Bug Summary

File:build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/include/llvm/IR/PatternMatch.h
Warning:line 1460, 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-store=region -analyzer-opt-analyze-nested-blocks -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/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm -resource-dir /usr/lib/llvm-15/lib/clang/15.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/InstCombine -I include -I /build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/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-15/lib/clang/15.0.0/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/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/= -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 -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/= -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-2022-04-20-140412-16051-1 -x c++ /build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/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 <= array_lengthof(TmpResult) + 1) &&(static_cast <bool> ((NextTmpIdx <= array_lengthof(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 580
, __extension__ __PRETTY_FUNCTION__))
580 "out-of-bound access")(static_cast <bool> ((NextTmpIdx <= array_lengthof(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 580
, __extension__ __PRETTY_FUNCTION__))
;
581
582 Value *Result;
583 if (!SimpVect.empty())
584 Result = createNaryFAdd(SimpVect, InstrQuota);
585 else {
586 // The addition is folded to 0.0.
587 Result = ConstantFP::get(Instr->getType(), 0.0);
588 }
589
590 return Result;
591}
592
593Value *FAddCombine::createNaryFAdd
594 (const AddendVect &Opnds, unsigned InstrQuota) {
595 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", 595
, __extension__ __PRETTY_FUNCTION__))
;
596
597 // Step 1: Check if the # of instructions needed exceeds the quota.
598
599 unsigned InstrNeeded = calcInstrNumber(Opnds);
600 if (InstrNeeded > InstrQuota)
601 return nullptr;
602
603 initCreateInstNum();
604
605 // step 2: Emit the N-ary addition.
606 // Note that at most three instructions are involved in Fadd-InstCombine: the
607 // addition in question, and at most two neighboring instructions.
608 // The resulting optimized addition should have at least one less instruction
609 // than the original addition expression tree. This implies that the resulting
610 // N-ary addition has at most two instructions, and we don't need to worry
611 // about tree-height when constructing the N-ary addition.
612
613 Value *LastVal = nullptr;
614 bool LastValNeedNeg = false;
615
616 // Iterate the addends, creating fadd/fsub using adjacent two addends.
617 for (const FAddend *Opnd : Opnds) {
618 bool NeedNeg;
619 Value *V = createAddendVal(*Opnd, NeedNeg);
620 if (!LastVal) {
621 LastVal = V;
622 LastValNeedNeg = NeedNeg;
623 continue;
624 }
625
626 if (LastValNeedNeg == NeedNeg) {
627 LastVal = createFAdd(LastVal, V);
628 continue;
629 }
630
631 if (LastValNeedNeg)
632 LastVal = createFSub(V, LastVal);
633 else
634 LastVal = createFSub(LastVal, V);
635
636 LastValNeedNeg = false;
637 }
638
639 if (LastValNeedNeg) {
640 LastVal = createFNeg(LastVal);
641 }
642
643#ifndef NDEBUG
644 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", 645
, __extension__ __PRETTY_FUNCTION__))
645 "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", 645
, __extension__ __PRETTY_FUNCTION__))
;
646#endif
647
648 return LastVal;
649}
650
651Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
652 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
653 if (Instruction *I = dyn_cast<Instruction>(V))
654 createInstPostProc(I);
655 return V;
656}
657
658Value *FAddCombine::createFNeg(Value *V) {
659 Value *NewV = Builder.CreateFNeg(V);
660 if (Instruction *I = dyn_cast<Instruction>(NewV))
661 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
662 return NewV;
663}
664
665Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
666 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
667 if (Instruction *I = dyn_cast<Instruction>(V))
668 createInstPostProc(I);
669 return V;
670}
671
672Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
673 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
674 if (Instruction *I = dyn_cast<Instruction>(V))
675 createInstPostProc(I);
676 return V;
677}
678
679void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
680 NewInstr->setDebugLoc(Instr->getDebugLoc());
681
682 // Keep track of the number of instruction created.
683 if (!NoNumber)
684 incCreateInstNum();
685
686 // Propagate fast-math flags
687 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
688}
689
690// Return the number of instruction needed to emit the N-ary addition.
691// NOTE: Keep this function in sync with createAddendVal().
692unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
693 unsigned OpndNum = Opnds.size();
694 unsigned InstrNeeded = OpndNum - 1;
695
696 // Adjust the number of instructions needed to emit the N-ary add.
697 for (const FAddend *Opnd : Opnds) {
698 if (Opnd->isConstant())
699 continue;
700
701 // The constant check above is really for a few special constant
702 // coefficients.
703 if (isa<UndefValue>(Opnd->getSymVal()))
704 continue;
705
706 const FAddendCoef &CE = Opnd->getCoef();
707 // Let the addend be "c * x". If "c == +/-1", the value of the addend
708 // is immediately available; otherwise, it needs exactly one instruction
709 // to evaluate the value.
710 if (!CE.isMinusOne() && !CE.isOne())
711 InstrNeeded++;
712 }
713 return InstrNeeded;
714}
715
716// Input Addend Value NeedNeg(output)
717// ================================================================
718// Constant C C false
719// <+/-1, V> V coefficient is -1
720// <2/-2, V> "fadd V, V" coefficient is -2
721// <C, V> "fmul V, C" false
722//
723// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
724Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
725 const FAddendCoef &Coeff = Opnd.getCoef();
726
727 if (Opnd.isConstant()) {
728 NeedNeg = false;
729 return Coeff.getValue(Instr->getType());
730 }
731
732 Value *OpndVal = Opnd.getSymVal();
733
734 if (Coeff.isMinusOne() || Coeff.isOne()) {
735 NeedNeg = Coeff.isMinusOne();
736 return OpndVal;
737 }
738
739 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
740 NeedNeg = Coeff.isMinusTwo();
741 return createFAdd(OpndVal, OpndVal);
742 }
743
744 NeedNeg = false;
745 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
746}
747
748// Checks if any operand is negative and we can convert add to sub.
749// This function checks for following negative patterns
750// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
751// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
752// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
753static Value *checkForNegativeOperand(BinaryOperator &I,
754 InstCombiner::BuilderTy &Builder) {
755 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
756
757 // This function creates 2 instructions to replace ADD, we need at least one
758 // of LHS or RHS to have one use to ensure benefit in transform.
759 if (!LHS->hasOneUse() && !RHS->hasOneUse())
760 return nullptr;
761
762 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
763 const APInt *C1 = nullptr, *C2 = nullptr;
764
765 // if ONE is on other side, swap
766 if (match(RHS, m_Add(m_Value(X), m_One())))
767 std::swap(LHS, RHS);
768
769 if (match(LHS, m_Add(m_Value(X), m_One()))) {
770 // if XOR on other side, swap
771 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
772 std::swap(X, RHS);
773
774 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
775 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
776 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
777 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
778 Value *NewAnd = Builder.CreateAnd(Z, *C1);
779 return Builder.CreateSub(RHS, NewAnd, "sub");
780 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
781 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
782 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
783 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
784 return Builder.CreateSub(RHS, NewOr, "sub");
785 }
786 }
787 }
788
789 // Restore LHS and RHS
790 LHS = I.getOperand(0);
791 RHS = I.getOperand(1);
792
793 // if XOR is on other side, swap
794 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
795 std::swap(LHS, RHS);
796
797 // C2 is ODD
798 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
799 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
800 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
801 if (C1->countTrailingZeros() == 0)
802 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
803 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
804 return Builder.CreateSub(RHS, NewOr, "sub");
805 }
806 return nullptr;
807}
808
809/// Wrapping flags may allow combining constants separated by an extend.
810static Instruction *foldNoWrapAdd(BinaryOperator &Add,
811 InstCombiner::BuilderTy &Builder) {
812 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
813 Type *Ty = Add.getType();
814 Constant *Op1C;
815 if (!match(Op1, m_Constant(Op1C)))
816 return nullptr;
817
818 // Try this match first because it results in an add in the narrow type.
819 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
820 Value *X;
821 const APInt *C1, *C2;
822 if (match(Op1, m_APInt(C1)) &&
823 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
824 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
825 Constant *NewC =
826 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
827 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
828 }
829
830 // More general combining of constants in the wide type.
831 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
832 Constant *NarrowC;
833 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
834 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
835 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
836 Value *WideX = Builder.CreateSExt(X, Ty);
837 return BinaryOperator::CreateAdd(WideX, NewC);
838 }
839 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
840 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
841 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
842 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
843 Value *WideX = Builder.CreateZExt(X, Ty);
844 return BinaryOperator::CreateAdd(WideX, NewC);
845 }
846
847 return nullptr;
848}
849
850Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
851 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
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 const APInt *C;
887 if (!match(Op1, m_APInt(C)))
888 return nullptr;
889
890 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
891 Constant *Op01C;
892 if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
893 haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
894 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
895
896 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
897 const APInt *C2;
898 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
899 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
900
901 if (C->isSignMask()) {
902 // If wrapping is not allowed, then the addition must set the sign bit:
903 // X + (signmask) --> X | signmask
904 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
905 return BinaryOperator::CreateOr(Op0, Op1);
906
907 // If wrapping is allowed, then the addition flips the sign bit of LHS:
908 // X + (signmask) --> X ^ signmask
909 return BinaryOperator::CreateXor(Op0, Op1);
910 }
911
912 // Is this add the last step in a convoluted sext?
913 // add(zext(xor i16 X, -32768), -32768) --> sext X
914 Type *Ty = Add.getType();
915 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
916 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
917 return CastInst::Create(Instruction::SExt, X, Ty);
918
919 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
920 // (X ^ signmask) + C --> (X + (signmask ^ C))
921 if (C2->isSignMask())
922 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
923
924 // If X has no high-bits set above an xor mask:
925 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
926 if (C2->isMask()) {
927 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
928 if ((*C2 | LHSKnown.Zero).isAllOnes())
929 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
930 }
931
932 // Look for a math+logic pattern that corresponds to sext-in-register of a
933 // value with cleared high bits. Convert that into a pair of shifts:
934 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
935 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
936 if (Op0->hasOneUse() && *C2 == -(*C)) {
937 unsigned BitWidth = Ty->getScalarSizeInBits();
938 unsigned ShAmt = 0;
939 if (C->isPowerOf2())
940 ShAmt = BitWidth - C->logBase2() - 1;
941 else if (C2->isPowerOf2())
942 ShAmt = BitWidth - C2->logBase2() - 1;
943 if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
944 0, &Add)) {
945 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
946 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
947 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
948 }
949 }
950 }
951
952 if (C->isOne() && Op0->hasOneUse()) {
953 // add (sext i1 X), 1 --> zext (not X)
954 // TODO: The smallest IR representation is (select X, 0, 1), and that would
955 // not require the one-use check. But we need to remove a transform in
956 // visitSelect and make sure that IR value tracking for select is equal or
957 // better than for these ops.
958 if (match(Op0, m_SExt(m_Value(X))) &&
959 X->getType()->getScalarSizeInBits() == 1)
960 return new ZExtInst(Builder.CreateNot(X), Ty);
961
962 // Shifts and add used to flip and mask off the low bit:
963 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
964 const APInt *C3;
965 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
966 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
967 Value *NotX = Builder.CreateNot(X);
968 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
969 }
970 }
971
972 // If all bits affected by the add are included in a high-bit-mask, do the
973 // add before the mask op:
974 // (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00
975 if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) &&
976 C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) {
977 Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C));
978 return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2));
979 }
980
981 return nullptr;
982}
983
984// Matches multiplication expression Op * C where C is a constant. Returns the
985// constant value in C and the other operand in Op. Returns true if such a
986// match is found.
987static bool MatchMul(Value *E, Value *&Op, APInt &C) {
988 const APInt *AI;
989 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
990 C = *AI;
991 return true;
992 }
993 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
994 C = APInt(AI->getBitWidth(), 1);
995 C <<= *AI;
996 return true;
997 }
998 return false;
999}
1000
1001// Matches remainder expression Op % C where C is a constant. Returns the
1002// constant value in C and the other operand in Op. Returns the signedness of
1003// the remainder operation in IsSigned. Returns true if such a match is
1004// found.
1005static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1006 const APInt *AI;
1007 IsSigned = false;
1008 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1009 IsSigned = true;
1010 C = *AI;
1011 return true;
1012 }
1013 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1014 C = *AI;
1015 return true;
1016 }
1017 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1018 C = *AI + 1;
1019 return true;
1020 }
1021 return false;
1022}
1023
1024// Matches division expression Op / C with the given signedness as indicated
1025// by IsSigned, where C is a constant. Returns the constant value in C and the
1026// other operand in Op. Returns true if such a match is found.
1027static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1028 const APInt *AI;
1029 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1030 C = *AI;
1031 return true;
1032 }
1033 if (!IsSigned) {
1034 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1035 C = *AI;
1036 return true;
1037 }
1038 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1039 C = APInt(AI->getBitWidth(), 1);
1040 C <<= *AI;
1041 return true;
1042 }
1043 }
1044 return false;
1045}
1046
1047// Returns whether C0 * C1 with the given signedness overflows.
1048static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1049 bool overflow;
1050 if (IsSigned)
1051 (void)C0.smul_ov(C1, overflow);
1052 else
1053 (void)C0.umul_ov(C1, overflow);
1054 return overflow;
1055}
1056
1057// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1058// does not overflow.
1059Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1060 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1061 Value *X, *MulOpV;
1062 APInt C0, MulOpC;
1063 bool IsSigned;
1064 // Match I = X % C0 + MulOpV * C0
1065 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1066 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1067 C0 == MulOpC) {
1068 Value *RemOpV;
1069 APInt C1;
1070 bool Rem2IsSigned;
1071 // Match MulOpC = RemOpV % C1
1072 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1073 IsSigned == Rem2IsSigned) {
1074 Value *DivOpV;
1075 APInt DivOpC;
1076 // Match RemOpV = X / C0
1077 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1078 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1079 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1080 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1081 : Builder.CreateURem(X, NewDivisor, "urem");
1082 }
1083 }
1084 }
1085
1086 return nullptr;
1087}
1088
1089/// Fold
1090/// (1 << NBits) - 1
1091/// Into:
1092/// ~(-(1 << NBits))
1093/// Because a 'not' is better for bit-tracking analysis and other transforms
1094/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1095static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1096 InstCombiner::BuilderTy &Builder) {
1097 Value *NBits;
1098 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1099 return nullptr;
1100
1101 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1102 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1103 // Be wary of constant folding.
1104 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1105 // Always NSW. But NUW propagates from `add`.
1106 BOp->setHasNoSignedWrap();
1107 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1108 }
1109
1110 return BinaryOperator::CreateNot(NotMask, I.getName());
1111}
1112
1113static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1114 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", 1114
, __extension__ __PRETTY_FUNCTION__))
;
1115 Type *Ty = I.getType();
1116 auto getUAddSat = [&]() {
1117 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1118 };
1119
1120 // add (umin X, ~Y), Y --> uaddsat X, Y
1121 Value *X, *Y;
1122 if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1123 m_Deferred(Y))))
1124 return CallInst::Create(getUAddSat(), { X, Y });
1125
1126 // add (umin X, ~C), C --> uaddsat X, C
1127 const APInt *C, *NotC;
1128 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1129 *C == ~*NotC)
1130 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1131
1132 return nullptr;
1133}
1134
1135Instruction *InstCombinerImpl::
1136 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1137 BinaryOperator &I) {
1138 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", 1141
, __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
1139 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", 1141
, __extension__ __PRETTY_FUNCTION__))
1140 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", 1141
, __extension__ __PRETTY_FUNCTION__))
1141 "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", 1141
, __extension__ __PRETTY_FUNCTION__))
;
1142
1143 // We have a subtraction/addition between a (potentially truncated) *logical*
1144 // right-shift of X and a "select".
1145 Value *X, *Select;
1146 Instruction *LowBitsToSkip, *Extract;
1147 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>>'
1148 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1149 m_Instruction(Extract))),
1150 m_Value(Select))))
5
Calling 'm_Value'
9
Returning from 'm_Value'
1151 return nullptr;
1152
1153 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1154 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
36
Assuming pointer value is null
37
Taking false branch
1155 return nullptr;
1156
1157 Type *XTy = X->getType();
1158 bool HadTrunc = I.getType() != XTy;
38
Assuming the condition is false
1159
1160 // If there was a truncation of extracted value, then we'll need to produce
1161 // one extra instruction, so we need to ensure one instruction will go away.
