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"
17#include "llvm/ADT/SmallVector.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/ValueTracking.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"
24#include "llvm/IR/Instructions.h"
25#include "llvm/IR/Operator.h"
26#include "llvm/IR/PatternMatch.h"
27#include "llvm/IR/Type.h"
28#include "llvm/IR/Value.h"
29#include "llvm/Support/AlignOf.h"
30#include "llvm/Support/Casting.h"
31#include "llvm/Support/KnownBits.h"
32#include "llvm/Transforms/InstCombine/InstCombiner.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
123 AlignedCharArrayUnion<APFloat> FpValBuf;
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
182 Value *simplify(Instruction *FAdd);
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
212 InstCombiner::BuilderTy &Builder;
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(RHS: That.getFpVal(), RM: RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(Sem: T.getSemantics());
286 getFpVal().add(RHS: T, RM: RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(RHS: createAPFloatFromInt(Sem: T.getSemantics(), Val: That.IntVal), RM: 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(Sem: Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(RHS: createAPFloatFromInt(Sem: Semantic, Val: That.IntVal),
319 RM: APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(RHS: That.getFpVal(), RM: 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, V: float(IntVal)) :
334 ConstantFP::get(Context&: Ty->getContext(), V: 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(i: 0);
358 Value *Opnd1 = I->getOperand(i: 1);
359 if ((C0 = dyn_cast<ConstantFP>(Val: Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Val: Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(Coefficient: 1, V: Opnd0);
368 else
369 Addend0.set(Coefficient: C0, V: nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(Coefficient: 1, V: Opnd1);
376 else
377 Addend.set(Coefficient: C1, V: 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(Coefficient: APFloat(C0->getValueAPF().getSemantics()), V: nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(i: 0);
392 Value *V1 = I->getOperand(i: 1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(Val: V0)) {
394 Addend0.set(Coefficient: C, V: V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(Val: V1)) {
399 Addend0.set(Coefficient: C, V: 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(ScaleAmt: Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(ScaleAmt: 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(Val: I, Addend0&: Opnd0, Addend1&: 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(Addend0&: Opnd0_0, Addend1&: 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(Addend0&: Opnd1_0, Addend1&: 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(Elt: &Opnd0_0);
460 AllOpnds.push_back(Elt: &Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(Elt: &Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(Elt: &Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(i: 0);
470 Value *V1 = I->getOperand(i: 1);
471 InstQuota = ((!isa<Constant>(Val: V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(Val: V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(V&: AllOpnds, InstrQuota: 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(Elt: &Opnd0);
491 AllOpnds.push_back(Elt: &Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(Elt: &Opnd1_1);
494
495 if (Value *R = simplifyFAdd(V&: AllOpnds, InstrQuota: 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(Elt: &Opnd1);
503 AllOpnds.push_back(Elt: &Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(Elt: &Opnd0_1);
506
507 if (Value *R = simplifyFAdd(V&: AllOpnds, InstrQuota: 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(Elt: 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(Elt: 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(N: StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(Elt: &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(Opnds: SimpVect, InstrQuota);
584 else {
585 // The addition is folded to 0.0.
586 Result = ConstantFP::get(Ty: Instr->getType(), V: 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(Vect: 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(A: *Opnd, NeedNeg);
619 if (!LastVal) {
620 LastVal = V;
621 LastValNeedNeg = NeedNeg;
622 continue;
623 }
624
625 if (LastValNeedNeg == NeedNeg) {
626 LastVal = createFAdd(Opnd0: LastVal, Opnd1: V);
627 continue;
628 }
629
630 if (LastValNeedNeg)
631 LastVal = createFSub(Opnd0: V, Opnd1: LastVal);
632 else
633 LastVal = createFSub(Opnd0: LastVal, Opnd1: V);
634
635 LastValNeedNeg = false;
636 }
637
638 if (LastValNeedNeg) {
639 LastVal = createFNeg(V: 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(L: Opnd0, R: Opnd1);
652 if (Instruction *I = dyn_cast<Instruction>(Val: V))
653 createInstPostProc(NewInst: I);
654 return V;
655}
656
657Value *FAddCombine::createFNeg(Value *V) {
658 Value *NewV = Builder.CreateFNeg(V);
659 if (Instruction *I = dyn_cast<Instruction>(Val: NewV))
660 createInstPostProc(NewInst: I, NoNumber: 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(L: Opnd0, R: Opnd1);
666 if (Instruction *I = dyn_cast<Instruction>(Val: V))
667 createInstPostProc(NewInst: I);
668 return V;
669}
670
671Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672 Value *V = Builder.CreateFMul(L: Opnd0, R: Opnd1);
673 if (Instruction *I = dyn_cast<Instruction>(Val: V))
674 createInstPostProc(NewInst: 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>(Val: 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(Ty: 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(Opnd0: OpndVal, Opnd1: OpndVal);
741 }
742
743 NeedNeg = false;
744 return createFMul(Opnd0: OpndVal, Opnd1: Coeff.getValue(Ty: 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
752static Value *checkForNegativeOperand(BinaryOperator &I,
753 InstCombiner::BuilderTy &Builder) {
754 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 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(V: RHS, P: m_Add(L: m_Value(V&: X), R: m_One())))
766 std::swap(a&: LHS, b&: RHS);
767
768 if (match(V: LHS, P: m_Add(L: m_Value(V&: X), R: m_One()))) {
769 // if XOR on other side, swap
770 if (match(V: RHS, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1))))
771 std::swap(a&: X, b&: RHS);
772
773 if (match(V: X, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: 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(V: Y, P: m_Or(L: m_Value(V&: Z), R: m_APInt(Res&: C2))) && (*C2 == ~(*C1))) {
777 Value *NewAnd = Builder.CreateAnd(LHS: Z, RHS: *C1);
778 return Builder.CreateSub(LHS: RHS, RHS: NewAnd, Name: "sub");
779 } else if (match(V: Y, P: m_And(L: m_Value(V&: Z), R: m_APInt(Res&: 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(LHS: Z, RHS: ~(*C1));
783 return Builder.CreateSub(LHS: RHS, RHS: NewOr, Name: "sub");
784 }
785 }
786 }
787
788 // Restore LHS and RHS
789 LHS = I.getOperand(i_nocapture: 0);
790 RHS = I.getOperand(i_nocapture: 1);
791
792 // if XOR is on other side, swap
793 if (match(V: RHS, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1))))
794 std::swap(a&: LHS, b&: 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(V: LHS, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1))))
800 if (C1->countr_zero() == 0)
801 if (match(V: Y, P: m_And(L: m_Value(V&: Z), R: m_APInt(Res&: C2))) && *C1 == (*C2 + 1)) {
802 Value *NewOr = Builder.CreateOr(LHS: Z, RHS: ~(*C2));
803 return Builder.CreateSub(LHS: RHS, RHS: NewOr, Name: "sub");
804 }
805 return nullptr;
806}
807
808/// Wrapping flags may allow combining constants separated by an extend.
809static Instruction *foldNoWrapAdd(BinaryOperator &Add,
810 InstCombiner::BuilderTy &Builder) {
811 Value *Op0 = Add.getOperand(i_nocapture: 0), *Op1 = Add.getOperand(i_nocapture: 1);
812 Type *Ty = Add.getType();
813 Constant *Op1C;
814 if (!match(V: Op1, P: m_Constant(C&: 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(V: Op1, P: m_APInt(Res&: C1)) &&
822 match(V: Op0, P: m_ZExt(Op: m_NUWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) &&
823 C1->isNegative() && C1->sge(RHS: -C2->sext(width: C1->getBitWidth()))) {
824 APInt NewC = *C2 + C1->trunc(width: C2->getBitWidth());
825 // If the smaller add will fold to zero, we don't need to check one use.
826 if (NewC.isZero())
827 return new ZExtInst(X, Ty);
828 // Otherwise only do this if the existing zero extend will be removed.
829 if (Op0->hasOneUse())
830 return new ZExtInst(
831 Builder.CreateNUWAdd(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: NewC)), Ty);
832 }
833
834 // More general combining of constants in the wide type.
835 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
836 // or (zext nneg (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
837 Constant *NarrowC;
838 if (match(V: Op0, P: m_OneUse(SubPattern: m_SExtLike(
839 Op: m_NSWAddLike(L: m_Value(V&: X), R: m_Constant(C&: NarrowC)))))) {
840 Value *WideC = Builder.CreateSExt(V: NarrowC, DestTy: Ty);
841 Value *NewC = Builder.CreateAdd(LHS: WideC, RHS: Op1C);
842 Value *WideX = Builder.CreateSExt(V: X, DestTy: Ty);
843 return BinaryOperator::CreateAdd(V1: WideX, V2: NewC);
844 }
845 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
846 if (match(V: Op0,
847 P: m_OneUse(SubPattern: m_ZExt(Op: m_NUWAddLike(L: m_Value(V&: X), R: m_Constant(C&: NarrowC)))))) {
848 Value *WideC = Builder.CreateZExt(V: NarrowC, DestTy: Ty);
849 Value *NewC = Builder.CreateAdd(LHS: WideC, RHS: Op1C);
850 Value *WideX = Builder.CreateZExt(V: X, DestTy: Ty);
851 return BinaryOperator::CreateAdd(V1: WideX, V2: NewC);
852 }
853 return nullptr;
854}
855
856Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
857 Value *Op0 = Add.getOperand(i_nocapture: 0), *Op1 = Add.getOperand(i_nocapture: 1);
858 Type *Ty = Add.getType();
859 Constant *Op1C;
860 if (!match(V: Op1, P: m_ImmConstant(C&: Op1C)))
861 return nullptr;
862
863 if (Instruction *NV = foldBinOpIntoSelectOrPhi(I&: Add))
864 return NV;
865
866 if (Instruction *FoldedLogic = foldBinOpSelectBinOp(Op&: Add))
867 return FoldedLogic;
868
869 Value *X;
870 Constant *Op00C;
871
872 // add (sub C1, X), C2 --> sub (add C1, C2), X
873 if (match(V: Op0, P: m_Sub(L: m_Constant(C&: Op00C), R: m_Value(V&: X))))
874 return BinaryOperator::CreateSub(V1: ConstantExpr::getAdd(C1: Op00C, C2: Op1C), V2: X);
875
876 Value *Y;
877
878 // add (sub X, Y), -1 --> add (not Y), X
879 if (match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y)))) &&
880 match(V: Op1, P: m_AllOnes()))
881 return BinaryOperator::CreateAdd(V1: Builder.CreateNot(V: Y), V2: X);
882
883 // zext(bool) + C -> bool ? C + 1 : C
884 if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) &&
885 X->getType()->getScalarSizeInBits() == 1)
886 return createSelectInstWithUnknownProfile(C: X, S1: InstCombiner::AddOne(C: Op1C),
887 S2: Op1);
888 // sext(bool) + C -> bool ? C - 1 : C
889 if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) &&
890 X->getType()->getScalarSizeInBits() == 1)
891 return createSelectInstWithUnknownProfile(C: X, S1: InstCombiner::SubOne(C: Op1C),
892 S2: Op1);
893
894 // ~X + C --> (C-1) - X
895 if (match(V: Op0, P: m_Not(V: m_Value(V&: X)))) {
896 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
897 auto *COne = ConstantInt::get(Ty: Op1C->getType(), V: 1);
898 bool WillNotSOV = willNotOverflowSignedSub(LHS: Op1C, RHS: COne, CxtI: Add);
899 BinaryOperator *Res =
900 BinaryOperator::CreateSub(V1: ConstantExpr::getSub(C1: Op1C, C2: COne), V2: X);
901 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
902 return Res;
903 }
904
905 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
906 const APInt *C;
907 unsigned BitWidth = Ty->getScalarSizeInBits();
908 if (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: X),
909 R: m_SpecificIntAllowPoison(V: BitWidth - 1)))) &&
910 match(V: Op1, P: m_One()))
911 return new ZExtInst(Builder.CreateIsNotNeg(Arg: X, Name: "isnotneg"), Ty);
912
913 if (!match(V: Op1, P: m_APInt(Res&: C)))
914 return nullptr;
915
916 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
917 Constant *Op01C;
918 if (match(V: Op0, P: m_DisjointOr(L: m_Value(V&: X), R: m_ImmConstant(C&: Op01C)))) {
919 BinaryOperator *NewAdd =
920 BinaryOperator::CreateAdd(V1: X, V2: ConstantExpr::getAdd(C1: Op01C, C2: Op1C));
921 NewAdd->setHasNoSignedWrap(Add.hasNoSignedWrap() &&
922 willNotOverflowSignedAdd(LHS: Op01C, RHS: Op1C, CxtI: Add));
923 NewAdd->setHasNoUnsignedWrap(Add.hasNoUnsignedWrap());
924 return NewAdd;
925 }
926
927 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
928 const APInt *C2;
929 if (match(V: Op0, P: m_Or(L: m_Value(), R: m_APInt(Res&: C2))) && *C2 == -*C)
930 return BinaryOperator::CreateXor(V1: Op0, V2: ConstantInt::get(Ty: Add.getType(), V: *C2));
931
932 if (C->isSignMask()) {
933 // If wrapping is not allowed, then the addition must set the sign bit:
934 // X + (signmask) --> X | signmask
935 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
936 return BinaryOperator::CreateOr(V1: Op0, V2: Op1);
937
938 // If wrapping is allowed, then the addition flips the sign bit of LHS:
939 // X + (signmask) --> X ^ signmask
940 return BinaryOperator::CreateXor(V1: Op0, V2: Op1);
941 }
942
943 // Is this add the last step in a convoluted sext?
944 // add(zext(xor i16 X, -32768), -32768) --> sext X
945 if (match(V: Op0, P: m_ZExt(Op: m_Xor(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) &&
946 C2->isMinSignedValue() && C2->sext(width: Ty->getScalarSizeInBits()) == *C)
947 return CastInst::Create(Instruction::SExt, S: X, Ty);
948
949 if (match(V: Op0, P: m_Xor(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) {
950 // (X ^ signmask) + C --> (X + (signmask ^ C))
951 if (C2->isSignMask())
952 return BinaryOperator::CreateAdd(V1: X, V2: ConstantInt::get(Ty, V: *C2 ^ *C));
953
954 // If X has no high-bits set above an xor mask:
955 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
956 if (C2->isMask()) {
957 KnownBits LHSKnown = computeKnownBits(V: X, CxtI: &Add);
958 if ((*C2 | LHSKnown.Zero).isAllOnes())
959 return BinaryOperator::CreateSub(V1: ConstantInt::get(Ty, V: *C2 + *C), V2: X);
960 }
961
962 // Look for a math+logic pattern that corresponds to sext-in-register of a
963 // value with cleared high bits. Convert that into a pair of shifts:
964 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
965 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
966 if (Op0->hasOneUse() && *C2 == -(*C)) {
967 unsigned BitWidth = Ty->getScalarSizeInBits();
968 unsigned ShAmt = 0;
969 if (C->isPowerOf2())
970 ShAmt = BitWidth - C->logBase2() - 1;
971 else if (C2->isPowerOf2())
972 ShAmt = BitWidth - C2->logBase2() - 1;
973 if (ShAmt &&
974 MaskedValueIsZero(V: X, Mask: APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: ShAmt), CxtI: &Add)) {
975 Constant *ShAmtC = ConstantInt::get(Ty, V: ShAmt);
976 Value *NewShl = Builder.CreateShl(LHS: X, RHS: ShAmtC, Name: "sext");
977 return BinaryOperator::CreateAShr(V1: NewShl, V2: ShAmtC);
978 }
979 }
980 }
981
982 if (C->isOne() && Op0->hasOneUse()) {
983 // add (sext i1 X), 1 --> zext (not X)
984 // TODO: The smallest IR representation is (select X, 0, 1), and that would
985 // not require the one-use check. But we need to remove a transform in
986 // visitSelect and make sure that IR value tracking for select is equal or
987 // better than for these ops.
