LLVM 19.0.0git
InstCombineAddSub.cpp
Go to the documentation of this file.
1//===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visit functions for add, fadd, sub, and fsub.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APFloat.h"
15#include "llvm/ADT/APInt.h"
16#include "llvm/ADT/STLExtras.h"
20#include "llvm/IR/Constant.h"
21#include "llvm/IR/Constants.h"
22#include "llvm/IR/InstrTypes.h"
23#include "llvm/IR/Instruction.h"
25#include "llvm/IR/Operator.h"
27#include "llvm/IR/Type.h"
28#include "llvm/IR/Value.h"
33#include <cassert>
34#include <utility>
35
36using namespace llvm;
37using namespace PatternMatch;
38
39#define DEBUG_TYPE "instcombine"
40
41namespace {
42
43 /// Class representing coefficient of floating-point addend.
44 /// This class needs to be highly efficient, which is especially true for
45 /// the constructor. As of I write this comment, the cost of the default
46 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
47 /// perform write-merging).
48 ///
49 class FAddendCoef {
50 public:
51 // The constructor has to initialize a APFloat, which is unnecessary for
52 // most addends which have coefficient either 1 or -1. So, the constructor
53 // is expensive. In order to avoid the cost of the constructor, we should
54 // reuse some instances whenever possible. The pre-created instances
55 // FAddCombine::Add[0-5] embodies this idea.
56 FAddendCoef() = default;
57 ~FAddendCoef();
58
59 // If possible, don't define operator+/operator- etc because these
60 // operators inevitably call FAddendCoef's constructor which is not cheap.
61 void operator=(const FAddendCoef &A);
62 void operator+=(const FAddendCoef &A);
63 void operator*=(const FAddendCoef &S);
64
65 void set(short C) {
66 assert(!insaneIntVal(C) && "Insane coefficient");
67 IsFp = false; IntVal = C;
68 }
69
70 void set(const APFloat& C);
71
72 void negate();
73
74 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75 Value *getValue(Type *) const;
76
77 bool isOne() const { return isInt() && IntVal == 1; }
78 bool isTwo() const { return isInt() && IntVal == 2; }
79 bool isMinusOne() const { return isInt() && IntVal == -1; }
80 bool isMinusTwo() const { return isInt() && IntVal == -2; }
81
82 private:
83 bool insaneIntVal(int V) { return V > 4 || V < -4; }
84
85 APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86
87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
89 }
90
91 const APFloat &getFpVal() const {
92 assert(IsFp && BufHasFpVal && "Incorret state");
93 return *getFpValPtr();
94 }
95
96 APFloat &getFpVal() {
97 assert(IsFp && BufHasFpVal && "Incorret state");
98 return *getFpValPtr();
99 }
100
101 bool isInt() const { return !IsFp; }
102
103 // If the coefficient is represented by an integer, promote it to a
104 // floating point.
105 void convertToFpType(const fltSemantics &Sem);
106
107 // Construct an APFloat from a signed integer.
108 // TODO: We should get rid of this function when APFloat can be constructed
109 // from an *SIGNED* integer.
110 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
111
112 bool IsFp = false;
113
114 // True iff FpValBuf contains an instance of APFloat.
115 bool BufHasFpVal = false;
116
117 // The integer coefficient of an individual addend is either 1 or -1,
118 // and we try to simplify at most 4 addends from neighboring at most
119 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120 // is overkill of this end.
121 short IntVal = 0;
122
124 };
125
126 /// FAddend is used to represent floating-point addend. An addend is
127 /// represented as <C, V>, where the V is a symbolic value, and C is a
128 /// constant coefficient. A constant addend is represented as <C, 0>.
129 class FAddend {
130 public:
131 FAddend() = default;
132
133 void operator+=(const FAddend &T) {
134 assert((Val == T.Val) && "Symbolic-values disagree");
135 Coeff += T.Coeff;
136 }
137
138 Value *getSymVal() const { return Val; }
139 const FAddendCoef &getCoef() const { return Coeff; }
140
141 bool isConstant() const { return Val == nullptr; }
142 bool isZero() const { return Coeff.isZero(); }
143
144 void set(short Coefficient, Value *V) {
145 Coeff.set(Coefficient);
146 Val = V;
147 }
148 void set(const APFloat &Coefficient, Value *V) {
149 Coeff.set(Coefficient);
150 Val = V;
151 }
152 void set(const ConstantFP *Coefficient, Value *V) {
153 Coeff.set(Coefficient->getValueAPF());
154 Val = V;
155 }
156
157 void negate() { Coeff.negate(); }
158
159 /// Drill down the U-D chain one step to find the definition of V, and
160 /// try to break the definition into one or two addends.
161 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
162
163 /// Similar to FAddend::drillDownOneStep() except that the value being
164 /// splitted is the addend itself.
165 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
166
167 private:
168 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
169
170 // This addend has the value of "Coeff * Val".
171 Value *Val = nullptr;
172 FAddendCoef Coeff;
173 };
174
175 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176 /// with its neighboring at most two instructions.
177 ///
178 class FAddCombine {
179 public:
180 FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
181
183
184 private:
185 using AddendVect = SmallVector<const FAddend *, 4>;
186
187 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
188
189 /// Convert given addend to a Value
190 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
191
192 /// Return the number of instructions needed to emit the N-ary addition.
193 unsigned calcInstrNumber(const AddendVect& Vect);
194
195 Value *createFSub(Value *Opnd0, Value *Opnd1);
196 Value *createFAdd(Value *Opnd0, Value *Opnd1);
197 Value *createFMul(Value *Opnd0, Value *Opnd1);
198 Value *createFNeg(Value *V);
199 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
201
202 // Debugging stuff are clustered here.
203 #ifndef NDEBUG
204 unsigned CreateInstrNum;
205 void initCreateInstNum() { CreateInstrNum = 0; }
206 void incCreateInstNum() { CreateInstrNum++; }
207 #else
208 void initCreateInstNum() {}
209 void incCreateInstNum() {}
210 #endif
211
213 Instruction *Instr = nullptr;
214 };
215
216} // end anonymous namespace
217
218//===----------------------------------------------------------------------===//
219//
220// Implementation of
221// {FAddendCoef, FAddend, FAddition, FAddCombine}.
222//
223//===----------------------------------------------------------------------===//
224FAddendCoef::~FAddendCoef() {
225 if (BufHasFpVal)
226 getFpValPtr()->~APFloat();
227}
228
229void FAddendCoef::set(const APFloat& C) {
230 APFloat *P = getFpValPtr();
231
232 if (isInt()) {
233 // As the buffer is meanless byte stream, we cannot call
234 // APFloat::operator=().
235 new(P) APFloat(C);
236 } else
237 *P = C;
238
239 IsFp = BufHasFpVal = true;
240}
241
242void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
243 if (!isInt())
244 return;
245
246 APFloat *P = getFpValPtr();
247 if (IntVal > 0)
248 new(P) APFloat(Sem, IntVal);
249 else {
250 new(P) APFloat(Sem, 0 - IntVal);
251 P->changeSign();
252 }
253 IsFp = BufHasFpVal = true;
254}
255
256APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
257 if (Val >= 0)
258 return APFloat(Sem, Val);
259
260 APFloat T(Sem, 0 - Val);
261 T.changeSign();
262
263 return T;
264}
265
266void FAddendCoef::operator=(const FAddendCoef &That) {
267 if (That.isInt())
268 set(That.IntVal);
269 else
270 set(That.getFpVal());
271}
272
273void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
276 if (isInt())
277 IntVal += That.IntVal;
278 else
279 getFpVal().add(That.getFpVal(), RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292}
293
294void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value");
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322}
323
324void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329}
330
331Value *FAddendCoef::getValue(Type *Ty) const {
332 return isInt() ?
333 ConstantFP::get(Ty, float(IntVal)) :
334 ConstantFP::get(Ty->getContext(), getFpVal());
335}
336
337// The definition of <Val> Addends
338// =========================================
339// A + B <1, A>, <1,B>
340// A - B <1, A>, <1,B>
341// 0 - B <-1, B>
342// C * A, <C, A>
343// A + C <1, A> <C, NULL>
344// 0 +/- 0 <0, NULL> (corner case)
345//
346// Legend: A and B are not constant, C is constant
347unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
351 return 0;
352
353 unsigned Opcode = I->getOpcode();
354
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356 ConstantFP *C0, *C1;
357 Value *Opnd0 = I->getOperand(0);
358 Value *Opnd1 = I->getOperand(1);
359 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(1, Opnd0);
368 else
369 Addend0.set(C0, nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(1, Opnd1);
376 else
377 Addend.set(C1, nullptr);
378 if (Opcode == Instruction::FSub)
379 Addend.negate();
380 }
381
382 if (Opnd0 || Opnd1)
383 return Opnd0 && Opnd1 ? 2 : 1;
384
385 // Both operands are zero. Weird!
386 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(0);
392 Value *V1 = I->getOperand(1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394 Addend0.set(C, V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399 Addend0.set(C, V0);
400 return 1;
401 }
402 }
403
404 return 0;
405}
406
407// Try to break *this* addend into two addends. e.g. Suppose this addend is
408// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409// i.e. <2.3, X> and <2.3, Y>.
410unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
412 if (isConstant())
413 return 0;
414
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
417 return BreakNum;
418
419 Addend0.Scale(Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(Coeff);
423
424 return BreakNum;
425}
426
427Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429 "Expected 'reassoc'+'nsz' instruction");
430
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
433 return nullptr;
434
435 assert((I->getOpcode() == Instruction::FAdd ||
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
494
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
506
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
508 return R;
509 }
510
511 return nullptr;
512}
513
514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends");
517
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
521
522 // Simplified addends are placed <SimpVect>.
523 AddendVect SimpVect;
524
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529
530 const FAddend *ThisAddend = Addends[SymIdx];
531 if (!ThisAddend) {
532 // This addend was processed before.
533 continue;
534 }
535
536 Value *Val = ThisAddend->getSymVal();
537
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543 // call later.
544
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(ThisAddend);
547
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(T);
561 }
562 }
563
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569 R += *SimpVect[Idx];
570
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(&R);
575 }
576 }
577 }
578
579 assert((NextTmpIdx <= std::size(TmpResult) + 1) && "out-of-bound access");
580
581 Value *Result;
582 if (!SimpVect.empty())
583 Result = createNaryFAdd(SimpVect, InstrQuota);
584 else {
585 // The addition is folded to 0.0.
