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
36 using namespace llvm;
37 using namespace PatternMatch;
38
39 #define DEBUG_TYPE "instcombine"
40
41 namespace {
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
set(short C)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
isZero() const74 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75 Value *getValue(Type *) const;
76
isOne() const77 bool isOne() const { return isInt() && IntVal == 1; }
isTwo() const78 bool isTwo() const { return isInt() && IntVal == 2; }
isMinusOne() const79 bool isMinusOne() const { return isInt() && IntVal == -1; }
isMinusTwo() const80 bool isMinusTwo() const { return isInt() && IntVal == -2; }
81
82 private:
insaneIntVal(int V)83 bool insaneIntVal(int V) { return V > 4 || V < -4; }
84
getFpValPtr()85 APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86
getFpValPtr() const87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
89 }
90
getFpVal() const91 const APFloat &getFpVal() const {
92 assert(IsFp && BufHasFpVal && "Incorret state");
93 return *getFpValPtr();
94 }
95
getFpVal()96 APFloat &getFpVal() {
97 assert(IsFp && BufHasFpVal && "Incorret state");
98 return *getFpValPtr();
99 }
100
isInt() const101 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
operator +=(const FAddend & T)133 void operator+=(const FAddend &T) {
134 assert((Val == T.Val) && "Symbolic-values disagree");
135 Coeff += T.Coeff;
136 }
137
getSymVal() const138 Value *getSymVal() const { return Val; }
getCoef() const139 const FAddendCoef &getCoef() const { return Coeff; }
140
isConstant() const141 bool isConstant() const { return Val == nullptr; }
isZero() const142 bool isZero() const { return Coeff.isZero(); }
143
set(short Coefficient,Value * V)144 void set(short Coefficient, Value *V) {
145 Coeff.set(Coefficient);
146 Val = V;
147 }
set(const APFloat & Coefficient,Value * V)148 void set(const APFloat &Coefficient, Value *V) {
149 Coeff.set(Coefficient);
150 Val = V;
151 }
set(const ConstantFP * Coefficient,Value * V)152 void set(const ConstantFP *Coefficient, Value *V) {
153 Coeff.set(Coefficient->getValueAPF());
154 Val = V;
155 }
156
negate()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:
Scale(const FAddendCoef & ScaleAmt)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:
FAddCombine(InstCombiner::BuilderTy & B)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;
initCreateInstNum()205 void initCreateInstNum() { CreateInstrNum = 0; }
incCreateInstNum()206 void incCreateInstNum() { CreateInstrNum++; }
207 #else
initCreateInstNum()208 void initCreateInstNum() {}
incCreateInstNum()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 //===----------------------------------------------------------------------===//
~FAddendCoef()224 FAddendCoef::~FAddendCoef() {
225 if (BufHasFpVal)
226 getFpValPtr()->~APFloat();
227 }
228
set(const APFloat & C)229 void 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
convertToFpType(const fltSemantics & Sem)242 void 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
createAPFloatFromInt(const fltSemantics & Sem,int Val)256 APFloat 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
operator =(const FAddendCoef & That)266 void FAddendCoef::operator=(const FAddendCoef &That) {
267 if (That.isInt())
268 set(That.IntVal);
269 else
270 set(That.getFpVal());
271 }
272
operator +=(const FAddendCoef & That)273 void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
276 if (isInt())
277 IntVal += That.IntVal;
278 else
279 getFpVal().add(That.getFpVal(), RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292 }
293
operator *=(const FAddendCoef & That)294 void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value");
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322 }
323
negate()324 void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329 }
330
getValue(Type * Ty) const331 Value *FAddendCoef::getValue(Type *Ty) const {
332 return isInt() ?
333 ConstantFP::get(Ty, float(IntVal)) :
334 ConstantFP::get(Ty->getContext(), getFpVal());
335 }
336
337 // The definition of <Val> Addends
338 // =========================================
339 // A + B <1, A>, <1,B>
340 // A - B <1, A>, <1,B>
341 // 0 - B <-1, B>
342 // C * A, <C, A>
343 // A + C <1, A> <C, NULL>
344 // 0 +/- 0 <0, NULL> (corner case)
345 //
346 // Legend: A and B are not constant, C is constant
drillValueDownOneStep(Value * Val,FAddend & Addend0,FAddend & Addend1)347 unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
351 return 0;
352
353 unsigned Opcode = I->getOpcode();
354
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356 ConstantFP *C0, *C1;
357 Value *Opnd0 = I->getOperand(0);
358 Value *Opnd1 = I->getOperand(1);
359 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(1, Opnd0);
368 else
369 Addend0.set(C0, nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(1, Opnd1);
376 else
377 Addend.set(C1, nullptr);
378 if (Opcode == Instruction::FSub)
379 Addend.negate();
380 }
381
382 if (Opnd0 || Opnd1)
383 return Opnd0 && Opnd1 ? 2 : 1;
384
385 // Both operands are zero. Weird!
386 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(0);
392 Value *V1 = I->getOperand(1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394 Addend0.set(C, V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399 Addend0.set(C, V0);
400 return 1;
401 }
402 }
403
404 return 0;
405 }
406
407 // Try to break *this* addend into two addends. e.g. Suppose this addend is
408 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409 // i.e. <2.3, X> and <2.3, Y>.
drillAddendDownOneStep(FAddend & Addend0,FAddend & Addend1) const410 unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
412 if (isConstant())
413 return 0;
414
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
417 return BreakNum;
418
419 Addend0.Scale(Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(Coeff);
423
424 return BreakNum;
425 }
426
simplify(Instruction * I)427 Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&
429 "Expected 'reassoc'+'nsz' instruction");
430
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
433 return nullptr;
434
435 assert((I->getOpcode() == Instruction::FAdd ||
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub");
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
494
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
506
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
508 return R;
509 }
510
511 return nullptr;
512 }
513
simplifyFAdd(AddendVect & Addends,unsigned InstrQuota)514 Value *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 // Points to the constant addend of the resulting simplified expression.
523 // If the resulting expr has constant-addend, this constant-addend is
524 // desirable to reside at the top of the resulting expression tree. Placing
525 // constant close to supper-expr(s) will potentially reveal some optimization
526 // opportunities in super-expr(s).
527 const FAddend *ConstAdd = nullptr;
528
529 // Simplified addends are placed <SimpVect>.
530 AddendVect SimpVect;
531
532 // The outer loop works on one symbolic-value at a time. Suppose the input
533 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
534 // The symbolic-values will be processed in this order: x, y, z.
535 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
536
537 const FAddend *ThisAddend = Addends[SymIdx];
538 if (!ThisAddend) {
539 // This addend was processed before.
540 continue;
541 }
542
543 Value *Val = ThisAddend->getSymVal();
544 unsigned StartIdx = SimpVect.size();
545 SimpVect.push_back(ThisAddend);
546
547 // The inner loop collects addends sharing same symbolic-value, and these
548 // addends will be later on folded into a single addend. Following above
549 // example, if the symbolic value "y" is being processed, the inner loop
550 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
551 // be later on folded into "<b1+b2, y>".
552 for (unsigned SameSymIdx = SymIdx + 1;
553 SameSymIdx < AddendNum; SameSymIdx++) {
554 const FAddend *T = Addends[SameSymIdx];
555 if (T && T->getSymVal() == Val) {
556 // Set null such that next iteration of the outer loop will not process
557 // this addend again.
558 Addends[SameSymIdx] = nullptr;
559 SimpVect.push_back(T);
560 }
561 }
562
563 // If multiple addends share same symbolic value, fold them together.
564 if (StartIdx + 1 != SimpVect.size()) {
565 FAddend &R = TmpResult[NextTmpIdx ++];
566 R = *SimpVect[StartIdx];
567 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
568 R += *SimpVect[Idx];
569
570 // Pop all addends being folded and push the resulting folded addend.
571 SimpVect.resize(StartIdx);
572 if (Val) {
573 if (!R.isZero()) {
574 SimpVect.push_back(&R);
575 }
576 } else {
577 // Don't push constant addend at this time. It will be the last element
578 // of <SimpVect>.