1162 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
1163 return nullptr;
1164
1165 // Extraction should extract high NBits bits, with shift amount calculated as:
1166 // low bits to skip = shift bitwidth - high bits to extract
1167 // The shift amount itself may be extended, and we need to look past zero-ext
1168 // when matching NBits, that will matter for matching later.
1169 Constant *C;
1170 Value *NBits;
1171 if (!match(
41
Taking false branch
1172 LowBitsToSkip,
1173 m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1174 !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
40
Assuming the condition is false
1175 APInt(C->getType()->getScalarSizeInBits(),
1176 X->getType()->getScalarSizeInBits()))))
1177 return nullptr;
1178
1179 // Sign-extending value can be zero-extended if we `sub`tract it,
1180 // or sign-extended otherwise.
1181 auto SkipExtInMagic = [&I](Value *&V) {
1182 if (I.getOpcode() == Instruction::Sub)
43
Taking true branch
1183 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>>>'
1184 else
1185 match(V, m_SExtOrSelf(m_Value(V)));
1186 };
59
Returning without writing to 'V'
1187
1188 // Now, finally validate the sign-extending magic.
1189 // `select` itself may be appropriately extended, look past that.
1190 SkipExtInMagic(Select);
42
Calling 'operator()'
60
Returning from 'operator()'
1191
1192 ICmpInst::Predicate Pred;
1193 const APInt *Thr;
1194 Value *SignExtendingValue, *Zero;
1195 bool ShouldSignext;
1196 // It must be a select between two values we will later establish to be a
1197 // sign-extending value and a zero constant. The condition guarding the
1198 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1199 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>>'
1200 m_Value(SignExtendingValue), m_Value(Zero))) ||
1201 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1202 return nullptr;
1203
1204 // icmp-select pair is commutative.
1205 if (!ShouldSignext)
1206 std::swap(SignExtendingValue, Zero);
1207
1208 // If we should not perform sign-extension then we must add/or/subtract zero.
1209 if (!match(Zero, m_Zero()))
1210 return nullptr;
1211 // Otherwise, it should be some constant, left-shifted by the same NBits we
1212 // had in `lshr`. Said left-shift can also be appropriately extended.
1213 // Again, we must look past zero-ext when looking for NBits.
1214 SkipExtInMagic(SignExtendingValue);
1215 Constant *SignExtendingValueBaseConstant;
1216 if (!match(SignExtendingValue,
1217 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1218 m_ZExtOrSelf(m_Specific(NBits)))))
1219 return nullptr;
1220 // If we `sub`, then the constant should be one, else it should be all-ones.
1221 if (I.getOpcode() == Instruction::Sub
1222 ? !match(SignExtendingValueBaseConstant, m_One())
1223 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1224 return nullptr;
1225
1226 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1227 Extract->getName() + ".sext");
1228 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1229 if (!HadTrunc)
1230 return NewAShr;
1231
1232 Builder.Insert(NewAShr);
1233 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1234}
1235
1236/// This is a specialization of a more general transform from
1237/// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
1238/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1239static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1240 InstCombiner::BuilderTy &Builder) {
1241 // TODO: Also handle mul by doubling the shift amount?
1242 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", 1244
, __extension__ __PRETTY_FUNCTION__))
1243 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", 1244
, __extension__ __PRETTY_FUNCTION__))
1244 "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", 1244
, __extension__ __PRETTY_FUNCTION__))
;
1245 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1246 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1247 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1248 return nullptr;
1249
1250 Value *X, *Y, *ShAmt;
1251 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1252 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1253 return nullptr;
1254
1255 // No-wrap propagates only when all ops have no-wrap.
1256 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1257 Op1->hasNoSignedWrap();
1258 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1259 Op1->hasNoUnsignedWrap();
1260
1261 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1262 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1263 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1264 NewI->setHasNoSignedWrap(HasNSW);
1265 NewI->setHasNoUnsignedWrap(HasNUW);
1266 }
1267 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1268 NewShl->setHasNoSignedWrap(HasNSW);
1269 NewShl->setHasNoUnsignedWrap(HasNUW);
1270 return NewShl;
1271}
1272
1273Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1274 if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1275 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1276 SQ.getWithInstruction(&I)))
1277 return replaceInstUsesWith(I, V);
1278
1279 if (SimplifyAssociativeOrCommutative(I))
1280 return &I;
1281
1282 if (Instruction *X = foldVectorBinop(I))
1283 return X;
1284
1285 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1286 return Phi;
1287
1288 // (A*B)+(A*C) -> A*(B+C) etc
1289 if (Value *V = SimplifyUsingDistributiveLaws(I))
1290 return replaceInstUsesWith(I, V);
1291
1292 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1293 return R;
1294
1295 if (Instruction *X = foldAddWithConstant(I))
1296 return X;
1297
1298 if (Instruction *X = foldNoWrapAdd(I, Builder))
1299 return X;
1300
1301 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1302 Type *Ty = I.getType();
1303 if (Ty->isIntOrIntVectorTy(1))
1304 return BinaryOperator::CreateXor(LHS, RHS);
1305
1306 // X + X --> X << 1
1307 if (LHS == RHS) {
1308 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1309 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1310 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1311 return Shl;
1312 }
1313
1314 Value *A, *B;
1315 if (match(LHS, m_Neg(m_Value(A)))) {
1316 // -A + -B --> -(A + B)
1317 if (match(RHS, m_Neg(m_Value(B))))
1318 return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1319
1320 // -A + B --> B - A
1321 return BinaryOperator::CreateSub(RHS, A);
1322 }
1323
1324 // A + -B --> A - B
1325 if (match(RHS, m_Neg(m_Value(B))))
1326 return BinaryOperator::CreateSub(LHS, B);
1327
1328 if (Value *V = checkForNegativeOperand(I, Builder))
1329 return replaceInstUsesWith(I, V);
1330
1331 // (A + 1) + ~B --> A - B
1332 // ~B + (A + 1) --> A - B
1333 // (~B + A) + 1 --> A - B
1334 // (A + ~B) + 1 --> A - B
1335 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1336 match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1337 return BinaryOperator::CreateSub(A, B);
1338
1339 // (A + RHS) + RHS --> A + (RHS << 1)
1340 if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1341 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1342
1343 // LHS + (A + LHS) --> A + (LHS << 1)
1344 if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1345 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1346
1347 {
1348 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1349 Constant *C1, *C2;
1350 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1351 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1352 (LHS->hasOneUse() || RHS->hasOneUse())) {
1353 Value *Sub = Builder.CreateSub(A, B);
1354 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1355 }
1356 }
1357
1358 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1359 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1360
1361 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1362 const APInt *C1, *C2;
1363 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1364 APInt one(C2->getBitWidth(), 1);
1365 APInt minusC1 = -(*C1);
1366 if (minusC1 == (one << *C2)) {
1367 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1368 return BinaryOperator::CreateSRem(RHS, NewRHS);
1369 }
1370 }
1371
1372 // A+B --> A|B iff A and B have no bits set in common.
1373 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1374 return BinaryOperator::CreateOr(LHS, RHS);
1375
1376 // add (select X 0 (sub n A)) A --> select X A n
1377 {
1378 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1379 Value *A = RHS;
1380 if (!SI) {
1381 SI = dyn_cast<SelectInst>(RHS);
1382 A = LHS;
1383 }
1384 if (SI && SI->hasOneUse()) {
1385 Value *TV = SI->getTrueValue();
1386 Value *FV = SI->getFalseValue();
1387 Value *N;
1388
1389 // Can we fold the add into the argument of the select?
1390 // We check both true and false select arguments for a matching subtract.
1391 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1392 // Fold the add into the true select value.
1393 return SelectInst::Create(SI->getCondition(), N, A);
1394
1395 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1396 // Fold the add into the false select value.
1397 return SelectInst::Create(SI->getCondition(), A, N);
1398 }
1399 }
1400
1401 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1402 return Ext;
1403
1404 // (add (xor A, B) (and A, B)) --> (or A, B)
1405 // (add (and A, B) (xor A, B)) --> (or A, B)
1406 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1407 m_c_And(m_Deferred(A), m_Deferred(B)))))
1408 return BinaryOperator::CreateOr(A, B);
1409
1410 // (add (or A, B) (and A, B)) --> (add A, B)
1411 // (add (and A, B) (or A, B)) --> (add A, B)
1412 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1413 m_c_And(m_Deferred(A), m_Deferred(B))))) {
1414 // Replacing operands in-place to preserve nuw/nsw flags.
1415 replaceOperand(I, 0, A);
1416 replaceOperand(I, 1, B);
1417 return &I;
1418 }
1419
1420 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1421 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1422 // computeKnownBits.
1423 bool Changed = false;
1424 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1425 Changed = true;
1426 I.setHasNoSignedWrap(true);
1427 }
1428 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1429 Changed = true;
1430 I.setHasNoUnsignedWrap(true);
1431 }
1432
1433 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1434 return V;
1435
1436 if (Instruction *V =
1437 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1438 return V;
1439
1440 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1441 return SatAdd;
1442
1443 // usub.sat(A, B) + B => umax(A, B)
1444 if (match(&I, m_c_BinOp(
1445 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1446 m_Deferred(B)))) {
1447 return replaceInstUsesWith(I,
1448 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1449 }
1450
1451 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1452 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1453 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1454 haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1455 return replaceInstUsesWith(
1456 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1457 {Builder.CreateOr(A, B)}));
1458
1459 return Changed ? &I : nullptr;
1460}
1461
1462/// Eliminate an op from a linear interpolation (lerp) pattern.
1463static Instruction *factorizeLerp(BinaryOperator &I,
1464 InstCombiner::BuilderTy &Builder) {
1465 Value *X, *Y, *Z;
1466 if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1467 m_OneUse(m_FSub(m_FPOne(),
1468 m_Value(Z))))),
1469 m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1470 return nullptr;
1471
1472 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1473 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1474 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1475 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1476}
1477
1478/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1479static Instruction *factorizeFAddFSub(BinaryOperator &I,
1480 InstCombiner::BuilderTy &Builder) {
1481 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", 1482
, __extension__ __PRETTY_FUNCTION__))
1482 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", 1482
, __extension__ __PRETTY_FUNCTION__))
;
1483 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", 1484
, __extension__ __PRETTY_FUNCTION__))
1484 "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", 1484
, __extension__ __PRETTY_FUNCTION__))
;
1485
1486 if (Instruction *Lerp = factorizeLerp(I, Builder))
1487 return Lerp;
1488
1489 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1490 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1491 return nullptr;
1492
1493 Value *X, *Y, *Z;
1494 bool IsFMul;
1495 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1496 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1497 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1498 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1499 IsFMul = true;
1500 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1501 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1502 IsFMul = false;
1503 else
1504 return nullptr;
1505
1506 // (X * Z) + (Y * Z) --> (X + Y) * Z
1507 // (X * Z) - (Y * Z) --> (X - Y) * Z
1508 // (X / Z) + (Y / Z) --> (X + Y) / Z
1509 // (X / Z) - (Y / Z) --> (X - Y) / Z
1510 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1511 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1512 : Builder.CreateFSubFMF(X, Y, &I);
1513
1514 // Bail out if we just created a denormal constant.
1515 // TODO: This is copied from a previous implementation. Is it necessary?
1516 const APFloat *C;
1517 if (match(XY, m_APFloat(C)) && !C->isNormal())
1518 return nullptr;
1519
1520 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1521 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1522}
1523
1524Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1525 if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1526 I.getFastMathFlags(),
1527 SQ.getWithInstruction(&I)))
1528 return replaceInstUsesWith(I, V);
1529
1530 if (SimplifyAssociativeOrCommutative(I))
1531 return &I;
1532
1533 if (Instruction *X = foldVectorBinop(I))
1534 return X;
1535
1536 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1537 return Phi;
1538
1539 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1540 return FoldedFAdd;
1541
1542 // (-X) + Y --> Y - X
1543 Value *X, *Y;
1544 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1545 return BinaryOperator::CreateFSubFMF(Y, X, &I);
1546
1547 // Similar to above, but look through fmul/fdiv for the negated term.
1548 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1549 Value *Z;
1550 if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1551 m_Value(Z)))) {
1552 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1553 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1554 }
1555 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1556 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1557 if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1558 m_Value(Z))) ||
1559 match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1560 m_Value(Z)))) {
1561 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1562 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1563 }
1564
1565 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1566 // integer add followed by a promotion.
1567 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1568 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1569 Value *LHSIntVal = LHSConv->getOperand(0);
1570 Type *FPType = LHSConv->getType();
1571
1572 // TODO: This check is overly conservative. In many cases known bits
1573 // analysis can tell us that the result of the addition has less significant
1574 // bits than the integer type can hold.
1575 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1576 Type *FScalarTy = FTy->getScalarType();
1577 Type *IScalarTy = ITy->getScalarType();
1578
1579 // Do we have enough bits in the significand to represent the result of
1580 // the integer addition?
1581 unsigned MaxRepresentableBits =
1582 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1583 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1584 };
1585
1586 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1587 // ... if the constant fits in the integer value. This is useful for things
1588 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1589 // requires a constant pool load, and generally allows the add to be better
1590 // instcombined.
1591 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1592 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1593 Constant *CI =
1594 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1595 if (LHSConv->hasOneUse() &&
1596 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1597 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1598 // Insert the new integer add.
1599 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1600 return new SIToFPInst(NewAdd, I.getType());
1601 }
1602 }
1603
1604 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1605 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1606 Value *RHSIntVal = RHSConv->getOperand(0);
1607 // It's enough to check LHS types only because we require int types to
1608 // be the same for this transform.
1609 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1610 // Only do this if x/y have the same type, if at least one of them has a
1611 // single use (so we don't increase the number of int->fp conversions),
1612 // and if the integer add will not overflow.
1613 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1614 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1615 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1616 // Insert the new integer add.
1617 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1618 return new SIToFPInst(NewAdd, I.getType());
1619 }
1620 }
1621 }
1622 }
1623
1624 // Handle specials cases for FAdd with selects feeding the operation
1625 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1626 return replaceInstUsesWith(I, V);
1627
1628 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1629 if (Instruction *F = factorizeFAddFSub(I, Builder))
1630 return F;
1631
1632 // Try to fold fadd into start value of reduction intrinsic.
1633 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1634 m_AnyZeroFP(), m_Value(X))),
1635 m_Value(Y)))) {
1636 // fadd (rdx 0.0, X), Y --> rdx Y, X
1637 return replaceInstUsesWith(
1638 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1639 {X->getType()}, {Y, X}, &I));
1640 }
1641 const APFloat *StartC, *C;
1642 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1643 m_APFloat(StartC), m_Value(X)))) &&
1644 match(RHS, m_APFloat(C))) {
1645 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1646 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1647 return replaceInstUsesWith(
1648 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1649 {X->getType()}, {NewStartC, X}, &I));
1650 }
1651
1652 // (X * MulC) + X --> X * (MulC + 1.0)
1653 Constant *MulC;
1654 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1655 m_Deferred(X)))) {
1656 MulC = ConstantExpr::getFAdd(MulC, ConstantFP::get(I.getType(), 1.0));
1657 return BinaryOperator::CreateFMulFMF(X, MulC, &I);
1658 }
1659
1660 if (Value *V = FAddCombine(Builder).simplify(&I))
1661 return replaceInstUsesWith(I, V);
1662 }
1663
1664 return nullptr;
1665}
1666
1667/// Optimize pointer differences into the same array into a size. Consider:
1668/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1669/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1670Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
1671 Type *Ty, bool IsNUW) {
1672 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1673 // this.
1674 bool Swapped = false;
1675 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1676 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1677 std::swap(LHS, RHS);
1678 Swapped = true;
1679 }
1680
1681 // Require at least one GEP with a common base pointer on both sides.
1682 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1683 // (gep X, ...) - X
1684 if (LHSGEP->getOperand(0) == RHS) {
1685 GEP1 = LHSGEP;
1686 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1687 // (gep X, ...) - (gep X, ...)
1688 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1689 RHSGEP->getOperand(0)->stripPointerCasts()) {
1690 GEP1 = LHSGEP;
1691 GEP2 = RHSGEP;
1692 }
1693 }
1694 }
1695
1696 if (!GEP1)
1697 return nullptr;
1698
1699 if (GEP2) {
1700 // (gep X, ...) - (gep X, ...)
