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 && "Incorrect state");
93 return *getFpValPtr();
94 }
95
96 APFloat &getFpVal() {
97 assert(IsFp && BufHasFpVal && "Incorrect state");
98 return *getFpValPtr();
99 }
100
101 bool isInt() const { return !IsFp; }
102
103 // If the coefficient is represented by an integer, promote it to a
104 // floating point.
105 void convertToFpType(const fltSemantics &Sem);
106
107 // Construct an APFloat from a signed integer.
108 // TODO: We should get rid of this function when APFloat can be constructed
109 // from an *SIGNED* integer.
110 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
111
112 bool IsFp = false;
113
114 // True iff FpValBuf contains an instance of APFloat.
115 bool BufHasFpVal = false;
116
117 // The integer coefficient of an individual addend is either 1 or -1,
118 // and we try to simplify at most 4 addends from neighboring at most
119 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120 // is overkill of this end.
121 short IntVal = 0;
122
123 AlignedCharArrayUnion<APFloat> FpValBuf;
124 };
125
126 /// FAddend is used to represent floating-point addend. An addend is
127 /// represented as <C, V>, where the V is a symbolic value, and C is a
128 /// constant coefficient. A constant addend is represented as <C, 0>.
129 class FAddend {
130 public:
131 FAddend() = default;
132
133 void operator+=(const FAddend &T) {
134 assert((Val == T.Val) && "Symbolic-values disagree");
135 Coeff += T.Coeff;
136 }
137
138 Value *getSymVal() const { return Val; }
139 const FAddendCoef &getCoef() const { return Coeff; }
140
141 bool isConstant() const { return Val == nullptr; }
142 bool isZero() const { return Coeff.isZero(); }
143
144 void set(short Coefficient, Value *V) {
145 Coeff.set(Coefficient);
146 Val = V;
147 }
148 void set(const APFloat &Coefficient, Value *V) {
149 Coeff.set(Coefficient);
150 Val = V;
151 }
152 void set(const ConstantFP *Coefficient, Value *V) {
153 Coeff.set(Coefficient->getValueAPF());
154 Val = V;
155 }
156
157 void negate() { Coeff.negate(); }
158
159 /// Drill down the U-D chain one step to find the definition of V, and
160 /// try to break the definition into one or two addends.
161 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
162
163 /// Similar to FAddend::drillDownOneStep() except that the value being
164 /// splitted is the addend itself.
165 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
166
167 private:
168 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
169
170 // This addend has the value of "Coeff * Val".
171 Value *Val = nullptr;
172 FAddendCoef Coeff;
173 };
174
175 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176 /// with its neighboring at most two instructions.
177 ///
178 class FAddCombine {
179 public:
180 FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
181
182 Value *simplify(Instruction *FAdd);
183
184 private:
185 using AddendVect = SmallVector<const FAddend *, 4>;
186
187 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
188
189 /// Convert given addend to a Value
190 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
191
192 /// Return the number of instructions needed to emit the N-ary addition.
193 unsigned calcInstrNumber(const AddendVect& Vect);
194
195 Value *createFSub(Value *Opnd0, Value *Opnd1);
196 Value *createFAdd(Value *Opnd0, Value *Opnd1);
197 Value *createFMul(Value *Opnd0, Value *Opnd1);
198 Value *createFNeg(Value *V);
199 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
201
202 // Debugging stuff are clustered here.
203 #ifndef NDEBUG
204 unsigned CreateInstrNum;
205 void initCreateInstNum() { CreateInstrNum = 0; }
206 void incCreateInstNum() { CreateInstrNum++; }
207 #else
208 void initCreateInstNum() {}
209 void incCreateInstNum() {}
210 #endif
211
212 InstCombiner::BuilderTy &Builder;
213 Instruction *Instr = nullptr;
214 };
215
216} // end anonymous namespace
217
218//===----------------------------------------------------------------------===//
219//
220// Implementation of
221// {FAddendCoef, FAddend, FAddition, FAddCombine}.
222//
223//===----------------------------------------------------------------------===//
224FAddendCoef::~FAddendCoef() {
225 if (BufHasFpVal)
226 getFpValPtr()->~APFloat();
227}
228
229void FAddendCoef::set(const APFloat& C) {
230 APFloat *P = getFpValPtr();
231
232 if (isInt()) {
233 // As the buffer is meanless byte stream, we cannot call
234 // APFloat::operator=().
235 new(P) APFloat(C);
236 } else
237 *P = C;
238
239 IsFp = BufHasFpVal = true;
240}
241
242void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
243 if (!isInt())
244 return;
245
246 APFloat *P = getFpValPtr();
247 if (IntVal > 0)
248 new(P) APFloat(Sem, IntVal);
249 else {
250 new(P) APFloat(Sem, 0 - IntVal);
251 P->changeSign();
252 }
253 IsFp = BufHasFpVal = true;
254}
255
256APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
257 if (Val >= 0)
258 return APFloat(Sem, Val);
259
260 APFloat T(Sem, 0 - Val);
261 T.changeSign();
262
263 return T;
264}
265
266void FAddendCoef::operator=(const FAddendCoef &That) {
267 if (That.isInt())
268 set(That.IntVal);
269 else
270 set(That.getFpVal());
271}
272
273void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
276 if (isInt())
277 IntVal += That.IntVal;
278 else
279 getFpVal().add(RHS: That.getFpVal(), RM: RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(Sem: T.getSemantics());
286 getFpVal().add(RHS: T, RM: RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(RHS: createAPFloatFromInt(Sem: T.getSemantics(), Val: That.IntVal), RM: RndMode);
292}
293
294void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value");
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Sem: Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(RHS: createAPFloatFromInt(Sem: Semantic, Val: That.IntVal),
319 RM: APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(RHS: That.getFpVal(), RM: APFloat::rmNearestTiesToEven);
322}
323
324void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329}
330
331Value *FAddendCoef::getValue(Type *Ty) const {
332 return isInt() ?
333 ConstantFP::get(Ty, V: float(IntVal)) :
334 ConstantFP::get(Context&: Ty->getContext(), V: getFpVal());
335}
336
337// The definition of <Val> Addends
338// =========================================
339// A + B <1, A>, <1,B>
340// A - B <1, A>, <1,B>
341// 0 - B <-1, B>
342// C * A, <C, A>
343// A + C <1, A> <C, NULL>
344// 0 +/- 0 <0, NULL> (corner case)
345//
346// Legend: A and B are not constant, C is constant
347unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
351 return 0;
352
353 unsigned Opcode = I->getOpcode();
354
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356 ConstantFP *C0, *C1;
357 Value *Opnd0 = I->getOperand(i: 0);
358 Value *Opnd1 = I->getOperand(i: 1);
359 if ((C0 = dyn_cast<ConstantFP>(Val: Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Val: Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(Coefficient: 1, V: Opnd0);
368 else
369 Addend0.set(Coefficient: C0, V: nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(Coefficient: 1, V: Opnd1);
376 else
377 Addend.set(Coefficient: C1, V: nullptr);
378 if (Opcode == Instruction::FSub)
379 Addend.negate();
380 }
381
382 if (Opnd0 || Opnd1)
383 return Opnd0 && Opnd1 ? 2 : 1;
384
385 // Both operands are zero. Weird!
386 Addend0.set(Coefficient: APFloat(C0->getValueAPF().getSemantics()), V: nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(i: 0);
392 Value *V1 = I->getOperand(i: 1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(Val: V0)) {
394 Addend0.set(Coefficient: C, V: V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(Val: V1)) {
399 Addend0.set(Coefficient: C, V: V0);
400 return 1;
401 }
402 }
403
404 return 0;
405}
406
407// Try to break *this* addend into two addends. e.g. Suppose this addend is
408// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409// i.e. <2.3, X> and <2.3, Y>.
410unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
412 if (isConstant())
413 return 0;
414
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
417 return BreakNum;
418
419 Addend0.Scale(ScaleAmt: Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(ScaleAmt: Coeff);
423
424 return BreakNum;
425}
426
427Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429 "Expected 'reassoc'+'nsz' instruction");
430
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
433 return nullptr;
434
435 assert((I->getOpcode() == Instruction::FAdd ||
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(Val: I, Addend0&: Opnd0, Addend1&: Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Addend0&: Opnd0_0, Addend1&: Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Addend0&: Opnd1_0, Addend1&: Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(Elt: &Opnd0_0);
460 AllOpnds.push_back(Elt: &Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(Elt: &Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(Elt: &Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(i: 0);
470 Value *V1 = I->getOperand(i: 1);
471 InstQuota = ((!isa<Constant>(Val: V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(Val: V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(V&: AllOpnds, InstrQuota: InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(Elt: &Opnd0);
491 AllOpnds.push_back(Elt: &Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(Elt: &Opnd1_1);
494
495 if (Value *R = simplifyFAdd(V&: AllOpnds, InstrQuota: 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(Elt: &Opnd1);
503 AllOpnds.push_back(Elt: &Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(Elt: &Opnd0_1);
506
507 if (Value *R = simplifyFAdd(V&: AllOpnds, InstrQuota: 1))
508 return R;
509 }
510
511 return nullptr;
512}
513
514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends");
517
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
521
522 // Simplified addends are placed <SimpVect>.
523 AddendVect SimpVect;
524
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529
530 const FAddend *ThisAddend = Addends[SymIdx];
531 if (!ThisAddend) {
532 // This addend was processed before.
533 continue;
534 }
535
536 Value *Val = ThisAddend->getSymVal();
537
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543 // call later.
544
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(Elt: ThisAddend);
547
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(Elt: T);
561 }
562 }
563
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569 R += *SimpVect[Idx];
570
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(N: StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(Elt: &R);
575 }
576 }
577 }
578
579 assert((NextTmpIdx <= std::size(TmpResult) + 1) && "out-of-bound access");
580
581 Value *Result;
582 if (!SimpVect.empty())
583 Result = createNaryFAdd(Opnds: SimpVect, InstrQuota);
584 else {
585 // The addition is folded to 0.0.
586 Result = ConstantFP::get(Ty: Instr->getType(), V: 0.0);
587 }
588
589 return Result;
590}
591
592Value *FAddCombine::createNaryFAdd
593 (const AddendVect &Opnds, unsigned InstrQuota) {
594 assert(!Opnds.empty() && "Expect at least one addend");
595
596 // Step 1: Check if the # of instructions needed exceeds the quota.
597
598 unsigned InstrNeeded = calcInstrNumber(Vect: Opnds);
599 if (InstrNeeded > InstrQuota)
600 return nullptr;
601
602 initCreateInstNum();
603
604 // step 2: Emit the N-ary addition.
605 // Note that at most three instructions are involved in Fadd-InstCombine: the
606 // addition in question, and at most two neighboring instructions.
607 // The resulting optimized addition should have at least one less instruction
608 // than the original addition expression tree. This implies that the resulting
609 // N-ary addition has at most two instructions, and we don't need to worry
610 // about tree-height when constructing the N-ary addition.
611
612 Value *LastVal = nullptr;
613 bool LastValNeedNeg = false;
614
615 // Iterate the addends, creating fadd/fsub using adjacent two addends.
616 for (const FAddend *Opnd : Opnds) {
617 bool NeedNeg;
618 Value *V = createAddendVal(A: *Opnd, NeedNeg);
619 if (!LastVal) {
620 LastVal = V;
621 LastValNeedNeg = NeedNeg;
622 continue;
623 }
624
625 if (LastValNeedNeg == NeedNeg) {
626 LastVal = createFAdd(Opnd0: LastVal, Opnd1: V);
627 continue;
628 }
629
630 if (LastValNeedNeg)
631 LastVal = createFSub(Opnd0: V, Opnd1: LastVal);
632 else
633 LastVal = createFSub(Opnd0: LastVal, Opnd1: V);
634
635 LastValNeedNeg = false;
636 }
637
638 if (LastValNeedNeg) {
639 LastVal = createFNeg(V: LastVal);
640 }
641
642#ifndef NDEBUG
643 assert(CreateInstrNum == InstrNeeded &&
644 "Inconsistent in instruction numbers");
645#endif
646
647 return LastVal;
648}
649
650Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
651 Value *V = Builder.CreateFSub(L: Opnd0, R: Opnd1);
652 if (Instruction *I = dyn_cast<Instruction>(Val: V))
653 createInstPostProc(NewInst: I);
654 return V;
655}
656
657Value *FAddCombine::createFNeg(Value *V) {
658 Value *NewV = Builder.CreateFNeg(V);
659 if (Instruction *I = dyn_cast<Instruction>(Val: NewV))
660 createInstPostProc(NewInst: I, NoNumber: true); // fneg's don't receive instruction numbers.
661 return NewV;
662}
663
664Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
665 Value *V = Builder.CreateFAdd(L: Opnd0, R: Opnd1);
666 if (Instruction *I = dyn_cast<Instruction>(Val: V))
667 createInstPostProc(NewInst: I);
668 return V;
669}
670
671Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672 Value *V = Builder.CreateFMul(L: Opnd0, R: Opnd1);
673 if (Instruction *I = dyn_cast<Instruction>(Val: V))
674 createInstPostProc(NewInst: I);
675 return V;
676}
677
678void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
679 NewInstr->setDebugLoc(Instr->getDebugLoc());
680
681 // Keep track of the number of instruction created.
682 if (!NoNumber)
683 incCreateInstNum();
684
685 // Propagate fast-math flags
686 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
687}
688
689// Return the number of instruction needed to emit the N-ary addition.
690// NOTE: Keep this function in sync with createAddendVal().
691unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
692 unsigned OpndNum = Opnds.size();
693 unsigned InstrNeeded = OpndNum - 1;
694
695 // Adjust the number of instructions needed to emit the N-ary add.
696 for (const FAddend *Opnd : Opnds) {
697 if (Opnd->isConstant())
698 continue;
699
700 // The constant check above is really for a few special constant
701 // coefficients.
702 if (isa<UndefValue>(Val: Opnd->getSymVal()))
703 continue;
704
705 const FAddendCoef &CE = Opnd->getCoef();
706 // Let the addend be "c * x". If "c == +/-1", the value of the addend
707 // is immediately available; otherwise, it needs exactly one instruction
708 // to evaluate the value.
709 if (!CE.isMinusOne() && !CE.isOne())
710 InstrNeeded++;
711 }
712 return InstrNeeded;
713}
714
715// Input Addend Value NeedNeg(output)
716// ================================================================
717// Constant C C false
718// <+/-1, V> V coefficient is -1
719// <2/-2, V> "fadd V, V" coefficient is -2
720// <C, V> "fmul V, C" false
721//
722// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
723Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
724 const FAddendCoef &Coeff = Opnd.getCoef();
725
726 if (Opnd.isConstant()) {
727 NeedNeg = false;
728 return Coeff.getValue(Ty: Instr->getType());
729 }
730
731 Value *OpndVal = Opnd.getSymVal();
732
733 if (Coeff.isMinusOne() || Coeff.isOne()) {
734 NeedNeg = Coeff.isMinusOne();
735 return OpndVal;
736 }
737
738 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
739 NeedNeg = Coeff.isMinusTwo();
740 return createFAdd(Opnd0: OpndVal, Opnd1: OpndVal);
741 }
742
743 NeedNeg = false;
744 return createFMul(Opnd0: OpndVal, Opnd1: Coeff.getValue(Ty: Instr->getType()));
745}
746
747// Checks if any operand is negative and we can convert add to sub.
748// This function checks for following negative patterns
749// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
750// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
751// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
752static Value *checkForNegativeOperand(BinaryOperator &I,
753 InstCombiner::BuilderTy &Builder) {
754 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
755
756 // This function creates 2 instructions to replace ADD, we need at least one
757 // of LHS or RHS to have one use to ensure benefit in transform.
758 if (!LHS->hasOneUse() && !RHS->hasOneUse())
759 return nullptr;
760
761 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
762 const APInt *C1 = nullptr, *C2 = nullptr;
763
764 // if ONE is on other side, swap
765 if (match(V: RHS, P: m_Add(L: m_Value(V&: X), R: m_One())))
766 std::swap(a&: LHS, b&: RHS);
767
768 if (match(V: LHS, P: m_Add(L: m_Value(V&: X), R: m_One()))) {
769 // if XOR on other side, swap
770 if (match(V: RHS, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1))))
771 std::swap(a&: X, b&: RHS);
772
773 if (match(V: X, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1)))) {
774 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
775 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
776 if (match(V: Y, P: m_Or(L: m_Value(V&: Z), R: m_APInt(Res&: C2))) && (*C2 == ~(*C1))) {
777 Value *NewAnd = Builder.CreateAnd(LHS: Z, RHS: *C1);
778 return Builder.CreateSub(LHS: RHS, RHS: NewAnd, Name: "sub");
779 } else if (match(V: Y, P: m_And(L: m_Value(V&: Z), R: m_APInt(Res&: C2))) && (*C1 == *C2)) {
780 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
781 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
782 Value *NewOr = Builder.CreateOr(LHS: Z, RHS: ~(*C1));
783 return Builder.CreateSub(LHS: RHS, RHS: NewOr, Name: "sub");
784 }
785 }
786 }
787
788 // Restore LHS and RHS
789 LHS = I.getOperand(i_nocapture: 0);
790 RHS = I.getOperand(i_nocapture: 1);
791
792 // if XOR is on other side, swap
793 if (match(V: RHS, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1))))
794 std::swap(a&: LHS, b&: RHS);
795
796 // C2 is ODD
797 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
798 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
799 if (match(V: LHS, P: m_Xor(L: m_Value(V&: Y), R: m_APInt(Res&: C1))))
800 if (C1->countr_zero() == 0)
801 if (match(V: Y, P: m_And(L: m_Value(V&: Z), R: m_APInt(Res&: C2))) && *C1 == (*C2 + 1)) {
802 Value *NewOr = Builder.CreateOr(LHS: Z, RHS: ~(*C2));
803 return Builder.CreateSub(LHS: RHS, RHS: NewOr, Name: "sub");
804 }
805 return nullptr;
806}
807
808/// Wrapping flags may allow combining constants separated by an extend.
809static Instruction *foldNoWrapAdd(BinaryOperator &Add,
810 InstCombiner::BuilderTy &Builder) {
811 Value *Op0 = Add.getOperand(i_nocapture: 0), *Op1 = Add.getOperand(i_nocapture: 1);
812 Type *Ty = Add.getType();
813 Constant *Op1C;
814 if (!match(V: Op1, P: m_Constant(C&: Op1C)))
815 return nullptr;
816
817 // Try this match first because it results in an add in the narrow type.
818 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
819 Value *X;
820 const APInt *C1, *C2;
821 if (match(V: Op1, P: m_APInt(Res&: C1)) &&
822 match(V: Op0, P: m_ZExt(Op: m_NUWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) &&
823 C1->isNegative() && C1->sge(RHS: -C2->sext(width: C1->getBitWidth()))) {
824 APInt NewC = *C2 + C1->trunc(width: C2->getBitWidth());
825 // If the smaller add will fold to zero, we don't need to check one use.
826 if (NewC.isZero())
827 return new ZExtInst(X, Ty);
828 // Otherwise only do this if the existing zero extend will be removed.
829 if (Op0->hasOneUse())
830 return new ZExtInst(
831 Builder.CreateNUWAdd(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: NewC)), Ty);
832 }
833
834 // More general combining of constants in the wide type.