988 if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) &&
989 X->getType()->getScalarSizeInBits() == 1)
990 return new ZExtInst(Builder.CreateNot(V: X), Ty);
991
992 // Shifts and add used to flip and mask off the low bit:
993 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
994 const APInt *C3;
995 if (match(V: Op0, P: m_AShr(L: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C2)), R: m_APInt(Res&: C3))) &&
996 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
997 Value *NotX = Builder.CreateNot(V: X);
998 return BinaryOperator::CreateAnd(V1: NotX, V2: ConstantInt::get(Ty, V: 1));
999 }
1000 }
1001
1002 // umax(X, C) + -C --> usub.sat(X, C)
1003 if (match(V: Op0, P: m_OneUse(SubPattern: m_UMax(L: m_Value(V&: X), R: m_SpecificInt(V: -*C)))))
1004 return replaceInstUsesWith(
1005 I&: Add, V: Builder.CreateBinaryIntrinsic(
1006 ID: Intrinsic::usub_sat, LHS: X, RHS: ConstantInt::get(Ty: Add.getType(), V: -*C)));
1007
1008 // Fold (add (zext (add X, -1)), 1) -> (zext X) if X is non-zero.
1009 // TODO: There's a general form for any constant on the outer add.
1010 if (C->isOne()) {
1011 if (match(V: Op0, P: m_ZExt(Op: m_Add(L: m_Value(V&: X), R: m_AllOnes())))) {
1012 const SimplifyQuery Q = SQ.getWithInstruction(I: &Add);
1013 if (llvm::isKnownNonZero(V: X, Q))
1014 return new ZExtInst(X, Ty);
1015 }
1016 }
1017
1018 return nullptr;
1019}
1020
1021// match variations of a^2 + 2*a*b + b^2
1022//
1023// to reuse the code between the FP and Int versions, the instruction OpCodes
1024// and constant types have been turned into template parameters.
1025//
1026// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1027// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1028// (we're matching `X<<1` instead of `X*2` for Int)
1029template <bool FP, typename Mul2Rhs>
1030static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1031 Value *&B) {
1032 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1033 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1034 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1035
1036 // (a * a) + (((a * 2) + b) * b)
1037 if (match(&I, m_c_BinOp(
1038 AddOp, m_OneUse(SubPattern: m_BinOp(Opcode: MulOp, L: m_Value(V&: A), R: m_Deferred(V: A))),
1039 m_OneUse(m_c_BinOp(
1040 MulOp,
1041 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(V: A), M2Rhs),
1042 m_Value(V&: B)),
1043 m_Deferred(V: B))))))
1044 return true;
1045
1046 // ((a * b) * 2) or ((a * 2) * b)
1047 // +
1048 // (a * a + b * b) or (b * b + a * a)
1049 return match(
1050 &I, m_c_BinOp(
1051 AddOp,
1052 m_CombineOr(
1053 m_OneUse(m_BinOp(
1054 Mul2Op, m_BinOp(Opcode: MulOp, L: m_Value(V&: A), R: m_Value(V&: B)), M2Rhs)),
1055 m_OneUse(m_c_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(V&: A), M2Rhs),
1056 m_Value(V&: B)))),
1057 m_OneUse(
1058 SubPattern: m_c_BinOp(Opcode: AddOp, L: m_BinOp(Opcode: MulOp, L: m_Deferred(V: A), R: m_Deferred(V: A)),
1059 R: m_BinOp(Opcode: MulOp, L: m_Deferred(V: B), R: m_Deferred(V: B))))));
1060}
1061
1062// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1063Instruction *InstCombinerImpl::foldSquareSumInt(BinaryOperator &I) {
1064 Value *A, *B;
1065 if (matchesSquareSum</*FP*/ false>(I, M2Rhs: m_SpecificInt(V: 1), A, B)) {
1066 Value *AB = Builder.CreateAdd(LHS: A, RHS: B);
1067 return BinaryOperator::CreateMul(V1: AB, V2: AB);
1068 }
1069 return nullptr;
1070}
1071
1072// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1073// Requires `nsz` and `reassoc`.
1074Instruction *InstCombinerImpl::foldSquareSumFP(BinaryOperator &I) {
1075 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1076 Value *A, *B;
1077 if (matchesSquareSum</*FP*/ true>(I, M2Rhs: m_SpecificFP(V: 2.0), A, B)) {
1078 Value *AB = Builder.CreateFAddFMF(L: A, R: B, FMFSource: &I);
1079 return BinaryOperator::CreateFMulFMF(V1: AB, V2: AB, FMFSource: &I);
1080 }
1081 return nullptr;
1082}
1083
1084// Matches multiplication expression Op * C where C is a constant. Returns the
1085// constant value in C and the other operand in Op. Returns true if such a
1086// match is found.
1087static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1088 const APInt *AI;
1089 if (match(V: E, P: m_Mul(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1090 C = *AI;
1091 return true;
1092 }
1093 if (match(V: E, P: m_Shl(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1094 C = APInt(AI->getBitWidth(), 1);
1095 C <<= *AI;
1096 return true;
1097 }
1098 return false;
1099}
1100
1101// Matches remainder expression Op % C where C is a constant. Returns the
1102// constant value in C and the other operand in Op. Returns the signedness of
1103// the remainder operation in IsSigned. Returns true if such a match is
1104// found.
1105static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1106 const APInt *AI;
1107 IsSigned = false;
1108 if (match(V: E, P: m_SRem(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1109 IsSigned = true;
1110 C = *AI;
1111 return true;
1112 }
1113 if (match(V: E, P: m_URem(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1114 C = *AI;
1115 return true;
1116 }
1117 if (match(V: E, P: m_And(L: m_Value(V&: Op), R: m_APInt(Res&: AI))) && (*AI + 1).isPowerOf2()) {
1118 C = *AI + 1;
1119 return true;
1120 }
1121 return false;
1122}
1123
1124// Matches division expression Op / C with the given signedness as indicated
1125// by IsSigned, where C is a constant. Returns the constant value in C and the
1126// other operand in Op. Returns true if such a match is found.
1127static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1128 const APInt *AI;
1129 if (IsSigned && match(V: E, P: m_SDiv(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1130 C = *AI;
1131 return true;
1132 }
1133 if (!IsSigned) {
1134 if (match(V: E, P: m_UDiv(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1135 C = *AI;
1136 return true;
1137 }
1138 if (match(V: E, P: m_LShr(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1139 C = APInt(AI->getBitWidth(), 1);
1140 C <<= *AI;
1141 return true;
1142 }
1143 }
1144 return false;
1145}
1146
1147// Returns whether C0 * C1 with the given signedness overflows.
1148static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1149 bool overflow;
1150 if (IsSigned)
1151 (void)C0.smul_ov(RHS: C1, Overflow&: overflow);
1152 else
1153 (void)C0.umul_ov(RHS: C1, Overflow&: overflow);
1154 return overflow;
1155}
1156
1157// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1158// does not overflow.
1159// Simplifies (X / C0) * C1 + (X % C0) * C2 to
1160// (X / C0) * (C1 - C2 * C0) + X * C2
1161Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1162 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
1163 Value *X, *MulOpV;
1164 APInt C0, MulOpC;
1165 bool IsSigned;
1166 // Match I = X % C0 + MulOpV * C0
1167 if (((MatchRem(E: LHS, Op&: X, C&: C0, IsSigned) && MatchMul(E: RHS, Op&: MulOpV, C&: MulOpC)) ||
1168 (MatchRem(E: RHS, Op&: X, C&: C0, IsSigned) && MatchMul(E: LHS, Op&: MulOpV, C&: MulOpC))) &&
1169 C0 == MulOpC) {
1170 Value *RemOpV;
1171 APInt C1;
1172 bool Rem2IsSigned;
1173 // Match MulOpC = RemOpV % C1
1174 if (MatchRem(E: MulOpV, Op&: RemOpV, C&: C1, IsSigned&: Rem2IsSigned) &&
1175 IsSigned == Rem2IsSigned) {
1176 Value *DivOpV;
1177 APInt DivOpC;
1178 // Match RemOpV = X / C0
1179 if (MatchDiv(E: RemOpV, Op&: DivOpV, C&: DivOpC, IsSigned) && X == DivOpV &&
1180 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1181 Value *NewDivisor = ConstantInt::get(Ty: X->getType(), V: C0 * C1);
1182 return IsSigned ? Builder.CreateSRem(LHS: X, RHS: NewDivisor, Name: "srem")
1183 : Builder.CreateURem(LHS: X, RHS: NewDivisor, Name: "urem");
1184 }
1185 }
1186 }
1187
1188 // Match I = (X / C0) * C1 + (X % C0) * C2
1189 Value *Div, *Rem;
1190 APInt C1, C2;
1191 if (!LHS->hasOneUse() || !MatchMul(E: LHS, Op&: Div, C&: C1))
1192 Div = LHS, C1 = APInt(I.getType()->getScalarSizeInBits(), 1);
1193 if (!RHS->hasOneUse() || !MatchMul(E: RHS, Op&: Rem, C&: C2))
1194 Rem = RHS, C2 = APInt(I.getType()->getScalarSizeInBits(), 1);
1195 if (match(V: Div, P: m_IRem(L: m_Value(), R: m_Value()))) {
1196 std::swap(a&: Div, b&: Rem);
1197 std::swap(a&: C1, b&: C2);
1198 }
1199 Value *DivOpV;
1200 APInt DivOpC;
1201 if (MatchRem(E: Rem, Op&: X, C&: C0, IsSigned) &&
1202 MatchDiv(E: Div, Op&: DivOpV, C&: DivOpC, IsSigned) && X == DivOpV && C0 == DivOpC &&
1203 // Avoid unprofitable replacement of and with mul.
1204 !(C1.isOne() && !IsSigned && DivOpC.isPowerOf2() && DivOpC != 2)) {
1205 APInt NewC = C1 - C2 * C0;
1206 if (!NewC.isZero() && !Rem->hasOneUse())
1207 return nullptr;
1208 if (!isGuaranteedNotToBeUndef(V: X, AC: &AC, CtxI: &I, DT: &DT))
1209 return nullptr;
1210 Value *MulXC2 = Builder.CreateMul(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: C2));
1211 if (NewC.isZero())
1212 return MulXC2;
1213 return Builder.CreateAdd(
1214 LHS: Builder.CreateMul(LHS: Div, RHS: ConstantInt::get(Ty: X->getType(), V: NewC)), RHS: MulXC2);
1215 }
1216
1217 return nullptr;
1218}
1219
1220/// Fold
1221/// (1 << NBits) - 1
1222/// Into:
1223/// ~(-(1 << NBits))
1224/// Because a 'not' is better for bit-tracking analysis and other transforms
1225/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1226static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1227 InstCombiner::BuilderTy &Builder) {
1228 Value *NBits;
1229 if (!match(V: &I, P: m_Add(L: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: NBits))), R: m_AllOnes())))
1230 return nullptr;
1231
1232 Constant *MinusOne = Constant::getAllOnesValue(Ty: NBits->getType());
1233 Value *NotMask = Builder.CreateShl(LHS: MinusOne, RHS: NBits, Name: "notmask");
1234 // Be wary of constant folding.
1235 if (auto *BOp = dyn_cast<BinaryOperator>(Val: NotMask)) {
1236 // Always NSW. But NUW propagates from `add`.
1237 BOp->setHasNoSignedWrap();
1238 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1239 }
1240
1241 return BinaryOperator::CreateNot(Op: NotMask, Name: I.getName());
1242}
1243
1244static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1245 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1246 Type *Ty = I.getType();
1247 auto getUAddSat = [&]() {
1248 return Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: Intrinsic::uadd_sat,
1249 Tys: Ty);
1250 };
1251
1252 // add (umin X, ~Y), Y --> uaddsat X, Y
1253 Value *X, *Y;
1254 if (match(V: &I, P: m_c_Add(L: m_c_UMin(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y))),
1255 R: m_Deferred(V: Y))))
1256 return CallInst::Create(Func: getUAddSat(), Args: { X, Y });
1257
1258 // add (umin X, ~C), C --> uaddsat X, C
1259 const APInt *C, *NotC;
1260 if (match(V: &I, P: m_Add(L: m_UMin(L: m_Value(V&: X), R: m_APInt(Res&: NotC)), R: m_APInt(Res&: C))) &&
1261 *C == ~*NotC)
1262 return CallInst::Create(Func: getUAddSat(), Args: { X, ConstantInt::get(Ty, V: *C) });
1263
1264 return nullptr;
1265}
1266
1267// Transform:
1268// (add A, (shl (neg B), Y))
1269// -> (sub A, (shl B, Y))
1270static Instruction *combineAddSubWithShlAddSub(InstCombiner::BuilderTy &Builder,
1271 const BinaryOperator &I) {
1272 Value *A, *B, *Cnt;
1273 if (match(V: &I,
1274 P: m_c_Add(L: m_OneUse(SubPattern: m_Shl(L: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: B))), R: m_Value(V&: Cnt))),
1275 R: m_Value(V&: A)))) {
1276 Value *NewShl = Builder.CreateShl(LHS: B, RHS: Cnt);
1277 return BinaryOperator::CreateSub(V1: A, V2: NewShl);
1278 }
1279 return nullptr;
1280}
1281
1282/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1283static Instruction *foldAddToAshr(BinaryOperator &Add) {
1284 // Division must be by power-of-2, but not the minimum signed value.
1285 Value *X;
1286 const APInt *DivC;
1287 if (!match(V: Add.getOperand(i_nocapture: 0), P: m_SDiv(L: m_Value(V&: X), R: m_Power2(V&: DivC))) ||
1288 DivC->isNegative())
1289 return nullptr;
1290
1291 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1292 // low bits set. It recognizes two canonical patterns:
1293 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1294 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1295 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1296 // Note that, by the time we end up here, if possible, ugt has been
1297 // canonicalized into eq.
1298 const APInt *MaskC, *MaskCCmp;
1299 CmpPredicate Pred;
1300 if (!match(V: Add.getOperand(i_nocapture: 1),
1301 P: m_SExt(Op: m_ICmp(Pred, L: m_And(L: m_Specific(V: X), R: m_APInt(Res&: MaskC)),
1302 R: m_APInt(Res&: MaskCCmp)))))
1303 return nullptr;
1304
1305 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1306 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1307 return nullptr;
1308
1309 APInt SMin = APInt::getSignedMinValue(numBits: Add.getType()->getScalarSizeInBits());
1310 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1311 ? (*MaskC == (SMin | (*DivC - 1)))
1312 : (*DivC == 2 && *MaskC == SMin + 1);
1313 if (!IsMaskValid)
1314 return nullptr;
1315
1316 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1317 return BinaryOperator::CreateAShr(
1318 V1: X, V2: ConstantInt::get(Ty: Add.getType(), V: DivC->exactLogBase2()));
1319}
1320
1321Instruction *InstCombinerImpl::foldAddLikeCommutative(Value *LHS, Value *RHS,
1322 bool NSW, bool NUW) {
1323 Value *A, *B, *C;
1324 if (match(V: LHS, P: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B))) &&
1325 match(V: RHS, P: m_Sub(L: m_Value(V&: C), R: m_Specific(V: A)))) {
1326 Instruction *R = BinaryOperator::CreateSub(V1: C, V2: B);
1327 bool NSWOut = NSW && match(V: LHS, P: m_NSWSub(L: m_Value(), R: m_Value())) &&
1328 match(V: RHS, P: m_NSWSub(L: m_Value(), R: m_Value()));
1329
1330 bool NUWOut = match(V: LHS, P: m_NUWSub(L: m_Value(), R: m_Value())) &&
1331 match(V: RHS, P: m_NUWSub(L: m_Value(), R: m_Value()));
1332 R->setHasNoSignedWrap(NSWOut);
1333 R->setHasNoUnsignedWrap(NUWOut);
1334 return R;
1335 }
1336
1337 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1338 const APInt *C1, *C2;
1339 if (match(V: LHS, P: m_Shl(L: m_SDiv(L: m_Specific(V: RHS), R: m_APInt(Res&: C1)), R: m_APInt(Res&: C2)))) {
1340 APInt One(C2->getBitWidth(), 1);
1341 APInt MinusC1 = -(*C1);
1342 if (MinusC1 == (One << *C2)) {
1343 Constant *NewRHS = ConstantInt::get(Ty: RHS->getType(), V: MinusC1);
1344 return BinaryOperator::CreateSRem(V1: RHS, V2: NewRHS);
1345 }
1346 }
1347
1348 return nullptr;
1349}
1350
1351Instruction *InstCombinerImpl::
1352 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1353 BinaryOperator &I) {
1354 assert((I.getOpcode() == Instruction::Add ||
1355 I.getOpcode() == Instruction::Or ||
1356 I.getOpcode() == Instruction::Sub) &&
1357 "Expecting add/or/sub instruction");
1358
1359 // We have a subtraction/addition between a (potentially truncated) *logical*
1360 // right-shift of X and a "select".
