LLVM 19.0.0git
InstCombineAddSub.cpp
Go to the documentation of this file.
1//===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visit functions for add, fadd, sub, and fsub.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APFloat.h"
15#include "llvm/ADT/APInt.h"
16#include "llvm/ADT/STLExtras.h"
20#include "llvm/IR/Constant.h"
21#include "llvm/IR/Constants.h"
22#include "llvm/IR/InstrTypes.h"
23#include "llvm/IR/Instruction.h"
25#include "llvm/IR/Operator.h"
27#include "llvm/IR/Type.h"
28#include "llvm/IR/Value.h"
33#include <cassert>
34#include <utility>
35
36using namespace llvm;
37using namespace PatternMatch;
38
39#define DEBUG_TYPE "instcombine"
40
41namespace {
42
43 /// Class representing coefficient of floating-point addend.
44 /// This class needs to be highly efficient, which is especially true for
45 /// the constructor. As of I write this comment, the cost of the default
46 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
47 /// perform write-merging).
48 ///
49 class FAddendCoef {
50 public:
51 // The constructor has to initialize a APFloat, which is unnecessary for
52 // most addends which have coefficient either 1 or -1. So, the constructor
53 // is expensive. In order to avoid the cost of the constructor, we should
54 // reuse some instances whenever possible. The pre-created instances
55 // FAddCombine::Add[0-5] embodies this idea.
56 FAddendCoef() = default;
57 ~FAddendCoef();
58
59 // If possible, don't define operator+/operator- etc because these
60 // operators inevitably call FAddendCoef's constructor which is not cheap.
61 void operator=(const FAddendCoef &A);
62 void operator+=(const FAddendCoef &A);
63 void operator*=(const FAddendCoef &S);
64
65 void set(short C) {
66 assert(!insaneIntVal(C) && "Insane coefficient");
67 IsFp = false; IntVal = C;
68 }
69
70 void set(const APFloat& C);
71
72 void negate();
73
74 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75 Value *getValue(Type *) const;
76
77 bool isOne() const { return isInt() && IntVal == 1; }
78 bool isTwo() const { return isInt() && IntVal == 2; }
79 bool isMinusOne() const { return isInt() && IntVal == -1; }
80 bool isMinusTwo() const { return isInt() && IntVal == -2; }
81
82 private:
83 bool insaneIntVal(int V) { return V > 4 || V < -4; }
84
85 APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86
87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
89 }
90
91 const APFloat &getFpVal() const {
92 assert(IsFp && BufHasFpVal && "Incorret state");
93 return *getFpValPtr();
94 }
95
96 APFloat &getFpVal() {
97 assert(IsFp && BufHasFpVal && "Incorret state");
98 return *getFpValPtr();
99 }
100
101 bool isInt() const { return !IsFp; }
102
103 // If the coefficient is represented by an integer, promote it to a
104 // floating point.
105 void convertToFpType(const fltSemantics &Sem);
106
107 // Construct an APFloat from a signed integer.
108 // TODO: We should get rid of this function when APFloat can be constructed
109 // from an *SIGNED* integer.
110 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
111
112 bool IsFp = false;
113
114 // True iff FpValBuf contains an instance of APFloat.
115 bool BufHasFpVal = false;
116
117 // The integer coefficient of an individual addend is either 1 or -1,
118 // and we try to simplify at most 4 addends from neighboring at most
119 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120 // is overkill of this end.
121 short IntVal = 0;
122
124 };
125
126 /// FAddend is used to represent floating-point addend. An addend is
127 /// represented as <C, V>, where the V is a symbolic value, and C is a
128 /// constant coefficient. A constant addend is represented as <C, 0>.
129 class FAddend {
130 public:
131 FAddend() = default;
132
133 void operator+=(const FAddend &T) {
134 assert((Val == T.Val) && "Symbolic-values disagree");
135 Coeff += T.Coeff;
136 }
137
138 Value *getSymVal() const { return Val; }
139 const FAddendCoef &getCoef() const { return Coeff; }
140
141 bool isConstant() const { return Val == nullptr; }
142 bool isZero() const { return Coeff.isZero(); }
143
144 void set(short Coefficient, Value *V) {
145 Coeff.set(Coefficient);
146 Val = V;
147 }
148 void set(const APFloat &Coefficient, Value *V) {
149 Coeff.set(Coefficient);
150 Val = V;
151 }
152 void set(const ConstantFP *Coefficient, Value *V) {
153 Coeff.set(Coefficient->getValueAPF());
154 Val = V;
155 }
156
157 void negate() { Coeff.negate(); }
158
159 /// Drill down the U-D chain one step to find the definition of V, and
160 /// try to break the definition into one or two addends.
161 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
162
163 /// Similar to FAddend::drillDownOneStep() except that the value being
164 /// splitted is the addend itself.
165 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
166
167 private:
168 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
169
170 // This addend has the value of "Coeff * Val".
171 Value *Val = nullptr;
172 FAddendCoef Coeff;
173 };
174
175 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176 /// with its neighboring at most two instructions.
177 ///
178 class FAddCombine {
179 public:
180 FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
181
183
184 private:
185 using AddendVect = SmallVector<const FAddend *, 4>;
186
187 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
188
189 /// Convert given addend to a Value
190 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
191
192 /// Return the number of instructions needed to emit the N-ary addition.
193 unsigned calcInstrNumber(const AddendVect& Vect);
194
195 Value *createFSub(Value *Opnd0, Value *Opnd1);
196 Value *createFAdd(Value *Opnd0, Value *Opnd1);
197 Value *createFMul(Value *Opnd0, Value *Opnd1);
198 Value *createFNeg(Value *V);
199 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
201
202 // Debugging stuff are clustered here.
203 #ifndef NDEBUG
204 unsigned CreateInstrNum;
205 void initCreateInstNum() { CreateInstrNum = 0; }
206 void incCreateInstNum() { CreateInstrNum++; }
207 #else
208 void initCreateInstNum() {}
209 void incCreateInstNum() {}
210 #endif
211
213 Instruction *Instr = nullptr;
214 };
215
216} // end anonymous namespace
217
218//===----------------------------------------------------------------------===//
219//
220// Implementation of
221// {FAddendCoef, FAddend, FAddition, FAddCombine}.
222//
223//===----------------------------------------------------------------------===//
224FAddendCoef::~FAddendCoef() {
225 if (BufHasFpVal)
226 getFpValPtr()->~APFloat();
227}
228
229void FAddendCoef::set(const APFloat& C) {
230 APFloat *P = getFpValPtr();
231
232 if (isInt()) {
233 // As the buffer is meanless byte stream, we cannot call
234 // APFloat::operator=().
235 new(P) APFloat(C);
236 } else
237 *P = C;
238
239 IsFp = BufHasFpVal = true;
240}
241
242void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
243 if (!isInt())
244 return;
245
246 APFloat *P = getFpValPtr();
247 if (IntVal > 0)
248 new(P) APFloat(Sem, IntVal);
249 else {
250 new(P) APFloat(Sem, 0 - IntVal);
251 P->changeSign();
252 }
253 IsFp = BufHasFpVal = true;
254}
255
256APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
257 if (Val >= 0)
258 return APFloat(Sem, Val);
259
260 APFloat T(Sem, 0 - Val);
261 T.changeSign();
262
263 return T;
264}
265
266void FAddendCoef::operator=(const FAddendCoef &That) {
267 if (That.isInt())
268 set(That.IntVal);
269 else
270 set(That.getFpVal());
271}
272
273void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
276 if (isInt())
277 IntVal += That.IntVal;
278 else
279 getFpVal().add(That.getFpVal(), RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292}
293
294void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value");
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322}
323
324void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329}
330
331Value *FAddendCoef::getValue(Type *Ty) const {
332 return isInt() ?
333 ConstantFP::get(Ty, float(IntVal)) :
334 ConstantFP::get(Ty->getContext(), getFpVal());
335}
336
337// The definition of <Val> Addends
338// =========================================
339// A + B <1, A>, <1,B>
340// A - B <1, A>, <1,B>
341// 0 - B <-1, B>
342// C * A, <C, A>
343// A + C <1, A> <C, NULL>
344// 0 +/- 0 <0, NULL> (corner case)
345//
346// Legend: A and B are not constant, C is constant
347unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
351 return 0;
352
353 unsigned Opcode = I->getOpcode();
354
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356 ConstantFP *C0, *C1;
357 Value *Opnd0 = I->getOperand(0);
358 Value *Opnd1 = I->getOperand(1);
359 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(1, Opnd0);
368 else
369 Addend0.set(C0, nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(1, Opnd1);
376 else
377 Addend.set(C1, nullptr);
378 if (Opcode == Instruction::FSub)
379 Addend.negate();
380 }
381
382 if (Opnd0 || Opnd1)
383 return Opnd0 && Opnd1 ? 2 : 1;
384
385 // Both operands are zero. Weird!
386 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(0);
392 Value *V1 = I->getOperand(1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394 Addend0.set(C, V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399 Addend0.set(C, V0);
400 return 1;
401 }
402 }
403
404 return 0;
405}
406
407// Try to break *this* addend into two addends. e.g. Suppose this addend is
408// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409// i.e. <2.3, X> and <2.3, Y>.
410unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
412 if (isConstant())
413 return 0;
414
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
417 return BreakNum;
418
419 Addend0.Scale(Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(Coeff);
423
424 return BreakNum;
425}
426
427Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429 "Expected 'reassoc'+'nsz' instruction");
430
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
433 return nullptr;
434
435 assert((I->getOpcode() == Instruction::FAdd ||
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
494
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
506
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
508 return R;
509 }
510
511 return nullptr;
512}
513
514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends");
517
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
521
522 // Simplified addends are placed <SimpVect>.
523 AddendVect SimpVect;
524
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529
530 const FAddend *ThisAddend = Addends[SymIdx];
531 if (!ThisAddend) {
532 // This addend was processed before.
533 continue;
534 }
535
536 Value *Val = ThisAddend->getSymVal();
537
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543 // call later.
544
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(ThisAddend);
547
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(T);
561 }
562 }
563
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569 R += *SimpVect[Idx];
570
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(&R);
575 }
576 }
577 }
578
579 assert((NextTmpIdx <= std::size(TmpResult) + 1) && "out-of-bound access");
580
581 Value *Result;
582 if (!SimpVect.empty())
583 Result = createNaryFAdd(SimpVect, InstrQuota);
584 else {
585 // The addition is folded to 0.0.
586 Result = ConstantFP::get(Instr->getType(), 0.0);
587 }
588
589 return Result;
590}
591
592Value *FAddCombine::createNaryFAdd
593 (const AddendVect &Opnds, unsigned InstrQuota) {
594 assert(!Opnds.empty() && "Expect at least one addend");
595
596 // Step 1: Check if the # of instructions needed exceeds the quota.
597
598 unsigned InstrNeeded = calcInstrNumber(Opnds);
599 if (InstrNeeded > InstrQuota)
600 return nullptr;
601
602 initCreateInstNum();
603
604 // step 2: Emit the N-ary addition.
605 // Note that at most three instructions are involved in Fadd-InstCombine: the
606 // addition in question, and at most two neighboring instructions.
607 // The resulting optimized addition should have at least one less instruction
608 // than the original addition expression tree. This implies that the resulting
609 // N-ary addition has at most two instructions, and we don't need to worry
610 // about tree-height when constructing the N-ary addition.
611
612 Value *LastVal = nullptr;
613 bool LastValNeedNeg = false;
614
615 // Iterate the addends, creating fadd/fsub using adjacent two addends.