1701 //
1702 // Avoid duplicating the arithmetic if there are more than one non-constant
1703 // indices between the two GEPs and either GEP has a non-constant index and
1704 // multiple users. If zero non-constant index, the result is a constant and
1705 // there is no duplication. If one non-constant index, the result is an add
1706 // or sub with a constant, which is no larger than the original code, and
1707 // there's no duplicated arithmetic, even if either GEP has multiple
1708 // users. If more than one non-constant indices combined, as long as the GEP
1709 // with at least one non-constant index doesn't have multiple users, there
1710 // is no duplication.
1711 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1712 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1713 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1714 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1715 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1716 return nullptr;
1717 }
1718 }
1719
1720 // Emit the offset of the GEP and an intptr_t.
1721 Value *Result = EmitGEPOffset(GEP1);
1722
1723 // If this is a single inbounds GEP and the original sub was nuw,
1724 // then the final multiplication is also nuw.
1725 if (auto *I = dyn_cast<Instruction>(Result))
1726 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1727 I->getOpcode() == Instruction::Mul)
1728 I->setHasNoUnsignedWrap();
1729
1730 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
1731 // If both GEPs are inbounds, then the subtract does not have signed overflow.
1732 if (GEP2) {
1733 Value *Offset = EmitGEPOffset(GEP2);
1734 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
1735 GEP1->isInBounds() && GEP2->isInBounds());
1736 }
1737
1738 // If we have p - gep(p, ...) then we have to negate the result.
1739 if (Swapped)
1740 Result = Builder.CreateNeg(Result, "diff.neg");
1741
1742 return Builder.CreateIntCast(Result, Ty, true);
1743}
1744
1745Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
1746 if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1747 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1748 SQ.getWithInstruction(&I)))
1749 return replaceInstUsesWith(I, V);
1750
1751 if (Instruction *X = foldVectorBinop(I))
1752 return X;
1753
1754 if (Instruction *Phi = foldBinopWithPhiOperands(I))
1755 return Phi;
1756
1757 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1758
1759 // If this is a 'B = x-(-A)', change to B = x+A.
1760 // We deal with this without involving Negator to preserve NSW flag.
1761 if (Value *V = dyn_castNegVal(Op1)) {
1762 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1763
1764 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1765 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", 1766
, __extension__ __PRETTY_FUNCTION__))
1766 "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", 1766
, __extension__ __PRETTY_FUNCTION__))
;
1767 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1768 Res->setHasNoSignedWrap(true);
1769 } else {
1770 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1771 Res->setHasNoSignedWrap(true);
1772 }
1773
1774 return Res;
1775 }
1776
1777 // Try this before Negator to preserve NSW flag.
1778 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1779 return R;
1780
1781 Constant *C;
1782 if (match(Op0, m_ImmConstant(C))) {
1783 Value *X;
1784 Constant *C2;
1785
1786 // C-(X+C2) --> (C-C2)-X
1787 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
1788 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1789 }
1790
1791 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
1792 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1793 return Ext;
1794
1795 bool Changed = false;
1796 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1797 Changed = true;
1798 I.setHasNoSignedWrap(true);
1799 }
1800 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1801 Changed = true;
1802 I.setHasNoUnsignedWrap(true);
1803 }
1804
1805 return Changed ? &I : nullptr;
1806 };
1807
1808 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
1809 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
1810 // a pure negation used by a select that looks like abs/nabs.
1811 bool IsNegation = match(Op0, m_ZeroInt());
1812 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
1813 const Instruction *UI = dyn_cast<Instruction>(U);
1814 if (!UI)
1815 return false;
1816 return match(UI,
1817 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
1818 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
1819 })) {
1820 if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
1821 return BinaryOperator::CreateAdd(NegOp1, Op0);
1822 }
1823 if (IsNegation)
1824 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
1825
1826 // (A*B)-(A*C) -> A*(B-C) etc
1827 if (Value *V = SimplifyUsingDistributiveLaws(I))
1828 return replaceInstUsesWith(I, V);
1829
1830 if (I.getType()->isIntOrIntVectorTy(1))
1831 return BinaryOperator::CreateXor(Op0, Op1);
1832
1833 // Replace (-1 - A) with (~A).
1834 if (match(Op0, m_AllOnes()))
1835 return BinaryOperator::CreateNot(Op1);
1836
1837 // (X + -1) - Y --> ~Y + X
1838 Value *X, *Y;
1839 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1840 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1841
1842 // Reassociate sub/add sequences to create more add instructions and
1843 // reduce dependency chains:
1844 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
1845 Value *Z;
1846 if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
1847 m_Value(Z))))) {
1848 Value *XZ = Builder.CreateAdd(X, Z);
1849 Value *YW = Builder.CreateAdd(Y, Op1);
1850 return BinaryOperator::CreateSub(XZ, YW);
1851 }
1852
1853 // ((X - Y) - Op1) --> X - (Y + Op1)
1854 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
1855 Value *Add = Builder.CreateAdd(Y, Op1);
1856 return BinaryOperator::CreateSub(X, Add);
1857 }
1858
1859 // (~X) - (~Y) --> Y - X
1860 // This is placed after the other reassociations and explicitly excludes a
1861 // sub-of-sub pattern to avoid infinite looping.
1862 if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
1863 isFreeToInvert(Op1, Op1->hasOneUse()) &&
1864 !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
1865 Value *NotOp0 = Builder.CreateNot(Op0);
1866 Value *NotOp1 = Builder.CreateNot(Op1);
1867 return BinaryOperator::CreateSub(NotOp1, NotOp0);
1868 }
1869
1870 auto m_AddRdx = [](Value *&Vec) {
1871 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
1872 };
1873 Value *V0, *V1;
1874 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
1875 V0->getType() == V1->getType()) {
1876 // Difference of sums is sum of differences:
1877 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
1878 Value *Sub = Builder.CreateSub(V0, V1);
1879 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
1880 {Sub->getType()}, {Sub});
1881 return replaceInstUsesWith(I, Rdx);
1882 }
1883
1884 if (Constant *C = dyn_cast<Constant>(Op0)) {
1885 Value *X;
1886 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1887 // C - (zext bool) --> bool ? C - 1 : C
1888 return SelectInst::Create(X, InstCombiner::SubOne(C), C);
1889 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1890 // C - (sext bool) --> bool ? C + 1 : C
1891 return SelectInst::Create(X, InstCombiner::AddOne(C), C);
1892
1893 // C - ~X == X + (1+C)
1894 if (match(Op1, m_Not(m_Value(X))))
1895 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
1896
1897 // Try to fold constant sub into select arguments.
1898 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1899 if (Instruction *R = FoldOpIntoSelect(I, SI))
1900 return R;
1901
1902 // Try to fold constant sub into PHI values.
1903 if (PHINode *PN = dyn_cast<PHINode>(Op1))
1904 if (Instruction *R = foldOpIntoPhi(I, PN))
1905 return R;
1906
1907 Constant *C2;
1908
1909 // C-(C2-X) --> X+(C-C2)
1910 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
1911 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
1912 }
1913
1914 const APInt *Op0C;
1915 if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) {
1916 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1917 // zero.
1918 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1919 if ((*Op0C | RHSKnown.Zero).isAllOnes())
1920 return BinaryOperator::CreateXor(Op1, Op0);
1921 }
1922
1923 {
1924 Value *Y;
1925 // X-(X+Y) == -Y X-(Y+X) == -Y
1926 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1927 return BinaryOperator::CreateNeg(Y);
1928
1929 // (X-Y)-X == -Y
1930 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1931 return BinaryOperator::CreateNeg(Y);
1932 }
1933
1934 // (sub (or A, B) (and A, B)) --> (xor A, B)
1935 {
1936 Value *A, *B;
1937 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1938 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1939 return BinaryOperator::CreateXor(A, B);
1940 }
1941
1942 // (sub (add A, B) (or A, B)) --> (and A, B)
1943 {
1944 Value *A, *B;
1945 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1946 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1947 return BinaryOperator::CreateAnd(A, B);
1948 }
1949
1950 // (sub (add A, B) (and A, B)) --> (or A, B)
1951 {
1952 Value *A, *B;
1953 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1954 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
1955 return BinaryOperator::CreateOr(A, B);
1956 }
1957
1958 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
1959 {
1960 Value *A, *B;
1961 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1962 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1963 (Op0->hasOneUse() || Op1->hasOneUse()))
1964 return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
1965 }
1966
1967 // (sub (or A, B), (xor A, B)) --> (and A, B)
1968 {
1969 Value *A, *B;
1970 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1971 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1972 return BinaryOperator::CreateAnd(A, B);
1973 }
1974
1975 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
1976 {
1977 Value *A, *B;
1978 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1979 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1980 (Op0->hasOneUse() || Op1->hasOneUse()))
1981 return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
1982 }
1983
1984 {
1985 Value *Y;
1986 // ((X | Y) - X) --> (~X & Y)
1987 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1988 return BinaryOperator::CreateAnd(
1989 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1990 }
1991
1992 {
1993 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
1994 Value *X;
1995 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
1996 m_OneUse(m_Neg(m_Value(X))))))) {
1997 return BinaryOperator::CreateNeg(Builder.CreateAnd(
1998 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
1999 }
2000 }
2001
2002 {
2003 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2004 Constant *C;
2005 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2006 return BinaryOperator::CreateNeg(
2007 Builder.CreateAnd(Op1, Builder.CreateNot(C)));
2008 }
2009 }
2010
2011 {
2012 // sub(add(X,Y), s/umin(X,Y)) --> s/umax(X,Y)
2013 // sub(add(X,Y), s/umax(X,Y)) --> s/umin(X,Y)
2014 // TODO: generalize to sub(add(Z,Y),umin(X,Y)) --> add(Z,usub.sat(Y,X))?
2015 if (auto *II = dyn_cast<MinMaxIntrinsic>(Op1)) {
2016 Value *X = II->getLHS();
2017 Value *Y = II->getRHS();
2018 if (match(Op0, m_c_Add(m_Specific(X), m_Specific(Y))) &&
2019 (Op0->hasOneUse() || Op1->hasOneUse())) {
2020 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(II->getIntrinsicID());
2021 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(InvID, X, Y);
2022 return replaceInstUsesWith(I, InvMaxMin);
2023 }
2024 }
2025 }
2026
2027 {
2028 // If we have a subtraction between some value and a select between
2029 // said value and something else, sink subtraction into select hands, i.e.:
2030 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2031 // ->
2032 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2033 // or
2034 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2035 // ->
2036 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2037 // This will result in select between new subtraction and 0.
2038 auto SinkSubIntoSelect =
2039 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2040 auto SubBuilder) -> Instruction * {
2041 Value *Cond, *TrueVal, *FalseVal;
2042 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2043 m_Value(FalseVal)))))
2044 return nullptr;
2045 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2046 return nullptr;
2047 // While it is really tempting to just create two subtractions and let
2048 // InstCombine fold one of those to 0, it isn't possible to do so
2049 // because of worklist visitation order. So ugly it is.
2050 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2051 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2052 Constant *Zero = Constant::getNullValue(Ty);
2053 SelectInst *NewSel =
2054 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2055 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2056 // Preserve prof metadata if any.
2057 NewSel->copyMetadata(cast<Instruction>(*Select));
2058 return NewSel;
2059 };
2060 if (Instruction *NewSel = SinkSubIntoSelect(
2061 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2062 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2063 return Builder->CreateSub(OtherHandOfSelect,
2064 /*OtherHandOfSub=*/Op1);
2065 }))
2066 return NewSel;
2067 if (Instruction *NewSel = SinkSubIntoSelect(
2068 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2069 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2070 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2071 OtherHandOfSelect);
2072 }))
2073 return NewSel;
2074 }
2075
2076 // (X - (X & Y)) --> (X & ~Y)
2077 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2078 (Op1->hasOneUse() || isa<Constant>(Y)))
2079 return BinaryOperator::CreateAnd(
2080 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2081
2082 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2083 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2084 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2085 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2086 // As long as Y is freely invertible, this will be neutral or a win.
2087 // Note: We don't generate the inverse max/min, just create the 'not' of
2088 // it and let other folds do the rest.
2089 if (match(Op0, m_Not(m_Value(X))) &&
2090 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2091 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2092 Value *Not = Builder.CreateNot(Op1);
2093 return BinaryOperator::CreateSub(Not, X);
2094 }
2095 if (match(Op1, m_Not(m_Value(X))) &&
2096 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2097 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2098 Value *Not = Builder.CreateNot(Op0);
2099 return BinaryOperator::CreateSub(X, Not);
2100 }
2101
2102 // Optimize pointer differences into the same array into a size. Consider:
2103 // &A[10] - &A[0]: we should compile this to "10".
2104 Value *LHSOp, *RHSOp;
2105 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2106 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2107 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2108 I.hasNoUnsignedWrap()))
2109 return replaceInstUsesWith(I, Res);
2110
2111 // trunc(p)-trunc(q) -> trunc(p-q)
2112 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2113 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2114 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2115 /* IsNUW */ false))
2116 return replaceInstUsesWith(I, Res);
2117
2118 // Canonicalize a shifty way to code absolute value to the common pattern.
2119 // There are 2 potential commuted variants.
2120 // We're relying on the fact that we only do this transform when the shift has
2121 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2122 // instructions).
2123 Value *A;
2124 const APInt *ShAmt;
2125 Type *Ty = I.getType();
2126 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2127 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2128 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2129 // B = ashr i32 A, 31 ; smear the sign bit
2130 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2131 // --> (A < 0) ? -A : A
2132 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2133 // Copy the nuw/nsw flags from the sub to the negate.
2134 Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2135 I.hasNoSignedWrap());
2136 return SelectInst::Create(Cmp, Neg, A);
2137 }
2138
2139 // If we are subtracting a low-bit masked subset of some value from an add
2140 // of that same value with no low bits changed, that is clearing some low bits
2141 // of the sum:
2142 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2143 const APInt *AddC, *AndC;
2144 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2145 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2146 unsigned BitWidth = Ty->getScalarSizeInBits();
2147 unsigned Cttz = AddC->countTrailingZeros();
2148 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2149 if ((HighMask & *AndC).isZero())
2150 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2151 }
2152
2153 if (Instruction *V =
2154 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2155 return V;
2156
2157 // X - usub.sat(X, Y) => umin(X, Y)
2158 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2159 m_Value(Y)))))
2160 return replaceInstUsesWith(
2161 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2162
2163 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2164 // TODO: The one-use restriction is not strictly necessary, but it may
2165 // require improving other pattern matching and/or codegen.
2166 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2167 return replaceInstUsesWith(
2168 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2169
2170 // Op0 - umin(X, Op0) --> usub.sat(Op0, X)
2171 if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
2172 return replaceInstUsesWith(
2173 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
2174
2175 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2176 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2177 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2178 return BinaryOperator::CreateNeg(USub);
2179 }
2180
2181 // umin(X, Op1) - Op1 --> 0 - usub.sat(Op1, X)
2182 if (match(Op0, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op1))))) {
2183 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op1, X});
2184 return BinaryOperator::CreateNeg(USub);
2185 }
2186
2187 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2188 if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
2189 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2190 return replaceInstUsesWith(
2191 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2192 {Builder.CreateNot(X)}));
2193
2194 return TryToNarrowDeduceFlags();
2195}
2196
2197/// This eliminates floating-point negation in either 'fneg(X)' or
2198/// 'fsub(-0.0, X)' form by combining into a constant operand.
2199static Instruction *foldFNegIntoConstant(Instruction &I) {
2200 // This is limited with one-use because fneg is assumed better for
2201 // reassociation and cheaper in codegen than fmul/fdiv.
2202 // TODO: Should the m_OneUse restriction be removed?
2203 Instruction *FNegOp;
2204 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2205 return nullptr;
2206
2207 Value *X;
2208 Constant *C;
2209
2210 // Fold negation into constant operand.
2211 // -(X * C) --> X * (-C)
2212 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2213 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
2214 // -(X / C) --> X / (-C)
2215 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2216 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
2217 // -(C / X) --> (-C) / X
2218 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) {
2219 Instruction *FDiv =
2220 BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
2221
2222 // Intersect 'nsz' and 'ninf' because those special value exceptions may not
2223 // apply to the fdiv. Everything else propagates from the fneg.
2224 // TODO: We could propagate nsz/ninf from fdiv alone?