579 ConstAdd = &R;
580 }
581 }
582 }
583
584 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
585 "out-of-bound access");
586
587 if (ConstAdd)
588 SimpVect.push_back(ConstAdd);
589
590 Value *Result;
591 if (!SimpVect.empty())
592 Result = createNaryFAdd(SimpVect, InstrQuota);
593 else {
594 // The addition is folded to 0.0.
595 Result = ConstantFP::get(Instr->getType(), 0.0);
596 }
597
598 return Result;
599 }
600
createNaryFAdd(const AddendVect & Opnds,unsigned InstrQuota)601 Value *FAddCombine::createNaryFAdd
602 (const AddendVect &Opnds, unsigned InstrQuota) {
603 assert(!Opnds.empty() && "Expect at least one addend");
604
605 // Step 1: Check if the # of instructions needed exceeds the quota.
606
607 unsigned InstrNeeded = calcInstrNumber(Opnds);
608 if (InstrNeeded > InstrQuota)
609 return nullptr;
610
611 initCreateInstNum();
612
613 // step 2: Emit the N-ary addition.
614 // Note that at most three instructions are involved in Fadd-InstCombine: the
615 // addition in question, and at most two neighboring instructions.
616 // The resulting optimized addition should have at least one less instruction
617 // than the original addition expression tree. This implies that the resulting
618 // N-ary addition has at most two instructions, and we don't need to worry
619 // about tree-height when constructing the N-ary addition.
620
621 Value *LastVal = nullptr;
622 bool LastValNeedNeg = false;
623
624 // Iterate the addends, creating fadd/fsub using adjacent two addends.
625 for (const FAddend *Opnd : Opnds) {
626 bool NeedNeg;
627 Value *V = createAddendVal(*Opnd, NeedNeg);
628 if (!LastVal) {
629 LastVal = V;
630 LastValNeedNeg = NeedNeg;
631 continue;
632 }
633
634 if (LastValNeedNeg == NeedNeg) {
635 LastVal = createFAdd(LastVal, V);
636 continue;
637 }
638
639 if (LastValNeedNeg)
640 LastVal = createFSub(V, LastVal);
641 else
642 LastVal = createFSub(LastVal, V);
643
644 LastValNeedNeg = false;
645 }
646
647 if (LastValNeedNeg) {
648 LastVal = createFNeg(LastVal);
649 }
650
651 #ifndef NDEBUG
652 assert(CreateInstrNum == InstrNeeded &&
653 "Inconsistent in instruction numbers");
654 #endif
655
656 return LastVal;
657 }
658
createFSub(Value * Opnd0,Value * Opnd1)659 Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
660 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
661 if (Instruction *I = dyn_cast<Instruction>(V))
662 createInstPostProc(I);
663 return V;
664 }
665
createFNeg(Value * V)666 Value *FAddCombine::createFNeg(Value *V) {
667 Value *NewV = Builder.CreateFNeg(V);
668 if (Instruction *I = dyn_cast<Instruction>(NewV))
669 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
670 return NewV;
671 }
672
createFAdd(Value * Opnd0,Value * Opnd1)673 Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
674 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
675 if (Instruction *I = dyn_cast<Instruction>(V))
676 createInstPostProc(I);
677 return V;
678 }
679
createFMul(Value * Opnd0,Value * Opnd1)680 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
681 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
682 if (Instruction *I = dyn_cast<Instruction>(V))
683 createInstPostProc(I);
684 return V;
685 }
686
createInstPostProc(Instruction * NewInstr,bool NoNumber)687 void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
688 NewInstr->setDebugLoc(Instr->getDebugLoc());
689
690 // Keep track of the number of instruction created.
691 if (!NoNumber)
692 incCreateInstNum();
693
694 // Propagate fast-math flags
695 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
696 }
697
698 // Return the number of instruction needed to emit the N-ary addition.
699 // NOTE: Keep this function in sync with createAddendVal().
calcInstrNumber(const AddendVect & Opnds)700 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
701 unsigned OpndNum = Opnds.size();
702 unsigned InstrNeeded = OpndNum - 1;
703
704 // The number of addends in the form of "(-1)*x".
705 unsigned NegOpndNum = 0;
706
707 // Adjust the number of instructions needed to emit the N-ary add.
708 for (const FAddend *Opnd : Opnds) {
709 if (Opnd->isConstant())
710 continue;
711
712 // The constant check above is really for a few special constant
713 // coefficients.
714 if (isa<UndefValue>(Opnd->getSymVal()))
715 continue;
716
717 const FAddendCoef &CE = Opnd->getCoef();
718 if (CE.isMinusOne() || CE.isMinusTwo())
719 NegOpndNum++;
720
721 // Let the addend be "c * x". If "c == +/-1", the value of the addend
722 // is immediately available; otherwise, it needs exactly one instruction
723 // to evaluate the value.
724 if (!CE.isMinusOne() && !CE.isOne())
725 InstrNeeded++;
726 }
727 return InstrNeeded;
728 }
729
730 // Input Addend Value NeedNeg(output)
731 // ================================================================
732 // Constant C C false
733 // <+/-1, V> V coefficient is -1
734 // <2/-2, V> "fadd V, V" coefficient is -2
735 // <C, V> "fmul V, C" false
736 //
737 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
createAddendVal(const FAddend & Opnd,bool & NeedNeg)738 Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
739 const FAddendCoef &Coeff = Opnd.getCoef();
740
741 if (Opnd.isConstant()) {
742 NeedNeg = false;
743 return Coeff.getValue(Instr->getType());
744 }
745
746 Value *OpndVal = Opnd.getSymVal();
747
748 if (Coeff.isMinusOne() || Coeff.isOne()) {
749 NeedNeg = Coeff.isMinusOne();
750 return OpndVal;
751 }
752
753 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
754 NeedNeg = Coeff.isMinusTwo();
755 return createFAdd(OpndVal, OpndVal);
756 }
757
758 NeedNeg = false;
759 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
760 }
761
762 // Checks if any operand is negative and we can convert add to sub.
763 // This function checks for following negative patterns
764 // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
765 // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
766 // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
checkForNegativeOperand(BinaryOperator & I,InstCombiner::BuilderTy & Builder)767 static Value *checkForNegativeOperand(BinaryOperator &I,
768 InstCombiner::BuilderTy &Builder) {
769 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
770
771 // This function creates 2 instructions to replace ADD, we need at least one
772 // of LHS or RHS to have one use to ensure benefit in transform.
773 if (!LHS->hasOneUse() && !RHS->hasOneUse())
774 return nullptr;
775
776 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
777 const APInt *C1 = nullptr, *C2 = nullptr;
778
779 // if ONE is on other side, swap
780 if (match(RHS, m_Add(m_Value(X), m_One())))
781 std::swap(LHS, RHS);
782
783 if (match(LHS, m_Add(m_Value(X), m_One()))) {
784 // if XOR on other side, swap
785 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
786 std::swap(X, RHS);
787
788 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
789 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
790 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
791 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
792 Value *NewAnd = Builder.CreateAnd(Z, *C1);
793 return Builder.CreateSub(RHS, NewAnd, "sub");
794 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
795 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
796 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
797 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
798 return Builder.CreateSub(RHS, NewOr, "sub");
799 }
800 }
801 }
802
803 // Restore LHS and RHS
804 LHS = I.getOperand(0);
805 RHS = I.getOperand(1);
806
807 // if XOR is on other side, swap
808 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
809 std::swap(LHS, RHS);
810
811 // C2 is ODD
812 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
813 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
814 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
815 if (C1->countTrailingZeros() == 0)
816 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
817 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
818 return Builder.CreateSub(RHS, NewOr, "sub");
819 }
820 return nullptr;
821 }
822
823 /// Wrapping flags may allow combining constants separated by an extend.
foldNoWrapAdd(BinaryOperator & Add,InstCombiner::BuilderTy & Builder)824 static Instruction *foldNoWrapAdd(BinaryOperator &Add,
825 InstCombiner::BuilderTy &Builder) {
826 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
827 Type *Ty = Add.getType();
828 Constant *Op1C;
829 if (!match(Op1, m_Constant(Op1C)))
830 return nullptr;
831
832 // Try this match first because it results in an add in the narrow type.