616 for (const FAddend *Opnd : Opnds) {
617 bool NeedNeg;
618 Value *V = createAddendVal(*Opnd, NeedNeg);
619 if (!LastVal) {
620 LastVal = V;
621 LastValNeedNeg = NeedNeg;
622 continue;
623 }
624
625 if (LastValNeedNeg == NeedNeg) {
626 LastVal = createFAdd(LastVal, V);
627 continue;
628 }
629
630 if (LastValNeedNeg)
631 LastVal = createFSub(V, LastVal);
632 else
633 LastVal = createFSub(LastVal, V);
634
635 LastValNeedNeg = false;
636 }
637
638 if (LastValNeedNeg) {
639 LastVal = createFNeg(LastVal);
640 }
641
642#ifndef NDEBUG
643 assert(CreateInstrNum == InstrNeeded &&
644 "Inconsistent in instruction numbers");
645#endif
646
647 return LastVal;
648}
649
650Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
651 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
652 if (Instruction *I = dyn_cast<Instruction>(V))
653 createInstPostProc(I);
654 return V;
655}
656
657Value *FAddCombine::createFNeg(Value *V) {
658 Value *NewV = Builder.CreateFNeg(V);
659 if (Instruction *I = dyn_cast<Instruction>(NewV))
660 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
661 return NewV;
662}
663
664Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
665 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
666 if (Instruction *I = dyn_cast<Instruction>(V))
667 createInstPostProc(I);
668 return V;
669}
670
671Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
673 if (Instruction *I = dyn_cast<Instruction>(V))
674 createInstPostProc(I);
675 return V;
676}
677
678void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
679 NewInstr->setDebugLoc(Instr->getDebugLoc());
680
681 // Keep track of the number of instruction created.
682 if (!NoNumber)
683 incCreateInstNum();
684
685 // Propagate fast-math flags
686 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
687}
688
689// Return the number of instruction needed to emit the N-ary addition.
690// NOTE: Keep this function in sync with createAddendVal().
691unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
692 unsigned OpndNum = Opnds.size();
693 unsigned InstrNeeded = OpndNum - 1;
694
695 // Adjust the number of instructions needed to emit the N-ary add.
696 for (const FAddend *Opnd : Opnds) {
697 if (Opnd->isConstant())
698 continue;
699
700 // The constant check above is really for a few special constant
701 // coefficients.
702 if (isa<UndefValue>(Opnd->getSymVal()))
703 continue;
704
705 const FAddendCoef &CE = Opnd->getCoef();
706 // Let the addend be "c * x". If "c == +/-1", the value of the addend
707 // is immediately available; otherwise, it needs exactly one instruction
708 // to evaluate the value.
709 if (!CE.isMinusOne() && !CE.isOne())
710 InstrNeeded++;
711 }
712 return InstrNeeded;
713}
714
715// Input Addend Value NeedNeg(output)
716// ================================================================
717// Constant C C false
718// <+/-1, V> V coefficient is -1
719// <2/-2, V> "fadd V, V" coefficient is -2
720// <C, V> "fmul V, C" false
721//
722// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
723Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
724 const FAddendCoef &Coeff = Opnd.getCoef();
725
726 if (Opnd.isConstant()) {
727 NeedNeg = false;
728 return Coeff.getValue(Instr->getType());
729 }
730
731 Value *OpndVal = Opnd.getSymVal();
732
733 if (Coeff.isMinusOne() || Coeff.isOne()) {
734 NeedNeg = Coeff.isMinusOne();
735 return OpndVal;
736 }
737
738 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
739 NeedNeg = Coeff.isMinusTwo();
740 return createFAdd(OpndVal, OpndVal);
741 }
742
743 NeedNeg = false;
744 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
745}
746
747// Checks if any operand is negative and we can convert add to sub.
748// This function checks for following negative patterns
749// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
750// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
751// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
753 InstCombiner::BuilderTy &Builder) {
754 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
755
756 // This function creates 2 instructions to replace ADD, we need at least one
757 // of LHS or RHS to have one use to ensure benefit in transform.
758 if (!LHS->hasOneUse() && !RHS->hasOneUse())
759 return nullptr;
760
761 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
762 const APInt *C1 = nullptr, *C2 = nullptr;
763
764 // if ONE is on other side, swap
765 if (match(RHS, m_Add(m_Value(X), m_One())))
766 std::swap(LHS, RHS);
767
768 if (match(LHS, m_Add(m_Value(X), m_One()))) {
769 // if XOR on other side, swap
770 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
771 std::swap(X, RHS);
772
773 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
774 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
775 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
776 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
777 Value *NewAnd = Builder.CreateAnd(Z, *C1);
778 return Builder.CreateSub(RHS, NewAnd, "sub");
779 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
780 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
781 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
782 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
783 return Builder.CreateSub(RHS, NewOr, "sub");
784 }
785 }
786 }
787
788 // Restore LHS and RHS
789 LHS = I.getOperand(0);
790 RHS = I.getOperand(1);
791
792 // if XOR is on other side, swap
793 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
794 std::swap(LHS, RHS);
795
796 // C2 is ODD
797 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
798 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
799 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
800 if (C1->countr_zero() == 0)
801 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
802 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
803 return Builder.CreateSub(RHS, NewOr, "sub");
804 }
805 return nullptr;
806}
807
808/// Wrapping flags may allow combining constants separated by an extend.
810 InstCombiner::BuilderTy &Builder) {
811 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
812 Type *Ty = Add.getType();
813 Constant *Op1C;
814 if (!match(Op1, m_Constant(Op1C)))
815 return nullptr;
816
817 // Try this match first because it results in an add in the narrow type.
818 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
819 Value *X;
820 const APInt *C1, *C2;
821 if (match(Op1, m_APInt(C1)) &&
822 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
823 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824 Constant *NewC =
825 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
826 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
827 }
828
829 // More general combining of constants in the wide type.
830 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
831 Constant *NarrowC;
832 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
833 Value *WideC = Builder.CreateSExt(NarrowC, Ty);
834 Value *NewC = Builder.CreateAdd(WideC, Op1C);
835 Value *WideX = Builder.CreateSExt(X, Ty);
836 return BinaryOperator::CreateAdd(WideX, NewC);
837 }
838 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
839 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
840 Value *WideC = Builder.CreateZExt(NarrowC, Ty);
841 Value *NewC = Builder.CreateAdd(WideC, Op1C);
842 Value *WideX = Builder.CreateZExt(X, Ty);
843 return BinaryOperator::CreateAdd(WideX, NewC);
844 }
845
846 return nullptr;
847}
848
850 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
851 Type *Ty = Add.getType();
852 Constant *Op1C;
853 if (!match(Op1, m_ImmConstant(Op1C)))
854 return nullptr;
855
857 return NV;
858
859 Value *X;
860 Constant *Op00C;
861
862 // add (sub C1, X), C2 --> sub (add C1, C2), X
863 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
864 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
865
866 Value *Y;
867
868 // add (sub X, Y), -1 --> add (not Y), X
869 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
870 match(Op1, m_AllOnes()))
871 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
872
873 // zext(bool) + C -> bool ? C + 1 : C
874 if (match(Op0, m_ZExt(m_Value(X))) &&
875 X->getType()->getScalarSizeInBits() == 1)
876 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
877 // sext(bool) + C -> bool ? C - 1 : C
878 if (match(Op0, m_SExt(m_Value(X))) &&
879 X->getType()->getScalarSizeInBits() == 1)
880 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
881
882 // ~X + C --> (C-1) - X
883 if (match(Op0, m_Not(m_Value(X)))) {
884 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
885 auto *COne = ConstantInt::get(Op1C->getType(), 1);
886 bool WillNotSOV = willNotOverflowSignedSub(Op1C, COne, Add);
887 BinaryOperator *Res =
888 BinaryOperator::CreateSub(ConstantExpr::getSub(Op1C, COne), X);
889 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
890 return Res;
891 }
892
893 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
894 const APInt *C;
895 unsigned BitWidth = Ty->getScalarSizeInBits();
896 if (match(Op0, m_OneUse(m_AShr(m_Value(X),
898 match(Op1, m_One()))
899 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
900
901 if (!match(Op1, m_APInt(C)))
902 return nullptr;
903
904 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
905 Constant *Op01C;
906 if (match(Op0, m_DisjointOr(m_Value(X), m_ImmConstant(Op01C))))
907 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
908
909 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
910 const APInt *C2;
911 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
912 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
913
914 if (C->isSignMask()) {
915 // If wrapping is not allowed, then the addition must set the sign bit:
916 // X + (signmask) --> X | signmask
917 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
918 return BinaryOperator::CreateOr(Op0, Op1);
919
920 // If wrapping is allowed, then the addition flips the sign bit of LHS:
921 // X + (signmask) --> X ^ signmask
922 return BinaryOperator::CreateXor(Op0, Op1);
923 }
924
925 // Is this add the last step in a convoluted sext?
926 // add(zext(xor i16 X, -32768), -32768) --> sext X
927 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
928 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
929 return CastInst::Create(Instruction::SExt, X, Ty);
930
931 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
932 // (X ^ signmask) + C --> (X + (signmask ^ C))
933 if (C2->isSignMask())
934 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
935
936 // If X has no high-bits set above an xor mask:
937 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
938 if (C2->isMask()) {
939 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
940 if ((*C2 | LHSKnown.Zero).isAllOnes())
941 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
942 }
943
944 // Look for a math+logic pattern that corresponds to sext-in-register of a
945 // value with cleared high bits. Convert that into a pair of shifts:
946 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
947 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
948 if (Op0->hasOneUse() && *C2 == -(*C)) {
949 unsigned BitWidth = Ty->getScalarSizeInBits();
950 unsigned ShAmt = 0;
951 if (C->isPowerOf2())
952 ShAmt = BitWidth - C->logBase2() - 1;
953 else if (C2->isPowerOf2())
954 ShAmt = BitWidth - C2->logBase2() - 1;
956 0, &Add)) {
957 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
958 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
959 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
960 }
961 }
962 }
963
964 if (C->isOne() && Op0->hasOneUse()) {
965 // add (sext i1 X), 1 --> zext (not X)
966 // TODO: The smallest IR representation is (select X, 0, 1), and that would
967 // not require the one-use check. But we need to remove a transform in
968 // visitSelect and make sure that IR value tracking for select is equal or
969 // better than for these ops.
970 if (match(Op0, m_SExt(m_Value(X))) &&
971 X->getType()->getScalarSizeInBits() == 1)
972 return new ZExtInst(Builder.CreateNot(X), Ty);
973
974 // Shifts and add used to flip and mask off the low bit:
975 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
976 const APInt *C3;
977 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
978 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
979 Value *NotX = Builder.CreateNot(X);
980 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
981 }
982 }
983
984 // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
985 // TODO: There's a general form for any constant on the outer add.
986 if (C->isOne()) {
987 if (match(Op0, m_ZExt(m_Add(m_Value(X), m_AllOnes())))) {
989 if (llvm::isKnownNonZero(X, DL, 0, Q.AC, Q.CxtI, Q.DT))
990 return new ZExtInst(X, Ty);
991 }
992 }
993
994 return nullptr;
995}
996
997// match variations of a^2 + 2*a*b + b^2
998//
999// to reuse the code between the FP and Int versions, the instruction OpCodes
1000// and constant types have been turned into template parameters.