586 Result = ConstantFP::get(Instr->getType(), 0.0);
587 }
588
589 return Result;
590}
591
592Value *FAddCombine::createNaryFAdd
593 (const AddendVect &Opnds, unsigned InstrQuota) {
594 assert(!Opnds.empty() && "Expect at least one addend");
595
596 // Step 1: Check if the # of instructions needed exceeds the quota.
597
598 unsigned InstrNeeded = calcInstrNumber(Opnds);
599 if (InstrNeeded > InstrQuota)
600 return nullptr;
601
602 initCreateInstNum();
603
604 // step 2: Emit the N-ary addition.
605 // Note that at most three instructions are involved in Fadd-InstCombine: the
606 // addition in question, and at most two neighboring instructions.
607 // The resulting optimized addition should have at least one less instruction
608 // than the original addition expression tree. This implies that the resulting
609 // N-ary addition has at most two instructions, and we don't need to worry
610 // about tree-height when constructing the N-ary addition.
611
612 Value *LastVal = nullptr;
613 bool LastValNeedNeg = false;
614
615 // Iterate the addends, creating fadd/fsub using adjacent two addends.
616 for (const FAddend *Opnd : Opnds) {
617 bool NeedNeg;
618 Value *V = createAddendVal(*Opnd, NeedNeg);
619 if (!LastVal) {
620 LastVal = V;
621 LastValNeedNeg = NeedNeg;
622 continue;
623 }
624
625 if (LastValNeedNeg == NeedNeg) {
626 LastVal = createFAdd(LastVal, V);
627 continue;
628 }
629
630 if (LastValNeedNeg)
631 LastVal = createFSub(V, LastVal);
632 else
633 LastVal = createFSub(LastVal, V);
634
635 LastValNeedNeg = false;
636 }
637
638 if (LastValNeedNeg) {
639 LastVal = createFNeg(LastVal);
640 }
641
642#ifndef NDEBUG
643 assert(CreateInstrNum == InstrNeeded &&
644 "Inconsistent in instruction numbers");
645#endif
646
647 return LastVal;
648}
649
650Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
651 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
652 if (Instruction *I = dyn_cast<Instruction>(V))
653 createInstPostProc(I);
654 return V;
655}
656
657Value *FAddCombine::createFNeg(Value *V) {
658 Value *NewV = Builder.CreateFNeg(V);
659 if (Instruction *I = dyn_cast<Instruction>(NewV))
660 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
661 return NewV;
662}
663
664Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
665 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
666 if (Instruction *I = dyn_cast<Instruction>(V))
667 createInstPostProc(I);
668 return V;
669}
670
671Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
673 if (Instruction *I = dyn_cast<Instruction>(V))
674 createInstPostProc(I);
675 return V;
676}
677
678void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
679 NewInstr->setDebugLoc(Instr->getDebugLoc());
680
681 // Keep track of the number of instruction created.
682 if (!NoNumber)
683 incCreateInstNum();
684
685 // Propagate fast-math flags
686 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
687}
688
689// Return the number of instruction needed to emit the N-ary addition.
690// NOTE: Keep this function in sync with createAddendVal().
691unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
692 unsigned OpndNum = Opnds.size();
693 unsigned InstrNeeded = OpndNum - 1;
694
695 // Adjust the number of instructions needed to emit the N-ary add.
696 for (const FAddend *Opnd : Opnds) {
697 if (Opnd->isConstant())
698 continue;
699
700 // The constant check above is really for a few special constant
701 // coefficients.
702 if (isa<UndefValue>(Opnd->getSymVal()))
703 continue;
704
705 const FAddendCoef &CE = Opnd->getCoef();
706 // Let the addend be "c * x". If "c == +/-1", the value of the addend
707 // is immediately available; otherwise, it needs exactly one instruction
708 // to evaluate the value.
709 if (!CE.isMinusOne() && !CE.isOne())
710 InstrNeeded++;
711 }
712 return InstrNeeded;
713}
714
715// Input Addend Value NeedNeg(output)
716// ================================================================
717// Constant C C false
718// <+/-1, V> V coefficient is -1
719// <2/-2, V> "fadd V, V" coefficient is -2
720// <C, V> "fmul V, C" false
721//
722// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
723Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
724 const FAddendCoef &Coeff = Opnd.getCoef();
725
726 if (Opnd.isConstant()) {
727 NeedNeg = false;
728 return Coeff.getValue(Instr->getType());
729 }
730
731 Value *OpndVal = Opnd.getSymVal();
732
733 if (Coeff.isMinusOne() || Coeff.isOne()) {
734 NeedNeg = Coeff.isMinusOne();
735 return OpndVal;
736 }
737
738 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
739 NeedNeg = Coeff.isMinusTwo();
740 return createFAdd(OpndVal, OpndVal);
741 }
742
743 NeedNeg = false;
744 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
745}
746
747// Checks if any operand is negative and we can convert add to sub.
748// This function checks for following negative patterns
749// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
750// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
751// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
753 InstCombiner::BuilderTy &Builder) {
754 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
755
756 // This function creates 2 instructions to replace ADD, we need at least one
757 // of LHS or RHS to have one use to ensure benefit in transform.
758 if (!LHS->hasOneUse() && !RHS->hasOneUse())
759 return nullptr;
760
761 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
762 const APInt *C1 = nullptr, *C2 = nullptr;
763
764 // if ONE is on other side, swap
765 if (match(RHS, m_Add(m_Value(X), m_One())))
766 std::swap(LHS, RHS);
767
768 if (match(LHS, m_Add(m_Value(X), m_One()))) {
769 // if XOR on other side, swap
770 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
771 std::swap(X, RHS);
772
773 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
774 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
775 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
776 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
777 Value *NewAnd = Builder.CreateAnd(Z, *C1);
778 return Builder.CreateSub(RHS, NewAnd, "sub");
779 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
780 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
781 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
782 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
783 return Builder.CreateSub(RHS, NewOr, "sub");
784 }
785 }
786 }
787
788 // Restore LHS and RHS
789 LHS = I.getOperand(0);
790 RHS = I.getOperand(1);
791
792 // if XOR is on other side, swap
793 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
794 std::swap(LHS, RHS);
795
796 // C2 is ODD
797 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
798 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
799 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
800 if (C1->countr_zero() == 0)
801 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
802 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
803 return Builder.CreateSub(RHS, NewOr, "sub");
804 }
805 return nullptr;
806}
807
808/// Wrapping flags may allow combining constants separated by an extend.
810 InstCombiner::BuilderTy &Builder) {
811 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
812 Type *Ty = Add.getType();
813 Constant *Op1C;
814 if (!match(Op1, m_Constant(Op1C)))
815 return nullptr;
816
817 // Try this match first because it results in an add in the narrow type.
818 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
819 Value *X;
820 const APInt *C1, *C2;
821 if (match(Op1, m_APInt(C1)) &&
823 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824 Constant *NewC =
825 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
826 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
827 }
828
829 // More general combining of constants in the wide type.
830 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
831 Constant *NarrowC;
832 if (match(Op0,
834 Value *WideC = Builder.CreateSExt(NarrowC, Ty);
835 Value *NewC = Builder.CreateAdd(WideC, Op1C);
836 Value *WideX = Builder.CreateSExt(X, Ty);
837 return BinaryOperator::CreateAdd(WideX, NewC);
838 }
839 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
840 if (match(Op0,
842 Value *WideC = Builder.CreateZExt(NarrowC, Ty);
843 Value *NewC = Builder.CreateAdd(WideC, Op1C);
844 Value *WideX = Builder.CreateZExt(X, Ty);
845 return BinaryOperator::CreateAdd(WideX, NewC);
846 }
847
848 return nullptr;
849}
850
852 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
853 Type *Ty = Add.getType();
854 Constant *Op1C;
855 if (!match(Op1, m_ImmConstant(Op1C)))
856 return nullptr;
857
859 return NV;
860
861 Value *X;
862 Constant *Op00C;
863
864 // add (sub C1, X), C2 --> sub (add C1, C2), X
865 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
866 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
867
868 Value *Y;
869
870 // add (sub X, Y), -1 --> add (not Y), X
871 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
872 match(Op1, m_AllOnes()))
873 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
874
875 // zext(bool) + C -> bool ? C + 1 : C
876 if (match(Op0, m_ZExt(m_Value(X))) &&
877 X->getType()->getScalarSizeInBits() == 1)
878 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
879 // sext(bool) + C -> bool ? C - 1 : C
880 if (match(Op0, m_SExt(m_Value(X))) &&
881 X->getType()->getScalarSizeInBits() == 1)
882 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
883
884 // ~X + C --> (C-1) - X
885 if (match(Op0, m_Not(m_Value(X)))) {
886 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
887 auto *COne = ConstantInt::get(Op1C->getType(), 1);
888 bool WillNotSOV = willNotOverflowSignedSub(Op1C, COne, Add);
889 BinaryOperator *Res =
890 BinaryOperator::CreateSub(ConstantExpr::getSub(Op1C, COne), X);
891 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
892 return Res;
893 }
894
895 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
896 const APInt *C;
897 unsigned BitWidth = Ty->getScalarSizeInBits();
898 if (match(Op0, m_OneUse(m_AShr(m_Value(X),
900 match(Op1, m_One()))
901 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
902
903 if (!match(Op1, m_APInt(C)))
904 return nullptr;
905
906 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
907 Constant *Op01C;
908 if (match(Op0, m_DisjointOr(m_Value(X), m_ImmConstant(Op01C))))
909 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
910
911 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
912 const APInt *C2;
913 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
914 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
915
916 if (C->isSignMask()) {
917 // If wrapping is not allowed, then the addition must set the sign bit:
918 // X + (signmask) --> X | signmask
919 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
920 return BinaryOperator::CreateOr(Op0, Op1);
921
922 // If wrapping is allowed, then the addition flips the sign bit of LHS:
923 // X + (signmask) --> X ^ signmask
924 return BinaryOperator::CreateXor(Op0, Op1);
925 }
926
927 // Is this add the last step in a convoluted sext?
