Bug Summary

File:build/source/llvm/include/llvm/IR/PatternMatch.h
Warning:line 1450, column 9
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name InstCombineAddSub.cpp -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/source/build-llvm -resource-dir /usr/lib/llvm-16/lib/clang/16.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Transforms/InstCombine -I /build/source/llvm/lib/Transforms/InstCombine -I include -I /build/source/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-16/lib/clang/16.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/source/build-llvm=build-llvm -fmacro-prefix-map=/build/source/= -fcoverage-prefix-map=/build/source/build-llvm=build-llvm -fcoverage-prefix-map=/build/source/= -source-date-epoch 1668078801 -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -Wno-misleading-indentation -std=c++17 -fdeprecated-macro -fdebug-compilation-dir=/build/source/build-llvm -fdebug-prefix-map=/build/source/build-llvm=build-llvm -fdebug-prefix-map=/build/source/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2022-11-10-135928-647445-1 -x c++ /build/source/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

/build/source/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

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" "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")(static_cast <bool> (!insaneIntVal(C) && "Insane coefficient"
) ? void (0) : __assert_fail ("!insaneIntVal(C) && \"Insane coefficient\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 66
, __extension__ __PRETTY_FUNCTION__))
;
67 IsFp = false; IntVal = C;
68 }
69
70 void set(const APFloat& C);
71
72 void negate();
73
74 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75 Value *getValue(Type *) const;
76
77 bool isOne() const { return isInt() && IntVal == 1; }
78 bool isTwo() const { return isInt() && IntVal == 2; }
79 bool isMinusOne() const { return isInt() && IntVal == -1; }
80 bool isMinusTwo() const { return isInt() && IntVal == -2; }
81
82 private:
83 bool insaneIntVal(int V) { return V > 4 || V < -4; }
84
85 APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86
87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
89 }
90
91 const APFloat &getFpVal() const {
92 assert(IsFp && BufHasFpVal && "Incorret state")(static_cast <bool> (IsFp && BufHasFpVal &&
"Incorret state") ? void (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 92
, __extension__ __PRETTY_FUNCTION__))
;
93 return *getFpValPtr();
94 }
95
96 APFloat &getFpVal() {
97 assert(IsFp && BufHasFpVal && "Incorret state")(static_cast <bool> (IsFp && BufHasFpVal &&
"Incorret state") ? void (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 97
, __extension__ __PRETTY_FUNCTION__))
;
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")(static_cast <bool> ((Val == T.Val) && "Symbolic-values disagree"
) ? void (0) : __assert_fail ("(Val == T.Val) && \"Symbolic-values disagree\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 134
, __extension__ __PRETTY_FUNCTION__))
;
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(That.getFpVal(), RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292}
293
294void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value")(static_cast <bool> (!insaneIntVal(Res) && "Insane int value"
) ? void (0) : __assert_fail ("!insaneIntVal(Res) && \"Insane int value\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 305
, __extension__ __PRETTY_FUNCTION__))
;
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322}
323
324void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329}
330
331Value *FAddendCoef::getValue(Type *Ty) const {
332 return isInt() ?
333 ConstantFP::get(Ty, float(IntVal)) :
334 ConstantFP::get(Ty->getContext(), getFpVal());
335}
336
337// The definition of <Val> Addends
338// =========================================
339// A + B <1, A>, <1,B>
340// A - B <1, A>, <1,B>
341// 0 - B <-1, B>
342// C * A, <C, A>
343// A + C <1, A> <C, NULL>
344// 0 +/- 0 <0, NULL> (corner case)
345//
346// Legend: A and B are not constant, C is constant
347unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
351 return 0;
352
353 unsigned Opcode = I->getOpcode();
354
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356 ConstantFP *C0, *C1;
357 Value *Opnd0 = I->getOperand(0);
358 Value *Opnd1 = I->getOperand(1);
359 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(1, Opnd0);
368 else
369 Addend0.set(C0, nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(1, Opnd1);
376 else
377 Addend.set(C1, nullptr);
378 if (Opcode == Instruction::FSub)
379 Addend.negate();
380 }
381
382 if (Opnd0 || Opnd1)
383 return Opnd0 && Opnd1 ? 2 : 1;
384
385 // Both operands are zero. Weird!
386 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(0);
392 Value *V1 = I->getOperand(1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394 Addend0.set(C, V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399 Addend0.set(C, V0);
400 return 1;
401 }
402 }
403
404 return 0;
405}
406
407// Try to break *this* addend into two addends. e.g. Suppose this addend is
408// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409// i.e. <2.3, X> and <2.3, Y>.
410unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
412 if (isConstant())
413 return 0;
414
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
417 return BreakNum;
418
419 Addend0.Scale(Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(Coeff);
423
424 return BreakNum;
425}
426
427Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&(static_cast <bool> (I->hasAllowReassoc() &&
I->hasNoSignedZeros() && "Expected 'reassoc'+'nsz' instruction"
) ? void (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 429
, __extension__ __PRETTY_FUNCTION__))
429 "Expected 'reassoc'+'nsz' instruction")(static_cast <bool> (I->hasAllowReassoc() &&
I->hasNoSignedZeros() && "Expected 'reassoc'+'nsz' instruction"
) ? void (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 429
, __extension__ __PRETTY_FUNCTION__))
;
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 ||(static_cast <bool> ((I->getOpcode() == Instruction::
FAdd || I->getOpcode() == Instruction::FSub) && "Expect add/sub"
) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 436
, __extension__ __PRETTY_FUNCTION__))
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub")(static_cast <bool> ((I->getOpcode() == Instruction::
FAdd || I->getOpcode() == Instruction::FSub) && "Expect add/sub"
) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 436
, __extension__ __PRETTY_FUNCTION__))
;
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
494
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
506
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
508 return R;
509 }
510
511 return nullptr;
512}
513
514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends")(static_cast <bool> (AddendNum <= 4 && "Too many addends"
) ? void (0) : __assert_fail ("AddendNum <= 4 && \"Too many addends\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 516
, __extension__ __PRETTY_FUNCTION__))
;
517
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
521
522 // Simplified addends are placed <SimpVect>.
523 AddendVect SimpVect;
524
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529
530 const FAddend *ThisAddend = Addends[SymIdx];
531 if (!ThisAddend) {
532 // This addend was processed before.
533 continue;
534 }
535
536 Value *Val = ThisAddend->getSymVal();
537
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543 // call later.
544
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(ThisAddend);
547
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(T);
561 }
562 }
563
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569 R += *SimpVect[Idx];
570
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(&R);
575 }
576 }
577 }
578
579 assert((NextTmpIdx <= std::size(TmpResult) + 1) && "out-of-bound access")(static_cast <bool> ((NextTmpIdx <= std::size(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
("(NextTmpIdx <= std::size(TmpResult) + 1) && \"out-of-bound access\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 579
, __extension__ __PRETTY_FUNCTION__))
;
580
581 Value *Result;
582 if (!SimpVect.empty())
583 Result = createNaryFAdd(SimpVect, InstrQuota);
584 else {
585 // The addition is folded to 0.0.
586 Result = ConstantFP::get(Instr->getType(), 0.0);
587 }
588
589 return Result;
590}
591
592Value *FAddCombine::createNaryFAdd
593 (const AddendVect &Opnds, unsigned InstrQuota) {
594 assert(!Opnds.empty() && "Expect at least one addend")(static_cast <bool> (!Opnds.empty() && "Expect at least one addend"
) ? void (0) : __assert_fail ("!Opnds.empty() && \"Expect at least one addend\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 594
, __extension__ __PRETTY_FUNCTION__))
;
595
596 // Step 1: Check if the # of instructions needed exceeds the quota.
597
598 unsigned InstrNeeded = calcInstrNumber(Opnds);
599 if (InstrNeeded > InstrQuota)
600 return nullptr;
601
602 initCreateInstNum();
603
604 // step 2: Emit the N-ary addition.
605 // Note that at most three instructions are involved in Fadd-InstCombine: the
606 // addition in question, and at most two neighboring instructions.
607 // The resulting optimized addition should have at least one less instruction
608 // than the original addition expression tree. This implies that the resulting
609 // N-ary addition has at most two instructions, and we don't need to worry
610 // about tree-height when constructing the N-ary addition.
611
612 Value *LastVal = nullptr;
613 bool LastValNeedNeg = false;
614
615 // Iterate the addends, creating fadd/fsub using adjacent two addends.
616 for (const FAddend *Opnd : Opnds) {
617 bool NeedNeg;
618 Value *V = createAddendVal(*Opnd, NeedNeg);
619 if (!LastVal) {
620 LastVal = V;
621 LastValNeedNeg = NeedNeg;
622 continue;
623 }
624
625 if (LastValNeedNeg == NeedNeg) {
626 LastVal = createFAdd(LastVal, V);
627 continue;
628 }
629
630 if (LastValNeedNeg)
631 LastVal = createFSub(V, LastVal);
632 else
633 LastVal = createFSub(LastVal, V);
634
635 LastValNeedNeg = false;
636 }
637
638 if (LastValNeedNeg) {
639 LastVal = createFNeg(LastVal);
640 }
641
642#ifndef NDEBUG
643 assert(CreateInstrNum == InstrNeeded &&(static_cast <bool> (CreateInstrNum == InstrNeeded &&
"Inconsistent in instruction numbers") ? void (0) : __assert_fail
("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 644
, __extension__ __PRETTY_FUNCTION__))
644 "Inconsistent in instruction numbers")(static_cast <bool> (CreateInstrNum == InstrNeeded &&
"Inconsistent in instruction numbers") ? void (0) : __assert_fail
("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 644
, __extension__ __PRETTY_FUNCTION__))
;
645#endif
646
647 return LastVal;
648}
649
650Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
651 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
652 if (Instruction *I = dyn_cast<Instruction>(V))
653 createInstPostProc(I);
654 return V;
655}
656
657Value *FAddCombine::createFNeg(Value *V) {
658 Value *NewV = Builder.CreateFNeg(V);
659 if (Instruction *I = dyn_cast<Instruction>(NewV))
660 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
661 return NewV;
662}
663
664Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
665 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
666 if (Instruction *I = dyn_cast<Instruction>(V))
667 createInstPostProc(I);
668 return V;
669}
670
671Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
672 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
673 if (Instruction *I = dyn_cast<Instruction>(V))
674 createInstPostProc(I);
675 return V;
676}
677
678void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
679 NewInstr->setDebugLoc(Instr->getDebugLoc());
680
681 // Keep track of the number of instruction created.
682 if (!NoNumber)
683 incCreateInstNum();
684
685 // Propagate fast-math flags
686 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
687}
688
689// Return the number of instruction needed to emit the N-ary addition.
690// NOTE: Keep this function in sync with createAddendVal().
691unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
692 unsigned OpndNum = Opnds.size();
693 unsigned InstrNeeded = OpndNum - 1;
694
695 // Adjust the number of instructions needed to emit the N-ary add.
696 for (const FAddend *Opnd : Opnds) {
697 if (Opnd->isConstant())
698 continue;
699
700 // The constant check above is really for a few special constant
701 // coefficients.
702 if (isa<UndefValue>(Opnd->getSymVal()))
703 continue;
704
705 const FAddendCoef &CE = Opnd->getCoef();
706 // Let the addend be "c * x". If "c == +/-1", the value of the addend
707 // is immediately available; otherwise, it needs exactly one instruction
708 // to evaluate the value.
709 if (!CE.isMinusOne() && !CE.isOne())
710 InstrNeeded++;
711 }
712 return InstrNeeded;
713}
714
715// Input Addend Value NeedNeg(output)
716// ================================================================
717// Constant C C false
718// <+/-1, V> V coefficient is -1
719// <2/-2, V> "fadd V, V" coefficient is -2
720// <C, V> "fmul V, C" false
721//
722// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
723Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
724 const FAddendCoef &Coeff = Opnd.getCoef();
725
726 if (Opnd.isConstant()) {
727 NeedNeg = false;
728 return Coeff.getValue(Instr->getType());
729 }
730
731 Value *OpndVal = Opnd.getSymVal();
732
733 if (Coeff.isMinusOne() || Coeff.isOne()) {
734 NeedNeg = Coeff.isMinusOne();
735 return OpndVal;
736 }
737
738 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
739 NeedNeg = Coeff.isMinusTwo();
740 return createFAdd(OpndVal, OpndVal);
741 }
742
743 NeedNeg = false;
744 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
745}
746
747// Checks if any operand is negative and we can convert add to sub.
748// This function checks for following negative patterns
749// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
750// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
751// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
752static Value *checkForNegativeOperand(BinaryOperator &I,
753 InstCombiner::BuilderTy &Builder) {
754 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
755
756 // This function creates 2 instructions to replace ADD, we need at least one
757 // of LHS or RHS to have one use to ensure benefit in transform.
758 if (!LHS->hasOneUse() && !RHS->hasOneUse())
759 return nullptr;
760
761 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
762 const APInt *C1 = nullptr, *C2 = nullptr;
763
764 // if ONE is on other side, swap
765 if (match(RHS, m_Add(m_Value(X), m_One())))
766 std::swap(LHS, RHS);
767
768 if (match(LHS, m_Add(m_Value(X), m_One()))) {
769 // if XOR on other side, swap
770 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
771 std::swap(X, RHS);
772
773 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
774 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
775 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
776 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
777 Value *NewAnd = Builder.CreateAnd(Z, *C1);
778 return Builder.CreateSub(RHS, NewAnd, "sub");
779 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
780 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
781 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
782 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
783 return Builder.CreateSub(RHS, NewOr, "sub");
784 }
785 }
786 }
787
788 // Restore LHS and RHS
789 LHS = I.getOperand(0);
790 RHS = I.getOperand(1);
791
792 // if XOR is on other side, swap
793 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
794 std::swap(LHS, RHS);
795
796 // C2 is ODD
797 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
798 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
799 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
800 if (C1->countTrailingZeros() == 0)
801 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
802 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
803 return Builder.CreateSub(RHS, NewOr, "sub");
804 }
805 return nullptr;
806}
807
808/// Wrapping flags may allow combining constants separated by an extend.
809static Instruction *foldNoWrapAdd(BinaryOperator &Add,
810 InstCombiner::BuilderTy &Builder) {
811 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
812 Type *Ty = Add.getType();
813 Constant *Op1C;
814 if (!match(Op1, m_Constant(Op1C)))
815 return nullptr;
816
817 // Try this match first because it results in an add in the narrow type.
818 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
819 Value *X;
820 const APInt *C1, *C2;
821 if (match(Op1, m_APInt(C1)) &&
822 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
823 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
824 Constant *NewC =
825 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
826 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
827 }
828
829 // More general combining of constants in the wide type.
830 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
831 Constant *NarrowC;
832 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
833 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
834 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
835 Value *WideX = Builder.CreateSExt(X, Ty);
836 return BinaryOperator::CreateAdd(WideX, NewC);
837 }
838 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
839 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
840 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
841 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
842 Value *WideX = Builder.CreateZExt(X, Ty);
843 return BinaryOperator::CreateAdd(WideX, NewC);
844 }
845
846 return nullptr;
847}
848
849Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
850 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
851 Type *Ty = Add.getType();
852 Constant *Op1C;
853 if (!match(Op1, m_ImmConstant(Op1C)))
854 return nullptr;
855
856 if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
857 return NV;
858
859 Value *X;
860 Constant *Op00C;
861
862 // add (sub C1, X), C2 --> sub (add C1, C2), X
863 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
864 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
865
866 Value *Y;
867
868 // add (sub X, Y), -1 --> add (not Y), X
869 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
870 match(Op1, m_AllOnes()))
871 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
872
873 // zext(bool) + C -> bool ? C + 1 : C
874 if (match(Op0, m_ZExt(m_Value(X))) &&
875 X->getType()->getScalarSizeInBits() == 1)
876 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
877 // sext(bool) + C -> bool ? C - 1 : C
878 if (match(Op0, m_SExt(m_Value(X))) &&
879 X->getType()->getScalarSizeInBits() == 1)
880 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
881
882 // ~X + C --> (C-1) - X
883 if (match(Op0, m_Not(m_Value(X))))
884 return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X);
885
886 // (iN X s>> (N - 1)) + 1 --> zext (X > -1)
887 const APInt *C;
888 unsigned BitWidth = Ty->getScalarSizeInBits();
889 if (match(Op0, m_OneUse(m_AShr(m_Value(X),
890 m_SpecificIntAllowUndef(BitWidth - 1)))) &&
891 match(Op1, m_One()))
892 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
893
894 if (!match(Op1, m_APInt(C)))
895 return nullptr;
896
897 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
898 Constant *Op01C;
899 if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
900 haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
901 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
902
903 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
904 const APInt *C2;
905 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
906 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
907
908 if (C->isSignMask()) {
909 // If wrapping is not allowed, then the addition must set the sign bit:
910 // X + (signmask) --> X | signmask
911 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
912 return BinaryOperator::CreateOr(Op0, Op1);
913
914 // If wrapping is allowed, then the addition flips the sign bit of LHS:
915 // X + (signmask) --> X ^ signmask
916 return BinaryOperator::CreateXor(Op0, Op1);
917 }
918
919 // Is this add the last step in a convoluted sext?
920 // add(zext(xor i16 X, -32768), -32768) --> sext X
921 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
922 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
923 return CastInst::Create(Instruction::SExt, X, Ty);
924
925 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
926 // (X ^ signmask) + C --> (X + (signmask ^ C))
927 if (C2->isSignMask())
928 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
929
930 // If X has no high-bits set above an xor mask:
931 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
932 if (C2->isMask()) {
933 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
934 if ((*C2 | LHSKnown.Zero).isAllOnes())
935 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
936 }
937
938 // Look for a math+logic pattern that corresponds to sext-in-register of a
939 // value with cleared high bits. Convert that into a pair of shifts:
940 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
941 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
942 if (Op0->hasOneUse() && *C2 == -(*C)) {
943 unsigned BitWidth = Ty->getScalarSizeInBits();
944 unsigned ShAmt = 0;
945 if (C->isPowerOf2())
946 ShAmt = BitWidth - C->logBase2() - 1;
947 else if (C2->isPowerOf2())
948 ShAmt = BitWidth - C2->logBase2() - 1;
949 if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
950 0, &Add)) {
951 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
952 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
953 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
954 }
955 }
956 }
957
958 if (C->isOne() && Op0->hasOneUse()) {
959 // add (sext i1 X), 1 --> zext (not X)
960 // TODO: The smallest IR representation is (select X, 0, 1), and that would
961 // not require the one-use check. But we need to remove a transform in
962 // visitSelect and make sure that IR value tracking for select is equal or
963 // better than for these ops.
964 if (match(Op0, m_SExt(m_Value(X))) &&
965 X->getType()->getScalarSizeInBits() == 1)
966 return new ZExtInst(Builder.CreateNot(X), Ty);
967
968 // Shifts and add used to flip and mask off the low bit:
969 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
970 const APInt *C3;
971 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
972 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
973 Value *NotX = Builder.CreateNot(X);
974 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
975 }
976 }
977
978 return nullptr;
979}
980
981// Matches multiplication expression Op * C where C is a constant. Returns the
982// constant value in C and the other operand in Op. Returns true if such a
983// match is found.
984static bool MatchMul(Value *E, Value *&Op, APInt &C) {
985 const APInt *AI;
986 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
987 C = *AI;
988 return true;
989 }
990 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
991 C = APInt(AI->getBitWidth(), 1);
992 C <<= *AI;
993 return true;
994 }
995 return false;
996}
997
998// Matches remainder expression Op % C where C is a constant. Returns the
999// constant value in C and the other operand in Op. Returns the signedness of
1000// the remainder operation in IsSigned. Returns true if such a match is
1001// found.