1001//
1002// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1003// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1004// (we're matching `X<<1` instead of `X*2` for Int)
1005template <bool FP, typename Mul2Rhs>
1006static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1007 Value *&B) {
1008 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1009 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1010 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1011
1012 // (a * a) + (((a * 2) + b) * b)
1013 if (match(&I, m_c_BinOp(
1014 AddOp, m_OneUse(m_BinOp(MulOp, m_Value(A), m_Deferred(A))),
1016 MulOp,
1017 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(A), M2Rhs),
1018 m_Value(B)),
1019 m_Deferred(B))))))
1020 return true;
1021
1022 // ((a * b) * 2) or ((a * 2) * b)
1023 // +
1024 // (a * a + b * b) or (b * b + a * a)
1025 return match(
1026 &I,
1027 m_c_BinOp(AddOp,
1030 Mul2Op, m_BinOp(MulOp, m_Value(A), m_Value(B)), M2Rhs)),
1031 m_OneUse(m_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(A), M2Rhs),
1032 m_Value(B)))),
1034 AddOp, m_BinOp(MulOp, m_Deferred(A), m_Deferred(A)),
1035 m_BinOp(MulOp, m_Deferred(B), m_Deferred(B))))));
1036}
1037
1038// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1040 Value *A, *B;
1041 if (matchesSquareSum</*FP*/ false>(I, m_SpecificInt(1), A, B)) {
1042 Value *AB = Builder.CreateAdd(A, B);
1043 return BinaryOperator::CreateMul(AB, AB);
1044 }
1045 return nullptr;
1046}
1047
1048// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1049// Requires `nsz` and `reassoc`.
1051 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1052 Value *A, *B;
1053 if (matchesSquareSum</*FP*/ true>(I, m_SpecificFP(2.0), A, B)) {
1054 Value *AB = Builder.CreateFAddFMF(A, B, &I);
1055 return BinaryOperator::CreateFMulFMF(AB, AB, &I);
1056 }
1057 return nullptr;
1058}
1059
1060// Matches multiplication expression Op * C where C is a constant. Returns the
1061// constant value in C and the other operand in Op. Returns true if such a
1062// match is found.
1063static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1064 const APInt *AI;
1065 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1066 C = *AI;
1067 return true;
1068 }
1069 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1070 C = APInt(AI->getBitWidth(), 1);
1071 C <<= *AI;
1072 return true;
1073 }
1074 return false;
1075}
1076
1077// Matches remainder expression Op % C where C is a constant. Returns the
1078// constant value in C and the other operand in Op. Returns the signedness of
1079// the remainder operation in IsSigned. Returns true if such a match is
1080// found.
1081static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1082 const APInt *AI;
1083 IsSigned = false;
1084 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1085 IsSigned = true;
1086 C = *AI;
1087 return true;
1088 }
1089 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1090 C = *AI;
1091 return true;
1092 }
1093 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1094 C = *AI + 1;
1095 return true;
1096 }
1097 return false;
1098}
1099
1100// Matches division expression Op / C with the given signedness as indicated
1101// by IsSigned, where C is a constant. Returns the constant value in C and the
1102// other operand in Op. Returns true if such a match is found.
1103static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1104 const APInt *AI;
1105 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1106 C = *AI;
1107 return true;
1108 }
1109 if (!IsSigned) {
1110 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1111 C = *AI;
1112 return true;
1113 }
1114 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1115 C = APInt(AI->getBitWidth(), 1);
1116 C <<= *AI;
1117 return true;
1118 }
1119 }
1120 return false;
1121}
1122
1123// Returns whether C0 * C1 with the given signedness overflows.
1124static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1125 bool overflow;
1126 if (IsSigned)
1127 (void)C0.smul_ov(C1, overflow);
1128 else
1129 (void)C0.umul_ov(C1, overflow);
1130 return overflow;
1131}
1132
1133// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1134// does not overflow.
1136 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1137 Value *X, *MulOpV;
1138 APInt C0, MulOpC;
1139 bool IsSigned;
1140 // Match I = X % C0 + MulOpV * C0
1141 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1142 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1143 C0 == MulOpC) {
1144 Value *RemOpV;
1145 APInt C1;
1146 bool Rem2IsSigned;
1147 // Match MulOpC = RemOpV % C1
1148 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1149 IsSigned == Rem2IsSigned) {
1150 Value *DivOpV;
1151 APInt DivOpC;
1152 // Match RemOpV = X / C0
1153 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1154 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1155 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1156 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1157 : Builder.CreateURem(X, NewDivisor, "urem");
1158 }
1159 }
1160 }
1161
1162 return nullptr;
1163}
1164
1165/// Fold
1166/// (1 << NBits) - 1
1167/// Into:
1168/// ~(-(1 << NBits))
1169/// Because a 'not' is better for bit-tracking analysis and other transforms
1170/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1172 InstCombiner::BuilderTy &Builder) {
1173 Value *NBits;
1174 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1175 return nullptr;
1176
1177 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1178 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1179 // Be wary of constant folding.
1180 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1181 // Always NSW. But NUW propagates from `add`.
1182 BOp->setHasNoSignedWrap();
1183 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1184 }
1185
1186 return BinaryOperator::CreateNot(NotMask, I.getName());
1187}
1188
1190 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1191 Type *Ty = I.getType();
1192 auto getUAddSat = [&]() {
1193 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1194 };
1195
1196 // add (umin X, ~Y), Y --> uaddsat X, Y
1197 Value *X, *Y;
1199 m_Deferred(Y))))
1200 return CallInst::Create(getUAddSat(), { X, Y });
1201
1202 // add (umin X, ~C), C --> uaddsat X, C
1203 const APInt *C, *NotC;
1204 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1205 *C == ~*NotC)
1206 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1207
1208 return nullptr;
1209}
1210
1211// Transform:
1212// (add A, (shl (neg B), Y))
1213// -> (sub A, (shl B, Y))
1215 const BinaryOperator &I) {
1216 Value *A, *B, *Cnt;
1217 if (match(&I,
1219 m_Value(A)))) {
1220 Value *NewShl = Builder.CreateShl(B, Cnt);
1221 return BinaryOperator::CreateSub(A, NewShl);
1222 }
1223 return nullptr;
1224}
1225
1226/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1228 // Division must be by power-of-2, but not the minimum signed value.
1229 Value *X;
1230 const APInt *DivC;
1231 if (!match(Add.getOperand(0), m_SDiv(m_Value(X), m_Power2(DivC))) ||
1232 DivC->isNegative())
1233 return nullptr;
1234
1235 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1236 // low bits set. It recognizes two canonical patterns:
1237 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1238 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1239 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1240 // Note that, by the time we end up here, if possible, ugt has been
1241 // canonicalized into eq.
1242 const APInt *MaskC, *MaskCCmp;
1244 if (!match(Add.getOperand(1),
1245 m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1246 m_APInt(MaskCCmp)))))
1247 return nullptr;
1248
1249 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1250 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1251 return nullptr;
1252
1253 APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1254 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1255 ? (*MaskC == (SMin | (*DivC - 1)))
1256 : (*DivC == 2 && *MaskC == SMin + 1);
1257 if (!IsMaskValid)
1258 return nullptr;
1259
1260 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1261 return BinaryOperator::CreateAShr(
1262 X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1263}
1264
1267 BinaryOperator &I) {
1268 assert((I.getOpcode() == Instruction::Add ||
1269 I.getOpcode() == Instruction::Or ||
1270 I.getOpcode() == Instruction::Sub) &&
1271 "Expecting add/or/sub instruction");
1272
1273 // We have a subtraction/addition between a (potentially truncated) *logical*
1274 // right-shift of X and a "select".
1275 Value *X, *Select;
1276 Instruction *LowBitsToSkip, *Extract;
1278 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1279 m_Instruction(Extract))),
1280 m_Value(Select))))
1281 return nullptr;
1282
1283 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1284 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1285 return nullptr;
1286
1287 Type *XTy = X->getType();
1288 bool HadTrunc = I.getType() != XTy;
1289
1290 // If there was a truncation of extracted value, then we'll need to produce
1291 // one extra instruction, so we need to ensure one instruction will go away.
1292 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1293 return nullptr;
1294
1295 // Extraction should extract high NBits bits, with shift amount calculated as:
1296 // low bits to skip = shift bitwidth - high bits to extract
1297 // The shift amount itself may be extended, and we need to look past zero-ext
1298 // when matching NBits, that will matter for matching later.
1299 Constant *C;
1300 Value *NBits;
1301 if (!match(
1302 LowBitsToSkip,
1305 APInt(C->getType()->getScalarSizeInBits(),
1306 X->getType()->getScalarSizeInBits()))))
1307 return nullptr;
1308
1309 // Sign-extending value can be zero-extended if we `sub`tract it,
1310 // or sign-extended otherwise.
1311 auto SkipExtInMagic = [&I](Value *&V) {
1312 if (I.getOpcode() == Instruction::Sub)
1313 match(V, m_ZExtOrSelf(m_Value(V)));
1314 else
1315 match(V, m_SExtOrSelf(m_Value(V)));
1316 };
1317
1318 // Now, finally validate the sign-extending magic.
1319 // `select` itself may be appropriately extended, look past that.
1320 SkipExtInMagic(Select);
1321
1323 const APInt *Thr;
1324 Value *SignExtendingValue, *Zero;
1325 bool ShouldSignext;
1326 // It must be a select between two values we will later establish to be a
1327 // sign-extending value and a zero constant. The condition guarding the
1328 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1329 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1330 m_Value(SignExtendingValue), m_Value(Zero))) ||
1331 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1332 return nullptr;
1333
1334 // icmp-select pair is commutative.
1335 if (!ShouldSignext)
1336 std::swap(SignExtendingValue, Zero);
1337
1338 // If we should not perform sign-extension then we must add/or/subtract zero.
1339 if (!match(Zero, m_Zero()))
1340 return nullptr;
1341 // Otherwise, it should be some constant, left-shifted by the same NBits we
1342 // had in `lshr`. Said left-shift can also be appropriately extended.
1343 // Again, we must look past zero-ext when looking for NBits.
1344 SkipExtInMagic(SignExtendingValue);
1345 Constant *SignExtendingValueBaseConstant;
1346 if (!match(SignExtendingValue,
1347 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1348 m_ZExtOrSelf(m_Specific(NBits)))))
1349 return nullptr;
1350 // If we `sub`, then the constant should be one, else it should be all-ones.
1351 if (I.getOpcode() == Instruction::Sub
1352 ? !match(SignExtendingValueBaseConstant, m_One())
1353 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1354 return nullptr;
1355
1356 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1357 Extract->getName() + ".sext");
1358 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1359 if (!HadTrunc)
1360 return NewAShr;
1361
1362 Builder.Insert(NewAShr);
1363 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1364}
1365
1366/// This is a specialization of a more general transform from
1367/// foldUsingDistributiveLaws. If that code can be made to work optimally
1368/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1370 InstCombiner::BuilderTy &Builder) {
1371 // TODO: Also handle mul by doubling the shift amount?
1372 assert((I.getOpcode() == Instruction::Add ||
1373 I.getOpcode() == Instruction::Sub) &&
1374 "Expected add/sub");
1375 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1376 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1377 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1378 return nullptr;
1379
1380 Value *X, *Y, *ShAmt;
1381 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1382 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1383 return nullptr;
1384
1385 // No-wrap propagates only when all ops have no-wrap.
1386 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1387 Op1->hasNoSignedWrap();
1388 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1389 Op1->hasNoUnsignedWrap();
1390
1391 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1392 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1393 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1394 NewI->setHasNoSignedWrap(HasNSW);
1395 NewI->setHasNoUnsignedWrap(HasNUW);
1396 }
1397 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1398 NewShl->setHasNoSignedWrap(HasNSW);
1399 NewShl->setHasNoUnsignedWrap(HasNUW);
1400 return NewShl;
1401}
1402
1403/// Reduce a sequence of masked half-width multiplies to a single multiply.
1404/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1406 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1407 // Skip the odd bitwidth types.
1408 if ((BitWidth & 0x1))
1409 return nullptr;
1410
1411 unsigned HalfBits = BitWidth >> 1;
1412 APInt HalfMask = APInt::getMaxValue(HalfBits);
1413
1414 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1415 Value *XLo, *YLo;
1416 Value *CrossSum;
1417 // Require one-use on the multiply to avoid increasing the number of
1418 // multiplications.
