Bug Summary

File:llvm/include/llvm/IR/PatternMatch.h
Warning:line 1233, column 9
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -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 -mthread-model posix -mframe-pointer=none -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -target-cpu x86-64 -dwarf-column-info -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/include -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-01-13-084841-49055-1 -x c++ /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/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 <cassert>
33#include <utility>
34
35using namespace llvm;
36using namespace PatternMatch;
37
38#define DEBUG_TYPE"instcombine" "instcombine"
39
40namespace {
41
42 /// Class representing coefficient of floating-point addend.
43 /// This class needs to be highly efficient, which is especially true for
44 /// the constructor. As of I write this comment, the cost of the default
45 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
46 /// perform write-merging).
47 ///
48 class FAddendCoef {
49 public:
50 // The constructor has to initialize a APFloat, which is unnecessary for
51 // most addends which have coefficient either 1 or -1. So, the constructor
52 // is expensive. In order to avoid the cost of the constructor, we should
53 // reuse some instances whenever possible. The pre-created instances
54 // FAddCombine::Add[0-5] embodies this idea.
55 FAddendCoef() = default;
56 ~FAddendCoef();
57
58 // If possible, don't define operator+/operator- etc because these
59 // operators inevitably call FAddendCoef's constructor which is not cheap.
60 void operator=(const FAddendCoef &A);
61 void operator+=(const FAddendCoef &A);
62 void operator*=(const FAddendCoef &S);
63
64 void set(short C) {
65 assert(!insaneIntVal(C) && "Insane coefficient")((!insaneIntVal(C) && "Insane coefficient") ? static_cast
<void> (0) : __assert_fail ("!insaneIntVal(C) && \"Insane coefficient\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 65, __PRETTY_FUNCTION__))
;
66 IsFp = false; IntVal = C;
67 }
68
69 void set(const APFloat& C);
70
71 void negate();
72
73 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
74 Value *getValue(Type *) const;
75
76 bool isOne() const { return isInt() && IntVal == 1; }
77 bool isTwo() const { return isInt() && IntVal == 2; }
78 bool isMinusOne() const { return isInt() && IntVal == -1; }
79 bool isMinusTwo() const { return isInt() && IntVal == -2; }
80
81 private:
82 bool insaneIntVal(int V) { return V > 4 || V < -4; }
83
84 APFloat *getFpValPtr()
85 { return reinterpret_cast<APFloat *>(&FpValBuf.buffer[0]); }
86
87 const APFloat *getFpValPtr() const
88 { return reinterpret_cast<const APFloat *>(&FpValBuf.buffer[0]); }
89
90 const APFloat &getFpVal() const {
91 assert(IsFp && BufHasFpVal && "Incorret state")((IsFp && BufHasFpVal && "Incorret state") ? static_cast
<void> (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 91, __PRETTY_FUNCTION__))
;
92 return *getFpValPtr();
93 }
94
95 APFloat &getFpVal() {
96 assert(IsFp && BufHasFpVal && "Incorret state")((IsFp && BufHasFpVal && "Incorret state") ? static_cast
<void> (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 96, __PRETTY_FUNCTION__))
;
97 return *getFpValPtr();
98 }
99
100 bool isInt() const { return !IsFp; }
101
102 // If the coefficient is represented by an integer, promote it to a
103 // floating point.
104 void convertToFpType(const fltSemantics &Sem);
105
106 // Construct an APFloat from a signed integer.
107 // TODO: We should get rid of this function when APFloat can be constructed
108 // from an *SIGNED* integer.
109 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
110
111 bool IsFp = false;
112
113 // True iff FpValBuf contains an instance of APFloat.
114 bool BufHasFpVal = false;
115
116 // The integer coefficient of an individual addend is either 1 or -1,
117 // and we try to simplify at most 4 addends from neighboring at most
118 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
119 // is overkill of this end.
120 short IntVal = 0;
121
122 AlignedCharArrayUnion<APFloat> FpValBuf;
123 };
124
125 /// FAddend is used to represent floating-point addend. An addend is
126 /// represented as <C, V>, where the V is a symbolic value, and C is a
127 /// constant coefficient. A constant addend is represented as <C, 0>.
128 class FAddend {
129 public:
130 FAddend() = default;
131
132 void operator+=(const FAddend &T) {
133 assert((Val == T.Val) && "Symbolic-values disagree")(((Val == T.Val) && "Symbolic-values disagree") ? static_cast
<void> (0) : __assert_fail ("(Val == T.Val) && \"Symbolic-values disagree\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 133, __PRETTY_FUNCTION__))
;
134 Coeff += T.Coeff;
135 }
136
137 Value *getSymVal() const { return Val; }
138 const FAddendCoef &getCoef() const { return Coeff; }
139
140 bool isConstant() const { return Val == nullptr; }
141 bool isZero() const { return Coeff.isZero(); }
142
143 void set(short Coefficient, Value *V) {
144 Coeff.set(Coefficient);
145 Val = V;
146 }
147 void set(const APFloat &Coefficient, Value *V) {
148 Coeff.set(Coefficient);
149 Val = V;
150 }
151 void set(const ConstantFP *Coefficient, Value *V) {
152 Coeff.set(Coefficient->getValueAPF());
153 Val = V;
154 }
155
156 void negate() { Coeff.negate(); }
157
158 /// Drill down the U-D chain one step to find the definition of V, and
159 /// try to break the definition into one or two addends.
160 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
161
162 /// Similar to FAddend::drillDownOneStep() except that the value being
163 /// splitted is the addend itself.
164 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
165
166 private:
167 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
168
169 // This addend has the value of "Coeff * Val".
170 Value *Val = nullptr;
171 FAddendCoef Coeff;
172 };
173
174 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
175 /// with its neighboring at most two instructions.
176 ///
177 class FAddCombine {
178 public:
179 FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
180
181 Value *simplify(Instruction *FAdd);
182
183 private:
184 using AddendVect = SmallVector<const FAddend *, 4>;
185
186 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
187
188 /// Convert given addend to a Value
189 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
190
191 /// Return the number of instructions needed to emit the N-ary addition.
192 unsigned calcInstrNumber(const AddendVect& Vect);
193
194 Value *createFSub(Value *Opnd0, Value *Opnd1);
195 Value *createFAdd(Value *Opnd0, Value *Opnd1);
196 Value *createFMul(Value *Opnd0, Value *Opnd1);
197 Value *createFNeg(Value *V);
198 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
199 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
200
201 // Debugging stuff are clustered here.
202 #ifndef NDEBUG
203 unsigned CreateInstrNum;
204 void initCreateInstNum() { CreateInstrNum = 0; }
205 void incCreateInstNum() { CreateInstrNum++; }
206 #else
207 void initCreateInstNum() {}
208 void incCreateInstNum() {}
209 #endif
210
211 InstCombiner::BuilderTy &Builder;
212 Instruction *Instr = nullptr;
213 };
214
215} // end anonymous namespace
216
217//===----------------------------------------------------------------------===//
218//
219// Implementation of
220// {FAddendCoef, FAddend, FAddition, FAddCombine}.
221//
222//===----------------------------------------------------------------------===//
223FAddendCoef::~FAddendCoef() {
224 if (BufHasFpVal)
225 getFpValPtr()->~APFloat();
226}
227
228void FAddendCoef::set(const APFloat& C) {
229 APFloat *P = getFpValPtr();
230
231 if (isInt()) {
232 // As the buffer is meanless byte stream, we cannot call
233 // APFloat::operator=().