833 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
834 Value *X;
835 const APInt *C1, *C2;
836 if (match(Op1, m_APInt(C1)) &&
837 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
838 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
839 Constant *NewC =
840 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
841 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
842 }
843
844 // More general combining of constants in the wide type.
845 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
846 Constant *NarrowC;
847 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
848 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
849 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
850 Value *WideX = Builder.CreateSExt(X, Ty);
851 return BinaryOperator::CreateAdd(WideX, NewC);
852 }
853 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
854 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
855 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
856 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
857 Value *WideX = Builder.CreateZExt(X, Ty);
858 return BinaryOperator::CreateAdd(WideX, NewC);
859 }
860
861 return nullptr;
862 }
863
foldAddWithConstant(BinaryOperator & Add)864 Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
865 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
866 Constant *Op1C;
867 if (!match(Op1, m_ImmConstant(Op1C)))
868 return nullptr;
869
870 if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
871 return NV;
872
873 Value *X;
874 Constant *Op00C;
875
876 // add (sub C1, X), C2 --> sub (add C1, C2), X
877 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
878 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
879
880 Value *Y;
881
882 // add (sub X, Y), -1 --> add (not Y), X
883 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
884 match(Op1, m_AllOnes()))
885 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
886
887 // zext(bool) + C -> bool ? C + 1 : C
888 if (match(Op0, m_ZExt(m_Value(X))) &&
889 X->getType()->getScalarSizeInBits() == 1)
890 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
891 // sext(bool) + C -> bool ? C - 1 : C
892 if (match(Op0, m_SExt(m_Value(X))) &&
893 X->getType()->getScalarSizeInBits() == 1)
894 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
895
896 // ~X + C --> (C-1) - X
897 if (match(Op0, m_Not(m_Value(X))))
898 return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X);
899
900 const APInt *C;
901 if (!match(Op1, m_APInt(C)))
902 return nullptr;
903
904 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
905 Constant *Op01C;
906 if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
907 haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
908 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
909
910 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
911 const APInt *C2;
912 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
913 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
914
915 if (C->isSignMask()) {
916 // If wrapping is not allowed, then the addition must set the sign bit:
917 // X + (signmask) --> X | signmask
918 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
919 return BinaryOperator::CreateOr(Op0, Op1);
920
921 // If wrapping is allowed, then the addition flips the sign bit of LHS:
922 // X + (signmask) --> X ^ signmask
923 return BinaryOperator::CreateXor(Op0, Op1);
924 }
925
926 // Is this add the last step in a convoluted sext?
927 // add(zext(xor i16 X, -32768), -32768) --> sext X
928 Type *Ty = Add.getType();
929 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
930 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
931 return CastInst::Create(Instruction::SExt, X, Ty);
932
933 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
934 // (X ^ signmask) + C --> (X + (signmask ^ C))
935 if (C2->isSignMask())
936 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
937
938 // If X has no high-bits set above an xor mask:
939 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
940 if (C2->isMask()) {
941 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
942 if ((*C2 | LHSKnown.Zero).isAllOnesValue())
943 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
944 }
945
946 // Look for a math+logic pattern that corresponds to sext-in-register of a
947 // value with cleared high bits. Convert that into a pair of shifts:
948 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
949 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
950 if (Op0->hasOneUse() && *C2 == -(*C)) {
951 unsigned BitWidth = Ty->getScalarSizeInBits();
952 unsigned ShAmt = 0;
953 if (C->isPowerOf2())
954 ShAmt = BitWidth - C->logBase2() - 1;
955 else if (C2->isPowerOf2())
956 ShAmt = BitWidth - C2->logBase2() - 1;
957 if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
958 0, &Add)) {
959 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
960 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
961 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
962 }
963 }
964 }
965
966 if (C->isOneValue() && Op0->hasOneUse()) {
967 // add (sext i1 X), 1 --> zext (not X)
968 // TODO: The smallest IR representation is (select X, 0, 1), and that would
969 // not require the one-use check. But we need to remove a transform in
970 // visitSelect and make sure that IR value tracking for select is equal or
971 // better than for these ops.
972 if (match(Op0, m_SExt(m_Value(X))) &&
973 X->getType()->getScalarSizeInBits() == 1)
974 return new ZExtInst(Builder.CreateNot(X), Ty);
975
976 // Shifts and add used to flip and mask off the low bit:
977 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
978 const APInt *C3;
979 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
980 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
981 Value *NotX = Builder.CreateNot(X);
982 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
983 }
984 }
985
986 // If all bits affected by the add are included in a high-bit-mask, do the
987 // add before the mask op:
988 // (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00
989 if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) &&
990 C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) {
991 Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C));
992 return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2));
993 }
994
995 return nullptr;
996 }
997
998 // Matches multiplication expression Op * C where C is a constant. Returns the
999 // constant value in C and the other operand in Op. Returns true if such a
1000 // match is found.
MatchMul(Value * E,Value * & Op,APInt & C)1001 static bool MatchMul(Value *E, Value *&Op, APInt &C) {
1002 const APInt *AI;
1003 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
1004 C = *AI;
1005 return true;
1006 }
1007 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1008 C = APInt(AI->getBitWidth(), 1);
1009 C <<= *AI;
1010 return true;
1011 }
1012 return false;
1013 }
1014
1015 // Matches remainder expression Op % C where C is a constant. Returns the
1016 // constant value in C and the other operand in Op. Returns the signedness of
1017 // the remainder operation in IsSigned. Returns true if such a match is
1018 // found.
MatchRem(Value * E,Value * & Op,APInt & C,bool & IsSigned)1019 static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1020 const APInt *AI;
1021 IsSigned = false;
1022 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1023 IsSigned = true;
1024 C = *AI;
1025 return true;
1026 }
1027 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1028 C = *AI;
1029 return true;
1030 }
1031 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1032 C = *AI + 1;
1033 return true;
1034 }
1035 return false;
1036 }
1037
1038 // Matches division expression Op / C with the given signedness as indicated
1039 // by IsSigned, where C is a constant. Returns the constant value in C and the
1040 // other operand in Op. Returns true if such a match is found.
MatchDiv(Value * E,Value * & Op,APInt & C,bool IsSigned)1041 static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1042 const APInt *AI;
1043 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1044 C = *AI;
1045 return true;
1046 }
1047 if (!IsSigned) {
1048 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1049 C = *AI;
1050 return true;
1051 }
1052 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1053 C = APInt(AI->getBitWidth(), 1);
1054 C <<= *AI;
1055 return true;
1056 }
1057 }
1058 return false;
1059 }
1060
1061 // Returns whether C0 * C1 with the given signedness overflows.
MulWillOverflow(APInt & C0,APInt & C1,bool IsSigned)1062 static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1063 bool overflow;
1064 if (IsSigned)
1065 (void)C0.smul_ov(C1, overflow);
1066 else
1067 (void)C0.umul_ov(C1, overflow);
1068 return overflow;
1069 }
1070
1071 // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1072 // does not overflow.
SimplifyAddWithRemainder(BinaryOperator & I)1073 Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1074 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1075 Value *X, *MulOpV;
1076 APInt C0, MulOpC;
1077 bool IsSigned;
1078 // Match I = X % C0 + MulOpV * C0
1079 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1080 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1081 C0 == MulOpC) {
1082 Value *RemOpV;
1083 APInt C1;
1084 bool Rem2IsSigned;
1085 // Match MulOpC = RemOpV % C1
1086 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1087 IsSigned == Rem2IsSigned) {
1088 Value *DivOpV;
1089 APInt DivOpC;
1090 // Match RemOpV = X / C0
1091 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1092 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1093 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1094 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1095 : Builder.CreateURem(X, NewDivisor, "urem");
1096 }
1097 }
1098 }
1099
1100 return nullptr;
1101 }
1102
1103 /// Fold
1104 /// (1 << NBits) - 1
1105 /// Into:
1106 /// ~(-(1 << NBits))
1107 /// Because a 'not' is better for bit-tracking analysis and other transforms
1108 /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
canonicalizeLowbitMask(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1109 static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1110 InstCombiner::BuilderTy &Builder) {
1111 Value *NBits;
1112 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1113 return nullptr;
1114
1115 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1116 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1117 // Be wary of constant folding.