1419 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1420 m_OneUse(m_Mul(m_Value(YLo), m_Value(XLo))))))
1421 return nullptr;
1422
1423 // XLo = X & HalfMask
1424 // YLo = Y & HalfMask
1425 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1426 // to enhance robustness
1427 Value *X, *Y;
1428 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1429 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1430 return nullptr;
1431
1432 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1433 // X' can be either X or XLo in the pattern (and the same for Y')
1434 if (match(CrossSum,
1439 return BinaryOperator::CreateMul(X, Y);
1440
1441 return nullptr;
1442}
1443
1445 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1446 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1448 return replaceInstUsesWith(I, V);
1449
1451 return &I;
1452
1454 return X;
1455
1457 return Phi;
1458
1459 // (A*B)+(A*C) -> A*(B+C) etc
1461 return replaceInstUsesWith(I, V);
1462
1463 if (Instruction *R = foldBoxMultiply(I))
1464 return R;
1465
1467 return R;
1468
1470 return X;
1471
1473 return X;
1474
1476 return R;
1477
1479 return R;
1480
1481 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1482 Type *Ty = I.getType();
1483 if (Ty->isIntOrIntVectorTy(1))
1484 return BinaryOperator::CreateXor(LHS, RHS);
1485
1486 // X + X --> X << 1
1487 if (LHS == RHS) {
1488 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1489 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1490 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1491 return Shl;
1492 }
1493
1494 Value *A, *B;
1495 if (match(LHS, m_Neg(m_Value(A)))) {
1496 // -A + -B --> -(A + B)
1497 if (match(RHS, m_Neg(m_Value(B))))
1499
1500 // -A + B --> B - A
1501 auto *Sub = BinaryOperator::CreateSub(RHS, A);
1502 auto *OB0 = cast<OverflowingBinaryOperator>(LHS);
1503 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1504
1505 return Sub;
1506 }
1507
1508 // A + -B --> A - B
1509 if (match(RHS, m_Neg(m_Value(B))))
1510 return BinaryOperator::CreateSub(LHS, B);
1511
1513 return replaceInstUsesWith(I, V);
1514
1515 // (A + 1) + ~B --> A - B
1516 // ~B + (A + 1) --> A - B
1517 // (~B + A) + 1 --> A - B
1518 // (A + ~B) + 1 --> A - B
1519 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1521 return BinaryOperator::CreateSub(A, B);
1522
1523 // (A + RHS) + RHS --> A + (RHS << 1)
1525 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1526
1527 // LHS + (A + LHS) --> A + (LHS << 1)
1529 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1530
1531 {
1532 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1533 Constant *C1, *C2;
1534 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1535 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1536 (LHS->hasOneUse() || RHS->hasOneUse())) {
1537 Value *Sub = Builder.CreateSub(A, B);
1538 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1539 }
1540
1541 // Canonicalize a constant sub operand as an add operand for better folding:
1542 // (C1 - A) + B --> (B - A) + C1
1544 m_Value(B)))) {
1545 Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1546 return BinaryOperator::CreateAdd(Sub, C1);
1547 }
1548 }
1549
1550 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1552
1553 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1554 const APInt *C1, *C2;
1555 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1556 APInt one(C2->getBitWidth(), 1);
1557 APInt minusC1 = -(*C1);
1558 if (minusC1 == (one << *C2)) {
1559 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1560 return BinaryOperator::CreateSRem(RHS, NewRHS);
1561 }
1562 }
1563
1564 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1565 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1566 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1567 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1568 return BinaryOperator::CreateAnd(A, NewMask);
1569 }
1570
1571 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1572 if ((match(RHS, m_ZExt(m_Value(A))) &&
1574 (match(LHS, m_ZExt(m_Value(A))) &&
1576 return new ZExtInst(B, LHS->getType());
1577
1578 // zext(A) + sext(A) --> 0 if A is i1
1580 A->getType()->isIntOrIntVectorTy(1))
1581 return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1582
1583 // A+B --> A|B iff A and B have no bits set in common.
1584 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1585 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(&I)))
1586 return BinaryOperator::CreateDisjointOr(LHS, RHS);
1587
1588 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1589 return Ext;
1590
1591 // (add (xor A, B) (and A, B)) --> (or A, B)
1592 // (add (and A, B) (xor A, B)) --> (or A, B)
1593 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1595 return BinaryOperator::CreateOr(A, B);
1596
1597 // (add (or A, B) (and A, B)) --> (add A, B)
1598 // (add (and A, B) (or A, B)) --> (add A, B)
1599 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1601 // Replacing operands in-place to preserve nuw/nsw flags.
1602 replaceOperand(I, 0, A);
1603 replaceOperand(I, 1, B);
1604 return &I;
1605 }
1606
1607 // (add A (or A, -A)) --> (and (add A, -1) A)
1608 // (add A (or -A, A)) --> (and (add A, -1) A)
1609 // (add (or A, -A) A) --> (and (add A, -1) A)
1610 // (add (or -A, A) A) --> (and (add A, -1) A)
1612 m_Deferred(A)))))) {
1613 Value *Add =
1615 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1616 return BinaryOperator::CreateAnd(Add, A);
1617 }
1618
1619 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1620 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1621 if (match(&I,
1623 m_AllOnes()))) {
1625 Value *Dec = Builder.CreateAdd(A, AllOnes);
1626 Value *Not = Builder.CreateXor(A, AllOnes);
1627 return BinaryOperator::CreateAnd(Dec, Not);
1628 }
1629
1630 // Disguised reassociation/factorization:
1631 // ~(A * C1) + A
1632 // ((A * -C1) - 1) + A
1633 // ((A * -C1) + A) - 1
1634 // (A * (1 - C1)) - 1
1635 if (match(&I,
1637 m_Deferred(A)))) {
1638 Type *Ty = I.getType();
1639 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1640 Value *NewMul = Builder.CreateMul(A, NewMulC);
1641 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1642 }
1643
1644 // (A * -2**C) + B --> B - (A << C)
1645 const APInt *NegPow2C;
1646 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1647 m_Value(B)))) {
1648 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1649 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1650 return BinaryOperator::CreateSub(B, Shl);
1651 }
1652
1653 // Canonicalize signum variant that ends in add:
1654 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1659 m_OneUse(m_ICmp(Pred, m_Specific(A), m_ZeroInt()))))) &&
1660 Pred == CmpInst::ICMP_SGT) {
1661 Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1662 Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1663 return BinaryOperator::CreateOr(LHS, Zext);
1664 }
1665
1666 if (Instruction *Ashr = foldAddToAshr(I))
1667 return Ashr;
1668
1669 // (~X) + (~Y) --> -2 - (X + Y)
1670 {
1671 // To ensure we can save instructions we need to ensure that we consume both
1672 // LHS/RHS (i.e they have a `not`).
1673 bool ConsumesLHS, ConsumesRHS;
1674 if (isFreeToInvert(LHS, LHS->hasOneUse(), ConsumesLHS) && ConsumesLHS &&
1675 isFreeToInvert(RHS, RHS->hasOneUse(), ConsumesRHS) && ConsumesRHS) {
1678 assert(NotLHS != nullptr && NotRHS != nullptr &&
1679 "isFreeToInvert desynced with getFreelyInverted");
1680 Value *LHSPlusRHS = Builder.CreateAdd(NotLHS, NotRHS);
1681 return BinaryOperator::CreateSub(
1682 ConstantInt::getSigned(RHS->getType(), -2), LHSPlusRHS);
1683 }
1684 }
1685
1687 return R;
1688
1689 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1690 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1691 // computeKnownBits.
1692 bool Changed = false;
1693 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHSCache, RHSCache, I)) {
1694 Changed = true;
1695 I.setHasNoSignedWrap(true);
1696 }
1697 if (!I.hasNoUnsignedWrap() &&
1698 willNotOverflowUnsignedAdd(LHSCache, RHSCache, I)) {
1699 Changed = true;
1700 I.setHasNoUnsignedWrap(true);
1701 }
1702
1704 return V;
1705
1706 if (Instruction *V =
1708 return V;
1709
1711 return SatAdd;
1712
1713 // usub.sat(A, B) + B => umax(A, B)
1714 if (match(&I, m_c_BinOp(
1715 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1716 m_Deferred(B)))) {
1717 return replaceInstUsesWith(I,
1718 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1719 }
1720
1721 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1722 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1723 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1725 return replaceInstUsesWith(
1726 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1727 {Builder.CreateOr(A, B)}));
1728
1729 // Fold the log2_ceil idiom:
1730 // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
1731 // -->
1732 // BW - ctlz(A - 1, false)
1733 const APInt *XorC;
1734 if (match(&I,
1735 m_c_Add(
1736 m_ZExt(m_ICmp(Pred, m_Intrinsic<Intrinsic::ctpop>(m_Value(A)),
1737 m_One())),
1740 m_Intrinsic<Intrinsic::ctlz>(m_Deferred(A), m_One())))),
1741 m_APInt(XorC))))))) &&
1742 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
1743 *XorC == A->getType()->getScalarSizeInBits() - 1) {
1744 Value *Sub = Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()));
1745 Value *Ctlz = Builder.CreateIntrinsic(Intrinsic::ctlz, {A->getType()},
1746 {Sub, Builder.getFalse()});
1747 Value *Ret = Builder.CreateSub(
1748 ConstantInt::get(A->getType(), A->getType()->getScalarSizeInBits()),
1749 Ctlz, "", /*HasNUW*/ true, /*HasNSW*/ true);
1750 return replaceInstUsesWith(I, Builder.CreateZExtOrTrunc(Ret, I.getType()));
1751 }
1752
1753 if (Instruction *Res = foldSquareSumInt(I))
1754 return Res;
1755
1756 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1757 return Res;
1758
1760 return Res;
1761
1762 return Changed ? &I : nullptr;
1763}
1764
1765/// Eliminate an op from a linear interpolation (lerp) pattern.
1767 InstCombiner::BuilderTy &Builder) {
1768 Value *X, *Y, *Z;
1771 m_Value(Z))))),
1773 return nullptr;
1774
1775 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1776 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1777 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1778 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1779}
1780
1781/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1783 InstCombiner::BuilderTy &Builder) {
1784 assert((I.getOpcode() == Instruction::FAdd ||
1785 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1786 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1787 "FP factorization requires FMF");
1788
1789 if (Instruction *Lerp = factorizeLerp(I, Builder))
1790 return Lerp;
1791
1792 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1793 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1794 return nullptr;
1795
1796 Value *X, *Y, *Z;
1797 bool IsFMul;
1798 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1799 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1800 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1801 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1802 IsFMul = true;
1803 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1804 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1805 IsFMul = false;
1806 else
1807 return nullptr;
1808
1809 // (X * Z) + (Y * Z) --> (X + Y) * Z
1810 // (X * Z) - (Y * Z) --> (X - Y) * Z
1811 // (X / Z) + (Y / Z) --> (X + Y) / Z
1812 // (X / Z) - (Y / Z) --> (X - Y) / Z
1813 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1814 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1815 : Builder.CreateFSubFMF(X, Y, &I);
1816
1817 // Bail out if we just created a denormal constant.
1818 // TODO: This is copied from a previous implementation. Is it necessary?
1819 const APFloat *C;
1820 if (match(XY, m_APFloat(C)) && !C->isNormal())
1821 return nullptr;
1822
1823 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1825}
1826
1828 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1829 I.getFastMathFlags(),
1831 return replaceInstUsesWith(I, V);
1832
1834 return &I;
1835
1837 return X;
1838
1840 return Phi;
1841
1842 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1843 return FoldedFAdd;
1844
1845 // (-X) + Y --> Y - X
1846 Value *X, *Y;
1847 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1849
1850 // Similar to above, but look through fmul/fdiv for the negated term.