1002static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1003 const APInt *AI;
1004 IsSigned = false;
1005 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1006 IsSigned = true;
1007 C = *AI;
1008 return true;
1009 }
1010 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1011 C = *AI;
1012 return true;
1013 }
1014 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1015 C = *AI + 1;
1016 return true;
1017 }
1018 return false;
1019}
1020
1021// Matches division expression Op / C with the given signedness as indicated
1022// by IsSigned, where C is a constant. Returns the constant value in C and the
1023// other operand in Op. Returns true if such a match is found.
1024static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1025 const APInt *AI;
1026 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1027 C = *AI;
1028 return true;
1029 }
1030 if (!IsSigned) {
1031 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1032 C = *AI;
1033 return true;
1034 }
1035 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1036 C = APInt(AI->getBitWidth(), 1);
1037 C <<= *AI;
1038 return true;
1039 }
1040 }
1041 return false;
1042}
1043
1044// Returns whether C0 * C1 with the given signedness overflows.
1045static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1046 bool overflow;
1047 if (IsSigned)
1048 (void)C0.smul_ov(C1, overflow);
1049 else
1050 (void)C0.umul_ov(C1, overflow);
1051 return overflow;
1052}
1053
1054// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1055// does not overflow.
1056Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1057 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1058 Value *X, *MulOpV;
1059 APInt C0, MulOpC;
1060 bool IsSigned;
1061 // Match I = X % C0 + MulOpV * C0
1062 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1063 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1064 C0 == MulOpC) {
1065 Value *RemOpV;
1066 APInt C1;
1067 bool Rem2IsSigned;
1068 // Match MulOpC = RemOpV % C1
1069 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1070 IsSigned == Rem2IsSigned) {
1071 Value *DivOpV;
1072 APInt DivOpC;
1073 // Match RemOpV = X / C0
1074 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1075 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1076 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1077 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1078 : Builder.CreateURem(X, NewDivisor, "urem");
1079 }
1080 }
1081 }
1082
1083 return nullptr;
1084}
1085
1086/// Fold
1087/// (1 << NBits) - 1
1088/// Into:
1089/// ~(-(1 << NBits))
1090/// Because a 'not' is better for bit-tracking analysis and other transforms
1091/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1092static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1093 InstCombiner::BuilderTy &Builder) {
1094 Value *NBits;
1095 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1096 return nullptr;
1097
1098 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1099 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1100 // Be wary of constant folding.
1101 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1102 // Always NSW. But NUW propagates from `add`.
1103 BOp->setHasNoSignedWrap();
1104 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1105 }
1106
1107 return BinaryOperator::CreateNot(NotMask, I.getName());
1108}
1109
1110static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1111 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction")(static_cast <bool> (I.getOpcode() == Instruction::Add &&
"Expecting add instruction") ? void (0) : __assert_fail ("I.getOpcode() == Instruction::Add && \"Expecting add instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1111
, __extension__ __PRETTY_FUNCTION__))
;
1112 Type *Ty = I.getType();
1113 auto getUAddSat = [&]() {
1114 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1115 };
1116
1117 // add (umin X, ~Y), Y --> uaddsat X, Y
1118 Value *X, *Y;
1119 if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1120 m_Deferred(Y))))
1121 return CallInst::Create(getUAddSat(), { X, Y });
1122
1123 // add (umin X, ~C), C --> uaddsat X, C
1124 const APInt *C, *NotC;
1125 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1126 *C == ~*NotC)
1127 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1128
1129 return nullptr;
1130}
1131
1132Instruction *InstCombinerImpl::
1133 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1134 BinaryOperator &I) {
1135 assert((I.getOpcode() == Instruction::Add ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1138
, __extension__ __PRETTY_FUNCTION__))
1
Assuming the condition is false
2
Assuming the condition is false
3
Assuming the condition is true
4
'?' condition is true
1136 I.getOpcode() == Instruction::Or ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1138
, __extension__ __PRETTY_FUNCTION__))
1137 I.getOpcode() == Instruction::Sub) &&(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1138
, __extension__ __PRETTY_FUNCTION__))
1138 "Expecting add/or/sub instruction")(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1138
, __extension__ __PRETTY_FUNCTION__))
;
1139
1140 // We have a subtraction/addition between a (potentially truncated) *logical*
1141 // right-shift of X and a "select".
1142 Value *X, *Select;
1143 Instruction *LowBitsToSkip, *Extract;
1144 if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd(
10
Calling 'm_c_BinOp<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>>'
12
Returning from 'm_c_BinOp<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>>'
13
Calling 'match<llvm::BinaryOperator, llvm::PatternMatch::AnyBinaryOp_match<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>, true>>'
35
Returning from 'match<llvm::BinaryOperator, llvm::PatternMatch::AnyBinaryOp_match<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>, true>>'
1145 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1146 m_Instruction(Extract))),
1147 m_Value(Select))))
5
Calling 'm_Value'
9
Returning from 'm_Value'
1148 return nullptr;
1149
1150 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1151 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
36
Assuming pointer value is null
37
Taking false branch
1152 return nullptr;
1153
1154 Type *XTy = X->getType();
1155 bool HadTrunc = I.getType() != XTy;
38
Assuming the condition is false
1156
1157 // If there was a truncation of extracted value, then we'll need to produce
1158 // one extra instruction, so we need to ensure one instruction will go away.
1159 if (HadTrunc
38.1
'HadTrunc' is false
38.1
'HadTrunc' is false
&& !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
39
Taking false branch
1160 return nullptr;
1161
1162 // Extraction should extract high NBits bits, with shift amount calculated as:
1163 // low bits to skip = shift bitwidth - high bits to extract
1164 // The shift amount itself may be extended, and we need to look past zero-ext
1165 // when matching NBits, that will matter for matching later.
1166 Constant *C;
1167 Value *NBits;
1168 if (!match(
41
Taking false branch
1169 LowBitsToSkip,
1170 m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1171 !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
40
Assuming the condition is false
1172 APInt(C->getType()->getScalarSizeInBits(),
1173 X->getType()->getScalarSizeInBits()))))
1174 return nullptr;
1175
1176 // Sign-extending value can be zero-extended if we `sub`tract it,
1177 // or sign-extended otherwise.
1178 auto SkipExtInMagic = [&I](Value *&V) {
1179 if (I.getOpcode() == Instruction::Sub)
43
Taking true branch
1180 match(V, m_ZExtOrSelf(m_Value(V)));
44
Calling 'm_Value'
46
Returning from 'm_Value'
47
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
55
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
56
Calling 'match<llvm::Value, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'
58
Returning from 'match<llvm::Value, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'
1181 else
1182 match(V, m_SExtOrSelf(m_Value(V)));
1183 };
59
Returning without writing to 'V'
1184
1185 // Now, finally validate the sign-extending magic.
1186 // `select` itself may be appropriately extended, look past that.
1187 SkipExtInMagic(Select);
42
Calling 'operator()'
60
Returning from 'operator()'
1188
1189 ICmpInst::Predicate Pred;
1190 const APInt *Thr;
1191 Value *SignExtendingValue, *Zero;
1192 bool ShouldSignext;
1193 // It must be a select between two values we will later establish to be a
1194 // sign-extending value and a zero constant. The condition guarding the
1195 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1196 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
61
Passing null pointer value via 1st parameter 'V'
62
Calling 'match<llvm::Value, llvm::PatternMatch::ThreeOps_match<llvm::PatternMatch::CmpClass_match<llvm::PatternMatch::specificval_ty, llvm::PatternMatch::apint_match, llvm::ICmpInst, llvm::CmpInst::Predicate, false>, llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Value>, 57>>'
1197 m_Value(SignExtendingValue), m_Value(Zero))) ||
1198 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1199 return nullptr;
1200
1201 // icmp-select pair is commutative.
1202 if (!ShouldSignext)
1203 std::swap(SignExtendingValue, Zero);
1204
1205 // If we should not perform sign-extension then we must add/or/subtract zero.
1206 if (!match(Zero, m_Zero()))
1207 return nullptr;
1208 // Otherwise, it should be some constant, left-shifted by the same NBits we
1209 // had in `lshr`. Said left-shift can also be appropriately extended.
1210 // Again, we must look past zero-ext when looking for NBits.
1211 SkipExtInMagic(SignExtendingValue);
1212 Constant *SignExtendingValueBaseConstant;
1213 if (!match(SignExtendingValue,
1214 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1215 m_ZExtOrSelf(m_Specific(NBits)))))
1216 return nullptr;
1217 // If we `sub`, then the constant should be one, else it should be all-ones.
1218 if (I.getOpcode() == Instruction::Sub
1219 ? !match(SignExtendingValueBaseConstant, m_One())
1220 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1221 return nullptr;
1222
1223 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1224 Extract->getName() + ".sext");
1225 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1226 if (!HadTrunc)
1227 return NewAShr;
1228
1229 Builder.Insert(NewAShr);
1230 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1231}
1232
1233/// This is a specialization of a more general transform from
1234/// foldUsingDistributiveLaws. If that code can be made to work optimally
1235/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1236static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1237 InstCombiner::BuilderTy &Builder) {
1238 // TODO: Also handle mul by doubling the shift amount?
1239 assert((I.getOpcode() == Instruction::Add ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1241
, __extension__ __PRETTY_FUNCTION__))
1240 I.getOpcode() == Instruction::Sub) &&(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1241
, __extension__ __PRETTY_FUNCTION__))
1241 "Expected add/sub")(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1241
, __extension__ __PRETTY_FUNCTION__))
;
1242 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1243 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1244 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1245 return nullptr;
1246
1247 Value *X, *Y, *ShAmt;
1248 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1249 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1250 return nullptr;
1251
1252 // No-wrap propagates only when all ops have no-wrap.
1253 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1254 Op1->hasNoSignedWrap();
1255 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1256 Op1->hasNoUnsignedWrap();
1257
1258 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1259 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1260 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1261 NewI->setHasNoSignedWrap(HasNSW);
1262 NewI->setHasNoUnsignedWrap(HasNUW);
1263 }
1264 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1265 NewShl->setHasNoSignedWrap(HasNSW);
1266 NewShl->setHasNoUnsignedWrap(HasNUW);
1267 return NewShl;
1268}
1269
1270/// Reduce a sequence of masked half-width multiplies to a single multiply.
1271/// ((XLow * YHigh) + (YLow * XHigh)) << HalfBits) + (XLow * YLow) --> X * Y
1272static Instruction *foldBoxMultiply(BinaryOperator &I) {
1273 unsigned BitWidth = I.getType()->getScalarSizeInBits();
1274 // Skip the odd bitwidth types.
1275 if ((BitWidth & 0x1))
1276 return nullptr;
1277
1278 unsigned HalfBits = BitWidth >> 1;
1279 APInt HalfMask = APInt::getMaxValue(HalfBits);
1280
1281 // ResLo = (CrossSum << HalfBits) + (YLo * XLo)
1282 Value *XLo, *YLo;
1283 Value *CrossSum;
1284 if (!match(&I, m_c_Add(m_Shl(m_Value(CrossSum), m_SpecificInt(HalfBits)),
1285 m_Mul(m_Value(YLo), m_Value(XLo)))))
1286 return nullptr;
1287
1288 // XLo = X & HalfMask
1289 // YLo = Y & HalfMask
1290 // TODO: Refactor with SimplifyDemandedBits or KnownBits known leading zeros
1291 // to enhance robustness
1292 Value *X, *Y;
1293 if (!match(XLo, m_And(m_Value(X), m_SpecificInt(HalfMask))) ||
1294 !match(YLo, m_And(m_Value(Y), m_SpecificInt(HalfMask))))
1295 return nullptr;
1296
1297 // CrossSum = (X' * (Y >> Halfbits)) + (Y' * (X >> HalfBits))
1298 // X' can be either X or XLo in the pattern (and the same for Y')
1299 if (match(CrossSum,
1300 m_c_Add(m_c_Mul(m_LShr(m_Specific(Y), m_SpecificInt(HalfBits)),
1301 m_CombineOr(m_Specific(X), m_Specific(XLo))),
1302 m_c_Mul(m_LShr(m_Specific(X), m_SpecificInt(HalfBits)),
1303 m_CombineOr(m_Specific(Y), m_Specific(YLo))))))
1304 return BinaryOperator::CreateMul(X, Y);
1305
1306 return nullptr;
1307}
1308
1309Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1310 if (Value *V = simplifyAddInst(I.getOperand(0), I.getOperand(1),
1311 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1312 SQ.getWithInstruction(&I)))
1313 return replaceInstUsesWith(I, V);
1314
1315 if (SimplifyAssociativeOrCommutative(I))
1316 return &I;
1317
1318 if (Instruction *X = foldVectorBinop(I))
1319 return X;
1320
1321 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1322 return Phi;
1323
1324 // (A*B)+(A*C) -> A*(B+C) etc
1325 if (Value *V = foldUsingDistributiveLaws(I))
1326 return replaceInstUsesWith(I, V);
1327
1328 if (Instruction *R = foldBoxMultiply(I))
1329 return R;
1330
1331 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1332 return R;
1333
1334 if (Instruction *X = foldAddWithConstant(I))
1335 return X;
1336
1337 if (Instruction *X = foldNoWrapAdd(I, Builder))
1338 return X;
1339
1340 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1341 Type *Ty = I.getType();
1342 if (Ty->isIntOrIntVectorTy(1))
1343 return BinaryOperator::CreateXor(LHS, RHS);
1344
1345 // X + X --> X << 1
1346 if (LHS == RHS) {
1347 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1348 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1349 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1350 return Shl;
1351 }
1352
1353 Value *A, *B;
1354 if (match(LHS, m_Neg(m_Value(A)))) {
1355 // -A + -B --> -(A + B)
1356 if (match(RHS, m_Neg(m_Value(B))))
1357 return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1358
1359 // -A + B --> B - A
1360 return BinaryOperator::CreateSub(RHS, A);
1361 }
1362
1363 // A + -B --> A - B
1364 if (match(RHS, m_Neg(m_Value(B))))
1365 return BinaryOperator::CreateSub(LHS, B);
1366
1367 if (Value *V = checkForNegativeOperand(I, Builder))
1368 return replaceInstUsesWith(I, V);
1369
1370 // (A + 1) + ~B --> A - B
1371 // ~B + (A + 1) --> A - B
1372 // (~B + A) + 1 --> A - B
1373 // (A + ~B) + 1 --> A - B
1374 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1375 match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1376 return BinaryOperator::CreateSub(A, B);
1377
1378 // (A + RHS) + RHS --> A + (RHS << 1)
1379 if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1380 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1381
1382 // LHS + (A + LHS) --> A + (LHS << 1)
1383 if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1384 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1385
1386 {
1387 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1388 Constant *C1, *C2;
1389 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1390 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1391 (LHS->hasOneUse() || RHS->hasOneUse())) {
1392 Value *Sub = Builder.CreateSub(A, B);
1393 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1394 }
1395 }
1396
1397 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1398 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1399
1400 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1401 const APInt *C1, *C2;
1402 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1403 APInt one(C2->getBitWidth(), 1);
1404 APInt minusC1 = -(*C1);
1405 if (minusC1 == (one << *C2)) {
1406 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1407 return BinaryOperator::CreateSRem(RHS, NewRHS);
1408 }
1409 }
1410
1411 // (A & 2^C1) + A => A & (2^C1 - 1) iff bit C1 in A is a sign bit
1412 if (match(&I, m_c_Add(m_And(m_Value(A), m_APInt(C1)), m_Deferred(A))) &&
1413 C1->isPowerOf2() && (ComputeNumSignBits(A) > C1->countLeadingZeros())) {
1414 Constant *NewMask = ConstantInt::get(RHS->getType(), *C1 - 1);
1415 return BinaryOperator::CreateAnd(A, NewMask);
1416 }
1417
1418 // A+B --> A|B iff A and B have no bits set in common.
1419 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1420 return BinaryOperator::CreateOr(LHS, RHS);
1421
1422 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1423 return Ext;
1424
1425 // (add (xor A, B) (and A, B)) --> (or A, B)
1426 // (add (and A, B) (xor A, B)) --> (or A, B)
1427 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1428 m_c_And(m_Deferred(A), m_Deferred(B)))))
1429 return BinaryOperator::CreateOr(A, B);
1430
1431 // (add (or A, B) (and A, B)) --> (add A, B)
1432 // (add (and A, B) (or A, B)) --> (add A, B)
1433 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1434 m_c_And(m_Deferred(A), m_Deferred(B))))) {
1435 // Replacing operands in-place to preserve nuw/nsw flags.
1436 replaceOperand(I, 0, A);
1437 replaceOperand(I, 1, B);
1438 return &I;
1439 }
1440
1441 // (add A (or A, -A)) --> (and (add A, -1) A)
1442 // (add A (or -A, A)) --> (and (add A, -1) A)
1443 // (add (or A, -A) A) --> (and (add A, -1) A)
1444 // (add (or -A, A) A) --> (and (add A, -1) A)
1445 if (match(&I, m_c_BinOp(m_Value(A), m_OneUse(m_c_Or(m_Neg(m_Deferred(A)),
1446 m_Deferred(A)))))) {
1447 Value *Add =
1448 Builder.CreateAdd(A, Constant::getAllOnesValue(A->getType()), "",
1449 I.hasNoUnsignedWrap(), I.hasNoSignedWrap());
1450 return BinaryOperator::CreateAnd(Add, A);
1451 }
1452
1453 // Canonicalize ((A & -A) - 1) --> ((A - 1) & ~A)
1454 // Forms all commutable operations, and simplifies ctpop -> cttz folds.
1455 if (match(&I,
1456 m_Add(m_OneUse(m_c_And(m_Value(A), m_OneUse(m_Neg(m_Deferred(A))))),
1457 m_AllOnes()))) {
1458 Constant *AllOnes = ConstantInt::getAllOnesValue(RHS->getType());
1459 Value *Dec = Builder.CreateAdd(A, AllOnes);
1460 Value *Not = Builder.CreateXor(A, AllOnes);
1461 return BinaryOperator::CreateAnd(Dec, Not);
1462 }
1463
1464 // Disguised reassociation/factorization:
1465 // ~(A * C1) + A
1466 // ((A * -C1) - 1) + A
1467 // ((A * -C1) + A) - 1
1468 // (A * (1 - C1)) - 1
1469 if (match(&I,
1470 m_c_Add(m_OneUse(m_Not(m_OneUse(m_Mul(m_Value(A), m_APInt(C1))))),
1471 m_Deferred(A)))) {
1472 Type *Ty = I.getType();
1473 Constant *NewMulC = ConstantInt::get(Ty, 1 - *C1);
1474 Value *NewMul = Builder.CreateMul(A, NewMulC);
1475 return BinaryOperator::CreateAdd(NewMul, ConstantInt::getAllOnesValue(Ty));
1476 }
1477
1478 // (A * -2**C) + B --> B - (A << C)
1479 const APInt *NegPow2C;
1480 if (match(&I, m_c_Add(m_OneUse(m_Mul(m_Value(A), m_NegatedPower2(NegPow2C))),
1481 m_Value(B)))) {
1482 Constant *ShiftAmtC = ConstantInt::get(Ty, NegPow2C->countTrailingZeros());
1483 Value *Shl = Builder.CreateShl(A, ShiftAmtC);
1484 return BinaryOperator::CreateSub(B, Shl);
1485 }
1486
1487 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1488 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1489 // computeKnownBits.
1490 bool Changed = false;
1491 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1492 Changed = true;
1493 I.setHasNoSignedWrap(true);
1494 }
1495 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1496 Changed = true;
1497 I.setHasNoUnsignedWrap(true);
1498 }
1499
1500 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1501 return V;
1502
1503 if (Instruction *V =
1504 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1505 return V;
1506
1507 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1508 return SatAdd;
1509
1510 // usub.sat(A, B) + B => umax(A, B)
1511 if (match(&I, m_c_BinOp(
1512 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1513 m_Deferred(B)))) {
1514 return replaceInstUsesWith(I,
1515 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1516 }
1517
1518 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1519 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1520 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1521 haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1522 return replaceInstUsesWith(
1523 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1524 {Builder.CreateOr(A, B)}));
1525
1526 return Changed ? &I : nullptr;
1527}
1528
1529/// Eliminate an op from a linear interpolation (lerp) pattern.