234 new(P) APFloat(C);
235 } else
236 *P = C;
237
238 IsFp = BufHasFpVal = true;
239}
240
241void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
242 if (!isInt())
243 return;
244
245 APFloat *P = getFpValPtr();
246 if (IntVal > 0)
247 new(P) APFloat(Sem, IntVal);
248 else {
249 new(P) APFloat(Sem, 0 - IntVal);
250 P->changeSign();
251 }
252 IsFp = BufHasFpVal = true;
253}
254
255APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
256 if (Val >= 0)
257 return APFloat(Sem, Val);
258
259 APFloat T(Sem, 0 - Val);
260 T.changeSign();
261
262 return T;
263}
264
265void FAddendCoef::operator=(const FAddendCoef &That) {
266 if (That.isInt())
267 set(That.IntVal);
268 else
269 set(That.getFpVal());
270}
271
272void FAddendCoef::operator+=(const FAddendCoef &That) {
273 enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
274 if (isInt() == That.isInt()) {
275 if (isInt())
276 IntVal += That.IntVal;
277 else
278 getFpVal().add(That.getFpVal(), RndMode);
279 return;
280 }
281
282 if (isInt()) {
283 const APFloat &T = That.getFpVal();
284 convertToFpType(T.getSemantics());
285 getFpVal().add(T, RndMode);
286 return;
287 }
288
289 APFloat &T = getFpVal();
290 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
291}
292
293void FAddendCoef::operator*=(const FAddendCoef &That) {
294 if (That.isOne())
295 return;
296
297 if (That.isMinusOne()) {
298 negate();
299 return;
300 }
301
302 if (isInt() && That.isInt()) {
303 int Res = IntVal * (int)That.IntVal;
304 assert(!insaneIntVal(Res) && "Insane int value")((!insaneIntVal(Res) && "Insane int value") ? static_cast
<void> (0) : __assert_fail ("!insaneIntVal(Res) && \"Insane int value\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 304, __PRETTY_FUNCTION__))
;
305 IntVal = Res;
306 return;
307 }
308
309 const fltSemantics &Semantic =
310 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
311
312 if (isInt())
313 convertToFpType(Semantic);
314 APFloat &F0 = getFpVal();
315
316 if (That.isInt())
317 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
318 APFloat::rmNearestTiesToEven);
319 else
320 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
321}
322
323void FAddendCoef::negate() {
324 if (isInt())
325 IntVal = 0 - IntVal;
326 else
327 getFpVal().changeSign();
328}
329
330Value *FAddendCoef::getValue(Type *Ty) const {
331 return isInt() ?
332 ConstantFP::get(Ty, float(IntVal)) :
333 ConstantFP::get(Ty->getContext(), getFpVal());
334}
335
336// The definition of <Val> Addends
337// =========================================
338// A + B <1, A>, <1,B>
339// A - B <1, A>, <1,B>
340// 0 - B <-1, B>
341// C * A, <C, A>
342// A + C <1, A> <C, NULL>
343// 0 +/- 0 <0, NULL> (corner case)
344//
345// Legend: A and B are not constant, C is constant
346unsigned FAddend::drillValueDownOneStep
347 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
348 Instruction *I = nullptr;
349 if (!Val || !(I = dyn_cast<Instruction>(Val)))
350 return 0;
351
352 unsigned Opcode = I->getOpcode();
353
354 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
355 ConstantFP *C0, *C1;
356 Value *Opnd0 = I->getOperand(0);
357 Value *Opnd1 = I->getOperand(1);
358 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
359 Opnd0 = nullptr;
360
361 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
362 Opnd1 = nullptr;
363
364 if (Opnd0) {
365 if (!C0)
366 Addend0.set(1, Opnd0);
367 else
368 Addend0.set(C0, nullptr);
369 }
370
371 if (Opnd1) {
372 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
373 if (!C1)
374 Addend.set(1, Opnd1);
375 else
376 Addend.set(C1, nullptr);
377 if (Opcode == Instruction::FSub)
378 Addend.negate();
379 }
380
381 if (Opnd0 || Opnd1)
382 return Opnd0 && Opnd1 ? 2 : 1;
383
384 // Both operands are zero. Weird!
385 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
386 return 1;
387 }
388
389 if (I->getOpcode() == Instruction::FMul) {
390 Value *V0 = I->getOperand(0);
391 Value *V1 = I->getOperand(1);
392 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
393 Addend0.set(C, V1);
394 return 1;
395 }
396
397 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
398 Addend0.set(C, V0);
399 return 1;
400 }
401 }
402
403 return 0;
404}
405
406// Try to break *this* addend into two addends. e.g. Suppose this addend is
407// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
408// i.e. <2.3, X> and <2.3, Y>.
409unsigned FAddend::drillAddendDownOneStep
410 (FAddend &Addend0, FAddend &Addend1) const {
411 if (isConstant())
412 return 0;
413
414 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
415 if (!BreakNum || Coeff.isOne())
416 return BreakNum;
417
418 Addend0.Scale(Coeff);
419
420 if (BreakNum == 2)
421 Addend1.Scale(Coeff);
422
423 return BreakNum;
424}
425
426Value *FAddCombine::simplify(Instruction *I) {
427 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&((I->hasAllowReassoc() && I->hasNoSignedZeros()
&& "Expected 'reassoc'+'nsz' instruction") ? static_cast
<void> (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 428, __PRETTY_FUNCTION__))
428 "Expected 'reassoc'+'nsz' instruction")((I->hasAllowReassoc() && I->hasNoSignedZeros()
&& "Expected 'reassoc'+'nsz' instruction") ? static_cast
<void> (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 428, __PRETTY_FUNCTION__))
;
429
430 // Currently we are not able to handle vector type.
431 if (I->getType()->isVectorTy())
432 return nullptr;
433
434 assert((I->getOpcode() == Instruction::FAdd ||(((I->getOpcode() == Instruction::FAdd || I->getOpcode(
) == Instruction::FSub) && "Expect add/sub") ? static_cast
<void> (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 435, __PRETTY_FUNCTION__))
435 I->getOpcode() == Instruction::FSub) && "Expect add/sub")(((I->getOpcode() == Instruction::FAdd || I->getOpcode(
) == Instruction::FSub) && "Expect add/sub") ? static_cast
<void> (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 435, __PRETTY_FUNCTION__))
;
436
437 // Save the instruction before calling other member-functions.
438 Instr = I;
439
440 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
441
442 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
443
444 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
445 unsigned Opnd0_ExpNum = 0;
446 unsigned Opnd1_ExpNum = 0;
447
448 if (!Opnd0.isConstant())
449 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
450
451 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
452 if (OpndNum == 2 && !Opnd1.isConstant())
453 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
454
455 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
456 if (Opnd0_ExpNum && Opnd1_ExpNum) {
457 AddendVect AllOpnds;
458 AllOpnds.push_back(&Opnd0_0);
459 AllOpnds.push_back(&Opnd1_0);
460 if (Opnd0_ExpNum == 2)
461 AllOpnds.push_back(&Opnd0_1);
462 if (Opnd1_ExpNum == 2)
463 AllOpnds.push_back(&Opnd1_1);
464
465 // Compute instruction quota. We should save at least one instruction.
466 unsigned InstQuota = 0;
467
468 Value *V0 = I->getOperand(0);
469 Value *V1 = I->getOperand(1);
470 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
471 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
472
473 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
474 return R;
475 }
476
477 if (OpndNum != 2) {
478 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
479 // splitted into two addends, say "V = X - Y", the instruction would have
480 // been optimized into "I = Y - X" in the previous steps.
481 //
482 const FAddendCoef &CE = Opnd0.getCoef();
483 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
484 }
485
486 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
487 if (Opnd1_ExpNum) {
488 AddendVect AllOpnds;
489 AllOpnds.push_back(&Opnd0);
490 AllOpnds.push_back(&Opnd1_0);
491 if (Opnd1_ExpNum == 2)
492 AllOpnds.push_back(&Opnd1_1);
493
494 if (Value *R = simplifyFAdd(AllOpnds, 1))
495 return R;
496 }
497
498 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
499 if (Opnd0_ExpNum) {
500 AddendVect AllOpnds;
501 AllOpnds.push_back(&Opnd1);
502 AllOpnds.push_back(&Opnd0_0);
503 if (Opnd0_ExpNum == 2)
504 AllOpnds.push_back(&Opnd0_1);
505
506 if (Value *R = simplifyFAdd(AllOpnds, 1))
507 return R;
508 }
509
510 return nullptr;
511}
512
513Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
514 unsigned AddendNum = Addends.size();
515 assert(AddendNum <= 4 && "Too many addends")((AddendNum <= 4 && "Too many addends") ? static_cast
<void> (0) : __assert_fail ("AddendNum <= 4 && \"Too many addends\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 515, __PRETTY_FUNCTION__))
;
516
517 // For saving intermediate results;
518 unsigned NextTmpIdx = 0;
519 FAddend TmpResult[3];
520
521 // Points to the constant addend of the resulting simplified expression.
522 // If the resulting expr has constant-addend, this constant-addend is
523 // desirable to reside at the top of the resulting expression tree. Placing
524 // constant close to supper-expr(s) will potentially reveal some optimization
525 // opportunities in super-expr(s).