1851 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1852 Value *Z;
1854 m_Value(Z)))) {
1855 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1856 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1857 }
1858 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1859 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1861 m_Value(Z))) ||
1863 m_Value(Z)))) {
1864 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1865 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1866 }
1867
1868 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1869 // integer add followed by a promotion.
1870 if (Instruction *R = foldFBinOpOfIntCasts(I))
1871 return R;
1872
1873 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1874 // Handle specials cases for FAdd with selects feeding the operation
1876 return replaceInstUsesWith(I, V);
1877
1878 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1880 return F;
1881
1883 return F;
1884
1885 // Try to fold fadd into start value of reduction intrinsic.
1886 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1887 m_AnyZeroFP(), m_Value(X))),
1888 m_Value(Y)))) {
1889 // fadd (rdx 0.0, X), Y --> rdx Y, X
1890 return replaceInstUsesWith(
1891 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1892 {X->getType()}, {Y, X}, &I));
1893 }
1894 const APFloat *StartC, *C;
1895 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1896 m_APFloat(StartC), m_Value(X)))) &&
1897 match(RHS, m_APFloat(C))) {
1898 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1899 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1900 return replaceInstUsesWith(
1901 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1902 {X->getType()}, {NewStartC, X}, &I));
1903 }
1904
1905 // (X * MulC) + X --> X * (MulC + 1.0)
1906 Constant *MulC;
1907 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1908 m_Deferred(X)))) {
1910 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1911 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1912 }
1913
1914 // (-X - Y) + (X + Z) --> Z - Y
1916 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1917 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1918
1919 if (Value *V = FAddCombine(Builder).simplify(&I))
1920 return replaceInstUsesWith(I, V);
1921 }
1922
1923 // minumum(X, Y) + maximum(X, Y) => X + Y.
1924 if (match(&I,
1925 m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1926 m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1927 m_Deferred(Y))))) {
1929 // We cannot preserve ninf if nnan flag is not set.
1930 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1931 // while in optimized version NaN + Inf and this is a poison with ninf flag.
1932 if (!Result->hasNoNaNs())
1933 Result->setHasNoInfs(false);
1934 return Result;
1935 }
1936
1937 return nullptr;
1938}
1939
1940/// Optimize pointer differences into the same array into a size. Consider:
1941/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1942/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1944 Type *Ty, bool IsNUW) {
1945 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1946 // this.
1947 bool Swapped = false;
1948 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1949 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1950 std::swap(LHS, RHS);
1951 Swapped = true;
1952 }
1953
1954 // Require at least one GEP with a common base pointer on both sides.
1955 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1956 // (gep X, ...) - X
1957 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1959 GEP1 = LHSGEP;
1960 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1961 // (gep X, ...) - (gep X, ...)
1962 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1963 RHSGEP->getOperand(0)->stripPointerCasts()) {
1964 GEP1 = LHSGEP;
1965 GEP2 = RHSGEP;
1966 }
1967 }
1968 }
1969
1970 if (!GEP1)
1971 return nullptr;
1972
1973 if (GEP2) {
1974 // (gep X, ...) - (gep X, ...)
1975 //
1976 // Avoid duplicating the arithmetic if there are more than one non-constant
1977 // indices between the two GEPs and either GEP has a non-constant index and
1978 // multiple users. If zero non-constant index, the result is a constant and
1979 // there is no duplication. If one non-constant index, the result is an add
1980 // or sub with a constant, which is no larger than the original code, and
1981 // there's no duplicated arithmetic, even if either GEP has multiple
1982 // users. If more than one non-constant indices combined, as long as the GEP
1983 // with at least one non-constant index doesn't have multiple users, there
1984 // is no duplication.
1985 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1986 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1987 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1988 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1989 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1990 return nullptr;
1991 }
1992 }
1993
1994 // Emit the offset of the GEP and an intptr_t.
1995 Value *Result = EmitGEPOffset(GEP1);
1996
1997 // If this is a single inbounds GEP and the original sub was nuw,
1998 // then the final multiplication is also nuw.
1999 if (auto *I = dyn_cast<Instruction>(Result))
2000 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
2001 I->getOpcode() == Instruction::Mul)
2002 I->setHasNoUnsignedWrap();
2003
2004 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2005 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2006 if (GEP2) {
2007 Value *Offset = EmitGEPOffset(GEP2);
2008 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2009 GEP1->isInBounds() && GEP2->isInBounds());
2010 }
2011
2012 // If we have p - gep(p, ...) then we have to negate the result.
2013 if (Swapped)
2014 Result = Builder.CreateNeg(Result, "diff.neg");
2015
2016 return Builder.CreateIntCast(Result, Ty, true);
2017}
2018
2020 InstCombiner::BuilderTy &Builder) {
2021 Value *Op0 = I.getOperand(0);
2022 Value *Op1 = I.getOperand(1);
2023 Type *Ty = I.getType();
2024 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2025 if (!MinMax)
2026 return nullptr;
2027
2028 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2029 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2030 Value *X = MinMax->getLHS();
2031 Value *Y = MinMax->getRHS();
2032 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2033 (Op0->hasOneUse() || Op1->hasOneUse())) {
2034 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2035 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2036 return CallInst::Create(F, {X, Y});
2037 }
2038
2039 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2040 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2041 Value *Z;
2042 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2043 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2044 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2045 return BinaryOperator::CreateAdd(X, USub);
2046 }
2047 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2048 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2049 return BinaryOperator::CreateAdd(X, USub);
2050 }
2051 }
2052
2053 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2054 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2055 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2056 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2057 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2058 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2059 return CallInst::Create(F, {Op0, Z});
2060 }
2061
2062 return nullptr;
2063}
2064
2066 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2067 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2069 return replaceInstUsesWith(I, V);
2070
2072 return X;
2073
2075 return Phi;
2076
2077 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2078
2079 // If this is a 'B = x-(-A)', change to B = x+A.
2080 // We deal with this without involving Negator to preserve NSW flag.
2081 if (Value *V = dyn_castNegVal(Op1)) {
2082 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2083
2084 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2085 assert(BO->getOpcode() == Instruction::Sub &&
2086 "Expected a subtraction operator!");
2087 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2088 Res->setHasNoSignedWrap(true);
2089 } else {
2090 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2091 Res->setHasNoSignedWrap(true);
2092 }
2093
2094 return Res;
2095 }
2096
2097 // Try this before Negator to preserve NSW flag.
2099 return R;
2100
2101 Constant *C;
2102 if (match(Op0, m_ImmConstant(C))) {
2103 Value *X;
2104 Constant *C2;
2105
2106 // C-(X+C2) --> (C-C2)-X
2107 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2108 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2109 // => (C-C2)-X can have NSW/NUW
2110 bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2111 BinaryOperator *Res =
2112 BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2113 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2114 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2115 WillNotSOV);
2116 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2117 OBO1->hasNoUnsignedWrap());
2118 return Res;
2119 }
2120 }
2121
2122 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2123 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2124 return Ext;
2125
2126 bool Changed = false;
2127 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2128 Changed = true;
2129 I.setHasNoSignedWrap(true);
2130 }
2131 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2132 Changed = true;
2133 I.setHasNoUnsignedWrap(true);
2134 }
2135
2136 return Changed ? &I : nullptr;
2137 };
2138
2139 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2140 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2141 // a pure negation used by a select that looks like abs/nabs.
2142 bool IsNegation = match(Op0, m_ZeroInt());
2143 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2144 const Instruction *UI = dyn_cast<Instruction>(U);
2145 if (!UI)
2146 return false;
2147 return match(UI,
2148 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2149 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2150 })) {
2151 if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2152 I.hasNoSignedWrap(),
2153 Op1, *this))
2154 return BinaryOperator::CreateAdd(NegOp1, Op0);
2155 }
2156 if (IsNegation)
2157 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2158
2159 // (A*B)-(A*C) -> A*(B-C) etc
2161 return replaceInstUsesWith(I, V);
2162
2163 if (I.getType()->isIntOrIntVectorTy(1))
2164 return BinaryOperator::CreateXor(Op0, Op1);
2165
2166 // Replace (-1 - A) with (~A).
2167 if (match(Op0, m_AllOnes()))
2168 return BinaryOperator::CreateNot(Op1);
2169
2170 // (X + -1) - Y --> ~Y + X
2171 Value *X, *Y;
2172 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2173 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2174
2175 // Reassociate sub/add sequences to create more add instructions and
2176 // reduce dependency chains:
2177 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2178 Value *Z;
2180 m_Value(Z))))) {
2181 Value *XZ = Builder.CreateAdd(X, Z);
2182 Value *YW = Builder.CreateAdd(Y, Op1);
2183 return BinaryOperator::CreateSub(XZ, YW);
2184 }
2185
2186 // ((X - Y) - Op1) --> X - (Y + Op1)
2187 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2188 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Op0);
2189 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2190 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2191 Value *Add = Builder.CreateAdd(Y, Op1, "", /* HasNUW */ HasNUW,
2192 /* HasNSW */ HasNSW);
2193 BinaryOperator *Sub = BinaryOperator::CreateSub(X, Add);
2194 Sub->setHasNoUnsignedWrap(HasNUW);
2195 Sub->setHasNoSignedWrap(HasNSW);
2196 return Sub;
2197 }
2198
2199 {
2200 // (X + Z) - (Y + Z) --> (X - Y)
2201 // This is done in other passes, but we want to be able to consume this
2202 // pattern in InstCombine so we can generate it without creating infinite
2203 // loops.
2204 if (match(Op0, m_Add(m_Value(X), m_Value(Z))) &&
2205 match(Op1, m_c_Add(m_Value(Y), m_Specific(Z))))
2206 return BinaryOperator::CreateSub(X, Y);
2207
2208 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2209 Constant *CX, *CY;
2210 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_ImmConstant(CX)))) &&
2211 match(Op1, m_OneUse(m_Add(m_Value(Y), m_ImmConstant(CY))))) {
2212 Value *OpsSub = Builder.CreateSub(X, Y);
2213 Constant *ConstsSub = ConstantExpr::getSub(CX, CY);
2214 return BinaryOperator::CreateAdd(OpsSub, ConstsSub);
2215 }
2216 }
2217
2218 // (~X) - (~Y) --> Y - X
2219 {
2220 // Need to ensure we can consume at least one of the `not` instructions,
2221 // otherwise this can inf loop.
2222 bool ConsumesOp0, ConsumesOp1;
2223 if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
2224 isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
2225 (ConsumesOp0 || ConsumesOp1)) {
2226 Value *NotOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
2227 Value *NotOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
2228 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2229 "isFreeToInvert desynced with getFreelyInverted");
2230 return BinaryOperator::CreateSub(NotOp1, NotOp0);
2231 }
2232 }
2233
2234 auto m_AddRdx = [](Value *&Vec) {
2235 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2236 };
2237 Value *V0, *V1;
2238 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2239 V0->getType() == V1->getType()) {
2240 // Difference of sums is sum of differences:
2241 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2242 Value *Sub = Builder.CreateSub(V0, V1);
2243 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2244 {Sub->getType()}, {Sub});
2245 return replaceInstUsesWith(I, Rdx);
2246 }
2247
2248 if (Constant *C = dyn_cast<Constant>(Op0)) {
2249 Value *X;
2250 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2251 // C - (zext bool) --> bool ? C - 1 : C
2253 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2254 // C - (sext bool) --> bool ? C + 1 : C
2256
2257 // C - ~X == X + (1+C)
2258 if (match(Op1, m_Not(m_Value(X))))
2259 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2260
2261 // Try to fold constant sub into select arguments.