1530static Instruction *factorizeLerp(BinaryOperator &I,
1531 InstCombiner::BuilderTy &Builder) {
1532 Value *X, *Y, *Z;
1533 if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1534 m_OneUse(m_FSub(m_FPOne(),
1535 m_Value(Z))))),
1536 m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1537 return nullptr;
1538
1539 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1540 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1541 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1542 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1543}
1544
1545/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1546static Instruction *factorizeFAddFSub(BinaryOperator &I,
1547 InstCombiner::BuilderTy &Builder) {
1548 assert((I.getOpcode() == Instruction::FAdd ||(static_cast <bool> ((I.getOpcode() == Instruction::FAdd
|| I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1549
, __extension__ __PRETTY_FUNCTION__))
1549 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub")(static_cast <bool> ((I.getOpcode() == Instruction::FAdd
|| I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1549
, __extension__ __PRETTY_FUNCTION__))
;
1550 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&(static_cast <bool> (I.hasAllowReassoc() && I.hasNoSignedZeros
() && "FP factorization requires FMF") ? void (0) : __assert_fail
("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1551
, __extension__ __PRETTY_FUNCTION__))
1551 "FP factorization requires FMF")(static_cast <bool> (I.hasAllowReassoc() && I.hasNoSignedZeros
() && "FP factorization requires FMF") ? void (0) : __assert_fail
("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1551
, __extension__ __PRETTY_FUNCTION__))
;
1552
1553 if (Instruction *Lerp = factorizeLerp(I, Builder))
1554 return Lerp;
1555
1556 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1557 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1558 return nullptr;
1559
1560 Value *X, *Y, *Z;
1561 bool IsFMul;
1562 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1563 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1564 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1565 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1566 IsFMul = true;
1567 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1568 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1569 IsFMul = false;
1570 else
1571 return nullptr;
1572
1573 // (X * Z) + (Y * Z) --> (X + Y) * Z
1574 // (X * Z) - (Y * Z) --> (X - Y) * Z
1575 // (X / Z) + (Y / Z) --> (X + Y) / Z
1576 // (X / Z) - (Y / Z) --> (X - Y) / Z
1577 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1578 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1579 : Builder.CreateFSubFMF(X, Y, &I);
1580
1581 // Bail out if we just created a denormal constant.
1582 // TODO: This is copied from a previous implementation. Is it necessary?
1583 const APFloat *C;
1584 if (match(XY, m_APFloat(C)) && !C->isNormal())
1585 return nullptr;
1586
1587 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1588 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1589}
1590
1591Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1592 if (Value *V = simplifyFAddInst(I.getOperand(0), I.getOperand(1),
1593 I.getFastMathFlags(),
1594 SQ.getWithInstruction(&I)))
1595 return replaceInstUsesWith(I, V);
1596
1597 if (SimplifyAssociativeOrCommutative(I))
1598 return &I;
1599
1600 if (Instruction *X = foldVectorBinop(I))
1601 return X;
1602
1603 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1604 return Phi;
1605
1606 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1607 return FoldedFAdd;
1608
1609 // (-X) + Y --> Y - X
1610 Value *X, *Y;
1611 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1612 return BinaryOperator::CreateFSubFMF(Y, X, &I);
1613
1614 // Similar to above, but look through fmul/fdiv for the negated term.
1615 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1616 Value *Z;
1617 if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1618 m_Value(Z)))) {
1619 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1620 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1621 }
1622 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1623 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1624 if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1625 m_Value(Z))) ||
1626 match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1627 m_Value(Z)))) {
1628 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1629 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1630 }
1631
1632 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1633 // integer add followed by a promotion.
1634 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1635 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1636 Value *LHSIntVal = LHSConv->getOperand(0);
1637 Type *FPType = LHSConv->getType();
1638
1639 // TODO: This check is overly conservative. In many cases known bits
1640 // analysis can tell us that the result of the addition has less significant
1641 // bits than the integer type can hold.
1642 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1643 Type *FScalarTy = FTy->getScalarType();
1644 Type *IScalarTy = ITy->getScalarType();
1645
1646 // Do we have enough bits in the significand to represent the result of
1647 // the integer addition?
1648 unsigned MaxRepresentableBits =
1649 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1650 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1651 };
1652
1653 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1654 // ... if the constant fits in the integer value. This is useful for things
1655 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1656 // requires a constant pool load, and generally allows the add to be better
1657 // instcombined.
1658 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1659 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1660 Constant *CI =
1661 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1662 if (LHSConv->hasOneUse() &&
1663 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1664 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1665 // Insert the new integer add.
1666 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1667 return new SIToFPInst(NewAdd, I.getType());
1668 }
1669 }
1670
1671 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1672 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1673 Value *RHSIntVal = RHSConv->getOperand(0);
1674 // It's enough to check LHS types only because we require int types to
1675 // be the same for this transform.
1676 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1677 // Only do this if x/y have the same type, if at least one of them has a
1678 // single use (so we don't increase the number of int->fp conversions),
1679 // and if the integer add will not overflow.
1680 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1681 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1682 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1683 // Insert the new integer add.
1684 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1685 return new SIToFPInst(NewAdd, I.getType());
1686 }
1687 }
1688 }
1689 }
1690
1691 // Handle specials cases for FAdd with selects feeding the operation
1692 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1693 return replaceInstUsesWith(I, V);
1694
1695 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1696 if (Instruction *F = factorizeFAddFSub(I, Builder))
1697 return F;
1698
1699 // Try to fold fadd into start value of reduction intrinsic.
1700 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1701 m_AnyZeroFP(), m_Value(X))),
1702 m_Value(Y)))) {
1703 // fadd (rdx 0.0, X), Y --> rdx Y, X
1704 return replaceInstUsesWith(
1705 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1706 {X->getType()}, {Y, X}, &I));
1707 }
1708 const APFloat *StartC, *C;
1709 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1710 m_APFloat(StartC), m_Value(X)))) &&
1711 match(RHS, m_APFloat(C))) {
1712 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1713 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1714 return replaceInstUsesWith(
1715 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1716 {X->getType()}, {NewStartC, X}, &I));
1717 }
1718
1719 // (X * MulC) + X --> X * (MulC + 1.0)
1720 Constant *MulC;
1721 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1722 m_Deferred(X)))) {
1723 if (Constant *NewMulC = ConstantFoldBinaryOpOperands(
1724 Instruction::FAdd, MulC, ConstantFP::get(I.getType(), 1.0), DL))
1725 return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
1726 }
1727
1728 // (-X - Y) + (X + Z) --> Z - Y
1729 if (match(&I, m_c_FAdd(m_FSub(m_FNeg(m_Value(X)), m_Value(Y)),
1730 m_c_FAdd(m_Deferred(X), m_Value(Z)))))
1731 return BinaryOperator::CreateFSubFMF(Z, Y, &I);
1732
1733 if (Value *V = FAddCombine(Builder).simplify(&I))
1734 return replaceInstUsesWith(I, V);
1735 }
1736
1737 return nullptr;
1738}
1739
1740/// Optimize pointer differences into the same array into a size. Consider:
1741/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1742/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1743Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
1744 Type *Ty, bool IsNUW) {
1745 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1746 // this.
1747 bool Swapped = false;
1748 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1749 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1750 std::swap(LHS, RHS);
1751 Swapped = true;
1752 }
1753
1754 // Require at least one GEP with a common base pointer on both sides.
1755 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1756 // (gep X, ...) - X
1757 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1758 RHS->stripPointerCasts()) {
1759 GEP1 = LHSGEP;
1760 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1761 // (gep X, ...) - (gep X, ...)
1762 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1763 RHSGEP->getOperand(0)->stripPointerCasts()) {
1764 GEP1 = LHSGEP;
1765 GEP2 = RHSGEP;
1766 }
1767 }
1768 }
1769
1770 if (!GEP1)
1771 return nullptr;
1772
1773 if (GEP2) {
1774 // (gep X, ...) - (gep X, ...)
1775 //
1776 // Avoid duplicating the arithmetic if there are more than one non-constant
1777 // indices between the two GEPs and either GEP has a non-constant index and
1778 // multiple users. If zero non-constant index, the result is a constant and
1779 // there is no duplication. If one non-constant index, the result is an add
1780 // or sub with a constant, which is no larger than the original code, and
1781 // there's no duplicated arithmetic, even if either GEP has multiple
1782 // users. If more than one non-constant indices combined, as long as the GEP
1783 // with at least one non-constant index doesn't have multiple users, there
1784 // is no duplication.
1785 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1786 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1787 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1788 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1789 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1790 return nullptr;
1791 }
1792 }
1793
1794 // Emit the offset of the GEP and an intptr_t.
1795 Value *Result = EmitGEPOffset(GEP1);
1796
1797 // If this is a single inbounds GEP and the original sub was nuw,
1798 // then the final multiplication is also nuw.
1799 if (auto *I = dyn_cast<Instruction>(Result))
1800 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1801 I->getOpcode() == Instruction::Mul)
1802 I->setHasNoUnsignedWrap();
1803
1804 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
1805 // If both GEPs are inbounds, then the subtract does not have signed overflow.
1806 if (GEP2) {
1807 Value *Offset = EmitGEPOffset(GEP2);
1808 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
1809 GEP1->isInBounds() && GEP2->isInBounds());
1810 }
1811
1812 // If we have p - gep(p, ...) then we have to negate the result.
1813 if (Swapped)
1814 Result = Builder.CreateNeg(Result, "diff.neg");
1815
1816 return Builder.CreateIntCast(Result, Ty, true);
1817}
1818
1819static Instruction *foldSubOfMinMax(BinaryOperator &I,
1820 InstCombiner::BuilderTy &Builder) {
1821 Value *Op0 = I.getOperand(0);
1822 Value *Op1 = I.getOperand(1);
1823 Type *Ty = I.getType();
1824 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op1);
1825 if (!MinMax)
1826 return nullptr;
1827
1828 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
1829 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
1830 Value *X = MinMax->getLHS();
1831 Value *Y = MinMax->getRHS();
1832 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
1833 (Op0->hasOneUse() || Op1->hasOneUse())) {
1834 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
1835 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
1836 return CallInst::Create(F, {X, Y});
1837 }
1838
1839 // sub(add(X,Y),umin(Y,Z)) --> add(X,usub.sat(Y,Z))
1840 // sub(add(X,Z),umin(Y,Z)) --> add(X,usub.sat(Z,Y))
1841 Value *Z;
1842 if (match(Op1, m_OneUse(m_UMin(m_Value(Y), m_Value(Z))))) {
1843 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Y), m_Value(X))))) {
1844 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Y, Z});
1845 return BinaryOperator::CreateAdd(X, USub);
1846 }
1847 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Z), m_Value(X))))) {
1848 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, Ty, {Z, Y});
1849 return BinaryOperator::CreateAdd(X, USub);
1850 }
1851 }
1852
1853 // sub Op0, smin((sub nsw Op0, Z), 0) --> smax Op0, Z
1854 // sub Op0, smax((sub nsw Op0, Z), 0) --> smin Op0, Z
1855 if (MinMax->isSigned() && match(Y, m_ZeroInt()) &&
1856 match(X, m_NSWSub(m_Specific(Op0), m_Value(Z)))) {
1857 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMax->getIntrinsicID());
1858 Function *F = Intrinsic::getDeclaration(I.getModule(), InvID, Ty);
1859 return CallInst::Create(F, {Op0, Z});
1860 }
1861
1862 return nullptr;
1863}
1864
1865Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
1866 if (Value *V = simplifySubInst(I.getOperand(0), I.getOperand(1),
1867 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1868 SQ.getWithInstruction(&I)))
1869 return replaceInstUsesWith(I, V);
1870
1871 if (Instruction *X = foldVectorBinop(I))
1872 return X;
1873
1874 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1875 return Phi;
1876
1877 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1878
1879 // If this is a 'B = x-(-A)', change to B = x+A.
1880 // We deal with this without involving Negator to preserve NSW flag.
1881 if (Value *V = dyn_castNegVal(Op1)) {
1882 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1883
1884 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1885 assert(BO->getOpcode() == Instruction::Sub &&(static_cast <bool> (BO->getOpcode() == Instruction::
Sub && "Expected a subtraction operator!") ? void (0)
: __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1886
, __extension__ __PRETTY_FUNCTION__))
1886 "Expected a subtraction operator!")(static_cast <bool> (BO->getOpcode() == Instruction::
Sub && "Expected a subtraction operator!") ? void (0)
: __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1886
, __extension__ __PRETTY_FUNCTION__))
;
1887 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1888 Res->setHasNoSignedWrap(true);
1889 } else {
1890 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1891 Res->setHasNoSignedWrap(true);
1892 }
1893
1894 return Res;
1895 }
1896
1897 // Try this before Negator to preserve NSW flag.
1898 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1899 return R;
1900
1901 Constant *C;
1902 if (match(Op0, m_ImmConstant(C))) {
1903 Value *X;
1904 Constant *C2;
1905
1906 // C-(X+C2) --> (C-C2)-X
1907 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
1908 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1909 }
1910
1911 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
1912 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1913 return Ext;
1914
1915 bool Changed = false;
1916 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1917 Changed = true;
1918 I.setHasNoSignedWrap(true);
1919 }
1920 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1921 Changed = true;
1922 I.setHasNoUnsignedWrap(true);
1923 }
1924
1925 return Changed ? &I : nullptr;
1926 };
1927
1928 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
1929 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
1930 // a pure negation used by a select that looks like abs/nabs.
1931 bool IsNegation = match(Op0, m_ZeroInt());
1932 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
1933 const Instruction *UI = dyn_cast<Instruction>(U);
1934 if (!UI)
1935 return false;
1936 return match(UI,
1937 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
1938 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
1939 })) {
1940 if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
1941 return BinaryOperator::CreateAdd(NegOp1, Op0);
1942 }
1943 if (IsNegation)
1944 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
1945
1946 // (A*B)-(A*C) -> A*(B-C) etc
1947 if (Value *V = foldUsingDistributiveLaws(I))
1948 return replaceInstUsesWith(I, V);
1949
1950 if (I.getType()->isIntOrIntVectorTy(1))
1951 return BinaryOperator::CreateXor(Op0, Op1);
1952
1953 // Replace (-1 - A) with (~A).
1954 if (match(Op0, m_AllOnes()))
1955 return BinaryOperator::CreateNot(Op1);
1956
1957 // (X + -1) - Y --> ~Y + X
1958 Value *X, *Y;
1959 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1960 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1961
1962 // Reassociate sub/add sequences to create more add instructions and
1963 // reduce dependency chains:
1964 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
1965 Value *Z;
1966 if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
1967 m_Value(Z))))) {
1968 Value *XZ = Builder.CreateAdd(X, Z);
1969 Value *YW = Builder.CreateAdd(Y, Op1);
1970 return BinaryOperator::CreateSub(XZ, YW);
1971 }
1972
1973 // ((X - Y) - Op1) --> X - (Y + Op1)
1974 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
1975 Value *Add = Builder.CreateAdd(Y, Op1);
1976 return BinaryOperator::CreateSub(X, Add);
1977 }
1978
1979 // (~X) - (~Y) --> Y - X
1980 // This is placed after the other reassociations and explicitly excludes a
1981 // sub-of-sub pattern to avoid infinite looping.
1982 if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
1983 isFreeToInvert(Op1, Op1->hasOneUse()) &&
1984 !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
1985 Value *NotOp0 = Builder.CreateNot(Op0);
1986 Value *NotOp1 = Builder.CreateNot(Op1);
1987 return BinaryOperator::CreateSub(NotOp1, NotOp0);
1988 }
1989
1990 auto m_AddRdx = [](Value *&Vec) {
1991 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
1992 };
1993 Value *V0, *V1;
1994 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
1995 V0->getType() == V1->getType()) {
1996 // Difference of sums is sum of differences:
1997 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
1998 Value *Sub = Builder.CreateSub(V0, V1);
1999 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
2000 {Sub->getType()}, {Sub});
2001 return replaceInstUsesWith(I, Rdx);
2002 }
2003
2004 if (Constant *C = dyn_cast<Constant>(Op0)) {
2005 Value *X;
2006 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2007 // C - (zext bool) --> bool ? C - 1 : C
2008 return SelectInst::Create(X, InstCombiner::SubOne(C), C);
2009 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
2010 // C - (sext bool) --> bool ? C + 1 : C
2011 return SelectInst::Create(X, InstCombiner::AddOne(C), C);
2012
2013 // C - ~X == X + (1+C)
2014 if (match(Op1, m_Not(m_Value(X))))
2015 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
2016
2017 // Try to fold constant sub into select arguments.
2018 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2019 if (Instruction *R = FoldOpIntoSelect(I, SI))
2020 return R;
2021
2022 // Try to fold constant sub into PHI values.
2023 if (PHINode *PN = dyn_cast<PHINode>(Op1))
2024 if (Instruction *R = foldOpIntoPhi(I, PN))
2025 return R;
2026
2027 Constant *C2;
2028
2029 // C-(C2-X) --> X+(C-C2)
2030 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
2031 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
2032 }
2033
2034 const APInt *Op0C;
2035 if (match(Op0, m_APInt(Op0C))) {
2036 if (Op0C->isMask()) {
2037 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
2038 // zero.
2039 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
2040 if ((*Op0C | RHSKnown.Zero).isAllOnes())
2041 return BinaryOperator::CreateXor(Op1, Op0);
2042 }
2043
2044 // C - ((C3 -nuw X) & C2) --> (C - (C2 & C3)) + (X & C2) when:
2045 // (C3 - ((C2 & C3) - 1)) is pow2
2046 // ((C2 + C3) & ((C2 & C3) - 1)) == ((C2 & C3) - 1)
2047 // C2 is negative pow2 || sub nuw
2048 const APInt *C2, *C3;
2049 BinaryOperator *InnerSub;
2050 if (match(Op1, m_OneUse(m_And(m_BinOp(InnerSub), m_APInt(C2)))) &&
2051 match(InnerSub, m_Sub(m_APInt(C3), m_Value(X))) &&
2052 (InnerSub->hasNoUnsignedWrap() || C2->isNegatedPowerOf2())) {
2053 APInt C2AndC3 = *C2 & *C3;
2054 APInt C2AndC3Minus1 = C2AndC3 - 1;
2055 APInt C2AddC3 = *C2 + *C3;
2056 if ((*C3 - C2AndC3Minus1).isPowerOf2() &&
2057 C2AndC3Minus1.isSubsetOf(C2AddC3)) {
2058 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(), *C2));
2059 return BinaryOperator::CreateAdd(
2060 And, ConstantInt::get(I.getType(), *Op0C - C2AndC3));
2061 }
2062 }
2063 }
2064
2065 {
2066 Value *Y;
2067 // X-(X+Y) == -Y X-(Y+X) == -Y
2068 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
2069 return BinaryOperator::CreateNeg(Y);
2070
2071 // (X-Y)-X == -Y
2072 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
2073 return BinaryOperator::CreateNeg(Y);
2074 }
2075
2076 // (sub (or A, B) (and A, B)) --> (xor A, B)
2077 {
2078 Value *A, *B;
2079 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2080 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2081 return BinaryOperator::CreateXor(A, B);
2082 }
2083
2084 // (sub (add A, B) (or A, B)) --> (and A, B)
2085 {
2086 Value *A, *B;
2087 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2088 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2089 return BinaryOperator::CreateAnd(A, B);
2090 }
2091
2092 // (sub (add A, B) (and A, B)) --> (or A, B)
2093 {
2094 Value *A, *B;
2095 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2096 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
2097 return BinaryOperator::CreateOr(A, B);
2098 }
2099
2100 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
2101 {
2102 Value *A, *B;
2103 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2104 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2105 (Op0->hasOneUse() || Op1->hasOneUse()))
2106 return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
2107 }
2108
2109 // (sub (or A, B), (xor A, B)) --> (and A, B)
2110 {
2111 Value *A, *B;
2112 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2113 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2114 return BinaryOperator::CreateAnd(A, B);
2115 }
2116
2117 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
2118 {
2119 Value *A, *B;
2120 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2121 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
2122 (Op0->hasOneUse() || Op1->hasOneUse()))
2123 return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
2124 }
2125
2126 {
2127 Value *Y;
2128 // ((X | Y) - X) --> (~X & Y)
2129 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
2130 return BinaryOperator::CreateAnd(
2131 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
2132 }
2133
2134 {
2135 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
2136 Value *X;
2137 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
2138 m_OneUse(m_Neg(m_Value(X))))))) {
2139 return BinaryOperator::CreateNeg(Builder.CreateAnd(
2140 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
2141 }
2142 }
2143
2144 {
2145 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2146 Constant *C;
2147 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2148 return BinaryOperator::CreateNeg(
2149 Builder.CreateAnd(Op1, Builder.CreateNot(C)));
2150 }
2151 }
2152
2153 if (Instruction *R = foldSubOfMinMax(I, Builder))
2154 return R;
2155
2156 {
2157 // If we have a subtraction between some value and a select between
2158 // said value and something else, sink subtraction into select hands, i.e.:
2159 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2160 // ->
2161 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2162 // or
2163 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2164 // ->
2165 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2166 // This will result in select between new subtraction and 0.