526 const FAddend *ConstAdd = nullptr;
527
528 // Simplified addends are placed <SimpVect>.
529 AddendVect SimpVect;
530
531 // The outer loop works on one symbolic-value at a time. Suppose the input
532 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
533 // The symbolic-values will be processed in this order: x, y, z.
534 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
535
536 const FAddend *ThisAddend = Addends[SymIdx];
537 if (!ThisAddend) {
538 // This addend was processed before.
539 continue;
540 }
541
542 Value *Val = ThisAddend->getSymVal();
543 unsigned StartIdx = SimpVect.size();
544 SimpVect.push_back(ThisAddend);
545
546 // The inner loop collects addends sharing same symbolic-value, and these
547 // addends will be later on folded into a single addend. Following above
548 // example, if the symbolic value "y" is being processed, the inner loop
549 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
550 // be later on folded into "<b1+b2, y>".
551 for (unsigned SameSymIdx = SymIdx + 1;
552 SameSymIdx < AddendNum; SameSymIdx++) {
553 const FAddend *T = Addends[SameSymIdx];
554 if (T && T->getSymVal() == Val) {
555 // Set null such that next iteration of the outer loop will not process
556 // this addend again.
557 Addends[SameSymIdx] = nullptr;
558 SimpVect.push_back(T);
559 }
560 }
561
562 // If multiple addends share same symbolic value, fold them together.
563 if (StartIdx + 1 != SimpVect.size()) {
564 FAddend &R = TmpResult[NextTmpIdx ++];
565 R = *SimpVect[StartIdx];
566 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
567 R += *SimpVect[Idx];
568
569 // Pop all addends being folded and push the resulting folded addend.
570 SimpVect.resize(StartIdx);
571 if (Val) {
572 if (!R.isZero()) {
573 SimpVect.push_back(&R);
574 }
575 } else {
576 // Don't push constant addend at this time. It will be the last element
577 // of <SimpVect>.
578 ConstAdd = &R;
579 }
580 }
581 }
582
583 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&(((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
"out-of-bound access") ? static_cast<void> (0) : __assert_fail
("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 584, __PRETTY_FUNCTION__))
584 "out-of-bound access")(((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
"out-of-bound access") ? static_cast<void> (0) : __assert_fail
("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 584, __PRETTY_FUNCTION__))
;
585
586 if (ConstAdd)
587 SimpVect.push_back(ConstAdd);
588
589 Value *Result;
590 if (!SimpVect.empty())
591 Result = createNaryFAdd(SimpVect, InstrQuota);
592 else {
593 // The addition is folded to 0.0.
594 Result = ConstantFP::get(Instr->getType(), 0.0);
595 }
596
597 return Result;
598}
599
600Value *FAddCombine::createNaryFAdd
601 (const AddendVect &Opnds, unsigned InstrQuota) {
602 assert(!Opnds.empty() && "Expect at least one addend")((!Opnds.empty() && "Expect at least one addend") ? static_cast
<void> (0) : __assert_fail ("!Opnds.empty() && \"Expect at least one addend\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 602, __PRETTY_FUNCTION__))
;
603
604 // Step 1: Check if the # of instructions needed exceeds the quota.
605
606 unsigned InstrNeeded = calcInstrNumber(Opnds);
607 if (InstrNeeded > InstrQuota)
608 return nullptr;
609
610 initCreateInstNum();
611
612 // step 2: Emit the N-ary addition.
613 // Note that at most three instructions are involved in Fadd-InstCombine: the
614 // addition in question, and at most two neighboring instructions.
615 // The resulting optimized addition should have at least one less instruction
616 // than the original addition expression tree. This implies that the resulting
617 // N-ary addition has at most two instructions, and we don't need to worry
618 // about tree-height when constructing the N-ary addition.
619
620 Value *LastVal = nullptr;
621 bool LastValNeedNeg = false;
622
623 // Iterate the addends, creating fadd/fsub using adjacent two addends.
624 for (const FAddend *Opnd : Opnds) {
625 bool NeedNeg;
626 Value *V = createAddendVal(*Opnd, NeedNeg);
627 if (!LastVal) {
628 LastVal = V;
629 LastValNeedNeg = NeedNeg;
630 continue;
631 }
632
633 if (LastValNeedNeg == NeedNeg) {
634 LastVal = createFAdd(LastVal, V);
635 continue;
636 }
637
638 if (LastValNeedNeg)
639 LastVal = createFSub(V, LastVal);
640 else
641 LastVal = createFSub(LastVal, V);
642
643 LastValNeedNeg = false;
644 }
645
646 if (LastValNeedNeg) {
647 LastVal = createFNeg(LastVal);
648 }
649
650#ifndef NDEBUG
651 assert(CreateInstrNum == InstrNeeded &&((CreateInstrNum == InstrNeeded && "Inconsistent in instruction numbers"
) ? static_cast<void> (0) : __assert_fail ("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 652, __PRETTY_FUNCTION__))
652 "Inconsistent in instruction numbers")((CreateInstrNum == InstrNeeded && "Inconsistent in instruction numbers"
) ? static_cast<void> (0) : __assert_fail ("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 652, __PRETTY_FUNCTION__))
;
653#endif
654
655 return LastVal;
656}
657
658Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
659 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
660 if (Instruction *I = dyn_cast<Instruction>(V))
661 createInstPostProc(I);
662 return V;
663}
664
665Value *FAddCombine::createFNeg(Value *V) {
666 Value *Zero = cast<Value>(ConstantFP::getZeroValueForNegation(V->getType()));
667 Value *NewV = createFSub(Zero, V);
668 if (Instruction *I = dyn_cast<Instruction>(NewV))
669 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
670 return NewV;
671}
672
673Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
674 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
675 if (Instruction *I = dyn_cast<Instruction>(V))
676 createInstPostProc(I);
677 return V;
678}
679
680Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
681 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
682 if (Instruction *I = dyn_cast<Instruction>(V))
683 createInstPostProc(I);
684 return V;
685}
686
687void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
688 NewInstr->setDebugLoc(Instr->getDebugLoc());
689
690 // Keep track of the number of instruction created.
691 if (!NoNumber)
692 incCreateInstNum();
693
694 // Propagate fast-math flags
695 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
696}
697
698// Return the number of instruction needed to emit the N-ary addition.
699// NOTE: Keep this function in sync with createAddendVal().
700unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
701 unsigned OpndNum = Opnds.size();
702 unsigned InstrNeeded = OpndNum - 1;
703
704 // The number of addends in the form of "(-1)*x".
705 unsigned NegOpndNum = 0;
706
707 // Adjust the number of instructions needed to emit the N-ary add.
708 for (const FAddend *Opnd : Opnds) {
709 if (Opnd->isConstant())
710 continue;
711
712 // The constant check above is really for a few special constant
713 // coefficients.
714 if (isa<UndefValue>(Opnd->getSymVal()))
715 continue;
716
717 const FAddendCoef &CE = Opnd->getCoef();
718 if (CE.isMinusOne() || CE.isMinusTwo())
719 NegOpndNum++;
720
721 // Let the addend be "c * x". If "c == +/-1", the value of the addend
722 // is immediately available; otherwise, it needs exactly one instruction
723 // to evaluate the value.
724 if (!CE.isMinusOne() && !CE.isOne())
725 InstrNeeded++;
726 }
727 if (NegOpndNum == OpndNum)
728 InstrNeeded++;
729 return InstrNeeded;
730}
731
732// Input Addend Value NeedNeg(output)
733// ================================================================
734// Constant C C false
735// <+/-1, V> V coefficient is -1
736// <2/-2, V> "fadd V, V" coefficient is -2
737// <C, V> "fmul V, C" false
738//
739// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
740Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
741 const FAddendCoef &Coeff = Opnd.getCoef();
742
743 if (Opnd.isConstant()) {
744 NeedNeg = false;
745 return Coeff.getValue(Instr->getType());
746 }
747
748 Value *OpndVal = Opnd.getSymVal();
749
750 if (Coeff.isMinusOne() || Coeff.isOne()) {
751 NeedNeg = Coeff.isMinusOne();
752 return OpndVal;
753 }
754
755 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
756 NeedNeg = Coeff.isMinusTwo();
757 return createFAdd(OpndVal, OpndVal);
758 }
759
760 NeedNeg = false;
761 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
762}
763
764// Checks if any operand is negative and we can convert add to sub.