2225 FastMathFlags FMF = I.getFastMathFlags();
2226 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2227 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2228 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2229 return FDiv;
2230 }
2231 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2232 // -(X + C) --> -X + -C --> -C - X
2233 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2234 return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I);
2235
2236 return nullptr;
2237}
2238
2239static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2240 InstCombiner::BuilderTy &Builder) {
2241 Value *FNeg;
2242 if (!match(&I, m_FNeg(m_Value(FNeg))))
2243 return nullptr;
2244
2245 Value *X, *Y;
2246 if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2247 return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2248
2249 if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2250 return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2251
2252 return nullptr;
2253}
2254
2255Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2256 Value *Op = I.getOperand(0);
2257
2258 if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
2259 getSimplifyQuery().getWithInstruction(&I)))
2260 return replaceInstUsesWith(I, V);
2261
2262 if (Instruction *X = foldFNegIntoConstant(I))
2263 return X;
2264
2265 Value *X, *Y;
2266
2267 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2268 if (I.hasNoSignedZeros() &&
2269 match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2270 return BinaryOperator::CreateFSubFMF(Y, X, &I);
2271
2272 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2273 return R;
2274
2275 // Try to eliminate fneg if at least 1 arm of the select is negated.
2276 Value *Cond;
2277 if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
2278 // Unlike most transforms, this one is not safe to propagate nsz unless
2279 // it is present on the original select. (We are conservatively intersecting
2280 // the nsz flags from the select and root fneg instruction.)
2281 auto propagateSelectFMF = [&](SelectInst *S, bool CommonOperand) {
2282 S->copyFastMathFlags(&I);
2283 if (auto *OldSel = dyn_cast<SelectInst>(Op))
2284 if (!OldSel->hasNoSignedZeros() && !CommonOperand &&
2285 !isGuaranteedNotToBeUndefOrPoison(OldSel->getCondition()))
2286 S->setHasNoSignedZeros(false);
2287 };
2288 // -(Cond ? -P : Y) --> Cond ? P : -Y
2289 Value *P;
2290 if (match(X, m_FNeg(m_Value(P)))) {
2291 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2292 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2293 propagateSelectFMF(NewSel, P == Y);
2294 return NewSel;
2295 }
2296 // -(Cond ? X : -P) --> Cond ? -X : P
2297 if (match(Y, m_FNeg(m_Value(P)))) {
2298 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2299 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2300 propagateSelectFMF(NewSel, P == X);
2301 return NewSel;
2302 }
2303 }
2304
2305 return nullptr;
2306}
2307
2308Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2309 if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
2310 I.getFastMathFlags(),
2311 getSimplifyQuery().getWithInstruction(&I)))
2312 return replaceInstUsesWith(I, V);
2313
2314 if (Instruction *X = foldVectorBinop(I))
2315 return X;
2316
2317 if (Instruction *Phi = foldBinopWithPhiOperands(I))
2318 return Phi;
2319
2320 // Subtraction from -0.0 is the canonical form of fneg.
2321 // fsub -0.0, X ==> fneg X
2322 // fsub nsz 0.0, X ==> fneg nsz X
2323 //
2324 // FIXME This matcher does not respect FTZ or DAZ yet:
2325 // fsub -0.0, Denorm ==> +-0
2326 // fneg Denorm ==> -Denorm
2327 Value *Op;
2328 if (match(&I, m_FNeg(m_Value(Op))))
2329 return UnaryOperator::CreateFNegFMF(Op, &I);
2330
2331 if (Instruction *X = foldFNegIntoConstant(I))
2332 return X;
2333
2334 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2335 return R;
2336
2337 Value *X, *Y;
2338 Constant *C;
2339
2340 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2341 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2342 // Canonicalize to fadd to make analysis easier.
2343 // This can also help codegen because fadd is commutative.
2344 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2345 // killed later. We still limit that particular transform with 'hasOneUse'
2346 // because an fneg is assumed better/cheaper than a generic fsub.
2347 if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2348 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2349 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2350 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2351 }
2352 }
2353
2354 // (-X) - Op1 --> -(X + Op1)
2355 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2356 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2357 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2358 return UnaryOperator::CreateFNegFMF(FAdd, &I);
2359 }
2360
2361 if (isa<Constant>(Op0))
2362 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2363 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2364 return NV;
2365
2366 // X - C --> X + (-C)
2367 // But don't transform constant expressions because there's an inverse fold
2368 // for X + (-Y) --> X - Y.
2369 if (match(Op1, m_ImmConstant(C)))
2370 return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
2371
2372 // X - (-Y) --> X + Y
2373 if (match(Op1, m_FNeg(m_Value(Y))))
2374 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2375
2376 // Similar to above, but look through a cast of the negated value:
2377 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2378 Type *Ty = I.getType();
2379 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2380 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2381
2382 // X - (fpext(-Y)) --> X + fpext(Y)
2383 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2384 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2385
2386 // Similar to above, but look through fmul/fdiv of the negated value:
2387 // Op0 - (-X * Y) --> Op0 + (X * Y)
2388 // Op0 - (Y * -X) --> Op0 + (X * Y)
2389 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2390 Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2391 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2392 }
2393 // Op0 - (-X / Y) --> Op0 + (X / Y)
2394 // Op0 - (X / -Y) --> Op0 + (X / Y)
2395 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2396 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2397 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2398 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2399 }
2400
2401 // Handle special cases for FSub with selects feeding the operation
2402 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2403 return replaceInstUsesWith(I, V);
2404
2405 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2406 // (Y - X) - Y --> -X
2407 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2408 return UnaryOperator::CreateFNegFMF(X, &I);
2409
2410 // Y - (X + Y) --> -X
2411 // Y - (Y + X) --> -X
2412 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2413 return UnaryOperator::CreateFNegFMF(X, &I);
2414
2415 // (X * C) - X --> X * (C - 1.0)
2416 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2417 Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
2418 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2419 }
2420 // X - (X * C) --> X * (1.0 - C)
2421 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2422 Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
2423 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2424 }
2425
2426 // Reassociate fsub/fadd sequences to create more fadd instructions and
2427 // reduce dependency chains:
2428 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2429 Value *Z;
2430 if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2431 m_Value(Z))))) {
2432 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2433 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2434 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2435 }
2436
2437 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2438 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2439 m_Value(Vec)));
2440 };
2441 Value *A0, *A1, *V0, *V1;
2442 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2443 V0->getType() == V1->getType()) {
2444 // Difference of sums is sum of differences:
2445 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2446 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2447 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2448 {Sub->getType()}, {A0, Sub}, &I);
2449 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2450 }
2451
2452 if (Instruction *F = factorizeFAddFSub(I, Builder))
2453 return F;
2454
2455 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2456 // functionality has been subsumed by simple pattern matching here and in
2457 // InstSimplify. We should let a dedicated reassociation pass handle more
2458 // complex pattern matching and remove this from InstCombine.
2459 if (Value *V = FAddCombine(Builder).simplify(&I))
2460 return replaceInstUsesWith(I, V);
2461
2462 // (X - Y) - Op1 --> X - (Y + Op1)
2463 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2464 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2465 return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
2466 }
2467 }
2468
2469 return nullptr;
2470}

/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/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() { return class_match<PoisonValue>(); }
140
141/// Match an arbitrary Constant and ignore it.
142inline class_match<Constant> m_Constant() { return class_match<Constant>(); }
143
144/// Match an arbitrary ConstantInt and ignore it.
145inline class_match<ConstantInt> m_ConstantInt() {
146 return class_match<ConstantInt>();
147}
148
149/// Match an arbitrary ConstantFP and ignore it.
150inline class_match<ConstantFP> m_ConstantFP() {
151 return class_match<ConstantFP>();
152}
153
154/// Match an arbitrary ConstantExpr and ignore it.
155inline class_match<ConstantExpr> m_ConstantExpr() {
156 return class_match<ConstantExpr>();
157}
158
159/// Match an arbitrary basic block value and ignore it.
160inline class_match<BasicBlock> m_BasicBlock() {
161 return class_match<BasicBlock>();
162}
163
164/// Inverting matcher
165template <typename Ty> struct match_unless {
166 Ty M;
167
168 match_unless(const Ty &Matcher) : M(Matcher) {}
169
170 template <typename ITy> bool match(ITy *V) { return !M.match(V); }
171};
172
173/// Match if the inner matcher does *NOT* match.
174template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) {
175 return match_unless<Ty>(M);
176}
177
178/// Matching combinators
179template <typename LTy, typename RTy> struct match_combine_or {
180 LTy L;
181 RTy R;
182
183 match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
184
185 template <typename ITy> bool match(ITy *V) {
186 if (L.match(V))
17
Taking false branch
187 return true;
188 if (R.match(V))
18
Calling 'match_combine_and::match'
31
Returning from 'match_combine_and::match'
32
Taking true branch
189 return true;
190 return false;
191 }
192};
193
194template <typename LTy, typename RTy> struct match_combine_and {
195 LTy L;
196 RTy R;
197
198 match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
199
200 template <typename ITy> bool match(ITy *V) {
201 if (L.match(V))
19
Calling 'BinaryOp_match::match'
28
Returning from 'BinaryOp_match::match'
29
Taking true branch
202 if (R.match(V))
30
Taking true branch
203 return true;
204 return false;
205 }
206};
207
208/// Combine two pattern matchers matching L || R
209template <typename LTy, typename RTy>
210inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) {
211 return match_combine_or<LTy, RTy>(L, R);
52
Returning without writing to 'L.Op.VR'
212}
213
214/// Combine two pattern matchers matching L && R
215template <typename LTy, typename RTy>
216inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) {
217 return match_combine_and<LTy, RTy>(L, R);
218}
219
220struct apint_match {
221 const APInt *&Res;
222 bool AllowUndef;
223
224 apint_match(const APInt *&Res, bool AllowUndef)
225 : Res(Res), AllowUndef(AllowUndef) {}
226
227 template <typename ITy> bool match(ITy *V) {
228 if (auto *CI = dyn_cast<ConstantInt>(V)) {
229 Res = &CI->getValue();
230 return true;
231 }
232 if (V->getType()->isVectorTy())
233 if (const auto *C = dyn_cast<Constant>(V))
234 if (auto *CI = dyn_cast_or_null<ConstantInt>(
235 C->getSplatValue(AllowUndef))) {
236 Res = &CI->getValue();
237 return true;
238 }
239 return false;
240 }
241};
242// Either constexpr if or renaming ConstantFP::getValueAPF to
243// ConstantFP::getValue is needed to do it via single template
244// function for both apint/apfloat.
245struct apfloat_match {
246 const APFloat *&Res;
247 bool AllowUndef;
248
249 apfloat_match(const APFloat *&Res, bool AllowUndef)
250 : Res(Res), AllowUndef(AllowUndef) {}
251
252 template <typename ITy> bool match(ITy *V) {
253 if (auto *CI = dyn_cast<ConstantFP>(V)) {
254 Res = &CI->getValueAPF();
255 return true;
256 }
257 if (V->getType()->isVectorTy())
258 if (const auto *C = dyn_cast<Constant>(V))
259 if (auto *CI = dyn_cast_or_null<ConstantFP>(
260 C->getSplatValue(AllowUndef))) {
261 Res = &CI->getValueAPF();
262 return true;
263 }
264 return false;
265 }
266};
267
268/// Match a ConstantInt or splatted ConstantVector, binding the
269/// specified pointer to the contained APInt.
270inline apint_match m_APInt(const APInt *&Res) {
271 // Forbid undefs by default to maintain previous behavior.
272 return apint_match(Res, /* AllowUndef */ false);
273}
274
275/// Match APInt while allowing undefs in splat vector constants.
276inline apint_match m_APIntAllowUndef(const APInt *&Res) {
277 return apint_match(Res, /* AllowUndef */ true);
278}
279
280/// Match APInt while forbidding undefs in splat vector constants.
281inline apint_match m_APIntForbidUndef(const APInt *&Res) {
282 return apint_match(Res, /* AllowUndef */ false);
283}
284
285/// Match a ConstantFP or splatted ConstantVector, binding the
286/// specified pointer to the contained APFloat.
287inline apfloat_match m_APFloat(const APFloat *&Res) {
288 // Forbid undefs by default to maintain previous behavior.
289 return apfloat_match(Res, /* AllowUndef */ false);
290}
291
292/// Match APFloat while allowing undefs in splat vector constants.
293inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) {
294 return apfloat_match(Res, /* AllowUndef */ true);
295}
296
297/// Match APFloat while forbidding undefs in splat vector constants.
298inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) {
299 return apfloat_match(Res, /* AllowUndef */ false);
300}
301
302template <int64_t Val> struct constantint_match {
303 template <typename ITy> bool match(ITy *V) {
304 if (const auto *CI = dyn_cast<ConstantInt>(V)) {
305 const APInt &CIV = CI->getValue();
306 if (Val >= 0)
307 return CIV == static_cast<uint64_t>(Val);
308 // If Val is negative, and CI is shorter than it, truncate to the right
309 // number of bits. If it is larger, then we have to sign extend. Just
310 // compare their negated values.
311 return -CIV == -Val;
312 }
313 return false;
314 }
315};
316
317/// Match a ConstantInt with a specific value.
318template <int64_t Val> inline constantint_match<Val> m_ConstantInt() {
319 return constantint_match<Val>();
320}
321
322/// This helper class is used to match constant scalars, vector splats,
323/// and fixed width vectors that satisfy a specified predicate.
324/// For fixed width vector constants, undefined elements are ignored.
325template <typename Predicate, typename ConstantVal>
326struct cstval_pred_ty : public Predicate {
327 template <typename ITy> bool match(ITy *V) {
328 if (const auto *CV = dyn_cast<ConstantVal>(V))
329 return this->isValue(CV->getValue());
330 if (const auto *VTy = dyn_cast<VectorType>(V->getType())) {
331 if (const auto *C = dyn_cast<Constant>(V)) {
332 if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue()))
333 return this->isValue(CV->getValue());
334
335 // Number of elements of a scalable vector unknown at compile time
336 auto *FVTy = dyn_cast<FixedVectorType>(VTy);
337 if (!FVTy)
338 return false;
339
340 // Non-splat vector constant: check each element for a match.
341 unsigned NumElts = FVTy->getNumElements();
342 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", 342, __extension__ __PRETTY_FUNCTION__
))
;
343 bool HasNonUndefElements = false;
344 for (unsigned i = 0; i != NumElts; ++i) {
345 Constant *Elt = C->getAggregateElement(i);
346 if (!Elt)
347 return false;
348 if (isa<UndefValue>(Elt))
349 continue;
350 auto *CV = dyn_cast<ConstantVal>(Elt);
351 if (!CV || !this->isValue(CV->getValue()))
352 return false;
353 HasNonUndefElements = true;
354 }
355 return HasNonUndefElements;
356 }
357 }
358 return false;
359 }
360};
361
362/// specialization of cstval_pred_ty for ConstantInt
363template <typename Predicate>
364using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>;
365
366/// specialization of cstval_pred_ty for ConstantFP
367template <typename Predicate>
368using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>;
369
370/// This helper class is used to match scalar and vector constants that
371/// satisfy a specified predicate, and bind them to an APInt.
372template <typename Predicate> struct api_pred_ty : public Predicate {
373 const APInt *&Res;
374
375 api_pred_ty(const APInt *&R) : Res(R) {}
376
377 template <typename ITy> bool match(ITy *V) {
378 if (const auto *CI = dyn_cast<ConstantInt>(V))
379 if (this->isValue(CI->getValue())) {
380 Res = &CI->getValue();
381 return true;
382 }
383 if (V->getType()->isVectorTy())
384 if (const auto *C = dyn_cast<Constant>(V))
385 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
386 if (this->isValue(CI->getValue())) {
387 Res = &CI->getValue();
388 return true;
389 }
390
391 return false;
392 }
393};
394
395/// This helper class is used to match scalar and vector constants that
396/// satisfy a specified predicate, and bind them to an APFloat.
397/// Undefs are allowed in splat vector constants.
398template <typename Predicate> struct apf_pred_ty : public Predicate {
399 const APFloat *&Res;
400
401 apf_pred_ty(const APFloat *&R) : Res(R) {}
402
403 template <typename ITy> bool match(ITy *V) {
404 if (const auto *CI = dyn_cast<ConstantFP>(V))
405 if (this->isValue(CI->getValue())) {
406 Res = &CI->getValue();
407 return true;
408 }
409 if (V->getType()->isVectorTy())
410 if (const auto *C = dyn_cast<Constant>(V))
411 if (auto *CI = dyn_cast_or_null<ConstantFP>(
412 C->getSplatValue(/* AllowUndef */ true)))
413 if (this->isValue(CI->getValue())) {
414 Res = &CI->getValue();
415 return true;
416 }
417
418 return false;
419 }
420};
421
422///////////////////////////////////////////////////////////////////////////////
423//
424// Encapsulate constant value queries for use in templated predicate matchers.