2262 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2263 if (Instruction *R = FoldOpIntoSelect(I, SI))
2264 return R;
2265
2266 // Try to fold constant sub into PHI values.
2267 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2268 if (Instruction *R = foldOpIntoPhi(I, PN))
2269 return R;
2270
2271 Constant *C2;
2272
2273 // C-(C2-X) --> X+(C-C2)
2274 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2275 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2276 }
2277
2278 const APInt *Op0C;
2279 if (match(Op0, m_APInt(Op0C))) {
2280 if (Op0C->isMask()) {
2281 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2282 // zero.
2283 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
2284 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2285 return BinaryOperator::CreateXor(Op1, Op0);
2286 }
2287
2288 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2289 // (C3 - ((C2 & C3) - 1)) is pow2
2290 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2291 // C2 is negative pow2 || sub nuw
2292 const APInt *C2, *C3;
2293 BinaryOperator *InnerSub;
2294 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2295 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2296 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2297 APInt C2AndC3 = *C2 & *C3;
2298 APInt C2AndC3Minus1 = C2AndC3 - 1;
2299 APInt C2AddC3 = *C2 + *C3;
2300 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2301 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2302 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2303 return BinaryOperator::CreateAdd(
2304 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2305 }
2306 }
2307 }
2308
2309 {
2310 Value *Y;
2311 // X-(X+Y) == -Y X-(Y+X) == -Y
2312 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2314
2315 // (X-Y)-X == -Y
2316 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2318 }
2319
2320 // (sub (or A, B) (and A, B)) --> (xor A, B)
2321 {
2322 Value *A, *B;
2323 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2324 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2325 return BinaryOperator::CreateXor(A, B);
2326 }
2327
2328 // (sub (add A, B) (or A, B)) --> (and A, B)
2329 {
2330 Value *A, *B;
2331 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2332 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2333 return BinaryOperator::CreateAnd(A, B);
2334 }
2335
2336 // (sub (add A, B) (and A, B)) --> (or A, B)
2337 {
2338 Value *A, *B;
2339 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2341 return BinaryOperator::CreateOr(A, B);
2342 }
2343
2344 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2345 {
2346 Value *A, *B;
2347 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2348 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2349 (Op0->hasOneUse() || Op1->hasOneUse()))
2351 }
2352
2353 // (sub (or A, B), (xor A, B)) --> (and A, B)
2354 {
2355 Value *A, *B;
2356 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2357 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2358 return BinaryOperator::CreateAnd(A, B);
2359 }
2360
2361 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2362 {
2363 Value *A, *B;
2364 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2365 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2366 (Op0->hasOneUse() || Op1->hasOneUse()))
2368 }
2369
2370 {
2371 Value *Y;
2372 // ((X | Y) - X) --> (~X & Y)
2373 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2374 return BinaryOperator::CreateAnd(
2375 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2376 }
2377
2378 {
2379 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2380 Value *X;
2381 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2382 m_OneUse(m_Neg(m_Value(X))))))) {
2384 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2385 }
2386 }
2387
2388 {
2389 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2390 Constant *C;
2391 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2394 }
2395 }
2396
2397 {
2398 // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2399 // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2400 Value *C, *X;
2401 auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2402 return match(LHS, m_OneUse(m_c_Xor(m_Value(X), m_Specific(RHS)))) &&
2403 match(RHS, m_SExt(m_Value(C))) &&
2404 (C->getType()->getScalarSizeInBits() == 1);
2405 };
2406 if (m_SubXorCmp(Op0, Op1))
2408 if (m_SubXorCmp(Op1, Op0))
2410 }
2411
2413 return R;
2414
2416 return R;
2417
2418 {
2419 // If we have a subtraction between some value and a select between
2420 // said value and something else, sink subtraction into select hands, i.e.:
2421 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2422 // ->
2423 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2424 // or
2425 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2426 // ->
2427 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2428 // This will result in select between new subtraction and 0.
2429 auto SinkSubIntoSelect =
2430 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2431 auto SubBuilder) -> Instruction * {
2432 Value *Cond, *TrueVal, *FalseVal;
2433 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2434 m_Value(FalseVal)))))
2435 return nullptr;
2436 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2437 return nullptr;
2438 // While it is really tempting to just create two subtractions and let
2439 // InstCombine fold one of those to 0, it isn't possible to do so
2440 // because of worklist visitation order. So ugly it is.
2441 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2442 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2443 Constant *Zero = Constant::getNullValue(Ty);
2444 SelectInst *NewSel =
2445 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2446 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2447 // Preserve prof metadata if any.
2448 NewSel->copyMetadata(cast<Instruction>(*Select));
2449 return NewSel;
2450 };
2451 if (Instruction *NewSel = SinkSubIntoSelect(
2452 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2453 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2454 return Builder->CreateSub(OtherHandOfSelect,
2455 /*OtherHandOfSub=*/Op1);
2456 }))
2457 return NewSel;
2458 if (Instruction *NewSel = SinkSubIntoSelect(
2459 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2460 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2461 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2462 OtherHandOfSelect);
2463 }))
2464 return NewSel;
2465 }
2466
2467 // (X - (X & Y)) --> (X & ~Y)
2468 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2469 (Op1->hasOneUse() || isa<Constant>(Y)))
2470 return BinaryOperator::CreateAnd(
2471 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2472
2473 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2474 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2475 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2476 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2477 // As long as Y is freely invertible, this will be neutral or a win.
2478 // Note: We don't generate the inverse max/min, just create the 'not' of
2479 // it and let other folds do the rest.
2480 if (match(Op0, m_Not(m_Value(X))) &&
2481 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2482 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2483 Value *Not = Builder.CreateNot(Op1);
2484 return BinaryOperator::CreateSub(Not, X);
2485 }
2486 if (match(Op1, m_Not(m_Value(X))) &&
2487 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2488 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2489 Value *Not = Builder.CreateNot(Op0);
2490 return BinaryOperator::CreateSub(X, Not);
2491 }
2492
2493 // Optimize pointer differences into the same array into a size. Consider:
2494 // &A[10] - &A[0]: we should compile this to "10".
2495 Value *LHSOp, *RHSOp;
2496 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2497 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2498 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2499 I.hasNoUnsignedWrap()))
2500 return replaceInstUsesWith(I, Res);
2501
2502 // trunc(p)-trunc(q) -> trunc(p-q)
2503 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2504 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2505 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2506 /* IsNUW */ false))
2507 return replaceInstUsesWith(I, Res);
2508
2509 // Canonicalize a shifty way to code absolute value to the common pattern.
2510 // There are 2 potential commuted variants.
2511 // We're relying on the fact that we only do this transform when the shift has
2512 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2513 // instructions).
2514 Value *A;
2515 const APInt *ShAmt;
2516 Type *Ty = I.getType();
2517 unsigned BitWidth = Ty->getScalarSizeInBits();
2518 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2519 Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2520 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2521 // B = ashr i32 A, 31 ; smear the sign bit
2522 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2523 // --> (A < 0) ? -A : A
2524 Value *IsNeg = Builder.CreateIsNeg(A);
2525 // Copy the nuw/nsw flags from the sub to the negate.
2526 Value *NegA = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2527 I.hasNoSignedWrap());
2528 return SelectInst::Create(IsNeg, NegA, A);
2529 }
2530
2531 // If we are subtracting a low-bit masked subset of some value from an add
2532 // of that same value with no low bits changed, that is clearing some low bits
2533 // of the sum:
2534 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2535 const APInt *AddC, *AndC;
2536 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2537 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2538 unsigned Cttz = AddC->countr_zero();
2539 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2540 if ((HighMask & *AndC).isZero())
2541 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2542 }
2543
2544 if (Instruction *V =
2546 return V;
2547
2548 // X - usub.sat(X, Y) => umin(X, Y)
2549 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2550 m_Value(Y)))))
2551 return replaceInstUsesWith(
2552 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2553
2554 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2555 // TODO: The one-use restriction is not strictly necessary, but it may
2556 // require improving other pattern matching and/or codegen.
2557 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2558 return replaceInstUsesWith(
2559 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2560
2561 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2562 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2563 return replaceInstUsesWith(
2564 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2565
2566 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2567 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2568 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2569 return BinaryOperator::CreateNeg(USub);
2570 }
2571
2572 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2573 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2574 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2575 return BinaryOperator::CreateNeg(USub);
2576 }
2577
2578 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2579 if (match(Op0, m_SpecificInt(BitWidth)) &&
2580 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2581 return replaceInstUsesWith(
2582 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2583 {Builder.CreateNot(X)}));
2584
2585 // Reduce multiplies for difference-of-squares by factoring:
2586 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2587 if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2588 match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2589 auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2590 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2591 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2592 OBO1->hasNoSignedWrap() && BitWidth > 2;
2593 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2594 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2595 Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2596 Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2597 Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2598 return replaceInstUsesWith(I, Mul);
2599 }
2600
2601 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2602 if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2604 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2605 Value *Sub =
2606 Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2607 Value *Call =
2608 Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2609 return replaceInstUsesWith(I, Call);
2610 }
2611 }
2612
2614 return Res;
2615
2616 return TryToNarrowDeduceFlags();
2617}
2618
2619/// This eliminates floating-point negation in either 'fneg(X)' or
2620/// 'fsub(-0.0, X)' form by combining into a constant operand.
2622 // This is limited with one-use because fneg is assumed better for
2623 // reassociation and cheaper in codegen than fmul/fdiv.
2624 // TODO: Should the m_OneUse restriction be removed?
2625 Instruction *FNegOp;
2626 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2627 return nullptr;
2628
2629 Value *X;
2630 Constant *C;
2631
2632 // Fold negation into constant operand.
2633 // -(X * C) --> X * (-C)
2634 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2635 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2636 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2637 // -(X / C) --> X / (-C)
2638 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2639 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2640 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2641 // -(C / X) --> (-C) / X
2642 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2643 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2645
2646 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2647 // not apply to the fdiv. Everything else propagates from the fneg.
2648 // TODO: We could propagate nsz/ninf from fdiv alone?
2649 FastMathFlags FMF = I.getFastMathFlags();
2650 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2651 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2652 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2653 return FDiv;
2654 }
2655 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2656 // -(X + C) --> -X + -C --> -C - X
2657 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2658 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2659 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2660
2661 return nullptr;
2662}
2663
2664Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2665 Instruction &FMFSource) {
2666 Value *X, *Y;
2667 if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2668 return cast<Instruction>(Builder.CreateFMulFMF(
2669 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2670 }
2671
2672 if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2673 return cast<Instruction>(Builder.CreateFDivFMF(
2674 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2675 }
2676
2677 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2678 // Make sure to preserve flags and metadata on the call.
2679 if (II->getIntrinsicID() == Intrinsic::ldexp) {
2680 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2683
2685 II->getCalledFunction(),
2686 {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2687 New->copyMetadata(*II);
2688 return New;
2689 }
2690 }
2691
2692 return nullptr;
2693}
2694
2696 Value *Op = I.getOperand(0);
2697
2698 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2699 getSimplifyQuery().getWithInstruction(&I)))
2700 return replaceInstUsesWith(I, V);
2701
2703 return X;
2704
2705 Value *X, *Y;
2706
2707 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2708 if (I.hasNoSignedZeros() &&
2711
2712 Value *OneUse;
2713 if (!match(Op, m_OneUse(m_Value(OneUse))))
2714 return nullptr;
2715
2716 if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2717 return replaceInstUsesWith(I, R);
2718
2719 // Try to eliminate fneg if at least 1 arm of the select is negated.
2720 Value *Cond;
2721 if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2722 // Unlike most transforms, this one is not safe to propagate nsz unless
2723 // it is present on the original select. We union the flags from the select
2724 // and fneg and then remove nsz if needed.