765// This function checks for following negative patterns
766// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
767// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
768// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
769static Value *checkForNegativeOperand(BinaryOperator &I,
770 InstCombiner::BuilderTy &Builder) {
771 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
772
773 // This function creates 2 instructions to replace ADD, we need at least one
774 // of LHS or RHS to have one use to ensure benefit in transform.
775 if (!LHS->hasOneUse() && !RHS->hasOneUse())
776 return nullptr;
777
778 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
779 const APInt *C1 = nullptr, *C2 = nullptr;
780
781 // if ONE is on other side, swap
782 if (match(RHS, m_Add(m_Value(X), m_One())))
783 std::swap(LHS, RHS);
784
785 if (match(LHS, m_Add(m_Value(X), m_One()))) {
786 // if XOR on other side, swap
787 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
788 std::swap(X, RHS);
789
790 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
791 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
792 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
793 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
794 Value *NewAnd = Builder.CreateAnd(Z, *C1);
795 return Builder.CreateSub(RHS, NewAnd, "sub");
796 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
797 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
798 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
799 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
800 return Builder.CreateSub(RHS, NewOr, "sub");
801 }
802 }
803 }
804
805 // Restore LHS and RHS
806 LHS = I.getOperand(0);
807 RHS = I.getOperand(1);
808
809 // if XOR is on other side, swap
810 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
811 std::swap(LHS, RHS);
812
813 // C2 is ODD
814 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
815 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
816 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
817 if (C1->countTrailingZeros() == 0)
818 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
819 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
820 return Builder.CreateSub(RHS, NewOr, "sub");
821 }
822 return nullptr;
823}
824
825/// Wrapping flags may allow combining constants separated by an extend.
826static Instruction *foldNoWrapAdd(BinaryOperator &Add,
827 InstCombiner::BuilderTy &Builder) {
828 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
829 Type *Ty = Add.getType();
830 Constant *Op1C;
831 if (!match(Op1, m_Constant(Op1C)))
832 return nullptr;
833
834 // Try this match first because it results in an add in the narrow type.
835 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
836 Value *X;
837 const APInt *C1, *C2;
838 if (match(Op1, m_APInt(C1)) &&
839 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
840 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
841 Constant *NewC =
842 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
843 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
844 }
845
846 // More general combining of constants in the wide type.
847 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
848 Constant *NarrowC;
849 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
850 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
851 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
852 Value *WideX = Builder.CreateSExt(X, Ty);
853 return BinaryOperator::CreateAdd(WideX, NewC);
854 }
855 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
856 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
857 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
858 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
859 Value *WideX = Builder.CreateZExt(X, Ty);
860 return BinaryOperator::CreateAdd(WideX, NewC);
861 }
862
863 return nullptr;
864}
865
866Instruction *InstCombiner::foldAddWithConstant(BinaryOperator &Add) {
867 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
868 Constant *Op1C;
869 if (!match(Op1, m_Constant(Op1C)))
870 return nullptr;
871
872 if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
873 return NV;
874
875 Value *X;
876 Constant *Op00C;
877
878 // add (sub C1, X), C2 --> sub (add C1, C2), X
879 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
880 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
881
882 Value *Y;
883
884 // add (sub X, Y), -1 --> add (not Y), X
885 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
886 match(Op1, m_AllOnes()))
887 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
888
889 // zext(bool) + C -> bool ? C + 1 : C
890 if (match(Op0, m_ZExt(m_Value(X))) &&
891 X->getType()->getScalarSizeInBits() == 1)
892 return SelectInst::Create(X, AddOne(Op1C), Op1);
893 // sext(bool) + C -> bool ? C - 1 : C
894 if (match(Op0, m_SExt(m_Value(X))) &&
895 X->getType()->getScalarSizeInBits() == 1)
896 return SelectInst::Create(X, SubOne(Op1C), Op1);
897
898 // ~X + C --> (C-1) - X
899 if (match(Op0, m_Not(m_Value(X))))
900 return BinaryOperator::CreateSub(SubOne(Op1C), X);
901
902 const APInt *C;
903 if (!match(Op1, m_APInt(C)))
904 return nullptr;
905
906 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
907 const APInt *C2;
908 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
909 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
910
911 if (C->isSignMask()) {
912 // If wrapping is not allowed, then the addition must set the sign bit:
913 // X + (signmask) --> X | signmask
914 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
915 return BinaryOperator::CreateOr(Op0, Op1);
916
917 // If wrapping is allowed, then the addition flips the sign bit of LHS:
918 // X + (signmask) --> X ^ signmask
919 return BinaryOperator::CreateXor(Op0, Op1);
920 }
921
922 // Is this add the last step in a convoluted sext?
923 // add(zext(xor i16 X, -32768), -32768) --> sext X
924 Type *Ty = Add.getType();
925 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
926 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
927 return CastInst::Create(Instruction::SExt, X, Ty);
928
929 if (C->isOneValue() && Op0->hasOneUse()) {
930 // add (sext i1 X), 1 --> zext (not X)
931 // TODO: The smallest IR representation is (select X, 0, 1), and that would
932 // not require the one-use check. But we need to remove a transform in
933 // visitSelect and make sure that IR value tracking for select is equal or
934 // better than for these ops.
935 if (match(Op0, m_SExt(m_Value(X))) &&
936 X->getType()->getScalarSizeInBits() == 1)
937 return new ZExtInst(Builder.CreateNot(X), Ty);
938
939 // Shifts and add used to flip and mask off the low bit:
940 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
941 const APInt *C3;
942 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
943 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
944 Value *NotX = Builder.CreateNot(X);
945 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
946 }
947 }
948
949 return nullptr;
950}
951
952// Matches multiplication expression Op * C where C is a constant. Returns the
953// constant value in C and the other operand in Op. Returns true if such a
954// match is found.
955static bool MatchMul(Value *E, Value *&Op, APInt &C) {
956 const APInt *AI;
957 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
958 C = *AI;
959 return true;
960 }
961 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
962 C = APInt(AI->getBitWidth(), 1);
963 C <<= *AI;
964 return true;
965 }
966 return false;
967}
968
969// Matches remainder expression Op % C where C is a constant. Returns the
970// constant value in C and the other operand in Op. Returns the signedness of
971// the remainder operation in IsSigned. Returns true if such a match is
972// found.
973static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
974 const APInt *AI;
975 IsSigned = false;
976 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
977 IsSigned = true;
978 C = *AI;
979 return true;
980 }
981 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
982 C = *AI;
983 return true;
984 }
985 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
986 C = *AI + 1;
987 return true;
988 }
989 return false;
990}
991
992// Matches division expression Op / C with the given signedness as indicated
993// by IsSigned, where C is a constant. Returns the constant value in C and the
994// other operand in Op. Returns true if such a match is found.
995static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
996 const APInt *AI;
997 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
998 C = *AI;
999 return true;
1000 }
1001 if (!IsSigned) {
1002 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1003 C = *AI;
1004 return true;
1005 }
1006 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1007 C = APInt(AI->getBitWidth(), 1);
1008 C <<= *AI;
1009 return true;
1010 }
1011 }
1012 return false;
1013}
1014
1015// Returns whether C0 * C1 with the given signedness overflows.
1016static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1017 bool overflow;
1018 if (IsSigned)
1019 (void)C0.smul_ov(C1, overflow);
1020 else
1021 (void)C0.umul_ov(C1, overflow);
1022 return overflow;
1023}
1024
1025// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1026// does not overflow.
1027Value *InstCombiner::SimplifyAddWithRemainder(BinaryOperator &I) {
1028 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1029 Value *X, *MulOpV;
1030 APInt C0, MulOpC;
1031 bool IsSigned;
1032 // Match I = X % C0 + MulOpV * C0
1033 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1034 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1035 C0 == MulOpC) {
1036 Value *RemOpV;
1037 APInt C1;
1038 bool Rem2IsSigned;
1039 // Match MulOpC = RemOpV % C1
1040 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1041 IsSigned == Rem2IsSigned) {
1042 Value *DivOpV;
1043 APInt DivOpC;
1044 // Match RemOpV = X / C0
1045 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1046 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1047 Value *NewDivisor =
1048 ConstantInt::get(X->getType()->getContext(), C0 * C1);
1049 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1050 : Builder.CreateURem(X, NewDivisor, "urem");
1051 }
1052 }
1053 }
1054
1055 return nullptr;
1056}
1057
1058/// Fold
1059/// (1 << NBits) - 1
1060/// Into:
1061/// ~(-(1 << NBits))
1062/// Because a 'not' is better for bit-tracking analysis and other transforms
1063/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1064static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1065 InstCombiner::BuilderTy &Builder) {
1066 Value *NBits;
1067 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1068 return nullptr;
1069
1070 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1071 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1072 // Be wary of constant folding.