1361 Value *X, *Select;
1362 Instruction *LowBitsToSkip, *Extract;
1363 if (!match(V: &I, P: m_c_BinOp(L: m_TruncOrSelf(Op: m_Instruction(
1364 I&: Extract, Match: m_LShr(L: m_Value(V&: X),
1365 R: m_Instruction(I&: LowBitsToSkip)))),
1366 R: m_Value(V&: Select))))
1367 return nullptr;
1368
1369 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1370 if (I.getOpcode() == Instruction::Sub && I.getOperand(i_nocapture: 1) != Select)
1371 return nullptr;
1372
1373 Type *XTy = X->getType();
1374 bool HadTrunc = I.getType() != XTy;
1375
1376 // If there was a truncation of extracted value, then we'll need to produce
1377 // one extra instruction, so we need to ensure one instruction will go away.
1378 if (HadTrunc && !match(V: &I, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value())))
1379 return nullptr;
1380
1381 // Extraction should extract high NBits bits, with shift amount calculated as:
1382 // low bits to skip = shift bitwidth - high bits to extract
1383 // The shift amount itself may be extended, and we need to look past zero-ext
1384 // when matching NBits, that will matter for matching later.
1385 Value *NBits;
1386 if (!match(V: LowBitsToSkip,
1387 P: m_ZExtOrSelf(Op: m_Sub(L: m_SpecificInt(V: XTy->getScalarSizeInBits()),
1388 R: m_ZExtOrSelf(Op: m_Value(V&: NBits))))))
1389 return nullptr;
1390
1391 // Sign-extending value can be zero-extended if we `sub`tract it,
1392 // or sign-extended otherwise.
1393 auto SkipExtInMagic = [&I](Value *&V) {
1394 if (I.getOpcode() == Instruction::Sub)
1395 match(V, P: m_ZExtOrSelf(Op: m_Value(V)));
1396 else
1397 match(V, P: m_SExtOrSelf(Op: m_Value(V)));
1398 };
1399
1400 // Now, finally validate the sign-extending magic.
1401 // `select` itself may be appropriately extended, look past that.
1402 SkipExtInMagic(Select);
1403
1404 CmpPredicate Pred;
1405 const APInt *Thr;
1406 Value *SignExtendingValue, *Zero;
1407 bool ShouldSignext;
1408 // It must be a select between two values we will later establish to be a
1409 // sign-extending value and a zero constant. The condition guarding the
1410 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1411 if (!match(V: Select, P: m_Select(C: m_ICmp(Pred, L: m_Specific(V: X), R: m_APInt(Res&: Thr)),
1412 L: m_Value(V&: SignExtendingValue), R: m_Value(V&: Zero))) ||
1413 !isSignBitCheck(Pred, RHS: *Thr, TrueIfSigned&: ShouldSignext))
1414 return nullptr;
1415
1416 // icmp-select pair is commutative.
1417 if (!ShouldSignext)
1418 std::swap(a&: SignExtendingValue, b&: Zero);
1419
1420 // If we should not perform sign-extension then we must add/or/subtract zero.
1421 if (!match(V: Zero, P: m_Zero()))
1422 return nullptr;
1423 // Otherwise, it should be some constant, left-shifted by the same NBits we
1424 // had in `lshr`. Said left-shift can also be appropriately extended.
1425 // Again, we must look past zero-ext when looking for NBits.
1426 SkipExtInMagic(SignExtendingValue);
1427 Constant *SignExtendingValueBaseConstant;
1428 if (!match(V: SignExtendingValue,
1429 P: m_Shl(L: m_Constant(C&: SignExtendingValueBaseConstant),
1430 R: m_ZExtOrSelf(Op: m_Specific(V: NBits)))))
1431 return nullptr;
1432 // If we `sub`, then the constant should be one, else it should be all-ones.
1433 if (I.getOpcode() == Instruction::Sub
1434 ? !match(V: SignExtendingValueBaseConstant, P: m_One())
1435 : !match(V: SignExtendingValueBaseConstant, P: m_AllOnes()))
1436 return nullptr;
1437
1438 auto *NewAShr = BinaryOperator::CreateAShr(V1: X, V2: LowBitsToSkip,
1439 Name: Extract->getName() + ".sext");
1440 NewAShr->copyIRFlags(V: Extract); // Preserve `exact`-ness.
1441 if (!HadTrunc)
1442 return NewAShr;
1443
1444 Builder.Insert(I: NewAShr);
1445 return TruncInst::CreateTruncOrBitCast(S: NewAShr, Ty: I.getType());
1446}
1447
1448/// This is a specialization of a more general transform from
1449/// foldUsingDistributiveLaws. If that code can be made to work optimally
1450/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1451static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1452 InstCombiner::BuilderTy &Builder) {
1453 // TODO: Also handle mul by doubling the shift amount?
1454 assert((I.getOpcode() == Instruction::Add ||
1455 I.getOpcode() == Instruction::Sub) &&
1456 "Expected add/sub");
1457 auto *Op0 = dyn_cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 0));
1458 auto *Op1 = dyn_cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 1));
1459 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1460 return nullptr;
1461
1462 Value *X, *Y, *ShAmt;
1463 if (!match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: ShAmt))) ||
1464 !match(V: Op1, P: m_Shl(L: m_Value(V&: Y), R: m_Specific(V: ShAmt))))
1465 return nullptr;
1466
1467 // No-wrap propagates only when all ops have no-wrap.
1468 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1469 Op1->hasNoSignedWrap();
1470 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1471 Op1->hasNoUnsignedWrap();
1472
1473 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1474 Value *NewMath = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: X, RHS: Y);
1475 if (auto *NewI = dyn_cast<BinaryOperator>(Val: NewMath)) {
1476 NewI->setHasNoSignedWrap(HasNSW);
1477 NewI->setHasNoUnsignedWrap(HasNUW);
1478 }
1479 auto *NewShl = BinaryOperator::CreateShl(V1: NewMath, V2: ShAmt);
1480 NewShl->setHasNoSignedWrap(HasNSW);
1481 NewShl->setHasNoUnsignedWrap(HasNUW);
1482 return NewShl;
1483}
1484
1485/// Reduce a sequence of masked half-width multiplies to a single multiply.
1486/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1487static Instruction *foldBoxMultiply(BinaryOperator &I) {
1488 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1489 // Skip the odd bitwidth types.
1490 if ((BitWidth & 0x1))
1491 return nullptr;
1492
1493 unsigned HalfBits = BitWidth >> 1;
1494 APInt HalfMask = APInt::getMaxValue(numBits: HalfBits);
1495
1496 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1497 Value *XLo, *YLo;
1498 Value *CrossSum;
1499 // Require one-use on the multiply to avoid increasing the number of
1500 // multiplications.
1501 if (!match(V: &I, P: m_c_Add(L: m_Shl(L: m_Value(V&: CrossSum), R: m_SpecificInt(V: HalfBits)),
1502 R: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: YLo), R: m_Value(V&: XLo))))))
1503 return nullptr;
1504
1505 // XLo = X & HalfMask
1506 // YLo = Y & HalfMask
1507 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1508 // to enhance robustness
1509 Value *X, *Y;
1510 if (!match(V: XLo, P: m_And(L: m_Value(V&: X), R: m_SpecificInt(V: HalfMask))) ||
1511 !match(V: YLo, P: m_And(L: m_Value(V&: Y), R: m_SpecificInt(V: HalfMask))))
1512 return nullptr;
1513
1514 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1515 // X' can be either X or XLo in the pattern (and the same for Y')
1516 if (match(V: CrossSum,
1517 P: m_c_Add(L: m_c_Mul(L: m_LShr(L: m_Specific(V: Y), R: m_SpecificInt(V: HalfBits)),
1518 R: m_CombineOr(L: m_Specific(V: X), R: m_Specific(V: XLo))),
1519 R: m_c_Mul(L: m_LShr(L: m_Specific(V: X), R: m_SpecificInt(V: HalfBits)),
1520 R: m_CombineOr(L: m_Specific(V: Y), R: m_Specific(V: YLo))))))
1521 return BinaryOperator::CreateMul(V1: X, V2: Y);
1522
1523 return nullptr;
1524}
1525
1526Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1527 if (Value *V = simplifyAddInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
1528 IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(),
1529 Q: SQ.getWithInstruction(I: &I)))
1530 return replaceInstUsesWith(I, V);
1531
1532 if (SimplifyAssociativeOrCommutative(I))
1533 return &I;
1534
1535 if (Instruction *X = foldVectorBinop(Inst&: I))
1536 return X;
1537
1538 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
1539 return Phi;
1540
1541 // (A*B)+(A*C) -> A*(B+C) etc
1542 if (Value *V = foldUsingDistributiveLaws(I))
1543 return replaceInstUsesWith(I, V);
1544
1545 if (Instruction *R = foldBoxMultiply(I))
1546 return R;
1547
1548 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1549 return R;
1550
1551 if (Instruction *X = foldAddWithConstant(Add&: I))
1552 return X;
1553
1554 if (Instruction *X = foldNoWrapAdd(Add&: I, Builder))
1555 return X;
1556
1557 if (Instruction *R = foldBinOpShiftWithShift(I))
1558 return R;
1559
1560 if (Instruction *R = combineAddSubWithShlAddSub(Builder, I))
1561 return R;
1562
1563 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
1564 if (Instruction *R = foldAddLikeCommutative(LHS, RHS, NSW: I.hasNoSignedWrap(),
1565 NUW: I.hasNoUnsignedWrap()))
1566 return R;
1567 if (Instruction *R = foldAddLikeCommutative(LHS: RHS, RHS: LHS, NSW: I.hasNoSignedWrap(),
1568 NUW: I.hasNoUnsignedWrap()))
1569 return R;
1570 Type *Ty = I.getType();
1571 if (Ty->isIntOrIntVectorTy(BitWidth: 1))
1572 return BinaryOperator::CreateXor(V1: LHS, V2: RHS);
1573
1574 // X + X --> X << 1
1575 if (LHS == RHS) {
1576 auto *Shl = BinaryOperator::CreateShl(V1: LHS, V2: ConstantInt::get(Ty, V: 1));
1577 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1578 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1579 return Shl;
1580 }
1581
1582 Value *A, *B;
1583 if (match(V: LHS, P: m_Neg(V: m_Value(V&: A)))) {
1584 // -A + -B --> -(A + B)
1585 if (match(V: RHS, P: m_Neg(V: m_Value(V&: B))))
1586 return BinaryOperator::CreateNeg(Op: Builder.CreateAdd(LHS: A, RHS: B));
1587
1588 // -A + B --> B - A
1589 auto *Sub = BinaryOperator::CreateSub(V1: RHS, V2: A);
1590 auto *OB0 = cast<OverflowingBinaryOperator>(Val: LHS);
1591 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1592
1593 return Sub;
1594 }
1595
1596 // A + -B --> A - B
1597 if (match(V: RHS, P: m_Neg(V: m_Value(V&: B)))) {
1598 auto *Sub = BinaryOperator::CreateSub(V1: LHS, V2: B);
1599 auto *OBO = cast<OverflowingBinaryOperator>(Val: RHS);
1600 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO->hasNoSignedWrap());
1601 return Sub;
1602 }
1603
1604 if (Value *V = checkForNegativeOperand(I, Builder))
1605 return replaceInstUsesWith(I, V);
1606
1607 // (A + 1) + ~B --> A - B
1608 // ~B + (A + 1) --> A - B
1609 // (~B + A) + 1 --> A - B
1610 // (A + ~B) + 1 --> A - B
1611 if (match(V: &I, P: m_c_BinOp(L: m_Add(L: m_Value(V&: A), R: m_One()), R: m_Not(V: m_Value(V&: B)))) ||
1612 match(V: &I, P: m_BinOp(L: m_c_Add(L: m_Not(V: m_Value(V&: B)), R: m_Value(V&: A)), R: m_One())))
1613 return BinaryOperator::CreateSub(V1: A, V2: B);
1614
1615 // (A + RHS) + RHS --> A + (RHS << 1)
1616 if (match(V: LHS, P: m_OneUse(SubPattern: m_c_Add(L: m_Value(V&: A), R: m_Specific(V: RHS)))))
1617 return BinaryOperator::CreateAdd(V1: A, V2: Builder.CreateShl(LHS: RHS, RHS: 1, Name: "reass.add"));
1618
1619 // LHS + (A + LHS) --> A + (LHS << 1)
1620 if (match(V: RHS, P: m_OneUse(SubPattern: m_c_Add(L: m_Value(V&: A), R: m_Specific(V: LHS)))))
1621 return BinaryOperator::CreateAdd(V1: A, V2: Builder.CreateShl(LHS, RHS: 1, Name: "reass.add"));
1622
1623 {
1624 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1625 Constant *C1, *C2;
1626 if (match(V: &I, P: m_c_Add(L: m_Add(L: m_Value(V&: A), R: m_ImmConstant(C&: C1)),
1627 R: m_Sub(L: m_ImmConstant(C&: C2), R: m_Value(V&: B)))) &&
1628 (LHS->hasOneUse() || RHS->hasOneUse())) {
1629 Value *Sub = Builder.CreateSub(LHS: A, RHS: B);
1630 return BinaryOperator::CreateAdd(V1: Sub, V2: ConstantExpr::getAdd(C1, C2));
1631 }
1632
1633 // Canonicalize a constant sub operand as an add operand for better folding:
1634 // (C1 - A) + B --> (B - A) + C1
1635 if (match(V: &I, P: m_c_Add(L: m_OneUse(SubPattern: m_Sub(L: m_ImmConstant(C&: C1), R: m_Value(V&: A))),
1636 R: m_Value(V&: B)))) {
1637 Value *Sub = Builder.CreateSub(LHS: B, RHS: A, Name: "reass.sub");
1638 return BinaryOperator::CreateAdd(V1: Sub, V2: C1);
1639 }
1640 }
1641
1642 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1643 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1644
1645 const APInt *C1;
1646 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1647 if (match(V: &I, P: m_c_Add(L: m_And(L: m_Value(V&: A), R: m_APInt(Res&: C1)), R: m_Deferred(V: A))) &&
1648 C1->isPowerOf2() && (ComputeNumSignBits(Op: A) > C1->countl_zero())) {
1649 Constant *NewMask = ConstantInt::get(Ty: RHS->getType(), V: *C1 - 1);
1650 return BinaryOperator::CreateAnd(V1: A, V2: NewMask);
1651 }
1652
1653 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1654 if ((match(V: RHS, P: m_ZExt(Op: m_Value(V&: A))) &&
1655 match(V: LHS, P: m_ZExt(Op: m_NUWSub(L: m_Value(V&: B), R: m_Specific(V: A))))) ||
1656 (match(V: LHS, P: m_ZExt(Op: m_Value(V&: A))) &&
1657 match(V: RHS, P: m_ZExt(Op: m_NUWSub(L: m_Value(V&: B), R: m_Specific(V: A))))))
1658 return new ZExtInst(B, LHS->getType());
1659
1660 // zext(A) + sext(A) --> 0 if A is i1
1661 if (match(V: &I, P: m_c_BinOp(L: m_ZExt(Op: m_Value(V&: A)), R: m_SExt(Op: m_Deferred(V: A)))) &&
1662 A->getType()->isIntOrIntVectorTy(BitWidth: 1))
1663 return replaceInstUsesWith(I, V: Constant::getNullValue(Ty: I.getType()));
1664
1665 // sext(A < B) + zext(A > B) => ucmp/scmp(A, B)
1666 CmpPredicate LTPred, GTPred;
1667 if (match(V: &I,
1668 P: m_c_Add(L: m_SExt(Op: m_c_ICmp(Pred&: LTPred, L: m_Value(V&: A), R: m_Value(V&: B))),
1669 R: m_ZExt(Op: m_c_ICmp(Pred&: GTPred, L: m_Deferred(V: A), R: m_Deferred(V: B))))) &&
1670 A->getType()->isIntOrIntVectorTy()) {
1671 if (ICmpInst::isGT(P: LTPred)) {
1672 std::swap(a&: LTPred, b&: GTPred);
1673 std::swap(a&: A, b&: B);
1674 }
1675
1676 if (ICmpInst::isLT(P: LTPred) && ICmpInst::isGT(P: GTPred) &&
1677 ICmpInst::isSigned(predicate: LTPred) == ICmpInst::isSigned(predicate: GTPred))
1678 return replaceInstUsesWith(
1679 I, V: Builder.CreateIntrinsic(
1680 RetTy: Ty,
1681 ID: ICmpInst::isSigned(predicate: LTPred) ? Intrinsic::scmp : Intrinsic::ucmp,
1682 Args: {A, B}));
1683 }
1684
1685 // A+B --> A|B iff A and B have no bits set in common.