2725 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2726 S->copyFastMathFlags(&I);
2727 if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2728 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2729 S->setFastMathFlags(FMF);
2730 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2731 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2732 S->setHasNoSignedZeros(false);
2733 }
2734 };
2735 // -(Cond ? -P : Y) --> Cond ? P : -Y
2736 Value *P;
2737 if (match(X, m_FNeg(m_Value(P)))) {
2738 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2739 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2740 propagateSelectFMF(NewSel, P == Y);
2741 return NewSel;
2742 }
2743 // -(Cond ? X : -P) --> Cond ? -X : P
2744 if (match(Y, m_FNeg(m_Value(P)))) {
2745 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2746 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2747 propagateSelectFMF(NewSel, P == X);
2748 return NewSel;
2749 }
2750 }
2751
2752 // fneg (copysign x, y) -> copysign x, (fneg y)
2753 if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2754 // The source copysign has an additional value input, so we can't propagate
2755 // flags the copysign doesn't also have.
2756 FastMathFlags FMF = I.getFastMathFlags();
2757 FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2758
2761
2762 Value *NegY = Builder.CreateFNeg(Y);
2763 Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2764 return replaceInstUsesWith(I, NewCopySign);
2765 }
2766
2767 return nullptr;
2768}
2769
2771 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2772 I.getFastMathFlags(),
2773 getSimplifyQuery().getWithInstruction(&I)))
2774 return replaceInstUsesWith(I, V);
2775
2777 return X;
2778
2780 return Phi;
2781
2782 // Subtraction from -0.0 is the canonical form of fneg.
2783 // fsub -0.0, X ==> fneg X
2784 // fsub nsz 0.0, X ==> fneg nsz X
2785 //
2786 // FIXME This matcher does not respect FTZ or DAZ yet:
2787 // fsub -0.0, Denorm ==> +-0
2788 // fneg Denorm ==> -Denorm
2789 Value *Op;
2790 if (match(&I, m_FNeg(m_Value(Op))))
2792
2794 return X;
2795
2796 if (Instruction *R = foldFBinOpOfIntCasts(I))
2797 return R;
2798
2799 Value *X, *Y;
2800 Constant *C;
2801
2802 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2803 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2804 // Canonicalize to fadd to make analysis easier.
2805 // This can also help codegen because fadd is commutative.
2806 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2807 // killed later. We still limit that particular transform with 'hasOneUse'
2808 // because an fneg is assumed better/cheaper than a generic fsub.
2809 if (I.hasNoSignedZeros() ||
2810 cannotBeNegativeZero(Op0, 0, getSimplifyQuery().getWithInstruction(&I))) {
2811 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2812 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2813 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2814 }
2815 }
2816
2817 // (-X) - Op1 --> -(X + Op1)
2818 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2819 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2820 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2822 }
2823
2824 if (isa<Constant>(Op0))
2825 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2826 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2827 return NV;
2828
2829 // X - C --> X + (-C)
2830 // But don't transform constant expressions because there's an inverse fold
2831 // for X + (-Y) --> X - Y.
2832 if (match(Op1, m_ImmConstant(C)))
2833 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2834 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2835
2836 // X - (-Y) --> X + Y
2837 if (match(Op1, m_FNeg(m_Value(Y))))
2838 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2839
2840 // Similar to above, but look through a cast of the negated value:
2841 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2842 Type *Ty = I.getType();
2843 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2845
2846 // X - (fpext(-Y)) --> X + fpext(Y)
2847 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2849
2850 // Similar to above, but look through fmul/fdiv of the negated value:
2851 // Op0 - (-X * Y) --> Op0 + (X * Y)
2852 // Op0 - (Y * -X) --> Op0 + (X * Y)
2853 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2855 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2856 }
2857 // Op0 - (-X / Y) --> Op0 + (X / Y)
2858 // Op0 - (X / -Y) --> Op0 + (X / Y)
2859 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2860 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2861 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2862 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2863 }
2864
2865 // Handle special cases for FSub with selects feeding the operation
2866 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2867 return replaceInstUsesWith(I, V);
2868
2869 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2870 // (Y - X) - Y --> -X
2871 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2873
2874 // Y - (X + Y) --> -X
2875 // Y - (Y + X) --> -X
2876 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2878
2879 // (X * C) - X --> X * (C - 1.0)
2880 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2882 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2883 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2884 }
2885 // X - (X * C) --> X * (1.0 - C)
2886 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2888 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2889 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2890 }
2891
2892 // Reassociate fsub/fadd sequences to create more fadd instructions and
2893 // reduce dependency chains:
2894 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2895 Value *Z;
2897 m_Value(Z))))) {
2898 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2899 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2900 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2901 }
2902
2903 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2904 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2905 m_Value(Vec)));
2906 };
2907 Value *A0, *A1, *V0, *V1;
2908 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2909 V0->getType() == V1->getType()) {
2910 // Difference of sums is sum of differences:
2911 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2912 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2913 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2914 {Sub->getType()}, {A0, Sub}, &I);
2915 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2916 }
2917
2919 return F;
2920
2921 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2922 // functionality has been subsumed by simple pattern matching here and in
2923 // InstSimplify. We should let a dedicated reassociation pass handle more
2924 // complex pattern matching and remove this from InstCombine.
2925 if (Value *V = FAddCombine(Builder).simplify(&I))
2926 return replaceInstUsesWith(I, V);
2927
2928 // (X - Y) - Op1 --> X - (Y + Op1)
2929 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2930 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2932 }
2933 }
2934
2935 return nullptr;
2936}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static bool isConstant(const MachineInstr &MI)
amdgpu AMDGPU Register Bank Select
This file declares a class to represent arbitrary precision floating point values and provide a varie...
This file implements a class to represent arbitrary precision integral constant values and operations...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
hexagon bit simplify
static Instruction * factorizeFAddFSub(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Factor a common operand out of fadd/fsub of fmul/fdiv.
static Instruction * foldAddToAshr(BinaryOperator &Add)
Try to reduce signed division by power-of-2 to an arithmetic shift right.
static bool MatchMul(Value *E, Value *&Op, APInt &C)
static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned)
static Instruction * foldFNegIntoConstant(Instruction &I, const DataLayout &DL)
This eliminates floating-point negation in either 'fneg(X)' or 'fsub(-0.0, X)' form by combining into...
static Instruction * combineAddSubWithShlAddSub(InstCombiner::BuilderTy &Builder, const BinaryOperator &I)
static Instruction * factorizeLerp(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Eliminate an op from a linear interpolation (lerp) pattern.
static Instruction * foldSubOfMinMax(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static Instruction * foldBoxMultiply(BinaryOperator &I)
Reduce a sequence of masked half-width multiplies to a single multiply.
static Value * checkForNegativeOperand(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned)
static Instruction * foldNoWrapAdd(BinaryOperator &Add, InstCombiner::BuilderTy &Builder)
Wrapping flags may allow combining constants separated by an extend.
static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A, Value *&B)
static Instruction * factorizeMathWithShlOps(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
This is a specialization of a more general transform from foldUsingDistributiveLaws.
static Instruction * canonicalizeLowbitMask(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Fold (1 << NBits) - 1 Into: ~(-(1 << NBits)) Because a 'not' is better for bit-tracking analysis and ...
static Instruction * foldToUnsignedSaturatedAdd(BinaryOperator &I)
static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned)
This file provides internal interfaces used to implement the InstCombine.
This file provides the interface for the instcombine pass implementation.
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:531
#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:1303
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1060
Class for arbitrary precision integers.
Definition: APInt.h:76
APInt umul_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1977
bool isNegatedPowerOf2() const
Check if this APInt's negated value is a power of two greater than zero.
Definition: APInt.h:427
bool isMinSignedValue() const
Determine if this is the smallest signed value.
Definition: APInt.h:401
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:906
static APInt getMaxValue(unsigned numBits)
Gets maximum unsigned value of APInt for specific bit width.
Definition: APInt.h:184
bool isSignMask() const
Check if the APInt's value is returned by getSignMask.
Definition: APInt.h:444
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1439
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:307
int32_t exactLogBase2() const
Definition: APInt.h:1725
unsigned countr_zero() const
Count the number of trailing zero bits.
Definition: APInt.h:1589
unsigned countl_zero() const
The APInt version of std::countl_zero.
Definition: APInt.h:1548
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:1703
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:1235
bool isPowerOf2() const
Check if this APInt's value is a power of two greater than zero.
Definition: APInt.h:418
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Constructs an APInt value that has the top hiBitsSet bits set.
Definition: APInt.h:274
bool sge(const APInt &RHS) const
Signed greater or equal comparison.
Definition: APInt.h:1215
static BinaryOperator * CreateFDivFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:328
static BinaryOperator * CreateNeg(Value *Op, const Twine &Name, BasicBlock::iterator InsertBefore)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static BinaryOperator * CreateFMulFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:323
static BinaryOperator * CreateFSubFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:318
static BinaryOperator * CreateFAddFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:313
static BinaryOperator * CreateNot(Value *Op, const Twine &Name, BasicBlock::iterator InsertBefore)
This class represents a function call, abstracting a target machine's calling convention.
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr, BasicBlock::iterator InsertBefore)
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name, BasicBlock::iterator InsertBefore)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass's ...
static CastInst * CreateTruncOrBitCast(Value *S, Type *Ty, const Twine &Name, BasicBlock::iterator InsertBefore)
Create a Trunc or BitCast cast instruction.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:965
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:988
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:992
@ ICMP_EQ
equal
Definition: InstrTypes.h:986
@ ICMP_NE
not equal
Definition: InstrTypes.h:987
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2544
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2537
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:267
const APFloat & getValueAPF() const
Definition: Constants.h:310
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:314
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
Definition: Constants.h:122
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:417
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:370
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
Convenience struct for specifying and reasoning about fast-math flags.
Definition: FMF.h:20
bool noSignedZeros() const
Definition: FMF.h:68
bool noInfs() const
Definition: FMF.h:67
bool isInBounds() const
Test whether this is an inbounds GEP, as defined by LangRef.html.
Definition: Operator.h:420
unsigned countNonConstantIndices() const
Definition: Operator.h:500
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:1541
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1715
Value * CreateSRem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1404
Value * CreateBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 2 operands which is mangled on the first type.
Definition: IRBuilder.cpp:921
Value * CreateFMulFMF(Value *L, Value *R, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1595
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:1568
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:1622
Value * CreateFPTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2084
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:2028
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:460
CallInst * CreateIntrinsic(Intrinsic::ID ID, ArrayRef< Type * > Types, ArrayRef< Value * > Args, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with Args, mangled using Types.
Definition: IRBuilder.cpp:932
Value * CreateFNegFMF(Value *V, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1739
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2022
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2542
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:305
Value * CreateNUWAdd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1334
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1748
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:2537
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1338
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1410
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="", bool IsNonNeg=false)
Definition: IRBuilder.h:2010
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1469
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1321
ConstantInt * getFalse()
Get the constant value for i1 false.
Definition: IRBuilder.h:465
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2532
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1491
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1660
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:2179
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2395
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2093
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1513
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1729
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1398
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1355
Value * CreateCopySign(Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create call to the copysign intrinsic.
Definition: IRBuilder.h:1016
Instruction * FoldOpIntoSelect(Instruction &Op, SelectInst *SI, bool FoldWithMultiUse=false)
Given an instruction with a select as one operand and a constant as the other operand,...
Instruction * foldBinOpOfSelectAndCastOfSelectCondition(BinaryOperator &I)
Tries to simplify binops of select and cast of the select condition.