1073 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1074 // Always NSW. But NUW propagates from `add`.
1075 BOp->setHasNoSignedWrap();
1076 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1077 }
1078
1079 return BinaryOperator::CreateNot(NotMask, I.getName());
1080}
1081
1082static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1083 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction")((I.getOpcode() == Instruction::Add && "Expecting add instruction"
) ? static_cast<void> (0) : __assert_fail ("I.getOpcode() == Instruction::Add && \"Expecting add instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1083, __PRETTY_FUNCTION__))
;
1084 Type *Ty = I.getType();
1085 auto getUAddSat = [&]() {
1086 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1087 };
1088
1089 // add (umin X, ~Y), Y --> uaddsat X, Y
1090 Value *X, *Y;
1091 if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1092 m_Deferred(Y))))
1093 return CallInst::Create(getUAddSat(), { X, Y });
1094
1095 // add (umin X, ~C), C --> uaddsat X, C
1096 const APInt *C, *NotC;
1097 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1098 *C == ~*NotC)
1099 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1100
1101 return nullptr;
1102}
1103
1104Instruction *
1105InstCombiner::canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1106 BinaryOperator &I) {
1107 assert((I.getOpcode() == Instruction::Add ||(((I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction
::Or || I.getOpcode() == Instruction::Sub) && "Expecting add/or/sub instruction"
) ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1110, __PRETTY_FUNCTION__))
1
Assuming the condition is false
2
Assuming the condition is false
3
Assuming the condition is true
4
'?' condition is true
1108 I.getOpcode() == Instruction::Or ||(((I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction
::Or || I.getOpcode() == Instruction::Sub) && "Expecting add/or/sub instruction"
) ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1110, __PRETTY_FUNCTION__))
1109 I.getOpcode() == Instruction::Sub) &&(((I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction
::Or || I.getOpcode() == Instruction::Sub) && "Expecting add/or/sub instruction"
) ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1110, __PRETTY_FUNCTION__))
1110 "Expecting add/or/sub instruction")(((I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction
::Or || I.getOpcode() == Instruction::Sub) && "Expecting add/or/sub instruction"
) ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1110, __PRETTY_FUNCTION__))
;
1111
1112 // We have a subtraction/addition between a (potentially truncated) *logical*
1113 // right-shift of X and a "select".
1114 Value *X, *Select;
1115 Instruction *LowBitsToSkip, *Extract;
1116 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>>'
36
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>>'
37
Taking false branch
1117 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1118 m_Instruction(Extract))),
1119 m_Value(Select))))
5
Calling 'm_Value'
9
Returning from 'm_Value'
1120 return nullptr;
1121
1122 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1123 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
38
Assuming pointer value is null
39
Taking false branch
1124 return nullptr;
1125
1126 Type *XTy = X->getType();
1127 bool HadTrunc = I.getType() != XTy;
40
Assuming the condition is false
1128
1129 // If there was a truncation of extracted value, then we'll need to produce
1130 // one extra instruction, so we need to ensure one instruction will go away.
1131 if (HadTrunc
40.1
'HadTrunc' is false
40.1
'HadTrunc' is false
&& !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
41
Taking false branch
1132 return nullptr;
1133
1134 // Extraction should extract high NBits bits, with shift amount calculated as:
1135 // low bits to skip = shift bitwidth - high bits to extract
1136 // The shift amount itself may be extended, and we need to look past zero-ext
1137 // when matching NBits, that will matter for matching later.
1138 Constant *C;
1139 Value *NBits;
1140 if (!match(
43
Taking false branch
1141 LowBitsToSkip,
1142 m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1143 !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
42
Assuming the condition is false
1144 APInt(C->getType()->getScalarSizeInBits(),
1145 X->getType()->getScalarSizeInBits()))))
1146 return nullptr;
1147
1148 // Sign-extending value can be zero-extended if we `sub`tract it,
1149 // or sign-extended otherwise.
1150 auto SkipExtInMagic = [&I](Value *&V) {
1151 if (I.getOpcode() == Instruction::Sub)
45
Taking true branch
1152 match(V, m_ZExtOrSelf(m_Value(V)));
46
Calling 'm_Value'
48
Returning from 'm_Value'
49
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
57
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
58
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> >>'
60
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> >>'
1153 else
1154 match(V, m_SExtOrSelf(m_Value(V)));
1155 };
61
Returning without writing to 'V'
1156
1157 // Now, finally validate the sign-extending magic.
1158 // `select` itself may be appropriately extended, look past that.
1159 SkipExtInMagic(Select);
44
Calling 'operator()'
62
Returning from 'operator()'
1160
1161 ICmpInst::Predicate Pred;
1162 const APInt *Thr;
1163 Value *SignExtendingValue, *Zero;
1164 bool ShouldSignext;
1165 // It must be a select between two values we will later establish to be a
1166 // sign-extending value and a zero constant. The condition guarding the
1167 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1168 if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
63
Passing null pointer value via 1st parameter 'V'
64
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>>'
1169 m_Value(SignExtendingValue), m_Value(Zero))) ||
1170 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1171 return nullptr;
1172
1173 // icmp-select pair is commutative.
1174 if (!ShouldSignext)
1175 std::swap(SignExtendingValue, Zero);
1176
1177 // If we should not perform sign-extension then we must add/or/subtract zero.
1178 if (!match(Zero, m_Zero()))
1179 return nullptr;
1180 // Otherwise, it should be some constant, left-shifted by the same NBits we
1181 // had in `lshr`. Said left-shift can also be appropriately extended.
1182 // Again, we must look past zero-ext when looking for NBits.
1183 SkipExtInMagic(SignExtendingValue);
1184 Constant *SignExtendingValueBaseConstant;
1185 if (!match(SignExtendingValue,
1186 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1187 m_ZExtOrSelf(m_Specific(NBits)))))
1188 return nullptr;
1189 // If we `sub`, then the constant should be one, else it should be all-ones.
1190 if (I.getOpcode() == Instruction::Sub
1191 ? !match(SignExtendingValueBaseConstant, m_One())
1192 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1193 return nullptr;
1194
1195 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1196 Extract->getName() + ".sext");
1197 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1198 if (!HadTrunc)
1199 return NewAShr;
1200
1201 Builder.Insert(NewAShr);
1202 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1203}
1204
1205Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1206 if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1207 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1208 SQ.getWithInstruction(&I)))
1209 return replaceInstUsesWith(I, V);
1210
1211 if (SimplifyAssociativeOrCommutative(I))
1212 return &I;
1213
1214 if (Instruction *X = foldVectorBinop(I))
1215 return X;
1216
1217 // (A*B)+(A*C) -> A*(B+C) etc
1218 if (Value *V = SimplifyUsingDistributiveLaws(I))
1219 return replaceInstUsesWith(I, V);
1220
1221 if (Instruction *X = foldAddWithConstant(I))
1222 return X;
1223
1224 if (Instruction *X = foldNoWrapAdd(I, Builder))
1225 return X;
1226
1227 // FIXME: This should be moved into the above helper function to allow these
1228 // transforms for general constant or constant splat vectors.
1229 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1230 Type *Ty = I.getType();
1231 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1232 Value *XorLHS = nullptr; ConstantInt *XorRHS = nullptr;
1233 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1234 unsigned TySizeBits = Ty->getScalarSizeInBits();
1235 const APInt &RHSVal = CI->getValue();
1236 unsigned ExtendAmt = 0;
1237 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1238 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1239 if (XorRHS->getValue() == -RHSVal) {
1240 if (RHSVal.isPowerOf2())
1241 ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
1242 else if (XorRHS->getValue().isPowerOf2())
1243 ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
1244 }
1245
1246 if (ExtendAmt) {
1247 APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
1248 if (!MaskedValueIsZero(XorLHS, Mask, 0, &I))
1249 ExtendAmt = 0;
1250 }
1251
1252 if (ExtendAmt) {
1253 Constant *ShAmt = ConstantInt::get(Ty, ExtendAmt);
1254 Value *NewShl = Builder.CreateShl(XorLHS, ShAmt, "sext");
1255 return BinaryOperator::CreateAShr(NewShl, ShAmt);
1256 }
1257
1258 // If this is a xor that was canonicalized from a sub, turn it back into
1259 // a sub and fuse this add with it.