835 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
836 // or (zext nneg (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
837 Constant *NarrowC;
838 if (match(V: Op0, P: m_OneUse(SubPattern: m_SExtLike(
839 Op: m_NSWAddLike(L: m_Value(V&: X), R: m_Constant(C&: NarrowC)))))) {
840 Value *WideC = Builder.CreateSExt(V: NarrowC, DestTy: Ty);
841 Value *NewC = Builder.CreateAdd(LHS: WideC, RHS: Op1C);
842 Value *WideX = Builder.CreateSExt(V: X, DestTy: Ty);
843 return BinaryOperator::CreateAdd(V1: WideX, V2: NewC);
844 }
845 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
846 if (match(V: Op0,
847 P: m_OneUse(SubPattern: m_ZExt(Op: m_NUWAddLike(L: m_Value(V&: X), R: m_Constant(C&: NarrowC)))))) {
848 Value *WideC = Builder.CreateZExt(V: NarrowC, DestTy: Ty);
849 Value *NewC = Builder.CreateAdd(LHS: WideC, RHS: Op1C);
850 Value *WideX = Builder.CreateZExt(V: X, DestTy: Ty);
851 return BinaryOperator::CreateAdd(V1: WideX, V2: NewC);
852 }
853 return nullptr;
854}
855
856Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
857 Value *Op0 = Add.getOperand(i_nocapture: 0), *Op1 = Add.getOperand(i_nocapture: 1);
858 Type *Ty = Add.getType();
859 Constant *Op1C;
860 if (!match(V: Op1, P: m_ImmConstant(C&: Op1C)))
861 return nullptr;
862
863 if (Instruction *NV = foldBinOpIntoSelectOrPhi(I&: Add))
864 return NV;
865
866 if (Instruction *FoldedLogic = foldBinOpSelectBinOp(Op&: Add))
867 return FoldedLogic;
868
869 Value *X;
870 Constant *Op00C;
871
872 // add (sub C1, X), C2 --> sub (add C1, C2), X
873 if (match(V: Op0, P: m_Sub(L: m_Constant(C&: Op00C), R: m_Value(V&: X))))
874 return BinaryOperator::CreateSub(V1: ConstantExpr::getAdd(C1: Op00C, C2: Op1C), V2: X);
875
876 Value *Y;
877
878 // add (sub X, Y), -1 --> add (not Y), X
879 if (match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y)))) &&
880 match(V: Op1, P: m_AllOnes()))
881 return BinaryOperator::CreateAdd(V1: Builder.CreateNot(V: Y), V2: X);
882
883 // zext(bool) + C -> bool ? C + 1 : C
884 if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) &&
885 X->getType()->getScalarSizeInBits() == 1)
886 return createSelectInstWithUnknownProfile(C: X, S1: InstCombiner::AddOne(C: Op1C),
887 S2: Op1);
888 // sext(bool) + C -> bool ? C - 1 : C
889 if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) &&
890 X->getType()->getScalarSizeInBits() == 1)
891 return createSelectInstWithUnknownProfile(C: X, S1: InstCombiner::SubOne(C: Op1C),
892 S2: Op1);
893
894 // ~X + C --> (C-1) - X
895 if (match(V: Op0, P: m_Not(V: m_Value(V&: X)))) {
896 // ~X + C has NSW and (C-1) won't oveflow => (C-1)-X can have NSW
897 auto *COne = ConstantInt::get(Ty: Op1C->getType(), V: 1);
898 bool WillNotSOV = willNotOverflowSignedSub(LHS: Op1C, RHS: COne, CxtI: Add);
899 BinaryOperator *Res =
900 BinaryOperator::CreateSub(V1: ConstantExpr::getSub(C1: Op1C, C2: COne), V2: X);
901 Res->setHasNoSignedWrap(Add.hasNoSignedWrap() && WillNotSOV);
902 return Res;
903 }
904
905 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
906 const APInt *C;
907 unsigned BitWidth = Ty->getScalarSizeInBits();
908 if (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: X),
909 R: m_SpecificIntAllowPoison(V: BitWidth - 1)))) &&
910 match(V: Op1, P: m_One()))
911 return new ZExtInst(Builder.CreateIsNotNeg(Arg: X, Name: "isnotneg"), Ty);
912
913 if (!match(V: Op1, P: m_APInt(Res&: C)))
914 return nullptr;
915
916 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
917 Constant *Op01C;
918 if (match(V: Op0, P: m_DisjointOr(L: m_Value(V&: X), R: m_ImmConstant(C&: Op01C)))) {
919 BinaryOperator *NewAdd =
920 BinaryOperator::CreateAdd(V1: X, V2: ConstantExpr::getAdd(C1: Op01C, C2: Op1C));
921 NewAdd->setHasNoSignedWrap(Add.hasNoSignedWrap() &&
922 willNotOverflowSignedAdd(LHS: Op01C, RHS: Op1C, CxtI: Add));
923 NewAdd->setHasNoUnsignedWrap(Add.hasNoUnsignedWrap());
924 return NewAdd;
925 }
926
927 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
928 const APInt *C2;
929 if (match(V: Op0, P: m_Or(L: m_Value(), R: m_APInt(Res&: C2))) && *C2 == -*C)
930 return BinaryOperator::CreateXor(V1: Op0, V2: ConstantInt::get(Ty: Add.getType(), V: *C2));
931
932 if (C->isSignMask()) {
933 // If wrapping is not allowed, then the addition must set the sign bit:
934 // X + (signmask) --> X | signmask
935 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
936 return BinaryOperator::CreateDisjointOr(V1: Op0, V2: Op1);
937
938 // If wrapping is allowed, then the addition flips the sign bit of LHS:
939 // X + (signmask) --> X ^ signmask
940 return BinaryOperator::CreateXor(V1: Op0, V2: Op1);
941 }
942
943 // Is this add the last step in a convoluted sext?
944 // add(zext(xor i16 X, -32768), -32768) --> sext X
945 if (match(V: Op0, P: m_ZExt(Op: m_Xor(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) &&
946 C2->isMinSignedValue() && C2->sext(width: Ty->getScalarSizeInBits()) == *C)
947 return CastInst::Create(Instruction::SExt, S: X, Ty);
948
949 if (match(V: Op0, P: m_Xor(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) {
950 // (X ^ signmask) + C --> (X + (signmask ^ C))
951 if (C2->isSignMask())
952 return BinaryOperator::CreateAdd(V1: X, V2: ConstantInt::get(Ty, V: *C2 ^ *C));
953
954 // If X has no bits set other than an xor mask,
955 // xor is equivalent to sub with no borrow between bits:
956 // add (xor X, C2), C --> sub (C2 + C), X
957 KnownBits LHSKnown = computeKnownBits(V: X, CxtI: &Add);
958 if ((*C2 | LHSKnown.Zero).isAllOnes())
959 return BinaryOperator::CreateSub(V1: ConstantInt::get(Ty, V: *C2 + *C), V2: X);
960
961 // Look for a math+logic pattern that corresponds to sext-in-register of a
962 // value with cleared high bits. Convert that into a pair of shifts:
963 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
964 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
965 if (Op0->hasOneUse() && *C2 == -(*C)) {
966 unsigned BitWidth = Ty->getScalarSizeInBits();
967 unsigned ShAmt = 0;
968 if (C->isPowerOf2())
969 ShAmt = BitWidth - C->logBase2() - 1;
970 else if (C2->isPowerOf2())
971 ShAmt = BitWidth - C2->logBase2() - 1;
972 if (ShAmt &&
973 MaskedValueIsZero(V: X, Mask: APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: ShAmt), CxtI: &Add)) {
974 Constant *ShAmtC = ConstantInt::get(Ty, V: ShAmt);
975 Value *NewShl = Builder.CreateShl(LHS: X, RHS: ShAmtC, Name: "sext");
976 return BinaryOperator::CreateAShr(V1: NewShl, V2: ShAmtC);
977 }
978 }
979 }
980
981 if (C->isOne() && Op0->hasOneUse()) {
982 // add (sext i1 X), 1 --> zext (not X)
983 // TODO: The smallest IR representation is (select X, 0, 1), and that would
984 // not require the one-use check. But we need to remove a transform in
985 // visitSelect and make sure that IR value tracking for select is equal or
986 // better than for these ops.
987 if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) &&
988 X->getType()->getScalarSizeInBits() == 1)
989 return new ZExtInst(Builder.CreateNot(V: X), Ty);
990
991 // Shifts and add used to flip and mask off the low bit:
992 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
993 const APInt *C3;
994 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))) &&
995 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
996 Value *NotX = Builder.CreateNot(V: X);
997 return BinaryOperator::CreateAnd(V1: NotX, V2: ConstantInt::get(Ty, V: 1));
998 }
999 }
1000
1001 // umax(X, C) + -C --> usub.sat(X, C)
1002 if (match(V: Op0, P: m_OneUse(SubPattern: m_UMax(Op0: m_Value(V&: X), Op1: m_SpecificInt(V: -*C)))))
1003 return replaceInstUsesWith(
1004 I&: Add, V: Builder.CreateBinaryIntrinsic(
1005 ID: Intrinsic::usub_sat, LHS: X, RHS: ConstantInt::get(Ty: Add.getType(), V: -*C)));
1006 // uadd.sat(X, C) + -C --> umin(X, ~C)
1007 // The saturating add gives X + C or UMAX, so subtracting C leaves X or
1008 // UMAX - C. Note UMAX - C == ~C.
1009 {
1010 APInt SatC = -*C;
1011 if (match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::uadd_sat>(
1012 Ops: m_Value(V&: X), Ops: m_SpecificInt(V: SatC)))))
1013 return replaceInstUsesWith(
1014 I&: Add, V: Builder.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: X,
1015 RHS: ConstantInt::get(Ty, V: ~SatC)));
1016 }
1017 // Fold (add (zext (add X, -C)), C) -> (zext X) if X u>= C.
1018 // Truncate C to the narrow type to avoid mismatched width comparisons.
1019 {
1020 const APInt *InnerC;
1021 if (match(V: Op0, P: m_ZExt(Op: m_Add(L: m_Value(V&: X), R: m_APIntAllowPoison(Res&: InnerC))))) {
1022 unsigned NarrowBW = InnerC->getBitWidth();
1023 if (C->isIntN(N: NarrowBW)) {
1024 APInt NarrowC = C->trunc(width: NarrowBW);
1025 const SimplifyQuery Q = SQ.getWithInstruction(I: &Add);
1026 if (*InnerC == -NarrowC &&
1027 (NarrowC.isOne()
1028 ? llvm::isKnownNonZero(V: X, Q)
1029 : computeKnownBits(V: X, CxtI: &Add).getMinValue().uge(RHS: NarrowC)))
1030 return new ZExtInst(X, Ty);
1031 }
1032 }
1033 }
1034
1035 return nullptr;
1036}
1037
1038// match variations of a^2 + 2*a*b + b^2
1039//
1040// to reuse the code between the FP and Int versions, the instruction OpCodes
1041// and constant types have been turned into template parameters.
1042//
1043// Mul2Rhs: The constant to perform the multiplicative equivalent of X*2 with;
1044// should be `m_SpecificFP(2.0)` for FP and `m_SpecificInt(1)` for Int
1045// (we're matching `X<<1` instead of `X*2` for Int)
1046template <bool FP, typename Mul2Rhs>
1047static bool matchesSquareSum(BinaryOperator &I, Mul2Rhs M2Rhs, Value *&A,
1048 Value *&B) {
1049 constexpr unsigned MulOp = FP ? Instruction::FMul : Instruction::Mul;
1050 constexpr unsigned AddOp = FP ? Instruction::FAdd : Instruction::Add;
1051 constexpr unsigned Mul2Op = FP ? Instruction::FMul : Instruction::Shl;
1052
1053 // (a * a) + (((a * 2) + b) * b)
1054 if (match(&I, m_c_BinOp(
1055 AddOp, m_OneUse(SubPattern: m_BinOp(Opcode: MulOp, L: m_Value(V&: A), R: m_Deferred(V: A))),
1056 m_OneUse(m_c_BinOp(
1057 MulOp,
1058 m_c_BinOp(AddOp, m_BinOp(Mul2Op, m_Deferred(V: A), M2Rhs),
1059 m_Value(V&: B)),
1060 m_Deferred(V: B))))))
1061 return true;
1062
1063 // ((a * b) * 2) or ((a * 2) * b)
1064 // +
1065 // (a * a + b * b) or (b * b + a * a)
1066 return match(
1067 &I, m_c_BinOp(
1068 AddOp,
1069 m_CombineOr(
1070 m_OneUse(m_BinOp(
1071 Mul2Op, m_BinOp(Opcode: MulOp, L: m_Value(V&: A), R: m_Value(V&: B)), M2Rhs)),
1072 m_OneUse(m_c_BinOp(MulOp, m_BinOp(Mul2Op, m_Value(V&: A), M2Rhs),
1073 m_Value(V&: B)))),
1074 m_OneUse(
1075 SubPattern: m_c_BinOp(Opcode: AddOp, L: m_BinOp(Opcode: MulOp, L: m_Deferred(V: A), R: m_Deferred(V: A)),
1076 R: m_BinOp(Opcode: MulOp, L: m_Deferred(V: B), R: m_Deferred(V: B))))));
1077}
1078
1079// Fold integer variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1080Instruction *InstCombinerImpl::foldSquareSumInt(BinaryOperator &I) {
1081 Value *A, *B;
1082 if (matchesSquareSum</*FP*/ false>(I, M2Rhs: m_SpecificInt(V: 1), A, B)) {
1083 Value *AB = Builder.CreateAdd(LHS: A, RHS: B);
1084 return BinaryOperator::CreateMul(V1: AB, V2: AB);
1085 }
1086 return nullptr;
1087}
1088
1089// Fold floating point variations of a^2 + 2*a*b + b^2 -> (a + b)^2
1090// Requires `nsz` and `reassoc`.
1091Instruction *InstCombinerImpl::foldSquareSumFP(BinaryOperator &I) {
1092 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() && "Assumption mismatch");
1093 Value *A, *B;
1094 if (matchesSquareSum</*FP*/ true>(I, M2Rhs: m_SpecificFP(V: 2.0), A, B)) {
1095 Value *AB = Builder.CreateFAddFMF(L: A, R: B, FMFSource: &I);
1096 return BinaryOperator::CreateFMulFMF(V1: AB, V2: AB, FMFSource: &I);
1097 }
1098 return nullptr;
1099}
1100
1101// Matches multiplication expression Op * C where C is a constant. Returns the
1102// constant value in C and the other operand in Op. Returns true if such a
1103// match is found.
1104static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1105 const APInt *AI;
1106 if (match(V: E, P: m_Mul(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1107 C = *AI;
1108 return true;
1109 }
1110 if (match(V: E, P: m_Shl(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1111 C = APInt(AI->getBitWidth(), 1);
1112 C <<= *AI;
1113 return true;
1114 }
1115 return false;
1116}
1117
1118// Matches remainder expression Op % C where C is a constant. Returns the
1119// constant value in C and the other operand in Op. Returns the signedness of
1120// the remainder operation in IsSigned. Returns true if such a match is
1121// found.
1122static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1123 const APInt *AI;
1124 IsSigned = false;
1125 if (match(V: E, P: m_SRem(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1126 IsSigned = true;
1127 C = *AI;
1128 return true;
1129 }
1130 if (match(V: E, P: m_URem(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1131 C = *AI;
1132 return true;
1133 }
1134 if (match(V: E, P: m_And(L: m_Value(V&: Op), R: m_APInt(Res&: AI))) && (*AI + 1).isPowerOf2()) {
1135 C = *AI + 1;
1136 return true;
1137 }
1138 return false;
1139}
1140
1141// Matches division expression Op / C with the given signedness as indicated
1142// by IsSigned, where C is a constant. Returns the constant value in C and the
1143// other operand in Op. Returns true if such a match is found.
1144static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1145 const APInt *AI;
1146 if (IsSigned && match(V: E, P: m_SDiv(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1147 C = *AI;
1148 return true;
1149 }
1150 if (!IsSigned) {
1151 if (match(V: E, P: m_UDiv(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1152 C = *AI;
1153 return true;
1154 }
1155 if (match(V: E, P: m_LShr(L: m_Value(V&: Op), R: m_APInt(Res&: AI)))) {
1156 C = APInt(AI->getBitWidth(), 1);
1157 C <<= *AI;
1158 return true;
1159 }
1160 }
1161 return false;
1162}
1163
1164// Returns whether C0 * C1 with the given signedness overflows.
1165static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1166 bool overflow;
1167 if (IsSigned)
1168 (void)C0.smul_ov(RHS: C1, Overflow&: overflow);
1169 else
1170 (void)C0.umul_ov(RHS: C1, Overflow&: overflow);
1171 return overflow;
1172}
1173
1174// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1175// does not overflow.
1176// Simplifies (X / C0) * C1 + (X % C0) * C2 to
1177// (X / C0) * (C1 - C2 * C0) + X * C2
1178Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1179 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
1180 Value *X, *MulOpV;
1181 APInt C0, MulOpC;
1182 bool IsSigned;
1183 // Match I = X % C0 + MulOpV * C0
1184 if (((MatchRem(E: LHS, Op&: X, C&: C0, IsSigned) && MatchMul(E: RHS, Op&: MulOpV, C&: MulOpC)) ||
1185 (MatchRem(E: RHS, Op&: X, C&: C0, IsSigned) && MatchMul(E: LHS, Op&: MulOpV, C&: MulOpC))) &&
1186 C0 == MulOpC) {
1187 Value *RemOpV;
1188 APInt C1;
1189 bool Rem2IsSigned;
1190 // Match MulOpC = RemOpV % C1
1191 if (MatchRem(E: MulOpV, Op&: RemOpV, C&: C1, IsSigned&: Rem2IsSigned) &&
1192 IsSigned == Rem2IsSigned) {
1193 Value *DivOpV;
1194 APInt DivOpC;
1195 // Match RemOpV = X / C0
1196 if (MatchDiv(E: RemOpV, Op&: DivOpV, C&: DivOpC, IsSigned) && X == DivOpV &&
1197 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1198 Value *NewDivisor = ConstantInt::get(Ty: X->getType(), V: C0 * C1);
1199 return IsSigned ? Builder.CreateSRem(LHS: X, RHS: NewDivisor, Name: "srem")
1200 : Builder.CreateURem(LHS: X, RHS: NewDivisor, Name: "urem");
1201 }
1202 }
1203 }
1204
1205 // Match I = (X / C0) * C1 + (X % C0) * C2.
1206 auto FoldDivRem = [&](Value *DivSide, Value *RemSide) -> Value * {
1207 Value *Div, *Rem;
1208 APInt C1, C2;
1209 if (!DivSide->hasOneUse() || !MatchMul(E: DivSide, Op&: Div, C&: C1))
1210 Div = DivSide, C1 = APInt(I.getType()->getScalarSizeInBits(), 1);
1211 if (!RemSide->hasOneUse() || !MatchMul(E: RemSide, Op&: Rem, C&: C2))
1212 Rem = RemSide, C2 = APInt(I.getType()->getScalarSizeInBits(), 1);
1213 Value *DivOpV;
1214 APInt DivOpC;
1215 if (MatchRem(E: Rem, Op&: X, C&: C0, IsSigned) &&
1216 MatchDiv(E: Div, Op&: DivOpV, C&: DivOpC, IsSigned) && X == DivOpV &&
1217 C0 == DivOpC &&
1218 // Avoid unprofitable replacement of and with mul.