928 // add(zext(xor i16 X, -32768), -32768) --> sext X
929 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
930 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
931 return CastInst::Create(Instruction::SExt, X, Ty);
932
933 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
934 // (X ^ signmask) + C --> (X + (signmask ^ C))
935 if (C2->isSignMask())
936 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
937
938 // If X has no high-bits set above an xor mask:
939 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
940 if (C2->isMask()) {
941 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
942 if ((*C2 | LHSKnown.Zero).isAllOnes())
943 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
944 }
945
946 // Look for a math+logic pattern that corresponds to sext-in-register of a
947 // value with cleared high bits. Convert that into a pair of shifts:
948 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
949 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
950 if (Op0->hasOneUse() && *C2 == -(*C)) {
951 unsigned BitWidth = Ty->getScalarSizeInBits();
952 unsigned ShAmt = 0;
953 if (C->isPowerOf2())
954 ShAmt = BitWidth - C->logBase2() - 1;
955 else if (C2->isPowerOf2())
956 ShAmt = BitWidth - C2->logBase2() - 1;
958 0, &Add)) {
959 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
960 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
961 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
962 }
963 }
964 }
965
966 if (C->isOne() && Op0->hasOneUse()) {
967 // add (sext i1 X), 1 --> zext (not X)
968 // TODO: The smallest IR representation is (select X, 0, 1), and that would
969 // not require the one-use check. But we need to remove a transform in
970 // visitSelect and make sure that IR value tracking for select is equal or
971 // better than for these ops.
972 if (match(Op0, m_SExt(m_Value(X))) &&
973 X->getType()->getScalarSizeInBits() == 1)
974 return new ZExtInst(Builder.CreateNot(X), Ty);
975
976 // Shifts and add used to flip and mask off the low bit:
977 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
978 const APInt *C3;
979 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
980 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
981 Value *NotX = Builder.CreateNot(X);
982 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
983 }
984 }
985
986 // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
987 // TODO: There's a general form for any constant on the outer add.
988 if (C->isOne()) {
989 if (match(Op0, m_ZExt(m_Add(m_Value(X), m_AllOnes())))) {
991 if (llvm::isKnownNonZero(X, DL, 0, Q.AC, Q.CxtI, Q.DT))
992 return new ZExtInst(X, Ty);
993 }
994 }
995
996 return nullptr;
997}
998
999// match variations of a^2 + 2*a*b + b^2
1000//
1001// to reuse the code between the FP and Int versions, the instruction OpCodes
1002// and constant types have been turned into template parameters.
1003//
1004// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1005// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1006// (we're matching `X<<1` instead of `X*2` for Int)
1007template <bool FP, typename Mul2Rhs>
1008static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1009 Value *&B) {
1010 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1011 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1012 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1013
1014 // (a * a) + (((a * 2) + b) * b)
1015 if (match(&I, m_c_BinOp(
1016 AddOp, m_OneUse(m_BinOp(MulOp, m_Value(A), m_Deferred(A))),
1018 MulOp,
1019 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(A), M2Rhs),
1020 m_Value(B)),
1021 m_Deferred(B))))))
1022 return true;
1023
1024 // ((a * b) * 2) or ((a * 2) * b)
1025 // +
1026 // (a * a + b * b) or (b * b + a * a)
1027 return match(
1028 &I,
1029 m_c_BinOp(AddOp,
1032 Mul2Op, m_BinOp(MulOp, m_Value(A), m_Value(B)), M2Rhs)),
1033 m_OneUse(m_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(A), M2Rhs),
1034 m_Value(B)))),
1036 AddOp, m_BinOp(MulOp, m_Deferred(A), m_Deferred(A)),
1037 m_BinOp(MulOp, m_Deferred(B), m_Deferred(B))))));
1038}
1039
1040// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1042 Value *A, *B;
1043 if (matchesSquareSum</*FP*/ false>(I, m_SpecificInt(1), A, B)) {
1044 Value *AB = Builder.CreateAdd(A, B);
1045 return BinaryOperator::CreateMul(AB, AB);
1046 }
1047 return nullptr;
1048}
1049
1050// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1051// Requires `nsz` and `reassoc`.
1053 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1054 Value *A, *B;
1055 if (matchesSquareSum</*FP*/ true>(I, m_SpecificFP(2.0), A, B)) {
1056 Value *AB = Builder.CreateFAddFMF(A, B, &I);
1057 return BinaryOperator::CreateFMulFMF(AB, AB, &I);
1058 }
1059 return nullptr;
1060}
1061
1062// Matches multiplication expression Op * C where C is a constant. Returns the
1063// constant value in C and the other operand in Op. Returns true if such a
1064// match is found.
1065static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1066 const APInt *AI;
1067 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1068 C = *AI;
1069 return true;
1070 }
1071 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1072 C = APInt(AI->getBitWidth(), 1);
1073 C <<= *AI;
1074 return true;
1075 }
1076 return false;
1077}
1078
1079// Matches remainder expression Op % C where C is a constant. Returns the
1080// constant value in C and the other operand in Op. Returns the signedness of
1081// the remainder operation in IsSigned. Returns true if such a match is
1082// found.
1083static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1084 const APInt *AI;
1085 IsSigned = false;
1086 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1087 IsSigned = true;
1088 C = *AI;
1089 return true;
1090 }
1091 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1092 C = *AI;
1093 return true;
1094 }
1095 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1096 C = *AI + 1;
1097 return true;
1098 }
1099 return false;
1100}
1101
1102// Matches division expression Op / C with the given signedness as indicated
1103// by IsSigned, where C is a constant. Returns the constant value in C and the
1104// other operand in Op. Returns true if such a match is found.
1105static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1106 const APInt *AI;
1107 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1108 C = *AI;
1109 return true;
1110 }
1111 if (!IsSigned) {
1112 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1113 C = *AI;
1114 return true;
1115 }
1116 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1117 C = APInt(AI->getBitWidth(), 1);
1118 C <<= *AI;
1119 return true;
1120 }
1121 }
1122 return false;
1123}
1124
1125// Returns whether C0 * C1 with the given signedness overflows.
1126static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1127 bool overflow;
1128 if (IsSigned)
1129 (void)C0.smul_ov(C1, overflow);
1130 else
1131 (void)C0.umul_ov(C1, overflow);
1132 return overflow;
1133}
1134
1135// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1136// does not overflow.
1138 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1139 Value *X, *MulOpV;
1140 APInt C0, MulOpC;
1141 bool IsSigned;
1142 // Match I = X % C0 + MulOpV * C0
1143 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1144 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1145 C0 == MulOpC) {
1146 Value *RemOpV;
1147 APInt C1;
1148 bool Rem2IsSigned;
1149 // Match MulOpC = RemOpV % C1
1150 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1151 IsSigned == Rem2IsSigned) {
1152 Value *DivOpV;
1153 APInt DivOpC;
1154 // Match RemOpV = X / C0
1155 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1156 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1157 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1158 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1159 : Builder.CreateURem(X, NewDivisor, "urem");
1160 }
1161 }
1162 }
1163
1164 return nullptr;
1165}
1166
1167/// Fold
1168/// (1 << NBits) - 1
1169/// Into:
1170/// ~(-(1 << NBits))
1171/// Because a 'not' is better for bit-tracking analysis and other transforms
1172/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1174 InstCombiner::BuilderTy &Builder) {
1175 Value *NBits;
1176 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1177 return nullptr;
1178
1179 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1180 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1181 // Be wary of constant folding.
1182 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1183 // Always NSW. But NUW propagates from `add`.
1184 BOp->setHasNoSignedWrap();
1185 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1186 }
1187
1188 return BinaryOperator::CreateNot(NotMask, I.getName());
1189}
1190
1192 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1193 Type *Ty = I.getType();
1194 auto getUAddSat = [&]() {
1195 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1196 };
1197
1198 // add (umin X, ~Y), Y --> uaddsat X, Y
1199 Value *X, *Y;
1201 m_Deferred(Y))))
1202 return CallInst::Create(getUAddSat(), { X, Y });
1203
1204 // add (umin X, ~C), C --> uaddsat X, C
1205 const APInt *C, *NotC;
1206 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1207 *C == ~*NotC)
1208 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1209
1210 return nullptr;
1211}
1212
1213// Transform:
1214// (add A, (shl (neg B), Y))
1215// -> (sub A, (shl B, Y))
1217 const BinaryOperator &I) {
1218 Value *A, *B, *Cnt;
1219 if (match(&I,
1221 m_Value(A)))) {
1222 Value *NewShl = Builder.CreateShl(B, Cnt);
1223 return BinaryOperator::CreateSub(A, NewShl);
1224 }
1225 return nullptr;
1226}
1227
1228/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1230 // Division must be by power-of-2, but not the minimum signed value.
1231 Value *X;
1232 const APInt *DivC;
1233 if (!match(Add.getOperand(0), m_SDiv(m_Value(X), m_Power2(DivC))) ||
1234 DivC->isNegative())
1235 return nullptr;
1236
1237 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1238 // low bits set. It recognizes two canonical patterns:
1239 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1240 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1241 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1242 // Note that, by the time we end up here, if possible, ugt has been
1243 // canonicalized into eq.
1244 const APInt *MaskC, *MaskCCmp;
1246 if (!match(Add.getOperand(1),
1247 m_SExt(m_ICmp(Pred, m_And(m_Specific(X), m_APInt(MaskC)),
1248 m_APInt(MaskCCmp)))))
1249 return nullptr;
1250
1251 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1252 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1253 return nullptr;
1254
1255 APInt SMin = APInt::getSignedMinValue(Add.getType()->getScalarSizeInBits());
1256 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1257 ? (*MaskC == (SMin | (*DivC - 1)))
1258 : (*DivC == 2 && *MaskC == SMin + 1);
1259 if (!IsMaskValid)
1260 return nullptr;
1261
1262 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1263 return BinaryOperator::CreateAShr(
1264 X, ConstantInt::get(Add.getType(), DivC->exactLogBase2()));
1265}
1266
1269 BinaryOperator &I) {
1270 assert((I.getOpcode() == Instruction::Add ||
1271 I.getOpcode() == Instruction::Or ||
1272 I.getOpcode() == Instruction::Sub) &&
1273 "Expecting add/or/sub instruction");
1274
1275 // We have a subtraction/addition between a (potentially truncated) *logical*
1276 // right-shift of X and a "select".
1277 Value *X, *Select;
1278 Instruction *LowBitsToSkip, *Extract;
1280 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1281 m_Instruction(Extract))),
1282 m_Value(Select))))
1283 return nullptr;
1284
1285 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1286 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1287 return nullptr;
1288
1289 Type *XTy = X->getType();
1290 bool HadTrunc = I.getType() != XTy;
1291
1292 // If there was a truncation of extracted value, then we'll need to produce
1293 // one extra instruction, so we need to ensure one instruction will go away.