2167 auto SinkSubIntoSelect =
2168 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2169 auto SubBuilder) -> Instruction * {
2170 Value *Cond, *TrueVal, *FalseVal;
2171 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2172 m_Value(FalseVal)))))
2173 return nullptr;
2174 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2175 return nullptr;
2176 // While it is really tempting to just create two subtractions and let
2177 // InstCombine fold one of those to 0, it isn't possible to do so
2178 // because of worklist visitation order. So ugly it is.
2179 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2180 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2181 Constant *Zero = Constant::getNullValue(Ty);
2182 SelectInst *NewSel =
2183 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2184 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2185 // Preserve prof metadata if any.
2186 NewSel->copyMetadata(cast<Instruction>(*Select));
2187 return NewSel;
2188 };
2189 if (Instruction *NewSel = SinkSubIntoSelect(
2190 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2191 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2192 return Builder->CreateSub(OtherHandOfSelect,
2193 /*OtherHandOfSub=*/Op1);
2194 }))
2195 return NewSel;
2196 if (Instruction *NewSel = SinkSubIntoSelect(
2197 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2198 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2199 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2200 OtherHandOfSelect);
2201 }))
2202 return NewSel;
2203 }
2204
2205 // (X - (X & Y)) --> (X & ~Y)
2206 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2207 (Op1->hasOneUse() || isa<Constant>(Y)))
2208 return BinaryOperator::CreateAnd(
2209 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2210
2211 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2212 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2213 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2214 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2215 // As long as Y is freely invertible, this will be neutral or a win.
2216 // Note: We don't generate the inverse max/min, just create the 'not' of
2217 // it and let other folds do the rest.
2218 if (match(Op0, m_Not(m_Value(X))) &&
2219 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2220 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2221 Value *Not = Builder.CreateNot(Op1);
2222 return BinaryOperator::CreateSub(Not, X);
2223 }
2224 if (match(Op1, m_Not(m_Value(X))) &&
2225 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2226 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2227 Value *Not = Builder.CreateNot(Op0);
2228 return BinaryOperator::CreateSub(X, Not);
2229 }
2230
2231 // Optimize pointer differences into the same array into a size. Consider:
2232 // &A[10] - &A[0]: we should compile this to "10".
2233 Value *LHSOp, *RHSOp;
2234 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2235 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2236 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2237 I.hasNoUnsignedWrap()))
2238 return replaceInstUsesWith(I, Res);
2239
2240 // trunc(p)-trunc(q) -> trunc(p-q)
2241 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2242 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2243 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2244 /* IsNUW */ false))
2245 return replaceInstUsesWith(I, Res);
2246
2247 // Canonicalize a shifty way to code absolute value to the common pattern.
2248 // There are 2 potential commuted variants.
2249 // We're relying on the fact that we only do this transform when the shift has
2250 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2251 // instructions).
2252 Value *A;
2253 const APInt *ShAmt;
2254 Type *Ty = I.getType();
2255 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2256 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2257 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2258 // B = ashr i32 A, 31 ; smear the sign bit
2259 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2260 // --> (A < 0) ? -A : A
2261 Value *IsNeg = Builder.CreateIsNeg(A);
2262 // Copy the nuw/nsw flags from the sub to the negate.
2263 Value *NegA = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2264 I.hasNoSignedWrap());
2265 return SelectInst::Create(IsNeg, NegA, A);
2266 }
2267
2268 // If we are subtracting a low-bit masked subset of some value from an add
2269 // of that same value with no low bits changed, that is clearing some low bits
2270 // of the sum:
2271 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2272 const APInt *AddC, *AndC;
2273 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2274 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2275 unsigned BitWidth = Ty->getScalarSizeInBits();
2276 unsigned Cttz = AddC->countTrailingZeros();
2277 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2278 if ((HighMask & *AndC).isZero())
2279 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2280 }
2281
2282 if (Instruction *V =
2283 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2284 return V;
2285
2286 // X - usub.sat(X, Y) => umin(X, Y)
2287 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2288 m_Value(Y)))))
2289 return replaceInstUsesWith(
2290 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2291
2292 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2293 // TODO: The one-use restriction is not strictly necessary, but it may
2294 // require improving other pattern matching and/or codegen.
2295 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2296 return replaceInstUsesWith(
2297 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2298
2299 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2300 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2301 return replaceInstUsesWith(
2302 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2303
2304 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2305 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2306 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2307 return BinaryOperator::CreateNeg(USub);
2308 }
2309
2310 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2311 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2312 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2313 return BinaryOperator::CreateNeg(USub);
2314 }
2315
2316 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2317 if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
2318 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2319 return replaceInstUsesWith(
2320 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2321 {Builder.CreateNot(X)}));
2322
2323 return TryToNarrowDeduceFlags();
2324}
2325
2326/// This eliminates floating-point negation in either 'fneg(X)' or
2327/// 'fsub(-0.0, X)' form by combining into a constant operand.
2328static Instruction *foldFNegIntoConstant(Instruction &I, const DataLayout &DL) {
2329 // This is limited with one-use because fneg is assumed better for
2330 // reassociation and cheaper in codegen than fmul/fdiv.
2331 // TODO: Should the m_OneUse restriction be removed?
2332 Instruction *FNegOp;
2333 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2334 return nullptr;
2335
2336 Value *X;
2337 Constant *C;
2338
2339 // Fold negation into constant operand.
2340 // -(X * C) --> X * (-C)
2341 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2342 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2343 return BinaryOperator::CreateFMulFMF(X, NegC, &I);
2344 // -(X / C) --> X / (-C)
2345 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2346 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2347 return BinaryOperator::CreateFDivFMF(X, NegC, &I);
2348 // -(C / X) --> (-C) / X
2349 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X))))
2350 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
2351 Instruction *FDiv = BinaryOperator::CreateFDivFMF(NegC, X, &I);
2352
2353 // Intersect 'nsz' and 'ninf' because those special value exceptions may
2354 // not apply to the fdiv. Everything else propagates from the fneg.
2355 // TODO: We could propagate nsz/ninf from fdiv alone?
2356 FastMathFlags FMF = I.getFastMathFlags();
2357 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2358 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2359 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2360 return FDiv;
2361 }
2362 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2363 // -(X + C) --> -X + -C --> -C - X
2364 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2365 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2366 return BinaryOperator::CreateFSubFMF(NegC, X, &I);
2367
2368 return nullptr;
2369}
2370
2371static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2372 InstCombiner::BuilderTy &Builder) {
2373 Value *FNeg;
2374 if (!match(&I, m_FNeg(m_Value(FNeg))))
2375 return nullptr;
2376
2377 Value *X, *Y;
2378 if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2379 return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2380
2381 if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2382 return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2383
2384 return nullptr;
2385}
2386
2387Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2388 Value *Op = I.getOperand(0);
2389
2390 if (Value *V = simplifyFNegInst(Op, I.getFastMathFlags(),
2391 getSimplifyQuery().getWithInstruction(&I)))
2392 return replaceInstUsesWith(I, V);
2393
2394 if (Instruction *X = foldFNegIntoConstant(I, DL))
2395 return X;
2396
2397 Value *X, *Y;
2398
2399 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2400 if (I.hasNoSignedZeros() &&
2401 match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2402 return BinaryOperator::CreateFSubFMF(Y, X, &I);
2403
2404 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2405 return R;
2406
2407 // Try to eliminate fneg if at least 1 arm of the select is negated.
2408 Value *Cond;
2409 if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
2410 // Unlike most transforms, this one is not safe to propagate nsz unless
2411 // it is present on the original select. (We are conservatively intersecting
2412 // the nsz flags from the select and root fneg instruction.)
2413 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2414 S->copyFastMathFlags(&I);
2415 if (auto *OldSel = dyn_cast<SelectInst>(Op))
2416 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2417 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2418 S->setHasNoSignedZeros(false);
2419 };
2420 // -(Cond ? -P : Y) --> Cond ? P : -Y
2421 Value *P;
2422 if (match(X, m_FNeg(m_Value(P)))) {
2423 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2424 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2425 propagateSelectFMF(NewSel, P == Y);
2426 return NewSel;
2427 }
2428 // -(Cond ? X : -P) --> Cond ? -X : P
2429 if (match(Y, m_FNeg(m_Value(P)))) {
2430 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2431 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2432 propagateSelectFMF(NewSel, P == X);
2433 return NewSel;
2434 }
2435 }
2436
2437 return nullptr;
2438}
2439
2440Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2441 if (Value *V = simplifyFSubInst(I.getOperand(0), I.getOperand(1),
2442 I.getFastMathFlags(),
2443 getSimplifyQuery().getWithInstruction(&I)))
2444 return replaceInstUsesWith(I, V);
2445
2446 if (Instruction *X = foldVectorBinop(I))
2447 return X;
2448
2449 if (Instruction *Phi = foldBinopWithPhiOperands(I))
2450 return Phi;
2451
2452 // Subtraction from -0.0 is the canonical form of fneg.
2453 // fsub -0.0, X ==> fneg X
2454 // fsub nsz 0.0, X ==> fneg nsz X
2455 //
2456 // FIXME This matcher does not respect FTZ or DAZ yet:
2457 // fsub -0.0, Denorm ==> +-0
2458 // fneg Denorm ==> -Denorm
2459 Value *Op;
2460 if (match(&I, m_FNeg(m_Value(Op))))
2461 return UnaryOperator::CreateFNegFMF(Op, &I);
2462
2463 if (Instruction *X = foldFNegIntoConstant(I, DL))
2464 return X;
2465
2466 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2467 return R;
2468
2469 Value *X, *Y;
2470 Constant *C;
2471
2472 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2473 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2474 // Canonicalize to fadd to make analysis easier.
2475 // This can also help codegen because fadd is commutative.
2476 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2477 // killed later. We still limit that particular transform with 'hasOneUse'
2478 // because an fneg is assumed better/cheaper than a generic fsub.
2479 if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2480 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2481 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2482 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2483 }
2484 }
2485
2486 // (-X) - Op1 --> -(X + Op1)
2487 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2488 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2489 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2490 return UnaryOperator::CreateFNegFMF(FAdd, &I);
2491 }
2492
2493 if (isa<Constant>(Op0))
2494 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2495 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2496 return NV;
2497
2498 // X - C --> X + (-C)
2499 // But don't transform constant expressions because there's an inverse fold
2500 // for X + (-Y) --> X - Y.
2501 if (match(Op1, m_ImmConstant(C)))
2502 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
2503 return BinaryOperator::CreateFAddFMF(Op0, NegC, &I);
2504
2505 // X - (-Y) --> X + Y
2506 if (match(Op1, m_FNeg(m_Value(Y))))
2507 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2508
2509 // Similar to above, but look through a cast of the negated value:
2510 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2511 Type *Ty = I.getType();
2512 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2513 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2514
2515 // X - (fpext(-Y)) --> X + fpext(Y)
2516 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2517 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2518
2519 // Similar to above, but look through fmul/fdiv of the negated value:
2520 // Op0 - (-X * Y) --> Op0 + (X * Y)
2521 // Op0 - (Y * -X) --> Op0 + (X * Y)
2522 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2523 Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2524 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2525 }
2526 // Op0 - (-X / Y) --> Op0 + (X / Y)
2527 // Op0 - (X / -Y) --> Op0 + (X / Y)
2528 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2529 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2530 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2531 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2532 }
2533
2534 // Handle special cases for FSub with selects feeding the operation
2535 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2536 return replaceInstUsesWith(I, V);
2537
2538 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2539 // (Y - X) - Y --> -X
2540 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2541 return UnaryOperator::CreateFNegFMF(X, &I);
2542
2543 // Y - (X + Y) --> -X
2544 // Y - (Y + X) --> -X
2545 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2546 return UnaryOperator::CreateFNegFMF(X, &I);
2547
2548 // (X * C) - X --> X * (C - 1.0)
2549 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2550 if (Constant *CSubOne = ConstantFoldBinaryOpOperands(
2551 Instruction::FSub, C, ConstantFP::get(Ty, 1.0), DL))
2552 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2553 }
2554 // X - (X * C) --> X * (1.0 - C)
2555 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2556 if (Constant *OneSubC = ConstantFoldBinaryOpOperands(
2557 Instruction::FSub, ConstantFP::get(Ty, 1.0), C, DL))
2558 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2559 }
2560
2561 // Reassociate fsub/fadd sequences to create more fadd instructions and
2562 // reduce dependency chains:
2563 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2564 Value *Z;
2565 if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2566 m_Value(Z))))) {
2567 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2568 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2569 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2570 }
2571
2572 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2573 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2574 m_Value(Vec)));
2575 };
2576 Value *A0, *A1, *V0, *V1;
2577 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2578 V0->getType() == V1->getType()) {
2579 // Difference of sums is sum of differences:
2580 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2581 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2582 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2583 {Sub->getType()}, {A0, Sub}, &I);
2584 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2585 }
2586
2587 if (Instruction *F = factorizeFAddFSub(I, Builder))
2588 return F;
2589
2590 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2591 // functionality has been subsumed by simple pattern matching here and in
2592 // InstSimplify. We should let a dedicated reassociation pass handle more
2593 // complex pattern matching and remove this from InstCombine.
2594 if (Value *V = FAddCombine(Builder).simplify(&I))
2595 return replaceInstUsesWith(I, V);
2596
2597 // (X - Y) - Op1 --> X - (Y + Op1)
2598 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2599 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2600 return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
2601 }
2602 }
2603
2604 return nullptr;
2605}

/build/source/llvm/include/llvm/IR/PatternMatch.h

1//===- PatternMatch.h - Match on the LLVM IR --------------------*- 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 provides a simple and efficient mechanism for performing general
10// tree-based pattern matches on the LLVM IR. The power of these routines is
11// that it allows you to write concise patterns that are expressive and easy to
12// understand. The other major advantage of this is that it allows you to
13// trivially capture/bind elements in the pattern to variables. For example,
14// you can do something like this:
15//
16// Value *Exp = ...
17// Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2)
18// if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)),
19// m_And(m_Value(Y), m_ConstantInt(C2))))) {
20// ... Pattern is matched and variables are bound ...
21// }
22//
23// This is primarily useful to things like the instruction combiner, but can
24// also be useful for static analysis tools or code generators.
25//
26//===----------------------------------------------------------------------===//
27
28#ifndef LLVM_IR_PATTERNMATCH_H
29#define LLVM_IR_PATTERNMATCH_H
30
31#include "llvm/ADT/APFloat.h"
32#include "llvm/ADT/APInt.h"
33#include "llvm/IR/Constant.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DataLayout.h"
36#include "llvm/IR/InstrTypes.h"
37#include "llvm/IR/Instruction.h"
38#include "llvm/IR/Instructions.h"
39#include "llvm/IR/IntrinsicInst.h"
40#include "llvm/IR/Intrinsics.h"
41#include "llvm/IR/Operator.h"
42#include "llvm/IR/Value.h"
43#include "llvm/Support/Casting.h"
44#include <cstdint>
45
46namespace llvm {
47namespace PatternMatch {
48
49template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) {
50 return const_cast<Pattern &>(P).match(V);
14
Calling 'AnyBinaryOp_match::match'
34
Returning from 'AnyBinaryOp_match::match'
57
Returning without writing to 'P.R.VR'
63
Passing null pointer value via 1st parameter 'V'
64
Calling 'ThreeOps_match::match'
51}
52
53template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) {
54 return const_cast<Pattern &>(P).match(Mask);
55}
56
57template <typename SubPattern_t> struct OneUse_match {
58 SubPattern_t SubPattern;
59
60 OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {}
61
62 template <typename OpTy> bool match(OpTy *V) {
63 return V->hasOneUse() && SubPattern.match(V);
64 }
65};
66
67template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) {
68 return SubPattern;
69}
70
71template <typename Class> struct class_match {
72 template <typename ITy> bool match(ITy *V) { return isa<Class>(V); }
73};
74
75/// Match an arbitrary value and ignore it.
76inline class_match<Value> m_Value() { return class_match<Value>(); }
77
78/// Match an arbitrary unary operation and ignore it.
79inline class_match<UnaryOperator> m_UnOp() {
80 return class_match<UnaryOperator>();
81}
82
83/// Match an arbitrary binary operation and ignore it.
84inline class_match<BinaryOperator> m_BinOp() {
85 return class_match<BinaryOperator>();
86}
87
88/// Matches any compare instruction and ignore it.
89inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); }
90
91struct undef_match {
92 static bool check(const Value *V) {
93 if (isa<UndefValue>(V))
94 return true;
95
96 const auto *CA = dyn_cast<ConstantAggregate>(V);
97 if (!CA)
98 return false;
99
100 SmallPtrSet<const ConstantAggregate *, 8> Seen;
101 SmallVector<const ConstantAggregate *, 8> Worklist;
102
103 // Either UndefValue, PoisonValue, or an aggregate that only contains
104 // these is accepted by matcher.
105 // CheckValue returns false if CA cannot satisfy this constraint.
106 auto CheckValue = [&](const ConstantAggregate *CA) {
107 for (const Value *Op : CA->operand_values()) {
108 if (isa<UndefValue>(Op))
109 continue;
110
111 const auto *CA = dyn_cast<ConstantAggregate>(Op);
112 if (!CA)
113 return false;
114 if (Seen.insert(CA).second)
115 Worklist.emplace_back(CA);
116 }
117
118 return true;
119 };
120
121 if (!CheckValue(CA))
122 return false;
123
124 while (!Worklist.empty()) {
125 if (!CheckValue(Worklist.pop_back_val()))
126 return false;
127 }
128 return true;
129 }
130 template <typename ITy> bool match(ITy *V) { return check(V); }
131};
132
133/// Match an arbitrary undef constant. This matches poison as well.
134/// If this is an aggregate and contains a non-aggregate element that is
135/// neither undef nor poison, the aggregate is not matched.
136inline auto m_Undef() { return undef_match(); }
137
138/// Match an arbitrary poison constant.
139inline class_match<PoisonValue> m_Poison() {
140 return class_match<PoisonValue>();
141}
142
143/// Match an arbitrary Constant and ignore it.