425// This allows checking if constants match using compound predicates and works
426// with vector constants, possibly with relaxed constraints. For example, ignore
427// undef values.
428//
429///////////////////////////////////////////////////////////////////////////////
430
431struct is_any_apint {
432 bool isValue(const APInt &C) { return true; }
433};
434/// Match an integer or vector with any integral constant.
435/// For vectors, this includes constants with undefined elements.
436inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() {
437 return cst_pred_ty<is_any_apint>();
438}
439
440struct is_all_ones {
441 bool isValue(const APInt &C) { return C.isAllOnes(); }
442};
443/// Match an integer or vector with all bits set.
444/// For vectors, this includes constants with undefined elements.
445inline cst_pred_ty<is_all_ones> m_AllOnes() {
446 return cst_pred_ty<is_all_ones>();
447}
448
449struct is_maxsignedvalue {
450 bool isValue(const APInt &C) { return C.isMaxSignedValue(); }
451};
452/// Match an integer or vector with values having all bits except for the high
453/// bit set (0x7f...).
454/// For vectors, this includes constants with undefined elements.
455inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() {
456 return cst_pred_ty<is_maxsignedvalue>();
457}
458inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) {
459 return V;
460}
461
462struct is_negative {
463 bool isValue(const APInt &C) { return C.isNegative(); }
464};
465/// Match an integer or vector of negative values.
466/// For vectors, this includes constants with undefined elements.
467inline cst_pred_ty<is_negative> m_Negative() {
468 return cst_pred_ty<is_negative>();
469}
470inline api_pred_ty<is_negative> m_Negative(const APInt *&V) {
471 return V;
472}
473
474struct is_nonnegative {
475 bool isValue(const APInt &C) { return C.isNonNegative(); }
476};
477/// Match an integer or vector of non-negative values.
478/// For vectors, this includes constants with undefined elements.
479inline cst_pred_ty<is_nonnegative> m_NonNegative() {
480 return cst_pred_ty<is_nonnegative>();
481}
482inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) {
483 return V;
484}
485
486struct is_strictlypositive {
487 bool isValue(const APInt &C) { return C.isStrictlyPositive(); }
488};
489/// Match an integer or vector of strictly positive values.
490/// For vectors, this includes constants with undefined elements.
491inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() {
492 return cst_pred_ty<is_strictlypositive>();
493}
494inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) {
495 return V;
496}
497
498struct is_nonpositive {
499 bool isValue(const APInt &C) { return C.isNonPositive(); }
500};
501/// Match an integer or vector of non-positive values.
502/// For vectors, this includes constants with undefined elements.
503inline cst_pred_ty<is_nonpositive> m_NonPositive() {
504 return cst_pred_ty<is_nonpositive>();
505}
506inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; }
507
508struct is_one {
509 bool isValue(const APInt &C) { return C.isOne(); }
510};
511/// Match an integer 1 or a vector with all elements equal to 1.
512/// For vectors, this includes constants with undefined elements.
513inline cst_pred_ty<is_one> m_One() {
514 return cst_pred_ty<is_one>();
515}
516
517struct is_zero_int {
518 bool isValue(const APInt &C) { return C.isZero(); }
519};
520/// Match an integer 0 or a vector with all elements equal to 0.
521/// For vectors, this includes constants with undefined elements.
522inline cst_pred_ty<is_zero_int> m_ZeroInt() {
523 return cst_pred_ty<is_zero_int>();
524}
525
526struct is_zero {
527 template <typename ITy> bool match(ITy *V) {
528 auto *C = dyn_cast<Constant>(V);
529 // FIXME: this should be able to do something for scalable vectors
530 return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C));
531 }
532};
533/// Match any null constant or a vector with all elements equal to 0.
534/// For vectors, this includes constants with undefined elements.
535inline is_zero m_Zero() {
536 return is_zero();
537}
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() {
545 return cst_pred_ty<is_power2>();
546}
547inline api_pred_ty<is_power2> m_Power2(const APInt *&V) {
548 return V;
549}
550
551struct is_negated_power2 {
552 bool isValue(const APInt &C) { return C.isNegatedPowerOf2(); }
553};
554/// Match a integer or vector negated power-of-2.
555/// For vectors, this includes constants with undefined elements.
556inline cst_pred_ty<is_negated_power2> m_NegatedPower2() {
557 return cst_pred_ty<is_negated_power2>();
558}
559inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) {
560 return V;
561}
562
563struct is_power2_or_zero {
564 bool isValue(const APInt &C) { return !C || C.isPowerOf2(); }
565};
566/// Match an integer or vector of 0 or power-of-2 values.
567/// For vectors, this includes constants with undefined elements.
568inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() {
569 return cst_pred_ty<is_power2_or_zero>();
570}
571inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) {
572 return V;
573}
574
575struct is_sign_mask {
576 bool isValue(const APInt &C) { return C.isSignMask(); }
577};
578/// Match an integer or vector with only the sign bit(s) set.
579/// For vectors, this includes constants with undefined elements.
580inline cst_pred_ty<is_sign_mask> m_SignMask() {
581 return cst_pred_ty<is_sign_mask>();
582}
583
584struct is_lowbit_mask {
585 bool isValue(const APInt &C) { return C.isMask(); }
586};
587/// Match an integer or vector with only the low bit(s) set.
588/// For vectors, this includes constants with undefined elements.
589inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() {
590 return cst_pred_ty<is_lowbit_mask>();
591}
592inline api_pred_ty<is_lowbit_mask> m_LowBitMask(const APInt *&V) {
593 return V;
594}
595
596struct icmp_pred_with_threshold {
597 ICmpInst::Predicate Pred;
598 const APInt *Thr;
599 bool isValue(const APInt &C) { return ICmpInst::compare(C, *Thr, Pred); }
600};
601/// Match an integer or vector with every element comparing 'pred' (eg/ne/...)
602/// to Threshold. For vectors, this includes constants with undefined elements.
603inline cst_pred_ty<icmp_pred_with_threshold>
604m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) {
605 cst_pred_ty<icmp_pred_with_threshold> P;
606 P.Pred = Predicate;
607 P.Thr = &Threshold;
608 return P;
609}
610
611struct is_nan {
612 bool isValue(const APFloat &C) { return C.isNaN(); }
613};
614/// Match an arbitrary NaN constant. This includes quiet and signalling nans.
615/// For vectors, this includes constants with undefined elements.
616inline cstfp_pred_ty<is_nan> m_NaN() {
617 return cstfp_pred_ty<is_nan>();
618}
619
620struct is_nonnan {
621 bool isValue(const APFloat &C) { return !C.isNaN(); }
622};
623/// Match a non-NaN FP constant.
624/// For vectors, this includes constants with undefined elements.
625inline cstfp_pred_ty<is_nonnan> m_NonNaN() {
626 return cstfp_pred_ty<is_nonnan>();
627}
628
629struct is_inf {
630 bool isValue(const APFloat &C) { return C.isInfinity(); }
631};
632/// Match a positive or negative infinity FP constant.
633/// For vectors, this includes constants with undefined elements.
634inline cstfp_pred_ty<is_inf> m_Inf() {
635 return cstfp_pred_ty<is_inf>();
636}
637
638struct is_noninf {
639 bool isValue(const APFloat &C) { return !C.isInfinity(); }
640};
641/// Match a non-infinity FP constant, i.e. finite or NaN.
642/// For vectors, this includes constants with undefined elements.
643inline cstfp_pred_ty<is_noninf> m_NonInf() {
644 return cstfp_pred_ty<is_noninf>();
645}
646
647struct is_finite {
648 bool isValue(const APFloat &C) { return C.isFinite(); }
649};
650/// Match a finite FP constant, i.e. not infinity or NaN.
651/// For vectors, this includes constants with undefined elements.
652inline cstfp_pred_ty<is_finite> m_Finite() {
653 return cstfp_pred_ty<is_finite>();
654}
655inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; }
656
657struct is_finitenonzero {
658 bool isValue(const APFloat &C) { return C.isFiniteNonZero(); }
659};
660/// Match a finite non-zero FP constant.
661/// For vectors, this includes constants with undefined elements.
662inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() {
663 return cstfp_pred_ty<is_finitenonzero>();
664}
665inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) {
666 return V;
667}
668
669struct is_any_zero_fp {
670 bool isValue(const APFloat &C) { return C.isZero(); }
671};
672/// Match a floating-point negative zero or positive zero.
673/// For vectors, this includes constants with undefined elements.
674inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() {
675 return cstfp_pred_ty<is_any_zero_fp>();
676}
677
678struct is_pos_zero_fp {
679 bool isValue(const APFloat &C) { return C.isPosZero(); }
680};
681/// Match a floating-point positive zero.
682/// For vectors, this includes constants with undefined elements.
683inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() {
684 return cstfp_pred_ty<is_pos_zero_fp>();
685}
686
687struct is_neg_zero_fp {
688 bool isValue(const APFloat &C) { return C.isNegZero(); }
689};
690/// Match a floating-point negative zero.
691/// For vectors, this includes constants with undefined elements.
692inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() {
693 return cstfp_pred_ty<is_neg_zero_fp>();
694}
695
696struct is_non_zero_fp {
697 bool isValue(const APFloat &C) { return C.isNonZero(); }
698};
699/// Match a floating-point non-zero.
700/// For vectors, this includes constants with undefined elements.
701inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() {
702 return cstfp_pred_ty<is_non_zero_fp>();
703}
704
705///////////////////////////////////////////////////////////////////////////////
706
707template <typename Class> struct bind_ty {
708 Class *&VR;
709
710 bind_ty(Class *&V) : VR(V) {}
711
712 template <typename ITy> bool match(ITy *V) {
713 if (auto *CV = dyn_cast<Class>(V)) {
714 VR = CV;
715 return true;
716 }
717 return false;
718 }
719};
720
721/// Match a value, capturing it if we match.
722inline 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'
723inline bind_ty<const Value> m_Value(const Value *&V) { return V; }
724
725/// Match an instruction, capturing it if we match.
726inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; }
727/// Match a unary operator, capturing it if we match.
728inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; }
729/// Match a binary operator, capturing it if we match.
730inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; }
731/// Match a with overflow intrinsic, capturing it if we match.
732inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; }
733inline bind_ty<const WithOverflowInst>
734m_WithOverflowInst(const WithOverflowInst *&I) {
735 return I;
736}
737
738/// Match a Constant, capturing the value if we match.
739inline bind_ty<Constant> m_Constant(Constant *&C) { return C; }
740
741/// Match a ConstantInt, capturing the value if we match.
742inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; }
743
744/// Match a ConstantFP, capturing the value if we match.
745inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; }
746
747/// Match a ConstantExpr, capturing the value if we match.
748inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; }
749
750/// Match a basic block value, capturing it if we match.
751inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; }
752inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) {
753 return V;
754}
755
756/// Match an arbitrary immediate Constant and ignore it.
757inline match_combine_and<class_match<Constant>,
758 match_unless<class_match<ConstantExpr>>>
759m_ImmConstant() {
760 return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr()));
761}
762
763/// Match an immediate Constant, capturing the value if we match.
764inline match_combine_and<bind_ty<Constant>,
765 match_unless<class_match<ConstantExpr>>>
766m_ImmConstant(Constant *&C) {
767 return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr()));
768}
769
770/// Match a specified Value*.
771struct specificval_ty {
772 const Value *Val;
773
774 specificval_ty(const Value *V) : Val(V) {}
775
776 template <typename ITy> bool match(ITy *V) { return V == Val; }
777};
778
779/// Match if we have a specific specified value.
780inline specificval_ty m_Specific(const Value *V) { return V; }
781
782/// Stores a reference to the Value *, not the Value * itself,
783/// thus can be used in commutative matchers.
784template <typename Class> struct deferredval_ty {
785 Class *const &Val;
786
787 deferredval_ty(Class *const &V) : Val(V) {}
788
789 template <typename ITy> bool match(ITy *const V) { return V == Val; }
790};
791
792/// Like m_Specific(), but works if the specific value to match is determined
793/// as part of the same match() expression. For example:
794/// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will
795/// bind X before the pattern match starts.
796/// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against
797/// whichever value m_Value(X) populated.
798inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; }
799inline deferredval_ty<const Value> m_Deferred(const Value *const &V) {
800 return V;
801}
802
803/// Match a specified floating point value or vector of all elements of
804/// that value.
805struct specific_fpval {
806 double Val;
807
808 specific_fpval(double V) : Val(V) {}
809
810 template <typename ITy> bool match(ITy *V) {
811 if (const auto *CFP = dyn_cast<ConstantFP>(V))
812 return CFP->isExactlyValue(Val);
813 if (V->getType()->isVectorTy())
814 if (const auto *C = dyn_cast<Constant>(V))
815 if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
816 return CFP->isExactlyValue(Val);
817 return false;
818 }
819};
820
821/// Match a specific floating point value or vector with all elements
822/// equal to the value.
823inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); }
824
825/// Match a float 1.0 or vector with all elements equal to 1.0.
826inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); }
827
828struct bind_const_intval_ty {
829 uint64_t &VR;
830
831 bind_const_intval_ty(uint64_t &V) : VR(V) {}
832
833 template <typename ITy> bool match(ITy *V) {
834 if (const auto *CV = dyn_cast<ConstantInt>(V))
835 if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) {
836 VR = CV->getZExtValue();
837 return true;
838 }
839 return false;
840 }
841};
842
843/// Match a specified integer value or vector of all elements of that
844/// value.
845template <bool AllowUndefs>
846struct specific_intval {
847 APInt Val;
848
849 specific_intval(APInt V) : Val(std::move(V)) {}
850
851 template <typename ITy> bool match(ITy *V) {
852 const auto *CI = dyn_cast<ConstantInt>(V);
853 if (!CI && V->getType()->isVectorTy())
854 if (const auto *C = dyn_cast<Constant>(V))
855 CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs));
856
857 return CI && APInt::isSameValue(CI->getValue(), Val);
858 }
859};
860
861/// Match a specific integer value or vector with all elements equal to
862/// the value.
863inline specific_intval<false> m_SpecificInt(APInt V) {
864 return specific_intval<false>(std::move(V));
865}
866
867inline specific_intval<false> m_SpecificInt(uint64_t V) {
868 return m_SpecificInt(APInt(64, V));
869}
870
871inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) {
872 return specific_intval<true>(std::move(V));
873}
874
875inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) {
876 return m_SpecificIntAllowUndef(APInt(64, V));
877}
878
879/// Match a ConstantInt and bind to its value. This does not match
880/// ConstantInts wider than 64-bits.
881inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; }
882
883/// Match a specified basic block value.
884struct specific_bbval {
885 BasicBlock *Val;
886
887 specific_bbval(BasicBlock *Val) : Val(Val) {}
888
889 template <typename ITy> bool match(ITy *V) {
890 const auto *BB = dyn_cast<BasicBlock>(V);
891 return BB && BB == Val;
892 }
893};
894
895/// Match a specific basic block value.
896inline specific_bbval m_SpecificBB(BasicBlock *BB) {
897 return specific_bbval(BB);
898}
899
900/// A commutative-friendly version of m_Specific().
901inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) {
902 return BB;
903}
904inline deferredval_ty<const BasicBlock>
905m_Deferred(const BasicBlock *const &BB) {
906 return BB;
907}
908
909//===----------------------------------------------------------------------===//
910// Matcher for any binary operator.
911//
912template <typename LHS_t, typename RHS_t, bool Commutable = false>
913struct AnyBinaryOp_match {
914 LHS_t L;
915 RHS_t R;
916
917 // The evaluation order is always stable, regardless of Commutability.
918 // The LHS is always matched first.
919 AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
920
921 template <typename OpTy> bool match(OpTy *V) {
922 if (auto *I
15.1
'I' is non-null
15.1
'I' is non-null
= dyn_cast<BinaryOperator>(V))
15
Assuming 'V' is a 'BinaryOperator'
923 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
16
Calling 'match_combine_or::match'
33
Returning from 'match_combine_or::match'
924 (Commutable && L.match(I->getOperand(1)) &&
925 R.match(I->getOperand(0)));
926 return false;
927 }
928};
929
930template <typename LHS, typename RHS>
931inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) {
932 return AnyBinaryOp_match<LHS, RHS>(L, R);
933}
934
935//===----------------------------------------------------------------------===//
936// Matcher for any unary operator.