1219 !(C1.isOne() && !IsSigned && DivOpC.isPowerOf2() && DivOpC != 2)) {
1220 APInt NewC = C1 - C2 * C0;
1221 if (!NewC.isZero() && !Rem->hasOneUse())
1222 return nullptr;
1223 if (!isGuaranteedNotToBeUndef(V: X, AC: &AC, CtxI: &I, DT: &DT))
1224 return nullptr;
1225 Value *MulXC2 = Builder.CreateMul(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: C2));
1226 if (NewC.isZero())
1227 return MulXC2;
1228 return Builder.CreateAdd(
1229 LHS: Builder.CreateMul(LHS: Div, RHS: ConstantInt::get(Ty: X->getType(), V: NewC)), RHS: MulXC2);
1230 }
1231 return nullptr;
1232 };
1233 if (Value *V = FoldDivRem(LHS, RHS))
1234 return V;
1235 if (Value *V = FoldDivRem(RHS, LHS))
1236 return V;
1237
1238 return nullptr;
1239}
1240
1241/// Fold
1242/// (1 << NBits) - 1
1243/// Into:
1244/// ~(-(1 << NBits))
1245/// Because a 'not' is better for bit-tracking analysis and other transforms
1246/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1247static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1248 InstCombiner::BuilderTy &Builder) {
1249 Value *NBits;
1250 if (!match(V: &I, P: m_Add(L: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: NBits))), R: m_AllOnes())))
1251 return nullptr;
1252
1253 Constant *MinusOne = Constant::getAllOnesValue(Ty: NBits->getType());
1254 Value *NotMask = Builder.CreateShl(LHS: MinusOne, RHS: NBits, Name: "notmask");
1255 // Be wary of constant folding.
1256 if (auto *BOp = dyn_cast<BinaryOperator>(Val: NotMask)) {
1257 // Always NSW. But NUW propagates from `add`.
1258 BOp->setHasNoSignedWrap();
1259 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1260 }
1261
1262 return BinaryOperator::CreateNot(Op: NotMask, Name: I.getName());
1263}
1264
1265static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1266 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1267 Type *Ty = I.getType();
1268 auto getUAddSat = [&]() {
1269 return Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: Intrinsic::uadd_sat,
1270 OverloadTys: Ty);
1271 };
1272
1273 // add (umin X, ~Y), Y --> uaddsat X, Y
1274 Value *X, *Y;
1275 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))),
1276 R: m_Deferred(V: Y))))
1277 return CallInst::Create(Func: getUAddSat(), Args: { X, Y });
1278
1279 // add (umin X, ~C), C --> uaddsat X, C
1280 const APInt *C, *NotC;
1281 if (match(V: &I, P: m_Add(L: m_UMin(Op0: m_Value(V&: X), Op1: m_APInt(Res&: NotC)), R: m_APInt(Res&: C))) &&
1282 *C == ~*NotC)
1283 return CallInst::Create(Func: getUAddSat(), Args: { X, ConstantInt::get(Ty, V: *C) });
1284
1285 return nullptr;
1286}
1287
1288// Transform:
1289// (add A, (shl (neg B), Y))
1290// -> (sub A, (shl B, Y))
1291static Instruction *combineAddSubWithShlAddSub(InstCombiner::BuilderTy &Builder,
1292 const BinaryOperator &I) {
1293 Value *A, *B, *Cnt;
1294 if (match(V: &I,
1295 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))),
1296 R: m_Value(V&: A)))) {
1297 Value *NewShl = Builder.CreateShl(LHS: B, RHS: Cnt);
1298 return BinaryOperator::CreateSub(V1: A, V2: NewShl);
1299 }
1300 return nullptr;
1301}
1302
1303/// Try to reduce signed division by power-of-2 to an arithmetic shift right.
1304static Instruction *foldAddToAshr(BinaryOperator &Add) {
1305 // Division must be by power-of-2, but not the minimum signed value.
1306 Value *X;
1307 const APInt *DivC;
1308 if (!match(V: Add.getOperand(i_nocapture: 0), P: m_SDiv(L: m_Value(V&: X), R: m_Power2(V&: DivC))) ||
1309 DivC->isNegative())
1310 return nullptr;
1311
1312 // Rounding is done by adding -1 if the dividend (X) is negative and has any
1313 // low bits set. It recognizes two canonical patterns:
1314 // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
1315 // pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
1316 // 2. For an 'eq' cmp, the pattern's: sext (icmp eq X & (SMIN + 1), SMIN + 1).
1317 // Note that, by the time we end up here, if possible, ugt has been
1318 // canonicalized into eq.
1319 const APInt *MaskC, *MaskCCmp;
1320 CmpPredicate Pred;
1321 if (!match(V: Add.getOperand(i_nocapture: 1),
1322 P: m_SExt(Op: m_ICmp(Pred, L: m_And(L: m_Specific(V: X), R: m_APInt(Res&: MaskC)),
1323 R: m_APInt(Res&: MaskCCmp)))))
1324 return nullptr;
1325
1326 if ((Pred != ICmpInst::ICMP_UGT || !MaskCCmp->isSignMask()) &&
1327 (Pred != ICmpInst::ICMP_EQ || *MaskCCmp != *MaskC))
1328 return nullptr;
1329
1330 APInt SMin = APInt::getSignedMinValue(numBits: Add.getType()->getScalarSizeInBits());
1331 bool IsMaskValid = Pred == ICmpInst::ICMP_UGT
1332 ? (*MaskC == (SMin | (*DivC - 1)))
1333 : (*DivC == 2 && *MaskC == SMin + 1);
1334 if (!IsMaskValid)
1335 return nullptr;
1336
1337 // (X / DivC) + sext ((X & (SMin | (DivC - 1)) >u SMin) --> X >>s log2(DivC)
1338 return BinaryOperator::CreateAShr(
1339 V1: X, V2: ConstantInt::get(Ty: Add.getType(), V: DivC->exactLogBase2()));
1340}
1341
1342Instruction *InstCombinerImpl::foldAddLikeCommutative(Value *LHS, Value *RHS,
1343 bool NSW, bool NUW) {
1344 Value *A, *B, *C;
1345 if (match(V: LHS, P: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B))) &&
1346 match(V: RHS, P: m_Sub(L: m_Value(V&: C), R: m_Specific(V: A)))) {
1347 Instruction *R = BinaryOperator::CreateSub(V1: C, V2: B);
1348 bool NSWOut = NSW && match(V: LHS, P: m_NSWSub(L: m_Value(), R: m_Value())) &&
1349 match(V: RHS, P: m_NSWSub(L: m_Value(), R: m_Value()));
1350
1351 bool NUWOut = match(V: LHS, P: m_NUWSub(L: m_Value(), R: m_Value())) &&
1352 match(V: RHS, P: m_NUWSub(L: m_Value(), R: m_Value()));
1353 R->setHasNoSignedWrap(NSWOut);
1354 R->setHasNoUnsignedWrap(NUWOut);
1355 return R;
1356 }
1357
1358 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1359 const APInt *C1, *C2;
1360 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)))) {
1361 APInt One(C2->getBitWidth(), 1);
1362 APInt MinusC1 = -(*C1);
1363 if (MinusC1 == (One << *C2)) {
1364 Constant *NewRHS = ConstantInt::get(Ty: RHS->getType(), V: MinusC1);
1365 return BinaryOperator::CreateSRem(V1: RHS, V2: NewRHS);
1366 }
1367 }
1368
1369 // (A + C) + (B & ~C) == A + (B | C)
1370 if (match(V: LHS, P: m_c_Add(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
1371 match(V: RHS, P: m_c_And(L: m_Value(V&: B), R: m_SpecificInt(V: ~*C1)))) {
1372 // Replacing one add with {or, add}. Avoid growth if both sides are shared.
1373 if (!LHS->hasOneUse() && !RHS->hasOneUse())
1374 return nullptr;
1375
1376 bool NSWOut = NSW && match(V: LHS, P: m_NSWAdd(L: m_Value(), R: m_Value()));
1377 bool NUWOut = NUW && match(V: LHS, P: m_NUWAdd(L: m_Value(), R: m_Value()));
1378 Value *NewOr =
1379 Builder.CreateOr(LHS: B, RHS: Constant::getIntegerValue(Ty: LHS->getType(), V: *C1));
1380 Instruction *NewAdd = BinaryOperator::CreateAdd(V1: A, V2: NewOr);
1381 NewAdd->setHasNoSignedWrap(NSWOut);
1382 NewAdd->setHasNoUnsignedWrap(NUWOut);
1383 return NewAdd;
1384 }
1385
1386 return nullptr;
1387}
1388
1389Instruction *InstCombinerImpl::
1390 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1391 BinaryOperator &I) {
1392 assert((I.getOpcode() == Instruction::Add ||
1393 I.getOpcode() == Instruction::Or ||
1394 I.getOpcode() == Instruction::Sub) &&
1395 "Expecting add/or/sub instruction");
1396
1397 // We have a subtraction/addition between a (potentially truncated) *logical*
1398 // right-shift of X and a "select".
1399 Value *X, *Select;
1400 Instruction *LowBitsToSkip, *Extract;
1401 if (!match(V: &I, P: m_c_BinOp(L: m_TruncOrSelf(Op: m_Instruction(
1402 I&: Extract, P: m_LShr(L: m_Value(V&: X),
1403 R: m_Instruction(I&: LowBitsToSkip)))),
1404 R: m_Value(V&: Select))))
1405 return nullptr;
1406
1407 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1408 if (I.getOpcode() == Instruction::Sub && I.getOperand(i_nocapture: 1) != Select)
1409 return nullptr;
1410
1411 Type *XTy = X->getType();
1412 bool HadTrunc = I.getType() != XTy;
1413
1414 // If there was a truncation of extracted value, then we'll need to produce
1415 // one extra instruction, so we need to ensure one instruction will go away.
1416 if (HadTrunc && !match(V: &I, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value())))
1417 return nullptr;
1418
1419 // Extraction should extract high NBits bits, with shift amount calculated as:
1420 // low bits to skip = shift bitwidth - high bits to extract
1421 // The shift amount itself may be extended, and we need to look past zero-ext
1422 // when matching NBits, that will matter for matching later.
1423 Value *NBits;
1424 if (!match(V: LowBitsToSkip,
1425 P: m_ZExtOrSelf(Op: m_Sub(L: m_SpecificInt(V: XTy->getScalarSizeInBits()),
1426 R: m_ZExtOrSelf(Op: m_Value(V&: NBits))))))
1427 return nullptr;
1428
1429 // Sign-extending value can be zero-extended if we `sub`tract it,
1430 // or sign-extended otherwise.
1431 auto SkipExtInMagic = [&I](Value *&V) {
1432 if (I.getOpcode() == Instruction::Sub)
1433 match(V, P: m_ZExtOrSelf(Op: m_Value(V)));
1434 else
1435 match(V, P: m_SExtOrSelf(Op: m_Value(V)));
1436 };
1437
1438 // Now, finally validate the sign-extending magic.
1439 // `select` itself may be appropriately extended, look past that.
1440 SkipExtInMagic(Select);
1441
1442 CmpPredicate Pred;
1443 const APInt *Thr;
1444 Value *SignExtendingValue, *Zero;
1445 bool ShouldSignext;
1446 // It must be a select between two values we will later establish to be a
1447 // sign-extending value and a zero constant. The condition guarding the
1448 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1449 if (!match(V: Select, P: m_Select(C: m_ICmp(Pred, L: m_Specific(V: X), R: m_APInt(Res&: Thr)),
1450 L: m_Value(V&: SignExtendingValue), R: m_Value(V&: Zero))) ||
1451 !isSignBitCheck(Pred, RHS: *Thr, TrueIfSigned&: ShouldSignext))
1452 return nullptr;
1453
1454 // icmp-select pair is commutative.
1455 if (!ShouldSignext)
1456 std::swap(a&: SignExtendingValue, b&: Zero);
1457
1458 // If we should not perform sign-extension then we must add/or/subtract zero.
1459 if (!match(V: Zero, P: m_Zero()))
1460 return nullptr;
1461 // Otherwise, it should be some constant, left-shifted by the same NBits we
1462 // had in `lshr`. Said left-shift can also be appropriately extended.
1463 // Again, we must look past zero-ext when looking for NBits.
1464 SkipExtInMagic(SignExtendingValue);
1465 Constant *SignExtendingValueBaseConstant;
1466 if (!match(V: SignExtendingValue,
1467 P: m_Shl(L: m_Constant(C&: SignExtendingValueBaseConstant),
1468 R: m_ZExtOrSelf(Op: m_Specific(V: NBits)))))
1469 return nullptr;
1470 // If we `sub`, then the constant should be one, else it should be all-ones.
1471 if (I.getOpcode() == Instruction::Sub
1472 ? !match(V: SignExtendingValueBaseConstant, P: m_One())
1473 : !match(V: SignExtendingValueBaseConstant, P: m_AllOnes()))
1474 return nullptr;
1475
1476 auto *NewAShr = BinaryOperator::CreateAShr(V1: X, V2: LowBitsToSkip,
1477 Name: Extract->getName() + ".sext");
1478 NewAShr->copyIRFlags(V: Extract); // Preserve `exact`-ness.
1479 if (!HadTrunc)
1480 return NewAShr;
1481
1482 Builder.Insert(I: NewAShr);
1483 return TruncInst::CreateTruncOrBitCast(S: NewAShr, Ty: I.getType());
1484}
1485
1486/// This is a specialization of a more general transform from
1487/// foldUsingDistributiveLaws. If that code can be made to work optimally
1488/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1489static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1490 InstCombiner::BuilderTy &Builder) {
1491 // TODO: Also handle mul by doubling the shift amount?
1492 assert((I.getOpcode() == Instruction::Add ||
1493 I.getOpcode() == Instruction::Sub) &&
1494 "Expected add/sub");
1495 auto *Op0 = dyn_cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 0));
1496 auto *Op1 = dyn_cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 1));
1497 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1498 return nullptr;
1499
1500 Value *X, *Y, *ShAmt;
1501 if (!match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: ShAmt))) ||
1502 !match(V: Op1, P: m_Shl(L: m_Value(V&: Y), R: m_Specific(V: ShAmt))))
1503 return nullptr;
1504
1505 // No-wrap propagates only when all ops have no-wrap.
1506 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1507 Op1->hasNoSignedWrap();
1508 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1509 Op1->hasNoUnsignedWrap();
1510
1511 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1512 Value *NewMath = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: X, RHS: Y);
1513 if (auto *NewI = dyn_cast<BinaryOperator>(Val: NewMath)) {
1514 NewI->setHasNoSignedWrap(HasNSW);
1515 NewI->setHasNoUnsignedWrap(HasNUW);
1516 }
1517 auto *NewShl = BinaryOperator::CreateShl(V1: NewMath, V2: ShAmt);
1518 NewShl->setHasNoSignedWrap(HasNSW);
1519 NewShl->setHasNoUnsignedWrap(HasNUW);
1520 return NewShl;
1521}
1522
1523/// Reduce a sequence of masked half-width multiplies to a single multiply.
1524/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1525static Instruction *foldBoxMultiply(BinaryOperator &I) {
1526 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1527 // Skip the odd bitwidth types.
1528 if ((BitWidth & 0x1))
1529 return nullptr;
1530
1531 unsigned HalfBits = BitWidth >> 1;
1532 APInt HalfMask = APInt::getMaxValue(numBits: HalfBits);
1533
1534 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1535 Value *XLo, *YLo;
1536 Value *CrossSum;
1537 // Require one-use on the multiply to avoid increasing the number of
1538 // multiplications.
1539 if (!match(V: &I, P: m_c_Add(L: m_Shl(L: m_Value(V&: CrossSum), R: m_SpecificInt(V: HalfBits)),
1540 R: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: YLo), R: m_Value(V&: XLo))))))
1541 return nullptr;
1542
1543 // XLo = X & HalfMask
1544 // YLo = Y & HalfMask
1545 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1546 // to enhance robustness
1547 Value *X, *Y;
1548 if (!match(V: XLo, P: m_And(L: m_Value(V&: X), R: m_SpecificInt(V: HalfMask))) ||
1549 !match(V: YLo, P: m_And(L: m_Value(V&: Y), R: m_SpecificInt(V: HalfMask))))
1550 return nullptr;
1551
1552 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1553 // X' can be either X or XLo in the pattern (and the same for Y')
1554 if (match(V: CrossSum,
1555 P: m_c_Add(L: m_c_Mul(L: m_LShr(L: m_Specific(V: Y), R: m_SpecificInt(V: HalfBits)),
1556 R: m_CombineOr(Ps: m_Specific(V: X), Ps: m_Specific(V: XLo))),
1557 R: m_c_Mul(L: m_LShr(L: m_Specific(V: X), R: m_SpecificInt(V: HalfBits)),
1558 R: m_CombineOr(Ps: m_Specific(V: Y), Ps: m_Specific(V: YLo))))))
1559 return BinaryOperator::CreateMul(V1: X, V2: Y);
1560
1561 return nullptr;
1562}
1563
1564/// Return true if X + (Y-1) is provably non-wrapping in X's type
1565static bool checkDivCeilNUW(Value *X, Value *Y, const SimplifyQuery &SQ) {
1566 ConstantRange CRX = computeConstantRange(V: X, /*ForSigned=*/false, SQ);
1567 ConstantRange CRY = computeConstantRange(V: Y, /*ForSigned=*/false, SQ);
1568 APInt MinY = CRY.getUnsignedMin();
1569 APInt MaxX = CRX.getUnsignedMax();
1570 APInt MaxY = CRY.getUnsignedMax();
1571
1572 return !MinY.isZero() && !MaxX.ugt(RHS: -MaxY);
1573}
1574
1575/// Fold the div_ceil idiom in both forms:
1576/// add(udiv(X, Y), zext(icmp ne(urem(X, Y), 0)))
1577/// -> udiv(add nuw(X, Y - 1), Y)
1578/// add(zext(udiv(X, Y)), zext(icmp ne(urem(X, Y), 0)))
1579/// -> zext(udiv(add nuw(X, Y - 1), Y))
1580/// The zext form applies when udiv/urem operate in a narrower type than the
1581/// add.