1118 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1119 // Always NSW. But NUW propagates from `add`.
1120 BOp->setHasNoSignedWrap();
1121 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1122 }
1123
1124 return BinaryOperator::CreateNot(NotMask, I.getName());
1125 }
1126
foldToUnsignedSaturatedAdd(BinaryOperator & I)1127 static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1128 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction");
1129 Type *Ty = I.getType();
1130 auto getUAddSat = [&]() {
1131 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1132 };
1133
1134 // add (umin X, ~Y), Y --> uaddsat X, Y
1135 Value *X, *Y;
1136 if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1137 m_Deferred(Y))))
1138 return CallInst::Create(getUAddSat(), { X, Y });
1139
1140 // add (umin X, ~C), C --> uaddsat X, C
1141 const APInt *C, *NotC;
1142 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1143 *C == ~*NotC)
1144 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1145
1146 return nullptr;
1147 }
1148
1149 Instruction *InstCombinerImpl::
canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(BinaryOperator & I)1150 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1151 BinaryOperator &I) {
1152 assert((I.getOpcode() == Instruction::Add ||
1153 I.getOpcode() == Instruction::Or ||
1154 I.getOpcode() == Instruction::Sub) &&
1155 "Expecting add/or/sub instruction");
1156
1157 // We have a subtraction/addition between a (potentially truncated) *logical*
1158 // right-shift of X and a "select".
1159 Value *X, *Select;
1160 Instruction *LowBitsToSkip, *Extract;
1161 if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd(
1162 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1163 m_Instruction(Extract))),
1164 m_Value(Select))))
1165 return nullptr;
1166
1167 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1168 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
1169 return nullptr;
1170
1171 Type *XTy = X->getType();
1172 bool HadTrunc = I.getType() != XTy;
1173
1174 // If there was a truncation of extracted value, then we'll need to produce
1175 // one extra instruction, so we need to ensure one instruction will go away.
1176 if (HadTrunc && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1177 return nullptr;
1178
1179 // Extraction should extract high NBits bits, with shift amount calculated as:
1180 // low bits to skip = shift bitwidth - high bits to extract
1181 // The shift amount itself may be extended, and we need to look past zero-ext
1182 // when matching NBits, that will matter for matching later.
1183 Constant *C;
1184 Value *NBits;
1185 if (!match(
1186 LowBitsToSkip,
1187 m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1188 !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1189 APInt(C->getType()->getScalarSizeInBits(),
1190 X->getType()->getScalarSizeInBits()))))
1191 return nullptr;
1192
1193 // Sign-extending value can be zero-extended if we `sub`tract it,
1194 // or sign-extended otherwise.
1195 auto SkipExtInMagic = [&I](Value *&V) {
1196 if (I.getOpcode() == Instruction::Sub)
1197 match(V, m_ZExtOrSelf(m_Value(V)));
1198 else
1199 match(V, m_SExtOrSelf(m_Value(V)));
1200 };
1201
1202 // Now, finally validate the sign-extending magic.
1203 // `select` itself may be appropriately extended, look past that.
1204 SkipExtInMagic(Select);
1205
1206 ICmpInst::Predicate Pred;
1207 const APInt *Thr;
1208 Value *SignExtendingValue, *Zero;
1209 bool ShouldSignext;
1210 // It must be a select between two values we will later establish to be a
1211 // sign-extending value and a zero constant. The condition guarding the
1212 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1213 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
1214 m_Value(SignExtendingValue), m_Value(Zero))) ||
1215 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1216 return nullptr;
1217
1218 // icmp-select pair is commutative.
1219 if (!ShouldSignext)
1220 std::swap(SignExtendingValue, Zero);
1221
1222 // If we should not perform sign-extension then we must add/or/subtract zero.
1223 if (!match(Zero, m_Zero()))
1224 return nullptr;
1225 // Otherwise, it should be some constant, left-shifted by the same NBits we
1226 // had in `lshr`. Said left-shift can also be appropriately extended.
1227 // Again, we must look past zero-ext when looking for NBits.
1228 SkipExtInMagic(SignExtendingValue);
1229 Constant *SignExtendingValueBaseConstant;
1230 if (!match(SignExtendingValue,
1231 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1232 m_ZExtOrSelf(m_Specific(NBits)))))
1233 return nullptr;
1234 // If we `sub`, then the constant should be one, else it should be all-ones.
1235 if (I.getOpcode() == Instruction::Sub
1236 ? !match(SignExtendingValueBaseConstant, m_One())
1237 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1238 return nullptr;
1239
1240 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1241 Extract->getName() + ".sext");
1242 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1243 if (!HadTrunc)
1244 return NewAShr;
1245
1246 Builder.Insert(NewAShr);
1247 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1248 }
1249
1250 /// This is a specialization of a more general transform from
1251 /// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
1252 /// for multi-use cases or propagating nsw/nuw, then we would not need this.
factorizeMathWithShlOps(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1253 static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1254 InstCombiner::BuilderTy &Builder) {
1255 // TODO: Also handle mul by doubling the shift amount?
1256 assert((I.getOpcode() == Instruction::Add ||
1257 I.getOpcode() == Instruction::Sub) &&
1258 "Expected add/sub");
1259 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1260 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1261 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1262 return nullptr;
1263
1264 Value *X, *Y, *ShAmt;
1265 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1266 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1267 return nullptr;
1268
1269 // No-wrap propagates only when all ops have no-wrap.
1270 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1271 Op1->hasNoSignedWrap();
1272 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1273 Op1->hasNoUnsignedWrap();
1274
1275 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1276 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1277 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1278 NewI->setHasNoSignedWrap(HasNSW);
1279 NewI->setHasNoUnsignedWrap(HasNUW);
1280 }
1281 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1282 NewShl->setHasNoSignedWrap(HasNSW);
1283 NewShl->setHasNoUnsignedWrap(HasNUW);
1284 return NewShl;
1285 }
1286
visitAdd(BinaryOperator & I)1287 Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1288 if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1289 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1290 SQ.getWithInstruction(&I)))
1291 return replaceInstUsesWith(I, V);
1292
1293 if (SimplifyAssociativeOrCommutative(I))
1294 return &I;
1295
1296 if (Instruction *X = foldVectorBinop(I))
1297 return X;
1298
1299 // (A*B)+(A*C) -> A*(B+C) etc
1300 if (Value *V = SimplifyUsingDistributiveLaws(I))
1301 return replaceInstUsesWith(I, V);
1302
1303 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1304 return R;
1305
1306 if (Instruction *X = foldAddWithConstant(I))
1307 return X;
1308
1309 if (Instruction *X = foldNoWrapAdd(I, Builder))
1310 return X;
1311
1312 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1313 Type *Ty = I.getType();
1314 if (Ty->isIntOrIntVectorTy(1))
1315 return BinaryOperator::CreateXor(LHS, RHS);
1316
1317 // X + X --> X << 1
1318 if (LHS == RHS) {
1319 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1320 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1321 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1322 return Shl;
1323 }
1324
1325 Value *A, *B;
1326 if (match(LHS, m_Neg(m_Value(A)))) {
1327 // -A + -B --> -(A + B)
1328 if (match(RHS, m_Neg(m_Value(B))))
1329 return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1330
1331 // -A + B --> B - A
1332 return BinaryOperator::CreateSub(RHS, A);
1333 }
1334
1335 // A + -B --> A - B
1336 if (match(RHS, m_Neg(m_Value(B))))
1337 return BinaryOperator::CreateSub(LHS, B);
1338
1339 if (Value *V = checkForNegativeOperand(I, Builder))
1340 return replaceInstUsesWith(I, V);
1341
1342 // (A + 1) + ~B --> A - B
1343 // ~B + (A + 1) --> A - B
1344 // (~B + A) + 1 --> A - B
1345 // (A + ~B) + 1 --> A - B
1346 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1347 match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1348 return BinaryOperator::CreateSub(A, B);
1349
1350 // (A + RHS) + RHS --> A + (RHS << 1)
1351 if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1352 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1353
1354 // LHS + (A + LHS) --> A + (LHS << 1)
1355 if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1356 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1357
1358 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1359 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1360
1361 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1362 const APInt *C1, *C2;
1363 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1364 APInt one(C2->getBitWidth(), 1);
1365 APInt minusC1 = -(*C1);
1366 if (minusC1 == (one << *C2)) {
1367 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1368 return BinaryOperator::CreateSRem(RHS, NewRHS);
1369 }
1370 }
1371
1372 // A+B --> A|B iff A and B have no bits set in common.