144inline class_match<Constant> m_Constant() { return class_match<Constant>(); }
145
146/// Match an arbitrary ConstantInt and ignore it.
147inline class_match<ConstantInt> m_ConstantInt() {
148 return class_match<ConstantInt>();
149}
150
151/// Match an arbitrary ConstantFP and ignore it.
152inline class_match<ConstantFP> m_ConstantFP() {
153 return class_match<ConstantFP>();
154}
155
156struct constantexpr_match {
157 template <typename ITy> bool match(ITy *V) {
158 auto *C = dyn_cast<Constant>(V);
159 return C && (isa<ConstantExpr>(C) || C->containsConstantExpression());
160 }
161};
162
163/// Match a constant expression or a constant that contains a constant
164/// expression.
165inline constantexpr_match m_ConstantExpr() { return constantexpr_match(); }
166
167/// Match an arbitrary basic block value and ignore it.
168inline class_match<BasicBlock> m_BasicBlock() {
169 return class_match<BasicBlock>();
170}
171
172/// Inverting matcher
173template <typename Ty> struct match_unless {
174 Ty M;
175
176 match_unless(const Ty &Matcher) : M(Matcher) {}
177
178 template <typename ITy> bool match(ITy *V) { return !M.match(V); }
179};
180
181/// Match if the inner matcher does *NOT* match.
182template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) {
183 return match_unless<Ty>(M);
184}
185
186/// Matching combinators
187template <typename LTy, typename RTy> struct match_combine_or {
188 LTy L;
189 RTy R;
190
191 match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
192
193 template <typename ITy> bool match(ITy *V) {
194 if (L.match(V))
17
Taking false branch
195 return true;
196 if (R.match(V))
18
Calling 'match_combine_and::match'
31
Returning from 'match_combine_and::match'
32
Taking true branch
197 return true;
198 return false;
199 }
200};
201
202template <typename LTy, typename RTy> struct match_combine_and {
203 LTy L;
204 RTy R;
205
206 match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
207
208 template <typename ITy> bool match(ITy *V) {
209 if (L.match(V))
19
Calling 'BinaryOp_match::match'
28
Returning from 'BinaryOp_match::match'
29
Taking true branch
210 if (R.match(V))
30
Taking true branch
211 return true;
212 return false;
213 }
214};
215
216/// Combine two pattern matchers matching L || R
217template <typename LTy, typename RTy>
218inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) {
219 return match_combine_or<LTy, RTy>(L, R);
52
Returning without writing to 'L.Op.VR'
220}
221
222/// Combine two pattern matchers matching L && R
223template <typename LTy, typename RTy>
224inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) {
225 return match_combine_and<LTy, RTy>(L, R);
226}
227
228struct apint_match {
229 const APInt *&Res;
230 bool AllowUndef;
231
232 apint_match(const APInt *&Res, bool AllowUndef)
233 : Res(Res), AllowUndef(AllowUndef) {}
234
235 template <typename ITy> bool match(ITy *V) {
236 if (auto *CI = dyn_cast<ConstantInt>(V)) {
237 Res = &CI->getValue();
238 return true;
239 }
240 if (V->getType()->isVectorTy())
241 if (const auto *C = dyn_cast<Constant>(V))
242 if (auto *CI =
243 dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndef))) {
244 Res = &CI->getValue();
245 return true;
246 }
247 return false;
248 }
249};
250// Either constexpr if or renaming ConstantFP::getValueAPF to
251// ConstantFP::getValue is needed to do it via single template
252// function for both apint/apfloat.
253struct apfloat_match {
254 const APFloat *&Res;
255 bool AllowUndef;
256
257 apfloat_match(const APFloat *&Res, bool AllowUndef)
258 : Res(Res), AllowUndef(AllowUndef) {}
259
260 template <typename ITy> bool match(ITy *V) {
261 if (auto *CI = dyn_cast<ConstantFP>(V)) {
262 Res = &CI->getValueAPF();
263 return true;
264 }
265 if (V->getType()->isVectorTy())
266 if (const auto *C = dyn_cast<Constant>(V))
267 if (auto *CI =
268 dyn_cast_or_null<ConstantFP>(C->getSplatValue(AllowUndef))) {
269 Res = &CI->getValueAPF();
270 return true;
271 }
272 return false;
273 }
274};
275
276/// Match a ConstantInt or splatted ConstantVector, binding the
277/// specified pointer to the contained APInt.
278inline apint_match m_APInt(const APInt *&Res) {
279 // Forbid undefs by default to maintain previous behavior.
280 return apint_match(Res, /* AllowUndef */ false);
281}
282
283/// Match APInt while allowing undefs in splat vector constants.
284inline apint_match m_APIntAllowUndef(const APInt *&Res) {
285 return apint_match(Res, /* AllowUndef */ true);
286}
287
288/// Match APInt while forbidding undefs in splat vector constants.
289inline apint_match m_APIntForbidUndef(const APInt *&Res) {
290 return apint_match(Res, /* AllowUndef */ false);
291}
292
293/// Match a ConstantFP or splatted ConstantVector, binding the
294/// specified pointer to the contained APFloat.
295inline apfloat_match m_APFloat(const APFloat *&Res) {
296 // Forbid undefs by default to maintain previous behavior.
297 return apfloat_match(Res, /* AllowUndef */ false);
298}
299
300/// Match APFloat while allowing undefs in splat vector constants.
301inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) {
302 return apfloat_match(Res, /* AllowUndef */ true);
303}
304
305/// Match APFloat while forbidding undefs in splat vector constants.
306inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) {
307 return apfloat_match(Res, /* AllowUndef */ false);
308}
309
310template <int64_t Val> struct constantint_match {
311 template <typename ITy> bool match(ITy *V) {
312 if (const auto *CI = dyn_cast<ConstantInt>(V)) {
313 const APInt &CIV = CI->getValue();
314 if (Val >= 0)
315 return CIV == static_cast<uint64_t>(Val);
316 // If Val is negative, and CI is shorter than it, truncate to the right
317 // number of bits. If it is larger, then we have to sign extend. Just
318 // compare their negated values.
319 return -CIV == -Val;
320 }
321 return false;
322 }
323};
324
325/// Match a ConstantInt with a specific value.
326template <int64_t Val> inline constantint_match<Val> m_ConstantInt() {
327 return constantint_match<Val>();
328}
329
330/// This helper class is used to match constant scalars, vector splats,
331/// and fixed width vectors that satisfy a specified predicate.
332/// For fixed width vector constants, undefined elements are ignored.
333template <typename Predicate, typename ConstantVal>
334struct cstval_pred_ty : public Predicate {
335 template <typename ITy> bool match(ITy *V) {
336 if (const auto *CV = dyn_cast<ConstantVal>(V))
337 return this->isValue(CV->getValue());
338 if (const auto *VTy = dyn_cast<VectorType>(V->getType())) {
339 if (const auto *C = dyn_cast<Constant>(V)) {
340 if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue()))
341 return this->isValue(CV->getValue());
342
343 // Number of elements of a scalable vector unknown at compile time
344 auto *FVTy = dyn_cast<FixedVectorType>(VTy);
345 if (!FVTy)
346 return false;
347
348 // Non-splat vector constant: check each element for a match.
349 unsigned NumElts = FVTy->getNumElements();
350 assert(NumElts != 0 && "Constant vector with no elements?")(static_cast <bool> (NumElts != 0 && "Constant vector with no elements?"
) ? void (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\""
, "llvm/include/llvm/IR/PatternMatch.h", 350, __extension__ __PRETTY_FUNCTION__
))
;
351 bool HasNonUndefElements = false;
352 for (unsigned i = 0; i != NumElts; ++i) {
353 Constant *Elt = C->getAggregateElement(i);
354 if (!Elt)
355 return false;
356 if (isa<UndefValue>(Elt))
357 continue;
358 auto *CV = dyn_cast<ConstantVal>(Elt);
359 if (!CV || !this->isValue(CV->getValue()))
360 return false;
361 HasNonUndefElements = true;
362 }
363 return HasNonUndefElements;
364 }
365 }
366 return false;
367 }
368};
369
370/// specialization of cstval_pred_ty for ConstantInt
371template <typename Predicate>
372using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>;
373
374/// specialization of cstval_pred_ty for ConstantFP
375template <typename Predicate>
376using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>;
377
378/// This helper class is used to match scalar and vector constants that
379/// satisfy a specified predicate, and bind them to an APInt.
380template <typename Predicate> struct api_pred_ty : public Predicate {
381 const APInt *&Res;
382
383 api_pred_ty(const APInt *&R) : Res(R) {}
384
385 template <typename ITy> bool match(ITy *V) {
386 if (const auto *CI = dyn_cast<ConstantInt>(V))
387 if (this->isValue(CI->getValue())) {
388 Res = &CI->getValue();
389 return true;
390 }
391 if (V->getType()->isVectorTy())
392 if (const auto *C = dyn_cast<Constant>(V))
393 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
394 if (this->isValue(CI->getValue())) {
395 Res = &CI->getValue();
396 return true;
397 }
398
399 return false;
400 }
401};
402
403/// This helper class is used to match scalar and vector constants that
404/// satisfy a specified predicate, and bind them to an APFloat.
405/// Undefs are allowed in splat vector constants.
406template <typename Predicate> struct apf_pred_ty : public Predicate {
407 const APFloat *&Res;
408
409 apf_pred_ty(const APFloat *&R) : Res(R) {}
410
411 template <typename ITy> bool match(ITy *V) {
412 if (const auto *CI = dyn_cast<ConstantFP>(V))
413 if (this->isValue(CI->getValue())) {
414 Res = &CI->getValue();
415 return true;
416 }
417 if (V->getType()->isVectorTy())
418 if (const auto *C = dyn_cast<Constant>(V))
419 if (auto *CI = dyn_cast_or_null<ConstantFP>(
420 C->getSplatValue(/* AllowUndef */ true)))
421 if (this->isValue(CI->getValue())) {
422 Res = &CI->getValue();
423 return true;
424 }
425
426 return false;
427 }
428};
429
430///////////////////////////////////////////////////////////////////////////////
431//
432// Encapsulate constant value queries for use in templated predicate matchers.
433// This allows checking if constants match using compound predicates and works
434// with vector constants, possibly with relaxed constraints. For example, ignore
435// undef values.
436//
437///////////////////////////////////////////////////////////////////////////////
438
439struct is_any_apint {
440 bool isValue(const APInt &C) { return true; }
441};
442/// Match an integer or vector with any integral constant.
443/// For vectors, this includes constants with undefined elements.
444inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() {
445 return cst_pred_ty<is_any_apint>();
446}
447
448struct is_all_ones {
449 bool isValue(const APInt &C) { return C.isAllOnes(); }
450};
451/// Match an integer or vector with all bits set.
452/// For vectors, this includes constants with undefined elements.
453inline cst_pred_ty<is_all_ones> m_AllOnes() {
454 return cst_pred_ty<is_all_ones>();
455}
456
457struct is_maxsignedvalue {
458 bool isValue(const APInt &C) { return C.isMaxSignedValue(); }
459};
460/// Match an integer or vector with values having all bits except for the high
461/// bit set (0x7f...).
462/// For vectors, this includes constants with undefined elements.
463inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() {
464 return cst_pred_ty<is_maxsignedvalue>();
465}
466inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) {
467 return V;
468}
469
470struct is_negative {
471 bool isValue(const APInt &C) { return C.isNegative(); }
472};
473/// Match an integer or vector of negative values.
474/// For vectors, this includes constants with undefined elements.
475inline cst_pred_ty<is_negative> m_Negative() {
476 return cst_pred_ty<is_negative>();
477}
478inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { return V; }
479
480struct is_nonnegative {
481 bool isValue(const APInt &C) { return C.isNonNegative(); }
482};
483/// Match an integer or vector of non-negative values.
484/// For vectors, this includes constants with undefined elements.
485inline cst_pred_ty<is_nonnegative> m_NonNegative() {
486 return cst_pred_ty<is_nonnegative>();
487}
488inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { return V; }
489
490struct is_strictlypositive {
491 bool isValue(const APInt &C) { return C.isStrictlyPositive(); }
492};
493/// Match an integer or vector of strictly positive values.
494/// For vectors, this includes constants with undefined elements.
495inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() {
496 return cst_pred_ty<is_strictlypositive>();
497}
498inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) {
499 return V;
500}
501
502struct is_nonpositive {
503 bool isValue(const APInt &C) { return C.isNonPositive(); }
504};
505/// Match an integer or vector of non-positive values.
506/// For vectors, this includes constants with undefined elements.
507inline cst_pred_ty<is_nonpositive> m_NonPositive() {
508 return cst_pred_ty<is_nonpositive>();
509}
510inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; }
511
512struct is_one {
513 bool isValue(const APInt &C) { return C.isOne(); }
514};
515/// Match an integer 1 or a vector with all elements equal to 1.
516/// For vectors, this includes constants with undefined elements.
517inline cst_pred_ty<is_one> m_One() { return cst_pred_ty<is_one>(); }
518
519struct is_zero_int {
520 bool isValue(const APInt &C) { return C.isZero(); }
521};
522/// Match an integer 0 or a vector with all elements equal to 0.
523/// For vectors, this includes constants with undefined elements.
524inline cst_pred_ty<is_zero_int> m_ZeroInt() {
525 return cst_pred_ty<is_zero_int>();
526}
527
528struct is_zero {
529 template <typename ITy> bool match(ITy *V) {
530 auto *C = dyn_cast<Constant>(V);
531 // FIXME: this should be able to do something for scalable vectors
532 return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C));
533 }
534};
535/// Match any null constant or a vector with all elements equal to 0.
536/// For vectors, this includes constants with undefined elements.
537inline is_zero m_Zero() { return is_zero(); }
538
539struct is_power2 {
540 bool isValue(const APInt &C) { return C.isPowerOf2(); }
541};
542/// Match an integer or vector power-of-2.
543/// For vectors, this includes constants with undefined elements.
544inline cst_pred_ty<is_power2> m_Power2() { return cst_pred_ty<is_power2>(); }
545inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { return V; }
546
547struct is_negated_power2 {
548 bool isValue(const APInt &C) { return C.isNegatedPowerOf2(); }
549};
550/// Match a integer or vector negated power-of-2.
551/// For vectors, this includes constants with undefined elements.
552inline cst_pred_ty<is_negated_power2> m_NegatedPower2() {
553 return cst_pred_ty<is_negated_power2>();
554}
555inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) {
556 return V;
557}
558
559struct is_power2_or_zero {
560 bool isValue(const APInt &C) { return !C || C.isPowerOf2(); }
561};
562/// Match an integer or vector of 0 or power-of-2 values.
563/// For vectors, this includes constants with undefined elements.
564inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() {
565 return cst_pred_ty<is_power2_or_zero>();
566}
567inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) {
568 return V;
569}
570
571struct is_sign_mask {
572 bool isValue(const APInt &C) { return C.isSignMask(); }
573};
574/// Match an integer or vector with only the sign bit(s) set.
575/// For vectors, this includes constants with undefined elements.
576inline cst_pred_ty<is_sign_mask> m_SignMask() {
577 return cst_pred_ty<is_sign_mask>();
578}
579
580struct is_lowbit_mask {
581 bool isValue(const APInt &C) { return C.isMask(); }
582};
583/// Match an integer or vector with only the low bit(s) set.
584/// For vectors, this includes constants with undefined elements.
585inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() {
586 return cst_pred_ty<is_lowbit_mask>();
587}
588inline api_pred_ty<is_lowbit_mask> m_LowBitMask(const APInt *&V) { return V; }
589
590struct icmp_pred_with_threshold {
591 ICmpInst::Predicate Pred;
592 const APInt *Thr;
593 bool isValue(const APInt &C) { return ICmpInst::compare(C, *Thr, Pred); }
594};
595/// Match an integer or vector with every element comparing 'pred' (eg/ne/...)
596/// to Threshold. For vectors, this includes constants with undefined elements.
597inline cst_pred_ty<icmp_pred_with_threshold>
598m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) {
599 cst_pred_ty<icmp_pred_with_threshold> P;
600 P.Pred = Predicate;
601 P.Thr = &Threshold;
602 return P;
603}
604
605struct is_nan {
606 bool isValue(const APFloat &C) { return C.isNaN(); }
607};
608/// Match an arbitrary NaN constant. This includes quiet and signalling nans.
609/// For vectors, this includes constants with undefined elements.
610inline cstfp_pred_ty<is_nan> m_NaN() { return cstfp_pred_ty<is_nan>(); }
611
612struct is_nonnan {
613 bool isValue(const APFloat &C) { return !C.isNaN(); }
614};
615/// Match a non-NaN FP constant.
616/// For vectors, this includes constants with undefined elements.
617inline cstfp_pred_ty<is_nonnan> m_NonNaN() {
618 return cstfp_pred_ty<is_nonnan>();
619}
620
621struct is_inf {
622 bool isValue(const APFloat &C) { return C.isInfinity(); }
623};
624/// Match a positive or negative infinity FP constant.
625/// For vectors, this includes constants with undefined elements.
626inline cstfp_pred_ty<is_inf> m_Inf() { return cstfp_pred_ty<is_inf>(); }
627
628struct is_noninf {
629 bool isValue(const APFloat &C) { return !C.isInfinity(); }
630};
631/// Match a non-infinity FP constant, i.e. finite or NaN.
632/// For vectors, this includes constants with undefined elements.
633inline cstfp_pred_ty<is_noninf> m_NonInf() {
634 return cstfp_pred_ty<is_noninf>();
635}
636
637struct is_finite {
638 bool isValue(const APFloat &C) { return C.isFinite(); }
639};
640/// Match a finite FP constant, i.e. not infinity or NaN.
641/// For vectors, this includes constants with undefined elements.
642inline cstfp_pred_ty<is_finite> m_Finite() {
643 return cstfp_pred_ty<is_finite>();
644}
645inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; }
646
647struct is_finitenonzero {
648 bool isValue(const APFloat &C) { return C.isFiniteNonZero(); }
649};
650/// Match a finite non-zero FP constant.
651/// For vectors, this includes constants with undefined elements.
652inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() {
653 return cstfp_pred_ty<is_finitenonzero>();
654}
655inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) {
656 return V;
657}
658
659struct is_any_zero_fp {
660 bool isValue(const APFloat &C) { return C.isZero(); }
661};
662/// Match a floating-point negative zero or positive zero.
663/// For vectors, this includes constants with undefined elements.
664inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() {
665 return cstfp_pred_ty<is_any_zero_fp>();
666}
667
668struct is_pos_zero_fp {
669 bool isValue(const APFloat &C) { return C.isPosZero(); }
670};
671/// Match a floating-point positive zero.
672/// For vectors, this includes constants with undefined elements.
673inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() {
674 return cstfp_pred_ty<is_pos_zero_fp>();
675}
676
677struct is_neg_zero_fp {
678 bool isValue(const APFloat &C) { return C.isNegZero(); }
679};
680/// Match a floating-point negative zero.
681/// For vectors, this includes constants with undefined elements.
682inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() {
683 return cstfp_pred_ty<is_neg_zero_fp>();
684}
685
686struct is_non_zero_fp {
687 bool isValue(const APFloat &C) { return C.isNonZero(); }
688};
689/// Match a floating-point non-zero.
690/// For vectors, this includes constants with undefined elements.
691inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() {
692 return cstfp_pred_ty<is_non_zero_fp>();
693}
694
695///////////////////////////////////////////////////////////////////////////////
696
697template <typename Class> struct bind_ty {
698 Class *&VR;
699
700 bind_ty(Class *&V) : VR(V) {}
701
702 template <typename ITy> bool match(ITy *V) {
703 if (auto *CV = dyn_cast<Class>(V)) {
704 VR = CV;
705 return true;
706 }
707 return false;
708 }
709};
710
711/// Match a value, capturing it if we match.