1686 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1687 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ: SQ.getWithInstruction(I: &I)))
1688 return BinaryOperator::CreateDisjointOr(V1: LHS, V2: RHS);
1689
1690 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1691 return Ext;
1692
1693 // (add (xor A, B) (and A, B)) --> (or A, B)
1694 // (add (and A, B) (xor A, B)) --> (or A, B)
1695 if (match(V: &I, P: m_c_BinOp(L: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)),
1696 R: m_c_And(L: m_Deferred(V: A), R: m_Deferred(V: B)))))
1697 return BinaryOperator::CreateOr(V1: A, V2: B);
1698
1699 // (add (or A, B) (and A, B)) --> (add A, B)
1700 // (add (and A, B) (or A, B)) --> (add A, B)
1701 if (match(V: &I, P: m_c_BinOp(L: m_Or(L: m_Value(V&: A), R: m_Value(V&: B)),
1702 R: m_c_And(L: m_Deferred(V: A), R: m_Deferred(V: B))))) {
1703 // Replacing operands in-place to preserve nuw/nsw flags.
1704 replaceOperand(I, OpNum: 0, V: A);
1705 replaceOperand(I, OpNum: 1, V: B);
1706 return &I;
1707 }
1708
1709 // (add A (or A, -A)) --> (and (add A, -1) A)
1710 // (add A (or -A, A)) --> (and (add A, -1) A)
1711 // (add (or A, -A) A) --> (and (add A, -1) A)
1712 // (add (or -A, A) A) --> (and (add A, -1) A)
1713 if (match(V: &I, P: m_c_BinOp(L: m_Value(V&: A), R: m_OneUse(SubPattern: m_c_Or(L: m_Neg(V: m_Deferred(V: A)),
1714 R: m_Deferred(V: A)))))) {
1715 Value *Add =
1716 Builder.CreateAdd(LHS: A, RHS: Constant::getAllOnesValue(Ty: A->getType()), Name: "",
1717 HasNUW: I.hasNoUnsignedWrap(), HasNSW: I.hasNoSignedWrap());
1718 return BinaryOperator::CreateAnd(V1: Add, V2: A);
1719 }
1720
1721 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1722 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1723 if (match(V: &I,
1724 P: m_Add(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_OneUse(SubPattern: m_Neg(V: m_Deferred(V: A))))),
1725 R: m_AllOnes()))) {
1726 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty: RHS->getType());
1727 Value *Dec = Builder.CreateAdd(LHS: A, RHS: AllOnes);
1728 Value *Not = Builder.CreateXor(LHS: A, RHS: AllOnes);
1729 return BinaryOperator::CreateAnd(V1: Dec, V2: Not);
1730 }
1731
1732 // Disguised reassociation/factorization:
1733 // ~(A * C1) + A
1734 // ((A * -C1) - 1) + A
1735 // ((A * -C1) + A) - 1
1736 // (A * (1 - C1)) - 1
1737 if (match(V: &I,
1738 P: m_c_Add(L: m_OneUse(SubPattern: m_Not(V: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: A), R: m_APInt(Res&: C1))))),
1739 R: m_Deferred(V: A)))) {
1740 Type *Ty = I.getType();
1741 Constant *NewMulC = ConstantInt::get(Ty, V: 1 - *C1);
1742 Value *NewMul = Builder.CreateMul(LHS: A, RHS: NewMulC);
1743 return BinaryOperator::CreateAdd(V1: NewMul, V2: ConstantInt::getAllOnesValue(Ty));
1744 }
1745
1746 // (A * -2**C) + B --> B - (A << C)
1747 const APInt *NegPow2C;
1748 if (match(V: &I, P: m_c_Add(L: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: A), R: m_NegatedPower2(V&: NegPow2C))),
1749 R: m_Value(V&: B)))) {
1750 Constant *ShiftAmtC = ConstantInt::get(Ty, V: NegPow2C->countr_zero());
1751 Value *Shl = Builder.CreateShl(LHS: A, RHS: ShiftAmtC);
1752 return BinaryOperator::CreateSub(V1: B, V2: Shl);
1753 }
1754
1755 // Canonicalize signum variant that ends in add:
1756 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1757 uint64_t BitWidth = Ty->getScalarSizeInBits();
1758 if (match(V: LHS, P: m_AShr(L: m_Value(V&: A), R: m_SpecificIntAllowPoison(V: BitWidth - 1))) &&
1759 match(V: RHS, P: m_OneUse(SubPattern: m_ZExt(Op: m_OneUse(SubPattern: m_SpecificICmp(
1760 MatchPred: CmpInst::ICMP_SGT, L: m_Specific(V: A), R: m_ZeroInt())))))) {
1761 Value *NotZero = Builder.CreateIsNotNull(Arg: A, Name: "isnotnull");
1762 Value *Zext = Builder.CreateZExt(V: NotZero, DestTy: Ty, Name: "isnotnull.zext");
1763 return BinaryOperator::CreateOr(V1: LHS, V2: Zext);
1764 }
1765
1766 {
1767 Value *Cond, *Ext;
1768 Constant *C;
1769 // (add X, (sext/zext (icmp eq X, C)))
1770 // -> (select (icmp eq X, C), (add C, (sext/zext 1)), X)
1771 auto CondMatcher =
1772 m_Value(V&: Cond, Match: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Deferred(V: A),
1773 R: m_ImmConstant(C)));
1774
1775 if (match(V: &I,
1776 P: m_c_Add(L: m_Value(V&: A), R: m_Value(V&: Ext, Match: m_ZExtOrSExt(Op: CondMatcher)))) &&
1777 Ext->hasOneUse()) {
1778 Value *Add = isa<ZExtInst>(Val: Ext) ? InstCombiner::AddOne(C)
1779 : InstCombiner::SubOne(C);
1780 return replaceInstUsesWith(I, V: Builder.CreateSelect(C: Cond, True: Add, False: A));
1781 }
1782 }
1783
1784 // (add (add A, 1), (sext (icmp ne A, 0))) => call umax(A, 1)
1785 if (match(V: LHS, P: m_Add(L: m_Value(V&: A), R: m_One())) &&
1786 match(V: RHS, P: m_OneUse(SubPattern: m_SExt(Op: m_OneUse(SubPattern: m_SpecificICmp(
1787 MatchPred: ICmpInst::ICMP_NE, L: m_Specific(V: A), R: m_ZeroInt())))))) {
1788 Value *OneConst = ConstantInt::get(Ty: A->getType(), V: 1);
1789 Value *UMax = Builder.CreateBinaryIntrinsic(ID: Intrinsic::umax, LHS: A, RHS: OneConst);
1790 return replaceInstUsesWith(I, V: UMax);
1791 }
1792
1793 if (Instruction *Ashr = foldAddToAshr(Add&: I))
1794 return Ashr;
1795
1796 // Ceiling division by power-of-2:
1797 // (X >> log2(N)) + zext(X & (N-1) != 0) --> (X + (N-1)) >> log2(N)
1798 // This is valid when adding (N-1) to X doesn't overflow.
1799 {
1800 Value *X;
1801 const APInt *ShiftAmt, *Mask;
1802 CmpPredicate Pred;
1803
1804 // Match: (X >> C) + zext((X & Mask) != 0)
1805 // or: zext((X & Mask) != 0) + (X >> C)
1806 if (match(V: &I, P: m_c_Add(L: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: ShiftAmt))),
1807 R: m_ZExt(Op: m_SpecificICmp(
1808 MatchPred: ICmpInst::ICMP_NE,
1809 L: m_And(L: m_Deferred(V: X), R: m_LowBitMask(V&: Mask)),
1810 R: m_ZeroInt())))) &&
1811 Mask->popcount() == *ShiftAmt) {
1812
1813 // Check if X + Mask doesn't overflow
1814 Constant *MaskC = ConstantInt::get(Ty: X->getType(), V: *Mask);
1815 if (willNotOverflowUnsignedAdd(LHS: X, RHS: MaskC, CxtI: I)) {
1816 // (X + Mask) >> ShiftAmt
1817 Value *Add = Builder.CreateNUWAdd(LHS: X, RHS: MaskC);
1818 return BinaryOperator::CreateLShr(
1819 V1: Add, V2: ConstantInt::get(Ty: X->getType(), V: *ShiftAmt));
1820 }
1821 }
1822 }
1823
1824 // (~X) + (~Y) --> -2 - (X + Y)
1825 {
1826 // To ensure we can save instructions we need to ensure that we consume both
1827 // LHS/RHS (i.e they have a `not`).
1828 bool ConsumesLHS, ConsumesRHS;
1829 if (isFreeToInvert(V: LHS, WillInvertAllUses: LHS->hasOneUse(), DoesConsume&: ConsumesLHS) && ConsumesLHS &&
1830 isFreeToInvert(V: RHS, WillInvertAllUses: RHS->hasOneUse(), DoesConsume&: ConsumesRHS) && ConsumesRHS) {
1831 Value *NotLHS = getFreelyInverted(V: LHS, WillInvertAllUses: LHS->hasOneUse(), Builder: &Builder);
1832 Value *NotRHS = getFreelyInverted(V: RHS, WillInvertAllUses: RHS->hasOneUse(), Builder: &Builder);
1833 assert(NotLHS != nullptr && NotRHS != nullptr &&
1834 "isFreeToInvert desynced with getFreelyInverted");
1835 Value *LHSPlusRHS = Builder.CreateAdd(LHS: NotLHS, RHS: NotRHS);
1836 return BinaryOperator::CreateSub(
1837 V1: ConstantInt::getSigned(Ty: RHS->getType(), V: -2), V2: LHSPlusRHS);
1838 }
1839 }
1840
1841 if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &I))
1842 return R;
1843
1844 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1845 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1846 // computeKnownBits.
1847 bool Changed = false;
1848 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS: LHSCache, RHS: RHSCache, CxtI: I)) {
1849 Changed = true;
1850 I.setHasNoSignedWrap(true);
1851 }
1852 if (!I.hasNoUnsignedWrap() &&
1853 willNotOverflowUnsignedAdd(LHS: LHSCache, RHS: RHSCache, CxtI: I)) {
1854 Changed = true;
1855 I.setHasNoUnsignedWrap(true);
1856 }
1857
1858 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1859 return V;
1860
1861 if (Instruction *V =
1862 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1863 return V;
1864
1865 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1866 return SatAdd;
1867
1868 // usub.sat(A, B) + B => umax(A, B)
1869 if (match(V: &I, P: m_c_BinOp(
1870 L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::usub_sat>(Op0: m_Value(V&: A), Op1: m_Value(V&: B))),
1871 R: m_Deferred(V: B)))) {
1872 return replaceInstUsesWith(I,
1873 V: Builder.CreateIntrinsic(ID: Intrinsic::umax, Types: {I.getType()}, Args: {A, B}));
1874 }
1875
1876 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1877 if (match(V: LHS, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: A)))) &&
1878 match(V: RHS, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: B)))) &&
1879 haveNoCommonBitsSet(LHSCache: A, RHSCache: B, SQ: SQ.getWithInstruction(I: &I)))
1880 return replaceInstUsesWith(
1881 I, V: Builder.CreateIntrinsic(ID: Intrinsic::ctpop, Types: {I.getType()},
1882 Args: {Builder.CreateOr(LHS: A, RHS: B)}));
1883
1884 // Fold the log2_ceil idiom:
1885 // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
1886 // -->
1887 // BW - ctlz(A - 1, false)
1888 const APInt *XorC;
1889 CmpPredicate Pred;
1890 if (match(V: &I,
1891 P: m_c_Add(
1892 L: m_ZExt(Op: m_ICmp(Pred, L: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: A)),
1893 R: m_One())),
1894 R: m_OneUse(SubPattern: m_ZExtOrSelf(Op: m_OneUse(SubPattern: m_Xor(
1895 L: m_OneUse(SubPattern: m_TruncOrSelf(Op: m_OneUse(
1896 SubPattern: m_Intrinsic<Intrinsic::ctlz>(Op0: m_Deferred(V: A), Op1: m_One())))),
1897 R: m_APInt(Res&: XorC))))))) &&
1898 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
1899 *XorC == A->getType()->getScalarSizeInBits() - 1) {
1900 Value *Sub = Builder.CreateAdd(LHS: A, RHS: Constant::getAllOnesValue(Ty: A->getType()));
1901 Value *Ctlz = Builder.CreateIntrinsic(ID: Intrinsic::ctlz, Types: {A->getType()},
1902 Args: {Sub, Builder.getFalse()});
1903 Value *Ret = Builder.CreateSub(
1904 LHS: ConstantInt::get(Ty: A->getType(), V: A->getType()->getScalarSizeInBits()),
1905 RHS: Ctlz, Name: "", /*HasNUW=*/true, /*HasNSW=*/true);
1906 return replaceInstUsesWith(I, V: Builder.CreateZExtOrTrunc(V: Ret, DestTy: I.getType()));
1907 }
1908
1909 if (Instruction *Res = foldSquareSumInt(I))
1910 return Res;
1911
1912 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
1913 return Res;
1914
1915 if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
1916 return Res;
1917
1918 // Re-enqueue users of the induction variable of add recurrence if we infer
1919 // new nuw/nsw flags.
1920 if (Changed) {
1921 PHINode *PHI;
1922 Value *Start, *Step;
1923 if (matchSimpleRecurrence(I: &I, P&: PHI, Start, Step))
1924 Worklist.pushUsersToWorkList(I&: *PHI);
1925 }
1926
1927 return Changed ? &I : nullptr;
1928}
1929
1930/// Eliminate an op from a linear interpolation (lerp) pattern.
1931static Instruction *factorizeLerp(BinaryOperator &I,
1932 InstCombiner::BuilderTy &Builder) {
1933 Value *X, *Y, *Z;
1934 if (!match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_c_FMul(L: m_Value(V&: Y),
1935 R: m_OneUse(SubPattern: m_FSub(L: m_FPOne(),
1936 R: m_Value(V&: Z))))),
1937 R: m_OneUse(SubPattern: m_c_FMul(L: m_Value(V&: X), R: m_Deferred(V: Z))))))
1938 return nullptr;
1939
1940 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1941 Value *XY = Builder.CreateFSubFMF(L: X, R: Y, FMFSource: &I);
1942 Value *MulZ = Builder.CreateFMulFMF(L: Z, R: XY, FMFSource: &I);
1943 return BinaryOperator::CreateFAddFMF(V1: Y, V2: MulZ, FMFSource: &I);
1944}
1945
1946/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1947static Instruction *factorizeFAddFSub(BinaryOperator &I,
1948 InstCombiner::BuilderTy &Builder) {
1949 assert((I.getOpcode() == Instruction::FAdd ||
1950 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1951 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1952 "FP factorization requires FMF");
1953
1954 if (Instruction *Lerp = factorizeLerp(I, Builder))
1955 return Lerp;
1956
1957 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
1958 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1959 return nullptr;
1960
1961 Value *X, *Y, *Z;
1962 bool IsFMul;
1963 if ((match(V: Op0, P: m_FMul(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
1964 match(V: Op1, P: m_c_FMul(L: m_Value(V&: Y), R: m_Specific(V: Z)))) ||
1965 (match(V: Op0, P: m_FMul(L: m_Value(V&: Z), R: m_Value(V&: X))) &&
1966 match(V: Op1, P: m_c_FMul(L: m_Value(V&: Y), R: m_Specific(V: Z)))))
1967 IsFMul = true;
1968 else if (match(V: Op0, P: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
1969 match(V: Op1, P: m_FDiv(L: m_Value(V&: Y), R: m_Specific(V: Z))))
1970 IsFMul = false;
1971 else
1972 return nullptr;
1973
1974 // (X * Z) + (Y * Z) --> (X + Y) * Z
1975 // (X * Z) - (Y * Z) --> (X - Y) * Z
1976 // (X / Z) + (Y / Z) --> (X + Y) / Z
1977 // (X / Z) - (Y / Z) --> (X - Y) / Z
1978 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1979 Value *XY = IsFAdd ? Builder.CreateFAddFMF(L: X, R: Y, FMFSource: &I)
1980 : Builder.CreateFSubFMF(L: X, R: Y, FMFSource: &I);
1981
1982 // Bail out if we just created a denormal constant.