937// TODO fuse unary, binary matcher into n-ary matcher
938//
939template <typename OP_t> struct AnyUnaryOp_match {
940 OP_t X;
941
942 AnyUnaryOp_match(const OP_t &X) : X(X) {}
943
944 template <typename OpTy> bool match(OpTy *V) {
945 if (auto *I = dyn_cast<UnaryOperator>(V))
946 return X.match(I->getOperand(0));
947 return false;
948 }
949};
950
951template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) {
952 return AnyUnaryOp_match<OP_t>(X);
953}
954
955//===----------------------------------------------------------------------===//
956// Matchers for specific binary operators.
957//
958
959template <typename LHS_t, typename RHS_t, unsigned Opcode,
960 bool Commutable = false>
961struct BinaryOp_match {
962 LHS_t L;
963 RHS_t R;
964
965 // The evaluation order is always stable, regardless of Commutability.
966 // The LHS is always matched first.
967 BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
968
969 template <typename OpTy> inline bool match(unsigned Opc, OpTy *V) {
970 if (V->getValueID() == Value::InstructionVal + Opc) {
21
Assuming the condition is true
22
Taking true branch
971 auto *I = cast<BinaryOperator>(V);
23
'V' is a 'BinaryOperator'
972 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
973 (Commutable && L.match(I->getOperand(1)) &&
974 R.match(I->getOperand(0)));
975 }
976 if (auto *CE = dyn_cast<ConstantExpr>(V))
977 return CE->getOpcode() == Opc &&
978 ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) ||
979 (Commutable && L.match(CE->getOperand(1)) &&
980 R.match(CE->getOperand(0))));
981 return false;
982 }
983
984 template <typename OpTy> bool match(OpTy *V) { return match(Opcode, V); }
20
Calling 'BinaryOp_match::match'
27
Returning from 'BinaryOp_match::match'
985};
986
987template <typename LHS, typename RHS>
988inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L,
989 const RHS &R) {
990 return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R);
991}
992
993template <typename LHS, typename RHS>
994inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L,
995 const RHS &R) {
996 return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R);
997}
998
999template <typename LHS, typename RHS>
1000inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L,
1001 const RHS &R) {
1002 return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R);
1003}
1004
1005template <typename LHS, typename RHS>
1006inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L,
1007 const RHS &R) {
1008 return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R);
1009}
1010
1011template <typename Op_t> struct FNeg_match {
1012 Op_t X;
1013
1014 FNeg_match(const Op_t &Op) : X(Op) {}
1015 template <typename OpTy> bool match(OpTy *V) {
1016 auto *FPMO = dyn_cast<FPMathOperator>(V);
1017 if (!FPMO) return false;
1018
1019 if (FPMO->getOpcode() == Instruction::FNeg)
1020 return X.match(FPMO->getOperand(0));
1021
1022 if (FPMO->getOpcode() == Instruction::FSub) {
1023 if (FPMO->hasNoSignedZeros()) {
1024 // With 'nsz', any zero goes.
1025 if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0)))
1026 return false;
1027 } else {
1028 // Without 'nsz', we need fsub -0.0, X exactly.
1029 if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0)))
1030 return false;
1031 }
1032
1033 return X.match(FPMO->getOperand(1));
1034 }
1035
1036 return false;
1037 }
1038};
1039
1040/// Match 'fneg X' as 'fsub -0.0, X'.
1041template <typename OpTy>
1042inline FNeg_match<OpTy>
1043m_FNeg(const OpTy &X) {
1044 return FNeg_match<OpTy>(X);
1045}
1046
1047/// Match 'fneg X' as 'fsub +-0.0, X'.
1048template <typename RHS>
1049inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub>
1050m_FNegNSZ(const RHS &X) {
1051 return m_FSub(m_AnyZeroFP(), X);
1052}
1053
1054template <typename LHS, typename RHS>
1055inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L,
1056 const RHS &R) {
1057 return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R);
1058}
1059
1060template <typename LHS, typename RHS>
1061inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L,
1062 const RHS &R) {
1063 return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R);
1064}
1065
1066template <typename LHS, typename RHS>
1067inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L,
1068 const RHS &R) {
1069 return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R);
1070}
1071
1072template <typename LHS, typename RHS>
1073inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L,
1074 const RHS &R) {
1075 return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R);
1076}
1077
1078template <typename LHS, typename RHS>
1079inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L,
1080 const RHS &R) {
1081 return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R);
1082}
1083
1084template <typename LHS, typename RHS>
1085inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L,
1086 const RHS &R) {
1087 return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R);
1088}
1089
1090template <typename LHS, typename RHS>
1091inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L,
1092 const RHS &R) {
1093 return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R);
1094}
1095
1096template <typename LHS, typename RHS>
1097inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L,
1098 const RHS &R) {
1099 return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R);
1100}
1101
1102template <typename LHS, typename RHS>
1103inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L,
1104 const RHS &R) {
1105 return BinaryOp_match<LHS, RHS, Instruction::And>(L, R);
1106}
1107
1108template <typename LHS, typename RHS>
1109inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L,
1110 const RHS &R) {
1111 return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R);
1112}
1113
1114template <typename LHS, typename RHS>
1115inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L,
1116 const RHS &R) {
1117 return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R);
1118}
1119
1120template <typename LHS, typename RHS>
1121inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L,
1122 const RHS &R) {
1123 return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R);
1124}
1125
1126template <typename LHS, typename RHS>
1127inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L,
1128 const RHS &R) {
1129 return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R);
1130}
1131
1132template <typename LHS, typename RHS>
1133inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L,
1134 const RHS &R) {
1135 return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R);
1136}
1137
1138template <typename LHS_t, typename RHS_t, unsigned Opcode,
1139 unsigned WrapFlags = 0>
1140struct OverflowingBinaryOp_match {
1141 LHS_t L;
1142 RHS_t R;
1143
1144 OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS)
1145 : L(LHS), R(RHS) {}
1146
1147 template <typename OpTy> bool match(OpTy *V) {
1148 if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) {
1149 if (Op->getOpcode() != Opcode)
1150 return false;
1151 if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) &&
1152 !Op->hasNoUnsignedWrap())
1153 return false;
1154 if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) &&
1155 !Op->hasNoSignedWrap())
1156 return false;
1157 return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1));
1158 }
1159 return false;
1160 }
1161};
1162
1163template <typename LHS, typename RHS>
1164inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1165 OverflowingBinaryOperator::NoSignedWrap>
1166m_NSWAdd(const LHS &L, const RHS &R) {
1167 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1168 OverflowingBinaryOperator::NoSignedWrap>(
1169 L, R);
1170}
1171template <typename LHS, typename RHS>
1172inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1173 OverflowingBinaryOperator::NoSignedWrap>
1174m_NSWSub(const LHS &L, const RHS &R) {
1175 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1176 OverflowingBinaryOperator::NoSignedWrap>(
1177 L, R);
1178}
1179template <typename LHS, typename RHS>
1180inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1181 OverflowingBinaryOperator::NoSignedWrap>
1182m_NSWMul(const LHS &L, const RHS &R) {
1183 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1184 OverflowingBinaryOperator::NoSignedWrap>(
1185 L, R);
1186}
1187template <typename LHS, typename RHS>
1188inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1189 OverflowingBinaryOperator::NoSignedWrap>
1190m_NSWShl(const LHS &L, const RHS &R) {
1191 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1192 OverflowingBinaryOperator::NoSignedWrap>(
1193 L, R);
1194}
1195
1196template <typename LHS, typename RHS>
1197inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1198 OverflowingBinaryOperator::NoUnsignedWrap>
1199m_NUWAdd(const LHS &L, const RHS &R) {
1200 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1201 OverflowingBinaryOperator::NoUnsignedWrap>(
1202 L, R);
1203}
1204template <typename LHS, typename RHS>
1205inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1206 OverflowingBinaryOperator::NoUnsignedWrap>
1207m_NUWSub(const LHS &L, const RHS &R) {
1208 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1209 OverflowingBinaryOperator::NoUnsignedWrap>(
1210 L, R);
1211}
1212template <typename LHS, typename RHS>
1213inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1214 OverflowingBinaryOperator::NoUnsignedWrap>
1215m_NUWMul(const LHS &L, const RHS &R) {
1216 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1217 OverflowingBinaryOperator::NoUnsignedWrap>(
1218 L, R);
1219}
1220template <typename LHS, typename RHS>
1221inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1222 OverflowingBinaryOperator::NoUnsignedWrap>
1223m_NUWShl(const LHS &L, const RHS &R) {
1224 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1225 OverflowingBinaryOperator::NoUnsignedWrap>(
1226 L, R);
1227}
1228
1229template <typename LHS_t, typename RHS_t, bool Commutable = false>
1230struct SpecificBinaryOp_match
1231 : public BinaryOp_match<LHS_t, RHS_t, 0, Commutable> {
1232 unsigned Opcode;
1233
1234 SpecificBinaryOp_match(unsigned Opcode, const LHS_t &LHS, const RHS_t &RHS)
1235 : BinaryOp_match<LHS_t, RHS_t, 0, Commutable>(LHS, RHS), Opcode(Opcode) {}
1236
1237 template <typename OpTy> bool match(OpTy *V) {
1238 return BinaryOp_match<LHS_t, RHS_t, 0, Commutable>::match(Opcode, V);
1239 }
1240};
1241
1242/// Matches a specific opcode.
1243template <typename LHS, typename RHS>
1244inline SpecificBinaryOp_match<LHS, RHS> m_BinOp(unsigned Opcode, const LHS &L,
1245 const RHS &R) {
1246 return SpecificBinaryOp_match<LHS, RHS>(Opcode, L, R);
1247}
1248
1249//===----------------------------------------------------------------------===//
1250// Class that matches a group of binary opcodes.
1251//
1252template <typename LHS_t, typename RHS_t, typename Predicate>
1253struct BinOpPred_match : Predicate {
1254 LHS_t L;
1255 RHS_t R;
1256
1257 BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1258
1259 template <typename OpTy> bool match(OpTy *V) {
1260 if (auto *I = dyn_cast<Instruction>(V))
1261 return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) &&
1262 R.match(I->getOperand(1));
1263 if (auto *CE = dyn_cast<ConstantExpr>(V))
1264 return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) &&
1265 R.match(CE->getOperand(1));
1266 return false;
1267 }
1268};
1269
1270struct is_shift_op {
1271 bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); }
1272};
1273
1274struct is_right_shift_op {
1275 bool isOpType(unsigned Opcode) {
1276 return Opcode == Instruction::LShr || Opcode == Instruction::AShr;
1277 }
1278};
1279
1280struct is_logical_shift_op {
1281 bool isOpType(unsigned Opcode) {
1282 return Opcode == Instruction::LShr || Opcode == Instruction::Shl;
1283 }
1284};
1285
1286struct is_bitwiselogic_op {
1287 bool isOpType(unsigned Opcode) {
1288 return Instruction::isBitwiseLogicOp(Opcode);
1289 }
1290};
1291
1292struct is_idiv_op {
1293 bool isOpType(unsigned Opcode) {
1294 return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
1295 }
1296};
1297
1298struct is_irem_op {
1299 bool isOpType(unsigned Opcode) {
1300 return Opcode == Instruction::SRem || Opcode == Instruction::URem;
1301 }
1302};
1303
1304/// Matches shift operations.
1305template <typename LHS, typename RHS>
1306inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L,
1307 const RHS &R) {
1308 return BinOpPred_match<LHS, RHS, is_shift_op>(L, R);
1309}
1310
1311/// Matches logical shift operations.
1312template <typename LHS, typename RHS>
1313inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L,
1314 const RHS &R) {
1315 return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R);
1316}
1317
1318/// Matches logical shift operations.
1319template <typename LHS, typename RHS>
1320inline BinOpPred_match<LHS, RHS, is_logical_shift_op>
1321m_LogicalShift(const LHS &L, const RHS &R) {
1322 return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R);
1323}
1324
1325/// Matches bitwise logic operations.
1326template <typename LHS, typename RHS>
1327inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op>
1328m_BitwiseLogic(const LHS &L, const RHS &R) {
1329 return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R);
1330}
1331
1332/// Matches integer division operations.
1333template <typename LHS, typename RHS>
1334inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L,
1335 const RHS &R) {
1336 return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R);
1337}
1338
1339/// Matches integer remainder operations.
1340template <typename LHS, typename RHS>
1341inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L,
1342 const RHS &R) {
1343 return BinOpPred_match<LHS, RHS, is_irem_op>(L, R);
1344}
1345
1346//===----------------------------------------------------------------------===//
1347// Class that matches exact binary ops.
1348//
1349template <typename SubPattern_t> struct Exact_match {
1350 SubPattern_t SubPattern;
1351
1352 Exact_match(const SubPattern_t &SP) : SubPattern(SP) {}
1353
1354 template <typename OpTy> bool match(OpTy *V) {
1355 if (auto *PEO = dyn_cast<PossiblyExactOperator>(V))
1356 return PEO->isExact() && SubPattern.match(V);
1357 return false;
1358 }
1359};
1360
1361template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) {
1362 return SubPattern;
1363}
1364
1365//===----------------------------------------------------------------------===//
1366// Matchers for CmpInst classes
1367//
1368
1369template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy,
1370 bool Commutable = false>
1371struct CmpClass_match {
1372 PredicateTy &Predicate;
1373 LHS_t L;
1374 RHS_t R;
1375
1376 // The evaluation order is always stable, regardless of Commutability.
1377 // The LHS is always matched first.
1378 CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS)
1379 : Predicate(Pred), L(LHS), R(RHS) {}
1380
1381 template <typename OpTy> bool match(OpTy *V) {
1382 if (auto *I = dyn_cast<Class>(V)) {
1383 if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) {
1384 Predicate = I->getPredicate();
1385 return true;
1386 } else if (Commutable && L.match(I->getOperand(1)) &&
1387 R.match(I->getOperand(0))) {
1388 Predicate = I->getSwappedPredicate();
1389 return true;
1390 }
1391 }
1392 return false;
1393 }
1394};
1395
1396template <typename LHS, typename RHS>
1397inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>
1398m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1399 return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R);
1400}
1401
1402template <typename LHS, typename RHS>
1403inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>
1404m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1405 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R);
1406}
1407
1408template <typename LHS, typename RHS>
1409inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>
1410m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1411 return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R);
1412}
1413
1414//===----------------------------------------------------------------------===//
1415// Matchers for instructions with a given opcode and number of operands.
1416//
1417
1418/// Matches instructions with Opcode and three operands.
1419template <typename T0, unsigned Opcode> struct OneOps_match {
1420 T0 Op1;
1421
1422 OneOps_match(const T0 &Op1) : Op1(Op1) {}
1423
1424 template <typename OpTy> bool match(OpTy *V) {
1425 if (V->getValueID() == Value::InstructionVal + Opcode) {
1426 auto *I = cast<Instruction>(V);
1427 return Op1.match(I->getOperand(0));
1428 }
1429 return false;
1430 }
1431};
1432
1433/// Matches instructions with Opcode and three operands.
1434template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match {
1435 T0 Op1;
1436 T1 Op2;
1437
1438 TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {}
1439
1440 template <typename OpTy> bool match(OpTy *V) {
1441 if (V->getValueID() == Value::InstructionVal + Opcode) {
1442 auto *I = cast<Instruction>(V);
1443 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1));
1444 }
1445 return false;
1446 }
1447};
1448
1449/// Matches instructions with Opcode and three operands.
1450template <typename T0, typename T1, typename T2, unsigned Opcode>
1451struct ThreeOps_match {
1452 T0 Op1;
1453 T1 Op2;
1454 T2 Op3;
1455
1456 ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3)
1457 : Op1(Op1), Op2(Op2), Op3(Op3) {}
1458
1459 template <typename OpTy> bool match(OpTy *V) {
1460 if (V->getValueID() == Value::InstructionVal + Opcode) {
65
Called C++ object pointer is null
1461 auto *I = cast<Instruction>(V);
1462 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1463 Op3.match(I->getOperand(2));
1464 }
1465 return false;
1466 }
1467};
1468
1469/// Matches SelectInst.