712inline bind_ty<Value> m_Value(Value *&V) { return V; }
6
Calling constructor for 'bind_ty<llvm::Value>'
7
Returning from constructor for 'bind_ty<llvm::Value>'
8
Returning without writing to 'V'
45
Returning without writing to 'V'
713inline bind_ty<const Value> m_Value(const Value *&V) { return V; }
714
715/// Match an instruction, capturing it if we match.
716inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; }
717/// Match a unary operator, capturing it if we match.
718inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; }
719/// Match a binary operator, capturing it if we match.
720inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; }
721/// Match a with overflow intrinsic, capturing it if we match.
722inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) {
723 return I;
724}
725inline bind_ty<const WithOverflowInst>
726m_WithOverflowInst(const WithOverflowInst *&I) {
727 return I;
728}
729
730/// Match a Constant, capturing the value if we match.
731inline bind_ty<Constant> m_Constant(Constant *&C) { return C; }
732
733/// Match a ConstantInt, capturing the value if we match.
734inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; }
735
736/// Match a ConstantFP, capturing the value if we match.
737inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; }
738
739/// Match a ConstantExpr, capturing the value if we match.
740inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; }
741
742/// Match a basic block value, capturing it if we match.
743inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; }
744inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) {
745 return V;
746}
747
748/// Match an arbitrary immediate Constant and ignore it.
749inline match_combine_and<class_match<Constant>,
750 match_unless<constantexpr_match>>
751m_ImmConstant() {
752 return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr()));
753}
754
755/// Match an immediate Constant, capturing the value if we match.
756inline match_combine_and<bind_ty<Constant>,
757 match_unless<constantexpr_match>>
758m_ImmConstant(Constant *&C) {
759 return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr()));
760}
761
762/// Match a specified Value*.
763struct specificval_ty {
764 const Value *Val;
765
766 specificval_ty(const Value *V) : Val(V) {}
767
768 template <typename ITy> bool match(ITy *V) { return V == Val; }
769};
770
771/// Match if we have a specific specified value.
772inline specificval_ty m_Specific(const Value *V) { return V; }
773
774/// Stores a reference to the Value *, not the Value * itself,
775/// thus can be used in commutative matchers.
776template <typename Class> struct deferredval_ty {
777 Class *const &Val;
778
779 deferredval_ty(Class *const &V) : Val(V) {}
780
781 template <typename ITy> bool match(ITy *const V) { return V == Val; }
782};
783
784/// Like m_Specific(), but works if the specific value to match is determined
785/// as part of the same match() expression. For example:
786/// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will
787/// bind X before the pattern match starts.
788/// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against
789/// whichever value m_Value(X) populated.
790inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; }
791inline deferredval_ty<const Value> m_Deferred(const Value *const &V) {
792 return V;
793}
794
795/// Match a specified floating point value or vector of all elements of
796/// that value.
797struct specific_fpval {
798 double Val;
799
800 specific_fpval(double V) : Val(V) {}
801
802 template <typename ITy> bool match(ITy *V) {
803 if (const auto *CFP = dyn_cast<ConstantFP>(V))
804 return CFP->isExactlyValue(Val);
805 if (V->getType()->isVectorTy())
806 if (const auto *C = dyn_cast<Constant>(V))
807 if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
808 return CFP->isExactlyValue(Val);
809 return false;
810 }
811};
812
813/// Match a specific floating point value or vector with all elements
814/// equal to the value.
815inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); }
816
817/// Match a float 1.0 or vector with all elements equal to 1.0.
818inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); }
819
820struct bind_const_intval_ty {
821 uint64_t &VR;
822
823 bind_const_intval_ty(uint64_t &V) : VR(V) {}
824
825 template <typename ITy> bool match(ITy *V) {
826 if (const auto *CV = dyn_cast<ConstantInt>(V))
827 if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) {
828 VR = CV->getZExtValue();
829 return true;
830 }
831 return false;
832 }
833};
834
835/// Match a specified integer value or vector of all elements of that
836/// value.
837template <bool AllowUndefs> struct specific_intval {
838 APInt Val;
839
840 specific_intval(APInt V) : Val(std::move(V)) {}
841
842 template <typename ITy> bool match(ITy *V) {
843 const auto *CI = dyn_cast<ConstantInt>(V);
844 if (!CI && V->getType()->isVectorTy())
845 if (const auto *C = dyn_cast<Constant>(V))
846 CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs));
847
848 return CI && APInt::isSameValue(CI->getValue(), Val);
849 }
850};
851
852/// Match a specific integer value or vector with all elements equal to
853/// the value.
854inline specific_intval<false> m_SpecificInt(APInt V) {
855 return specific_intval<false>(std::move(V));
856}
857
858inline specific_intval<false> m_SpecificInt(uint64_t V) {
859 return m_SpecificInt(APInt(64, V));
860}
861
862inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) {
863 return specific_intval<true>(std::move(V));
864}
865
866inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) {
867 return m_SpecificIntAllowUndef(APInt(64, V));
868}
869
870/// Match a ConstantInt and bind to its value. This does not match
871/// ConstantInts wider than 64-bits.
872inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; }
873
874/// Match a specified basic block value.
875struct specific_bbval {
876 BasicBlock *Val;
877
878 specific_bbval(BasicBlock *Val) : Val(Val) {}
879
880 template <typename ITy> bool match(ITy *V) {
881 const auto *BB = dyn_cast<BasicBlock>(V);
882 return BB && BB == Val;
883 }
884};
885
886/// Match a specific basic block value.
887inline specific_bbval m_SpecificBB(BasicBlock *BB) {
888 return specific_bbval(BB);
889}
890
891/// A commutative-friendly version of m_Specific().
892inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) {
893 return BB;
894}
895inline deferredval_ty<const BasicBlock>
896m_Deferred(const BasicBlock *const &BB) {
897 return BB;
898}
899
900//===----------------------------------------------------------------------===//
901// Matcher for any binary operator.
902//
903template <typename LHS_t, typename RHS_t, bool Commutable = false>
904struct AnyBinaryOp_match {
905 LHS_t L;
906 RHS_t R;
907
908 // The evaluation order is always stable, regardless of Commutability.
909 // The LHS is always matched first.
910 AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
911
912 template <typename OpTy> bool match(OpTy *V) {
913 if (auto *I
15.1
'I' is non-null
15.1
'I' is non-null
= dyn_cast<BinaryOperator>(V))
15
Assuming 'V' is a 'CastReturnType'
914 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
16
Calling 'match_combine_or::match'
33
Returning from 'match_combine_or::match'
915 (Commutable && L.match(I->getOperand(1)) &&
916 R.match(I->getOperand(0)));
917 return false;
918 }
919};
920
921template <typename LHS, typename RHS>
922inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) {
923 return AnyBinaryOp_match<LHS, RHS>(L, R);
924}
925
926//===----------------------------------------------------------------------===//
927// Matcher for any unary operator.
928// TODO fuse unary, binary matcher into n-ary matcher
929//
930template <typename OP_t> struct AnyUnaryOp_match {
931 OP_t X;
932
933 AnyUnaryOp_match(const OP_t &X) : X(X) {}
934
935 template <typename OpTy> bool match(OpTy *V) {
936 if (auto *I = dyn_cast<UnaryOperator>(V))
937 return X.match(I->getOperand(0));
938 return false;
939 }
940};
941
942template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) {
943 return AnyUnaryOp_match<OP_t>(X);
944}
945
946//===----------------------------------------------------------------------===//
947// Matchers for specific binary operators.
948//
949
950template <typename LHS_t, typename RHS_t, unsigned Opcode,
951 bool Commutable = false>
952struct BinaryOp_match {
953 LHS_t L;
954 RHS_t R;
955
956 // The evaluation order is always stable, regardless of Commutability.
957 // The LHS is always matched first.
958 BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
959
960 template <typename OpTy> inline bool match(unsigned Opc, OpTy *V) {
961 if (V->getValueID() == Value::InstructionVal + Opc) {
21
Assuming the condition is true
22
Taking true branch
962 auto *I = cast<BinaryOperator>(V);
23
'V' is a 'CastReturnType'
963 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
24
Value assigned to 'Select'
25
Assuming the condition is true
26
Assuming the condition is true
964 (Commutable && L.match(I->getOperand(1)) &&
965 R.match(I->getOperand(0)));
966 }
967 if (auto *CE = dyn_cast<ConstantExpr>(V))
968 return CE->getOpcode() == Opc &&
969 ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) ||
970 (Commutable && L.match(CE->getOperand(1)) &&
971 R.match(CE->getOperand(0))));
972 return false;
973 }
974
975 template <typename OpTy> bool match(OpTy *V) { return match(Opcode, V); }
20
Calling 'BinaryOp_match::match'
27
Returning from 'BinaryOp_match::match'
976};
977
978template <typename LHS, typename RHS>
979inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L,
980 const RHS &R) {
981 return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R);
982}
983
984template <typename LHS, typename RHS>
985inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L,
986 const RHS &R) {
987 return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R);
988}
989
990template <typename LHS, typename RHS>
991inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L,
992 const RHS &R) {
993 return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R);
994}
995
996template <typename LHS, typename RHS>
997inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L,
998 const RHS &R) {
999 return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R);
1000}
1001
1002template <typename Op_t> struct FNeg_match {
1003 Op_t X;
1004
1005 FNeg_match(const Op_t &Op) : X(Op) {}
1006 template <typename OpTy> bool match(OpTy *V) {
1007 auto *FPMO = dyn_cast<FPMathOperator>(V);
1008 if (!FPMO)
1009 return false;
1010
1011 if (FPMO->getOpcode() == Instruction::FNeg)
1012 return X.match(FPMO->getOperand(0));
1013
1014 if (FPMO->getOpcode() == Instruction::FSub) {
1015 if (FPMO->hasNoSignedZeros()) {
1016 // With 'nsz', any zero goes.
1017 if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0)))
1018 return false;
1019 } else {
1020 // Without 'nsz', we need fsub -0.0, X exactly.
1021 if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0)))
1022 return false;
1023 }
1024
1025 return X.match(FPMO->getOperand(1));
1026 }
1027
1028 return false;
1029 }
1030};
1031
1032/// Match 'fneg X' as 'fsub -0.0, X'.
1033template <typename OpTy> inline FNeg_match<OpTy> m_FNeg(const OpTy &X) {
1034 return FNeg_match<OpTy>(X);
1035}
1036
1037/// Match 'fneg X' as 'fsub +-0.0, X'.
1038template <typename RHS>
1039inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub>
1040m_FNegNSZ(const RHS &X) {
1041 return m_FSub(m_AnyZeroFP(), X);
1042}
1043
1044template <typename LHS, typename RHS>
1045inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L,
1046 const RHS &R) {
1047 return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R);
1048}
1049
1050template <typename LHS, typename RHS>
1051inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L,
1052 const RHS &R) {
1053 return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R);
1054}
1055
1056template <typename LHS, typename RHS>
1057inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L,
1058 const RHS &R) {
1059 return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R);
1060}
1061
1062template <typename LHS, typename RHS>
1063inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L,
1064 const RHS &R) {
1065 return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R);
1066}
1067
1068template <typename LHS, typename RHS>
1069inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L,
1070 const RHS &R) {
1071 return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R);
1072}
1073
1074template <typename LHS, typename RHS>
1075inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L,
1076 const RHS &R) {
1077 return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R);
1078}
1079
1080template <typename LHS, typename RHS>
1081inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L,
1082 const RHS &R) {
1083 return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R);
1084}
1085
1086template <typename LHS, typename RHS>
1087inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L,
1088 const RHS &R) {
1089 return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R);
1090}
1091
1092template <typename LHS, typename RHS>
1093inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L,
1094 const RHS &R) {
1095 return BinaryOp_match<LHS, RHS, Instruction::And>(L, R);
1096}
1097
1098template <typename LHS, typename RHS>
1099inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L,
1100 const RHS &R) {
1101 return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R);
1102}
1103
1104template <typename LHS, typename RHS>
1105inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L,
1106 const RHS &R) {
1107 return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R);
1108}
1109
1110template <typename LHS, typename RHS>
1111inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L,
1112 const RHS &R) {
1113 return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R);
1114}
1115
1116template <typename LHS, typename RHS>
1117inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L,
1118 const RHS &R) {
1119 return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R);
1120}
1121
1122template <typename LHS, typename RHS>
1123inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L,
1124 const RHS &R) {
1125 return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R);
1126}
1127
1128template <typename LHS_t, typename RHS_t, unsigned Opcode,
1129 unsigned WrapFlags = 0>
1130struct OverflowingBinaryOp_match {
1131 LHS_t L;
1132 RHS_t R;
1133
1134 OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS)
1135 : L(LHS), R(RHS) {}
1136
1137 template <typename OpTy> bool match(OpTy *V) {
1138 if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) {
1139 if (Op->getOpcode() != Opcode)
1140 return false;
1141 if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) &&
1142 !Op->hasNoUnsignedWrap())
1143 return false;
1144 if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) &&
1145 !Op->hasNoSignedWrap())
1146 return false;
1147 return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1));
1148 }
1149 return false;
1150 }
1151};
1152
1153template <typename LHS, typename RHS>
1154inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1155 OverflowingBinaryOperator::NoSignedWrap>
1156m_NSWAdd(const LHS &L, const RHS &R) {
1157 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1158 OverflowingBinaryOperator::NoSignedWrap>(L,
1159 R);
1160}
1161template <typename LHS, typename RHS>
1162inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1163 OverflowingBinaryOperator::NoSignedWrap>
1164m_NSWSub(const LHS &L, const RHS &R) {
1165 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1166 OverflowingBinaryOperator::NoSignedWrap>(L,
1167 R);
1168}
1169template <typename LHS, typename RHS>
1170inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1171 OverflowingBinaryOperator::NoSignedWrap>
1172m_NSWMul(const LHS &L, const RHS &R) {
1173 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1174 OverflowingBinaryOperator::NoSignedWrap>(L,
1175 R);
1176}
1177template <typename LHS, typename RHS>
1178inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1179 OverflowingBinaryOperator::NoSignedWrap>
1180m_NSWShl(const LHS &L, const RHS &R) {
1181 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1182 OverflowingBinaryOperator::NoSignedWrap>(L,
1183 R);
1184}
1185
1186template <typename LHS, typename RHS>
1187inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1188 OverflowingBinaryOperator::NoUnsignedWrap>
1189m_NUWAdd(const LHS &L, const RHS &R) {
1190 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1191 OverflowingBinaryOperator::NoUnsignedWrap>(
1192 L, R);
1193}
1194template <typename LHS, typename RHS>
1195inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1196 OverflowingBinaryOperator::NoUnsignedWrap>
1197m_NUWSub(const LHS &L, const RHS &R) {
1198 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1199 OverflowingBinaryOperator::NoUnsignedWrap>(
1200 L, R);
1201}
1202template <typename LHS, typename RHS>
1203inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1204 OverflowingBinaryOperator::NoUnsignedWrap>
1205m_NUWMul(const LHS &L, const RHS &R) {
1206 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1207 OverflowingBinaryOperator::NoUnsignedWrap>(
1208 L, R);
1209}
1210template <typename LHS, typename RHS>
1211inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1212 OverflowingBinaryOperator::NoUnsignedWrap>
1213m_NUWShl(const LHS &L, const RHS &R) {
1214 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1215 OverflowingBinaryOperator::NoUnsignedWrap>(
1216 L, R);
1217}
1218
1219template <typename LHS_t, typename RHS_t, bool Commutable = false>
1220struct SpecificBinaryOp_match
1221 : public BinaryOp_match<LHS_t, RHS_t, 0, Commutable> {
1222 unsigned Opcode;
1223
1224 SpecificBinaryOp_match(unsigned Opcode, const LHS_t &LHS, const RHS_t &RHS)
1225 : BinaryOp_match<LHS_t, RHS_t, 0, Commutable>(LHS, RHS), Opcode(Opcode) {}
1226
1227 template <typename OpTy> bool match(OpTy *V) {
1228 return BinaryOp_match<LHS_t, RHS_t, 0, Commutable>::match(Opcode, V);
1229 }
1230};
1231
1232/// Matches a specific opcode.
1233template <typename LHS, typename RHS>
1234inline SpecificBinaryOp_match<LHS, RHS> m_BinOp(unsigned Opcode, const LHS &L,
1235 const RHS &R) {
1236 return SpecificBinaryOp_match<LHS, RHS>(Opcode, L, R);
1237}
1238
1239//===----------------------------------------------------------------------===//
1240// Class that matches a group of binary opcodes.
1241//
1242template <typename LHS_t, typename RHS_t, typename Predicate>
1243struct BinOpPred_match : Predicate {
1244 LHS_t L;
1245 RHS_t R;
1246
1247 BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1248
1249 template <typename OpTy> bool match(OpTy *V) {
1250 if (auto *I = dyn_cast<Instruction>(V))
1251 return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) &&
1252 R.match(I->getOperand(1));
1253 if (auto *CE = dyn_cast<ConstantExpr>(V))
1254 return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) &&
1255 R.match(CE->getOperand(1));
1256 return false;
1257 }
1258};
1259
1260struct is_shift_op {
1261 bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); }
1262};
1263
1264struct is_right_shift_op {
1265 bool isOpType(unsigned Opcode) {
1266 return Opcode == Instruction::LShr || Opcode == Instruction::AShr;
1267 }
1268};
1269
1270struct is_logical_shift_op {
1271 bool isOpType(unsigned Opcode) {
1272 return Opcode == Instruction::LShr || Opcode == Instruction::Shl;
1273 }
1274};
1275
1276struct is_bitwiselogic_op {
1277 bool isOpType(unsigned Opcode) {
1278 return Instruction::isBitwiseLogicOp(Opcode);
1279 }
1280};
1281
1282struct is_idiv_op {
1283 bool isOpType(unsigned Opcode) {
1284 return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
1285 }
1286};
1287
1288struct is_irem_op {
1289 bool isOpType(unsigned Opcode) {
1290 return Opcode == Instruction::SRem || Opcode == Instruction::URem;
1291 }
1292};
1293
1294/// Matches shift operations.
1295template <typename LHS, typename RHS>
1296inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L,
1297 const RHS &R) {
1298 return BinOpPred_match<LHS, RHS, is_shift_op>(L, R);
1299}
1300
1301/// Matches logical shift operations.
1302template <typename LHS, typename RHS>
1303inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L,
1304 const RHS &R) {
1305 return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R);
1306}
1307
1308/// Matches logical shift operations.
1309template <typename LHS, typename RHS>
1310inline BinOpPred_match<LHS, RHS, is_logical_shift_op>
1311m_LogicalShift(const LHS &L, const RHS &R) {
1312 return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R);
1313}
1314
1315/// Matches bitwise logic operations.
1316template <typename LHS, typename RHS>
1317inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op>
1318m_BitwiseLogic(const LHS &L, const RHS &R) {
1319 return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R);
1320}
1321
1322/// Matches integer division operations.
1323template <typename LHS, typename RHS>
1324inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L,
1325 const RHS &R) {
1326 return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R);
1327}
1328
1329/// Matches integer remainder operations.
1330template <typename LHS, typename RHS>
1331inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L,
1332 const RHS &R) {
1333 return BinOpPred_match<LHS, RHS, is_irem_op>(L, R);
1334}
1335
1336//===----------------------------------------------------------------------===//
1337// Class that matches exact binary ops.