1294 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1295 return nullptr;
1296
1297 // Extraction should extract high NBits bits, with shift amount calculated as:
1298 // low bits to skip = shift bitwidth - high bits to extract
1299 // The shift amount itself may be extended, and we need to look past zero-ext
1300 // when matching NBits, that will matter for matching later.
1301 Constant *C;
1302 Value *NBits;
1303 if (!match(
1304 LowBitsToSkip,
1307 APInt(C->getType()->getScalarSizeInBits(),
1308 X->getType()->getScalarSizeInBits()))))
1309 return nullptr;
1310
1311 // Sign-extending value can be zero-extended if we `sub`tract it,
1312 // or sign-extended otherwise.
1313 auto SkipExtInMagic = [&I](Value *&V) {
1314 if (I.getOpcode() == Instruction::Sub)
1315 match(V, m_ZExtOrSelf(m_Value(V)));
1316 else
1317 match(V, m_SExtOrSelf(m_Value(V)));
1318 };
1319
1320 // Now, finally validate the sign-extending magic.
1321 // `select` itself may be appropriately extended, look past that.
1322 SkipExtInMagic(Select);
1323
1325 const APInt *Thr;
1326 Value *SignExtendingValue, *Zero;
1327 bool ShouldSignext;
1328 // It must be a select between two values we will later establish to be a
1329 // sign-extending value and a zero constant. The condition guarding the
1330 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1331 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1332 m_Value(SignExtendingValue), m_Value(Zero))) ||
1333 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1334 return nullptr;
1335
1336 // icmp-select pair is commutative.
1337 if (!ShouldSignext)
1338 std::swap(SignExtendingValue, Zero);
1339
1340 // If we should not perform sign-extension then we must add/or/subtract zero.
1341 if (!match(Zero, m_Zero()))
1342 return nullptr;
1343 // Otherwise, it should be some constant, left-shifted by the same NBits we
1344 // had in `lshr`. Said left-shift can also be appropriately extended.
1345 // Again, we must look past zero-ext when looking for NBits.
1346 SkipExtInMagic(SignExtendingValue);
1347 Constant *SignExtendingValueBaseConstant;
1348 if (!match(SignExtendingValue,
1349 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1350 m_ZExtOrSelf(m_Specific(NBits)))))
1351 return nullptr;
1352 // If we `sub`, then the constant should be one, else it should be all-ones.
1353 if (I.getOpcode() == Instruction::Sub
1354 ? !match(SignExtendingValueBaseConstant, m_One())
1355 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1356 return nullptr;
1357
1358 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1359 Extract->getName() + ".sext");
1360 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1361 if (!HadTrunc)
1362 return NewAShr;
1363
1364 Builder.Insert(NewAShr);
1365 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1366}
1367
1368/// This is a specialization of a more general transform from
1369/// foldUsingDistributiveLaws. If that code can be made to work optimally
1370/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1372 InstCombiner::BuilderTy &Builder) {
1373 // TODO: Also handle mul by doubling the shift amount?
1374 assert((I.getOpcode() == Instruction::Add ||
1375 I.getOpcode() == Instruction::Sub) &&
1376 "Expected add/sub");
1377 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1378 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1379 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1380 return nullptr;
1381
1382 Value *X, *Y, *ShAmt;
1383 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1384 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1385 return nullptr;
1386
1387 // No-wrap propagates only when all ops have no-wrap.
1388 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1389 Op1->hasNoSignedWrap();
1390 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1391 Op1->hasNoUnsignedWrap();
1392
1393 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1394 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1395 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1396 NewI->setHasNoSignedWrap(HasNSW);
1397 NewI->setHasNoUnsignedWrap(HasNUW);
1398 }
1399 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1400 NewShl->setHasNoSignedWrap(HasNSW);
1401 NewShl->setHasNoUnsignedWrap(HasNUW);
1402 return NewShl;
1403}
1404
1405/// Reduce a sequence of masked half-width multiplies to a single multiply.
1406/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1408 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1409 // Skip the odd bitwidth types.
1410 if ((BitWidth & 0x1))
1411 return nullptr;
1412
1413 unsigned HalfBits = BitWidth >> 1;
1414 APInt HalfMask = APInt::getMaxValue(HalfBits);
1415
1416 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1417 Value *XLo, *YLo;
1418 Value *CrossSum;
1419 // Require one-use on the multiply to avoid increasing the number of
1420 // multiplications.
1421 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1422 m_OneUse(m_Mul(m_Value(YLo), m_Value(XLo))))))
1423 return nullptr;
1424
1425 // XLo = X & HalfMask
1426 // YLo = Y & HalfMask
1427 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1428 // to enhance robustness
1429 Value *X, *Y;
1430 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1431 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1432 return nullptr;
1433
1434 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1435 // X' can be either X or XLo in the pattern (and the same for Y')
1436 if (match(CrossSum,
1441 return BinaryOperator::CreateMul(X, Y);
1442
1443 return nullptr;
1444}
1445
1447 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1448 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1450 return replaceInstUsesWith(I, V);
1451
1453 return &I;
1454
1456 return X;
1457
1459 return Phi;
1460
1461 // (A*B)+(A*C) -> A*(B+C) etc
1463 return replaceInstUsesWith(I, V);
1464
1465 if (Instruction *R = foldBoxMultiply(I))
1466 return R;
1467
1469 return R;
1470
1472 return X;
1473
1475 return X;
1476
1478 return R;
1479
1481 return R;
1482
1483 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1484 Type *Ty = I.getType();
1485 if (Ty->isIntOrIntVectorTy(1))
1486 return BinaryOperator::CreateXor(LHS, RHS);
1487
1488 // X + X --> X << 1
1489 if (LHS == RHS) {
1490 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1491 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1492 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1493 return Shl;
1494 }
1495
1496 Value *A, *B;
1497 if (match(LHS, m_Neg(m_Value(A)))) {
1498 // -A + -B --> -(A + B)
1499 if (match(RHS, m_Neg(m_Value(B))))
1501
1502 // -A + B --> B - A
1503 auto *Sub = BinaryOperator::CreateSub(RHS, A);
1504 auto *OB0 = cast<OverflowingBinaryOperator>(LHS);
1505 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1506
1507 return Sub;
1508 }
1509
1510 // A + -B --> A - B
1511 if (match(RHS, m_Neg(m_Value(B))))
1512 return BinaryOperator::CreateSub(LHS, B);
1513
1515 return replaceInstUsesWith(I, V);
1516
1517 // (A + 1) + ~B --> A - B
1518 // ~B + (A + 1) --> A - B
1519 // (~B + A) + 1 --> A - B
1520 // (A + ~B) + 1 --> A - B
1521 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1523 return BinaryOperator::CreateSub(A, B);
1524
1525 // (A + RHS) + RHS --> A + (RHS << 1)
1527 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1528
1529 // LHS + (A + LHS) --> A + (LHS << 1)
1531 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1532
1533 {
1534 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1535 Constant *C1, *C2;
1536 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1537 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1538 (LHS->hasOneUse() || RHS->hasOneUse())) {
1539 Value *Sub = Builder.CreateSub(A, B);
1540 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1541 }
1542
1543 // Canonicalize a constant sub operand as an add operand for better folding:
1544 // (C1 - A) + B --> (B - A) + C1
1546 m_Value(B)))) {
1547 Value *Sub = Builder.CreateSub(B, A, "reass.sub");
1548 return BinaryOperator::CreateAdd(Sub, C1);
1549 }
1550 }
1551
1552 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1554
1555 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1556 const APInt *C1, *C2;
1557 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1558 APInt one(C2->getBitWidth(), 1);
1559 APInt minusC1 = -(*C1);
1560 if (minusC1 == (one << *C2)) {
1561 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1562 return BinaryOperator::CreateSRem(RHS, NewRHS);
1563 }
1564 }
1565
1566 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1567 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1568 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countl_zero())) {
1569 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1570 return BinaryOperator::CreateAnd(A, NewMask);
1571 }
1572
1573 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1574 if ((match(RHS, m_ZExt(m_Value(A))) &&
1576 (match(LHS, m_ZExt(m_Value(A))) &&
1578 return new ZExtInst(B, LHS->getType());
1579
1580 // zext(A) + sext(A) --> 0 if A is i1
1582 A->getType()->isIntOrIntVectorTy(1))
1583 return replaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1584
1585 // A+B --> A|B iff A and B have no bits set in common.
1586 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1587 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(&I)))
1588 return BinaryOperator::CreateDisjointOr(LHS, RHS);
1589
1590 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1591 return Ext;
1592
1593 // (add (xor A, B) (and A, B)) --> (or A, B)
1594 // (add (and A, B) (xor A, B)) --> (or A, B)
1595 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1597 return BinaryOperator::CreateOr(A, B);
1598
1599 // (add (or A, B) (and A, B)) --> (add A, B)
1600 // (add (and A, B) (or A, B)) --> (add A, B)
1601 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1603 // Replacing operands in-place to preserve nuw/nsw flags.
1604 replaceOperand(I, 0, A);
1605 replaceOperand(I, 1, B);
1606 return &I;
1607 }
1608
1609 // (add A (or A, -A)) --> (and (add A, -1) A)
1610 // (add A (or -A, A)) --> (and (add A, -1) A)
1611 // (add (or A, -A) A) --> (and (add A, -1) A)
1612 // (add (or -A, A) A) --> (and (add A, -1) A)
1614 m_Deferred(A)))))) {
1615 Value *Add =
1617 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1618 return BinaryOperator::CreateAnd(Add, A);
1619 }
1620
1621 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1622 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1623 if (match(&I,
1625 m_AllOnes()))) {
1627 Value *Dec = Builder.CreateAdd(A, AllOnes);
1628 Value *Not = Builder.CreateXor(A, AllOnes);
1629 return BinaryOperator::CreateAnd(Dec, Not);
1630 }
1631
1632 // Disguised reassociation/factorization:
1633 // ~(A * C1) + A
1634 // ((A * -C1) - 1) + A
1635 // ((A * -C1) + A) - 1
1636 // (A * (1 - C1)) - 1
1637 if (match(&I,
1639 m_Deferred(A)))) {
1640 Type *Ty = I.getType();
1641 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1642 Value *NewMul = Builder.CreateMul(A, NewMulC);
1643 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1644 }
1645
1646 // (A * -2**C) + B --> B - (A << C)
1647 const APInt *NegPow2C;
1648 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1649 m_Value(B)))) {
1650 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countr_zero());
1651 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1652 return BinaryOperator::CreateSub(B, Shl);
1653 }
1654
1655 // Canonicalize signum variant that ends in add:
1656 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1661 m_OneUse(m_ICmp(Pred, m_Specific(A), m_ZeroInt()))))) &&
1662 Pred == CmpInst::ICMP_SGT) {
1663 Value *NotZero = Builder.CreateIsNotNull(A, "isnotnull");
1664 Value *Zext = Builder.CreateZExt(NotZero, Ty, "isnotnull.zext");
1665 return BinaryOperator::CreateOr(LHS, Zext);
1666 }
1667
1668 if (Instruction *Ashr = foldAddToAshr(I))
1669 return Ashr;
1670
1671 // (~X) + (~Y) --> -2 - (X + Y)
1672 {
1673 // To ensure we can save instructions we need to ensure that we consume both
1674 // LHS/RHS (i.e they have a `not`).