1983 // TODO: This is copied from a previous implementation. Is it necessary?
1984 const APFloat *C;
1985 if (match(V: XY, P: m_APFloat(Res&: C)) && !C->isNormal())
1986 return nullptr;
1987
1988 return IsFMul ? BinaryOperator::CreateFMulFMF(V1: XY, V2: Z, FMFSource: &I)
1989 : BinaryOperator::CreateFDivFMF(V1: XY, V2: Z, FMFSource: &I);
1990}
1991
1992Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1993 if (Value *V = simplifyFAddInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
1994 FMF: I.getFastMathFlags(),
1995 Q: SQ.getWithInstruction(I: &I)))
1996 return replaceInstUsesWith(I, V);
1997
1998 if (SimplifyAssociativeOrCommutative(I))
1999 return &I;
2000
2001 if (Instruction *X = foldVectorBinop(Inst&: I))
2002 return X;
2003
2004 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
2005 return Phi;
2006
2007 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
2008 return FoldedFAdd;
2009
2010 // B = fadd A, 0.0
2011 // Z = Op B
2012 // can be transformed into
2013 // Z = Op A
2014 // Where Op is such that we can ignore sign of 0 in fadd
2015 Value *A;
2016 if (match(V: &I, P: m_OneUse(SubPattern: m_FAdd(L: m_Value(V&: A), R: m_AnyZeroFP()))) &&
2017 canIgnoreSignBitOfZero(U: *I.use_begin()))
2018 return replaceInstUsesWith(I, V: A);
2019
2020 // (-X) + Y --> Y - X
2021 Value *X, *Y;
2022 if (match(V: &I, P: m_c_FAdd(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y))))
2023 return BinaryOperator::CreateFSubFMF(V1: Y, V2: X, FMFSource: &I);
2024
2025 // Similar to above, but look through fmul/fdiv for the negated term.
2026 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
2027 Value *Z;
2028 if (match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_c_FMul(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y))),
2029 R: m_Value(V&: Z)))) {
2030 Value *XY = Builder.CreateFMulFMF(L: X, R: Y, FMFSource: &I);
2031 return BinaryOperator::CreateFSubFMF(V1: Z, V2: XY, FMFSource: &I);
2032 }
2033 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
2034 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
2035 if (match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_FDiv(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y))),
2036 R: m_Value(V&: Z))) ||
2037 match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_FNeg(X: m_Value(V&: Y)))),
2038 R: m_Value(V&: Z)))) {
2039 Value *XY = Builder.CreateFDivFMF(L: X, R: Y, FMFSource: &I);
2040 return BinaryOperator::CreateFSubFMF(V1: Z, V2: XY, FMFSource: &I);
2041 }
2042
2043 // Check for (fadd double (sitofp x), y), see if we can merge this into an
2044 // integer add followed by a promotion.
2045 if (Instruction *R = foldFBinOpOfIntCasts(I))
2046 return R;
2047
2048 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
2049 // Handle specials cases for FAdd with selects feeding the operation
2050 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
2051 return replaceInstUsesWith(I, V);
2052
2053 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2054 if (Instruction *F = factorizeFAddFSub(I, Builder))
2055 return F;
2056
2057 if (Instruction *F = foldSquareSumFP(I))
2058 return F;
2059
2060 // Try to fold fadd into start value of reduction intrinsic.
2061 if (match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_fadd>(
2062 Op0: m_AnyZeroFP(), Op1: m_Value(V&: X))),
2063 R: m_Value(V&: Y)))) {
2064 // fadd (rdx 0.0, X), Y --> rdx Y, X
2065 return replaceInstUsesWith(
2066 I, V: Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_fadd,
2067 Types: {X->getType()}, Args: {Y, X}, FMFSource: &I));
2068 }
2069 const APFloat *StartC, *C;
2070 if (match(V: LHS, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_fadd>(
2071 Op0: m_APFloat(Res&: StartC), Op1: m_Value(V&: X)))) &&
2072 match(V: RHS, P: m_APFloat(Res&: C))) {
2073 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
2074 Constant *NewStartC = ConstantFP::get(Ty: I.getType(), V: *C + *StartC);
2075 return replaceInstUsesWith(
2076 I, V: Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_fadd,
2077 Types: {X->getType()}, Args: {NewStartC, X}, FMFSource: &I));
2078 }
2079
2080 // (X * MulC) + X --> X * (MulC + 1.0)
2081 Constant *MulC;
2082 if (match(V: &I, P: m_c_FAdd(L: m_FMul(L: m_Value(V&: X), R: m_ImmConstant(C&: MulC)),
2083 R: m_Deferred(V: X)))) {
2084 if (Constant *NewMulC = ConstantFoldBinaryOpOperands(
2085 Opcode: Instruction::FAdd, LHS: MulC, RHS: ConstantFP::get(Ty: I.getType(), V: 1.0), DL))
2086 return BinaryOperator::CreateFMulFMF(V1: X, V2: NewMulC, FMFSource: &I);
2087 }
2088
2089 // (-X - Y) + (X + Z) --> Z - Y
2090 if (match(V: &I, P: m_c_FAdd(L: m_FSub(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y)),
2091 R: m_c_FAdd(L: m_Deferred(V: X), R: m_Value(V&: Z)))))
2092 return BinaryOperator::CreateFSubFMF(V1: Z, V2: Y, FMFSource: &I);
2093
2094 if (Value *V = FAddCombine(Builder).simplify(I: &I))
2095 return replaceInstUsesWith(I, V);
2096 }
2097
2098 // minumum(X, Y) + maximum(X, Y) => X + Y.
2099 if (match(V: &I,
2100 P: m_c_FAdd(L: m_Intrinsic<Intrinsic::maximum>(Op0: m_Value(V&: X), Op1: m_Value(V&: Y)),
2101 R: m_c_Intrinsic<Intrinsic::minimum>(Op0: m_Deferred(V: X),
2102 Op1: m_Deferred(V: Y))))) {
2103 BinaryOperator *Result = BinaryOperator::CreateFAddFMF(V1: X, V2: Y, FMFSource: &I);
2104 // We cannot preserve ninf if nnan flag is not set.
2105 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
2106 // while in optimized version NaN + Inf and this is a poison with ninf flag.
2107 if (!Result->hasNoNaNs())
2108 Result->setHasNoInfs(false);
2109 return Result;
2110 }
2111
2112 return nullptr;
2113}
2114
2115CommonPointerBase CommonPointerBase::compute(Value *LHS, Value *RHS) {
2116 CommonPointerBase Base;
2117
2118 if (LHS->getType() != RHS->getType())
2119 return Base;
2120
2121 // Collect all base pointers of LHS.
2122 SmallPtrSet<Value *, 16> Ptrs;
2123 Value *Ptr = LHS;
2124 while (true) {
2125 Ptrs.insert(Ptr);
2126 if (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr))
2127 Ptr = GEP->getPointerOperand();
2128 else
2129 break;
2130 }
2131
2132 // Find common base and collect RHS GEPs.
2133 bool First = true;
2134 while (true) {
2135 if (Ptrs.contains(Ptr: RHS)) {
2136 Base.Ptr = RHS;
2137 break;
2138 }
2139
2140 if (auto *GEP = dyn_cast<GEPOperator>(Val: RHS)) {
2141 Base.RHSGEPs.push_back(Elt: GEP);
2142 if (First) {
2143 First = false;
2144 Base.RHSNW = GEP->getNoWrapFlags();
2145 } else {
2146 Base.RHSNW = Base.RHSNW.intersectForOffsetAdd(Other: GEP->getNoWrapFlags());
2147 }
2148 RHS = GEP->getPointerOperand();
2149 } else {
2150 // No common base.
2151 return Base;
2152 }
2153 }
2154
2155 // Collect LHS GEPs.
2156 First = true;
2157 while (true) {
2158 if (LHS == Base.Ptr)
2159 break;
2160
2161 auto *GEP = cast<GEPOperator>(Val: LHS);
2162 Base.LHSGEPs.push_back(Elt: GEP);
2163 if (First) {
2164 First = false;
2165 Base.LHSNW = GEP->getNoWrapFlags();
2166 } else {
2167 Base.LHSNW = Base.LHSNW.intersectForOffsetAdd(Other: GEP->getNoWrapFlags());
2168 }
2169 LHS = GEP->getPointerOperand();
2170 }
2171
2172 return Base;
2173}
2174
2175bool CommonPointerBase::isExpensive() const {
2176 unsigned NumGEPs = 0;
2177 auto ProcessGEPs = [&NumGEPs](ArrayRef<GEPOperator *> GEPs) {
2178 bool SeenMultiUse = false;
2179 for (GEPOperator *GEP : GEPs) {
2180 // Only count multi-use GEPs, excluding the first one. For the first one,
2181 // we will directly reuse the offset. For one-use GEPs, their offset will
2182 // be folded into a multi-use GEP.
2183 if (!GEP->hasOneUse()) {
2184 if (SeenMultiUse)
2185 ++NumGEPs;
2186 SeenMultiUse = true;
2187 }
2188 }
2189 };
2190 ProcessGEPs(LHSGEPs);
2191 ProcessGEPs(RHSGEPs);
2192 return NumGEPs > 2;
2193}
2194
2195/// Optimize pointer differences into the same array into a size. Consider:
2196/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
2197/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
2198Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
2199 Type *Ty, bool IsNUW) {
2200 CommonPointerBase Base = CommonPointerBase::compute(LHS, RHS);
2201 if (!Base.Ptr || Base.isExpensive())
2202 return nullptr;
2203
2204 // To avoid duplicating the offset arithmetic, rewrite the GEP to use the
2205 // computed offset.
2206 // TODO: We should probably do this even if there is only one GEP.
2207 bool RewriteGEPs = !Base.LHSGEPs.empty() && !Base.RHSGEPs.empty();
2208
2209 Type *IdxTy = DL.getIndexType(PtrTy: LHS->getType());
2210 Value *Result = EmitGEPOffsets(GEPs: Base.LHSGEPs, NW: Base.LHSNW, IdxTy, RewriteGEPs);
2211 Value *Offset2 = EmitGEPOffsets(GEPs: Base.RHSGEPs, NW: Base.RHSNW, IdxTy, RewriteGEPs);
2212
2213 // If this is a single inbounds GEP and the original sub was nuw,
2214 // then the final multiplication is also nuw.
2215 if (auto *I = dyn_cast<OverflowingBinaryOperator>(Val: Result))
2216 if (IsNUW && match(V: Offset2, P: m_Zero()) && Base.LHSNW.isInBounds() &&
2217 (I->use_empty() || I->hasOneUse()) && I->hasNoSignedWrap() &&
2218 !I->hasNoUnsignedWrap() &&
2219 ((I->getOpcode() == Instruction::Mul &&
2220 match(V: I->getOperand(i_nocapture: 1), P: m_NonNegative())) ||
2221 I->getOpcode() == Instruction::Shl))
2222 cast<Instruction>(Val: I)->setHasNoUnsignedWrap();
2223
2224 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2225 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2226 // If both GEPs are nuw and the original sub is nuw, the new sub is also nuw.
2227 if (!match(V: Offset2, P: m_Zero())) {
2228 Result =
2229 Builder.CreateSub(LHS: Result, RHS: Offset2, Name: "gepdiff",
2230 HasNUW: IsNUW && Base.LHSNW.hasNoUnsignedWrap() &&
2231 Base.RHSNW.hasNoUnsignedWrap(),
2232 HasNSW: Base.LHSNW.isInBounds() && Base.RHSNW.isInBounds());
2233 }
2234
2235 return Builder.CreateIntCast(V: Result, DestTy: Ty, isSigned: true);
2236}
2237
2238static Instruction *foldSubOfMinMax(BinaryOperator &I,
2239 InstCombiner::BuilderTy &Builder) {
2240 Value *Op0 = I.getOperand(i_nocapture: 0);
2241 Value *Op1 = I.getOperand(i_nocapture: 1);
2242 Type *Ty = I.getType();
2243 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op1);
2244 if (!MinMax)
2245 return nullptr;
2246
2247 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2248 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2249 Value *X = MinMax->getLHS();
2250 Value *Y = MinMax->getRHS();
2251 if (match(V: Op0, P: m_c_Add(L: m_Specific(V: X), R: m_Specific(V: Y))) &&
2252 (Op0->hasOneUse() || Op1->hasOneUse())) {
2253 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: MinMax->getIntrinsicID());
2254 Function *F = Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: InvID, Tys: Ty);
2255 return CallInst::Create(Func: F, Args: {X, Y});
2256 }
2257
2258 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2259 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2260 Value *Z;
2261 if (match(V: Op1, P: m_OneUse(SubPattern: m_UMin(L: m_Value(V&: Y), R: m_Value(V&: Z))))) {
2262 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Y), R: m_Value(V&: X))))) {
2263 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, Types: Ty, Args: {Y, Z});
2264 return BinaryOperator::CreateAdd(V1: X, V2: USub);
2265 }
2266 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Z), R: m_Value(V&: X))))) {
2267 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, Types: Ty, Args: {Z, Y});
2268 return BinaryOperator::CreateAdd(V1: X, V2: USub);
2269 }
2270 }
2271
2272 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2273 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2274 if (MinMax->isSigned() && match(V: Y, P: m_ZeroInt()) &&
2275 match(V: X, P: m_NSWSub(L: m_Specific(V: Op0), R: m_Value(V&: Z)))) {
2276 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: MinMax->getIntrinsicID());
2277 Function *F = Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: InvID, Tys: Ty);
2278 return CallInst::Create(Func: F, Args: {Op0, Z});
2279 }
2280
2281 return nullptr;
2282}
2283
2284Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
2285 if (Value *V = simplifySubInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
2286 IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(),
2287 Q: SQ.getWithInstruction(I: &I)))
2288 return replaceInstUsesWith(I, V);
2289
2290 if (Instruction *X = foldVectorBinop(Inst&: I))
2291 return X;
2292
2293 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
2294 return Phi;
2295
2296 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
2297
2298 // If this is a 'B = x-(-A)', change to B = x+A.
2299 // We deal with this without involving Negator to preserve NSW flag.
2300 if (Value *V = dyn_castNegVal(V: Op1)) {
2301 BinaryOperator *Res = BinaryOperator::CreateAdd(V1: Op0, V2: V);
2302
2303 if (const auto *BO = dyn_cast<BinaryOperator>(Val: Op1)) {
2304 assert(BO->getOpcode() == Instruction::Sub &&
2305 "Expected a subtraction operator!");
2306 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2307 Res->setHasNoSignedWrap(true);
2308 } else {
2309 if (cast<Constant>(Val: Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2310 Res->setHasNoSignedWrap(true);
2311 }
2312
2313 return Res;
2314 }
2315
2316 // Try this before Negator to preserve NSW flag.
2317 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
2318 return R;
2319
2320 Constant *C;
2321 if (match(V: Op0, P: m_ImmConstant(C))) {
2322 Value *X;
2323 Constant *C2;
2324
2325 // C-(X+C2) --> (C-C2)-X
2326 if (match(V: Op1, P: m_Add(L: m_Value(V&: X), R: m_ImmConstant(C&: C2)))) {
2327 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2328 // => (C-C2)-X can have NSW/NUW
2329 bool WillNotSOV = willNotOverflowSignedSub(LHS: C, RHS: C2, CxtI: I);
2330 BinaryOperator *Res =
2331 BinaryOperator::CreateSub(V1: ConstantExpr::getSub(C1: C, C2), V2: X);
2332 auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1);
2333 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO1->hasNoSignedWrap() &&
2334 WillNotSOV);
2335 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() &&
2336 OBO1->hasNoUnsignedWrap());
2337 return Res;
2338 }
2339 }
2340
2341 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2342 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2343 return Ext;
2344
2345 bool Changed = false;
2346 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(LHS: Op0, RHS: Op1, CxtI: I)) {
2347 Changed = true;
2348 I.setHasNoSignedWrap(true);
2349 }
2350 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(LHS: Op0, RHS: Op1, CxtI: I)) {
2351 Changed = true;
2352 I.setHasNoUnsignedWrap(true);
2353 }
2354
2355 return Changed ? &I : nullptr;
2356 };
2357
2358 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2359 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2360 // a pure negation used by a select that looks like abs/nabs.
