 // While it is really tempting to just create two subtractions and let
2387 // InstCombine fold one of those to 0, it isn't possible to do so
2388 // because of worklist visitation order. So ugly it is.
2389 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2390 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2391 Constant *Zero = Constant::getNullValue(Ty);
2392 SelectInst *NewSel =
2393 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2394 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2395 // Preserve prof metadata if any.
2396 NewSel->copyMetadata(cast<Instruction>(*Select));
2397 return NewSel;
2398 };
2399 if (Instruction *NewSel = SinkSubIntoSelect(
2400 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2401 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2402 return Builder->CreateSub(OtherHandOfSelect,
2403 /*OtherHandOfSub=*/Op1);
2404 }))
2405 return NewSel;
2406 if (Instruction *NewSel = SinkSubIntoSelect(
2407 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2408 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2409 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2410 OtherHandOfSelect);
2411 }))
2412 return NewSel;
2413 }
2414
2415 // (X - (X & Y)) --> (X & ~Y)
2416 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2417 (Op1->hasOneUse() || isa<Constant>(Y)))
2418 return BinaryOperator::CreateAnd(
2419 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2420
2421 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2422 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2423 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2424 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2425 // As long as Y is freely invertible, this will be neutral or a win.
2426 // Note: We don't generate the inverse max/min, just create the 'not' of
2427 // it and let other folds do the rest.
2428 if (match(Op0, m_Not(m_Value(X))) &&
2429 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2430 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2431 Value *Not = Builder.CreateNot(Op1);
2432 return BinaryOperator::CreateSub(Not, X);
2433 }
2434 if (match(Op1, m_Not(m_Value(X))) &&
2435 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2436 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2437 Value *Not = Builder.CreateNot(Op0);
2438 return BinaryOperator::CreateSub(X, Not);
2439 }
2440
2441 // Optimize pointer differences into the same array into a size. Consider:
2442 // &A[10] - &A[0]: we should compile this to "10".
2443 Value *LHSOp, *RHSOp;
2444 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2445 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2446 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2447 I.hasNoUnsignedWrap()))
2448 return replaceInstUsesWith(I, Res);
2449
2450 // trunc(p)-trunc(q) -> trunc(p-q)
2451 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2452 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2453 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2454 /* IsNUW */ false))
2455 return replaceInstUsesWith(I, Res);
2456
2457 // Canonicalize a shifty way to code absolute value to the common pattern.
2458 // There are 2 potential commuted variants.
2459 // We're relying on the fact that we only do this transform when the shift has
2460 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2461 // instructions).
2462 Value *A;
2463 const APInt *ShAmt;
2464 Type *Ty = I.getType();
2465 unsigned BitWidth = Ty->getScalarSizeInBits();
2466 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2467 Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2468 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2469 // B = ashr i32 A, 31 ; smear the sign bit
2470 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2471 // --> (A < 0) ? -A : A
2472 Value *IsNeg = Builder.CreateIsNeg(A);
2473 // Copy the nuw/nsw flags from the sub to the negate.
2474 Value *NegA = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2475 I.hasNoSignedWrap());
2476 return SelectInst::Create(IsNeg, NegA, A);
2477 }
2478
2479 // If we are subtracting a low-bit masked subset of some value from an add
2480 // of that same value with no low bits changed, that is clearing some low bits
2481 // of the sum:
2482 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2483 const APInt *AddC, *AndC;
2484 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2485 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2486 unsigned Cttz = AddC->countr_zero();
2487 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2488 if ((HighMask & *AndC).isZero())
2489 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2490 }
2491
2492 if (Instruction *V =
2494 return V;
2495
2496 // X - usub.sat(X, Y) => umin(X, Y)
2497 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2498 m_Value(Y)))))
2499 return replaceInstUsesWith(
2500 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2501
2502 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2503 // TODO: The one-use restriction is not strictly necessary, but it may
2504 // require improving other pattern matching and/or codegen.