1675 bool ConsumesLHS, ConsumesRHS;
1676 if (isFreeToInvert(LHS, LHS->hasOneUse(), ConsumesLHS) && ConsumesLHS &&
1677 isFreeToInvert(RHS, RHS->hasOneUse(), ConsumesRHS) && ConsumesRHS) {
1680 assert(NotLHS != nullptr && NotRHS != nullptr &&
1681 "isFreeToInvert desynced with getFreelyInverted");
1682 Value *LHSPlusRHS = Builder.CreateAdd(NotLHS, NotRHS);
1683 return BinaryOperator::CreateSub(
1684 ConstantInt::getSigned(RHS->getType(), -2), LHSPlusRHS);
1685 }
1686 }
1687
1689 return R;
1690
1691 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1692 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1693 // computeKnownBits.
1694 bool Changed = false;
1695 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHSCache, RHSCache, I)) {
1696 Changed = true;
1697 I.setHasNoSignedWrap(true);
1698 }
1699 if (!I.hasNoUnsignedWrap() &&
1700 willNotOverflowUnsignedAdd(LHSCache, RHSCache, I)) {
1701 Changed = true;
1702 I.setHasNoUnsignedWrap(true);
1703 }
1704
1706 return V;
1707
1708 if (Instruction *V =
1710 return V;
1711
1713 return SatAdd;
1714
1715 // usub.sat(A, B) + B => umax(A, B)
1716 if (match(&I, m_c_BinOp(
1717 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1718 m_Deferred(B)))) {
1719 return replaceInstUsesWith(I,
1720 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1721 }
1722
1723 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1724 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1725 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1727 return replaceInstUsesWith(
1728 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1729 {Builder.CreateOr(A, B)}));
1730
1731 // Fold the log2_ceil idiom:
1732 // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
1733 // -->
1734 // BW - ctlz(A - 1, false)
1735 const APInt *XorC;
1736 if (match(&I,
1737 m_c_Add(
1738 m_ZExt(m_ICmp(Pred, m_Intrinsic<Intrinsic::ctpop>(m_Value(A)),
1739 m_One())),
1742 m_Intrinsic<Intrinsic::ctlz>(m_Deferred(A), m_One())))),
1743 m_APInt(XorC))))))) &&
1744 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
1745 *XorC == A->getType()->getScalarSizeInBits() - 1) {
1746 Value *Sub = Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()));
1747 Value *Ctlz = Builder.CreateIntrinsic(Intrinsic::ctlz, {A->getType()},
1748 {Sub, Builder.getFalse()});
1749 Value *Ret = Builder.CreateSub(
1750 ConstantInt::get(A->getType(), A->getType()->getScalarSizeInBits()),
1751 Ctlz, "", /*HasNUW*/ true, /*HasNSW*/ true);
1752 return replaceInstUsesWith(I, Builder.CreateZExtOrTrunc(Ret, I.getType()));
1753 }
1754
1755 if (Instruction *Res = foldSquareSumInt(I))
1756 return Res;
1757
1758 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1759 return Res;
1760
1762 return Res;
1763
1764 return Changed ? &I : nullptr;
1765}
1766
1767/// Eliminate an op from a linear interpolation (lerp) pattern.
1769 InstCombiner::BuilderTy &Builder) {
1770 Value *X, *Y, *Z;
1773 m_Value(Z))))),
1775 return nullptr;
1776
1777 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1778 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1779 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1780 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1781}
1782
1783/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1785 InstCombiner::BuilderTy &Builder) {
1786 assert((I.getOpcode() == Instruction::FAdd ||
1787 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1788 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1789 "FP factorization requires FMF");
1790
1791 if (Instruction *Lerp = factorizeLerp(I, Builder))
1792 return Lerp;
1793
1794 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1795 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1796 return nullptr;
1797
1798 Value *X, *Y, *Z;
1799 bool IsFMul;
1800 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1801 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1802 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1803 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1804 IsFMul = true;
1805 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1806 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1807 IsFMul = false;
1808 else
1809 return nullptr;
1810
1811 // (X * Z) + (Y * Z) --> (X + Y) * Z
1812 // (X * Z) - (Y * Z) --> (X - Y) * Z
1813 // (X / Z) + (Y / Z) --> (X + Y) / Z
1814 // (X / Z) - (Y / Z) --> (X - Y) / Z
1815 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1816 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1817 : Builder.CreateFSubFMF(X, Y, &I);
1818
1819 // Bail out if we just created a denormal constant.
1820 // TODO: This is copied from a previous implementation. Is it necessary?
1821 const APFloat *C;
1822 if (match(XY, m_APFloat(C)) && !C->isNormal())
1823 return nullptr;
1824
1825 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1827}
1828
1830 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1831 I.getFastMathFlags(),
1833 return replaceInstUsesWith(I, V);
1834
1836 return &I;
1837
1839 return X;
1840
1842 return Phi;
1843
1844 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1845 return FoldedFAdd;
1846
1847 // (-X) + Y --> Y - X
1848 Value *X, *Y;
1849 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1851
1852 // Similar to above, but look through fmul/fdiv for the negated term.
1853 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1854 Value *Z;
1856 m_Value(Z)))) {
1857 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1858 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1859 }
1860 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1861 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1863 m_Value(Z))) ||
1865 m_Value(Z)))) {
1866 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1867 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1868 }
1869
1870 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1871 // integer add followed by a promotion.
1872 if (Instruction *R = foldFBinOpOfIntCasts(I))
1873 return R;
1874
1875 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1876 // Handle specials cases for FAdd with selects feeding the operation
1878 return replaceInstUsesWith(I, V);
1879
1880 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1882 return F;
1883
1885 return F;
1886
1887 // Try to fold fadd into start value of reduction intrinsic.
1888 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1889 m_AnyZeroFP(), m_Value(X))),
1890 m_Value(Y)))) {
1891 // fadd (rdx 0.0, X), Y --> rdx Y, X
1892 return replaceInstUsesWith(
1893 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1894 {X->getType()}, {Y, X}, &I));
1895 }
1896 const APFloat *StartC, *C;
1897 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1898 m_APFloat(StartC), m_Value(X)))) &&
1899 match(RHS, m_APFloat(C))) {
1900 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1901 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1902 return replaceInstUsesWith(
1903 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1904 {X->getType()}, {NewStartC, X}, &I));
1905 }
1906
1907 // (X * MulC) + X --> X * (MulC + 1.0)
1908 Constant *MulC;
1909 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1910 m_Deferred(X)))) {
1912 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1913 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1914 }
1915
1916 // (-X - Y) + (X + Z) --> Z - Y
1918 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1919 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1920
1921 if (Value *V = FAddCombine(Builder).simplify(&I))
1922 return replaceInstUsesWith(I, V);
1923 }
1924
1925 // minumum(X, Y) + maximum(X, Y) => X + Y.
1926 if (match(&I,
1927 m_c_FAdd(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
1928 m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
1929 m_Deferred(Y))))) {
1931 // We cannot preserve ninf if nnan flag is not set.
1932 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
1933 // while in optimized version NaN + Inf and this is a poison with ninf flag.
1934 if (!Result->hasNoNaNs())
1935 Result->setHasNoInfs(false);
1936 return Result;
1937 }
1938
1939 return nullptr;
1940}
1941
1942/// Optimize pointer differences into the same array into a size. Consider:
1943/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1944/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1946 Type *Ty, bool IsNUW) {
1947 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1948 // this.
1949 bool Swapped = false;
1950 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1951 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1952 std::swap(LHS, RHS);
1953 Swapped = true;
1954 }
1955
1956 // Require at least one GEP with a common base pointer on both sides.
1957 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1958 // (gep X, ...) - X
1959 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1961 GEP1 = LHSGEP;
1962 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1963 // (gep X, ...) - (gep X, ...)
1964 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1965 RHSGEP->getOperand(0)->stripPointerCasts()) {
1966 GEP1 = LHSGEP;
1967 GEP2 = RHSGEP;
1968 }
1969 }
1970 }
1971
1972 if (!GEP1)
1973 return nullptr;
1974
1975 if (GEP2) {
1976 // (gep X, ...) - (gep X, ...)
1977 //
1978 // Avoid duplicating the arithmetic if there are more than one non-constant
1979 // indices between the two GEPs and either GEP has a non-constant index and
1980 // multiple users. If zero non-constant index, the result is a constant and
1981 // there is no duplication. If one non-constant index, the result is an add
1982 // or sub with a constant, which is no larger than the original code, and
1983 // there's no duplicated arithmetic, even if either GEP has multiple
1984 // users. If more than one non-constant indices combined, as long as the GEP
1985 // with at least one non-constant index doesn't have multiple users, there
1986 // is no duplication.
1987 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1988 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1989 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1990 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1991 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1992 return nullptr;
1993 }
1994 }
1995
1996 // Emit the offset of the GEP and an intptr_t.
1997 Value *Result = EmitGEPOffset(GEP1);
1998
1999 // If this is a single inbounds GEP and the original sub was nuw,
2000 // then the final multiplication is also nuw.
2001 if (auto *I = dyn_cast<Instruction>(Result))
2002 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
2003 I->getOpcode() == Instruction::Mul)
2004 I->setHasNoUnsignedWrap();
2005
2006 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2007 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2008 if (GEP2) {
2009 Value *Offset = EmitGEPOffset(GEP2);
2010 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
2011 GEP1->isInBounds() && GEP2->isInBounds());
2012 }
2013
2014 // If we have p - gep(p, ...) then we have to negate the result.