1582Instruction *InstCombinerImpl::foldDivCeil(BinaryOperator &I) {
1583 Value *X, *Y;
1584
1585 auto UDivPat = m_OneUse(SubPattern: m_UDiv(L: m_Value(V&: X), R: m_Value(V&: Y)));
1586 auto URemPat = m_OneUse(SubPattern: m_URem(L: m_Deferred(V: X), R: m_Deferred(V: Y)));
1587 auto ICmpPat = m_OneUse(SubPattern: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, L: URemPat, R: m_Zero()));
1588 auto DivPat = m_OneUse(SubPattern: m_ZExtOrSelf(Op: UDivPat));
1589 auto ZExtCmpPat = m_OneUse(SubPattern: m_ZExt(Op: ICmpPat));
1590
1591 if (!match(V: &I, P: m_c_Add(L: DivPat, R: ZExtCmpPat)) || !checkDivCeilNUW(X, Y, SQ))
1592 return nullptr;
1593
1594 Value *YMinusOne =
1595 Builder.CreateAdd(LHS: Y, RHS: ConstantInt::getAllOnesValue(Ty: Y->getType()));
1596 Value *NUWAdd = Builder.CreateNUWAdd(LHS: X, RHS: YMinusOne);
1597 if (X->getType() != I.getType()) {
1598 Value *Div = Builder.CreateUDiv(LHS: NUWAdd, RHS: Y);
1599 return new ZExtInst(Div, I.getType());
1600 }
1601 return BinaryOperator::CreateUDiv(V1: NUWAdd, V2: Y);
1602}
1603
1604Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1605 if (Value *V = simplifyAddInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
1606 IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(),
1607 Q: SQ.getWithInstruction(I: &I)))
1608 return replaceInstUsesWith(I, V);
1609
1610 if (SimplifyAssociativeOrCommutative(I))
1611 return &I;
1612
1613 if (Instruction *X = foldVectorBinop(Inst&: I))
1614 return X;
1615
1616 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
1617 return Phi;
1618
1619 // (A*B)+(A*C) -> A*(B+C) etc
1620 if (Value *V = foldUsingDistributiveLaws(I))
1621 return replaceInstUsesWith(I, V);
1622
1623 if (Instruction *R = foldBoxMultiply(I))
1624 return R;
1625
1626 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1627 return R;
1628
1629 if (Instruction *X = foldAddWithConstant(Add&: I))
1630 return X;
1631
1632 if (Instruction *X = foldNoWrapAdd(Add&: I, Builder))
1633 return X;
1634
1635 if (Instruction *R = foldBinOpShiftWithShift(I))
1636 return R;
1637
1638 if (Instruction *R = combineAddSubWithShlAddSub(Builder, I))
1639 return R;
1640
1641 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
1642 if (Instruction *R = foldAddLikeCommutative(LHS, RHS, NSW: I.hasNoSignedWrap(),
1643 NUW: I.hasNoUnsignedWrap()))
1644 return R;
1645 if (Instruction *R = foldAddLikeCommutative(LHS: RHS, RHS: LHS, NSW: I.hasNoSignedWrap(),
1646 NUW: I.hasNoUnsignedWrap()))
1647 return R;
1648 Type *Ty = I.getType();
1649 if (Ty->isIntOrIntVectorTy(BitWidth: 1))
1650 return BinaryOperator::CreateXor(V1: LHS, V2: RHS);
1651
1652 // X + X --> X << 1
1653 if (LHS == RHS) {
1654 auto *Shl = BinaryOperator::CreateShl(V1: LHS, V2: ConstantInt::get(Ty, V: 1));
1655 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1656 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1657 return Shl;
1658 }
1659
1660 Value *A, *B;
1661 if (match(V: LHS, P: m_Neg(V: m_Value(V&: A)))) {
1662 // -A + -B --> -(A + B)
1663 if (match(V: RHS, P: m_Neg(V: m_Value(V&: B))))
1664 return BinaryOperator::CreateNeg(Op: Builder.CreateAdd(LHS: A, RHS: B));
1665
1666 // -A + B --> B - A
1667 auto *Sub = BinaryOperator::CreateSub(V1: RHS, V2: A);
1668 auto *OB0 = cast<OverflowingBinaryOperator>(Val: LHS);
1669 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OB0->hasNoSignedWrap());
1670
1671 return Sub;
1672 }
1673
1674 // A + -B --> A - B
1675 if (match(V: RHS, P: m_Neg(V: m_Value(V&: B)))) {
1676 auto *Sub = BinaryOperator::CreateSub(V1: LHS, V2: B);
1677 auto *OBO = cast<OverflowingBinaryOperator>(Val: RHS);
1678 Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO->hasNoSignedWrap());
1679 return Sub;
1680 }
1681
1682 if (Value *V = checkForNegativeOperand(I, Builder))
1683 return replaceInstUsesWith(I, V);
1684
1685 // (A + 1) + ~B --> A - B
1686 // ~B + (A + 1) --> A - B
1687 // (~B + A) + 1 --> A - B
1688 // (A + ~B) + 1 --> A - B
1689 // This relies on the ~B == -1-B identity.
1690 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)))) ||
1691 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())))
1692 return BinaryOperator::CreateSub(V1: A, V2: B);
1693
1694 {
1695 // (A + C) + ~B --> A - B + (C-1)
1696 // ~B + (A + C) --> A - B + (C-1)
1697 // (~B + A) + C --> A - B + (C-1)
1698 // (A + ~B) + C --> A - B + (C-1)
1699 // With constant C, subtraction of one is free, so we replace three ops
1700 // (two adds and a bitwise-not) with two (sub and add).
1701 const APInt *C;
1702 if (match(V: &I, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Add(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: C))),
1703 R: m_Not(V: m_Value(V&: B)))) ||
1704 match(V: &I, P: m_BinOp(L: m_OneUse(SubPattern: m_c_Add(L: m_Not(V: m_Value(V&: B)), R: m_Value(V&: A))),
1705 R: m_APIntAllowPoison(Res&: C)))) {
1706 Value *Sub = Builder.CreateSub(LHS: A, RHS: B);
1707 return BinaryOperator::CreateAdd(V1: Sub, V2: ConstantInt::get(Ty, V: *C - 1));
1708 }
1709 }
1710
1711 // (A + RHS) + RHS --> A + (RHS << 1)
1712 if (match(V: LHS, P: m_OneUse(SubPattern: m_c_Add(L: m_Value(V&: A), R: m_Specific(V: RHS)))))
1713 return BinaryOperator::CreateAdd(V1: A, V2: Builder.CreateShl(LHS: RHS, RHS: 1, Name: "reass.add"));
1714
1715 // LHS + (A + LHS) --> A + (LHS << 1)
1716 if (match(V: RHS, P: m_OneUse(SubPattern: m_c_Add(L: m_Value(V&: A), R: m_Specific(V: LHS)))))
1717 return BinaryOperator::CreateAdd(V1: A, V2: Builder.CreateShl(LHS, RHS: 1, Name: "reass.add"));
1718
1719 {
1720 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1721 Constant *C1, *C2;
1722 if (match(V: &I, P: m_c_Add(L: m_Add(L: m_Value(V&: A), R: m_ImmConstant(C&: C1)),
1723 R: m_Sub(L: m_ImmConstant(C&: C2), R: m_Value(V&: B)))) &&
1724 (LHS->hasOneUse() || RHS->hasOneUse())) {
1725 Value *Sub = Builder.CreateSub(LHS: A, RHS: B);
1726 return BinaryOperator::CreateAdd(V1: Sub, V2: ConstantExpr::getAdd(C1, C2));
1727 }
1728
1729 // Canonicalize a constant sub operand as an add operand for better folding:
1730 // (C1 - A) + B --> (B - A) + C1
1731 if (match(V: &I, P: m_c_Add(L: m_OneUse(SubPattern: m_Sub(L: m_ImmConstant(C&: C1), R: m_Value(V&: A))),
1732 R: m_Value(V&: B)))) {
1733 Value *Sub = Builder.CreateSub(LHS: B, RHS: A, Name: "reass.sub");
1734 return BinaryOperator::CreateAdd(V1: Sub, V2: C1);
1735 }
1736 }
1737
1738 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1739 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1740
1741 const APInt *C1;
1742 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1743 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))) &&
1744 C1->isPowerOf2() && (ComputeNumSignBits(Op: A) > C1->countl_zero())) {
1745 Constant *NewMask = ConstantInt::get(Ty: RHS->getType(), V: *C1 - 1);
1746 return BinaryOperator::CreateAnd(V1: A, V2: NewMask);
1747 }
1748
1749 // ZExt (B - A) + ZExt(A) --> ZExt(B)
1750 if ((match(V: RHS, P: m_ZExt(Op: m_Value(V&: A))) &&
1751 match(V: LHS, P: m_ZExt(Op: m_NUWSub(L: m_Value(V&: B), R: m_Specific(V: A))))) ||
1752 (match(V: LHS, P: m_ZExt(Op: m_Value(V&: A))) &&
1753 match(V: RHS, P: m_ZExt(Op: m_NUWSub(L: m_Value(V&: B), R: m_Specific(V: A))))))
1754 return new ZExtInst(B, LHS->getType());
1755
1756 // zext(A) + sext(A) --> 0 if A is i1
1757 if (match(V: &I, P: m_c_BinOp(L: m_ZExt(Op: m_Value(V&: A)), R: m_SExt(Op: m_Deferred(V: A)))) &&
1758 A->getType()->isIntOrIntVectorTy(BitWidth: 1))
1759 return replaceInstUsesWith(I, V: Constant::getNullValue(Ty: I.getType()));
1760
1761 // sext(A < B) + zext(A > B) => ucmp/scmp(A, B)
1762 CmpPredicate LTPred, GTPred;
1763 if (match(V: &I,
1764 P: m_c_Add(L: m_SExt(Op: m_c_ICmp(Pred&: LTPred, L: m_Value(V&: A), R: m_Value(V&: B))),
1765 R: m_ZExt(Op: m_c_ICmp(Pred&: GTPred, L: m_Deferred(V: A), R: m_Deferred(V: B))))) &&
1766 A->getType()->isIntOrIntVectorTy()) {
1767 if (ICmpInst::isGT(P: LTPred)) {
1768 std::swap(a&: LTPred, b&: GTPred);
1769 std::swap(a&: A, b&: B);
1770 }
1771
1772 if (ICmpInst::isLT(P: LTPred) && ICmpInst::isGT(P: GTPred) &&
1773 ICmpInst::isSigned(Pred: LTPred) == ICmpInst::isSigned(Pred: GTPred))
1774 return replaceInstUsesWith(
1775 I, V: Builder.CreateIntrinsic(
1776 RetTy: Ty,
1777 ID: ICmpInst::isSigned(Pred: LTPred) ? Intrinsic::scmp : Intrinsic::ucmp,
1778 Args: {A, B}));
1779 }
1780
1781 // A+B --> A|B iff A and B have no bits set in common.
1782 WithCache<const Value *> LHSCache(LHS), RHSCache(RHS);
1783 switch (
1784 getNoCommonBitsSetResult(LHSCache, RHSCache, SQ: SQ.getWithInstruction(I: &I))) {
1785 case NoCommonBitsSetResult::Known:
1786 return BinaryOperator::CreateDisjointOr(V1: LHS, V2: RHS);
1787 case NoCommonBitsSetResult::OnlyIfUndefIgnored:
1788 return BinaryOperator::CreateOr(V1: LHS, V2: RHS);
1789 case NoCommonBitsSetResult::Unknown:
1790 break;
1791 }
1792
1793 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1794 return Ext;
1795
1796 // (add (xor A, B) (and A, B)) --> (or A, B)
1797 // (add (and A, B) (xor A, B)) --> (or A, B)
1798 if (match(V: &I, P: m_c_BinOp(L: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)),
1799 R: m_c_And(L: m_Deferred(V: A), R: m_Deferred(V: B)))))
1800 return BinaryOperator::CreateOr(V1: A, V2: B);
1801
1802 // (add (or A, B) (and A, B)) --> (add A, B)
1803 // (add (and A, B) (or A, B)) --> (add A, B)
1804 if (match(V: &I, P: m_c_BinOp(L: m_Or(L: m_Value(V&: A), R: m_Value(V&: B)),
1805 R: m_c_And(L: m_Deferred(V: A), R: m_Deferred(V: B))))) {
1806 // Replacing operands in-place to preserve nuw/nsw flags.
1807 replaceOperand(I, OpNum: 0, V: A);
1808 replaceOperand(I, OpNum: 1, V: B);
1809 return &I;
1810 }
1811
1812 // (add A (or A, -A)) --> (and (add A, -1) A)
1813 // (add A (or -A, A)) --> (and (add A, -1) A)
1814 // (add (or A, -A) A) --> (and (add A, -1) A)
1815 // (add (or -A, A) A) --> (and (add A, -1) A)
1816 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)),
1817 R: m_Deferred(V: A)))))) {
1818 Value *Add =
1819 Builder.CreateAdd(LHS: A, RHS: Constant::getAllOnesValue(Ty: A->getType()), Name: "",
1820 HasNUW: I.hasNoUnsignedWrap(), HasNSW: I.hasNoSignedWrap());
1821 return BinaryOperator::CreateAnd(V1: Add, V2: A);
1822 }
1823
1824 // Align-up idiom:
1825 // X + ((-X) & (C - 1)) --> (X + C - 1) & -C
1826 // ((X - 1) | (C - 1)) + 1 -> (X + C - 1) & -C
1827 // For a power-of-two C. Note -C == ~(C - 1), so the mask is simply the
1828 // inverted low-bit mask.
1829 {
1830 const APInt *LowMask;
1831 if (match(V: &I,
1832 P: m_c_Add(L: m_OneUse(SubPattern: m_And(L: m_Neg(V: m_Value(V&: A)), R: m_LowBitMask(V&: LowMask))),
1833 R: m_Deferred(V: A))) ||
1834 match(V: &I, P: m_Add(L: m_OneUse(SubPattern: m_Or(L: m_Add(L: m_Value(V&: A), R: m_AllOnes()),
1835 R: m_LowBitMask(V&: LowMask))),
1836 R: m_One()))
1837
1838 ) {
1839 Value *NewAdd = Builder.CreateAdd(LHS: A, RHS: ConstantInt::get(Ty, V: *LowMask));
1840 return BinaryOperator::CreateAnd(V1: NewAdd, V2: ConstantInt::get(Ty, V: ~*LowMask));
1841 }
1842 }
1843
1844 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1845 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1846 if (match(V: &I,
1847 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))))),
1848 R: m_AllOnes()))) {
1849 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty: RHS->getType());
1850 Value *Dec = Builder.CreateAdd(LHS: A, RHS: AllOnes);
1851 Value *Not = Builder.CreateXor(LHS: A, RHS: AllOnes);
1852 return BinaryOperator::CreateAnd(V1: Dec, V2: Not);
1853 }
1854
1855 // Disguised reassociation/factorization:
1856 // ~(A * C1) + A
1857 // ((A * -C1) - 1) + A
1858 // ((A * -C1) + A) - 1
1859 // (A * (1 - C1)) - 1
1860 if (match(V: &I,
1861 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))))),
1862 R: m_Deferred(V: A)))) {
1863 Type *Ty = I.getType();
1864 Constant *NewMulC = ConstantInt::get(Ty, V: 1 - *C1);
1865 Value *NewMul = Builder.CreateMul(LHS: A, RHS: NewMulC);
1866 return BinaryOperator::CreateAdd(V1: NewMul, V2: ConstantInt::getAllOnesValue(Ty));
1867 }
1868
1869 // (A * -2**C) + B --> B - (A << C)
1870 const APInt *NegPow2C;
1871 if (match(V: &I, P: m_c_Add(L: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: A), R: m_NegatedPower2(V&: NegPow2C))),
1872 R: m_Value(V&: B)))) {
1873 Constant *ShiftAmtC = ConstantInt::get(Ty, V: NegPow2C->countr_zero());
1874 Value *Shl = Builder.CreateShl(LHS: A, RHS: ShiftAmtC);
1875 return BinaryOperator::CreateSub(V1: B, V2: Shl);
1876 }
1877
1878 // Canonicalize signum variant that ends in add:
1879 // (A s>> (BW - 1)) + (zext (A s> 0)) --> (A s>> (BW - 1)) | (zext (A != 0))
1880 uint64_t BitWidth = Ty->getScalarSizeInBits();
1881 if (match(V: LHS, P: m_AShr(L: m_Value(V&: A), R: m_SpecificIntAllowPoison(V: BitWidth - 1))) &&
1882 match(V: RHS, P: m_OneUse(SubPattern: m_ZExt(Op: m_OneUse(SubPattern: m_SpecificICmp(
1883 MatchPred: CmpInst::ICMP_SGT, L: m_Specific(V: A), R: m_ZeroInt())))))) {
1884 Value *NotZero = Builder.CreateIsNotNull(Arg: A, Name: "isnotnull");
1885 Value *Zext = Builder.CreateZExt(V: NotZero, DestTy: Ty, Name: "isnotnull.zext");
1886 return BinaryOperator::CreateOr(V1: LHS, V2: Zext);
1887 }
1888
1889 {
1890 Value *Cond, *Ext;
1891 Constant *C;
1892 // (add X, (sext/zext (icmp eq X, C)))
1893 // -> (select (icmp eq X, C), (add C, (sext/zext 1)), X)
1894 auto CondMatcher =
1895 m_Value(V&: Cond, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Deferred(V: A),
1896 R: m_ImmConstant(C)));
1897
1898 if (match(V: &I,
1899 P: m_c_Add(L: m_Value(V&: A), R: m_Value(V&: Ext, P: m_ZExtOrSExt(Op: CondMatcher)))) &&
1900 Ext->hasOneUse()) {
1901 Value *Add = isa<ZExtInst>(Val: Ext) ? InstCombiner::AddOne(C)
1902 : InstCombiner::SubOne(C);
1903 return replaceInstUsesWith(I, V: Builder.CreateSelect(C: Cond, True: Add, False: A));
1904 }
1905 }
1906
1907 // (add (add A, 1), (sext (icmp ne A, 0))) => call umax(A, 1)
1908 if (match(V: LHS, P: m_Add(L: m_Value(V&: A), R: m_One())) &&
1909 match(V: RHS, P: m_OneUse(SubPattern: m_SExt(Op: m_OneUse(SubPattern: m_SpecificICmp(
1910 MatchPred: ICmpInst::ICMP_NE, L: m_Specific(V: A), R: m_ZeroInt())))))) {
1911 Value *OneConst = ConstantInt::get(Ty: A->getType(), V: 1);
1912 Value *UMax = Builder.CreateBinaryIntrinsic(ID: Intrinsic::umax, LHS: A, RHS: OneConst);
1913 return replaceInstUsesWith(I, V: UMax);
1914 }
1915
1916 if (Instruction *Ashr = foldAddToAshr(Add&: I))
1917 return Ashr;
1918
1919 // Ceiling division by power-of-2:
1920 // (X >> log2(N)) + zext(X & (N-1) != 0) --> (X + (N-1)) >> log2(N)
1921 // This is valid when adding (N-1) to X doesn't overflow.
1922 {
1923 Value *X;
1924 const APInt *ShiftAmt, *Mask;
1925 CmpPredicate Pred;
1926
1927 // Match: (X >> C) + zext((X & Mask) != 0)
1928 // or: zext((X & Mask) != 0) + (X >> C)
1929 if (match(V: &I,
1930 P: m_c_Add(L: m_ZExt(Op: m_SpecificICmp(
1931 MatchPred: ICmpInst::ICMP_NE,
1932 L: m_And(L: m_Value(V&: X), R: m_LowBitMask(V&: Mask)), R: m_ZeroInt())),
1933 R: m_OneUse(SubPattern: m_ZExtOrSelf(Op: m_OneUse(
1934 SubPattern: m_LShr(L: m_Deferred(V: X), R: m_APInt(Res&: ShiftAmt))))))) &&
1935 Mask->popcount() == *ShiftAmt) {
1936
1937 // Check if X + Mask doesn't overflow
1938 unsigned Xbits = X->getType()->getScalarSizeInBits();
1939 unsigned Ibits = Ty->getScalarSizeInBits();
1940 bool NeedZext = Ibits > Xbits;
1941 Constant *MaskC = ConstantInt::get(Ty, V: Mask->zext(width: Ibits));
1942 if (NeedZext || willNotOverflowUnsignedAdd(LHS: X, RHS: MaskC, CxtI: I)) {
1943 if (NeedZext)
1944 X = Builder.CreateZExt(V: X, DestTy: Ty);
1945 // (X + Mask) >> ShiftAmt
1946 Value *Add = Builder.CreateNUWAdd(LHS: X, RHS: MaskC);
1947 return BinaryOperator::CreateLShr(
1948 V1: Add, V2: ConstantInt::get(Ty, V: ShiftAmt->zext(width: Ibits)));
1949 }
1950 }
1951 }
1952
1953 // (~X) + (~Y) --> -2 - (X + Y)
1954 {
1955 // To ensure we can save instructions we need to ensure that we consume both
1956 // LHS/RHS (i.e they have a `not`).