Instruction * visitAdd(BinaryOperator &I)
Instruction * canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(BinaryOperator &I)
Instruction * foldBinOpIntoSelectOrPhi(BinaryOperator &I)
This is a convenience wrapper function for the above two functions.
bool SimplifyAssociativeOrCommutative(BinaryOperator &I)
Performs a few simplifications for operators which are associative or commutative.
Value * foldUsingDistributiveLaws(BinaryOperator &I)
Tries to simplify binary operations which some other binary operation distributes over.
Instruction * foldBinOpShiftWithShift(BinaryOperator &I)
Instruction * foldSquareSumInt(BinaryOperator &I)
Instruction * foldOpIntoPhi(Instruction &I, PHINode *PN)
Given a binary operator, cast instruction, or select which has a PHI node as operand #0,...
Instruction * foldSquareSumFP(BinaryOperator &I)
Instruction * visitSub(BinaryOperator &I)
Value * OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty, bool isNUW)
Optimize pointer differences into the same array into a size.
Instruction * visitFAdd(BinaryOperator &I)
Instruction * foldBinopWithPhiOperands(BinaryOperator &BO)
For a binary operator with 2 phi operands, try to hoist the binary operation before the phi.
Instruction * tryFoldInstWithCtpopWithNot(Instruction *I)
Value * SimplifyAddWithRemainder(BinaryOperator &I)
Tries to simplify add operations using the definition of remainder.
Instruction * foldAddWithConstant(BinaryOperator &Add)
Instruction * foldVectorBinop(BinaryOperator &Inst)
Canonicalize the position of binops relative to shufflevector.
Value * SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS, Value *RHS)
Instruction * visitFNeg(UnaryOperator &I)
Instruction * visitFSub(BinaryOperator &I)
SimplifyQuery SQ
Definition: InstCombiner.h:76
bool isFreeToInvert(Value *V, bool WillInvertAllUses, bool &DoesConsume)
Return true if the specified value is free to invert (apply ~ to).
Definition: InstCombiner.h:231
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:385
static Constant * SubOne(Constant *C)
Subtract one from a Constant.
Definition: InstCombiner.h:179
const DataLayout & DL
Definition: InstCombiner.h:75
unsigned ComputeNumSignBits(const Value *Op, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:451
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:409
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
Definition: InstCombiner.h:430
BuilderTy & Builder
Definition: InstCombiner.h:60
bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:446
Value * getFreelyInverted(Value *V, bool WillInvertAllUses, BuilderTy *Builder, bool &DoesConsume)
Definition: InstCombiner.h:212
const SimplifyQuery & getSimplifyQuery() const
Definition: InstCombiner.h:341
static Constant * AddOne(Constant *C)
Add one to a Constant.
Definition: InstCombiner.h:174
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nuw flag on this instruction, which must be an operator which supports this flag.
bool hasNoUnsignedWrap() const LLVM_READONLY
Determine whether the no unsigned wrap flag is set.
void copyFastMathFlags(FastMathFlags FMF)
Convenience function for transferring all fast-math flag values to this instruction,...
void setHasNoSignedZeros(bool B)
Set or clear the no-signed-zeros flag on this instruction, which must be an operator which supports t...
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag.
void setFastMathFlags(FastMathFlags FMF)
Convenience function for setting multiple fast-math flags on this instruction, which must be an opera...
void setHasNoInfs(bool B)
Set or clear the no-infs flag on this instruction, which must be an operator which supports this flag...
FastMathFlags getFastMathFlags() const LLVM_READONLY
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:450
void copyMetadata(const Instruction &SrcInst, ArrayRef< unsigned > WL=ArrayRef< unsigned >())
Copy metadata from SrcInst to this instruction.
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:47
static Value * Negate(bool LHSIsZero, bool IsNSW, Value *Root, InstCombinerImpl &IC)
Attempt to negate Root.
Utility class for integer operators which may exhibit overflow - Add, Sub, Mul, and Shl.
Definition: Operator.h:75
bool hasNoSignedWrap() const
Test whether this operation is known to never undergo signed overflow, aka the nsw property.
Definition: Operator.h:108
bool hasNoUnsignedWrap() const
Test whether this operation is known to never undergo unsigned overflow, aka the nuw property.
Definition: Operator.h:102
This class represents the LLVM 'select' instruction.
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr, BasicBlock::iterator InsertBefore, Instruction *MDFrom=nullptr)
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1209
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:234
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
static UnaryOperator * CreateFNegFMF(Value *Op, Instruction *FMFSource, const Twine &Name, BasicBlock::iterator InsertBefore)
Definition: InstrTypes.h:189
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
bool hasOneUse() const
Return true if there is exactly one use of this value.
Definition: Value.h:434
bool hasNUsesOrMore(unsigned N) const
Return true if this value has N uses or more.
Definition: Value.cpp:153
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs and address space casts.
Definition: Value.cpp:693
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
This class represents zero extension of integer types.
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
Function * getDeclaration(Module *M, ID id, ArrayRef< Type * > Tys=std::nullopt)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1451
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:477
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:903
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:100
BinaryOp_match< LHS, RHS, Instruction::FMul, true > m_c_FMul(const LHS &L, const RHS &R)
Matches FMul with LHS and RHS in either order.
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:568
BinaryOp_match< LHS, RHS, Instruction::URem > m_URem(const LHS &L, const RHS &R)
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:160
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)
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)
match_combine_or< CastInst_match< OpTy, ZExtInst >, OpTy > m_ZExtOrSelf(const OpTy &Op)
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
bind_ty< Instruction > m_Instruction(Instruction *&I)
Match an instruction, capturing it if we match.
Definition: PatternMatch.h:765
cstfp_pred_ty< is_any_zero_fp > m_AnyZeroFP()
Match a floating-point negative zero or positive zero.
Definition: PatternMatch.h:713
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:821
DisjointOr_match< LHS, RHS > m_DisjointOr(const LHS &L, const RHS &R)
cst_pred_ty< is_one > m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:541
ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)
Matches SelectInst.
match_combine_or< CastInst_match< OpTy, SExtInst >, OpTy > m_SExtOrSelf(const OpTy &Op)
specific_fpval m_SpecificFP(double V)
Match a specific floating point value or vector with all elements equal to the value.
Definition: PatternMatch.h:864
match_combine_and< LTy, RTy > m_CombineAnd(const LTy &L, const RTy &R)
Combine two pattern matchers matching L && R.
Definition: PatternMatch.h:240
CastOperator_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
BinaryOp_match< LHS, RHS, Instruction::Xor, true > m_c_Xor(const LHS &L, const RHS &R)
Matches an Xor with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::FAdd > m_FAdd(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::Mul > m_Mul(const LHS &L, const RHS &R)
deferredval_ty< Value > m_Deferred(Value *const &V)
Like m_Specific(), but works if the specific value to match is determined as part of the same match()...
Definition: PatternMatch.h:839
cst_pred_ty< is_zero_int > m_ZeroInt()
Match an integer 0 or a vector with all elements equal to 0.
Definition: PatternMatch.h:548
CmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R)
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:67
MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty, true > m_c_SMin(const LHS &L, const RHS &R)
Matches an SMin with LHS and RHS in either order.
BinaryOp_match< cst_pred_ty< is_zero_int >, ValTy, Instruction::Sub > m_Neg(const ValTy &V)
Matches a 'Neg' as 'sub 0, V'.
match_combine_and< class_match< Constant >, match_unless< constantexpr_match > > m_ImmConstant()
Match an arbitrary immediate Constant and ignore it.
Definition: PatternMatch.h:800
MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty, true > m_c_UMax(const LHS &L, const RHS &R)
Matches a UMax with LHS and RHS in either order.
CastInst_match< OpTy, FPExtInst > m_FPExt(const OpTy &Op)
CastInst_match< OpTy, ZExtInst > m_ZExt(const OpTy &Op)
Matches ZExt.
BinaryOp_match< LHS, RHS, Instruction::UDiv > m_UDiv(const LHS &L, const RHS &R)
cst_pred_ty< is_negated_power2 > m_NegatedPower2()
Match a integer or vector negated power-of-2.
Definition: PatternMatch.h:576
specific_fpval m_FPOne()
Match a float 1.0 or vector with all elements equal to 1.0.
Definition: PatternMatch.h:867
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty, true > m_c_UMin(const LHS &L, const RHS &R)
Matches a UMin with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)
Matches a Add with LHS and RHS in either order.
MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty, true > m_c_SMax(const LHS &L, const RHS &R)
Matches an SMax with LHS and RHS in either order.
match_combine_or< match_combine_or< MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty, true >, MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty, true > >, match_combine_or< MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty, true >, MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty, true > > > m_c_MaxOrMin(const LHS &L, const RHS &R)
match_combine_or< CastOperator_match< OpTy, Instruction::Trunc >, OpTy > m_TruncOrSelf(const OpTy &Op)
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
specific_intval< true > m_SpecificIntAllowUndef(APInt V)
Definition: PatternMatch.h:911
OverflowingBinaryOp_match< LHS, RHS, Instruction::Sub, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWSub(const LHS &L, const RHS &R)
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt.
Definition: PatternMatch.h:294
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:92
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
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:311
CastInst_match< OpTy, FPTruncInst > m_FPTrunc(const OpTy &Op)
BinaryOp_match< LHS, RHS, Instruction::SRem > m_SRem(const LHS &L, const RHS &R)
BinaryOp_match< cst_pred_ty< is_all_ones >, ValTy, Instruction::Xor, true > m_Not(const ValTy &V)
Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
CastInst_match< OpTy, SExtInst > m_SExt(const OpTy &Op)
Matches SExt.
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:561
BinaryOp_match< LHS, RHS, Instruction::Or, true > m_c_Or(const LHS &L, const RHS &R)
Matches an Or with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::Mul, true > m_c_Mul(const LHS &L, const RHS &R)
Matches a Mul with LHS and RHS in either order.
m_Intrinsic_Ty< Opnd0, Opnd1 >::Ty m_CopySign(const Opnd0 &Op0, const Opnd1 &Op1)
CastOperator_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
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:234
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:647
@ CE
Windows NT (Windows on ARM)
NodeAddr< InstrNode * > Instr
Definition: RDFGraph.h:389
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
bool haveNoCommonBitsSet(const WithCache< const Value * > &LHSCache, const WithCache< const Value * > &RHSCache, const SimplifyQuery &SQ)
Return true if LHS and RHS have no common bits set.
Intrinsic::ID getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID)
@ Offset
Definition: DWP.cpp:456
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
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.
std::string & operator+=(std::string &buffer, StringRef string)
Definition: StringRef.h:899
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.
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:1745
Constant * ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, Constant *RHS, const DataLayout &DL)
Attempt to constant fold a binary operation with the specified operands.
@ Mul
Product of integers.
@ FMul
Product of floats.
@ And
Bitwise or logical AND of integers.
@ SMin
Signed integer min implemented in terms of select(cmp()).
@ Add
Sum of integers.
@ FAdd
Sum of floats.
DWARFExpression::Operation Op
RoundingMode
Rounding mode.
bool isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Return true if this function can prove that V does not have undef bits and is never poison.
constexpr unsigned BitWidth
Definition: BitmaskEnum.h:191
bool cannotBeNegativeZero(const Value *V, unsigned Depth, const SimplifyQuery &SQ)
Return true if we can prove that the specified FP value is never equal to -0.0.
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:860
A suitably aligned and sized character array member which can hold elements of any type.
Definition: AlignOf.h:27
const Instruction * CxtI
Definition: SimplifyQuery.h:65
const DominatorTree * DT
Definition: SimplifyQuery.h:63
SimplifyQuery getWithInstruction(const Instruction *I) const
Definition: SimplifyQuery.h:96
AssumptionCache * AC
Definition: SimplifyQuery.h:64