1470template <typename Cond, typename LHS, typename RHS>
1471inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select>
1472m_Select(const Cond &C, const LHS &L, const RHS &R) {
1473 return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R);
1474}
1475
1476/// This matches a select of two constants, e.g.:
1477/// m_SelectCst<-1, 0>(m_Value(V))
1478template <int64_t L, int64_t R, typename Cond>
1479inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>,
1480 Instruction::Select>
1481m_SelectCst(const Cond &C) {
1482 return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>());
1483}
1484
1485/// Matches FreezeInst.
1486template <typename OpTy>
1487inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) {
1488 return OneOps_match<OpTy, Instruction::Freeze>(Op);
1489}
1490
1491/// Matches InsertElementInst.
1492template <typename Val_t, typename Elt_t, typename Idx_t>
1493inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>
1494m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) {
1495 return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>(
1496 Val, Elt, Idx);
1497}
1498
1499/// Matches ExtractElementInst.
1500template <typename Val_t, typename Idx_t>
1501inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>
1502m_ExtractElt(const Val_t &Val, const Idx_t &Idx) {
1503 return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx);
1504}
1505
1506/// Matches shuffle.
1507template <typename T0, typename T1, typename T2> struct Shuffle_match {
1508 T0 Op1;
1509 T1 Op2;
1510 T2 Mask;
1511
1512 Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask)
1513 : Op1(Op1), Op2(Op2), Mask(Mask) {}
1514
1515 template <typename OpTy> bool match(OpTy *V) {
1516 if (auto *I = dyn_cast<ShuffleVectorInst>(V)) {
1517 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1518 Mask.match(I->getShuffleMask());
1519 }
1520 return false;
1521 }
1522};
1523
1524struct m_Mask {
1525 ArrayRef<int> &MaskRef;
1526 m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1527 bool match(ArrayRef<int> Mask) {
1528 MaskRef = Mask;
1529 return true;
1530 }
1531};
1532
1533struct m_ZeroMask {
1534 bool match(ArrayRef<int> Mask) {
1535 return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; });
1536 }
1537};
1538
1539struct m_SpecificMask {
1540 ArrayRef<int> &MaskRef;
1541 m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1542 bool match(ArrayRef<int> Mask) { return MaskRef == Mask; }
1543};
1544
1545struct m_SplatOrUndefMask {
1546 int &SplatIndex;
1547 m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {}
1548 bool match(ArrayRef<int> Mask) {
1549 auto First = find_if(Mask, [](int Elem) { return Elem != -1; });
1550 if (First == Mask.end())
1551 return false;
1552 SplatIndex = *First;
1553 return all_of(Mask,
1554 [First](int Elem) { return Elem == *First || Elem == -1; });
1555 }
1556};
1557
1558/// Matches ShuffleVectorInst independently of mask value.
1559template <typename V1_t, typename V2_t>
1560inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>
1561m_Shuffle(const V1_t &v1, const V2_t &v2) {
1562 return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2);
1563}
1564
1565template <typename V1_t, typename V2_t, typename Mask_t>
1566inline Shuffle_match<V1_t, V2_t, Mask_t>
1567m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) {
1568 return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask);
1569}
1570
1571/// Matches LoadInst.
1572template <typename OpTy>
1573inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) {
1574 return OneOps_match<OpTy, Instruction::Load>(Op);
1575}
1576
1577/// Matches StoreInst.
1578template <typename ValueOpTy, typename PointerOpTy>
1579inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>
1580m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) {
1581 return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp,
1582 PointerOp);
1583}
1584
1585//===----------------------------------------------------------------------===//
1586// Matchers for CastInst classes
1587//
1588
1589template <typename Op_t, unsigned Opcode> struct CastClass_match {
1590 Op_t Op;
1591
1592 CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
1593
1594 template <typename OpTy> bool match(OpTy *V) {
1595 if (auto *O = dyn_cast<Operator>(V))
1596 return O->getOpcode() == Opcode && Op.match(O->getOperand(0));
1597 return false;
1598 }
1599};
1600
1601/// Matches BitCast.
1602template <typename OpTy>
1603inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) {
1604 return CastClass_match<OpTy, Instruction::BitCast>(Op);
1605}
1606
1607/// Matches PtrToInt.
1608template <typename OpTy>
1609inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) {
1610 return CastClass_match<OpTy, Instruction::PtrToInt>(Op);
1611}
1612
1613/// Matches IntToPtr.
1614template <typename OpTy>
1615inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) {
1616 return CastClass_match<OpTy, Instruction::IntToPtr>(Op);
1617}
1618
1619/// Matches Trunc.
1620template <typename OpTy>
1621inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) {
1622 return CastClass_match<OpTy, Instruction::Trunc>(Op);
1623}
1624
1625template <typename OpTy>
1626inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy>
1627m_TruncOrSelf(const OpTy &Op) {
1628 return m_CombineOr(m_Trunc(Op), Op);
1629}
1630
1631/// Matches SExt.
1632template <typename OpTy>
1633inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) {
1634 return CastClass_match<OpTy, Instruction::SExt>(Op);
1635}
1636
1637/// Matches ZExt.
1638template <typename OpTy>
1639inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) {
1640 return CastClass_match<OpTy, Instruction::ZExt>(Op);
49
Returning without writing to 'Op.VR'
1641}
1642
1643template <typename OpTy>
1644inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy>
1645m_ZExtOrSelf(const OpTy &Op) {
1646 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'
1647}
1648
1649template <typename OpTy>
1650inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy>
1651m_SExtOrSelf(const OpTy &Op) {
1652 return m_CombineOr(m_SExt(Op), Op);
1653}
1654
1655template <typename OpTy>
1656inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1657 CastClass_match<OpTy, Instruction::SExt>>
1658m_ZExtOrSExt(const OpTy &Op) {
1659 return m_CombineOr(m_ZExt(Op), m_SExt(Op));
1660}
1661
1662template <typename OpTy>
1663inline match_combine_or<
1664 match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1665 CastClass_match<OpTy, Instruction::SExt>>,
1666 OpTy>
1667m_ZExtOrSExtOrSelf(const OpTy &Op) {
1668 return m_CombineOr(m_ZExtOrSExt(Op), Op);
1669}
1670
1671template <typename OpTy>
1672inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) {
1673 return CastClass_match<OpTy, Instruction::UIToFP>(Op);
1674}
1675
1676template <typename OpTy>
1677inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
1678 return CastClass_match<OpTy, Instruction::SIToFP>(Op);
1679}
1680
1681template <typename OpTy>
1682inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) {
1683 return CastClass_match<OpTy, Instruction::FPToUI>(Op);
1684}
1685
1686template <typename OpTy>
1687inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) {
1688 return CastClass_match<OpTy, Instruction::FPToSI>(Op);
1689}
1690
1691template <typename OpTy>
1692inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
1693 return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
1694}
1695
1696template <typename OpTy>
1697inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
1698 return CastClass_match<OpTy, Instruction::FPExt>(Op);
1699}
1700
1701//===----------------------------------------------------------------------===//
1702// Matchers for control flow.
1703//
1704
1705struct br_match {
1706 BasicBlock *&Succ;
1707
1708 br_match(BasicBlock *&Succ) : Succ(Succ) {}
1709
1710 template <typename OpTy> bool match(OpTy *V) {
1711 if (auto *BI = dyn_cast<BranchInst>(V))
1712 if (BI->isUnconditional()) {
1713 Succ = BI->getSuccessor(0);
1714 return true;
1715 }
1716 return false;
1717 }
1718};
1719
1720inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); }
1721
1722template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1723struct brc_match {
1724 Cond_t Cond;
1725 TrueBlock_t T;
1726 FalseBlock_t F;
1727
1728 brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f)
1729 : Cond(C), T(t), F(f) {}
1730
1731 template <typename OpTy> bool match(OpTy *V) {
1732 if (auto *BI = dyn_cast<BranchInst>(V))
1733 if (BI->isConditional() && Cond.match(BI->getCondition()))
1734 return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1));
1735 return false;
1736 }
1737};
1738
1739template <typename Cond_t>
1740inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>
1741m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) {
1742 return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>(
1743 C, m_BasicBlock(T), m_BasicBlock(F));
1744}
1745
1746template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1747inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t>
1748m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) {
1749 return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F);
1750}
1751
1752//===----------------------------------------------------------------------===//
1753// Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y).
1754//
1755
1756template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t,
1757 bool Commutable = false>
1758struct MaxMin_match {
1759 using PredType = Pred_t;
1760 LHS_t L;
1761 RHS_t R;
1762
1763 // The evaluation order is always stable, regardless of Commutability.
1764 // The LHS is always matched first.
1765 MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1766
1767 template <typename OpTy> bool match(OpTy *V) {
1768 if (auto *II = dyn_cast<IntrinsicInst>(V)) {
1769 Intrinsic::ID IID = II->getIntrinsicID();
1770 if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) ||
1771 (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) ||
1772 (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) ||
1773 (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) {
1774 Value *LHS = II->getOperand(0), *RHS = II->getOperand(1);
1775 return (L.match(LHS) && R.match(RHS)) ||
1776 (Commutable && L.match(RHS) && R.match(LHS));
1777 }
1778 }
1779 // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x".
1780 auto *SI = dyn_cast<SelectInst>(V);
1781 if (!SI)
1782 return false;
1783 auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition());
1784 if (!Cmp)
1785 return false;
1786 // At this point we have a select conditioned on a comparison. Check that
1787 // it is the values returned by the select that are being compared.
1788 auto *TrueVal = SI->getTrueValue();
1789 auto *FalseVal = SI->getFalseValue();
1790 auto *LHS = Cmp->getOperand(0);
1791 auto *RHS = Cmp->getOperand(1);
1792 if ((TrueVal != LHS || FalseVal != RHS) &&
1793 (TrueVal != RHS || FalseVal != LHS))
1794 return false;
1795 typename CmpInst_t::Predicate Pred =
1796 LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate();
1797 // Does "(x pred y) ? x : y" represent the desired max/min operation?
1798 if (!Pred_t::match(Pred))
1799 return false;
1800 // It does! Bind the operands.
1801 return (L.match(LHS) && R.match(RHS)) ||
1802 (Commutable && L.match(RHS) && R.match(LHS));
1803 }
1804};
1805
1806/// Helper class for identifying signed max predicates.
1807struct smax_pred_ty {
1808 static bool match(ICmpInst::Predicate Pred) {
1809 return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE;
1810 }
1811};
1812
1813/// Helper class for identifying signed min predicates.
1814struct smin_pred_ty {
1815 static bool match(ICmpInst::Predicate Pred) {
1816 return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE;
1817 }
1818};
1819
1820/// Helper class for identifying unsigned max predicates.
1821struct umax_pred_ty {
1822 static bool match(ICmpInst::Predicate Pred) {
1823 return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE;
1824 }
1825};
1826
1827/// Helper class for identifying unsigned min predicates.
1828struct umin_pred_ty {
1829 static bool match(ICmpInst::Predicate Pred) {
1830 return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE;
1831 }
1832};
1833
1834/// Helper class for identifying ordered max predicates.
1835struct ofmax_pred_ty {
1836 static bool match(FCmpInst::Predicate Pred) {
1837 return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE;
1838 }
1839};
1840
1841/// Helper class for identifying ordered min predicates.
1842struct ofmin_pred_ty {
1843 static bool match(FCmpInst::Predicate Pred) {
1844 return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE;
1845 }
1846};
1847
1848/// Helper class for identifying unordered max predicates.
1849struct ufmax_pred_ty {
1850 static bool match(FCmpInst::Predicate Pred) {
1851 return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE;
1852 }
1853};
1854
1855/// Helper class for identifying unordered min predicates.
1856struct ufmin_pred_ty {
1857 static bool match(FCmpInst::Predicate Pred) {
1858 return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE;
1859 }
1860};
1861
1862template <typename LHS, typename RHS>
1863inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L,
1864 const RHS &R) {
1865 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R);
1866}
1867
1868template <typename LHS, typename RHS>
1869inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L,
1870 const RHS &R) {
1871 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R);
1872}
1873
1874template <typename LHS, typename RHS>
1875inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L,
1876 const RHS &R) {
1877 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R);
1878}
1879
1880template <typename LHS, typename RHS>
1881inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L,
1882 const RHS &R) {
1883 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R);
1884}
1885
1886template <typename LHS, typename RHS>
1887inline match_combine_or<
1888 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>,
1889 MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>,
1890 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>,
1891 MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>>
1892m_MaxOrMin(const LHS &L, const RHS &R) {
1893 return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)),
1894 m_CombineOr(m_UMax(L, R), m_UMin(L, R)));
1895}
1896
1897/// Match an 'ordered' floating point maximum function.
1898/// Floating point has one special value 'NaN'. Therefore, there is no total
1899/// order. However, if we can ignore the 'NaN' value (for example, because of a
1900/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1901/// semantics. In the presence of 'NaN' we have to preserve the original
1902/// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate.
1903///
1904/// max(L, R) iff L and R are not NaN
1905/// m_OrdFMax(L, R) = R iff L or R are NaN
1906template <typename LHS, typename RHS>
1907inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L,
1908 const RHS &R) {
1909 return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R);
1910}
1911
1912/// Match an 'ordered' floating point minimum function.
1913/// Floating point has one special value 'NaN'. Therefore, there is no total
1914/// order. However, if we can ignore the 'NaN' value (for example, because of a
1915/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1916/// semantics. In the presence of 'NaN' we have to preserve the original
1917/// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate.
1918///
1919/// min(L, R) iff L and R are not NaN
1920/// m_OrdFMin(L, R) = R iff L or R are NaN
1921template <typename LHS, typename RHS>
1922inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L,
1923 const RHS &R) {
1924 return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R);
1925}
1926
1927/// Match an 'unordered' floating point maximum function.
1928/// Floating point has one special value 'NaN'. Therefore, there is no total
1929/// order. However, if we can ignore the 'NaN' value (for example, because of a
1930/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1931/// semantics. In the presence of 'NaN' we have to preserve the original
1932/// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate.
1933///
1934/// max(L, R) iff L and R are not NaN
1935/// m_UnordFMax(L, R) = L iff L or R are NaN
1936template <typename LHS, typename RHS>
1937inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>
1938m_UnordFMax(const LHS &L, const RHS &R) {
1939 return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R);
1940}
1941
1942/// Match an 'unordered' floating point minimum function.
1943/// Floating point has one special value 'NaN'. Therefore, there is no total
1944/// order. However, if we can ignore the 'NaN' value (for example, because of a
1945/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1946/// semantics. In the presence of 'NaN' we have to preserve the original
1947/// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate.
1948///
1949/// min(L, R) iff L and R are not NaN
1950/// m_UnordFMin(L, R) = L iff L or R are NaN
1951template <typename LHS, typename RHS>
1952inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>
1953m_UnordFMin(const LHS &L, const RHS &R) {
1954 return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R);
1955}
1956
1957//===----------------------------------------------------------------------===//
1958// Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b
1959// Note that S might be matched to other instructions than AddInst.
1960//
1961
1962template <typename LHS_t, typename RHS_t, typename Sum_t>
1963struct UAddWithOverflow_match {
1964 LHS_t L;
1965 RHS_t R;
1966 Sum_t S;
1967
1968 UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S)
1969 : L(L), R(R), S(S) {}
1970
1971 template <typename OpTy> bool match(OpTy *V) {
1972 Value *ICmpLHS, *ICmpRHS;
1973 ICmpInst::Predicate Pred;
1974 if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V))
1975 return false;
1976
1977 Value *AddLHS, *AddRHS;
1978 auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS));
1979
1980 // (a + b) u< a, (a + b) u< b
1981 if (Pred == ICmpInst::ICMP_ULT)
1982 if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS))
1983 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1984
1985 // a >u (a + b), b >u (a + b)
1986 if (Pred == ICmpInst::ICMP_UGT)
1987 if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS))
1988 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1989
1990 Value *Op1;
1991 auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes()));
1992 // (a ^ -1) <u b
1993 if (Pred == ICmpInst::ICMP_ULT) {
1994 if (XorExpr.match(ICmpLHS))
1995 return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS);
1996 }
1997 // b > u (a ^ -1)
1998 if (Pred == ICmpInst::ICMP_UGT) {
1999 if (XorExpr.match(ICmpRHS))
2000 return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS);
2001 }
2002
2003 // Match special-case for increment-by-1.