1373 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1374 return BinaryOperator::CreateOr(LHS, RHS);
1375
1376 // add (select X 0 (sub n A)) A --> select X A n
1377 {
1378 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1379 Value *A = RHS;
1380 if (!SI) {
1381 SI = dyn_cast<SelectInst>(RHS);
1382 A = LHS;
1383 }
1384 if (SI && SI->hasOneUse()) {
1385 Value *TV = SI->getTrueValue();
1386 Value *FV = SI->getFalseValue();
1387 Value *N;
1388
1389 // Can we fold the add into the argument of the select?
1390 // We check both true and false select arguments for a matching subtract.
1391 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1392 // Fold the add into the true select value.
1393 return SelectInst::Create(SI->getCondition(), N, A);
1394
1395 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1396 // Fold the add into the false select value.
1397 return SelectInst::Create(SI->getCondition(), A, N);
1398 }
1399 }
1400
1401 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1402 return Ext;
1403
1404 // (add (xor A, B) (and A, B)) --> (or A, B)
1405 // (add (and A, B) (xor A, B)) --> (or A, B)
1406 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1407 m_c_And(m_Deferred(A), m_Deferred(B)))))
1408 return BinaryOperator::CreateOr(A, B);
1409
1410 // (add (or A, B) (and A, B)) --> (add A, B)
1411 // (add (and A, B) (or A, B)) --> (add A, B)
1412 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1413 m_c_And(m_Deferred(A), m_Deferred(B))))) {
1414 // Replacing operands in-place to preserve nuw/nsw flags.
1415 replaceOperand(I, 0, A);
1416 replaceOperand(I, 1, B);
1417 return &I;
1418 }
1419
1420 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1421 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1422 // computeKnownBits.
1423 bool Changed = false;
1424 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1425 Changed = true;
1426 I.setHasNoSignedWrap(true);
1427 }
1428 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1429 Changed = true;
1430 I.setHasNoUnsignedWrap(true);
1431 }
1432
1433 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1434 return V;
1435
1436 if (Instruction *V =
1437 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1438 return V;
1439
1440 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1441 return SatAdd;
1442
1443 // usub.sat(A, B) + B => umax(A, B)
1444 if (match(&I, m_c_BinOp(
1445 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1446 m_Deferred(B)))) {
1447 return replaceInstUsesWith(I,
1448 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1449 }
1450
1451 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1452 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1453 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1454 haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1455 return replaceInstUsesWith(
1456 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1457 {Builder.CreateOr(A, B)}));
1458
1459 return Changed ? &I : nullptr;
1460 }
1461
1462 /// Eliminate an op from a linear interpolation (lerp) pattern.
factorizeLerp(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1463 static Instruction *factorizeLerp(BinaryOperator &I,
1464 InstCombiner::BuilderTy &Builder) {
1465 Value *X, *Y, *Z;
1466 if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1467 m_OneUse(m_FSub(m_FPOne(),
1468 m_Value(Z))))),
1469 m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1470 return nullptr;
1471
1472 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1473 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1474 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1475 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1476 }
1477
1478 /// Factor a common operand out of fadd/fsub of fmul/fdiv.
factorizeFAddFSub(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1479 static Instruction *factorizeFAddFSub(BinaryOperator &I,
1480 InstCombiner::BuilderTy &Builder) {
1481 assert((I.getOpcode() == Instruction::FAdd ||
1482 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub");
1483 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&
1484 "FP factorization requires FMF");
1485
1486 if (Instruction *Lerp = factorizeLerp(I, Builder))
1487 return Lerp;
1488
1489 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1490 Value *X, *Y, *Z;
1491 bool IsFMul;
1492 if ((match(Op0, m_OneUse(m_FMul(m_Value(X), m_Value(Z)))) &&
1493 match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))) ||
1494 (match(Op0, m_OneUse(m_FMul(m_Value(Z), m_Value(X)))) &&
1495 match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))))
1496 IsFMul = true;
1497 else if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Z)))) &&
1498 match(Op1, m_OneUse(m_FDiv(m_Value(Y), m_Specific(Z)))))
1499 IsFMul = false;
1500 else
1501 return nullptr;
1502
1503 // (X * Z) + (Y * Z) --> (X + Y) * Z
1504 // (X * Z) - (Y * Z) --> (X - Y) * Z
1505 // (X / Z) + (Y / Z) --> (X + Y) / Z
1506 // (X / Z) - (Y / Z) --> (X - Y) / Z
1507 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1508 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1509 : Builder.CreateFSubFMF(X, Y, &I);
1510
1511 // Bail out if we just created a denormal constant.
1512 // TODO: This is copied from a previous implementation. Is it necessary?
1513 const APFloat *C;
1514 if (match(XY, m_APFloat(C)) && !C->isNormal())
1515 return nullptr;
1516
1517 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1518 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1519 }
1520
visitFAdd(BinaryOperator & I)1521 Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1522 if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1523 I.getFastMathFlags(),
1524 SQ.getWithInstruction(&I)))
1525 return replaceInstUsesWith(I, V);
1526
1527 if (SimplifyAssociativeOrCommutative(I))
1528 return &I;
1529
1530 if (Instruction *X = foldVectorBinop(I))
1531 return X;
1532
1533 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1534 return FoldedFAdd;
1535
1536 // (-X) + Y --> Y - X
1537 Value *X, *Y;
1538 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1539 return BinaryOperator::CreateFSubFMF(Y, X, &I);
1540
1541 // Similar to above, but look through fmul/fdiv for the negated term.
1542 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1543 Value *Z;
1544 if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1545 m_Value(Z)))) {
1546 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1547 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1548 }
1549 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1550 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1551 if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1552 m_Value(Z))) ||
1553 match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1554 m_Value(Z)))) {
1555 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1556 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1557 }
1558
1559 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1560 // integer add followed by a promotion.
1561 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1562 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1563 Value *LHSIntVal = LHSConv->getOperand(0);
1564 Type *FPType = LHSConv->getType();
1565
1566 // TODO: This check is overly conservative. In many cases known bits
1567 // analysis can tell us that the result of the addition has less significant
1568 // bits than the integer type can hold.
1569 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1570 Type *FScalarTy = FTy->getScalarType();
1571 Type *IScalarTy = ITy->getScalarType();
1572
1573 // Do we have enough bits in the significand to represent the result of
1574 // the integer addition?
1575 unsigned MaxRepresentableBits =
1576 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1577 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1578 };
1579
1580 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1581 // ... if the constant fits in the integer value. This is useful for things
1582 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1583 // requires a constant pool load, and generally allows the add to be better
1584 // instcombined.
1585 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1586 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1587 Constant *CI =
1588 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1589 if (LHSConv->hasOneUse() &&
1590 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1591 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1592 // Insert the new integer add.
1593 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1594 return new SIToFPInst(NewAdd, I.getType());
1595 }
1596 }
1597
1598 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1599 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1600 Value *RHSIntVal = RHSConv->getOperand(0);
1601 // It's enough to check LHS types only because we require int types to
1602 // be the same for this transform.
1603 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1604 // Only do this if x/y have the same type, if at least one of them has a
1605 // single use (so we don't increase the number of int->fp conversions),
1606 // and if the integer add will not overflow.
1607 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1608 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1609 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1610 // Insert the new integer add.
1611 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1612 return new SIToFPInst(NewAdd, I.getType());
1613 }
1614 }
1615 }
1616 }
1617
1618 // Handle specials cases for FAdd with selects feeding the operation
1619 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1620 return replaceInstUsesWith(I, V);
1621
1622 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1623 if (Instruction *F = factorizeFAddFSub(I, Builder))
1624 return F;
1625
1626 // Try to fold fadd into start value of reduction intrinsic.