1957 bool ConsumesLHS, ConsumesRHS;
1958 if (isFreeToInvert(V: LHS, WillInvertAllUses: LHS->hasOneUse(), DoesConsume&: ConsumesLHS) && ConsumesLHS &&
1959 isFreeToInvert(V: RHS, WillInvertAllUses: RHS->hasOneUse(), DoesConsume&: ConsumesRHS) && ConsumesRHS) {
1960 Value *NotLHS = getFreelyInverted(V: LHS, WillInvertAllUses: LHS->hasOneUse(), Builder: &Builder);
1961 Value *NotRHS = getFreelyInverted(V: RHS, WillInvertAllUses: RHS->hasOneUse(), Builder: &Builder);
1962 assert(NotLHS != nullptr && NotRHS != nullptr &&
1963 "isFreeToInvert desynced with getFreelyInverted");
1964 Value *LHSPlusRHS = Builder.CreateAdd(LHS: NotLHS, RHS: NotRHS);
1965 return BinaryOperator::CreateSub(
1966 V1: ConstantInt::getSigned(Ty: RHS->getType(), V: -2), V2: LHSPlusRHS);
1967 }
1968 }
1969
1970 if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &I))
1971 return R;
1972
1973 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1974 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1975 // computeKnownBits.
1976 bool Changed = false;
1977 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS: LHSCache, RHS: RHSCache, CxtI: I)) {
1978 Changed = true;
1979 I.setHasNoSignedWrap(true);
1980 }
1981 if (!I.hasNoUnsignedWrap() &&
1982 willNotOverflowUnsignedAdd(LHS: LHSCache, RHS: RHSCache, CxtI: I)) {
1983 Changed = true;
1984 I.setHasNoUnsignedWrap(true);
1985 }
1986
1987 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1988 return V;
1989
1990 if (Instruction *V =
1991 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1992 return V;
1993
1994 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1995 return SatAdd;
1996
1997 // usub.sat(A, B) + B => umax(A, B)
1998 if (match(V: &I, P: m_c_BinOp(
1999 L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::usub_sat>(Ops: m_Value(V&: A), Ops: m_Value(V&: B))),
2000 R: m_Deferred(V: B)))) {
2001 return replaceInstUsesWith(I,
2002 V: Builder.CreateIntrinsic(ID: Intrinsic::umax, OverloadTypes: {I.getType()}, Args: {A, B}));
2003 }
2004
2005 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
2006 if (match(V: LHS, P: m_OneUse(SubPattern: m_Ctpop(Op0: m_Value(V&: A)))) &&
2007 match(V: RHS, P: m_OneUse(SubPattern: m_Ctpop(Op0: m_Value(V&: B)))) &&
2008 haveNoCommonBitsSet(LHSCache: A, RHSCache: B, SQ: SQ.getWithInstruction(I: &I)))
2009 return replaceInstUsesWith(
2010 I, V: Builder.CreateIntrinsic(ID: Intrinsic::ctpop, OverloadTypes: {I.getType()},
2011 Args: {Builder.CreateDisjointOr(LHS: A, RHS: B)}));
2012
2013 // Fold the log2_ceil idiom:
2014 // zext(ctpop(A) >u/!= 1) + (ctlz(A, true) ^ (BW - 1))
2015 // -->
2016 // BW - ctlz(A - 1, false)
2017 const APInt *XorC;
2018 CmpPredicate Pred;
2019 if (match(V: &I, P: m_c_Add(L: m_ZExt(Op: m_ICmp(Pred, L: m_Ctpop(Op0: m_Value(V&: A)), R: m_One())),
2020 R: m_OneUse(SubPattern: m_ZExtOrSelf(Op: m_OneUse(
2021 SubPattern: m_Xor(L: m_OneUse(SubPattern: m_TruncOrSelf(Op: m_OneUse(
2022 SubPattern: m_Ctlz(Op0: m_Deferred(V: A), Op1: m_One())))),
2023 R: m_APInt(Res&: XorC))))))) &&
2024 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_NE) &&
2025 *XorC == A->getType()->getScalarSizeInBits() - 1) {
2026 Value *Sub = Builder.CreateAdd(LHS: A, RHS: Constant::getAllOnesValue(Ty: A->getType()));
2027 Value *Ctlz = Builder.CreateIntrinsic(ID: Intrinsic::ctlz, OverloadTypes: {A->getType()},
2028 Args: {Sub, Builder.getFalse()});
2029 Value *Ret = Builder.CreateSub(
2030 LHS: ConstantInt::get(Ty: A->getType(), V: A->getType()->getScalarSizeInBits()),
2031 RHS: Ctlz, Name: "", /*HasNUW=*/true, /*HasNSW=*/true);
2032 return replaceInstUsesWith(I, V: Builder.CreateZExtOrTrunc(V: Ret, DestTy: I.getType()));
2033 }
2034
2035 if (Instruction *Res = foldSquareSumInt(I))
2036 return Res;
2037
2038 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
2039 return Res;
2040
2041 if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
2042 return Res;
2043
2044 if (Instruction *Res = foldDivCeil(I))
2045 return Res;
2046
2047 // Re-enqueue users of the induction variable of add recurrence if we infer
2048 // new nuw/nsw flags.
2049 if (Changed) {
2050 PHINode *PHI;
2051 Value *Start, *Step;
2052 if (matchSimpleRecurrence(I: &I, P&: PHI, Start, Step))
2053 Worklist.pushUsersToWorkList(I&: *PHI);
2054 }
2055
2056 return Changed ? &I : nullptr;
2057}
2058
2059/// Eliminate an op from a linear interpolation (lerp) pattern.
2060static Instruction *factorizeLerp(BinaryOperator &I,
2061 InstCombiner::BuilderTy &Builder) {
2062 Value *X, *Y, *Z;
2063 if (!match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_c_FMul(L: m_Value(V&: Y),
2064 R: m_OneUse(SubPattern: m_FSub(L: m_FPOne(),
2065 R: m_Value(V&: Z))))),
2066 R: m_OneUse(SubPattern: m_c_FMul(L: m_Value(V&: X), R: m_Deferred(V: Z))))))
2067 return nullptr;
2068
2069 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
2070 Value *XY = Builder.CreateFSubFMF(L: X, R: Y, FMFSource: &I);
2071 Value *MulZ = Builder.CreateFMulFMF(L: Z, R: XY, FMFSource: &I);
2072 return BinaryOperator::CreateFAddFMF(V1: Y, V2: MulZ, FMFSource: &I);
2073}
2074
2075/// Factor a common operand out of fadd/fsub of fmul/fdiv.
2076static Instruction *factorizeFAddFSub(BinaryOperator &I,
2077 InstCombiner::BuilderTy &Builder) {
2078 assert((I.getOpcode() == Instruction::FAdd ||
2079 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
2080 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
2081 "FP factorization requires FMF");
2082
2083 if (Instruction *Lerp = factorizeLerp(I, Builder))
2084 return Lerp;
2085
2086 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
2087 if (!Op0->hasOneUse() || !Op1->hasOneUse())
2088 return nullptr;
2089
2090 Value *X, *Y, *Z;
2091 bool IsFMul;
2092 if ((match(V: Op0, P: m_FMul(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
2093 match(V: Op1, P: m_c_FMul(L: m_Value(V&: Y), R: m_Specific(V: Z)))) ||
2094 (match(V: Op0, P: m_FMul(L: m_Value(V&: Z), R: m_Value(V&: X))) &&
2095 match(V: Op1, P: m_c_FMul(L: m_Value(V&: Y), R: m_Specific(V: Z)))))
2096 IsFMul = true;
2097 else if (match(V: Op0, P: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
2098 match(V: Op1, P: m_FDiv(L: m_Value(V&: Y), R: m_Specific(V: Z))))
2099 IsFMul = false;
2100 else
2101 return nullptr;
2102
2103 // (X * Z) + (Y * Z) --> (X + Y) * Z
2104 // (X * Z) - (Y * Z) --> (X - Y) * Z
2105 // (X / Z) + (Y / Z) --> (X + Y) / Z
2106 // (X / Z) - (Y / Z) --> (X - Y) / Z
2107 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
2108 Value *XY = IsFAdd ? Builder.CreateFAddFMF(L: X, R: Y, FMFSource: &I)
2109 : Builder.CreateFSubFMF(L: X, R: Y, FMFSource: &I);
2110
2111 // Bail out if we just created a denormal constant.
2112 // TODO: This is copied from a previous implementation. Is it necessary?
2113 const APFloat *C;
2114 if (match(V: XY, P: m_APFloat(Res&: C)) && !C->isNormal())
2115 return nullptr;
2116
2117 return IsFMul ? BinaryOperator::CreateFMulFMF(V1: XY, V2: Z, FMFSource: &I)
2118 : BinaryOperator::CreateFDivFMF(V1: XY, V2: Z, FMFSource: &I);
2119}
2120
2121Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
2122 if (Value *V = simplifyFAddInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
2123 FMF: I.getFastMathFlags(),
2124 Q: SQ.getWithInstruction(I: &I)))
2125 return replaceInstUsesWith(I, V);
2126
2127 if (SimplifyAssociativeOrCommutative(I))
2128 return &I;
2129
2130 if (Instruction *X = foldVectorBinop(Inst&: I))
2131 return X;
2132
2133 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
2134 return Phi;
2135
2136 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
2137 return FoldedFAdd;
2138
2139 // B = fadd A, 0.0
2140 // Z = Op B
2141 // can be transformed into
2142 // Z = Op A
2143 // Where Op is such that we can ignore sign of 0 in fadd
2144 Value *A;
2145 if (match(V: &I, P: m_OneUse(SubPattern: m_FAdd(L: m_Value(V&: A), R: m_AnyZeroFP()))) &&
2146 canIgnoreSignBitOfZero(U: *I.use_begin()))
2147 return replaceInstUsesWith(I, V: A);
2148
2149 // (-X) + Y --> Y - X
2150 Value *X, *Y;
2151 if (match(V: &I, P: m_c_FAdd(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y))))
2152 return BinaryOperator::CreateFSubFMF(V1: Y, V2: X, FMFSource: &I);
2153
2154 // Similar to above, but look through fmul/fdiv for the negated term.
2155 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
2156 Value *Z;
2157 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))),
2158 R: m_Value(V&: Z)))) {
2159 Value *XY = Builder.CreateFMulFMF(L: X, R: Y, FMFSource: &I);
2160 return BinaryOperator::CreateFSubFMF(V1: Z, V2: XY, FMFSource: &I);
2161 }
2162 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
2163 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
2164 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))),
2165 R: m_Value(V&: Z))) ||
2166 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)))),
2167 R: m_Value(V&: Z)))) {
2168 Value *XY = Builder.CreateFDivFMF(L: X, R: Y, FMFSource: &I);
2169 return BinaryOperator::CreateFSubFMF(V1: Z, V2: XY, FMFSource: &I);
2170 }
2171
2172 // Check for (fadd double (sitofp x), y), see if we can merge this into an
2173 // integer add followed by a promotion.
2174 if (Instruction *R = foldFBinOpOfIntCasts(I))
2175 return R;
2176
2177 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
2178 // Handle specials cases for FAdd with selects feeding the operation
2179 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
2180 return replaceInstUsesWith(I, V);
2181
2182 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2183 if (Instruction *F = factorizeFAddFSub(I, Builder))
2184 return F;
2185
2186 if (Instruction *F = foldSquareSumFP(I))
2187 return F;
2188
2189 // Try to fold fadd into start value of reduction intrinsic.
2190 if (match(V: &I, P: m_c_FAdd(L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_fadd>(
2191 Ops: m_AnyZeroFP(), Ops: m_Value(V&: X))),
2192 R: m_Value(V&: Y)))) {
2193 // fadd (rdx 0.0, X), Y --> rdx Y, X
2194 return replaceInstUsesWith(
2195 I, V: Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_fadd,
2196 OverloadTypes: {X->getType()}, Args: {Y, X}, FMFSource: &I));
2197 }
2198 const APFloat *StartC, *C;
2199 if (match(V: LHS, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_fadd>(
2200 Ops: m_APFloat(Res&: StartC), Ops: m_Value(V&: X)))) &&
2201 match(V: RHS, P: m_APFloat(Res&: C))) {
2202 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
2203 Constant *NewStartC = ConstantFP::get(Ty: I.getType(), V: *C + *StartC);
2204 return replaceInstUsesWith(
2205 I, V: Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_fadd,
2206 OverloadTypes: {X->getType()}, Args: {NewStartC, X}, FMFSource: &I));
2207 }
2208
2209 // (X * MulC) + X --> X * (MulC + 1.0)
2210 Constant *MulC;
2211 if (match(V: &I, P: m_c_FAdd(L: m_FMul(L: m_Value(V&: X), R: m_ImmConstant(C&: MulC)),
2212 R: m_Deferred(V: X)))) {
2213 if (Constant *NewMulC = ConstantFoldBinaryOpOperands(
2214 Opcode: Instruction::FAdd, LHS: MulC, RHS: ConstantFP::get(Ty: I.getType(), V: 1.0), DL))
2215 return BinaryOperator::CreateFMulFMF(V1: X, V2: NewMulC, FMFSource: &I);
2216 }
2217
2218 // (-X - Y) + (X + Z) --> Z - Y
2219 if (match(V: &I, P: m_c_FAdd(L: m_FSub(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y)),
2220 R: m_c_FAdd(L: m_Deferred(V: X), R: m_Value(V&: Z)))))
2221 return BinaryOperator::CreateFSubFMF(V1: Z, V2: Y, FMFSource: &I);
2222
2223 if (Value *V = FAddCombine(Builder).simplify(I: &I))
2224 return replaceInstUsesWith(I, V);
2225 }
2226
2227 // minumum(X, Y) + maximum(X, Y) => X + Y.
2228 if (match(V: &I,
2229 P: m_c_FAdd(L: m_Intrinsic<Intrinsic::maximum>(Ops: m_Value(V&: X), Ops: m_Value(V&: Y)),
2230 R: m_c_Intrinsic<Intrinsic::minimum>(Op0: m_Deferred(V: X),
2231 Op1: m_Deferred(V: Y))))) {
2232 BinaryOperator *Result = BinaryOperator::CreateFAddFMF(V1: X, V2: Y, FMFSource: &I);
2233 // We cannot preserve ninf if nnan flag is not set.
2234 // If X is NaN and Y is Inf then in original program we had NaN + NaN,
2235 // while in optimized version NaN + Inf and this is a poison with ninf flag.
2236 if (!Result->hasNoNaNs())
2237 Result->setHasNoInfs(false);
2238 return Result;
2239 }
2240
2241 return nullptr;
2242}
2243
2244CommonPointerBase CommonPointerBase::compute(Value *LHS, Value *RHS) {
2245 CommonPointerBase Base;
2246
2247 if (LHS->getType() != RHS->getType())
2248 return Base;
2249
2250 // Collect all base pointers of LHS.
2251 SmallPtrSet<Value *, 16> Ptrs;
2252 Value *Ptr = LHS;
2253 while (true) {
2254 Ptrs.insert(Ptr);
2255 if (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr))
2256 Ptr = GEP->getPointerOperand();
2257 else
2258 break;
2259 }
2260
2261 // Find common base and collect RHS GEPs.
2262 bool First = true;
2263 while (true) {
2264 if (Ptrs.contains(Ptr: RHS)) {
2265 Base.Ptr = RHS;
2266 break;
2267 }
2268
2269 if (auto *GEP = dyn_cast<GEPOperator>(Val: RHS)) {
2270 Base.RHSGEPs.push_back(Elt: GEP);
2271 if (First) {
2272 First = false;
2273 Base.RHSNW = GEP->getNoWrapFlags();
2274 } else {
2275 Base.RHSNW = Base.RHSNW.intersectForOffsetAdd(Other: GEP->getNoWrapFlags());
2276 }
2277 RHS = GEP->getPointerOperand();
2278 } else {
2279 // No common base.
2280 return Base;
2281 }
2282 }
2283
2284 // Collect LHS GEPs.
2285 First = true;
2286 while (true) {
2287 if (LHS == Base.Ptr)
2288 break;
2289
2290 auto *GEP = cast<GEPOperator>(Val: LHS);
2291 Base.LHSGEPs.push_back(Elt: GEP);
2292 if (First) {
2293 First = false;
2294 Base.LHSNW = GEP->getNoWrapFlags();
2295 } else {
2296 Base.LHSNW = Base.LHSNW.intersectForOffsetAdd(Other: GEP->getNoWrapFlags());
2297 }
2298 LHS = GEP->getPointerOperand();
2299 }
2300
2301 return Base;
2302}
2303
2304bool CommonPointerBase::isExpensive() const {
2305 unsigned NumGEPs = 0;
2306 auto ProcessGEPs = [&NumGEPs](ArrayRef<GEPOperator *> GEPs) {
2307 bool SeenMultiUse = false;
2308 for (GEPOperator *GEP : GEPs) {
2309 // Only count multi-use GEPs, excluding the first one. For the first one,
2310 // we will directly reuse the offset. For one-use GEPs, their offset will
2311 // be folded into a multi-use GEP.
2312 if (!GEP->hasOneUse()) {
2313 if (SeenMultiUse)
2314 ++NumGEPs;
2315 SeenMultiUse = true;
2316 }
2317 }
2318 };
2319 ProcessGEPs(LHSGEPs);
2320 ProcessGEPs(RHSGEPs);
2321 return NumGEPs > 2;
2322}
2323
2324/// Optimize pointer differences into the same array into a size. Consider:
2325/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
2326/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
2327Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
2328 Type *Ty, bool IsNUW) {
2329 CommonPointerBase Base = CommonPointerBase::compute(LHS, RHS);
2330 if (!Base.Ptr || Base.isExpensive())
2331 return nullptr;
2332
2333 // To avoid duplicating the offset arithmetic, rewrite the GEP to use the
2334 // computed offset.
2335 // TODO: We should probably do this even if there is only one GEP.
2336 bool RewriteGEPs = !Base.LHSGEPs.empty() && !Base.RHSGEPs.empty();
2337
2338 Type *IdxTy = DL.getIndexType(PtrTy: LHS->getType());
2339 Value *Result = EmitGEPOffsets(GEPs: Base.LHSGEPs, NW: Base.LHSNW, IdxTy, RewriteGEPs);
2340 Value *Offset2 = EmitGEPOffsets(GEPs: Base.RHSGEPs, NW: Base.RHSNW, IdxTy, RewriteGEPs);
2341
2342 // If this is a single inbounds GEP and the original sub was nuw,
2343 // then the final multiplication is also nuw.