1260 if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
1261 KnownBits LHSKnown = computeKnownBits(XorLHS, 0, &I);
1262 if ((XorRHS->getValue() | LHSKnown.Zero).isAllOnesValue())
1263 return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
1264 XorLHS);
1265 }
1266 // (X + signmask) + C could have gotten canonicalized to (X^signmask) + C,
1267 // transform them into (X + (signmask ^ C))
1268 if (XorRHS->getValue().isSignMask())
1269 return BinaryOperator::CreateAdd(XorLHS,
1270 ConstantExpr::getXor(XorRHS, CI));
1271 }
1272 }
1273
1274 if (Ty->isIntOrIntVectorTy(1))
1275 return BinaryOperator::CreateXor(LHS, RHS);
1276
1277 // X + X --> X << 1
1278 if (LHS == RHS) {
1279 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1280 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1281 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1282 return Shl;
1283 }
1284
1285 Value *A, *B;
1286 if (match(LHS, m_Neg(m_Value(A)))) {
1287 // -A + -B --> -(A + B)
1288 if (match(RHS, m_Neg(m_Value(B))))
1289 return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1290
1291 // -A + B --> B - A
1292 return BinaryOperator::CreateSub(RHS, A);
1293 }
1294
1295 // A + -B --> A - B
1296 if (match(RHS, m_Neg(m_Value(B))))
1297 return BinaryOperator::CreateSub(LHS, B);
1298
1299 if (Value *V = checkForNegativeOperand(I, Builder))
1300 return replaceInstUsesWith(I, V);
1301
1302 // (A + 1) + ~B --> A - B
1303 // ~B + (A + 1) --> A - B
1304 // (~B + A) + 1 --> A - B
1305 // (A + ~B) + 1 --> A - B
1306 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1307 match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1308 return BinaryOperator::CreateSub(A, B);
1309
1310 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1311 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1312
1313 // A+B --> A|B iff A and B have no bits set in common.
1314 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1315 return BinaryOperator::CreateOr(LHS, RHS);
1316
1317 // FIXME: We already did a check for ConstantInt RHS above this.
1318 // FIXME: Is this pattern covered by another fold? No regression tests fail on
1319 // removal.
1320 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1321 // (X & FF00) + xx00 -> (X+xx00) & FF00
1322 Value *X;
1323 ConstantInt *C2;
1324 if (LHS->hasOneUse() &&
1325 match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
1326 CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
1327 // See if all bits from the first bit set in the Add RHS up are included
1328 // in the mask. First, get the rightmost bit.
1329 const APInt &AddRHSV = CRHS->getValue();
1330
1331 // Form a mask of all bits from the lowest bit added through the top.
1332 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
1333
1334 // See if the and mask includes all of these bits.
1335 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
1336
1337 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1338 // Okay, the xform is safe. Insert the new add pronto.
1339 Value *NewAdd = Builder.CreateAdd(X, CRHS, LHS->getName());
1340 return BinaryOperator::CreateAnd(NewAdd, C2);
1341 }
1342 }
1343 }
1344
1345 // add (select X 0 (sub n A)) A --> select X A n
1346 {
1347 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1348 Value *A = RHS;
1349 if (!SI) {
1350 SI = dyn_cast<SelectInst>(RHS);
1351 A = LHS;
1352 }
1353 if (SI && SI->hasOneUse()) {
1354 Value *TV = SI->getTrueValue();
1355 Value *FV = SI->getFalseValue();
1356 Value *N;
1357
1358 // Can we fold the add into the argument of the select?
1359 // We check both true and false select arguments for a matching subtract.
1360 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1361 // Fold the add into the true select value.
1362 return SelectInst::Create(SI->getCondition(), N, A);
1363
1364 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1365 // Fold the add into the false select value.
1366 return SelectInst::Create(SI->getCondition(), A, N);
1367 }
1368 }
1369
1370 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1371 return Ext;
1372
1373 // (add (xor A, B) (and A, B)) --> (or A, B)
1374 // (add (and A, B) (xor A, B)) --> (or A, B)
1375 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1376 m_c_And(m_Deferred(A), m_Deferred(B)))))
1377 return BinaryOperator::CreateOr(A, B);
1378
1379 // (add (or A, B) (and A, B)) --> (add A, B)
1380 // (add (and A, B) (or A, B)) --> (add A, B)
1381 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1382 m_c_And(m_Deferred(A), m_Deferred(B))))) {
1383 I.setOperand(0, A);
1384 I.setOperand(1, B);
1385 return &I;
1386 }
1387
1388 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1389 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1390 // computeKnownBits.
1391 bool Changed = false;
1392 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1393 Changed = true;
1394 I.setHasNoSignedWrap(true);
1395 }
1396 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1397 Changed = true;
1398 I.setHasNoUnsignedWrap(true);
1399 }
1400
1401 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1402 return V;
1403
1404 if (Instruction *V =
1405 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1406 return V;
1407
1408 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1409 return SatAdd;
1410
1411 return Changed ? &I : nullptr;
1412}
1413
1414/// Eliminate an op from a linear interpolation (lerp) pattern.
1415static Instruction *factorizeLerp(BinaryOperator &I,
1416 InstCombiner::BuilderTy &Builder) {
1417 Value *X, *Y, *Z;
1418 if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1419 m_OneUse(m_FSub(m_FPOne(),
1420 m_Value(Z))))),
1421 m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1422 return nullptr;
1423
1424 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1425 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1426 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1427 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1428}
1429
1430/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1431static Instruction *factorizeFAddFSub(BinaryOperator &I,
1432 InstCombiner::BuilderTy &Builder) {
1433 assert((I.getOpcode() == Instruction::FAdd ||(((I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction
::FSub) && "Expecting fadd/fsub") ? static_cast<void
> (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1434, __PRETTY_FUNCTION__))
1434 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub")(((I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction
::FSub) && "Expecting fadd/fsub") ? static_cast<void
> (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1434, __PRETTY_FUNCTION__))
;
1435 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&((I.hasAllowReassoc() && I.hasNoSignedZeros() &&
"FP factorization requires FMF") ? static_cast<void> (
0) : __assert_fail ("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1436, __PRETTY_FUNCTION__))
1436 "FP factorization requires FMF")((I.hasAllowReassoc() && I.hasNoSignedZeros() &&
"FP factorization requires FMF") ? static_cast<void> (
0) : __assert_fail ("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1436, __PRETTY_FUNCTION__))
;
1437
1438 if (Instruction *Lerp = factorizeLerp(I, Builder))
1439 return Lerp;
1440
1441 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1442 Value *X, *Y, *Z;
1443 bool IsFMul;
1444 if ((match(Op0, m_OneUse(m_FMul(m_Value(X), m_Value(Z)))) &&
1445 match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))) ||
1446 (match(Op0, m_OneUse(m_FMul(m_Value(Z), m_Value(X)))) &&
1447 match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))))
1448 IsFMul = true;
1449 else if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Z)))) &&
1450 match(Op1, m_OneUse(m_FDiv(m_Value(Y), m_Specific(Z)))))
1451 IsFMul = false;
1452 else
1453 return nullptr;
1454
1455 // (X * Z) + (Y * Z) --> (X + Y) * Z
1456 // (X * Z) - (Y * Z) --> (X - Y) * Z
1457 // (X / Z) + (Y / Z) --> (X + Y) / Z
1458 // (X / Z) - (Y / Z) --> (X - Y) / Z
1459 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1460 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1461 : Builder.CreateFSubFMF(X, Y, &I);
1462
1463 // Bail out if we just created a denormal constant.
1464 // TODO: This is copied from a previous implementation. Is it necessary?