2505 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2506 return replaceInstUsesWith(
2507 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2508
2509 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2510 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2511 return replaceInstUsesWith(
2512 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2513
2514 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2515 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2516 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2517 return BinaryOperator::CreateNeg(USub);
2518 }
2519
2520 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2521 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2522 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2523 return BinaryOperator::CreateNeg(USub);
2524 }
2525
2526 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2527 if (match(Op0, m_SpecificInt(BitWidth)) &&
2528 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2529 return replaceInstUsesWith(
2530 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2531 {Builder.CreateNot(X)}));
2532
2533 // Reduce multiplies for difference-of-squares by factoring:
2534 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2535 if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2536 match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2537 auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2538 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2539 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2540 OBO1->hasNoSignedWrap() && BitWidth > 2;
2541 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2542 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2543 Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2544 Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2545 Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2546 return replaceInstUsesWith(I, Mul);
2547 }
2548
2549 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2550 if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2552 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2553 Value *Sub =
2554 Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2555 Value *Call =
2556 Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2557 return replaceInstUsesWith(I, Call);
2558 }
2559 }
2560
2562 return Res;
2563
2564 return TryToNarrowDeduceFlags();
2565}
2566
2567/// This eliminates floating-point negation in either 'fneg(X)' or
2568/// 'fsub(-0.0, X)' form by combining into a constant operand.
2570 // This is limited with one-use because fneg is assumed better for
2571 // reassociation and cheaper in codegen than fmul/fdiv.
2572 // TODO: Should the m_OneUse restriction be removed?
2573 Instruction *FNegOp;
2574 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2575 return nullptr;
2576
2577 Value *X;
2578 Constant *C;
2579
2580 // Fold negation into constant operand.
2581 // -(X * C) --> X * (-C)
2582 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2583 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2584 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2585 // -(X / C) --> X / (-C)
2586 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2587 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2588 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2589 // -(C / X) --> (-C) / X
2590 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2591 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2593
2594 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2595 // not apply to the fdiv. Everything else propagates from the fneg.
2596 // TODO: We could propagate nsz/ninf from fdiv alone?
2597 FastMathFlags FMF = I.getFastMathFlags();
2598 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2599 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2600 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2601 return FDiv;
2602 }
2603 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2604 // -(X + C) --> -X + -C --> -C - X
2605 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2606 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2607 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2608
2609 return nullptr;
2610}
2611
2612Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2613 Instruction &FMFSource) {
2614 Value *X, *Y;
2615 if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2616 return cast<Instruction>(Builder.CreateFMulFMF(
2617 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2618 }
2619
2620 if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2621 return cast<Instruction>(Builder.CreateFDivFMF(
2622 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2623 }
2624
2625 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2626 // Make sure to preserve flags and metadata on the call.
2627 if (II->getIntrinsicID() == Intrinsic::ldexp) {
2628 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2631
2633 II->getCalledFunction(),
2634 {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2635 New->copyMetadata(*II);
2636 return New;
2637 }
2638 }
2639
2640 return nullptr;
2641}
2642
2644 Value *Op = I.getOperand(0);
2645
2646 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2647 getSimplifyQuery().getWithInstruction(&I)))
2648 return replaceInstUsesWith(I, V);
2649
2651 return X;
2652
2653 Value *X, *Y;
2654
2655 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2656 if (I.hasNoSignedZeros() &&
2659
2660 Value *OneUse;
2661 if (!match(Op, m_OneUse(m_Value(OneUse))))
2662 return nullptr;
2663
2664 if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2665 return replaceInstUsesWith(I, R);
2666
2667 // Try to eliminate fneg if at least 1 arm of the select is negated.
2668 Value *Cond;
2669 if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2670 // Unlike most transforms, this one is not safe to propagate nsz unless
2671 // it is present on the original select. We union the flags from the select
2672 // and fneg and then remove nsz if needed.