1338//
1339template <typename SubPattern_t> struct Exact_match {
1340 SubPattern_t SubPattern;
1341
1342 Exact_match(const SubPattern_t &SP) : SubPattern(SP) {}
1343
1344 template <typename OpTy> bool match(OpTy *V) {
1345 if (auto *PEO = dyn_cast<PossiblyExactOperator>(V))
1346 return PEO->isExact() && SubPattern.match(V);
1347 return false;
1348 }
1349};
1350
1351template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) {
1352 return SubPattern;
1353}
1354
1355//===----------------------------------------------------------------------===//
1356// Matchers for CmpInst classes
1357//
1358
1359template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy,
1360 bool Commutable = false>
1361struct CmpClass_match {
1362 PredicateTy &Predicate;
1363 LHS_t L;
1364 RHS_t R;
1365
1366 // The evaluation order is always stable, regardless of Commutability.
1367 // The LHS is always matched first.
1368 CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS)
1369 : Predicate(Pred), L(LHS), R(RHS) {}
1370
1371 template <typename OpTy> bool match(OpTy *V) {
1372 if (auto *I = dyn_cast<Class>(V)) {
1373 if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) {
1374 Predicate = I->getPredicate();
1375 return true;
1376 } else if (Commutable && L.match(I->getOperand(1)) &&
1377 R.match(I->getOperand(0))) {
1378 Predicate = I->getSwappedPredicate();
1379 return true;
1380 }
1381 }
1382 return false;
1383 }
1384};
1385
1386template <typename LHS, typename RHS>
1387inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>
1388m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1389 return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R);
1390}
1391
1392template <typename LHS, typename RHS>
1393inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>
1394m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1395 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R);
1396}
1397
1398template <typename LHS, typename RHS>
1399inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>
1400m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1401 return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R);
1402}
1403
1404//===----------------------------------------------------------------------===//
1405// Matchers for instructions with a given opcode and number of operands.
1406//
1407
1408/// Matches instructions with Opcode and three operands.
1409template <typename T0, unsigned Opcode> struct OneOps_match {
1410 T0 Op1;
1411
1412 OneOps_match(const T0 &Op1) : Op1(Op1) {}
1413
1414 template <typename OpTy> bool match(OpTy *V) {
1415 if (V->getValueID() == Value::InstructionVal + Opcode) {
1416 auto *I = cast<Instruction>(V);
1417 return Op1.match(I->getOperand(0));
1418 }
1419 return false;
1420 }
1421};
1422
1423/// Matches instructions with Opcode and three operands.
1424template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match {
1425 T0 Op1;
1426 T1 Op2;
1427
1428 TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {}
1429
1430 template <typename OpTy> bool match(OpTy *V) {
1431 if (V->getValueID() == Value::InstructionVal + Opcode) {
1432 auto *I = cast<Instruction>(V);
1433 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1));
1434 }
1435 return false;
1436 }
1437};
1438
1439/// Matches instructions with Opcode and three operands.
1440template <typename T0, typename T1, typename T2, unsigned Opcode>
1441struct ThreeOps_match {
1442 T0 Op1;
1443 T1 Op2;
1444 T2 Op3;
1445
1446 ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3)
1447 : Op1(Op1), Op2(Op2), Op3(Op3) {}
1448
1449 template <typename OpTy> bool match(OpTy *V) {
1450 if (V->getValueID() == Value::InstructionVal + Opcode) {
65
Called C++ object pointer is null
1451 auto *I = cast<Instruction>(V);
1452 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1453 Op3.match(I->getOperand(2));
1454 }
1455 return false;
1456 }
1457};
1458
1459/// Matches SelectInst.
1460template <typename Cond, typename LHS, typename RHS>
1461inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select>
1462m_Select(const Cond &C, const LHS &L, const RHS &R) {
1463 return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R);
1464}
1465
1466/// This matches a select of two constants, e.g.:
1467/// m_SelectCst<-1, 0>(m_Value(V))
1468template <int64_t L, int64_t R, typename Cond>
1469inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>,
1470 Instruction::Select>
1471m_SelectCst(const Cond &C) {
1472 return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>());
1473}
1474
1475/// Matches FreezeInst.
1476template <typename OpTy>
1477inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) {
1478 return OneOps_match<OpTy, Instruction::Freeze>(Op);
1479}
1480
1481/// Matches InsertElementInst.
1482template <typename Val_t, typename Elt_t, typename Idx_t>
1483inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>
1484m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) {
1485 return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>(
1486 Val, Elt, Idx);
1487}
1488
1489/// Matches ExtractElementInst.
1490template <typename Val_t, typename Idx_t>
1491inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>
1492m_ExtractElt(const Val_t &Val, const Idx_t &Idx) {
1493 return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx);
1494}
1495
1496/// Matches shuffle.
1497template <typename T0, typename T1, typename T2> struct Shuffle_match {
1498 T0 Op1;
1499 T1 Op2;
1500 T2 Mask;
1501
1502 Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask)
1503 : Op1(Op1), Op2(Op2), Mask(Mask) {}
1504
1505 template <typename OpTy> bool match(OpTy *V) {
1506 if (auto *I = dyn_cast<ShuffleVectorInst>(V)) {
1507 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1508 Mask.match(I->getShuffleMask());
1509 }
1510 return false;
1511 }
1512};
1513
1514struct m_Mask {
1515 ArrayRef<int> &MaskRef;
1516 m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1517 bool match(ArrayRef<int> Mask) {
1518 MaskRef = Mask;
1519 return true;
1520 }
1521};
1522
1523struct m_ZeroMask {
1524 bool match(ArrayRef<int> Mask) {
1525 return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; });
1526 }
1527};
1528
1529struct m_SpecificMask {
1530 ArrayRef<int> &MaskRef;
1531 m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1532 bool match(ArrayRef<int> Mask) { return MaskRef == Mask; }
1533};
1534
1535struct m_SplatOrUndefMask {
1536 int &SplatIndex;
1537 m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {}
1538 bool match(ArrayRef<int> Mask) {
1539 auto First = find_if(Mask, [](int Elem) { return Elem != -1; });
1540 if (First == Mask.end())
1541 return false;
1542 SplatIndex = *First;
1543 return all_of(Mask,
1544 [First](int Elem) { return Elem == *First || Elem == -1; });
1545 }
1546};
1547
1548/// Matches ShuffleVectorInst independently of mask value.
1549template <typename V1_t, typename V2_t>
1550inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>
1551m_Shuffle(const V1_t &v1, const V2_t &v2) {
1552 return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2);
1553}
1554
1555template <typename V1_t, typename V2_t, typename Mask_t>
1556inline Shuffle_match<V1_t, V2_t, Mask_t>
1557m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) {
1558 return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask);
1559}
1560
1561/// Matches LoadInst.
1562template <typename OpTy>
1563inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) {
1564 return OneOps_match<OpTy, Instruction::Load>(Op);
1565}
1566
1567/// Matches StoreInst.
1568template <typename ValueOpTy, typename PointerOpTy>
1569inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>
1570m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) {
1571 return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp,
1572 PointerOp);
1573}
1574
1575//===----------------------------------------------------------------------===//
1576// Matchers for CastInst classes
1577//
1578
1579template <typename Op_t, unsigned Opcode> struct CastClass_match {
1580 Op_t Op;
1581
1582 CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
1583
1584 template <typename OpTy> bool match(OpTy *V) {
1585 if (auto *O = dyn_cast<Operator>(V))
1586 return O->getOpcode() == Opcode && Op.match(O->getOperand(0));
1587 return false;
1588 }
1589};
1590
1591/// Matches BitCast.
1592template <typename OpTy>
1593inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) {
1594 return CastClass_match<OpTy, Instruction::BitCast>(Op);
1595}
1596
1597/// Matches PtrToInt.
1598template <typename OpTy>
1599inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) {
1600 return CastClass_match<OpTy, Instruction::PtrToInt>(Op);
1601}
1602
1603/// Matches IntToPtr.
1604template <typename OpTy>
1605inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) {
1606 return CastClass_match<OpTy, Instruction::IntToPtr>(Op);
1607}
1608
1609/// Matches Trunc.
1610template <typename OpTy>
1611inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) {
1612 return CastClass_match<OpTy, Instruction::Trunc>(Op);
1613}
1614
1615template <typename OpTy>
1616inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy>
1617m_TruncOrSelf(const OpTy &Op) {
1618 return m_CombineOr(m_Trunc(Op), Op);
1619}
1620
1621/// Matches SExt.
1622template <typename OpTy>
1623inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) {
1624 return CastClass_match<OpTy, Instruction::SExt>(Op);
1625}
1626
1627/// Matches ZExt.
1628template <typename OpTy>
1629inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) {
1630 return CastClass_match<OpTy, Instruction::ZExt>(Op);
49
Returning without writing to 'Op.VR'
1631}
1632
1633template <typename OpTy>
1634inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy>
1635m_ZExtOrSelf(const OpTy &Op) {
1636 return m_CombineOr(m_ZExt(Op), Op);
48
Calling 'm_ZExt<llvm::PatternMatch::bind_ty<llvm::Value>>'
50
Returning from 'm_ZExt<llvm::PatternMatch::bind_ty<llvm::Value>>'
51
Calling 'm_CombineOr<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>'
53
Returning from 'm_CombineOr<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>'
54
Returning without writing to 'Op.VR'
1637}
1638
1639template <typename OpTy>
1640inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy>
1641m_SExtOrSelf(const OpTy &Op) {
1642 return m_CombineOr(m_SExt(Op), Op);
1643}
1644
1645template <typename OpTy>
1646inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1647 CastClass_match<OpTy, Instruction::SExt>>
1648m_ZExtOrSExt(const OpTy &Op) {
1649 return m_CombineOr(m_ZExt(Op), m_SExt(Op));
1650}
1651
1652template <typename OpTy>
1653inline match_combine_or<
1654 match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1655 CastClass_match<OpTy, Instruction::SExt>>,
1656 OpTy>
1657m_ZExtOrSExtOrSelf(const OpTy &Op) {
1658 return m_CombineOr(m_ZExtOrSExt(Op), Op);
1659}
1660
1661template <typename OpTy>
1662inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) {
1663 return CastClass_match<OpTy, Instruction::UIToFP>(Op);
1664}
1665
1666template <typename OpTy>
1667inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
1668 return CastClass_match<OpTy, Instruction::SIToFP>(Op);
1669}
1670
1671template <typename OpTy>
1672inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) {
1673 return CastClass_match<OpTy, Instruction::FPToUI>(Op);
1674}
1675
1676template <typename OpTy>
1677inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) {
1678 return CastClass_match<OpTy, Instruction::FPToSI>(Op);
1679}
1680
1681template <typename OpTy>
1682inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
1683 return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
1684}
1685
1686template <typename OpTy>
1687inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
1688 return CastClass_match<OpTy, Instruction::FPExt>(Op);
1689}
1690
1691//===----------------------------------------------------------------------===//
1692// Matchers for control flow.
1693//
1694
1695struct br_match {
1696 BasicBlock *&Succ;
1697
1698 br_match(BasicBlock *&Succ) : Succ(Succ) {}
1699
1700 template <typename OpTy> bool match(OpTy *V) {
1701 if (auto *BI = dyn_cast<BranchInst>(V))
1702 if (BI->isUnconditional()) {
1703 Succ = BI->getSuccessor(0);
1704 return true;
1705 }
1706 return false;
1707 }
1708};
1709
1710inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); }
1711
1712template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1713struct brc_match {
1714 Cond_t Cond;
1715 TrueBlock_t T;
1716 FalseBlock_t F;
1717
1718 brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f)
1719 : Cond(C), T(t), F(f) {}
1720
1721 template <typename OpTy> bool match(OpTy *V) {
1722 if (auto *BI = dyn_cast<BranchInst>(V))
1723 if (BI->isConditional() && Cond.match(BI->getCondition()))
1724 return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1));
1725 return false;
1726 }
1727};
1728
1729template <typename Cond_t>
1730inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>
1731m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) {
1732 return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>(
1733 C, m_BasicBlock(T), m_BasicBlock(F));
1734}
1735
1736template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1737inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t>
1738m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) {
1739 return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F);
1740}
1741
1742//===----------------------------------------------------------------------===//
1743// Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y).
1744//
1745
1746template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t,
1747 bool Commutable = false>
1748struct MaxMin_match {
1749 using PredType = Pred_t;
1750 LHS_t L;
1751 RHS_t R;
1752
1753 // The evaluation order is always stable, regardless of Commutability.
1754 // The LHS is always matched first.
1755 MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1756
1757 template <typename OpTy> bool match(OpTy *V) {
1758 if (auto *II = dyn_cast<IntrinsicInst>(V)) {
1759 Intrinsic::ID IID = II->getIntrinsicID();
1760 if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) ||
1761 (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) ||
1762 (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) ||
1763 (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) {
1764 Value *LHS = II->getOperand(0), *RHS = II->getOperand(1);
1765 return (L.match(LHS) && R.match(RHS)) ||
1766 (Commutable && L.match(RHS) && R.match(LHS));
1767 }
1768 }
1769 // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x".
1770 auto *SI = dyn_cast<SelectInst>(V);
1771 if (!SI)
1772 return false;
1773 auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition());
1774 if (!Cmp)
1775 return false;
1776 // At this point we have a select conditioned on a comparison. Check that
1777 // it is the values returned by the select that are being compared.
1778 auto *TrueVal = SI->getTrueValue();
1779 auto *FalseVal = SI->getFalseValue();
1780 auto *LHS = Cmp->getOperand(0);
1781 auto *RHS = Cmp->getOperand(1);
1782 if ((TrueVal != LHS || FalseVal != RHS) &&
1783 (TrueVal != RHS || FalseVal != LHS))
1784 return false;
1785 typename CmpInst_t::Predicate Pred =
1786 LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate();
1787 // Does "(x pred y) ? x : y" represent the desired max/min operation?
1788 if (!Pred_t::match(Pred))
1789 return false;
1790 // It does! Bind the operands.
1791 return (L.match(LHS) && R.match(RHS)) ||
1792 (Commutable && L.match(RHS) && R.match(LHS));
1793 }
1794};
1795
1796/// Helper class for identifying signed max predicates.
1797struct smax_pred_ty {
1798 static bool match(ICmpInst::Predicate Pred) {
1799 return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE;
1800 }
1801};
1802
1803/// Helper class for identifying signed min predicates.
1804struct smin_pred_ty {
1805 static bool match(ICmpInst::Predicate Pred) {
1806 return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE;
1807 }
1808};
1809
1810/// Helper class for identifying unsigned max predicates.
1811struct umax_pred_ty {
1812 static bool match(ICmpInst::Predicate Pred) {
1813 return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE;
1814 }
1815};
1816
1817/// Helper class for identifying unsigned min predicates.
1818struct umin_pred_ty {
1819 static bool match(ICmpInst::Predicate Pred) {
1820 return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE;
1821 }
1822};
1823
1824/// Helper class for identifying ordered max predicates.
1825struct ofmax_pred_ty {
1826 static bool match(FCmpInst::Predicate Pred) {
1827 return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE;
1828 }
1829};
1830
1831/// Helper class for identifying ordered min predicates.
1832struct ofmin_pred_ty {
1833 static bool match(FCmpInst::Predicate Pred) {
1834 return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE;
1835 }
1836};
1837
1838/// Helper class for identifying unordered max predicates.
1839struct ufmax_pred_ty {
1840 static bool match(FCmpInst::Predicate Pred) {
1841 return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE;
1842 }
1843};
1844
1845/// Helper class for identifying unordered min predicates.
1846struct ufmin_pred_ty {
1847 static bool match(FCmpInst::Predicate Pred) {
1848 return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE;
1849 }
1850};
1851
1852template <typename LHS, typename RHS>
1853inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L,
1854 const RHS &R) {
1855 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R);
1856}
1857
1858template <typename LHS, typename RHS>
1859inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L,
1860 const RHS &R) {
1861 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R);
1862}
1863
1864template <typename LHS, typename RHS>
1865inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L,
1866 const RHS &R) {
1867 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R);
1868}
1869
1870template <typename LHS, typename RHS>
1871inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L,
1872 const RHS &R) {
1873 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R);
1874}
1875
1876template <typename LHS, typename RHS>
1877inline match_combine_or<
1878 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>,
1879 MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>,
1880 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>,
1881 MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>>
1882m_MaxOrMin(const LHS &L, const RHS &R) {
1883 return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)),
1884 m_CombineOr(m_UMax(L, R), m_UMin(L, R)));
1885}
1886
1887/// Match an 'ordered' floating point maximum function.
1888/// Floating point has one special value 'NaN'. Therefore, there is no total
1889/// order. However, if we can ignore the 'NaN' value (for example, because of a
1890/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1891/// semantics. In the presence of 'NaN' we have to preserve the original
1892/// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate.
1893///
1894/// max(L, R) iff L and R are not NaN
1895/// m_OrdFMax(L, R) = R iff L or R are NaN
1896template <typename LHS, typename RHS>
1897inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L,
1898 const RHS &R) {
1899 return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R);
1900}
1901
1902/// Match an 'ordered' floating point minimum function.
1903/// Floating point has one special value 'NaN'. Therefore, there is no total
1904/// order. However, if we can ignore the 'NaN' value (for example, because of a
1905/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1906/// semantics. In the presence of 'NaN' we have to preserve the original
1907/// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate.
1908///
1909/// min(L, R) iff L and R are not NaN
1910/// m_OrdFMin(L, R) = R iff L or R are NaN
1911template <typename LHS, typename RHS>
1912inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L,
1913 const RHS &R) {
1914 return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R);
1915}
1916
1917/// Match an 'unordered' floating point maximum function.
1918/// Floating point has one special value 'NaN'. Therefore, there is no total
1919/// order. However, if we can ignore the 'NaN' value (for example, because of a
1920/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1921/// semantics. In the presence of 'NaN' we have to preserve the original
1922/// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate.
1923///
1924/// max(L, R) iff L and R are not NaN
1925/// m_UnordFMax(L, R) = L iff L or R are NaN
1926template <typename LHS, typename RHS>
1927inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>
1928m_UnordFMax(const LHS &L, const RHS &R) {
1929 return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R);
1930}
1931
1932/// Match an 'unordered' floating point minimum function.
1933/// Floating point has one special value 'NaN'. Therefore, there is no total
1934/// order. However, if we can ignore the 'NaN' value (for example, because of a
1935/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1936/// semantics. In the presence of 'NaN' we have to preserve the original
1937/// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate.
1938///
1939/// min(L, R) iff L and R are not NaN
1940/// m_UnordFMin(L, R) = L iff L or R are NaN
1941template <typename LHS, typename RHS>
1942inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>
1943m_UnordFMin(const LHS &L, const RHS &R) {
1944 return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R);
1945}
1946
1947//===----------------------------------------------------------------------===//
1948// Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b
1949// Note that S might be matched to other instructions than AddInst.
1950//
1951
1952template <typename LHS_t, typename RHS_t, typename Sum_t>
1953struct UAddWithOverflow_match {
1954 LHS_t L;
1955 RHS_t R;
1956 Sum_t S;
1957
1958 UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S)
1959 : L(L), R(R), S(S) {}
1960
1961 template <typename OpTy> bool match(OpTy *V) {
1962 Value *ICmpLHS, *ICmpRHS;
1963 ICmpInst::Predicate Pred;
1964 if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V))
1965 return false;
1966
1967 Value *AddLHS, *AddRHS;
1968 auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS));
1969
1970 // (a + b) u< a, (a + b) u< b
1971 if (Pred == ICmpInst::ICMP_ULT)
1972 if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS))
1973 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1974
1975 // a >u (a + b), b >u (a + b)
1976 if (Pred == ICmpInst::ICMP_UGT)
1977 if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS))
1978 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1979
1980 Value *Op1;
1981 auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes()));
1982 // (a ^ -1) <u b
1983 if (Pred == ICmpInst::ICMP_ULT) {
1984 if (XorExpr.match(ICmpLHS))
1985 return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS);
1986 }
1987 // b > u (a ^ -1)
1988 if (Pred == ICmpInst::ICMP_UGT) {
1989 if (XorExpr.match(ICmpRHS))
1990 return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS);
1991 }
1992
1993 // Match special-case for increment-by-1.