1627 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1628 m_AnyZeroFP(), m_Value(X))),
1629 m_Value(Y)))) {
1630 // fadd (rdx 0.0, X), Y --> rdx Y, X
1631 return replaceInstUsesWith(
1632 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1633 {X->getType()}, {Y, X}, &I));
1634 }
1635 const APFloat *StartC, *C;
1636 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1637 m_APFloat(StartC), m_Value(X)))) &&
1638 match(RHS, m_APFloat(C))) {
1639 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1640 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1641 return replaceInstUsesWith(
1642 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1643 {X->getType()}, {NewStartC, X}, &I));
1644 }
1645
1646 if (Value *V = FAddCombine(Builder).simplify(&I))
1647 return replaceInstUsesWith(I, V);
1648 }
1649
1650 return nullptr;
1651 }
1652
1653 /// Optimize pointer differences into the same array into a size. Consider:
1654 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1655 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
OptimizePointerDifference(Value * LHS,Value * RHS,Type * Ty,bool IsNUW)1656 Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
1657 Type *Ty, bool IsNUW) {
1658 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1659 // this.
1660 bool Swapped = false;
1661 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1662 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1663 std::swap(LHS, RHS);
1664 Swapped = true;
1665 }
1666
1667 // Require at least one GEP with a common base pointer on both sides.
1668 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1669 // (gep X, ...) - X
1670 if (LHSGEP->getOperand(0) == RHS) {
1671 GEP1 = LHSGEP;
1672 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1673 // (gep X, ...) - (gep X, ...)
1674 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1675 RHSGEP->getOperand(0)->stripPointerCasts()) {
1676 GEP1 = LHSGEP;
1677 GEP2 = RHSGEP;
1678 }
1679 }
1680 }
1681
1682 if (!GEP1)
1683 return nullptr;
1684
1685 if (GEP2) {
1686 // (gep X, ...) - (gep X, ...)
1687 //
1688 // Avoid duplicating the arithmetic if there are more than one non-constant
1689 // indices between the two GEPs and either GEP has a non-constant index and
1690 // multiple users. If zero non-constant index, the result is a constant and
1691 // there is no duplication. If one non-constant index, the result is an add
1692 // or sub with a constant, which is no larger than the original code, and
1693 // there's no duplicated arithmetic, even if either GEP has multiple
1694 // users. If more than one non-constant indices combined, as long as the GEP
1695 // with at least one non-constant index doesn't have multiple users, there
1696 // is no duplication.
1697 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1698 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1699 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1700 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1701 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1702 return nullptr;
1703 }
1704 }
1705
1706 // Emit the offset of the GEP and an intptr_t.
1707 Value *Result = EmitGEPOffset(GEP1);
1708
1709 // If this is a single inbounds GEP and the original sub was nuw,
1710 // then the final multiplication is also nuw.
1711 if (auto *I = dyn_cast<Instruction>(Result))
1712 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1713 I->getOpcode() == Instruction::Mul)
1714 I->setHasNoUnsignedWrap();
1715
1716 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
1717 // If both GEPs are inbounds, then the subtract does not have signed overflow.
1718 if (GEP2) {
1719 Value *Offset = EmitGEPOffset(GEP2);
1720 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
1721 GEP1->isInBounds() && GEP2->isInBounds());
1722 }
1723
1724 // If we have p - gep(p, ...) then we have to negate the result.
1725 if (Swapped)
1726 Result = Builder.CreateNeg(Result, "diff.neg");
1727
1728 return Builder.CreateIntCast(Result, Ty, true);
1729 }
1730
visitSub(BinaryOperator & I)1731 Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
1732 if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1733 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1734 SQ.getWithInstruction(&I)))
1735 return replaceInstUsesWith(I, V);
1736
1737 if (Instruction *X = foldVectorBinop(I))
1738 return X;
1739
1740 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1741
1742 // If this is a 'B = x-(-A)', change to B = x+A.
1743 // We deal with this without involving Negator to preserve NSW flag.
1744 if (Value *V = dyn_castNegVal(Op1)) {
1745 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1746
1747 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1748 assert(BO->getOpcode() == Instruction::Sub &&
1749 "Expected a subtraction operator!");
1750 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1751 Res->setHasNoSignedWrap(true);
1752 } else {
1753 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1754 Res->setHasNoSignedWrap(true);
1755 }
1756
1757 return Res;
1758 }
1759
1760 // Try this before Negator to preserve NSW flag.
1761 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1762 return R;
1763
1764 Constant *C;
1765 if (match(Op0, m_ImmConstant(C))) {
1766 Value *X;
1767 Constant *C2;
1768
1769 // C-(X+C2) --> (C-C2)-X
1770 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
1771 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1772 }
1773
1774 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
1775 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1776 return Ext;
1777
1778 bool Changed = false;
1779 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1780 Changed = true;
1781 I.setHasNoSignedWrap(true);
1782 }
1783 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1784 Changed = true;
1785 I.setHasNoUnsignedWrap(true);
1786 }
1787
1788 return Changed ? &I : nullptr;
1789 };
1790
1791 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
1792 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
1793 // a pure negation used by a select that looks like abs/nabs.
1794 bool IsNegation = match(Op0, m_ZeroInt());
1795 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
1796 const Instruction *UI = dyn_cast<Instruction>(U);
1797 if (!UI)
1798 return false;
1799 return match(UI,
1800 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
1801 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
1802 })) {
1803 if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
1804 return BinaryOperator::CreateAdd(NegOp1, Op0);
1805 }
1806 if (IsNegation)
1807 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
1808
1809 // (A*B)-(A*C) -> A*(B-C) etc
1810 if (Value *V = SimplifyUsingDistributiveLaws(I))
1811 return replaceInstUsesWith(I, V);
1812
1813 if (I.getType()->isIntOrIntVectorTy(1))
1814 return BinaryOperator::CreateXor(Op0, Op1);
1815
1816 // Replace (-1 - A) with (~A).
1817 if (match(Op0, m_AllOnes()))
1818 return BinaryOperator::CreateNot(Op1);
1819
1820 // (~X) - (~Y) --> Y - X
1821 Value *X, *Y;
1822 if (match(Op0, m_Not(m_Value(X))) && match(Op1, m_Not(m_Value(Y))))
1823 return BinaryOperator::CreateSub(Y, X);
1824
1825 // (X + -1) - Y --> ~Y + X
1826 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1827 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1828
1829 // Reassociate sub/add sequences to create more add instructions and
1830 // reduce dependency chains:
1831 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
1832 Value *Z;
1833 if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
1834 m_Value(Z))))) {
1835 Value *XZ = Builder.CreateAdd(X, Z);
1836 Value *YW = Builder.CreateAdd(Y, Op1);
1837 return BinaryOperator::CreateSub(XZ, YW);
1838 }
1839
1840 // ((X - Y) - Op1) --> X - (Y + Op1)
1841 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
1842 Value *Add = Builder.CreateAdd(Y, Op1);
1843 return BinaryOperator::CreateSub(X, Add);
1844 }
1845
1846 auto m_AddRdx = [](Value *&Vec) {
1847 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
1848 };
1849 Value *V0, *V1;
1850 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
1851 V0->getType() == V1->getType()) {
1852 // Difference of sums is sum of differences:
1853 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
1854 Value *Sub = Builder.CreateSub(V0, V1);
1855 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
1856 {Sub->getType()}, {Sub});
1857 return replaceInstUsesWith(I, Rdx);
1858 }
1859
1860 if (Constant *C = dyn_cast<Constant>(Op0)) {
1861 Value *X;
1862 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1863 // C - (zext bool) --> bool ? C - 1 : C
1864 return SelectInst::Create(X, InstCombiner::SubOne(C), C);
1865 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1866 // C - (sext bool) --> bool ? C + 1 : C
1867 return SelectInst::Create(X, InstCombiner::AddOne(C), C);
1868
1869 // C - ~X == X + (1+C)
1870 if (match(Op1, m_Not(m_Value(X))))
1871 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
1872
1873 // Try to fold constant sub into select arguments.
1874 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1875 if (Instruction *R = FoldOpIntoSelect(I, SI))
1876 return R;
1877
1878 // Try to fold constant sub into PHI values.