1465 const APFloat *C;
1466 if (match(XY, m_APFloat(C)) && !C->isNormal())
1467 return nullptr;
1468
1469 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1470 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1471}
1472
1473Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
1474 if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1475 I.getFastMathFlags(),
1476 SQ.getWithInstruction(&I)))
1477 return replaceInstUsesWith(I, V);
1478
1479 if (SimplifyAssociativeOrCommutative(I))
1480 return &I;
1481
1482 if (Instruction *X = foldVectorBinop(I))
1483 return X;
1484
1485 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1486 return FoldedFAdd;
1487
1488 // (-X) + Y --> Y - X
1489 Value *X, *Y;
1490 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1491 return BinaryOperator::CreateFSubFMF(Y, X, &I);
1492
1493 // Similar to above, but look through fmul/fdiv for the negated term.
1494 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1495 Value *Z;
1496 if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1497 m_Value(Z)))) {
1498 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1499 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1500 }
1501 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1502 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1503 if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1504 m_Value(Z))) ||
1505 match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1506 m_Value(Z)))) {
1507 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1508 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1509 }
1510
1511 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1512 // integer add followed by a promotion.
1513 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1514 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1515 Value *LHSIntVal = LHSConv->getOperand(0);
1516 Type *FPType = LHSConv->getType();
1517
1518 // TODO: This check is overly conservative. In many cases known bits
1519 // analysis can tell us that the result of the addition has less significant
1520 // bits than the integer type can hold.
1521 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1522 Type *FScalarTy = FTy->getScalarType();
1523 Type *IScalarTy = ITy->getScalarType();
1524
1525 // Do we have enough bits in the significand to represent the result of
1526 // the integer addition?
1527 unsigned MaxRepresentableBits =
1528 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1529 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1530 };
1531
1532 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1533 // ... if the constant fits in the integer value. This is useful for things
1534 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1535 // requires a constant pool load, and generally allows the add to be better
1536 // instcombined.
1537 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1538 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1539 Constant *CI =
1540 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1541 if (LHSConv->hasOneUse() &&
1542 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1543 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1544 // Insert the new integer add.
1545 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1546 return new SIToFPInst(NewAdd, I.getType());
1547 }
1548 }
1549
1550 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1551 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1552 Value *RHSIntVal = RHSConv->getOperand(0);
1553 // It's enough to check LHS types only because we require int types to
1554 // be the same for this transform.
1555 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1556 // Only do this if x/y have the same type, if at least one of them has a
1557 // single use (so we don't increase the number of int->fp conversions),
1558 // and if the integer add will not overflow.
1559 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1560 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1561 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1562 // Insert the new integer add.
1563 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1564 return new SIToFPInst(NewAdd, I.getType());
1565 }
1566 }
1567 }
1568 }
1569
1570 // Handle specials cases for FAdd with selects feeding the operation
1571 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1572 return replaceInstUsesWith(I, V);
1573
1574 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1575 if (Instruction *F = factorizeFAddFSub(I, Builder))
1576 return F;
1577 if (Value *V = FAddCombine(Builder).simplify(&I))
1578 return replaceInstUsesWith(I, V);
1579 }
1580
1581 return nullptr;
1582}
1583
1584/// Optimize pointer differences into the same array into a size. Consider:
1585/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1586/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1587Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1588 Type *Ty, bool IsNUW) {
1589 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1590 // this.
1591 bool Swapped = false;
1592 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1593
1594 // For now we require one side to be the base pointer "A" or a constant
1595 // GEP derived from it.
1596 if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1597 // (gep X, ...) - X
1598 if (LHSGEP->getOperand(0) == RHS) {
1599 GEP1 = LHSGEP;
1600 Swapped = false;
1601 } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1602 // (gep X, ...) - (gep X, ...)
1603 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1604 RHSGEP->getOperand(0)->stripPointerCasts()) {
1605 GEP2 = RHSGEP;
1606 GEP1 = LHSGEP;
1607 Swapped = false;
1608 }
1609 }
1610 }
1611
1612 if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1613 // X - (gep X, ...)
1614 if (RHSGEP->getOperand(0) == LHS) {
1615 GEP1 = RHSGEP;
1616 Swapped = true;
1617 } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1618 // (gep X, ...) - (gep X, ...)
1619 if (RHSGEP->getOperand(0)->stripPointerCasts() ==
1620 LHSGEP->getOperand(0)->stripPointerCasts()) {
1621 GEP2 = LHSGEP;
1622 GEP1 = RHSGEP;
1623 Swapped = true;
1624 }
1625 }
1626 }
1627
1628 if (!GEP1)
1629 // No GEP found.
1630 return nullptr;
1631
1632 if (GEP2) {
1633 // (gep X, ...) - (gep X, ...)
1634 //
1635 // Avoid duplicating the arithmetic if there are more than one non-constant
1636 // indices between the two GEPs and either GEP has a non-constant index and
1637 // multiple users. If zero non-constant index, the result is a constant and
1638 // there is no duplication. If one non-constant index, the result is an add
1639 // or sub with a constant, which is no larger than the original code, and
1640 // there's no duplicated arithmetic, even if either GEP has multiple
1641 // users. If more than one non-constant indices combined, as long as the GEP
1642 // with at least one non-constant index doesn't have multiple users, there
1643 // is no duplication.
1644 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1645 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1646 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1647 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1648 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1649 return nullptr;
1650 }
1651 }
1652
1653 // Emit the offset of the GEP and an intptr_t.
1654 Value *Result = EmitGEPOffset(GEP1);
1655
1656 // If this is a single inbounds GEP and the original sub was nuw,
1657 // then the final multiplication is also nuw. We match an extra add zero
1658 // here, because that's what EmitGEPOffset() generates.
1659 Instruction *I;
1660 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1661 match(Result, m_Add(m_Instruction(I), m_Zero())) &&
1662 I->getOpcode() == Instruction::Mul)
1663 I->setHasNoUnsignedWrap();
1664
1665 // If we had a constant expression GEP on the other side offsetting the
1666 // pointer, subtract it from the offset we have.
1667 if (GEP2) {
1668 Value *Offset = EmitGEPOffset(GEP2);
1669 Result = Builder.CreateSub(Result, Offset);
1670 }
1671
1672 // If we have p - gep(p, ...) then we have to negate the result.
1673 if (Swapped)
1674 Result = Builder.CreateNeg(Result, "diff.neg");
1675
1676 return Builder.CreateIntCast(Result, Ty, true);
1677}
1678
1679Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1680 if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1681 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1682 SQ.getWithInstruction(&I)))
1683 return replaceInstUsesWith(I, V);
1684
1685 if (Instruction *X = foldVectorBinop(I))
1686 return X;
1687
1688 // (A*B)-(A*C) -> A*(B-C) etc
1689 if (Value *V = SimplifyUsingDistributiveLaws(I))
1690 return replaceInstUsesWith(I, V);
1691
1692 // If this is a 'B = x-(-A)', change to B = x+A.
1693 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1694 if (Value *V = dyn_castNegVal(Op1)) {
1695 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1696
1697 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1698 assert(BO->getOpcode() == Instruction::Sub &&((BO->getOpcode() == Instruction::Sub && "Expected a subtraction operator!"
) ? static_cast<void> (0) : __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1699, __PRETTY_FUNCTION__))
1699 "Expected a subtraction operator!")((BO->getOpcode() == Instruction::Sub && "Expected a subtraction operator!"
) ? static_cast<void> (0) : __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1699, __PRETTY_FUNCTION__))
;
1700 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1701 Res->setHasNoSignedWrap(true);
1702 } else {
1703 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1704 Res->setHasNoSignedWrap(true);
1705 }
1706
1707 return Res;
1708 }
1709
1710 if (I.getType()->isIntOrIntVectorTy(1))
1711 return BinaryOperator::CreateXor(Op0, Op1);
1712
1713 // Replace (-1 - A) with (~A).