2344 if (auto *I = dyn_cast<OverflowingBinaryOperator>(Val: Result))
2345 if (IsNUW && match(V: Offset2, P: m_Zero()) && Base.LHSNW.isInBounds() &&
2346 (I->use_empty() || I->hasOneUse()) && I->hasNoSignedWrap() &&
2347 !I->hasNoUnsignedWrap() &&
2348 ((I->getOpcode() == Instruction::Mul &&
2349 match(V: I->getOperand(i_nocapture: 1), P: m_NonNegative())) ||
2350 I->getOpcode() == Instruction::Shl))
2351 cast<Instruction>(Val: I)->setHasNoUnsignedWrap();
2352
2353 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
2354 // If both GEPs are inbounds, then the subtract does not have signed overflow.
2355 // If both GEPs are nuw and the original sub is nuw, the new sub is also nuw.
2356 if (!match(V: Offset2, P: m_Zero())) {
2357 Result =
2358 Builder.CreateSub(LHS: Result, RHS: Offset2, Name: "gepdiff",
2359 HasNUW: IsNUW && Base.LHSNW.hasNoUnsignedWrap() &&
2360 Base.RHSNW.hasNoUnsignedWrap(),
2361 HasNSW: Base.LHSNW.isInBounds() && Base.RHSNW.isInBounds());
2362 }
2363
2364 return Builder.CreateIntCast(V: Result, DestTy: Ty, isSigned: true);
2365}
2366
2367static Instruction *foldSubOfMinMax(BinaryOperator &I,
2368 InstCombiner::BuilderTy &Builder) {
2369 Value *Op0 = I.getOperand(i_nocapture: 0);
2370 Value *Op1 = I.getOperand(i_nocapture: 1);
2371 Type *Ty = I.getType();
2372 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op1);
2373 if (!MinMax)
2374 return nullptr;
2375
2376 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2377 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2378 Value *X = MinMax->getLHS();
2379 Value *Y = MinMax->getRHS();
2380 if (match(V: Op0, P: m_c_Add(L: m_Specific(V: X), R: m_Specific(V: Y))) &&
2381 (Op0->hasOneUse() || Op1->hasOneUse())) {
2382 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: MinMax->getIntrinsicID());
2383 Function *F = Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: InvID, OverloadTys: Ty);
2384 return CallInst::Create(Func: F, Args: {X, Y});
2385 }
2386
2387 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
2388 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
2389 Value *Z;
2390 if (match(V: Op1, P: m_OneUse(SubPattern: m_UMin(Op0: m_Value(V&: Y), Op1: m_Value(V&: Z))))) {
2391 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Y), R: m_Value(V&: X))))) {
2392 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, OverloadTypes: Ty, Args: {Y, Z});
2393 return BinaryOperator::CreateAdd(V1: X, V2: USub);
2394 }
2395 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Z), R: m_Value(V&: X))))) {
2396 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, OverloadTypes: Ty, Args: {Z, Y});
2397 return BinaryOperator::CreateAdd(V1: X, V2: USub);
2398 }
2399 }
2400
2401 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
2402 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
2403 if (MinMax->isSigned() && match(V: Y, P: m_ZeroInt()) &&
2404 match(V: X, P: m_NSWSub(L: m_Specific(V: Op0), R: m_Value(V&: Z)))) {
2405 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: MinMax->getIntrinsicID());
2406 Function *F = Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: InvID, OverloadTys: Ty);
2407 return CallInst::Create(Func: F, Args: {Op0, Z});
2408 }
2409
2410 return nullptr;
2411}
2412
2413Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
2414 if (Value *V = simplifySubInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
2415 IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(),
2416 Q: SQ.getWithInstruction(I: &I)))
2417 return replaceInstUsesWith(I, V);
2418
2419 if (Instruction *X = foldVectorBinop(Inst&: I))
2420 return X;
2421
2422 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
2423 return Phi;
2424
2425 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
2426
2427 // If this is a 'B = x-(-A)', change to B = x+A.
2428 // We deal with this without involving Negator to preserve NSW flag.
2429 if (Value *V = dyn_castNegVal(V: Op1)) {
2430 BinaryOperator *Res = BinaryOperator::CreateAdd(V1: Op0, V2: V);
2431
2432 if (const auto *BO = dyn_cast<BinaryOperator>(Val: Op1)) {
2433 assert(BO->getOpcode() == Instruction::Sub &&
2434 "Expected a subtraction operator!");
2435 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
2436 Res->setHasNoSignedWrap(true);
2437 } else {
2438 if (cast<Constant>(Val: Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
2439 Res->setHasNoSignedWrap(true);
2440 }
2441
2442 return Res;
2443 }
2444
2445 // Try this before Negator to preserve NSW flag.
2446 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
2447 return R;
2448
2449 Constant *C;
2450 if (match(V: Op0, P: m_ImmConstant(C))) {
2451 Value *X;
2452 Constant *C2;
2453
2454 // C-(X+C2) --> (C-C2)-X
2455 if (match(V: Op1, P: m_AddLike(L: m_Value(V&: X), R: m_ImmConstant(C&: C2)))) {
2456 // C-C2 never overflow, and C-(X+C2), (X+C2) has NSW/NUW
2457 // => (C-C2)-X can have NSW/NUW
2458 bool WillNotSOV = willNotOverflowSignedSub(LHS: C, RHS: C2, CxtI: I);
2459 BinaryOperator *Res =
2460 BinaryOperator::CreateSub(V1: ConstantExpr::getSub(C1: C, C2), V2: X);
2461
2462 // or disjoint is equivalent to add nuw nsw.
2463 bool Op1NSW = true;
2464 bool Op1NUW = true;
2465
2466 if (auto *OBO1 = dyn_cast<OverflowingBinaryOperator>(Val: Op1)) {
2467 Op1NSW = OBO1->hasNoSignedWrap();
2468 Op1NUW = OBO1->hasNoUnsignedWrap();
2469 }
2470
2471 Res->setHasNoSignedWrap(I.hasNoSignedWrap() && Op1NSW && WillNotSOV);
2472 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() && Op1NUW);
2473 return Res;
2474 }
2475 }
2476
2477 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
2478 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2479 return Ext;
2480
2481 bool Changed = false;
2482 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(LHS: Op0, RHS: Op1, CxtI: I)) {
2483 Changed = true;
2484 I.setHasNoSignedWrap(true);
2485 }
2486 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(LHS: Op0, RHS: Op1, CxtI: I)) {
2487 Changed = true;
2488 I.setHasNoUnsignedWrap(true);
2489 }
2490
2491 return Changed ? &I : nullptr;
2492 };
2493
2494 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
2495 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
2496 // a pure negation used by a select that looks like abs/nabs.
2497 bool IsNegation = match(V: Op0, P: m_ZeroInt());
2498 if (!IsNegation || none_of(Range: I.users(), P: match_fn(P: m_c_Select(L: m_Specific(V: Op1),
2499 R: m_Specific(V: &I))))) {
2500 if (Value *NegOp1 = Negator::Negate(LHSIsZero: IsNegation, /* IsNSW */ IsNegation &&
2501 I.hasNoSignedWrap(),
2502 Root: Op1, IC&: *this))
2503 return BinaryOperator::CreateAdd(V1: NegOp1, V2: Op0);
2504 }
2505 if (IsNegation)
2506 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
2507
2508 // (A*B)-(A*C) -> A*(B-C) etc
2509 if (Value *V = foldUsingDistributiveLaws(I))
2510 return replaceInstUsesWith(I, V);
2511
2512 if (I.getType()->isIntOrIntVectorTy(BitWidth: 1))
2513 return BinaryOperator::CreateXor(V1: Op0, V2: Op1);
2514
2515 // Replace (-1 - A) with (~A).
2516 if (match(V: Op0, P: m_AllOnes()))
2517 return BinaryOperator::CreateNot(Op: Op1);
2518
2519 // (X + -1) - Y --> ~Y + X
2520 Value *X, *Y;
2521 if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_AllOnes()))))
2522 return BinaryOperator::CreateAdd(V1: Builder.CreateNot(V: Op1), V2: X);
2523
2524 // if (C1 & C2) == C2 then (X & C1) - (X & C2) -> X & (C1 ^ C2)
2525 Constant *C1, *C2;
2526 if (match(V: Op0, P: m_And(L: m_Value(V&: X), R: m_ImmConstant(C&: C1))) &&
2527 match(V: Op1, P: m_And(L: m_Specific(V: X), R: m_ImmConstant(C&: C2)))) {
2528 Value *AndC = ConstantFoldBinaryInstruction(Opcode: Instruction::And, V1: C1, V2: C2);
2529 if (C2->isElementWiseEqual(Y: AndC))
2530 return BinaryOperator::CreateAnd(
2531 V1: X, V2: ConstantFoldBinaryInstruction(Opcode: Instruction::Xor, V1: C1, V2: C2));
2532 }
2533
2534 // Reassociate sub/add sequences to create more add instructions and
2535 // reduce dependency chains:
2536 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2537 Value *Z;
2538 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))),
2539 R: m_Value(V&: Z))))) {
2540 Value *XZ = Builder.CreateAdd(LHS: X, RHS: Z);
2541 Value *YW = Builder.CreateAdd(LHS: Y, RHS: Op1);
2542 return BinaryOperator::CreateSub(V1: XZ, V2: YW);
2543 }
2544
2545 // ((X - Y) - Op1) --> X - (Y + Op1)
2546 if (match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
2547 OverflowingBinaryOperator *LHSSub = cast<OverflowingBinaryOperator>(Val: Op0);
2548 bool HasNUW = I.hasNoUnsignedWrap() && LHSSub->hasNoUnsignedWrap();
2549 bool HasNSW = HasNUW && I.hasNoSignedWrap() && LHSSub->hasNoSignedWrap();
2550 Value *Add = Builder.CreateAdd(LHS: Y, RHS: Op1, Name: "", /*HasNUW=*/HasNUW,
2551 /*HasNSW=*/HasNSW);
2552 BinaryOperator *Sub = BinaryOperator::CreateSub(V1: X, V2: Add);
2553 Sub->setHasNoUnsignedWrap(HasNUW);
2554 Sub->setHasNoSignedWrap(HasNSW);
2555 return Sub;
2556 }
2557
2558 // (X + C0) - (Y + C1) --> (X - Y) + (C0 - C1)
2559 {
2560 Constant *CX, *CY;
2561 if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_ImmConstant(C&: CX)))) &&
2562 match(V: Op1, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: Y), R: m_ImmConstant(C&: CY))))) {
2563 Value *OpsSub = Builder.CreateSub(LHS: X, RHS: Y);
2564 Constant *ConstsSub = ConstantExpr::getSub(C1: CX, C2: CY);
2565 return BinaryOperator::CreateAdd(V1: OpsSub, V2: ConstsSub);
2566 }
2567 }
2568
2569 // (X + Z) - (Y + Z) --> (X - Y)
2570 {
2571 Value *W, *Z;
2572 if (match(V: Op0, P: m_AddLike(L: m_Value(V&: W), R: m_Value(V&: X))) &&
2573 match(V: Op1, P: m_AddLike(L: m_Value(V&: Y), R: m_Value(V&: Z)))) {
2574 Instruction *R = nullptr;
2575 if (W == Y)
2576 R = BinaryOperator::CreateSub(V1: X, V2: Z);
2577 else if (W == Z)
2578 R = BinaryOperator::CreateSub(V1: X, V2: Y);
2579 else if (X == Y)
2580 R = BinaryOperator::CreateSub(V1: W, V2: Z);
2581 else if (X == Z)
2582 R = BinaryOperator::CreateSub(V1: W, V2: Y);
2583 if (R) {
2584 bool NSW = I.hasNoSignedWrap() &&
2585 match(V: Op0, P: m_NSWAddLike(L: m_Value(), R: m_Value())) &&
2586 match(V: Op1, P: m_NSWAddLike(L: m_Value(), R: m_Value()));
2587
2588 bool NUW = I.hasNoUnsignedWrap() &&
2589 match(V: Op1, P: m_NUWAddLike(L: m_Value(), R: m_Value()));
2590 R->setHasNoSignedWrap(NSW);
2591 R->setHasNoUnsignedWrap(NUW);
2592 return R;
2593 }
2594 }
2595 }
2596
2597 // (~X) - (~Y) --> Y - X
2598 {
2599 // Need to ensure we can consume at least one of the `not` instructions,
2600 // otherwise this can inf loop.
2601 bool ConsumesOp0, ConsumesOp1;
2602 if (isFreeToInvert(V: Op0, WillInvertAllUses: Op0->hasOneUse(), DoesConsume&: ConsumesOp0) &&
2603 isFreeToInvert(V: Op1, WillInvertAllUses: Op1->hasOneUse(), DoesConsume&: ConsumesOp1) &&
2604 (ConsumesOp0 || ConsumesOp1)) {
2605 Value *NotOp0 = getFreelyInverted(V: Op0, WillInvertAllUses: Op0->hasOneUse(), Builder: &Builder);
2606 Value *NotOp1 = getFreelyInverted(V: Op1, WillInvertAllUses: Op1->hasOneUse(), Builder: &Builder);
2607 assert(NotOp0 != nullptr && NotOp1 != nullptr &&
2608 "isFreeToInvert desynced with getFreelyInverted");
2609 return BinaryOperator::CreateSub(V1: NotOp1, V2: NotOp0);
2610 }
2611 }
2612
2613 auto m_AddRdx = [](Value *&Vec) {
2614 return m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_add>(Ops: m_Value(V&: Vec)));
2615 };
2616 Value *V0, *V1;
2617 if (match(V: Op0, P: m_AddRdx(V0)) && match(V: Op1, P: m_AddRdx(V1)) &&
2618 V0->getType() == V1->getType()) {
2619 // Difference of sums is sum of differences:
2620 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
2621 Value *Sub = Builder.CreateSub(LHS: V0, RHS: V1);
2622 Value *Rdx = Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_add,
2623 OverloadTypes: {Sub->getType()}, Args: {Sub});
2624 return replaceInstUsesWith(I, V: Rdx);
2625 }
2626
2627 if (Constant *C = dyn_cast<Constant>(Val: Op0)) {
2628 Value *X;
2629 if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))
2630 // C - (zext bool) --> bool ? C - 1 : C
2631 return SelectInst::Create(C: X, S1: InstCombiner::SubOne(C), S2: C);
2632 if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))
2633 // C - (sext bool) --> bool ? C + 1 : C
2634 return SelectInst::Create(C: X, S1: InstCombiner::AddOne(C), S2: C);
2635
2636 // C - ~X == X + (1+C)
2637 if (match(V: Op1, P: m_Not(V: m_Value(V&: X))))
2638 return BinaryOperator::CreateAdd(V1: X, V2: InstCombiner::AddOne(C));
2639
2640 // Try to fold constant sub into select arguments.
2641 if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1))
2642 if (Instruction *R = FoldOpIntoSelect(Op&: I, SI))
2643 return R;
2644
2645 // Try to fold constant sub into PHI values.
2646 if (PHINode *PN = dyn_cast<PHINode>(Val: Op1))
2647 if (Instruction *R = foldOpIntoPhi(I, PN))
2648 return R;
2649
2650 Constant *C2;
2651
2652 // C-(C2-X) --> X+(C-C2)
2653 if (match(V: Op1, P: m_Sub(L: m_ImmConstant(C&: C2), R: m_Value(V&: X))))
2654 return BinaryOperator::CreateAdd(V1: X, V2: ConstantExpr::getSub(C1: C, C2));
2655 }
2656
2657 const APInt *Op0C;
2658 if (match(V: Op0, P: m_APInt(Res&: Op0C))) {
2659 if (Op0C->isMask()) {
2660 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2661 // zero. We don't use information from dominating conditions so this
2662 // transform is easier to reverse if necessary.
2663 KnownBits RHSKnown = llvm::computeKnownBits(
2664 V: Op1, Q: SQ.getWithInstruction(I: &I).getWithoutDomCondCache());
2665 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2666 return BinaryOperator::CreateXor(V1: Op1, V2: Op0);
2667 }
2668
2669 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2670 // (C3 - ((C2 & C3) - 1)) is pow2
2671 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2672 // C2 is negative pow2 || sub nuw
2673 const APInt *C2, *C3;
2674 BinaryOperator *InnerSub;
2675 if (match(V: Op1, P: m_OneUse(SubPattern: m_And(L: m_BinOp(I&: InnerSub), R: m_APInt(Res&: C2)))) &&
2676 match(V: InnerSub, P: m_Sub(L: m_APInt(Res&: C3), R: m_Value(V&: X))) &&
2677 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2678 APInt C2AndC3 = *C2 & *C3;
2679 APInt C2AndC3Minus1 = C2AndC3 - 1;
2680 APInt C2AddC3 = *C2 + *C3;
2681 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2682 C2AndC3Minus1.isSubsetOf(RHS: C2AddC3)) {
2683 Value *And = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty: I.getType(), V: *C2));
2684 return BinaryOperator::CreateAdd(
2685 V1: And, V2: ConstantInt::get(Ty: I.getType(), V: *Op0C - C2AndC3));
2686 }
2687 }
2688 }
2689
2690 {
2691 Value *Y;
2692 // X-(X+Y) == -Y X-(Y+X) == -Y
2693 if (match(V: Op1, P: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: Y))))
2694 return BinaryOperator::CreateNeg(Op: Y);
2695
2696 // (X-Y)-X == -Y
2697 if (match(V: Op0, P: m_Sub(L: m_Specific(V: Op1), R: m_Value(V&: Y))))
2698 return BinaryOperator::CreateNeg(Op: Y);
2699 }
2700
2701 // (sub (or A, B) (and A, B)) --> (xor A, B)
2702 {
2703 Value *A, *B;
2704 if (match(V: Op1, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2705 match(V: Op0, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2706 return BinaryOperator::CreateXor(V1: A, V2: B);
2707 }
2708
2709 // (sub (add A, B) (or A, B)) --> (and A, B)
2710 {
2711 Value *A, *B;
2712 if (match(V: Op0, P: m_Add(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2713 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2714 return BinaryOperator::CreateAnd(V1: A, V2: B);
2715 }
2716
2717 // (sub (add A, B) (and A, B)) --> (or A, B)
2718 {
2719 Value *A, *B;
2720 if (match(V: Op0, P: m_Add(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2721 match(V: Op1, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2722 return BinaryOperator::CreateOr(V1: A, V2: B);
2723 }
2724
2725 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2726 {
2727 Value *A, *B;
2728 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2729 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))) &&
2730 (Op0->hasOneUse() || Op1->hasOneUse()))
2731 return BinaryOperator::CreateNeg(Op: Builder.CreateXor(LHS: A, RHS: B));
2732 }
2733
2734 // (sub (or A, B), (xor A, B)) --> (and A, B)
2735 {
2736 Value *A, *B;
2737 if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2738 match(V: Op0, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2739 return BinaryOperator::CreateAnd(V1: A, V2: B);
2740 }
2741
2742 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2743 {
2744 Value *A, *B;
2745 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2746 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))) &&
2747 (Op0->hasOneUse() || Op1->hasOneUse()))
2748 return BinaryOperator::CreateNeg(Op: Builder.CreateAnd(LHS: A, RHS: B));
2749 }
2750
2751 {
2752 Value *Y;
2753 // ((X | Y) - X) --> (~X & Y)
2754 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Value(V&: Y), R: m_Specific(V: Op1)))))
2755 return BinaryOperator::CreateAnd(
2756 V1: Y, V2: Builder.CreateNot(V: Op1, Name: Op1->getName() + ".not"));
2757 }
2758
2759 {
2760 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2761 Value *X;
2762 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Specific(V: Op1),
2763 R: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))))))) {
2764 return BinaryOperator::CreateNeg(Op: Builder.CreateAnd(
2765 LHS: Op1, RHS: Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: I.getType()))));
2766 }
2767 }
2768
2769 {
2770 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2771 Constant *C;
2772 if (match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Specific(V: Op1), R: m_Constant(C))))) {
2773 return BinaryOperator::CreateNeg(
2774 Op: Builder.CreateAnd(LHS: Op1, RHS: Builder.CreateNot(V: C)));
2775 }
2776 }
2777
2778 {
2779 // (sub (xor X, (sext C)), (sext C)) => (select C, (neg X), X)
2780 // (sub (sext C), (xor X, (sext C))) => (select C, X, (neg X))
2781 Value *C, *X;
2782 auto m_SubXorCmp = [&C, &X](Value *LHS, Value *RHS) {
2783 return match(V: LHS, P: m_OneUse(SubPattern: m_c_Xor(L: m_Value(V&: X), R: m_Specific(V: RHS)))) &&
2784 match(V: RHS, P: m_SExt(Op: m_Value(V&: C))) &&
2785 (C->getType()->getScalarSizeInBits() == 1);
2786 };
2787 if (m_SubXorCmp(Op0, Op1))
2788 return createSelectInstWithUnknownProfile(C, S1: Builder.CreateNeg(V: X), S2: X);
2789 if (m_SubXorCmp(Op1, Op0))
2790 return createSelectInstWithUnknownProfile(C, S1: X, S2: Builder.CreateNeg(V: X));
2791 }
2792
2793 if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &I))
2794 return R;
2795
2796 if (Instruction *R = foldSubOfMinMax(I, Builder))
2797 return R;
2798
2799 {
2800 // If we have a subtraction between some value and a select between
2801 // said value and something else, sink subtraction into select hands, i.e.:
2802 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2803 // ->
2804 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2805 // or
2806 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2807 // ->
2808 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2809 // This will result in select between new subtraction and 0.
