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