2015 if (Swapped)
2016 Result = Builder.CreateNeg(Result, "diff.neg");
2017
2018 return Builder.CreateIntCast(Result, Ty, true);
2019}
2020
2022 InstCombiner::BuilderTy &Builder) {
2023 Value *Op0 = I.getOperand(0);
2024 Value *Op1 = I.getOperand(1);
2025 Type *Ty = I.getType();
2026 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
2027 if (!MinMax)
2028 return nullptr;
2029
2030 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2031 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2032 Value *X = MinMax->getLHS();
2033 Value *Y = MinMax->getRHS();
2034 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2035 (Op0->hasOneUse() || Op1->hasOneUse())) {
2036 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2037 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2038 return CallInst::Create(F, {X, Y});
2039 }
2040
2041 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2042 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2043 Value *Z;
2044 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
2045 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
2046 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
2047 return BinaryOperator::CreateAdd(X, USub);
2048 }
2049 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
2050 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
2051 return BinaryOperator::CreateAdd(X, USub);
2052 }
2053 }
2054
2055 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2056 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2057 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
2058 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
2059 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
2060 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
2061 return CallInst::Create(F, {Op0, Z});
2062 }
2063
2064 return nullptr;
2065}
2066
2068 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
2069 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
2071 return replaceInstUsesWith(I, V);
2072
2074 return X;
2075
2077 return Phi;
2078
2079 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2080
2081 // If this is a 'B = x-(-A)', change to B = x+A.
2082 // We deal with this without involving Negator to preserve NSW flag.
2083 if (Value *V = dyn_castNegVal(Op1)) {
2084 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2085
2086 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
2087 assert(BO->getOpcode() == Instruction::Sub &&
2088 "Expected a subtraction operator!");
2089 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2090 Res->setHasNoSignedWrap(true);
2091 } else {
2092 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2093 Res->setHasNoSignedWrap(true);
2094 }
2095
2096 return Res;
2097 }
2098
2099 // Try this before Negator to preserve NSW flag.
2101 return R;
2102
2103 Constant *C;
2104 if (match(Op0, m_ImmConstant(C))) {
2105 Value *X;
2106 Constant *C2;
2107
2108 // C-(X+C2) --> (C-C2)-X
2109 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) {
2110 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2111 // => (C-C2)-X can have NSW/NUW
2112 bool WillNotSOV = willNotOverflowSignedSub(C, C2, I);
2113 BinaryOperator *Res =
2114 BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
2115 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2116 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2117 WillNotSOV);
2118 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2119 OBO1->hasNoUnsignedWrap());
2120 return Res;
2121 }
2122 }
2123
2124 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2125 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2126 return Ext;
2127
2128 bool Changed = false;
2129 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2130 Changed = true;
2131 I.setHasNoSignedWrap(true);
2132 }
2133 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2134 Changed = true;
2135 I.setHasNoUnsignedWrap(true);
2136 }
2137
2138 return Changed ? &I : nullptr;
2139 };
2140
2141 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2142 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2143 // a pure negation used by a select that looks like abs/nabs.
2144 bool IsNegation = match(Op0, m_ZeroInt());
2145 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
2146 const Instruction *UI = dyn_cast<Instruction>(U);
2147 if (!UI)
2148 return false;
2149 return match(UI,
2150 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
2151 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
2152 })) {
2153 if (Value *NegOp1 = Negator::Negate(IsNegation, /* IsNSW */ IsNegation &&
2154 I.hasNoSignedWrap(),
2155 Op1, *this))
2156 return BinaryOperator::CreateAdd(NegOp1, Op0);
2157 }
2158 if (IsNegation)
2159 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2160
2161 // (A*B)-(A*C) -> A*(B-C) etc
2163 return replaceInstUsesWith(I, V);
2164
2165 if (I.getType()->isIntOrIntVectorTy(1))
2166 return BinaryOperator::CreateXor(Op0, Op1);
2167
2168 // Replace (-1 - A) with (~A).
2169 if (match(Op0, m_AllOnes()))
2170 return BinaryOperator::CreateNot(Op1);
2171
2172 // (X + -1) - Y --> ~Y + X
2173 Value *X, *Y;
2174 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
2175 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
2176
2177 // Reassociate sub/add sequences to create more add instructions and
2178 // reduce dependency chains:
2179 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2180 Value *Z;
2182 m_Value(Z))))) {
2183 Value *XZ = Builder.CreateAdd(X, Z);
2184 Value *YW = Builder.CreateAdd(Y, Op1);
2185 return BinaryOperator::CreateSub(XZ, YW);
2186 }
2187
2188 // ((X - Y) - Op1) --> X - (Y + Op1)
2189 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
2190 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Op0);
2191 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2192 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2193 Value *Add = Builder.CreateAdd(Y, Op1, "", /* HasNUW */ HasNUW,
2194 /* HasNSW */ HasNSW);
2195 BinaryOperator *Sub = BinaryOperator::CreateSub(X, Add);
2196 Sub->setHasNoUnsignedWrap(HasNUW);
2197 Sub->setHasNoSignedWrap(HasNSW);
2198 return Sub;
2199 }
2200
2201 {
2202 // (X + Z) - (Y + Z) --> (X - Y)
2203 // This is done in other passes, but we want to be able to consume this
2204 // pattern in InstCombine so we can generate it without creating infinite
2205 // loops.
2206 if (match(Op0, m_Add(m_Value(X), m_Value(Z))) &&
2207 match(Op1, m_c_Add(m_Value(Y), m_Specific(Z))))
2208 return BinaryOperator::CreateSub(X, Y);
2209
2210 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2211 Constant *CX, *CY;
2212 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_ImmConstant(CX)))) &&
2213 match(Op1, m_OneUse(m_Add(m_Value(Y), m_ImmConstant(CY))))) {
2214 Value *OpsSub = Builder.CreateSub(X, Y);
2215 Constant *ConstsSub = ConstantExpr::getSub(CX, CY);
2216 return BinaryOperator::CreateAdd(OpsSub, ConstsSub);
2217 }
2218 }
2219
2220 // (~X) - (~Y) --> Y - X
2221 {
2222 // Need to ensure we can consume at least one of the `not` instructions,
2223 // otherwise this can inf loop.
2224 bool ConsumesOp0, ConsumesOp1;
2225 if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
2226 isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
2227 (ConsumesOp0 || ConsumesOp1)) {
2228 Value *NotOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
2229 Value *NotOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
2230 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2231 "isFreeToInvert desynced with getFreelyInverted");
2232 return BinaryOperator::CreateSub(NotOp1, NotOp0);
2233 }
2234 }
2235
2236 auto m_AddRdx = [](Value *&Vec) {
2237 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
2238 };
2239 Value *V0, *V1;
2240 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
2241 V0->getType() == V1->getType()) {
2242 // Difference of sums is sum of differences:
2243 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2244 Value *Sub = Builder.CreateSub(V0, V1);
2245 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2246 {Sub->getType()}, {Sub});
2247 return replaceInstUsesWith(I, Rdx);
2248 }
2249
2250 if (Constant *C = dyn_cast<Constant>(Op0)) {
2251 Value *X;
2252 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2253 // C - (zext bool) --> bool ? C - 1 : C
2255 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2256 // C - (sext bool) --> bool ? C + 1 : C
2258
2259 // C - ~X == X + (1+C)
2260 if (match(Op1, m_Not(m_Value(X))))
2261 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2262
2263 // Try to fold constant sub into select arguments.
2264 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2265 if (Instruction *R = FoldOpIntoSelect(I, SI))
2266 return R;
2267
2268 // Try to fold constant sub into PHI values.
2269 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2270 if (Instruction *R = foldOpIntoPhi(I, PN))
2271 return R;
2272
2273 Constant *C2;
2274
2275 // C-(C2-X) --> X+(C-C2)
2276 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2277 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2278 }
2279
2280 const APInt *Op0C;
2281 if (match(Op0, m_APInt(Op0C))) {
2282 if (Op0C->isMask()) {
2283 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2284 // zero.
2285 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
2286 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2287 return BinaryOperator::CreateXor(Op1, Op0);
2288 }
2289
2290 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2291 // (C3 - ((C2 & C3) - 1)) is pow2
2292 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2293 // C2 is negative pow2 || sub nuw
2294 const APInt *C2, *C3;
2295 BinaryOperator *InnerSub;
2296 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2297 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2298 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2299 APInt C2AndC3 = *C2 & *C3;
2300 APInt C2AndC3Minus1 = C2AndC3 - 1;
2301 APInt C2AddC3 = *C2 + *C3;
2302 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2303 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2304 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2305 return BinaryOperator::CreateAdd(
2306 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2307 }
2308 }
2309 }
2310
2311 {
2312 Value *Y;
2313 // X-(X+Y) == -Y X-(Y+X) == -Y
2314 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2316
2317 // (X-Y)-X == -Y
2318 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2320 }
2321
2322 // (sub (or A, B) (and A, B)) --> (xor A, B)
2323 {
2324 Value *A, *B;
2325 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2326 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2327 return BinaryOperator::CreateXor(A, B);
2328 }
2329
2330 // (sub (add A, B) (or A, B)) --> (and A, B)
2331 {
2332 Value *A, *B;
2333 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2334 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2335 return BinaryOperator::CreateAnd(A, B);
2336 }
2337
2338 // (sub (add A, B) (and A, B)) --> (or A, B)
2339 {
2340 Value *A, *B;
2341 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2343 return BinaryOperator::CreateOr(A, B);
2344 }
2345
2346 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2347 {
2348 Value *A, *B;
2349 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2350 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2351 (Op0->hasOneUse() || Op1->hasOneUse()))
2353 }
2354
2355 // (sub (or A, B), (xor A, B)) --> (and A, B)
2356 {
2357 Value *A, *B;
2358 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2359 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2360 return BinaryOperator::CreateAnd(A, B);
2361 }
2362
2363 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2364 {
2365 Value *A, *B;
2366 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2367 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2368 (Op0->hasOneUse() || Op1->hasOneUse()))
2370 }
2371
2372 {
2373 Value *Y;
2374 // ((X | Y) - X) --> (~X & Y)
2375 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2376 return BinaryOperator::CreateAnd(
2377 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2378 }
2379
2380 {
2381 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2382 Value *X;
2383 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2384 m_OneUse(m_Neg(m_Value(X))))))) {
2386 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2387 }
2388 }
2389
2390 {
2391 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2392 Constant *C;
2393 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2396 }
2397 }
2398
2399 {
2400 // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2401 // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2402 Value *C, *X;
2403 auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2404 return match(LHS, m_OneUse(m_c_Xor(m_Value(X), m_Specific(RHS)))) &&
2405 match(RHS, m_SExt(m_Value(C))) &&
2406 (C->getType()->getScalarSizeInBits() == 1);
2407 };
2408 if (m_SubXorCmp(Op0, Op1))
2410 if (m_SubXorCmp(Op1, Op0))
2412 }
2413
2415 return R;
2416
2418 return R;
2419
2420 {
2421 // If we have a subtraction between some value and a select between
2422 // said value and something else, sink subtraction into select hands, i.e.:
2423 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2424 // ->
2425 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2426 // or
2427 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2428 // ->
2429 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2430 // This will result in select between new subtraction and 0.