2810 auto SinkSubIntoSelect =
2811 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2812 auto SubBuilder) -> Instruction * {
2813 Value *Cond, *TrueVal, *FalseVal;
2814 if (!match(V: Select, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TrueVal),
2815 R: m_Value(V&: FalseVal)))))
2816 return nullptr;
2817 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2818 return nullptr;
2819 // While it is really tempting to just create two subtractions and let
2820 // InstCombine fold one of those to 0, it isn't possible to do so
2821 // because of worklist visitation order. So ugly it is.
2822 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2823 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2824 Constant *Zero = Constant::getNullValue(Ty);
2825 SelectInst *NewSel =
2826 SelectInst::Create(C: Cond, S1: OtherHandOfSubIsTrueVal ? Zero : NewSub,
2827 S2: OtherHandOfSubIsTrueVal ? NewSub : Zero);
2828 // Preserve prof metadata if any.
2829 NewSel->copyMetadata(SrcInst: cast<Instruction>(Val&: *Select));
2830 return NewSel;
2831 };
2832 if (Instruction *NewSel = SinkSubIntoSelect(
2833 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2834 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2835 return Builder->CreateSub(LHS: OtherHandOfSelect,
2836 /*OtherHandOfSub=*/RHS: Op1);
2837 }))
2838 return NewSel;
2839 if (Instruction *NewSel = SinkSubIntoSelect(
2840 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2841 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2842 return Builder->CreateSub(/*OtherHandOfSub=*/LHS: Op0,
2843 RHS: OtherHandOfSelect);
2844 }))
2845 return NewSel;
2846 }
2847
2848 // (X - (X & Y)) --> (X & ~Y)
2849 if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value(V&: Y))) &&
2850 (Op1->hasOneUse() || isa<Constant>(Val: Y)))
2851 return BinaryOperator::CreateAnd(
2852 V1: Op0, V2: Builder.CreateNot(V: Y, Name: Y->getName() + ".not"));
2853
2854 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2855 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2856 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2857 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2858 // As long as Y is freely invertible, this will be neutral or a win.
2859 // Note: We don't generate the inverse max/min, just create the 'not' of
2860 // it and let other folds do the rest.
2861 if (match(V: Op0, P: m_Not(V: m_Value(V&: X))) &&
2862 match(V: Op1, P: m_c_MaxOrMin(L: m_Specific(V: Op0), R: m_Value(V&: Y))) &&
2863 !Op0->hasNUsesOrMore(N: 3) && isFreeToInvert(V: Y, WillInvertAllUses: Y->hasOneUse())) {
2864 Value *Not = Builder.CreateNot(V: Op1);
2865 return BinaryOperator::CreateSub(V1: Not, V2: X);
2866 }
2867 if (match(V: Op1, P: m_Not(V: m_Value(V&: X))) &&
2868 match(V: Op0, P: m_c_MaxOrMin(L: m_Specific(V: Op1), R: m_Value(V&: Y))) &&
2869 !Op1->hasNUsesOrMore(N: 3) && isFreeToInvert(V: Y, WillInvertAllUses: Y->hasOneUse())) {
2870 Value *Not = Builder.CreateNot(V: Op0);
2871 return BinaryOperator::CreateSub(V1: X, V2: Not);
2872 }
2873
2874 // min(X+1, Y) - min(X, Y) --> zext X < Y
2875 // Replacing a sub and at least one min with an icmp
2876 // and a zext is a potential improvement.
2877 if (match(V: Op0, P: m_c_SMin(L: m_NSWAddLike(L: m_Value(V&: X), R: m_One()), R: m_Value(V&: Y))) &&
2878 match(V: Op1, P: m_c_SMin(L: m_Specific(V: X), R: m_Specific(V: Y))) &&
2879 I.getType()->getScalarSizeInBits() != 1 &&
2880 (Op0->hasOneUse() || Op1->hasOneUse())) {
2881 Value *Cond = Builder.CreateICmpSLT(LHS: X, RHS: Y);
2882 return new ZExtInst(Cond, I.getType());
2883 }
2884 if (match(V: Op0, P: m_c_UMin(L: m_NUWAddLike(L: m_Value(V&: X), R: m_One()), R: m_Value(V&: Y))) &&
2885 match(V: Op1, P: m_c_UMin(L: m_Specific(V: X), R: m_Specific(V: Y))) &&
2886 I.getType()->getScalarSizeInBits() != 1 &&
2887 (Op0->hasOneUse() || Op1->hasOneUse())) {
2888 Value *Cond = Builder.CreateICmpULT(LHS: X, RHS: Y);
2889 return new ZExtInst(Cond, I.getType());
2890 }
2891
2892 // Optimize pointer differences into the same array into a size. Consider:
2893 // &A[10] - &A[0]: we should compile this to "10".
2894 Value *LHSOp, *RHSOp;
2895 if (match(V: Op0, P: m_PtrToIntOrAddr(Op: m_Value(V&: LHSOp))) &&
2896 match(V: Op1, P: m_PtrToIntOrAddr(Op: m_Value(V&: RHSOp))))
2897 if (Value *Res = OptimizePointerDifference(LHS: LHSOp, RHS: RHSOp, Ty: I.getType(),
2898 IsNUW: I.hasNoUnsignedWrap()))
2899 return replaceInstUsesWith(I, V: Res);
2900
2901 // trunc(p)-trunc(q) -> trunc(p-q)
2902 if (match(V: Op0, P: m_Trunc(Op: m_PtrToIntOrAddr(Op: m_Value(V&: LHSOp)))) &&
2903 match(V: Op1, P: m_Trunc(Op: m_PtrToIntOrAddr(Op: m_Value(V&: RHSOp)))))
2904 if (Value *Res = OptimizePointerDifference(LHS: LHSOp, RHS: RHSOp, Ty: I.getType(),
2905 /* IsNUW */ false))
2906 return replaceInstUsesWith(I, V: Res);
2907
2908 auto MatchSubOfZExtOfPtrToIntOrAddr = [&]() {
2909 if (match(V: Op0, P: m_ZExt(Op: m_PtrToIntSameSize(DL, Op: m_Value(V&: LHSOp)))) &&
2910 match(V: Op1, P: m_ZExt(Op: m_PtrToIntSameSize(DL, Op: m_Value(V&: RHSOp)))))
2911 return true;
2912 if (match(V: Op0, P: m_ZExt(Op: m_PtrToAddr(Op: m_Value(V&: LHSOp)))) &&
2913 match(V: Op1, P: m_ZExt(Op: m_PtrToAddr(Op: m_Value(V&: RHSOp)))))
2914 return true;
2915 // Special case for non-canonical ptrtoint in constant expression,
2916 // where the zext has been folded into the ptrtoint.
2917 if (match(V: Op0, P: m_ZExt(Op: m_PtrToIntSameSize(DL, Op: m_Value(V&: LHSOp)))) &&
2918 match(V: Op1, P: m_PtrToInt(Op: m_Value(V&: RHSOp))))
2919 return true;
2920 return false;
2921 };
2922 if (MatchSubOfZExtOfPtrToIntOrAddr()) {
2923 if (auto *GEP = dyn_cast<GEPOperator>(Val: LHSOp)) {
2924 if (GEP->getPointerOperand() == RHSOp) {
2925 if (GEP->hasNoUnsignedWrap() || GEP->hasNoUnsignedSignedWrap()) {
2926 Value *Offset = EmitGEPOffset(GEP);
2927 Value *Res = GEP->hasNoUnsignedWrap()
2928 ? Builder.CreateZExt(
2929 V: Offset, DestTy: I.getType(), Name: "",
2930 /*IsNonNeg=*/GEP->hasNoUnsignedSignedWrap())
2931 : Builder.CreateSExt(V: Offset, DestTy: I.getType());
2932 return replaceInstUsesWith(I, V: Res);
2933 }
2934 }
2935 }
2936 }
2937
2938 // Canonicalize a shifty way to code absolute value to the common pattern.
2939 // There are 2 potential commuted variants.
2940 // We're relying on the fact that we only do this transform when the shift has
2941 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2942 // instructions).
2943 Value *A;
2944 const APInt *ShAmt;
2945 Type *Ty = I.getType();
2946 unsigned BitWidth = Ty->getScalarSizeInBits();
2947 if (match(V: Op1, P: m_AShr(L: m_Value(V&: A), R: m_APInt(Res&: ShAmt))) &&
2948 Op1->hasNUses(N: 2) && *ShAmt == BitWidth - 1 &&
2949 match(V: Op0, P: m_OneUse(SubPattern: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: Op1))))) {
2950 // B = ashr i32 A, 31 ; smear the sign bit
2951 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2952 // --> (A < 0) ? -A : A
2953 Value *IsNeg = Builder.CreateIsNeg(Arg: A);
2954 // Copy the nsw flags from the sub to the negate.
2955 Value *NegA = I.hasNoUnsignedWrap()
2956 ? Constant::getNullValue(Ty: A->getType())
2957 : Builder.CreateNeg(V: A, Name: "", HasNSW: I.hasNoSignedWrap());
2958 return SelectInst::Create(C: IsNeg, S1: NegA, S2: A);
2959 }
2960
2961 // If we are subtracting a low-bit masked subset of some value from an add
2962 // of that same value with no low bits changed, that is clearing some low bits
2963 // of the sum:
2964 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2965 const APInt *AddC, *AndC;
2966 if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: AddC))) &&
2967 match(V: Op1, P: m_And(L: m_Specific(V: X), R: m_APInt(Res&: AndC)))) {
2968 unsigned Cttz = AddC->countr_zero();
2969 APInt HighMask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - Cttz));
2970 if ((HighMask & *AndC).isZero())
2971 return BinaryOperator::CreateAnd(V1: Op0, V2: ConstantInt::get(Ty, V: ~(*AndC)));
2972 }
2973
2974 if (Instruction *V =
2975 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2976 return V;
2977
2978 // X - usub.sat(X, Y) => umin(X, Y)
2979 if (match(V: Op1, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::usub_sat>(Ops: m_Specific(V: Op0),
2980 Ops: m_Value(V&: Y)))))
2981 return replaceInstUsesWith(
2982 I, V: Builder.CreateIntrinsic(ID: Intrinsic::umin, OverloadTypes: {I.getType()}, Args: {Op0, Y}));
2983
2984 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2985 // TODO: The one-use restriction is not strictly necessary, but it may
2986 // require improving other pattern matching and/or codegen.
2987 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_UMax(L: m_Value(V&: X), R: m_Specific(V: Op1)))))
2988 return replaceInstUsesWith(
2989 I, V: Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, OverloadTypes: {Ty}, Args: {X, Op1}));
2990
2991 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2992 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_UMin(L: m_Value(V&: X), R: m_Specific(V: Op0)))))
2993 return replaceInstUsesWith(
2994 I, V: Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, OverloadTypes: {Ty}, Args: {Op0, X}));
2995
2996 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2997 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_UMax(L: m_Value(V&: X), R: m_Specific(V: Op0))))) {
2998 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, OverloadTypes: {Ty}, Args: {X, Op0});
2999 return BinaryOperator::CreateNeg(Op: USub);
3000 }
3001
3002 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
3003 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_UMin(L: m_Value(V&: X), R: m_Specific(V: Op1))))) {
3004 Value *USub = Builder.CreateIntrinsic(ID: Intrinsic::usub_sat, OverloadTypes: {Ty}, Args: {Op1, X});
3005 return BinaryOperator::CreateNeg(Op: USub);
3006 }
3007
3008 // C - ctpop(X) => ctpop(~X) if C is bitwidth
3009 if (match(V: Op0, P: m_SpecificInt(V: BitWidth)) &&
3010 match(V: Op1, P: m_OneUse(SubPattern: m_Ctpop(Op0: m_Value(V&: X)))))
3011 return replaceInstUsesWith(
3012 I, V: Builder.CreateIntrinsic(ID: Intrinsic::ctpop, OverloadTypes: {I.getType()},
3013 Args: {Builder.CreateNot(V: X)}));
3014
3015 // Reduce multiplies for difference-of-squares by factoring:
3016 // (X * X) - (Y * Y) --> (X + Y) * (X - Y)
3017 if (match(V: Op0, P: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: X), R: m_Deferred(V: X)))) &&
3018 match(V: Op1, P: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: Y), R: m_Deferred(V: Y))))) {
3019 auto *OBO0 = cast<OverflowingBinaryOperator>(Val: Op0);
3020 auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1);
3021 bool PropagateNSW = I.hasNoSignedWrap() && OBO0->hasNoSignedWrap() &&
3022 OBO1->hasNoSignedWrap() && BitWidth > 2;
3023 bool PropagateNUW = I.hasNoUnsignedWrap() && OBO0->hasNoUnsignedWrap() &&
3024 OBO1->hasNoUnsignedWrap() && BitWidth > 1;
3025 Value *Add = Builder.CreateAdd(LHS: X, RHS: Y, Name: "add", HasNUW: PropagateNUW, HasNSW: PropagateNSW);
3026 Value *Sub = Builder.CreateSub(LHS: X, RHS: Y, Name: "sub", HasNUW: PropagateNUW, HasNSW: PropagateNSW);
3027 Value *Mul = Builder.CreateMul(LHS: Add, RHS: Sub, Name: "", HasNUW: PropagateNUW, HasNSW: PropagateNSW);
3028 return replaceInstUsesWith(I, V: Mul);
3029 }
3030
3031 // max(X,Y) nsw/nuw - min(X,Y) --> abs(X nsw - Y)
3032 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_SMax(L: m_Value(V&: X), R: m_Value(V&: Y)))) &&
3033 match(V: Op1, P: m_OneUse(SubPattern: m_c_SMin(L: m_Specific(V: X), R: m_Specific(V: Y))))) {
3034 if (I.hasNoUnsignedWrap() || I.hasNoSignedWrap()) {
3035 Value *Sub =
3036 Builder.CreateSub(LHS: X, RHS: Y, Name: "sub", /*HasNUW=*/false, /*HasNSW=*/true);
3037 Value *Call =
3038 Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: Sub, RHS: Builder.getTrue());
3039 return replaceInstUsesWith(I, V: Call);
3040 }
3041 }
3042
3043 if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
3044 return Res;
3045
3046 // (sub (sext (add nsw (X, Y)), sext (X))) --> (sext (Y))
3047 if (match(V: Op1, P: m_SExtLike(Op: m_Value(V&: X))) &&
3048 match(V: Op0, P: m_SExtLike(Op: m_c_NSWAdd(L: m_Specific(V: X), R: m_Value(V&: Y))))) {
3049 Value *SExtY = Builder.CreateSExt(V: Y, DestTy: I.getType());
3050 return replaceInstUsesWith(I, V: SExtY);
3051 }
3052
3053 // (sub[ nsw] (sext (add nsw (X, Y)), sext (add nsw (X, Z)))) -->
3054 // --> (sub[ nsw] (sext (Y), sext (Z)))
3055 {
3056 Value *Z, *Add0, *Add1;
3057 if (match(V: Op0, P: m_SExtLike(Op: m_Value(V&: Add0))) &&
3058 match(V: Op1, P: m_SExtLike(Op: m_Value(V&: Add1))) &&
3059 ((match(V: Add0, P: m_NSWAdd(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
3060 match(V: Add1, P: m_c_NSWAdd(L: m_Specific(V: X), R: m_Value(V&: Z)))) ||
3061 (match(V: Add0, P: m_NSWAdd(L: m_Value(V&: Y), R: m_Value(V&: X))) &&
3062 match(V: Add1, P: m_c_NSWAdd(L: m_Specific(V: X), R: m_Value(V&: Z)))))) {
3063 unsigned NumOfNewInstrs = 0;
3064 // Non-constant Y, Z require new SExt.
3065 NumOfNewInstrs += !isa<Constant>(Val: Y) ? 1 : 0;
3066 NumOfNewInstrs += !isa<Constant>(Val: Z) ? 1 : 0;
3067 // Check if we can trade some of the old instructions for the new ones.
3068 unsigned NumOfDeadInstrs = 0;
3069 if (Op0->hasOneUse()) {
3070 // If Op0 (sext) has multiple uses, then we keep it
3071 // and the add that it uses, otherwise, we can remove
3072 // the sext and probably the add (depending on the number of its uses).
3073 ++NumOfDeadInstrs;
3074 NumOfDeadInstrs += Add0->hasOneUse() ? 1 : 0;
3075 }
3076 if (Op1->hasOneUse()) {
3077 ++NumOfDeadInstrs;
3078 NumOfDeadInstrs += Add1->hasOneUse() ? 1 : 0;
3079 }
3080 if (NumOfDeadInstrs >= NumOfNewInstrs) {
3081 Value *SExtY = Builder.CreateSExt(V: Y, DestTy: I.getType());
3082 Value *SExtZ = Builder.CreateSExt(V: Z, DestTy: I.getType());
3083 Value *Sub = Builder.CreateSub(LHS: SExtY, RHS: SExtZ, Name: "",
3084 /*HasNUW=*/false,
3085 /*HasNSW=*/I.hasNoSignedWrap());
3086 return replaceInstUsesWith(I, V: Sub);
3087 }
3088 }
3089 }
3090
3091 return TryToNarrowDeduceFlags();
3092}
3093
3094/// This eliminates floating-point negation in either 'fneg(X)' or
3095/// 'fsub(-0.0, X)' form by combining into a constant operand.