2361 bool IsNegation = match(V: Op0, P: m_ZeroInt());
2362 if (!IsNegation || none_of(Range: I.users(), P: match_fn(P: m_c_Select(L: m_Specific(V: Op1),
2363 R: m_Specific(V: &I))))) {
2364 if (Value *NegOp1 = Negator::Negate(LHSIsZero: IsNegation, /* IsNSW */ IsNegation &&
2365 I.hasNoSignedWrap(),
2366 Root: Op1, IC&: *this))
2367 return BinaryOperator::CreateAdd(V1: NegOp1, V2: Op0);
2368 }
2369 if (IsNegation)
2370 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2371
2372 // (A*B)-(A*C) -> A*(B-C) etc
2373 if (Value *V = foldUsingDistributiveLaws(I))
2374 return replaceInstUsesWith(I, V);
2375
2376 if (I.getType()->isIntOrIntVectorTy(BitWidth: 1))
2377 return BinaryOperator::CreateXor(V1: Op0, V2: Op1);
2378
2379 // Replace (-1 - A) with (~A).
2380 if (match(V: Op0, P: m_AllOnes()))
2381 return BinaryOperator::CreateNot(Op: Op1);
2382
2383 // (X + -1) - Y --> ~Y + X
2384 Value *X, *Y;
2385 if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_AllOnes()))))
2386 return BinaryOperator::CreateAdd(V1: Builder.CreateNot(V: Op1), V2: X);
2387
2388 // if (C1 & C2) == C2 then (X & C1) - (X & C2) -> X & (C1 ^ C2)
2389 Constant *C1, *C2;
2390 if (match(V: Op0, P: m_And(L: m_Value(V&: X), R: m_ImmConstant(C&: C1))) &&
2391 match(V: Op1, P: m_And(L: m_Specific(V: X), R: m_ImmConstant(C&: C2)))) {
2392 Value *AndC = ConstantFoldBinaryInstruction(Opcode: Instruction::And, V1: C1, V2: C2);
2393 if (C2->isElementWiseEqual(Y: AndC))
2394 return BinaryOperator::CreateAnd(
2395 V1: X, V2: ConstantFoldBinaryInstruction(Opcode: Instruction::Xor, V1: C1, V2: C2));
2396 }
2397
2398 // Reassociate sub/add sequences to create more add instructions and
2399 // reduce dependency chains:
2400 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2401 Value *Z;
2402 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y))),
2403 R: m_Value(V&: Z))))) {
2404 Value *XZ = Builder.CreateAdd(LHS: X, RHS: Z);
2405 Value *YW = Builder.CreateAdd(LHS: Y, RHS: Op1);
2406 return BinaryOperator::CreateSub(V1: XZ, V2: YW);
2407 }
2408
2409 // ((X - Y) - Op1) --> X - (Y + Op1)
2410 if (match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
2411 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Val: Op0);
2412 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2413 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2414 Value *Add = Builder.CreateAdd(LHS: Y, RHS: Op1, Name: "", /*HasNUW=*/HasNUW,
2415 /*HasNSW=*/HasNSW);
2416 BinaryOperator *Sub = BinaryOperator::CreateSub(V1: X, V2: Add);
2417 Sub->setHasNoUnsignedWrap(HasNUW);
2418 Sub->setHasNoSignedWrap(HasNSW);
2419 return Sub;
2420 }
2421
2422 {
2423 // (X + Z) - (Y + Z) --> (X - Y)
2424 // This is done in other passes, but we want to be able to consume this
2425 // pattern in InstCombine so we can generate it without creating infinite
2426 // loops.
2427 if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
2428 match(V: Op1, P: m_c_Add(L: m_Value(V&: Y), R: m_Specific(V: Z))))
2429 return BinaryOperator::CreateSub(V1: X, V2: Y);
2430
2431 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2432 Constant *CX, *CY;
2433 if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_ImmConstant(C&: CX)))) &&
2434 match(V: Op1, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: Y), R: m_ImmConstant(C&: CY))))) {
2435 Value *OpsSub = Builder.CreateSub(LHS: X, RHS: Y);
2436 Constant *ConstsSub = ConstantExpr::getSub(C1: CX, C2: CY);
2437 return BinaryOperator::CreateAdd(V1: OpsSub, V2: ConstsSub);
2438 }
2439 }
2440
2441 {
2442 Value *W, *Z;
2443 if (match(V: Op0, P: m_AddLike(L: m_Value(V&: W), R: m_Value(V&: X))) &&
2444 match(V: Op1, P: m_AddLike(L: m_Value(V&: Y), R: m_Value(V&: Z)))) {
2445 Instruction *R = nullptr;
2446 if (W == Y)
2447 R = BinaryOperator::CreateSub(V1: X, V2: Z);
2448 else if (W == Z)
2449 R = BinaryOperator::CreateSub(V1: X, V2: Y);
2450 else if (X == Y)
2451 R = BinaryOperator::CreateSub(V1: W, V2: Z);
2452 else if (X == Z)
2453 R = BinaryOperator::CreateSub(V1: W, V2: Y);
2454 if (R) {
2455 bool NSW = I.hasNoSignedWrap() &&
2456 match(V: Op0, P: m_NSWAddLike(L: m_Value(), R: m_Value())) &&
2457 match(V: Op1, P: m_NSWAddLike(L: m_Value(), R: m_Value()));
2458
2459 bool NUW = I.hasNoUnsignedWrap() &&
2460 match(V: Op1, P: m_NUWAddLike(L: m_Value(), R: m_Value()));
2461 R->setHasNoSignedWrap(NSW);
2462 R->setHasNoUnsignedWrap(NUW);
2463 return R;
2464 }
2465 }
2466 }
2467
2468 // (~X) - (~Y) --> Y - X
2469 {
2470 // Need to ensure we can consume at least one of the `not` instructions,
2471 // otherwise this can inf loop.
2472 bool ConsumesOp0, ConsumesOp1;
2473 if (isFreeToInvert(V: Op0, WillInvertAllUses: Op0->hasOneUse(), DoesConsume&: ConsumesOp0) &&
2474 isFreeToInvert(V: Op1, WillInvertAllUses: Op1->hasOneUse(), DoesConsume&: ConsumesOp1) &&
2475 (ConsumesOp0 || ConsumesOp1)) {
2476 Value *NotOp0 = getFreelyInverted(V: Op0, WillInvertAllUses: Op0->hasOneUse(), Builder: &Builder);
2477 Value *NotOp1 = getFreelyInverted(V: Op1, WillInvertAllUses: Op1->hasOneUse(), Builder: &Builder);
2478 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2479 "isFreeToInvert desynced with getFreelyInverted");
2480 return BinaryOperator::CreateSub(V1: NotOp1, V2: NotOp0);
2481 }
2482 }
2483
2484 auto m_AddRdx = [](Value *&Vec) {
2485 return m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_add>(Op0: m_Value(V&: Vec)));
2486 };
2487 Value *V0, *V1;
2488 if (match(V: Op0, P: m_AddRdx(V0)) && match(V: Op1, P: m_AddRdx(V1)) &&
2489 V0->getType() == V1->getType()) {
2490 // Difference of sums is sum of differences:
2491 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2492 Value *Sub = Builder.CreateSub(LHS: V0, RHS: V1);
2493 Value *Rdx = Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_add,
2494 Types: {Sub->getType()}, Args: {Sub});
2495 return replaceInstUsesWith(I, V: Rdx);
2496 }
2497
2498 if (Constant *C = dyn_cast<Constant>(Val: Op0)) {
2499 Value *X;
2500 if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))
2501 // C - (zext bool) --> bool ? C - 1 : C
2502 return SelectInst::Create(C: X, S1: InstCombiner::SubOne(C), S2: C);
2503 if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))
2504 // C - (sext bool) --> bool ? C + 1 : C
2505 return SelectInst::Create(C: X, S1: InstCombiner::AddOne(C), S2: C);
2506
2507 // C - ~X == X + (1+C)
2508 if (match(V: Op1, P: m_Not(V: m_Value(V&: X))))
2509 return BinaryOperator::CreateAdd(V1: X, V2: InstCombiner::AddOne(C));
2510
2511 // Try to fold constant sub into select arguments.
2512 if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1))
2513 if (Instruction *R = FoldOpIntoSelect(Op&: I, SI))
2514 return R;
2515
2516 // Try to fold constant sub into PHI values.
2517 if (PHINode *PN = dyn_cast<PHINode>(Val: Op1))
2518 if (Instruction *R = foldOpIntoPhi(I, PN))
2519 return R;
2520
2521 Constant *C2;
2522
2523 // C-(C2-X) --> X+(C-C2)
2524 if (match(V: Op1, P: m_Sub(L: m_ImmConstant(C&: C2), R: m_Value(V&: X))))
2525 return BinaryOperator::CreateAdd(V1: X, V2: ConstantExpr::getSub(C1: C, C2));
2526 }
2527
2528 const APInt *Op0C;
2529 if (match(V: Op0, P: m_APInt(Res&: Op0C))) {
2530 if (Op0C->isMask()) {
2531 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2532 // zero. We don't use information from dominating conditions so this
2533 // transform is easier to reverse if necessary.
2534 KnownBits RHSKnown = llvm::computeKnownBits(
2535 V: Op1, Q: SQ.getWithInstruction(I: &I).getWithoutDomCondCache());
2536 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2537 return BinaryOperator::CreateXor(V1: Op1, V2: Op0);
2538 }
2539
2540 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2541 // (C3 - ((C2 & C3) - 1)) is pow2
2542 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2543 // C2 is negative pow2 || sub nuw
2544 const APInt *C2, *C3;
2545 BinaryOperator *InnerSub;
2546 if (match(V: Op1, P: m_OneUse(SubPattern: m_And(L: m_BinOp(I&: InnerSub), R: m_APInt(Res&: C2)))) &&
2547 match(V: InnerSub, P: m_Sub(L: m_APInt(Res&: C3), R: m_Value(V&: X))) &&
2548 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2549 APInt C2AndC3 = *C2 & *C3;
2550 APInt C2AndC3Minus1 = C2AndC3 - 1;
2551 APInt C2AddC3 = *C2 + *C3;
2552 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2553 C2AndC3Minus1.isSubsetOf(RHS: C2AddC3)) {
2554 Value *And = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty: I.getType(), V: *C2));
2555 return BinaryOperator::CreateAdd(
2556 V1: And, V2: ConstantInt::get(Ty: I.getType(), V: *Op0C - C2AndC3));
2557 }
2558 }
2559 }
2560
2561 {
2562 Value *Y;
2563 // X-(X+Y) == -Y X-(Y+X) == -Y
2564 if (match(V: Op1, P: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: Y))))
2565 return BinaryOperator::CreateNeg(Op: Y);
2566
2567 // (X-Y)-X == -Y
2568 if (match(V: Op0, P: m_Sub(L: m_Specific(V: Op1), R: m_Value(V&: Y))))
2569 return BinaryOperator::CreateNeg(Op: Y);
2570 }
2571
2572 // (sub (or A, B) (and A, B)) --> (xor A, B)
2573 {
2574 Value *A, *B;
2575 if (match(V: Op1, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2576 match(V: Op0, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2577 return BinaryOperator::CreateXor(V1: A, V2: B);
2578 }
2579
2580 // (sub (add A, B) (or A, B)) --> (and A, B)
2581 {
2582 Value *A, *B;
2583 if (match(V: Op0, P: m_Add(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2584 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2585 return BinaryOperator::CreateAnd(V1: A, V2: B);
2586 }
2587
2588 // (sub (add A, B) (and A, B)) --> (or A, B)
2589 {
2590 Value *A, *B;
2591 if (match(V: Op0, P: m_Add(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2592 match(V: Op1, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2593 return BinaryOperator::CreateOr(V1: A, V2: B);
2594 }
2595
2596 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2597 {
2598 Value *A, *B;
2599 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2600 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))) &&
2601 (Op0->hasOneUse() || Op1->hasOneUse()))
2602 return BinaryOperator::CreateNeg(Op: Builder.CreateXor(LHS: A, RHS: B));
2603 }
2604
2605 // (sub (or A, B), (xor A, B)) --> (and A, B)
2606 {
2607 Value *A, *B;
2608 if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2609 match(V: Op0, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2610 return BinaryOperator::CreateAnd(V1: A, V2: B);
2611 }
2612
2613 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2614 {
2615 Value *A, *B;
2616 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2617 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))) &&
2618 (Op0->hasOneUse() || Op1->hasOneUse()))
2619 return BinaryOperator::CreateNeg(Op: Builder.CreateAnd(LHS: A, RHS: B));
2620 }
2621
2622 {
2623 Value *Y;
2624 // ((X | Y) - X) --> (~X & Y)
2625 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Value(V&: Y), R: m_Specific(V: Op1)))))
2626 return BinaryOperator::CreateAnd(
2627 V1: Y, V2: Builder.CreateNot(V: Op1, Name: Op1->getName() + ".not"));
2628 }
2629
2630 {
2631 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2632 Value *X;
2633 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Specific(V: Op1),
2634 R: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))))))) {
2635 return BinaryOperator::CreateNeg(Op: Builder.CreateAnd(
2636 LHS: Op1, RHS: Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: I.getType()))));
2637 }
2638 }
2639
2640 {
2641 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2642 Constant *C;
2643 if (match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Specific(V: Op1), R: m_Constant(C))))) {
2644 return BinaryOperator::CreateNeg(
2645 Op: Builder.CreateAnd(LHS: Op1, RHS: Builder.CreateNot(V: C)));
2646 }
2647 }
2648
2649 {
2650 // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2651 // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2652 Value *C, *X;
2653 auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2654 return match(V: LHS, P: m_OneUse(SubPattern: m_c_Xor(L: m_Value(V&: X), R: m_Specific(V: RHS)))) &&
2655 match(V: RHS, P: m_SExt(Op: m_Value(V&: C))) &&
2656 (C->getType()->getScalarSizeInBits() == 1);
2657 };
2658 if (m_SubXorCmp(Op0, Op1))
2659 return SelectInst::Create(C, S1: Builder.CreateNeg(V: X), S2: X);
2660 if (m_SubXorCmp(Op1, Op0))
2661 return SelectInst::Create(C, S1: X, S2: Builder.CreateNeg(V: X));
2662 }
2663
2664 if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &I))
2665 return R;
2666
2667 if (Instruction *R = foldSubOfMinMax(I, Builder))
2668 return R;
2669
2670 {
2671 // If we have a subtraction between some value and a select between
2672 // said value and something else, sink subtraction into select hands, i.e.:
2673 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2674 // ->
2675 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2676 // or
2677 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2678 // ->
2679 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2680 // This will result in select between new subtraction and 0.
2681 auto SinkSubIntoSelect =
2682 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2683 auto SubBuilder) -> Instruction * {
2684 Value *Cond, *TrueVal, *FalseVal;
2685 if (!match(V: Select, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TrueVal),
2686 R: m_Value(V&: FalseVal)))))
2687 return nullptr;
2688 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2689 return nullptr;
2690 // While it is really tempting to just create two subtractions and let
2691 // InstCombine fold one of those to 0, it isn't possible to do so
2692 // because of worklist visitation order. So ugly it is.
2693 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2694 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2695 Constant *Zero = Constant::getNullValue(Ty);
2696 SelectInst *NewSel =
2697 SelectInst::Create(C: Cond, S1: OtherHandOfSubIsTrueVal ? Zero : NewSub,
2698 S2: OtherHandOfSubIsTrueVal ? NewSub : Zero);
2699 // Preserve prof metadata if any.
2700 NewSel->copyMetadata(SrcInst: cast<Instruction>(Val&: *Select));
2701 return NewSel;
2702 };
2703 if (Instruction *NewSel = SinkSubIntoSelect(
2704 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2705 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2706 return Builder->CreateSub(LHS: OtherHandOfSelect,
2707 /*OtherHandOfSub=*/RHS: Op1);
2708 }))
2709 return NewSel;
2710 if (Instruction *NewSel = SinkSubIntoSelect(
2711 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2712 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2713 return Builder->CreateSub(/*OtherHandOfSub=*/LHS: Op0,
2714 RHS: OtherHandOfSelect);
2715 }))
2716 return NewSel;
2717 }
2718
2719 // (X - (X & Y)) --> (X & ~Y)
2720 if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value(V&: Y))) &&
2721 (Op1->hasOneUse() || isa<Constant>(Val: Y)))
2722 return BinaryOperator::CreateAnd(
2723 V1: Op0, V2: Builder.CreateNot(V: Y, Name: Y->getName() + ".not"));
2724
2725 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2726 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2727 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2728 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2729 // As long as Y is freely invertible, this will be neutral or a win.