2431 auto SinkSubIntoSelect =
2432 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2433 auto SubBuilder) -> Instruction * {
2434 Value *Cond, *TrueVal, *FalseVal;
2435 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2436 m_Value(FalseVal)))))
2437 return nullptr;
2438 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2439 return nullptr;
2440 // While it is really tempting to just create two subtractions and let
2441 // InstCombine fold one of those to 0, it isn't possible to do so
2442 // because of worklist visitation order. So ugly it is.
2443 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2444 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2445 Constant *Zero = Constant::getNullValue(Ty);
2446 SelectInst *NewSel =
2447 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2448 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2449 // Preserve prof metadata if any.
2450 NewSel->copyMetadata(cast<Instruction>(*Select));
2451 return NewSel;
2452 };
2453 if (Instruction *NewSel = SinkSubIntoSelect(
2454 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2455 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2456 return Builder->CreateSub(OtherHandOfSelect,
2457 /*OtherHandOfSub=*/Op1);
2458 }))
2459 return NewSel;
2460 if (Instruction *NewSel = SinkSubIntoSelect(
2461 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2462 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2463 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2464 OtherHandOfSelect);
2465 }))
2466 return NewSel;
2467 }
2468
2469 // (X - (X & Y)) --> (X & ~Y)
2470 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2471 (Op1->hasOneUse() || isa<Constant>(Y)))
2472 return BinaryOperator::CreateAnd(
2473 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2474
2475 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2476 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2477 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2478 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2479 // As long as Y is freely invertible, this will be neutral or a win.
2480 // Note: We don't generate the inverse max/min, just create the 'not' of
2481 // it and let other folds do the rest.
2482 if (match(Op0, m_Not(m_Value(X))) &&
2483 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2484 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2485 Value *Not = Builder.CreateNot(Op1);
2486 return BinaryOperator::CreateSub(Not, X);
2487 }
2488 if (match(Op1, m_Not(m_Value(X))) &&
2489 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2490 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2491 Value *Not = Builder.CreateNot(Op0);
2492 return BinaryOperator::CreateSub(X, Not);
2493 }
2494
2495 // Optimize pointer differences into the same array into a size. Consider:
2496 // &A[10] - &A[0]: we should compile this to "10".
2497 Value *LHSOp, *RHSOp;
2498 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2499 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2500 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2501 I.hasNoUnsignedWrap()))
2502 return replaceInstUsesWith(I, Res);
2503
2504 // trunc(p)-trunc(q) -> trunc(p-q)
2505 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2506 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2507 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2508 /* IsNUW */ false))
2509 return replaceInstUsesWith(I, Res);
2510
2511 // Canonicalize a shifty way to code absolute value to the common pattern.
2512 // There are 2 potential commuted variants.
2513 // We're relying on the fact that we only do this transform when the shift has
2514 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2515 // instructions).
2516 Value *A;
2517 const APInt *ShAmt;
2518 Type *Ty = I.getType();
2519 unsigned BitWidth = Ty->getScalarSizeInBits();
2520 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2521 Op1->hasNUses(2) && *ShAmt == BitWidth - 1 &&
2522 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2523 // B = ashr i32 A, 31 ; smear the sign bit
2524 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2525 // --> (A < 0) ? -A : A
2526 Value *IsNeg = Builder.CreateIsNeg(A);
2527 // Copy the nsw flags from the sub to the negate.
2528 Value *NegA = I.hasNoUnsignedWrap()
2529 ? Constant::getNullValue(A->getType())
2530 : Builder.CreateNeg(A, "", I.hasNoSignedWrap());
2531 return SelectInst::Create(IsNeg, NegA, A);
2532 }
2533
2534 // If we are subtracting a low-bit masked subset of some value from an add
2535 // of that same value with no low bits changed, that is clearing some low bits
2536 // of the sum:
2537 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2538 const APInt *AddC, *AndC;
2539 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2540 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2541 unsigned Cttz = AddC->countr_zero();
2542 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2543 if ((HighMask & *AndC).isZero())
2544 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2545 }
2546
2547 if (Instruction *V =
2549 return V;
2550
2551 // X - usub.sat(X, Y) => umin(X, Y)
2552 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2553 m_Value(Y)))))
2554 return replaceInstUsesWith(
2555 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2556
2557 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2558 // TODO: The one-use restriction is not strictly necessary, but it may
2559 // require improving other pattern matching and/or codegen.
2560 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2561 return replaceInstUsesWith(
2562 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2563
2564 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2565 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2566 return replaceInstUsesWith(
2567 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2568
2569 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2570 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2571 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2572 return BinaryOperator::CreateNeg(USub);
2573 }
2574
2575 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2576 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2577 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2578 return BinaryOperator::CreateNeg(USub);
2579 }
2580
2581 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2582 if (match(Op0, m_SpecificInt(BitWidth)) &&
2583 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2584 return replaceInstUsesWith(
2585 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2586 {Builder.CreateNot(X)}));
2587
2588 // Reduce multiplies for difference-of-squares by factoring:
2589 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2590 if (match(Op0, m_OneUse(m_Mul(m_Value(X), m_Deferred(X)))) &&
2591 match(Op1, m_OneUse(m_Mul(m_Value(Y), m_Deferred(Y))))) {
2592 auto *OBO0 = cast<OverflowingBinaryOperator>(Op0);
2593 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2594 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2595 OBO1->hasNoSignedWrap() && BitWidth > 2;
2596 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2597 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2598 Value *Add = Builder.CreateAdd(X, Y, "add", PropagateNUW, PropagateNSW);
2599 Value *Sub = Builder.CreateSub(X, Y, "sub", PropagateNUW, PropagateNSW);
2600 Value *Mul = Builder.CreateMul(Add, Sub, "", PropagateNUW, PropagateNSW);
2601 return replaceInstUsesWith(I, Mul);
2602 }
2603
2604 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2605 if (match(Op0, m_OneUse(m_c_SMax(m_Value(X), m_Value(Y)))) &&
2607 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2608 Value *Sub =
2609 Builder.CreateSub(X, Y, "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2610 Value *Call =
2611 Builder.CreateBinaryIntrinsic(Intrinsic::abs, Sub, Builder.getTrue());
2612 return replaceInstUsesWith(I, Call);
2613 }
2614 }
2615
2617 return Res;
2618
2619 return TryToNarrowDeduceFlags();
2620}
2621
2622/// This eliminates floating-point negation in either 'fneg(X)' or
2623/// 'fsub(-0.0, X)' form by combining into a constant operand.
2625 // This is limited with one-use because fneg is assumed better for
2626 // reassociation and cheaper in codegen than fmul/fdiv.
2627 // TODO: Should the m_OneUse restriction be removed?
2628 Instruction *FNegOp;
2629 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2630 return nullptr;
2631
2632 Value *X;
2633 Constant *C;
2634
2635 // Fold negation into constant operand.
2636 // -(X * C) --> X * (-C)
2637 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2638 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2639 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2640 // -(X / C) --> X / (-C)
2641 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2642 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2643 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2644 // -(C / X) --> (-C) / X
2645 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2646 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2648
2649 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2650 // not apply to the fdiv. Everything else propagates from the fneg.
2651 // TODO: We could propagate nsz/ninf from fdiv alone?
2652 FastMathFlags FMF = I.getFastMathFlags();
2653 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2654 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2655 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2656 return FDiv;
2657 }
2658 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2659 // -(X + C) --> -X + -C --> -C - X
2660 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2661 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2662 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2663
2664 return nullptr;
2665}
2666
2667Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
2668 Instruction &FMFSource) {
2669 Value *X, *Y;
2670 if (match(FNegOp, m_FMul(m_Value(X), m_Value(Y)))) {
2671 return cast<Instruction>(Builder.CreateFMulFMF(
2672 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2673 }
2674
2675 if (match(FNegOp, m_FDiv(m_Value(X), m_Value(Y)))) {
2676 return cast<Instruction>(Builder.CreateFDivFMF(
2677 Builder.CreateFNegFMF(X, &FMFSource), Y, &FMFSource));
2678 }
2679
2680 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(FNegOp)) {
2681 // Make sure to preserve flags and metadata on the call.
2682 if (II->getIntrinsicID() == Intrinsic::ldexp) {
2683 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
2686
2688 II->getCalledFunction(),
2689 {Builder.CreateFNeg(II->getArgOperand(0)), II->getArgOperand(1)});
2690 New->copyMetadata(*II);
2691 return New;
2692 }
2693 }
2694
2695 return nullptr;
2696}
2697
2699 Value *Op = I.getOperand(0);
2700
2701 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2702 getSimplifyQuery().getWithInstruction(&I)))
2703 return replaceInstUsesWith(I, V);
2704
2706 return X;
2707
2708 Value *X, *Y;
2709
2710 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2711 if (I.hasNoSignedZeros() &&
2714
2715 Value *OneUse;
2716 if (!match(Op, m_OneUse(m_Value(OneUse))))
2717 return nullptr;
2718
2719 if (Instruction *R = hoistFNegAboveFMulFDiv(OneUse, I))
2720 return replaceInstUsesWith(I, R);
2721
2722 // Try to eliminate fneg if at least 1 arm of the select is negated.
2723 Value *Cond;
2724 if (match(OneUse, m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))) {
2725 // Unlike most transforms, this one is not safe to propagate nsz unless
2726 // it is present on the original select. We union the flags from the select
2727 // and fneg and then remove nsz if needed.