2673 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2674 S->copyFastMathFlags(&I);
2675 if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2676 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2677 S->setFastMathFlags(FMF);
2678 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2679 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2680 S->setHasNoSignedZeros(false);
2681 }
2682 };
2683 // -(Cond ? -P : Y) --> Cond ? P : -Y
2684 Value *P;
2685 if (match(X, m_FNeg(m_Value(P)))) {
2686 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2687 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2688 propagateSelectFMF(NewSel, P == Y);
2689 return NewSel;
2690 }
2691 // -(Cond ? X : -P) --> Cond ? -X : P
2692 if (match(Y, m_FNeg(m_Value(P)))) {
2693 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2694 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2695 propagateSelectFMF(NewSel, P == X);
2696 return NewSel;
2697 }
2698 }
2699
2700 // fneg (copysign x, y) -> copysign x, (fneg y)
2701 if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2702 // The source copysign has an additional value input, so we can't propagate
2703 // flags the copysign doesn't also have.
2704 FastMathFlags FMF = I.getFastMathFlags();
2705 FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2706
2709
2710 Value *NegY = Builder.CreateFNeg(Y);
2711 Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2712 return replaceInstUsesWith(I, NewCopySign);
2713 }
2714
2715 return nullptr;
2716}
2717
2719 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2720 I.getFastMathFlags(),
2721 getSimplifyQuery().getWithInstruction(&I)))
2722 return replaceInstUsesWith(I, V);
2723
2725 return X;
2726
2728 return Phi;
2729
2730 // Subtraction from -0.0 is the canonical form of fneg.
2731 // fsub -0.0, X ==> fneg X
2732 // fsub nsz 0.0, X ==> fneg nsz X
2733 //
2734 // FIXME This matcher does not respect FTZ or DAZ yet:
2735 // fsub -0.0, Denorm ==> +-0
2736 // fneg Denorm ==> -Denorm
2737 Value *Op;
2738 if (match(&I, m_FNeg(m_Value(Op))))
2740
2742 return X;
2743
2744 Value *X, *Y;
2745 Constant *C;
2746
2747 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2748 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2749 // Canonicalize to fadd to make analysis easier.
2750 // This can also help codegen because fadd is commutative.
2751 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2752 // killed later. We still limit that particular transform with 'hasOneUse'
2753 // because an fneg is assumed better/cheaper than a generic fsub.
2754 if (I.hasNoSignedZeros() || cannotBeNegativeZero(Op0, SQ.DL, SQ.TLI)) {
2755 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2756 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2757 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2758 }
2759 }
2760
2761 // (-X) - Op1 --> -(X + Op1)
2762 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2763 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2764 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2766 }
2767
2768 if (isa<Constant>(Op0))
2769 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2770 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2771 return NV;
2772
2773 // X - C --> X + (-C)
2774 // But don't transform constant expressions because there's an inverse fold
2775 // for X + (-Y) --> X - Y.
2776 if (match(Op1, m_ImmConstant(C)))
2777 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2778 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2779
2780 // X - (-Y) --> X + Y
2781 if (match(Op1, m_FNeg(m_Value(Y))))
2782 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2783
2784 // Similar to above, but look through a cast of the negated value:
2785 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2786 Type *Ty = I.getType();
2787 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2789
2790 // X - (fpext(-Y)) --> X + fpext(Y)
2791 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2793
2794 // Similar to above, but look through fmul/fdiv of the negated value:
2795 // Op0 - (-X * Y) --> Op0 + (X * Y)
2796 // Op0 - (Y * -X) --> Op0 + (X * Y)
2797 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2799 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2800 }
2801 // Op0 - (-X / Y) --> Op0 + (X / Y)
2802 // Op0 - (X / -Y) --> Op0 + (X / Y)
2803 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2804 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2805 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2806 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2807 }
2808
2809 // Handle special cases for FSub with selects feeding the operation
2810 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2811 return replaceInstUsesWith(I, V);
2812
2813 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2814 // (Y - X) - Y --> -X
2815 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2817
2818 // Y - (X + Y) --> -X
2819 // Y - (Y + X) --> -X
2820 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2822
2823 // (X * C) - X --> X * (C - 1.0)
2824 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2826 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2827 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2828 }
2829 // X - (X * C) --> X * (1.0 - C)
2830 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2832 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2833 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2834 }
2835
2836 // Reassociate fsub/fadd sequences to create more fadd instructions and
2837 // reduce dependency chains:
2838 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2839 Value *Z;
2841 m_Value(Z))))) {
2842 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2843 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2844 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2845 }
2846
2847 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2848 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2849 m_Value(Vec)));
2850 };
2851 Value *A0, *A1, *V0, *V1;
2852 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2853 V0->getType() == V1->getType()) {
2854 // Difference of sums is sum of differences:
2855 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2856 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2857 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2858 {Sub->getType()}, {A0, Sub}, &I);
2859 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2860 }
2861
2863 return F;
2864
2865 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2866 // functionality has been subsumed by simple pattern matching here and in
2867 // InstSimplify. We should let a dedicated reassociation pass handle more
2868 // complex pattern matching and remove this from InstCombine.