2730 // Note: We don't generate the inverse max/min, just create the 'not' of
2731 // it and let other folds do the rest.
2732 if (match(V: Op0, P: m_Not(V: m_Value(V&: X))) &&
2733 match(V: Op1, P: m_c_MaxOrMin(L: m_Specific(V: Op0), R: m_Value(V&: Y))) &&
2734 !Op0->hasNUsesOrMore(N: 3) && isFreeToInvert(V: Y, WillInvertAllUses: Y->hasOneUse())) {
2735 Value *Not = Builder.CreateNot(V: Op1);
2736 return BinaryOperator::CreateSub(V1: Not, V2: X);
2737 }
2738 if (match(V: Op1, P: m_Not(V: m_Value(V&: X))) &&
2739 match(V: Op0, P: m_c_MaxOrMin(L: m_Specific(V: Op1), R: m_Value(V&: Y))) &&
2740 !Op1->hasNUsesOrMore(N: 3) && isFreeToInvert(V: Y, WillInvertAllUses: Y->hasOneUse())) {
2741 Value *Not = Builder.CreateNot(V: Op0);
2742 return BinaryOperator::CreateSub(V1: X, V2: Not);
2743 }
2744
2745 // min(X+1, Y) - min(X, Y) --> zext X < Y
2746 // Replacing a sub and at least one min with an icmp
2747 // and a zext is a potential improvement.
2748 if (match(V: Op0, P: m_c_SMin(L: m_NSWAddLike(L: m_Value(V&: X), R: m_One()), R: m_Value(V&: Y))) &&
2749 match(V: Op1, P: m_c_SMin(L: m_Specific(V: X), R: m_Specific(V: Y))) &&
2750 I.getType()->getScalarSizeInBits() != 1 &&
2751 (Op0->hasOneUse() || Op1->hasOneUse())) {
2752 Value *Cond = Builder.CreateICmpSLT(LHS: X, RHS: Y);
2753 return new ZExtInst(Cond, I.getType());
2754 }
2755 if (match(V: Op0, P: m_c_UMin(L: m_NUWAddLike(L: m_Value(V&: X), R: m_One()), R: m_Value(V&: Y))) &&
2756 match(V: Op1, P: m_c_UMin(L: m_Specific(V: X), R: m_Specific(V: Y))) &&
2757 I.getType()->getScalarSizeInBits() != 1 &&
2758 (Op0->hasOneUse() || Op1->hasOneUse())) {
2759 Value *Cond = Builder.CreateICmpULT(LHS: X, RHS: Y);
2760 return new ZExtInst(Cond, I.getType());
2761 }
2762
2763 // Optimize pointer differences into the same array into a size. Consider:
2764 // &A[10] - &A[0]: we should compile this to "10".
2765 Value *LHSOp, *RHSOp;
2766 if (match(V: Op0, P: m_PtrToIntOrAddr(Op: m_Value(V&: LHSOp))) &&
2767 match(V: Op1, P: m_PtrToIntOrAddr(Op: m_Value(V&: RHSOp))))
2768 if (Value *Res = OptimizePointerDifference(LHS: LHSOp, RHS: RHSOp, Ty: I.getType(),
2769 IsNUW: I.hasNoUnsignedWrap()))
2770 return replaceInstUsesWith(I, V: Res);
2771
2772 // trunc(p)-trunc(q) -> trunc(p-q)
2773 if (match(V: Op0, P: m_Trunc(Op: m_PtrToIntOrAddr(Op: m_Value(V&: LHSOp)))) &&
2774 match(V: Op1, P: m_Trunc(Op: m_PtrToIntOrAddr(Op: m_Value(V&: RHSOp)))))
2775 if (Value *Res = OptimizePointerDifference(LHS: LHSOp, RHS: RHSOp, Ty: I.getType(),
2776 /* IsNUW */ false))
2777 return replaceInstUsesWith(I, V: Res);
2778
2779 auto MatchSubOfZExtOfPtrToIntOrAddr = [&]() {
2780 if (match(V: Op0, P: m_ZExt(Op: m_PtrToIntSameSize(DL, Op: m_Value(V&: LHSOp)))) &&
2781 match(V: Op1, P: m_ZExt(Op: m_PtrToIntSameSize(DL, Op: m_Value(V&: RHSOp)))))
2782 return true;
2783 if (match(V: Op0, P: m_ZExt(Op: m_PtrToAddr(Op: m_Value(V&: LHSOp)))) &&
2784 match(V: Op1, P: m_ZExt(Op: m_PtrToAddr(Op: m_Value(V&: RHSOp)))))
2785 return true;
2786 // Special case for non-canonical ptrtoint in constant expression,
2787 // where the zext has been folded into the ptrtoint.
2788 if (match(V: Op0, P: m_ZExt(Op: m_PtrToIntSameSize(DL, Op: m_Value(V&: LHSOp)))) &&
2789 match(V: Op1, P: m_PtrToInt(Op: m_Value(V&: RHSOp))))
2790 return true;
2791 return false;
2792 };
2793 if (MatchSubOfZExtOfPtrToIntOrAddr()) {
2794 if (auto *GEP = dyn_cast<GEPOperator>(Val: LHSOp)) {
2795 if (GEP->getPointerOperand() == RHSOp) {
2796 if (GEP->hasNoUnsignedWrap() || GEP->hasNoUnsignedSignedWrap()) {
2797 Value *Offset = EmitGEPOffset(GEP);
2798 Value *Res = GEP->hasNoUnsignedWrap()
2799 ? Builder.CreateZExt(
2800 V: Offset, DestTy: I.getType(), Name: "",
2801 /*IsNonNeg=*/GEP->hasNoUnsignedSignedWrap())
2802 : Builder.CreateSExt(V: Offset, DestTy: I.getType());
2803 return replaceInstUsesWith(I, V: Res);
2804 }
2805 }
2806 }
2807 }
2808
2809 // Canonicalize a shifty way to code absolute value to the common pattern.
2810 // There are 2 potential commuted variants.
2811 // We're relying on the fact that we only do this transform when the shift has
2812 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2813 // instructions).
2814 Value *A;
2815 const APInt *ShAmt;
2816 Type *Ty = I.getType();
2817 unsigned BitWidth = Ty->getScalarSizeInBits();
2818 if (match(V: Op1, P: m_AShr(L: m_Value(V&: A), R: m_APInt(Res&: ShAmt))) &&
2819 Op1->hasNUses(N: 2) && *ShAmt == BitWidth - 1 &&
2820 match(V: Op0, P: m_OneUse(SubPattern: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: Op1))))) {
2821 // B = ashr i32 A, 31 ; smear the sign bit
2822 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2823 // --> (A < 0) ? -A : A
2824 Value *IsNeg = Builder.CreateIsNeg(Arg: A);
2825 // Copy the nsw flags from the sub to the negate.
2826 Value *NegA = I.hasNoUnsignedWrap()
2827 ? Constant::getNullValue(Ty: A->getType())
2828 : Builder.CreateNeg(V: A, Name: "", HasNSW: I.hasNoSignedWrap());
2829 return SelectInst::Create(C: IsNeg, S1: NegA, S2: A);
2830 }
2831
2832 // If we are subtracting a low-bit masked subset of some value from an add
2833 // of that same value with no low bits changed, that is clearing some low bits
2834 // of the sum:
2835 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2836 const APInt *AddC, *AndC;
2837 if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: AddC))) &&
2838 match(V: Op1, P: m_And(L: m_Specific(V: X), R: m_APInt(Res&: AndC)))) {
2839 unsigned Cttz = AddC->countr_zero();
2840 APInt HighMask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - Cttz));
2841 if ((HighMask & *AndC).isZero())
2842 return BinaryOperator::CreateAnd(V1: Op0, V2: ConstantInt::get(Ty, V: ~(*AndC)));
2843 }
2844
2845 if (Instruction *V =
2846 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2847 return V;
2848
2849 // X - usub.sat(X, Y) => umin(X, Y)
2850 if (match(V: Op1, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::usub_sat>(Op0: m_Specific(V: Op0),
2851 Op1: m_Value(V&: Y)))))
2852 return replaceInstUsesWith(
2853 I, V: Builder.CreateIntrinsic(ID: Intrinsic::umin, Types: {I.getType()}, Args: {Op0, Y}));
2854
2855 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2856 // TODO: The one-use restriction is not strictly necessary, but it may
2857 // require improving other pattern matching and/or codegen.
2858 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_UMax(L: m_Value(V&: X), R: m_Specific(V: Op1)))))
2859 return replaceInstUsesWith(
2860 I, V: Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, Types: {Ty}, Args: {X, Op1}));
2861
2862 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2863 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_UMin(L: m_Value(V&: X), R: m_Specific(V: Op0)))))
2864 return replaceInstUsesWith(
2865 I, V: Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, Types: {Ty}, Args: {Op0, X}));
2866
2867 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2868 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_UMax(L: m_Value(V&: X), R: m_Specific(V: Op0))))) {
2869 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, Types: {Ty}, Args: {X, Op0});
2870 return BinaryOperator::CreateNeg(Op: USub);
2871 }
2872
2873 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2874 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_UMin(L: m_Value(V&: X), R: m_Specific(V: Op1))))) {
2875 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, Types: {Ty}, Args: {Op1, X});
2876 return BinaryOperator::CreateNeg(Op: USub);
2877 }
2878
2879 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2880 if (match(V: Op0, P: m_SpecificInt(V: BitWidth)) &&
2881 match(V: Op1, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: X)))))
2882 return replaceInstUsesWith(
2883 I, V: Builder.CreateIntrinsic(ID: Intrinsic::ctpop, Types: {I.getType()},
2884 Args: {Builder.CreateNot(V: X)}));
2885
2886 // Reduce multiplies for difference-of-squares by factoring:
2887 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
2888 if (match(V: Op0, P: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: X), R: m_Deferred(V: X)))) &&
2889 match(V: Op1, P: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: Y), R: m_Deferred(V: Y))))) {
2890 auto *OBO0 = cast<OverflowingBinaryOperator>(Val: Op0);
2891 auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1);
2892 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
2893 OBO1->hasNoSignedWrap() && BitWidth > 2;
2894 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
2895 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
2896 Value *Add = Builder.CreateAdd(LHS: X, RHS: Y, Name: "add", HasNUW: PropagateNUW, HasNSW: PropagateNSW);
2897 Value *Sub = Builder.CreateSub(LHS: X, RHS: Y, Name: "sub", HasNUW: PropagateNUW, HasNSW: PropagateNSW);
2898 Value *Mul = Builder.CreateMul(LHS: Add, RHS: Sub, Name: "", HasNUW: PropagateNUW, HasNSW: PropagateNSW);
2899 return replaceInstUsesWith(I, V: Mul);
2900 }
2901
2902 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
2903 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_SMax(L: m_Value(V&: X), R: m_Value(V&: Y)))) &&
2904 match(V: Op1, P: m_OneUse(SubPattern: m_c_SMin(L: m_Specific(V: X), R: m_Specific(V: Y))))) {
2905 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
2906 Value *Sub =
2907 Builder.CreateSub(LHS: X, RHS: Y, Name: "sub", /*HasNUW=*/false, /*HasNSW=*/true);
2908 Value *Call =
2909 Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: Sub, RHS: Builder.getTrue());
2910 return replaceInstUsesWith(I, V: Call);
2911 }
2912 }
2913
2914 if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
2915 return Res;
2916
2917 // (sub (sext (add nsw (X, Y)), sext (X))) --> (sext (Y))
2918 if (match(V: Op1, P: m_SExtLike(Op: m_Value(V&: X))) &&
2919 match(V: Op0, P: m_SExtLike(Op: m_c_NSWAdd(L: m_Specific(V: X), R: m_Value(V&: Y))))) {
2920 Value *SExtY = Builder.CreateSExt(V: Y, DestTy: I.getType());
2921 return replaceInstUsesWith(I, V: SExtY);
2922 }
2923
2924 // (sub[ nsw] (sext (add nsw (X, Y)), sext (add nsw (X, Z)))) -->
2925 // --> (sub[ nsw] (sext (Y), sext (Z)))
2926 {
2927 Value *Z, *Add0, *Add1;
2928 if (match(V: Op0, P: m_SExtLike(Op: m_Value(V&: Add0))) &&
2929 match(V: Op1, P: m_SExtLike(Op: m_Value(V&: Add1))) &&
2930 ((match(V: Add0, P: m_NSWAdd(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
2931 match(V: Add1, P: m_c_NSWAdd(L: m_Specific(V: X), R: m_Value(V&: Z)))) ||
2932 (match(V: Add0, P: m_NSWAdd(L: m_Value(V&: Y), R: m_Value(V&: X))) &&
2933 match(V: Add1, P: m_c_NSWAdd(L: m_Specific(V: X), R: m_Value(V&: Z)))))) {
2934 unsigned NumOfNewInstrs = 0;
2935 // Non-constant Y, Z require new SExt.
2936 NumOfNewInstrs += !isa<Constant>(Val: Y) ? 1 : 0;
2937 NumOfNewInstrs += !isa<Constant>(Val: Z) ? 1 : 0;
2938 // Check if we can trade some of the old instructions for the new ones.
2939 unsigned NumOfDeadInstrs = 0;
2940 if (Op0->hasOneUse()) {
2941 // If Op0 (sext) has multiple uses, then we keep it
2942 // and the add that it uses, otherwise, we can remove
2943 // the sext and probably the add (depending on the number of its uses).
2944 ++NumOfDeadInstrs;
2945 NumOfDeadInstrs += Add0->hasOneUse() ? 1 : 0;
2946 }
2947 if (Op1->hasOneUse()) {
2948 ++NumOfDeadInstrs;
2949 NumOfDeadInstrs += Add1->hasOneUse() ? 1 : 0;
2950 }
2951 if (NumOfDeadInstrs >= NumOfNewInstrs) {
2952 Value *SExtY = Builder.CreateSExt(V: Y, DestTy: I.getType());
2953 Value *SExtZ = Builder.CreateSExt(V: Z, DestTy: I.getType());
2954 Value *Sub = Builder.CreateSub(LHS: SExtY, RHS: SExtZ, Name: "",
2955 /*HasNUW=*/false,
2956 /*HasNSW=*/I.hasNoSignedWrap());
2957 return replaceInstUsesWith(I, V: Sub);
2958 }
2959 }
2960 }
2961
2962 return TryToNarrowDeduceFlags();
2963}
2964
2965/// This eliminates floating-point negation in either 'fneg(X)' or
2966/// 'fsub(-0.0, X)' form by combining into a constant operand.
2967static Instruction *foldFNegIntoConstant(Instruction &I, const DataLayout &DL) {
2968 // This is limited with one-use because fneg is assumed better for
2969 // reassociation and cheaper in codegen than fmul/fdiv.
2970 // TODO: Should the m_OneUse restriction be removed?
2971 Instruction *FNegOp;
2972 if (!match(V: &I, P: m_FNeg(X: m_OneUse(SubPattern: m_Instruction(I&: FNegOp)))))
2973 return nullptr;
2974
2975 Value *X;
2976 Constant *C;
2977
2978 // Fold negation into constant operand.
2979 // -(X * C) --> X * (-C)
2980 if (match(V: FNegOp, P: m_FMul(L: m_Value(V&: X), R: m_Constant(C))))
2981 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) {
2982 FastMathFlags FNegF = I.getFastMathFlags();
2983 FastMathFlags OpF = FNegOp->getFastMathFlags();
2984 FastMathFlags FMF = FastMathFlags::unionValue(LHS: FNegF, RHS: OpF) |
2985 FastMathFlags::intersectRewrite(LHS: FNegF, RHS: OpF);
2986 FMF.setNoInfs(FNegF.noInfs() && OpF.noInfs());
2987 return BinaryOperator::CreateFMulFMF(V1: X, V2: NegC, FMF);
2988 }
2989 // -(X / C) --> X / (-C)
2990 if (match(V: FNegOp, P: m_FDiv(L: m_Value(V&: X), R: m_Constant(C)))) {
2991 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) {
2992 Instruction *FDiv = BinaryOperator::CreateFDivFMF(V1: X, V2: NegC, FMFSource: &I);
2993 FDiv->copyMetadata(SrcInst: *FNegOp);
2994 return FDiv;
2995 }
2996 }
2997 // -(C / X) --> (-C) / X
2998 if (match(V: FNegOp, P: m_FDiv(L: m_Constant(C), R: m_Value(V&: X))))
2999 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) {
3000 Instruction *FDiv = BinaryOperator::CreateFDivFMF(V1: NegC, V2: X, FMFSource: &I);
3001
3002 // Intersect 'nsz' and 'ninf' because those special value exceptions may
3003 // not apply to the fdiv. Everything else propagates from the fneg.