2728 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2729 S->copyFastMathFlags(&I);
2730 if (auto *OldSel = dyn_cast<SelectInst>(Op)) {
2731 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
2732 S->setFastMathFlags(FMF);
2733 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2734 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2735 S->setHasNoSignedZeros(false);
2736 }
2737 };
2738 // -(Cond ? -P : Y) --> Cond ? P : -Y
2739 Value *P;
2740 if (match(X, m_FNeg(m_Value(P)))) {
2741 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2742 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2743 propagateSelectFMF(NewSel, P == Y);
2744 return NewSel;
2745 }
2746 // -(Cond ? X : -P) --> Cond ? -X : P
2747 if (match(Y, m_FNeg(m_Value(P)))) {
2748 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2749 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2750 propagateSelectFMF(NewSel, P == X);
2751 return NewSel;
2752 }
2753 }
2754
2755 // fneg (copysign x, y) -> copysign x, (fneg y)
2756 if (match(OneUse, m_CopySign(m_Value(X), m_Value(Y)))) {
2757 // The source copysign has an additional value input, so we can't propagate
2758 // flags the copysign doesn't also have.
2759 FastMathFlags FMF = I.getFastMathFlags();
2760 FMF &= cast<FPMathOperator>(OneUse)->getFastMathFlags();
2761
2764
2765 Value *NegY = Builder.CreateFNeg(Y);
2766 Value *NewCopySign = Builder.CreateCopySign(X, NegY);
2767 return replaceInstUsesWith(I, NewCopySign);
2768 }
2769
2770 return nullptr;
2771}
2772
2774 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2775 I.getFastMathFlags(),
2776 getSimplifyQuery().getWithInstruction(&I)))
2777 return replaceInstUsesWith(I, V);
2778
2780 return X;
2781
2783 return Phi;
2784
2785 // Subtraction from -0.0 is the canonical form of fneg.
2786 // fsub -0.0, X ==> fneg X
2787 // fsub nsz 0.0, X ==> fneg nsz X
2788 //
2789 // FIXME This matcher does not respect FTZ or DAZ yet:
2790 // fsub -0.0, Denorm ==> +-0
2791 // fneg Denorm ==> -Denorm
2792 Value *Op;
2793 if (match(&I, m_FNeg(m_Value(Op))))
2795
2797 return X;
2798
2799 if (Instruction *R = foldFBinOpOfIntCasts(I))
2800 return R;
2801
2802 Value *X, *Y;
2803 Constant *C;
2804
2805 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2806 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2807 // Canonicalize to fadd to make analysis easier.
2808 // This can also help codegen because fadd is commutative.
2809 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2810 // killed later. We still limit that particular transform with 'hasOneUse'
2811 // because an fneg is assumed better/cheaper than a generic fsub.
2812 if (I.hasNoSignedZeros() ||
2813 cannotBeNegativeZero(Op0, 0, getSimplifyQuery().getWithInstruction(&I))) {
2814 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2815 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2816 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2817 }
2818 }
2819
2820 // (-X) - Op1 --> -(X + Op1)
2821 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2822 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2823 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2825 }
2826
2827 if (isa<Constant>(Op0))
2828 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2829 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2830 return NV;
2831
2832 // X - C --> X + (-C)
2833 // But don't transform constant expressions because there's an inverse fold
2834 // for X + (-Y) --> X - Y.
2835 if (match(Op1, m_ImmConstant(C)))
2836 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2837 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2838
2839 // X - (-Y) --> X + Y
2840 if (match(Op1, m_FNeg(m_Value(Y))))
2841 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2842
2843 // Similar to above, but look through a cast of the negated value:
2844 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2845 Type *Ty = I.getType();
2846 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2848
2849 // X - (fpext(-Y)) --> X + fpext(Y)
2850 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2852
2853 // Similar to above, but look through fmul/fdiv of the negated value:
2854 // Op0 - (-X * Y) --> Op0 + (X * Y)
2855 // Op0 - (Y * -X) --> Op0 + (X * Y)
2856 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2858 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2859 }
2860 // Op0 - (-X / Y) --> Op0 + (X / Y)
2861 // Op0 - (X / -Y) --> Op0 + (X / Y)
2862 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2863 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2864 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2865 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2866 }
2867
2868 // Handle special cases for FSub with selects feeding the operation
2869 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2870 return replaceInstUsesWith(I, V);
2871
2872 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2873 // (Y - X) - Y --> -X
2874 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2876
2877 // Y - (X + Y) --> -X
2878 // Y - (Y + X) --> -X
2879 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2881
2882 // (X * C) - X --> X * (C - 1.0)
2883 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2885 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2886 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2887 }
2888 // X - (X * C) --> X * (1.0 - C)
2889 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2891 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2892 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2893 }
2894
2895 // Reassociate fsub/fadd sequences to create more fadd instructions and
2896 // reduce dependency chains:
2897 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2898 Value *Z;
2900 m_Value(Z))))) {
2901 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2902 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2903 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2904 }
2905
2906 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2907 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2908 m_Value(Vec)));
2909 };
2910 Value *A0, *A1, *V0, *V1;
2911 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2912 V0->getType() == V1->getType()) {
2913 // Difference of sums is sum of differences:
2914 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2915 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2916 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2917 {Sub->getType()}, {A0, Sub}, &I);
2918 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2919 }
2920
2922 return F;
2923
2924 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2925 // functionality has been subsumed by simple pattern matching here and in
2926 // InstSimplify. We should let a dedicated reassociation pass handle more
2927 // complex pattern matching and remove this from InstCombine.
2928 if (Value *V = FAddCombine(Builder).simplify(&I))
2929 return replaceInstUsesWith(I, V);
2930
2931 // (X - Y) - Op1 --> X - (Y + Op1)
2932 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2933 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2935 }
2936 }
2937
2938 return nullptr;
2939}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static bool isConstant(const MachineInstr &MI)
amdgpu AMDGPU Register Bank Select
This file declares a class to represent arbitrary precision floating point values and provide a varie...
This file implements a class to represent arbitrary precision integral constant values and operations...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
hexagon bit simplify
static Instruction * factorizeFAddFSub(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Factor a common operand out of fadd/fsub of fmul/fdiv.
static Instruction * foldAddToAshr(BinaryOperator &Add)
Try to reduce signed division by power-of-2 to an arithmetic shift right.
static bool MatchMul(Value *E, Value *&Op, APInt &C)
static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned)
static Instruction * foldFNegIntoConstant(Instruction &I, const DataLayout &DL)
This eliminates floating-point negation in either 'fneg(X)' or 'fsub(-0.0, X)' form by combining into...
static Instruction * combineAddSubWithShlAddSub(InstCombiner::BuilderTy &Builder, const BinaryOperator &I)
static Instruction * factorizeLerp(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Eliminate an op from a linear interpolation (lerp) pattern.
static Instruction * foldSubOfMinMax(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static Instruction * foldBoxMultiply(BinaryOperator &I)
Reduce a sequence of masked half-width multiplies to a single multiply.
static Value * checkForNegativeOperand(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned)
static Instruction * foldNoWrapAdd(BinaryOperator &Add, InstCombiner::BuilderTy &Builder)
Wrapping flags may allow combining constants separated by an extend.
static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A, Value *&B)
static Instruction * factorizeMathWithShlOps(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
This is a specialization of a more general transform from foldUsingDistributiveLaws.
static Instruction * canonicalizeLowbitMask(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
Fold (1 << NBits) - 1 Into: ~(-(1 << NBits)) Because a 'not' is better for bit-tracking analysis and ...
static Instruction * foldToUnsignedSaturatedAdd(BinaryOperator &I)
static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned)
This file provides internal interfaces used to implement the InstCombine.
This file provides the interface for the instcombine pass implementation.
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp: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:329
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:324
static BinaryOperator * CreateFSubFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:319
static BinaryOperator * CreateFAddFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:314
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:960
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:983
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:987
@ ICMP_EQ
equal
Definition: InstrTypes.h:981
@ ICMP_NE
not equal
Definition: InstrTypes.h:982
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2542
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2535
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:268
const APFloat & getValueAPF() const
Definition: Constants.h:311
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:315
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
Definition: Constants.h:123
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:419
unsigned countNonConstantIndices() const
Definition: Operator.h:495
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 * 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:2079
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:2023
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:1734
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2017
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2537
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 * CreateNeg(Value *V, const Twine &Name="", bool HasNSW=false)
Definition: IRBuilder.h:1715
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1743
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:145
Value * CreateIsNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg < 0.
Definition: IRBuilder.h:2532
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h: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:2005
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:2527
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:2174
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2390
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2088
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1513
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1724
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:451
void copyMetadata(const Instruction &SrcInst, ArrayRef< unsigned > WL=ArrayRef< unsigned >())
Copy metadata from SrcInst to this instruction.
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h: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:190
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:1459
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)
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:100
BinaryOp_match< LHS, RHS, Instruction::FMul, true > m_c_FMul(const LHS &L, const RHS &R)
Matches FMul with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::AShr > m_AShr(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::FSub > m_FSub(const LHS &L, const RHS &R)
cst_pred_ty< is_power2 > m_Power2()
Match an integer or vector power-of-2.
Definition: PatternMatch.h: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)
specific_intval< false > m_SpecificInt(const APInt &V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:918
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)
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
match_combine_or< OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoSignedWrap >, DisjointOr_match< LHS, RHS > > m_NSWAddLike(const LHS &L, const RHS &R)
Match either "add nsw" or "or disjoint".
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:92
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
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)
specific_intval< true > m_SpecificIntAllowUndef(const APInt &V)
Definition: PatternMatch.h:926
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.
match_combine_or< OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap >, DisjointOr_match< LHS, RHS > > m_NUWAddLike(const LHS &L, const RHS &R)
Match either "add nuw" or "or disjoint".
BinaryOp_match< LHS, RHS, Instruction::Mul, true > m_c_Mul(const LHS &L, const RHS &R)
Matches a Mul with LHS and RHS in either order.
m_Intrinsic_Ty< Opnd0, Opnd1 >::Ty m_CopySign(const Opnd0 &Op0, const Opnd1 &Op1)
CastOperator_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty > m_UMin(const LHS &L, const RHS &R)
match_combine_or< LTy, RTy > m_CombineOr(const LTy &L, const RTy &R)
Combine two pattern matchers matching L || R.
Definition: PatternMatch.h: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