2869 if (Value *V = FAddCombine(Builder).simplify(&I))
2870 return replaceInstUsesWith(I, V);
2871
2872 // (X - Y) - Op1 --> X - (Y + Op1)
2873 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2874 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2876 }
2877 }
2878
2879 return nullptr;
2880}
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...
assume Assume Builder
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
const fltSemantics & getSemantics() const
Definition: APFloat.h:1301
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1058
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:711
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:734
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:738
@ ICMP_EQ
equal
Definition: InstrTypes.h:732
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2573
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2103
static Constant * getSIToFP(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2152
static Constant * getSExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2089
static Constant * getFPToSI(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2174
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2566
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:389
unsigned countNonConstantIndices() const
Definition: Operator.h:469
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:1538
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1712
Value * CreateSRem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1401
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:1592
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:1565
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:1619
Value * CreateFPTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2063
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:452
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:941
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:1736
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2521
Value * CreateNSWAdd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1327
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:297
CallInst * CreateCopySign(Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create call to the copysign intrinsic.
Definition: IRBuilder.h:1013
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1997
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1745
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:2516
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1335
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1407
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:932
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1466
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1318
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2511
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:2158
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2374
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2072
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1510
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1726
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1395
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1352
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:421
static bool isFreeToInvert(Value *V, bool WillInvertAllUses)
Return true if the specified value is free to invert (apply ~ to).
Definition: InstCombiner.h:239
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
Definition: InstCombiner.h:211
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:487
AssumptionCache & AC
Definition: InstCombiner.h:71
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:445
DominatorTree & DT
Definition: InstCombiner.h:73
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:169
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
Definition: InstCombiner.h:466
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:482
const SimplifyQuery & getSimplifyQuery() const
Definition: InstCombiner.h:377
static Constant * AddOne(Constant *C)
Add one to a Constant.
Definition: InstCombiner.h:206
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:389
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.
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:688
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:1422
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
match_combine_or< CastClass_match< OpTy, Instruction::ZExt >, OpTy > m_ZExtOrSelf(const OpTy &Op)
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.
cst_pred_ty< is_sign_mask > m_SignMask()
Match an integer or vector with only the sign bit(s) set.
Definition: PatternMatch.h:584
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)
CastClass_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
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)
CastClass_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
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
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
CastClass_match< OpTy, Instruction::FPTrunc > m_FPTrunc(const OpTy &Op)
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'.
CastClass_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
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.
CastClass_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
BinaryOp_match< LHS, RHS, Instruction::UDiv > m_UDiv(const LHS &L, const RHS &R)
match_combine_or< CastClass_match< OpTy, Instruction::Trunc >, OpTy > m_TruncOrSelf(const OpTy &Op)
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.
CastClass_match< OpTy, Instruction::FPExt > m_FPExt(const OpTy &Op)
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)
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
BinaryOp_match< LHS, RHS, Instruction::SRem > m_SRem(const LHS &L, const RHS &R)
match_combine_or< CastClass_match< OpTy, Instruction::SExt >, OpTy > m_SExtOrSelf(const OpTy &Op)
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)
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)
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
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:1741
Constant * ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, Constant *RHS, const DataLayout &DL)
Attempt to constant fold a binary operation with the specified operands.
bool haveNoCommonBitsSet(const Value *LHS, const Value *RHS, const DataLayout &DL, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Return true if LHS and RHS have no common bits set.
@ 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:184
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
const Instruction * CxtI
const DominatorTree * DT
AssumptionCache * AC
SimplifyQuery getWithInstruction(Instruction *I) const
const TargetLibraryInfo * TLI