3004 // TODO: We could propagate nsz/ninf from fdiv alone?
3005 FastMathFlags FMF = I.getFastMathFlags();
3006 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
3007 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
3008 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
3009 FDiv->copyMetadata(SrcInst: *FNegOp);
3010 return FDiv;
3011 }
3012 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
3013 // -(X + C) --> -X + -C --> -C - X
3014 if (I.hasNoSignedZeros() && match(V: FNegOp, P: m_FAdd(L: m_Value(V&: X), R: m_Constant(C))))
3015 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL))
3016 return BinaryOperator::CreateFSubFMF(V1: NegC, V2: X, FMFSource: &I);
3017
3018 return nullptr;
3019}
3020
3021Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
3022 Instruction &FMFSource) {
3023 Value *X, *Y;
3024 if (match(V: FNegOp, P: m_FMul(L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3025 // Push into RHS which is more likely to simplify (const or another fneg).
3026 // FIXME: It would be better to invert the transform.
3027 return cast<Instruction>(Val: Builder.CreateFMulFMF(
3028 L: X, R: Builder.CreateFNegFMF(V: Y, FMFSource: &FMFSource), FMFSource: &FMFSource));
3029 }
3030
3031 if (match(V: FNegOp, P: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3032 auto *FDiv = cast<Instruction>(Val: Builder.CreateFDivFMF(
3033 L: Builder.CreateFNegFMF(V: X, FMFSource: &FMFSource), R: Y, FMFSource: &FMFSource));
3034 FDiv->copyMetadata(SrcInst: *cast<Instruction>(Val: FNegOp));
3035 return FDiv;
3036 }
3037
3038 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: FNegOp)) {
3039 // Make sure to preserve flags and metadata on the call.
3040 if (II->getIntrinsicID() == Intrinsic::ldexp) {
3041 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
3042 CallInst *New =
3043 Builder.CreateCall(Callee: II->getCalledFunction(),
3044 Args: {Builder.CreateFNegFMF(V: II->getArgOperand(i: 0), FMFSource: FMF),
3045 II->getArgOperand(i: 1)});
3046 New->setFastMathFlags(FMF);
3047 New->copyMetadata(SrcInst: *II);
3048 return New;
3049 }
3050 }
3051
3052 return nullptr;
3053}
3054
3055Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
3056 Value *Op = I.getOperand(i_nocapture: 0);
3057
3058 if (Value *V = simplifyFNegInst(Op, FMF: I.getFastMathFlags(),
3059 Q: getSimplifyQuery().getWithInstruction(I: &I)))
3060 return replaceInstUsesWith(I, V);
3061
3062 if (Instruction *X = foldFNegIntoConstant(I, DL))
3063 return X;
3064
3065 Value *X, *Y;
3066
3067 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
3068 if (I.hasNoSignedZeros() &&
3069 match(V: Op, P: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y)))))
3070 return BinaryOperator::CreateFSubFMF(V1: Y, V2: X, FMFSource: &I);
3071
3072 Value *OneUse;
3073 if (!match(V: Op, P: m_OneUse(SubPattern: m_Value(V&: OneUse))))
3074 return nullptr;
3075
3076 if (Instruction *R = hoistFNegAboveFMulFDiv(FNegOp: OneUse, FMFSource&: I))
3077 return replaceInstUsesWith(I, V: R);
3078
3079 // Try to eliminate fneg if at least 1 arm of the select is negated.
3080 Value *Cond;
3081 if (match(V: OneUse, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3082 // Unlike most transforms, this one is not safe to propagate nsz unless
3083 // it is present on the original select. We union the flags from the select
3084 // and fneg and then remove nsz if needed.
3085 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
3086 S->copyFastMathFlags(I: &I);
3087 if (auto *OldSel = dyn_cast<SelectInst>(Val: Op)) {
3088 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
3089 S->setFastMathFlags(FMF);
3090 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
3091 !isGuaranteedNotToBeUndefOrPoison(V: OldSel->getCondition()))
3092 S->setHasNoSignedZeros(false);
3093 }
3094 };
3095 // -(Cond ? -P : Y) --> Cond ? P : -Y
3096 Value *P;
3097 if (match(V: X, P: m_FNeg(X: m_Value(V&: P)))) {
3098 Value *NegY = Builder.CreateFNegFMF(V: Y, FMFSource: &I, Name: Y->getName() + ".neg");
3099 SelectInst *NewSel = SelectInst::Create(C: Cond, S1: P, S2: NegY);
3100 propagateSelectFMF(NewSel, P == Y);
3101 return NewSel;
3102 }
3103 // -(Cond ? X : -P) --> Cond ? -X : P
3104 if (match(V: Y, P: m_FNeg(X: m_Value(V&: P)))) {
3105 Value *NegX = Builder.CreateFNegFMF(V: X, FMFSource: &I, Name: X->getName() + ".neg");
3106 SelectInst *NewSel = SelectInst::Create(C: Cond, S1: NegX, S2: P);
3107 propagateSelectFMF(NewSel, P == X);
3108 return NewSel;
3109 }
3110
3111 // -(Cond ? X : C) --> Cond ? -X : -C
3112 // -(Cond ? C : Y) --> Cond ? -C : -Y
3113 if (match(V: X, P: m_ImmConstant()) || match(V: Y, P: m_ImmConstant())) {
3114 Value *NegX = Builder.CreateFNegFMF(V: X, FMFSource: &I, Name: X->getName() + ".neg");
3115 Value *NegY = Builder.CreateFNegFMF(V: Y, FMFSource: &I, Name: Y->getName() + ".neg");
3116 SelectInst *NewSel = SelectInst::Create(C: Cond, S1: NegX, S2: NegY);
3117 propagateSelectFMF(NewSel, /*CommonOperand=*/true);
3118 return NewSel;
3119 }
3120 }
3121
3122 // fneg (copysign x, y) -> copysign x, (fneg y)
3123 if (match(V: OneUse, P: m_CopySign(Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) {
3124 // The source copysign has an additional value input, so we can't propagate
3125 // flags the copysign doesn't also have.
3126 FastMathFlags FMF = I.getFastMathFlags();
3127 FMF &= cast<FPMathOperator>(Val: OneUse)->getFastMathFlags();
3128 Value *NegY = Builder.CreateFNegFMF(V: Y, FMFSource: FMF);
3129 Value *NewCopySign = Builder.CreateCopySign(LHS: X, RHS: NegY, FMFSource: FMF);
3130 return replaceInstUsesWith(I, V: NewCopySign);
3131 }
3132
3133 // fneg (shuffle x, Mask) --> shuffle (fneg x), Mask
3134 ArrayRef<int> Mask;
3135 if (match(V: OneUse, P: m_Shuffle(v1: m_Value(V&: X), v2: m_Poison(), mask: m_Mask(Mask))))
3136 return new ShuffleVectorInst(Builder.CreateFNegFMF(V: X, FMFSource: &I), Mask);
3137
3138 // fneg (reverse x) --> reverse (fneg x)
3139 if (match(V: OneUse, P: m_VecReverse(Op0: m_Value(V&: X)))) {
3140 Value *Reverse = Builder.CreateVectorReverse(V: Builder.CreateFNegFMF(V: X, FMFSource: &I));
3141 return replaceInstUsesWith(I, V: Reverse);
3142 }
3143
3144 return nullptr;
3145}
3146
3147Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
3148 if (Value *V = simplifyFSubInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
3149 FMF: I.getFastMathFlags(),
3150 Q: getSimplifyQuery().getWithInstruction(I: &I)))
3151 return replaceInstUsesWith(I, V);
3152
3153 if (Instruction *X = foldVectorBinop(Inst&: I))
3154 return X;
3155
3156 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
3157 return Phi;
3158
3159 // Subtraction from -0.0 is the canonical form of fneg.
3160 // fsub -0.0, X ==> fneg X
3161 // fsub nsz 0.0, X ==> fneg nsz X
3162 //
3163 // FIXME This matcher does not respect FTZ or DAZ yet:
3164 // fsub -0.0, Denorm ==> +-0
3165 // fneg Denorm ==> -Denorm
3166 Value *Op;
3167 if (match(V: &I, P: m_FNeg(X: m_Value(V&: Op))))
3168 return UnaryOperator::CreateFNegFMF(Op, FMFSource: &I);
3169
3170 if (Instruction *X = foldFNegIntoConstant(I, DL))
3171 return X;
3172
3173 if (Instruction *R = foldFBinOpOfIntCasts(I))
3174 return R;
3175
3176 Value *X, *Y;
3177 Constant *C;
3178
3179 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
3180 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
3181 // Canonicalize to fadd to make analysis easier.
3182 // This can also help codegen because fadd is commutative.
3183 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
3184 // killed later. We still limit that particular transform with 'hasOneUse'
3185 // because an fneg is assumed better/cheaper than a generic fsub.
3186 if (I.hasNoSignedZeros() ||
3187 cannotBeNegativeZero(V: Op0, SQ: getSimplifyQuery().getWithInstruction(I: &I))) {
3188 if (match(V: Op1, P: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
3189 Value *NewSub = Builder.CreateFSubFMF(L: Y, R: X, FMFSource: &I);
3190 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: NewSub, FMFSource: &I);
3191 }
3192 }
3193
3194 // (-X) - Op1 --> -(X + Op1)
3195 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Val: Op0) &&
3196 match(V: Op0, P: m_OneUse(SubPattern: m_FNeg(X: m_Value(V&: X))))) {
3197 Value *FAdd = Builder.CreateFAddFMF(L: X, R: Op1, FMFSource: &I);
3198 return UnaryOperator::CreateFNegFMF(Op: FAdd, FMFSource: &I);
3199 }
3200
3201 if (isa<Constant>(Val: Op0))
3202 if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1))
3203 if (Instruction *NV = FoldOpIntoSelect(Op&: I, SI))
3204 return NV;
3205
3206 // X - C --> X + (-C)
3207 // But don't transform constant expressions because there's an inverse fold
3208 // for X + (-Y) --> X - Y.
3209 if (match(V: Op1, P: m_ImmConstant(C)))
3210 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL))
3211 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: NegC, FMFSource: &I);
3212
3213 // X - (-Y) --> X + Y
3214 if (match(V: Op1, P: m_FNeg(X: m_Value(V&: Y))))
3215 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: Y, FMFSource: &I);
3216
3217 // Similar to above, but look through a cast of the negated value:
3218 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
3219 Type *Ty = I.getType();
3220 if (match(V: Op1, P: m_OneUse(SubPattern: m_FPTrunc(Op: m_FNeg(X: m_Value(V&: Y))))))
3221 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: Builder.CreateFPTrunc(V: Y, DestTy: Ty), FMFSource: &I);
3222
3223 // X - (fpext(-Y)) --> X + fpext(Y)
3224 if (match(V: Op1, P: m_OneUse(SubPattern: m_FPExt(Op: m_FNeg(X: m_Value(V&: Y))))))
3225 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: Builder.CreateFPExt(V: Y, DestTy: Ty), FMFSource: &I);
3226
3227 // Similar to above, but look through fmul/fdiv of the negated value:
3228 // Op0 - (-X * Y) --> Op0 + (X * Y)
3229 // Op0 - (Y * -X) --> Op0 + (X * Y)
3230 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_FMul(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y))))) {
3231 Value *FMul = Builder.CreateFMulFMF(L: X, R: Y, FMFSource: &I);
3232 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: FMul, FMFSource: &I);
3233 }
3234 // Op0 - (-X / Y) --> Op0 + (X / Y)
3235 // Op0 - (X / -Y) --> Op0 + (X / Y)
3236 if (match(V: Op1, P: m_OneUse(SubPattern: m_FDiv(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y)))) ||
3237 match(V: Op1, P: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_FNeg(X: m_Value(V&: Y)))))) {
3238 Value *FDiv = Builder.CreateFDivFMF(L: X, R: Y, FMFSource: &I);
3239 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: FDiv, FMFSource: &I);
3240 }
3241
3242 // Handle special cases for FSub with selects feeding the operation
3243 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS: Op0, RHS: Op1))
3244 return replaceInstUsesWith(I, V);
3245
3246 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
3247 // (Y - X) - Y --> -X
3248 if (match(V: Op0, P: m_FSub(L: m_Specific(V: Op1), R: m_Value(V&: X))))
3249 return UnaryOperator::CreateFNegFMF(Op: X, FMFSource: &I);
3250
3251 // Y - (X + Y) --> -X
3252 // Y - (Y + X) --> -X
3253 if (match(V: Op1, P: m_c_FAdd(L: m_Specific(V: Op0), R: m_Value(V&: X))))
3254 return UnaryOperator::CreateFNegFMF(Op: X, FMFSource: &I);
3255
3256 // (X * C) - X --> X * (C - 1.0)
3257 if (match(V: Op0, P: m_FMul(L: m_Specific(V: Op1), R: m_Constant(C)))) {
3258 if (Constant *CSubOne = ConstantFoldBinaryOpOperands(
3259 Opcode: Instruction::FSub, LHS: C, RHS: ConstantFP::get(Ty, V: 1.0), DL))
3260 return BinaryOperator::CreateFMulFMF(V1: Op1, V2: CSubOne, FMFSource: &I);
3261 }
3262 // X - (X * C) --> X * (1.0 - C)
3263 if (match(V: Op1, P: m_FMul(L: m_Specific(V: Op0), R: m_Constant(C)))) {
3264 if (Constant *OneSubC = ConstantFoldBinaryOpOperands(
3265 Opcode: Instruction::FSub, LHS: ConstantFP::get(Ty, V: 1.0), RHS: C, DL))
3266 return BinaryOperator::CreateFMulFMF(V1: Op0, V2: OneSubC, FMFSource: &I);
3267 }
3268
3269 // Reassociate fsub/fadd sequences to create more fadd instructions and
3270 // reduce dependency chains:
3271 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
3272 Value *Z;
3273 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_FAdd(L: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y))),
3274 R: m_Value(V&: Z))))) {
3275 Value *XZ = Builder.CreateFAddFMF(L: X, R: Z, FMFSource: &I);
3276 Value *YW = Builder.CreateFAddFMF(L: Y, R: Op1, FMFSource: &I);
3277 return BinaryOperator::CreateFSubFMF(V1: XZ, V2: YW, FMFSource: &I);
3278 }
3279
3280 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
3281 return m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_fadd>(Op0: m_Value(V&: Sum),
3282 Op1: m_Value(V&: Vec)));
3283 };
3284 Value *A0, *A1, *V0, *V1;
3285 if (match(V: Op0, P: m_FaddRdx(A0, V0)) && match(V: Op1, P: m_FaddRdx(A1, V1)) &&
3286 V0->getType() == V1->getType()) {
3287 // Difference of sums is sum of differences:
3288 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
3289 Value *Sub = Builder.CreateFSubFMF(L: V0, R: V1, FMFSource: &I);
3290 Value *Rdx = Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_fadd,
3291 Types: {Sub->getType()}, Args: {A0, Sub}, FMFSource: &I);
3292 return BinaryOperator::CreateFSubFMF(V1: Rdx, V2: A1, FMFSource: &I);
3293 }
3294
3295 if (Instruction *F = factorizeFAddFSub(I, Builder))
3296 return F;
3297
3298 // TODO: This performs reassociative folds for FP ops. Some fraction of the
3299 // functionality has been subsumed by simple pattern matching here and in
3300 // InstSimplify. We should let a dedicated reassociation pass handle more
3301 // complex pattern matching and remove this from InstCombine.
3302 if (Value *V = FAddCombine(Builder).simplify(I: &I))
3303 return replaceInstUsesWith(I, V);
3304
3305 // (X - Y) - Op1 --> X - (Y + Op1)
3306 if (match(V: Op0, P: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
3307 Value *FAdd = Builder.CreateFAddFMF(L: Y, R: Op1, FMFSource: &I);
3308 return BinaryOperator::CreateFSubFMF(V1: X, V2: FAdd, FMFSource: &I);
3309 }
3310 }
3311
3312 return nullptr;
3313}
3314