1879 if (PHINode *PN = dyn_cast<PHINode>(Op1))
1880 if (Instruction *R = foldOpIntoPhi(I, PN))
1881 return R;
1882
1883 Constant *C2;
1884
1885 // C-(C2-X) --> X+(C-C2)
1886 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
1887 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
1888 }
1889
1890 const APInt *Op0C;
1891 if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) {
1892 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1893 // zero.
1894 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1895 if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
1896 return BinaryOperator::CreateXor(Op1, Op0);
1897 }
1898
1899 {
1900 Value *Y;
1901 // X-(X+Y) == -Y X-(Y+X) == -Y
1902 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1903 return BinaryOperator::CreateNeg(Y);
1904
1905 // (X-Y)-X == -Y
1906 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1907 return BinaryOperator::CreateNeg(Y);
1908 }
1909
1910 // (sub (or A, B) (and A, B)) --> (xor A, B)
1911 {
1912 Value *A, *B;
1913 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1914 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1915 return BinaryOperator::CreateXor(A, B);
1916 }
1917
1918 // (sub (add A, B) (or A, B)) --> (and A, B)
1919 {
1920 Value *A, *B;
1921 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1922 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1923 return BinaryOperator::CreateAnd(A, B);
1924 }
1925
1926 // (sub (add A, B) (and A, B)) --> (or A, B)
1927 {
1928 Value *A, *B;
1929 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1930 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
1931 return BinaryOperator::CreateOr(A, B);
1932 }
1933
1934 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
1935 {
1936 Value *A, *B;
1937 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1938 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1939 (Op0->hasOneUse() || Op1->hasOneUse()))
1940 return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
1941 }
1942
1943 // (sub (or A, B), (xor A, B)) --> (and A, B)
1944 {
1945 Value *A, *B;
1946 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1947 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1948 return BinaryOperator::CreateAnd(A, B);
1949 }
1950
1951 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
1952 {
1953 Value *A, *B;
1954 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1955 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1956 (Op0->hasOneUse() || Op1->hasOneUse()))
1957 return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
1958 }
1959
1960 {
1961 Value *Y;
1962 // ((X | Y) - X) --> (~X & Y)
1963 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1964 return BinaryOperator::CreateAnd(
1965 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1966 }
1967
1968 {
1969 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
1970 Value *X;
1971 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
1972 m_OneUse(m_Neg(m_Value(X))))))) {
1973 return BinaryOperator::CreateNeg(Builder.CreateAnd(
1974 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
1975 }
1976 }
1977
1978 {
1979 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
1980 Constant *C;
1981 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
1982 return BinaryOperator::CreateNeg(
1983 Builder.CreateAnd(Op1, Builder.CreateNot(C)));
1984 }
1985 }
1986
1987 {
1988 // If we have a subtraction between some value and a select between
1989 // said value and something else, sink subtraction into select hands, i.e.:
1990 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
1991 // ->
1992 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
1993 // or
1994 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
1995 // ->
1996 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
1997 // This will result in select between new subtraction and 0.
1998 auto SinkSubIntoSelect =
1999 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2000 auto SubBuilder) -> Instruction * {
2001 Value *Cond, *TrueVal, *FalseVal;
2002 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2003 m_Value(FalseVal)))))
2004 return nullptr;
2005 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2006 return nullptr;
2007 // While it is really tempting to just create two subtractions and let
2008 // InstCombine fold one of those to 0, it isn't possible to do so
2009 // because of worklist visitation order. So ugly it is.
2010 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2011 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2012 Constant *Zero = Constant::getNullValue(Ty);
2013 SelectInst *NewSel =
2014 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2015 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2016 // Preserve prof metadata if any.
2017 NewSel->copyMetadata(cast<Instruction>(*Select));
2018 return NewSel;
2019 };
2020 if (Instruction *NewSel = SinkSubIntoSelect(
2021 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2022 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2023 return Builder->CreateSub(OtherHandOfSelect,
2024 /*OtherHandOfSub=*/Op1);
2025 }))
2026 return NewSel;
2027 if (Instruction *NewSel = SinkSubIntoSelect(
2028 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2029 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2030 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2031 OtherHandOfSelect);
2032 }))
2033 return NewSel;
2034 }
2035
2036 // (X - (X & Y)) --> (X & ~Y)
2037 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2038 (Op1->hasOneUse() || isa<Constant>(Y)))
2039 return BinaryOperator::CreateAnd(
2040 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2041
2042 {
2043 // ~A - Min/Max(~A, O) -> Max/Min(A, ~O) - A
2044 // ~A - Min/Max(O, ~A) -> Max/Min(A, ~O) - A
2045 // Min/Max(~A, O) - ~A -> A - Max/Min(A, ~O)
2046 // Min/Max(O, ~A) - ~A -> A - Max/Min(A, ~O)
2047 // So long as O here is freely invertible, this will be neutral or a win.
2048 Value *LHS, *RHS, *A;
2049 Value *NotA = Op0, *MinMax = Op1;
2050 SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2051 if (!SelectPatternResult::isMinOrMax(SPF)) {
2052 NotA = Op1;
2053 MinMax = Op0;
2054 SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2055 }
2056 if (SelectPatternResult::isMinOrMax(SPF) &&
2057 match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
2058 if (NotA == LHS)
2059 std::swap(LHS, RHS);
2060 // LHS is now O above and expected to have at least 2 uses (the min/max)
2061 // NotA is epected to have 2 uses from the min/max and 1 from the sub.
2062 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
2063 !NotA->hasNUsesOrMore(4)) {
2064 // Note: We don't generate the inverse max/min, just create the not of
2065 // it and let other folds do the rest.
2066 Value *Not = Builder.CreateNot(MinMax);
2067 if (NotA == Op0)
2068 return BinaryOperator::CreateSub(Not, A);
2069 else
2070 return BinaryOperator::CreateSub(A, Not);
2071 }
2072 }
2073 }
2074
2075 // Optimize pointer differences into the same array into a size. Consider:
2076 // &A[10] - &A[0]: we should compile this to "10".
2077 Value *LHSOp, *RHSOp;
2078 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2079 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2080 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2081 I.hasNoUnsignedWrap()))
2082 return replaceInstUsesWith(I, Res);
2083
2084 // trunc(p)-trunc(q) -> trunc(p-q)
2085 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2086 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2087 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2088 /* IsNUW */ false))
2089 return replaceInstUsesWith(I, Res);
2090
2091 // Canonicalize a shifty way to code absolute value to the common pattern.
2092 // There are 2 potential commuted variants.
2093 // We're relying on the fact that we only do this transform when the shift has
2094 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2095 // instructions).
2096 Value *A;
2097 const APInt *ShAmt;
2098 Type *Ty = I.getType();
2099 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2100 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2101 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2102 // B = ashr i32 A, 31 ; smear the sign bit
2103 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2104 // --> (A < 0) ? -A : A
2105 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2106 // Copy the nuw/nsw flags from the sub to the negate.
2107 Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2108 I.hasNoSignedWrap());
2109 return SelectInst::Create(Cmp, Neg, A);
2110 }
2111
2112 // If we are subtracting a low-bit masked subset of some value from an add
2113 // of that same value with no low bits changed, that is clearing some low bits
2114 // of the sum:
2115 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2116 const APInt *AddC, *AndC;
2117 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2118 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2119 unsigned BitWidth = Ty->getScalarSizeInBits();
2120 unsigned Cttz = AddC->countTrailingZeros();
2121 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2122 if ((HighMask & *AndC).isNullValue())
2123 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2124 }
2125
2126 if (Instruction *V =
2127 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2128 return V;
2129
2130 // X - usub.sat(X, Y) => umin(X, Y)
2131 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2132 m_Value(Y)))))
2133 return replaceInstUsesWith(
2134 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2135
2136 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2137 if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
2138 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2139 return replaceInstUsesWith(
2140 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2141 {Builder.CreateNot(X)}));
2142
2143 return TryToNarrowDeduceFlags();
2144 }
2145
2146 /// This eliminates floating-point negation in either 'fneg(X)' or
2147 /// 'fsub(-0.0, X)' form by combining into a constant operand.
foldFNegIntoConstant(Instruction & I)2148 static Instruction *foldFNegIntoConstant(Instruction &I) {
2149 // This is limited with one-use because fneg is assumed better for
2150 // reassociation and cheaper in codegen than fmul/fdiv.