1994 if (Pred == ICmpInst::ICMP_EQ) {
1995 // (a + 1) == 0
1996 // (1 + a) == 0
1997 if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) &&
1998 (m_One().match(AddLHS) || m_One().match(AddRHS)))
1999 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
2000 // 0 == (a + 1)
2001 // 0 == (1 + a)
2002 if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) &&
2003 (m_One().match(AddLHS) || m_One().match(AddRHS)))
2004 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
2005 }
2006
2007 return false;
2008 }
2009};
2010
2011/// Match an icmp instruction checking for unsigned overflow on addition.
2012///
2013/// S is matched to the addition whose result is being checked for overflow, and
2014/// L and R are matched to the LHS and RHS of S.
2015template <typename LHS_t, typename RHS_t, typename Sum_t>
2016UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>
2017m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) {
2018 return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S);
2019}
2020
2021template <typename Opnd_t> struct Argument_match {
2022 unsigned OpI;
2023 Opnd_t Val;
2024
2025 Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {}
2026
2027 template <typename OpTy> bool match(OpTy *V) {
2028 // FIXME: Should likely be switched to use `CallBase`.
2029 if (const auto *CI = dyn_cast<CallInst>(V))
2030 return Val.match(CI->getArgOperand(OpI));
2031 return false;
2032 }
2033};
2034
2035/// Match an argument.
2036template <unsigned OpI, typename Opnd_t>
2037inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) {
2038 return Argument_match<Opnd_t>(OpI, Op);
2039}
2040
2041/// Intrinsic matchers.
2042struct IntrinsicID_match {
2043 unsigned ID;
2044
2045 IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {}
2046
2047 template <typename OpTy> bool match(OpTy *V) {
2048 if (const auto *CI = dyn_cast<CallInst>(V))
2049 if (const auto *F = CI->getCalledFunction())
2050 return F->getIntrinsicID() == ID;
2051 return false;
2052 }
2053};
2054
2055/// Intrinsic matches are combinations of ID matchers, and argument
2056/// matchers. Higher arity matcher are defined recursively in terms of and-ing
2057/// them with lower arity matchers. Here's some convenient typedefs for up to
2058/// several arguments, and more can be added as needed
2059template <typename T0 = void, typename T1 = void, typename T2 = void,
2060 typename T3 = void, typename T4 = void, typename T5 = void,
2061 typename T6 = void, typename T7 = void, typename T8 = void,
2062 typename T9 = void, typename T10 = void>
2063struct m_Intrinsic_Ty;
2064template <typename T0> struct m_Intrinsic_Ty<T0> {
2065 using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>;
2066};
2067template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> {
2068 using Ty =
2069 match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>;
2070};
2071template <typename T0, typename T1, typename T2>
2072struct m_Intrinsic_Ty<T0, T1, T2> {
2073 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty,
2074 Argument_match<T2>>;
2075};
2076template <typename T0, typename T1, typename T2, typename T3>
2077struct m_Intrinsic_Ty<T0, T1, T2, T3> {
2078 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty,
2079 Argument_match<T3>>;
2080};
2081
2082template <typename T0, typename T1, typename T2, typename T3, typename T4>
2083struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> {
2084 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty,
2085 Argument_match<T4>>;
2086};
2087
2088template <typename T0, typename T1, typename T2, typename T3, typename T4,
2089 typename T5>
2090struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> {
2091 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty,
2092 Argument_match<T5>>;
2093};
2094
2095/// Match intrinsic calls like this:
2096/// m_Intrinsic<Intrinsic::fabs>(m_Value(X))
2097template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() {
2098 return IntrinsicID_match(IntrID);
2099}
2100
2101/// Matches MaskedLoad Intrinsic.
2102template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3>
2103inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty
2104m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2,
2105 const Opnd3 &Op3) {
2106 return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3);
2107}
2108
2109/// Matches MaskedGather Intrinsic.
2110template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3>
2111inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty
2112m_MaskedGather(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2,
2113 const Opnd3 &Op3) {
2114 return m_Intrinsic<Intrinsic::masked_gather>(Op0, Op1, Op2, Op3);
2115}
2116
2117template <Intrinsic::ID IntrID, typename T0>
2118inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) {
2119 return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0));
2120}
2121
2122template <Intrinsic::ID IntrID, typename T0, typename T1>
2123inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0,
2124 const T1 &Op1) {
2125 return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1));
2126}
2127
2128template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2>
2129inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty
2130m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) {
2131 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2));
2132}
2133
2134template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2135 typename T3>
2136inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty
2137m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) {
2138 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3));
2139}
2140
2141template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2142 typename T3, typename T4>
2143inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty
2144m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2145 const T4 &Op4) {
2146 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3),
2147 m_Argument<4>(Op4));
2148}
2149
2150template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2151 typename T3, typename T4, typename T5>
2152inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty
2153m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2154 const T4 &Op4, const T5 &Op5) {
2155 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4),
2156 m_Argument<5>(Op5));
2157}
2158
2159// Helper intrinsic matching specializations.
2160template <typename Opnd0>
2161inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) {
2162 return m_Intrinsic<Intrinsic::bitreverse>(Op0);
2163}
2164
2165template <typename Opnd0>
2166inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) {
2167 return m_Intrinsic<Intrinsic::bswap>(Op0);
2168}
2169
2170template <typename Opnd0>
2171inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) {
2172 return m_Intrinsic<Intrinsic::fabs>(Op0);
2173}
2174
2175template <typename Opnd0>
2176inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) {
2177 return m_Intrinsic<Intrinsic::canonicalize>(Op0);
2178}
2179
2180template <typename Opnd0, typename Opnd1>
2181inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0,
2182 const Opnd1 &Op1) {
2183 return m_Intrinsic<Intrinsic::minnum>(Op0, Op1);
2184}
2185
2186template <typename Opnd0, typename Opnd1>
2187inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0,
2188 const Opnd1 &Op1) {
2189 return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1);
2190}
2191
2192template <typename Opnd0, typename Opnd1, typename Opnd2>
2193inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2194m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2195 return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2);
2196}
2197
2198template <typename Opnd0, typename Opnd1, typename Opnd2>
2199inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2200m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2201 return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2);
2202}
2203
2204template <typename Opnd0>
2205inline typename m_Intrinsic_Ty<Opnd0>::Ty m_Sqrt(const Opnd0 &Op0) {
2206 return m_Intrinsic<Intrinsic::sqrt>(Op0);
2207}
2208
2209//===----------------------------------------------------------------------===//
2210// Matchers for two-operands operators with the operators in either order
2211//
2212
2213/// Matches a BinaryOperator with LHS and RHS in either order.
2214template <typename LHS, typename RHS>
2215inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) {
2216 return AnyBinaryOp_match<LHS, RHS, true>(L, R);
11
Returning without writing to 'R.VR'
2217}
2218
2219/// Matches an ICmp with a predicate over LHS and RHS in either order.
2220/// Swaps the predicate if operands are commuted.
2221template <typename LHS, typename RHS>
2222inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>
2223m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
2224 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L,
2225 R);
2226}
2227
2228/// Matches a specific opcode with LHS and RHS in either order.
2229template <typename LHS, typename RHS>
2230inline SpecificBinaryOp_match<LHS, RHS, true>
2231m_c_BinOp(unsigned Opcode, const LHS &L, const RHS &R) {
2232 return SpecificBinaryOp_match<LHS, RHS, true>(Opcode, L, R);
2233}
2234
2235/// Matches a Add with LHS and RHS in either order.
2236template <typename LHS, typename RHS>
2237inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L,
2238 const RHS &R) {
2239 return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R);
2240}
2241
2242/// Matches a Mul with LHS and RHS in either order.
2243template <typename LHS, typename RHS>
2244inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L,
2245 const RHS &R) {
2246 return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R);
2247}
2248
2249/// Matches an And with LHS and RHS in either order.
2250template <typename LHS, typename RHS>
2251inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L,
2252 const RHS &R) {
2253 return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R);
2254}
2255
2256/// Matches an Or with LHS and RHS in either order.
2257template <typename LHS, typename RHS>
2258inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L,
2259 const RHS &R) {
2260 return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R);
2261}
2262
2263/// Matches an Xor with LHS and RHS in either order.
2264template <typename LHS, typename RHS>
2265inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L,
2266 const RHS &R) {
2267 return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R);
2268}
2269
2270/// Matches a 'Neg' as 'sub 0, V'.
2271template <typename ValTy>
2272inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub>
2273m_Neg(const ValTy &V) {
2274 return m_Sub(m_ZeroInt(), V);
2275}
2276
2277/// Matches a 'Neg' as 'sub nsw 0, V'.
2278template <typename ValTy>
2279inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy,
2280 Instruction::Sub,
2281 OverflowingBinaryOperator::NoSignedWrap>
2282m_NSWNeg(const ValTy &V) {
2283 return m_NSWSub(m_ZeroInt(), V);
2284}
2285
2286/// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
2287template <typename ValTy>
2288inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true>
2289m_Not(const ValTy &V) {
2290 return m_c_Xor(V, m_AllOnes());
2291}
2292
2293template <typename ValTy> struct NotForbidUndef_match {
2294 ValTy Val;
2295 NotForbidUndef_match(const ValTy &V) : Val(V) {}
2296
2297 template <typename OpTy> bool match(OpTy *V) {
2298 // We do not use m_c_Xor because that could match an arbitrary APInt that is
2299 // not -1 as C and then fail to match the other operand if it is -1.
2300 // This code should still work even when both operands are constants.
2301 Value *X;
2302 const APInt *C;
2303 if (m_Xor(m_Value(X), m_APIntForbidUndef(C)).match(V) && C->isAllOnes())
2304 return Val.match(X);
2305 if (m_Xor(m_APIntForbidUndef(C), m_Value(X)).match(V) && C->isAllOnes())
2306 return Val.match(X);
2307 return false;
2308 }
2309};
2310
2311/// Matches a bitwise 'not' as 'xor V, -1' or 'xor -1, V'. For vectors, the
2312/// constant value must be composed of only -1 scalar elements.
2313template <typename ValTy>
2314inline NotForbidUndef_match<ValTy> m_NotForbidUndef(const ValTy &V) {
2315 return NotForbidUndef_match<ValTy>(V);
2316}
2317
2318/// Matches an SMin with LHS and RHS in either order.
2319template <typename LHS, typename RHS>
2320inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>
2321m_c_SMin(const LHS &L, const RHS &R) {
2322 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R);
2323}
2324/// Matches an SMax with LHS and RHS in either order.
2325template <typename LHS, typename RHS>
2326inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>
2327m_c_SMax(const LHS &L, const RHS &R) {
2328 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R);
2329}
2330/// Matches a UMin with LHS and RHS in either order.
2331template <typename LHS, typename RHS>
2332inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>
2333m_c_UMin(const LHS &L, const RHS &R) {
2334 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R);
2335}
2336/// Matches a UMax with LHS and RHS in either order.
2337template <typename LHS, typename RHS>
2338inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>
2339m_c_UMax(const LHS &L, const RHS &R) {
2340 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R);
2341}
2342
2343template <typename LHS, typename RHS>
2344inline match_combine_or<
2345 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>,
2346 MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>,
2347 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>,
2348 MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>>
2349m_c_MaxOrMin(const LHS &L, const RHS &R) {
2350 return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)),
2351 m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R)));
2352}
2353
2354/// Matches FAdd with LHS and RHS in either order.
2355template <typename LHS, typename RHS>
2356inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true>
2357m_c_FAdd(const LHS &L, const RHS &R) {
2358 return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R);
2359}
2360
2361/// Matches FMul with LHS and RHS in either order.
2362template <typename LHS, typename RHS>
2363inline BinaryOp_match<LHS, RHS, Instruction::FMul, true>
2364m_c_FMul(const LHS &L, const RHS &R) {
2365 return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R);
2366}
2367
2368template <typename Opnd_t> struct Signum_match {
2369 Opnd_t Val;
2370 Signum_match(const Opnd_t &V) : Val(V) {}
2371
2372 template <typename OpTy> bool match(OpTy *V) {
2373 unsigned TypeSize = V->getType()->getScalarSizeInBits();
2374 if (TypeSize == 0)
2375 return false;
2376
2377 unsigned ShiftWidth = TypeSize - 1;
2378 Value *OpL = nullptr, *OpR = nullptr;
2379
2380 // This is the representation of signum we match:
2381 //
2382 // signum(x) == (x >> 63) | (-x >>u 63)
2383 //
2384 // An i1 value is its own signum, so it's correct to match
2385 //
2386 // signum(x) == (x >> 0) | (-x >>u 0)
2387 //
2388 // for i1 values.
2389
2390 auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth));
2391 auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth));
2392 auto Signum = m_Or(LHS, RHS);
2393
2394 return Signum.match(V) && OpL == OpR && Val.match(OpL);
2395 }
2396};
2397
2398/// Matches a signum pattern.
2399///
2400/// signum(x) =
2401/// x > 0 -> 1
2402/// x == 0 -> 0
2403/// x < 0 -> -1
2404template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) {
2405 return Signum_match<Val_t>(V);
2406}
2407
2408template <int Ind, typename Opnd_t> struct ExtractValue_match {
2409 Opnd_t Val;
2410 ExtractValue_match(const Opnd_t &V) : Val(V) {}
2411
2412 template <typename OpTy> bool match(OpTy *V) {
2413 if (auto *I = dyn_cast<ExtractValueInst>(V)) {
2414 // If Ind is -1, don't inspect indices
2415 if (Ind != -1 &&
2416 !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind))
2417 return false;
2418 return Val.match(I->getAggregateOperand());
2419 }
2420 return false;
2421 }
2422};
2423
2424/// Match a single index ExtractValue instruction.
2425/// For example m_ExtractValue<1>(...)
2426template <int Ind, typename Val_t>
2427inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) {
2428 return ExtractValue_match<Ind, Val_t>(V);
2429}
2430
2431/// Match an ExtractValue instruction with any index.
2432/// For example m_ExtractValue(...)
2433template <typename Val_t>
2434inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) {
2435 return ExtractValue_match<-1, Val_t>(V);
2436}
2437
2438/// Matcher for a single index InsertValue instruction.
2439template <int Ind, typename T0, typename T1> struct InsertValue_match {
2440 T0 Op0;
2441 T1 Op1;
2442
2443 InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {}
2444
2445 template <typename OpTy> bool match(OpTy *V) {
2446 if (auto *I = dyn_cast<InsertValueInst>(V)) {
2447 return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) &&
2448 I->getNumIndices() == 1 && Ind == I->getIndices()[0];
2449 }
2450 return false;
2451 }
2452};
2453
2454/// Matches a single index InsertValue instruction.
2455template <int Ind, typename Val_t, typename Elt_t>
2456inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val,
2457 const Elt_t &Elt) {
2458 return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt);
2459}
2460
2461/// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or
2462/// the constant expression
2463/// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>`
2464/// under the right conditions determined by DataLayout.
2465struct VScaleVal_match {
2466 const DataLayout &DL;
2467 VScaleVal_match(const DataLayout &DL) : DL(DL) {}
2468
2469 template <typename ITy> bool match(ITy *V) {
2470 if (m_Intrinsic<Intrinsic::vscale>().match(V))
2471 return true;
2472
2473 Value *Ptr;
2474 if (m_PtrToInt(m_Value(Ptr)).match(V)) {
2475 if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
2476 auto *DerefTy = GEP->getSourceElementType();
2477 if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) &&
2478 m_Zero().match(GEP->getPointerOperand()) &&
2479 m_SpecificInt(1).match(GEP->idx_begin()->get()) &&
2480 DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8)
2481 return true;
2482 }
2483 }
2484
2485 return false;
2486 }
2487};
2488
2489inline VScaleVal_match m_VScale(const DataLayout &DL) {
2490 return VScaleVal_match(DL);
2491}
2492
2493template <typename LHS, typename RHS, unsigned Opcode, bool Commutable = false>
2494struct LogicalOp_match {
2495 LHS L;
2496 RHS R;
2497
2498 LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {}
2499
2500 template <typename T> bool match(T *V) {
2501 auto *I = dyn_cast<Instruction>(V);
2502 if (!I || !I->getType()->isIntOrIntVectorTy(1))
2503 return false;
2504
2505 if (I->getOpcode() == Opcode) {
2506 auto *Op0 = I->getOperand(0);
2507 auto *Op1 = I->getOperand(1);
2508 return (L.match(Op0) && R.match(Op1)) ||
2509 (Commutable && L.match(Op1) && R.match(Op0));
2510 }
2511
2512 if (auto *Select = dyn_cast<SelectInst>(I)) {
2513 auto *Cond = Select->getCondition();
2514 auto *TVal = Select->getTrueValue();
2515 auto *FVal = Select->getFalseValue();
2516
2517 // Don't match a scalar select of bool vectors.
2518 // Transforms expect a single type for operands if this matches.
2519 if (Cond->getType() != Select->getType())
2520 return false;
2521
2522 if (Opcode == Instruction::And) {
2523 auto *C = dyn_cast<Constant>(FVal);
2524 if (C && C->isNullValue())
2525 return (L.match(Cond) && R.match(TVal)) ||
2526 (Commutable && L.match(TVal) && R.match(Cond));
2527 } else {
2528 assert(Opcode == Instruction::Or)(static_cast <bool> (Opcode == Instruction::Or) ? void (
0) : __assert_fail ("Opcode == Instruction::Or", "llvm/include/llvm/IR/PatternMatch.h"
, 2528, __extension__ __PRETTY_FUNCTION__))
;
2529 auto *C = dyn_cast<Constant>(TVal);
2530 if (C && C->isOneValue())
2531 return (L.match(Cond) && R.match(FVal)) ||
2532 (Commutable && L.match(FVal) && R.match(Cond));
2533 }
2534 }
2535
2536 return false;
2537 }
2538};
2539
2540/// Matches L && R either in the form of L & R or L ? R : false.
2541/// Note that the latter form is poison-blocking.
2542template <typename LHS, typename RHS>
2543inline LogicalOp_match<LHS, RHS, Instruction::And> m_LogicalAnd(const LHS &L,
2544 const RHS &R) {
2545 return LogicalOp_match<LHS, RHS, Instruction::And>(L, R);
2546}
2547
2548/// Matches L && R where L and R are arbitrary values.
2549inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); }
2550
2551/// Matches L && R with LHS and RHS in either order.
2552template <typename LHS, typename RHS>
2553inline LogicalOp_match<LHS, RHS, Instruction::And, true>
2554m_c_LogicalAnd(const LHS &L, const RHS &R) {
2555 return LogicalOp_match<LHS, RHS, Instruction::And, true>(L, R);
2556}
2557
2558/// Matches L || R either in the form of L | R or L ? true : R.
2559/// Note that the latter form is poison-blocking.
2560template <typename LHS, typename RHS>
2561inline LogicalOp_match<LHS, RHS, Instruction::Or> m_LogicalOr(const LHS &L,
2562 const RHS &R) {
2563 return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R);
2564}
2565
2566/// Matches L || R where L and R are arbitrary values.
2567inline auto m_LogicalOr() { return m_LogicalOr(m_Value(), m_Value()); }
2568
2569/// Matches L || R with LHS and RHS in either order.
2570template <typename LHS, typename RHS>
2571inline LogicalOp_match<LHS, RHS, Instruction::Or, true>
2572m_c_LogicalOr(const LHS &L, const RHS &R) {
2573 return LogicalOp_match<LHS, RHS, Instruction::Or, true>(L, R);
2574}
2575
2576} // end namespace PatternMatch
2577} // end namespace llvm
2578
2579#endif // LLVM_IR_PATTERNMATCH_H