2004 if (Pred == ICmpInst::ICMP_EQ) {
2005 // (a + 1) == 0
2006 // (1 + a) == 0
2007 if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) &&
2008 (m_One().match(AddLHS) || m_One().match(AddRHS)))
2009 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
2010 // 0 == (a + 1)
2011 // 0 == (1 + a)
2012 if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) &&
2013 (m_One().match(AddLHS) || m_One().match(AddRHS)))
2014 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
2015 }
2016
2017 return false;
2018 }
2019};
2020
2021/// Match an icmp instruction checking for unsigned overflow on addition.
2022///
2023/// S is matched to the addition whose result is being checked for overflow, and
2024/// L and R are matched to the LHS and RHS of S.
2025template <typename LHS_t, typename RHS_t, typename Sum_t>
2026UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>
2027m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) {
2028 return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S);
2029}
2030
2031template <typename Opnd_t> struct Argument_match {
2032 unsigned OpI;
2033 Opnd_t Val;
2034
2035 Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {}
2036
2037 template <typename OpTy> bool match(OpTy *V) {
2038 // FIXME: Should likely be switched to use `CallBase`.
2039 if (const auto *CI = dyn_cast<CallInst>(V))
2040 return Val.match(CI->getArgOperand(OpI));
2041 return false;
2042 }
2043};
2044
2045/// Match an argument.
2046template <unsigned OpI, typename Opnd_t>
2047inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) {
2048 return Argument_match<Opnd_t>(OpI, Op);
2049}
2050
2051/// Intrinsic matchers.
2052struct IntrinsicID_match {
2053 unsigned ID;
2054
2055 IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {}
2056
2057 template <typename OpTy> bool match(OpTy *V) {
2058 if (const auto *CI = dyn_cast<CallInst>(V))
2059 if (const auto *F = CI->getCalledFunction())
2060 return F->getIntrinsicID() == ID;
2061 return false;
2062 }
2063};
2064
2065/// Intrinsic matches are combinations of ID matchers, and argument
2066/// matchers. Higher arity matcher are defined recursively in terms of and-ing
2067/// them with lower arity matchers. Here's some convenient typedefs for up to
2068/// several arguments, and more can be added as needed
2069template <typename T0 = void, typename T1 = void, typename T2 = void,
2070 typename T3 = void, typename T4 = void, typename T5 = void,
2071 typename T6 = void, typename T7 = void, typename T8 = void,
2072 typename T9 = void, typename T10 = void>
2073struct m_Intrinsic_Ty;
2074template <typename T0> struct m_Intrinsic_Ty<T0> {
2075 using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>;
2076};
2077template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> {
2078 using Ty =
2079 match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>;
2080};
2081template <typename T0, typename T1, typename T2>
2082struct m_Intrinsic_Ty<T0, T1, T2> {
2083 using Ty =
2084 match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty,
2085 Argument_match<T2>>;
2086};
2087template <typename T0, typename T1, typename T2, typename T3>
2088struct m_Intrinsic_Ty<T0, T1, T2, T3> {
2089 using Ty =
2090 match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty,
2091 Argument_match<T3>>;
2092};
2093
2094template <typename T0, typename T1, typename T2, typename T3, typename T4>
2095struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> {
2096 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty,
2097 Argument_match<T4>>;
2098};
2099
2100template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5>
2101struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> {
2102 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty,
2103 Argument_match<T5>>;
2104};
2105
2106/// Match intrinsic calls like this:
2107/// m_Intrinsic<Intrinsic::fabs>(m_Value(X))
2108template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() {
2109 return IntrinsicID_match(IntrID);
2110}
2111
2112/// Matches MaskedLoad Intrinsic.
2113template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3>
2114inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty
2115m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2,
2116 const Opnd3 &Op3) {
2117 return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3);
2118}
2119
2120/// Matches MaskedGather Intrinsic.
2121template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3>
2122inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty
2123m_MaskedGather(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2,
2124 const Opnd3 &Op3) {
2125 return m_Intrinsic<Intrinsic::masked_gather>(Op0, Op1, Op2, Op3);
2126}
2127
2128template <Intrinsic::ID IntrID, typename T0>
2129inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) {
2130 return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0));
2131}
2132
2133template <Intrinsic::ID IntrID, typename T0, typename T1>
2134inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0,
2135 const T1 &Op1) {
2136 return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1));
2137}
2138
2139template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2>
2140inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty
2141m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) {
2142 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2));
2143}
2144
2145template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2146 typename T3>
2147inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty
2148m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) {
2149 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3));
2150}
2151
2152template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2153 typename T3, typename T4>
2154inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty
2155m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2156 const T4 &Op4) {
2157 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3),
2158 m_Argument<4>(Op4));
2159}
2160
2161template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2162 typename T3, typename T4, typename T5>
2163inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty
2164m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2165 const T4 &Op4, const T5 &Op5) {
2166 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4),
2167 m_Argument<5>(Op5));
2168}
2169
2170// Helper intrinsic matching specializations.
2171template <typename Opnd0>
2172inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) {
2173 return m_Intrinsic<Intrinsic::bitreverse>(Op0);
2174}
2175
2176template <typename Opnd0>
2177inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) {
2178 return m_Intrinsic<Intrinsic::bswap>(Op0);
2179}
2180
2181template <typename Opnd0>
2182inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) {
2183 return m_Intrinsic<Intrinsic::fabs>(Op0);
2184}
2185
2186template <typename Opnd0>
2187inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) {
2188 return m_Intrinsic<Intrinsic::canonicalize>(Op0);
2189}
2190
2191template <typename Opnd0, typename Opnd1>
2192inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0,
2193 const Opnd1 &Op1) {
2194 return m_Intrinsic<Intrinsic::minnum>(Op0, Op1);
2195}
2196
2197template <typename Opnd0, typename Opnd1>
2198inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0,
2199 const Opnd1 &Op1) {
2200 return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1);
2201}
2202
2203template <typename Opnd0, typename Opnd1, typename Opnd2>
2204inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2205m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2206 return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2);
2207}
2208
2209template <typename Opnd0, typename Opnd1, typename Opnd2>
2210inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2211m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2212 return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2);
2213}
2214
2215//===----------------------------------------------------------------------===//
2216// Matchers for two-operands operators with the operators in either order
2217//
2218
2219/// Matches a BinaryOperator with LHS and RHS in either order.
2220template <typename LHS, typename RHS>
2221inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) {
2222 return AnyBinaryOp_match<LHS, RHS, true>(L, R);
11
Returning without writing to 'R.VR'
2223}
2224
2225/// Matches an ICmp with a predicate over LHS and RHS in either order.
2226/// Swaps the predicate if operands are commuted.
2227template <typename LHS, typename RHS>
2228inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>
2229m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
2230 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L,
2231 R);
2232}
2233
2234/// Matches a specific opcode with LHS and RHS in either order.
2235template <typename LHS, typename RHS>
2236inline SpecificBinaryOp_match<LHS, RHS, true>
2237m_c_BinOp(unsigned Opcode, const LHS &L, const RHS &R) {
2238 return SpecificBinaryOp_match<LHS, RHS, true>(Opcode, L, R);
2239}
2240
2241/// Matches a Add with LHS and RHS in either order.
2242template <typename LHS, typename RHS>
2243inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L,
2244 const RHS &R) {
2245 return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R);
2246}
2247
2248/// Matches a Mul with LHS and RHS in either order.
2249template <typename LHS, typename RHS>
2250inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L,
2251 const RHS &R) {
2252 return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R);
2253}
2254
2255/// Matches an And with LHS and RHS in either order.
2256template <typename LHS, typename RHS>
2257inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L,
2258 const RHS &R) {
2259 return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R);
2260}
2261
2262/// Matches an Or with LHS and RHS in either order.
2263template <typename LHS, typename RHS>
2264inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L,
2265 const RHS &R) {
2266 return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R);
2267}
2268
2269/// Matches an Xor with LHS and RHS in either order.
2270template <typename LHS, typename RHS>
2271inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L,
2272 const RHS &R) {
2273 return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R);
2274}
2275
2276/// Matches a 'Neg' as 'sub 0, V'.
2277template <typename ValTy>
2278inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub>
2279m_Neg(const ValTy &V) {
2280 return m_Sub(m_ZeroInt(), V);
2281}
2282
2283/// Matches a 'Neg' as 'sub nsw 0, V'.
2284template <typename ValTy>
2285inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy,
2286 Instruction::Sub,
2287 OverflowingBinaryOperator::NoSignedWrap>
2288m_NSWNeg(const ValTy &V) {
2289 return m_NSWSub(m_ZeroInt(), V);
2290}
2291
2292/// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
2293template <typename ValTy>
2294inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true>
2295m_Not(const ValTy &V) {
2296 return m_c_Xor(V, m_AllOnes());
2297}
2298
2299template <typename ValTy> struct NotForbidUndef_match {
2300 ValTy Val;
2301 NotForbidUndef_match(const ValTy &V) : Val(V) {}
2302
2303 template <typename OpTy> bool match(OpTy *V) {
2304 // We do not use m_c_Xor because that could match an arbitrary APInt that is
2305 // not -1 as C and then fail to match the other operand if it is -1.
2306 // This code should still work even when both operands are constants.
2307 Value *X;
2308 const APInt *C;
2309 if (m_Xor(m_Value(X), m_APIntForbidUndef(C)).match(V) && C->isAllOnes())
2310 return Val.match(X);
2311 if (m_Xor(m_APIntForbidUndef(C), m_Value(X)).match(V) && C->isAllOnes())
2312 return Val.match(X);
2313 return false;
2314 }
2315};
2316
2317/// Matches a bitwise 'not' as 'xor V, -1' or 'xor -1, V'. For vectors, the
2318/// constant value must be composed of only -1 scalar elements.
2319template <typename ValTy>
2320inline NotForbidUndef_match<ValTy> m_NotForbidUndef(const ValTy &V) {
2321 return NotForbidUndef_match<ValTy>(V);
2322}
2323
2324/// Matches an SMin with LHS and RHS in either order.
2325template <typename LHS, typename RHS>
2326inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>
2327m_c_SMin(const LHS &L, const RHS &R) {
2328 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R);
2329}
2330/// Matches an SMax with LHS and RHS in either order.
2331template <typename LHS, typename RHS>
2332inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>
2333m_c_SMax(const LHS &L, const RHS &R) {
2334 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R);
2335}
2336/// Matches a UMin with LHS and RHS in either order.
2337template <typename LHS, typename RHS>
2338inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>
2339m_c_UMin(const LHS &L, const RHS &R) {
2340 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R);
2341}
2342/// Matches a UMax with LHS and RHS in either order.
2343template <typename LHS, typename RHS>
2344inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>
2345m_c_UMax(const LHS &L, const RHS &R) {
2346 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R);
2347}
2348
2349template <typename LHS, typename RHS>
2350inline match_combine_or<
2351 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>,
2352 MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>,
2353 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>,
2354 MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>>
2355m_c_MaxOrMin(const LHS &L, const RHS &R) {
2356 return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)),
2357 m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R)));
2358}
2359
2360/// Matches FAdd with LHS and RHS in either order.
2361template <typename LHS, typename RHS>
2362inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true>
2363m_c_FAdd(const LHS &L, const RHS &R) {
2364 return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R);
2365}
2366
2367/// Matches FMul with LHS and RHS in either order.
2368template <typename LHS, typename RHS>
2369inline BinaryOp_match<LHS, RHS, Instruction::FMul, true>
2370m_c_FMul(const LHS &L, const RHS &R) {
2371 return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R);
2372}
2373
2374template <typename Opnd_t> struct Signum_match {
2375 Opnd_t Val;
2376 Signum_match(const Opnd_t &V) : Val(V) {}
2377
2378 template <typename OpTy> bool match(OpTy *V) {
2379 unsigned TypeSize = V->getType()->getScalarSizeInBits();
2380 if (TypeSize == 0)
2381 return false;
2382
2383 unsigned ShiftWidth = TypeSize - 1;
2384 Value *OpL = nullptr, *OpR = nullptr;
2385
2386 // This is the representation of signum we match:
2387 //
2388 // signum(x) == (x >> 63) | (-x >>u 63)
2389 //
2390 // An i1 value is its own signum, so it's correct to match
2391 //
2392 // signum(x) == (x >> 0) | (-x >>u 0)
2393 //
2394 // for i1 values.
2395
2396 auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth));
2397 auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth));
2398 auto Signum = m_Or(LHS, RHS);
2399
2400 return Signum.match(V) && OpL == OpR && Val.match(OpL);
2401 }
2402};
2403
2404/// Matches a signum pattern.
2405///
2406/// signum(x) =
2407/// x > 0 -> 1
2408/// x == 0 -> 0
2409/// x < 0 -> -1
2410template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) {
2411 return Signum_match<Val_t>(V);
2412}
2413
2414template <int Ind, typename Opnd_t> struct ExtractValue_match {
2415 Opnd_t Val;
2416 ExtractValue_match(const Opnd_t &V) : Val(V) {}
2417
2418 template <typename OpTy> bool match(OpTy *V) {
2419 if (auto *I = dyn_cast<ExtractValueInst>(V)) {
2420 // If Ind is -1, don't inspect indices
2421 if (Ind != -1 &&
2422 !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind))
2423 return false;
2424 return Val.match(I->getAggregateOperand());
2425 }
2426 return false;
2427 }
2428};
2429
2430/// Match a single index ExtractValue instruction.
2431/// For example m_ExtractValue<1>(...)
2432template <int Ind, typename Val_t>
2433inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) {
2434 return ExtractValue_match<Ind, Val_t>(V);
2435}
2436
2437/// Match an ExtractValue instruction with any index.
2438/// For example m_ExtractValue(...)
2439template <typename Val_t>
2440inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) {
2441 return ExtractValue_match<-1, Val_t>(V);
2442}
2443
2444/// Matcher for a single index InsertValue instruction.
2445template <int Ind, typename T0, typename T1> struct InsertValue_match {
2446 T0 Op0;
2447 T1 Op1;
2448
2449 InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {}
2450
2451 template <typename OpTy> bool match(OpTy *V) {
2452 if (auto *I = dyn_cast<InsertValueInst>(V)) {
2453 return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) &&
2454 I->getNumIndices() == 1 && Ind == I->getIndices()[0];
2455 }
2456 return false;
2457 }
2458};
2459
2460/// Matches a single index InsertValue instruction.
2461template <int Ind, typename Val_t, typename Elt_t>
2462inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val,
2463 const Elt_t &Elt) {
2464 return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt);
2465}
2466
2467/// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or
2468/// the constant expression
2469/// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>`
2470/// under the right conditions determined by DataLayout.
2471struct VScaleVal_match {
2472 const DataLayout &DL;
2473 VScaleVal_match(const DataLayout &DL) : DL(DL) {}
2474
2475 template <typename ITy> bool match(ITy *V) {
2476 if (m_Intrinsic<Intrinsic::vscale>().match(V))
2477 return true;
2478
2479 Value *Ptr;
2480 if (m_PtrToInt(m_Value(Ptr)).match(V)) {
2481 if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
2482 auto *DerefTy = GEP->getSourceElementType();
2483 if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) &&
2484 m_Zero().match(GEP->getPointerOperand()) &&
2485 m_SpecificInt(1).match(GEP->idx_begin()->get()) &&
2486 DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8)
2487 return true;
2488 }
2489 }
2490
2491 return false;
2492 }
2493};
2494
2495inline VScaleVal_match m_VScale(const DataLayout &DL) {
2496 return VScaleVal_match(DL);
2497}
2498
2499template <typename LHS, typename RHS, unsigned Opcode, bool Commutable = false>
2500struct LogicalOp_match {
2501 LHS L;
2502 RHS R;
2503
2504 LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {}
2505
2506 template <typename T> bool match(T *V) {
2507 auto *I = dyn_cast<Instruction>(V);
2508 if (!I || !I->getType()->isIntOrIntVectorTy(1))
2509 return false;
2510
2511 if (I->getOpcode() == Opcode) {
2512 auto *Op0 = I->getOperand(0);
2513 auto *Op1 = I->getOperand(1);
2514 return (L.match(Op0) && R.match(Op1)) ||
2515 (Commutable && L.match(Op1) && R.match(Op0));
2516 }
2517
2518 if (auto *Select = dyn_cast<SelectInst>(I)) {
2519 auto *Cond = Select->getCondition();
2520 auto *TVal = Select->getTrueValue();
2521 auto *FVal = Select->getFalseValue();
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>
2544m_LogicalAnd(const LHS &L, 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>
2562m_LogicalOr(const LHS &L, 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