3096static Instruction *foldFNegIntoConstant(Instruction &I, const DataLayout &DL) {
3097 // This is limited with one-use because fneg is assumed better for
3098 // reassociation and cheaper in codegen than fmul/fdiv.
3099 // TODO: Should the m_OneUse restriction be removed?
3100 Instruction *FNegOp;
3101 if (!match(V: &I, P: m_FNeg(X: m_OneUse(SubPattern: m_Instruction(I&: FNegOp)))))
3102 return nullptr;
3103
3104 Value *X;
3105 Constant *C;
3106
3107 // Fold negation into constant operand.
3108 // -(X * C) --> X * (-C)
3109 if (match(V: FNegOp, P: m_FMul(L: m_Value(V&: X), R: m_Constant(C))))
3110 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) {
3111 FastMathFlags FNegF = I.getFastMathFlags();
3112 FastMathFlags OpF = FNegOp->getFastMathFlags();
3113 FastMathFlags FMF = FastMathFlags::unionValue(LHS: FNegF, RHS: OpF) |
3114 FastMathFlags::intersectRewrite(LHS: FNegF, RHS: OpF);
3115 FMF.setNoInfs(FNegF.noInfs() && OpF.noInfs());
3116 return BinaryOperator::CreateFMulFMF(V1: X, V2: NegC, FMF);
3117 }
3118 // -(X / C) --> X / (-C)
3119 if (match(V: FNegOp, P: m_FDiv(L: m_Value(V&: X), R: m_Constant(C)))) {
3120 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) {
3121 Instruction *FDiv = BinaryOperator::CreateFDivFMF(V1: X, V2: NegC, FMFSource: &I);
3122
3123 // Intersect 'nsz' and 'ninf' because those special value exceptions may
3124 // not apply to the fdiv. Everything else propagates from the fneg.
3125 FastMathFlags FMF = I.getFastMathFlags();
3126 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
3127 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
3128 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
3129 FDiv->copyMetadata(SrcInst: *FNegOp);
3130 return FDiv;
3131 }
3132 }
3133 // -(C / X) --> (-C) / X
3134 if (match(V: FNegOp, P: m_FDiv(L: m_Constant(C), R: m_Value(V&: X))))
3135 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) {
3136 Instruction *FDiv = BinaryOperator::CreateFDivFMF(V1: NegC, V2: X, FMFSource: &I);
3137
3138 // Intersect 'nsz' and 'ninf' because those special value exceptions may
3139 // not apply to the fdiv. Everything else propagates from the fneg.
3140 // TODO: We could propagate nsz/ninf from fdiv alone?
3141 FastMathFlags FMF = I.getFastMathFlags();
3142 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
3143 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
3144 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
3145 FDiv->copyMetadata(SrcInst: *FNegOp);
3146 return FDiv;
3147 }
3148 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
3149 // -(X + C) --> -X + -C --> -C - X
3150 if (I.hasNoSignedZeros() && match(V: FNegOp, P: m_FAdd(L: m_Value(V&: X), R: m_Constant(C))))
3151 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL))
3152 return BinaryOperator::CreateFSubFMF(V1: NegC, V2: X, FMFSource: &I);
3153
3154 return nullptr;
3155}
3156
3157Instruction *InstCombinerImpl::hoistFNegAboveFMulFDiv(Value *FNegOp,
3158 Instruction &FMFSource) {
3159 Value *X, *Y;
3160 if (match(V: FNegOp, P: m_FMul(L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3161 // Push into RHS which is more likely to simplify (const or another fneg).
3162 // FIXME: It would be better to invert the transform.
3163 return cast<Instruction>(Val: Builder.CreateFMulFMF(
3164 L: X, R: Builder.CreateFNegFMF(V: Y, FMFSource: &FMFSource), FMFSource: &FMFSource));
3165 }
3166
3167 if (match(V: FNegOp, P: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3168 auto *FDiv = cast<Instruction>(Val: Builder.CreateFDivFMF(
3169 L: Builder.CreateFNegFMF(V: X, FMFSource: &FMFSource), R: Y, FMFSource: &FMFSource));
3170 FDiv->copyMetadata(SrcInst: *cast<Instruction>(Val: FNegOp));
3171 return FDiv;
3172 }
3173
3174 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: FNegOp)) {
3175 // Make sure to preserve flags and metadata on the call.
3176 if (II->getIntrinsicID() == Intrinsic::ldexp) {
3177 FastMathFlags FMF = FMFSource.getFastMathFlags() | II->getFastMathFlags();
3178 CallInst *New =
3179 Builder.CreateCall(Callee: II->getCalledFunction(),
3180 Args: {Builder.CreateFNegFMF(V: II->getArgOperand(i: 0), FMFSource: FMF),
3181 II->getArgOperand(i: 1)});
3182 New->setFastMathFlags(FMF);
3183 New->copyMetadata(SrcInst: *II);
3184 return New;
3185 }
3186 }
3187
3188 return nullptr;
3189}
3190
3191Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
3192 Value *Op = I.getOperand(i_nocapture: 0);
3193
3194 if (Value *V = simplifyFNegInst(Op, FMF: I.getFastMathFlags(),
3195 Q: getSimplifyQuery().getWithInstruction(I: &I)))
3196 return replaceInstUsesWith(I, V);
3197
3198 if (Instruction *X = foldFNegIntoConstant(I, DL))
3199 return X;
3200
3201 Value *X, *Y;
3202
3203 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
3204 if (I.hasNoSignedZeros() &&
3205 match(V: Op, P: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y)))))
3206 return BinaryOperator::CreateFSubFMF(V1: Y, V2: X, FMFSource: &I);
3207
3208 Value *OneUse;
3209 if (!match(V: Op, P: m_OneUse(SubPattern: m_Value(V&: OneUse))))
3210 return nullptr;
3211
3212 if (Instruction *R = hoistFNegAboveFMulFDiv(FNegOp: OneUse, FMFSource&: I))
3213 return replaceInstUsesWith(I, V: R);
3214
3215 // Try to eliminate fneg if at least 1 arm of the select is negated.
3216 Value *Cond;
3217 if (match(V: OneUse, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3218 // Unlike most transforms, this one is not safe to propagate nsz unless
3219 // it is present on the original select. We union the flags from the select
3220 // and fneg and then remove nsz if needed.
3221 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
3222 S->copyFastMathFlags(I: &I);
3223 if (auto *OldSel = dyn_cast<SelectInst>(Val: Op)) {
3224 FastMathFlags FMF = I.getFastMathFlags() | OldSel->getFastMathFlags();
3225 S->setFastMathFlags(FMF);
3226 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
3227 !isGuaranteedNotToBeUndefOrPoison(V: OldSel->getCondition()))
3228 S->setHasNoSignedZeros(false);
3229 }
3230 };
3231 // -(Cond ? -P : Y) --> Cond ? P : -Y
3232 Value *P;
3233 if (match(V: X, P: m_FNeg(X: m_Value(V&: P)))) {
3234 Value *NegY = Builder.CreateFNegFMF(V: Y, FMFSource: &I, Name: Y->getName() + ".neg");
3235 SelectInst *NewSel = SelectInst::Create(C: Cond, S1: P, S2: NegY);
3236 propagateSelectFMF(NewSel, P == Y);
3237 return NewSel;
3238 }
3239 // -(Cond ? X : -P) --> Cond ? -X : P
3240 if (match(V: Y, P: m_FNeg(X: m_Value(V&: P)))) {
3241 Value *NegX = Builder.CreateFNegFMF(V: X, FMFSource: &I, Name: X->getName() + ".neg");
3242 SelectInst *NewSel = SelectInst::Create(C: Cond, S1: NegX, S2: P);
3243 propagateSelectFMF(NewSel, P == X);
3244 return NewSel;
3245 }
3246
3247 // -(Cond ? X : C) --> Cond ? -X : -C
3248 // -(Cond ? C : Y) --> Cond ? -C : -Y
3249 if (match(V: X, P: m_ImmConstant()) || match(V: Y, P: m_ImmConstant())) {
3250 Value *NegX = Builder.CreateFNegFMF(V: X, FMFSource: &I, Name: X->getName() + ".neg");
3251 Value *NegY = Builder.CreateFNegFMF(V: Y, FMFSource: &I, Name: Y->getName() + ".neg");
3252 SelectInst *NewSel = SelectInst::Create(C: Cond, S1: NegX, S2: NegY);
3253 propagateSelectFMF(NewSel, /*CommonOperand=*/true);
3254 return NewSel;
3255 }
3256 }
3257
3258 // fneg (copysign x, y) -> copysign x, (fneg y)
3259 if (match(V: OneUse, P: m_CopySign(Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) {
3260 // The source copysign has an additional value input, so we can't propagate
3261 // flags the copysign doesn't also have.
3262 FastMathFlags FMF = I.getFastMathFlags();
3263 FMF &= cast<FPMathOperator>(Val: OneUse)->getFastMathFlags();
3264 Value *NegY = Builder.CreateFNegFMF(V: Y, FMFSource: FMF);
3265 Value *NewCopySign = Builder.CreateCopySign(LHS: X, RHS: NegY, FMFSource: FMF);
3266 return replaceInstUsesWith(I, V: NewCopySign);
3267 }
3268
3269 // fneg (shuffle x, Mask) --> shuffle (fneg x), Mask
3270 ArrayRef<int> Mask;
3271 if (match(V: OneUse, P: m_Shuffle(v1: m_Value(V&: X), v2: m_Poison(), mask: m_Mask(Mask))))
3272 return new ShuffleVectorInst(Builder.CreateFNegFMF(V: X, FMFSource: &I), Mask);
3273
3274 // fneg (reverse x) --> reverse (fneg x)
3275 if (match(V: OneUse, P: m_VecReverse(Op0: m_Value(V&: X)))) {
3276 Value *Reverse = Builder.CreateVectorReverse(V: Builder.CreateFNegFMF(V: X, FMFSource: &I));
3277 return replaceInstUsesWith(I, V: Reverse);
3278 }
3279
3280 return nullptr;
3281}
3282
3283Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
3284 if (Value *V = simplifyFSubInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
3285 FMF: I.getFastMathFlags(),
3286 Q: getSimplifyQuery().getWithInstruction(I: &I)))
3287 return replaceInstUsesWith(I, V);
3288
3289 if (Instruction *X = foldVectorBinop(Inst&: I))
3290 return X;
3291
3292 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
3293 return Phi;
3294
3295 // Subtraction from -0.0 is the canonical form of fneg.
3296 // fsub -0.0, X ==> fneg X
3297 // fsub nsz 0.0, X ==> fneg nsz X
3298 //
3299 // FIXME This matcher does not respect FTZ or DAZ yet:
3300 // fsub -0.0, Denorm ==> +-0
3301 // fneg Denorm ==> -Denorm
3302 Value *Op;
3303 if (match(V: &I, P: m_FNeg(X: m_Value(V&: Op))))
3304 return UnaryOperator::CreateFNegFMF(Op, FMFSource: &I);
3305
3306 if (Instruction *X = foldFNegIntoConstant(I, DL))
3307 return X;
3308
3309 if (Instruction *R = foldFBinOpOfIntCasts(I))
3310 return R;
3311
3312 Value *X, *Y;
3313 Constant *C;
3314
3315 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
3316 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
3317 // Canonicalize to fadd to make analysis easier.
3318 // This can also help codegen because fadd is commutative.
3319 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
3320 // killed later. We still limit that particular transform with 'hasOneUse'
3321 // because an fneg is assumed better/cheaper than a generic fsub.
3322 if (I.hasNoSignedZeros() ||
3323 cannotBeNegativeZero(V: Op0, SQ: getSimplifyQuery().getWithInstruction(I: &I))) {
3324 if (match(V: Op1, P: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
3325 Value *NewSub = Builder.CreateFSubFMF(L: Y, R: X, FMFSource: &I);
3326 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: NewSub, FMFSource: &I);
3327 }
3328 }
3329
3330 // (-X) - Op1 --> -(X + Op1)
3331 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Val: Op0) &&
3332 match(V: Op0, P: m_OneUse(SubPattern: m_FNeg(X: m_Value(V&: X))))) {
3333 Value *FAdd = Builder.CreateFAddFMF(L: X, R: Op1, FMFSource: &I);
3334 return UnaryOperator::CreateFNegFMF(Op: FAdd, FMFSource: &I);
3335 }
3336
3337 if (isa<Constant>(Val: Op0))
3338 if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1))
3339 if (Instruction *NV = FoldOpIntoSelect(Op&: I, SI))
3340 return NV;
3341
3342 // X - C --> X + (-C)
3343 // But don't transform constant expressions because there's an inverse fold
3344 // for X + (-Y) --> X - Y.
3345 if (match(V: Op1, P: m_ImmConstant(C)))
3346 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL))
3347 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: NegC, FMFSource: &I);
3348
3349 // X - (-Y) --> X + Y
3350 if (match(V: Op1, P: m_FNeg(X: m_Value(V&: Y))))
3351 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: Y, FMFSource: &I);
3352
3353 // Similar to above, but look through a cast of the negated value:
3354 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
3355 Type *Ty = I.getType();
3356 if (match(V: Op1, P: m_OneUse(SubPattern: m_FPTrunc(Op: m_FNeg(X: m_Value(V&: Y))))))
3357 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: Builder.CreateFPTrunc(V: Y, DestTy: Ty), FMFSource: &I);
3358
3359 // X - (fpext(-Y)) --> X + fpext(Y)
3360 if (match(V: Op1, P: m_OneUse(SubPattern: m_FPExt(Op: m_FNeg(X: m_Value(V&: Y))))))
3361 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: Builder.CreateFPExt(V: Y, DestTy: Ty), FMFSource: &I);
3362
3363 // Similar to above, but look through fmul/fdiv of the negated value:
3364 // Op0 - (-X * Y) --> Op0 + (X * Y)
3365 // Op0 - (Y * -X) --> Op0 + (X * Y)
3366 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_FMul(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y))))) {
3367 Value *FMul = Builder.CreateFMulFMF(L: X, R: Y, FMFSource: &I);
3368 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: FMul, FMFSource: &I);
3369 }
3370 // Op0 - (-X / Y) --> Op0 + (X / Y)
3371 // Op0 - (X / -Y) --> Op0 + (X / Y)
3372 if (match(V: Op1, P: m_OneUse(SubPattern: m_FDiv(L: m_FNeg(X: m_Value(V&: X)), R: m_Value(V&: Y)))) ||
3373 match(V: Op1, P: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_FNeg(X: m_Value(V&: Y)))))) {
3374 Value *FDiv = Builder.CreateFDivFMF(L: X, R: Y, FMFSource: &I);
3375 return BinaryOperator::CreateFAddFMF(V1: Op0, V2: FDiv, FMFSource: &I);
3376 }
3377
3378 // Handle special cases for FSub with selects feeding the operation
3379 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS: Op0, RHS: Op1))
3380 return replaceInstUsesWith(I, V);
3381
3382 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
3383 // (Y - X) - Y --> -X
3384 if (match(V: Op0, P: m_FSub(L: m_Specific(V: Op1), R: m_Value(V&: X))))
3385 return UnaryOperator::CreateFNegFMF(Op: X, FMFSource: &I);
3386
3387 // Y - (X + Y) --> -X
3388 // Y - (Y + X) --> -X
3389 if (match(V: Op1, P: m_c_FAdd(L: m_Specific(V: Op0), R: m_Value(V&: X))))
3390 return UnaryOperator::CreateFNegFMF(Op: X, FMFSource: &I);
3391
3392 // (X * C) - X --> X * (C - 1.0)
3393 if (match(V: Op0, P: m_FMul(L: m_Specific(V: Op1), R: m_Constant(C)))) {
3394 if (Constant *CSubOne = ConstantFoldBinaryOpOperands(
3395 Opcode: Instruction::FSub, LHS: C, RHS: ConstantFP::get(Ty, V: 1.0), DL))
3396 return BinaryOperator::CreateFMulFMF(V1: Op1, V2: CSubOne, FMFSource: &I);
3397 }
3398 // X - (X * C) --> X * (1.0 - C)
3399 if (match(V: Op1, P: m_FMul(L: m_Specific(V: Op0), R: m_Constant(C)))) {
3400 if (Constant *OneSubC = ConstantFoldBinaryOpOperands(
3401 Opcode: Instruction::FSub, LHS: ConstantFP::get(Ty, V: 1.0), RHS: C, DL))
3402 return BinaryOperator::CreateFMulFMF(V1: Op0, V2: OneSubC, FMFSource: &I);
3403 }
3404
3405 // Reassociate fsub/fadd sequences to create more fadd instructions and
3406 // reduce dependency chains:
3407 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
3408 Value *Z;
3409 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))),
3410 R: m_Value(V&: Z))))) {
3411 Value *XZ = Builder.CreateFAddFMF(L: X, R: Z, FMFSource: &I);
3412 Value *YW = Builder.CreateFAddFMF(L: Y, R: Op1, FMFSource: &I);
3413 return BinaryOperator::CreateFSubFMF(V1: XZ, V2: YW, FMFSource: &I);
3414 }
3415
3416 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
3417 return m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_fadd>(Ops: m_Value(V&: Sum),
3418 Ops: m_Value(V&: Vec)));
3419 };
3420 Value *A0, *A1, *V0, *V1;
3421 if (match(V: Op0, P: m_FaddRdx(A0, V0)) && match(V: Op1, P: m_FaddRdx(A1, V1)) &&
3422 V0->getType() == V1->getType()) {
3423 // Difference of sums is sum of differences:
3424 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
3425 Value *Sub = Builder.CreateFSubFMF(L: V0, R: V1, FMFSource: &I);
3426 Value *Rdx = Builder.CreateIntrinsic(ID: Intrinsic::vector_reduce_fadd,
3427 OverloadTypes: {Sub->getType()}, Args: {A0, Sub}, FMFSource: &I);
3428 return BinaryOperator::CreateFSubFMF(V1: Rdx, V2: A1, FMFSource: &I);
3429 }
3430
3431 if (Instruction *F = factorizeFAddFSub(I, Builder))
3432 return F;
3433
3434 // TODO: This performs reassociative folds for FP ops. Some fraction of the
3435 // functionality has been subsumed by simple pattern matching here and in
3436 // InstSimplify. We should let a dedicated reassociation pass handle more
3437 // complex pattern matching and remove this from InstCombine.
3438 if (Value *V = FAddCombine(Builder).simplify(I: &I))
3439 return replaceInstUsesWith(I, V);
3440
3441 // (X - Y) - Op1 --> X - (Y + Op1)
3442 if (match(V: Op0, P: m_OneUse(SubPattern: m_FSub(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
3443 Value *FAdd = Builder.CreateFAddFMF(L: Y, R: Op1, FMFSource: &I);
3444 return BinaryOperator::CreateFSubFMF(V1: X, V2: FAdd, FMFSource: &I);
3445 }
3446 }
3447
3448 return nullptr;
3449}
3450