2151 // TODO: Should the m_OneUse restriction be removed?
2152 Instruction *FNegOp;
2153 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2154 return nullptr;
2155
2156 Value *X;
2157 Constant *C;
2158
2159 // Fold negation into constant operand.
2160 // -(X * C) --> X * (-C)
2161 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2162 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
2163 // -(X / C) --> X / (-C)
2164 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2165 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
2166 // -(C / X) --> (-C) / X
2167 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) {
2168 Instruction *FDiv =
2169 BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
2170
2171 // Intersect 'nsz' and 'ninf' because those special value exceptions may not
2172 // apply to the fdiv. Everything else propagates from the fneg.
2173 // TODO: We could propagate nsz/ninf from fdiv alone?
2174 FastMathFlags FMF = I.getFastMathFlags();
2175 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2176 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() & OpFMF.noSignedZeros());
2177 FDiv->setHasNoInfs(FMF.noInfs() & OpFMF.noInfs());
2178 return FDiv;
2179 }
2180 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2181 // -(X + C) --> -X + -C --> -C - X
2182 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2183 return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I);
2184
2185 return nullptr;
2186 }
2187
hoistFNegAboveFMulFDiv(Instruction & I,InstCombiner::BuilderTy & Builder)2188 static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2189 InstCombiner::BuilderTy &Builder) {
2190 Value *FNeg;
2191 if (!match(&I, m_FNeg(m_Value(FNeg))))
2192 return nullptr;
2193
2194 Value *X, *Y;
2195 if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2196 return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2197
2198 if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2199 return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2200
2201 return nullptr;
2202 }
2203
visitFNeg(UnaryOperator & I)2204 Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2205 Value *Op = I.getOperand(0);
2206
2207 if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
2208 getSimplifyQuery().getWithInstruction(&I)))
2209 return replaceInstUsesWith(I, V);
2210
2211 if (Instruction *X = foldFNegIntoConstant(I))
2212 return X;
2213
2214 Value *X, *Y;
2215
2216 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2217 if (I.hasNoSignedZeros() &&
2218 match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2219 return BinaryOperator::CreateFSubFMF(Y, X, &I);
2220
2221 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2222 return R;
2223
2224 // Try to eliminate fneg if at least 1 arm of the select is negated.
2225 Value *Cond;
2226 if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
2227 // Unlike most transforms, this one is not safe to propagate nsz unless
2228 // it is present on the original select. (We are conservatively intersecting
2229 // the nsz flags from the select and root fneg instruction.)
2230 auto propagateSelectFMF = [&](SelectInst *S) {
2231 S->copyFastMathFlags(&I);
2232 if (auto *OldSel = dyn_cast<SelectInst>(Op))
2233 if (!OldSel->hasNoSignedZeros())
2234 S->setHasNoSignedZeros(false);
2235 };
2236 // -(Cond ? -P : Y) --> Cond ? P : -Y
2237 Value *P;
2238 if (match(X, m_FNeg(m_Value(P)))) {
2239 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2240 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2241 propagateSelectFMF(NewSel);
2242 return NewSel;
2243 }
2244 // -(Cond ? X : -P) --> Cond ? -X : P
2245 if (match(Y, m_FNeg(m_Value(P)))) {
2246 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2247 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2248 propagateSelectFMF(NewSel);
2249 return NewSel;
2250 }
2251 }
2252
2253 return nullptr;
2254 }
2255
visitFSub(BinaryOperator & I)2256 Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2257 if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
2258 I.getFastMathFlags(),
2259 getSimplifyQuery().getWithInstruction(&I)))
2260 return replaceInstUsesWith(I, V);
2261
2262 if (Instruction *X = foldVectorBinop(I))
2263 return X;
2264
2265 // Subtraction from -0.0 is the canonical form of fneg.
2266 // fsub -0.0, X ==> fneg X
2267 // fsub nsz 0.0, X ==> fneg nsz X
2268 //
2269 // FIXME This matcher does not respect FTZ or DAZ yet:
2270 // fsub -0.0, Denorm ==> +-0
2271 // fneg Denorm ==> -Denorm
2272 Value *Op;
2273 if (match(&I, m_FNeg(m_Value(Op))))
2274 return UnaryOperator::CreateFNegFMF(Op, &I);
2275
2276 if (Instruction *X = foldFNegIntoConstant(I))
2277 return X;
2278
2279 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2280 return R;
2281
2282 Value *X, *Y;
2283 Constant *C;
2284
2285 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2286 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2287 // Canonicalize to fadd to make analysis easier.
2288 // This can also help codegen because fadd is commutative.
2289 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2290 // killed later. We still limit that particular transform with 'hasOneUse'
2291 // because an fneg is assumed better/cheaper than a generic fsub.
2292 if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2293 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2294 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2295 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2296 }
2297 }
2298
2299 // (-X) - Op1 --> -(X + Op1)
2300 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2301 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2302 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2303 return UnaryOperator::CreateFNegFMF(FAdd, &I);
2304 }
2305
2306 if (isa<Constant>(Op0))
2307 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2308 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2309 return NV;
2310
2311 // X - C --> X + (-C)
2312 // But don't transform constant expressions because there's an inverse fold
2313 // for X + (-Y) --> X - Y.
2314 if (match(Op1, m_ImmConstant(C)))
2315 return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
2316
2317 // X - (-Y) --> X + Y
2318 if (match(Op1, m_FNeg(m_Value(Y))))
2319 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2320
2321 // Similar to above, but look through a cast of the negated value:
2322 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2323 Type *Ty = I.getType();
2324 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2325 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2326
2327 // X - (fpext(-Y)) --> X + fpext(Y)
2328 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2329 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2330
2331 // Similar to above, but look through fmul/fdiv of the negated value:
2332 // Op0 - (-X * Y) --> Op0 + (X * Y)
2333 // Op0 - (Y * -X) --> Op0 + (X * Y)
2334 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2335 Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2336 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2337 }
2338 // Op0 - (-X / Y) --> Op0 + (X / Y)
2339 // Op0 - (X / -Y) --> Op0 + (X / Y)
2340 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2341 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2342 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2343 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2344 }
2345
2346 // Handle special cases for FSub with selects feeding the operation
2347 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2348 return replaceInstUsesWith(I, V);
2349
2350 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2351 // (Y - X) - Y --> -X
2352 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2353 return UnaryOperator::CreateFNegFMF(X, &I);
2354
2355 // Y - (X + Y) --> -X
2356 // Y - (Y + X) --> -X
2357 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2358 return UnaryOperator::CreateFNegFMF(X, &I);
2359
2360 // (X * C) - X --> X * (C - 1.0)
2361 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2362 Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
2363 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2364 }
2365 // X - (X * C) --> X * (1.0 - C)
2366 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2367 Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
2368 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2369 }
2370
2371 // Reassociate fsub/fadd sequences to create more fadd instructions and
2372 // reduce dependency chains:
2373 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2374 Value *Z;
2375 if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2376 m_Value(Z))))) {
2377 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2378 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2379 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2380 }
2381
2382 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2383 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2384 m_Value(Vec)));
2385 };
2386 Value *A0, *A1, *V0, *V1;
2387 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2388 V0->getType() == V1->getType()) {
2389 // Difference of sums is sum of differences:
2390 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2391 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2392 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2393 {Sub->getType()}, {A0, Sub}, &I);
2394 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2395 }
2396
2397 if (Instruction *F = factorizeFAddFSub(I, Builder))
2398 return F;
2399
2400 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2401 // functionality has been subsumed by simple pattern matching here and in
2402 // InstSimplify. We should let a dedicated reassociation pass handle more
2403 // complex pattern matching and remove this from InstCombine.
2404 if (Value *V = FAddCombine(Builder).simplify(&I))
2405 return replaceInstUsesWith(I, V);
2406
2407 // (X - Y) - Op1 --> X - (Y + Op1)
2408 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2409 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2410 return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
2411 }
2412 }
2413
2414 return nullptr;
2415 }
2416