1714 if (match(Op0, m_AllOnes()))
1715 return BinaryOperator::CreateNot(Op1);
1716
1717 // (~X) - (~Y) --> Y - X
1718 Value *X, *Y;
1719 if (match(Op0, m_Not(m_Value(X))) && match(Op1, m_Not(m_Value(Y))))
1720 return BinaryOperator::CreateSub(Y, X);
1721
1722 // (X + -1) - Y --> ~Y + X
1723 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1724 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1725
1726 // Y - (X + 1) --> ~X + Y
1727 if (match(Op1, m_OneUse(m_Add(m_Value(X), m_One()))))
1728 return BinaryOperator::CreateAdd(Builder.CreateNot(X), Op0);
1729
1730 // Y - ~X --> (X + 1) + Y
1731 if (match(Op1, m_OneUse(m_Not(m_Value(X))))) {
1732 return BinaryOperator::CreateAdd(
1733 Builder.CreateAdd(Op0, ConstantInt::get(I.getType(), 1)), X);
1734 }
1735
1736 if (Constant *C = dyn_cast<Constant>(Op0)) {
1737 bool IsNegate = match(C, m_ZeroInt());
1738 Value *X;
1739 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1740 // 0 - (zext bool) --> sext bool
1741 // C - (zext bool) --> bool ? C - 1 : C
1742 if (IsNegate)
1743 return CastInst::CreateSExtOrBitCast(X, I.getType());
1744 return SelectInst::Create(X, SubOne(C), C);
1745 }
1746 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1747 // 0 - (sext bool) --> zext bool
1748 // C - (sext bool) --> bool ? C + 1 : C
1749 if (IsNegate)
1750 return CastInst::CreateZExtOrBitCast(X, I.getType());
1751 return SelectInst::Create(X, AddOne(C), C);
1752 }
1753
1754 // C - ~X == X + (1+C)
1755 if (match(Op1, m_Not(m_Value(X))))
1756 return BinaryOperator::CreateAdd(X, AddOne(C));
1757
1758 // Try to fold constant sub into select arguments.
1759 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1760 if (Instruction *R = FoldOpIntoSelect(I, SI))
1761 return R;
1762
1763 // Try to fold constant sub into PHI values.
1764 if (PHINode *PN = dyn_cast<PHINode>(Op1))
1765 if (Instruction *R = foldOpIntoPhi(I, PN))
1766 return R;
1767
1768 Constant *C2;
1769
1770 // C-(C2-X) --> X+(C-C2)
1771 if (match(Op1, m_Sub(m_Constant(C2), m_Value(X))))
1772 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
1773
1774 // C-(X+C2) --> (C-C2)-X
1775 if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
1776 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1777 }
1778
1779 const APInt *Op0C;
1780 if (match(Op0, m_APInt(Op0C))) {
1781
1782 if (Op0C->isNullValue()) {
1783 Value *Op1Wide;
1784 match(Op1, m_TruncOrSelf(m_Value(Op1Wide)));
1785 bool HadTrunc = Op1Wide != Op1;
1786 bool NoTruncOrTruncIsOneUse = !HadTrunc || Op1->hasOneUse();
1787 unsigned BitWidth = Op1Wide->getType()->getScalarSizeInBits();
1788
1789 Value *X;
1790 const APInt *ShAmt;
1791 // -(X >>u 31) -> (X >>s 31)
1792 if (NoTruncOrTruncIsOneUse &&
1793 match(Op1Wide, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1794 *ShAmt == BitWidth - 1) {
1795 Value *ShAmtOp = cast<Instruction>(Op1Wide)->getOperand(1);
1796 Instruction *NewShift = BinaryOperator::CreateAShr(X, ShAmtOp);
1797 NewShift->copyIRFlags(Op1Wide);
1798 if (!HadTrunc)
1799 return NewShift;
1800 Builder.Insert(NewShift);
1801 return TruncInst::CreateTruncOrBitCast(NewShift, Op1->getType());
1802 }
1803 // -(X >>s 31) -> (X >>u 31)
1804 if (NoTruncOrTruncIsOneUse &&
1805 match(Op1Wide, m_AShr(m_Value(X), m_APInt(ShAmt))) &&
1806 *ShAmt == BitWidth - 1) {
1807 Value *ShAmtOp = cast<Instruction>(Op1Wide)->getOperand(1);
1808 Instruction *NewShift = BinaryOperator::CreateLShr(X, ShAmtOp);
1809 NewShift->copyIRFlags(Op1Wide);
1810 if (!HadTrunc)
1811 return NewShift;
1812 Builder.Insert(NewShift);
1813 return TruncInst::CreateTruncOrBitCast(NewShift, Op1->getType());
1814 }
1815
1816 if (!HadTrunc && Op1->hasOneUse()) {
1817 Value *LHS, *RHS;
1818 SelectPatternFlavor SPF = matchSelectPattern(Op1, LHS, RHS).Flavor;
1819 if (SPF == SPF_ABS || SPF == SPF_NABS) {
1820 // This is a negate of an ABS/NABS pattern. Just swap the operands
1821 // of the select.
1822 cast<SelectInst>(Op1)->swapValues();
1823 // Don't swap prof metadata, we didn't change the branch behavior.
1824 return replaceInstUsesWith(I, Op1);
1825 }
1826 }
1827 }
1828
1829 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1830 // zero.
1831 if (Op0C->isMask()) {
1832 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1833 if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
1834 return BinaryOperator::CreateXor(Op1, Op0);
1835 }
1836 }
1837
1838 {
1839 Value *Y;
1840 // X-(X+Y) == -Y X-(Y+X) == -Y
1841 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1842 return BinaryOperator::CreateNeg(Y);
1843
1844 // (X-Y)-X == -Y
1845 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1846 return BinaryOperator::CreateNeg(Y);
1847 }
1848
1849 // (sub (or A, B) (and A, B)) --> (xor A, B)
1850 {
1851 Value *A, *B;
1852 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1853 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1854 return BinaryOperator::CreateXor(A, B);
1855 }
1856
1857 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
1858 {
1859 Value *A, *B;
1860 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1861 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1862 (Op0->hasOneUse() || Op1->hasOneUse()))
1863 return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
1864 }
1865
1866 // (sub (or A, B), (xor A, B)) --> (and A, B)
1867 {
1868 Value *A, *B;
1869 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1870 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1871 return BinaryOperator::CreateAnd(A, B);
1872 }
1873
1874 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
1875 {
1876 Value *A, *B;
1877 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1878 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1879 (Op0->hasOneUse() || Op1->hasOneUse()))
1880 return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
1881 }
1882
1883 {
1884 Value *Y;
1885 // ((X | Y) - X) --> (~X & Y)
1886 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1887 return BinaryOperator::CreateAnd(
1888 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1889 }
1890
1891 {
1892 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
1893 Value *X;
1894 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
1895 m_OneUse(m_Neg(m_Value(X))))))) {
1896 return BinaryOperator::CreateNeg(Builder.CreateAnd(
1897 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
1898 }
1899 }
1900
1901 {
1902 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
1903 Constant *C;
1904 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
1905 return BinaryOperator::CreateNeg(
1906 Builder.CreateAnd(Op1, Builder.CreateNot(C)));
1907 }
1908 }
1909
1910 {
1911 // If we have a subtraction between some value and a select between
1912 // said value and something else, sink subtraction into select hands, i.e.:
1913 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
1914 // ->
1915 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
1916 // or
1917 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
1918 // ->
1919 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
1920 // This will result in select between new subtraction and 0.
1921 auto SinkSubIntoSelect =
1922 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
1923 auto SubBuilder) -> Instruction * {
1924 Value *Cond, *TrueVal, *FalseVal;
1925 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
1926 m_Value(FalseVal)))))
1927 return nullptr;
1928 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
1929 return nullptr;
1930 // While it is really tempting to just create two subtractions and let
1931 // InstCombine fold one of those to 0, it isn't possible to do so
1932 // because of worklist visitation order. So ugly it is.
1933 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
1934 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
1935 Constant *Zero = Constant::getNullValue(Ty);
1936 SelectInst *NewSel =
1937 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
1938 OtherHandOfSubIsTrueVal ? NewSub : Zero);
1939 // Preserve prof metadata if any.
1940 NewSel->copyMetadata(cast<Instruction>(*Select));
1941 return NewSel;
1942 };
1943 if (Instruction *NewSel = SinkSubIntoSelect(
1944 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
1945 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
1946 return Builder->CreateSub(OtherHandOfSelect,
1947 /*OtherHandOfSub=*/Op1);
1948 }))
1949 return NewSel;
1950 if (Instruction *NewSel = SinkSubIntoSelect(
1951 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
1952 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
1953 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
1954 OtherHandOfSelect);
1955 }))
1956 return NewSel;
1957 }
1958
1959 if (Op1->hasOneUse()) {
1960 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
1961 Constant *C = nullptr;
1962
1963 // (X - (Y - Z)) --> (X + (Z - Y)).
1964 if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
1965 return BinaryOperator::CreateAdd(Op0,
1966 Builder.CreateSub(Z, Y, Op1->getName()));
1967
1968 // (X - (X & Y)) --> (X & ~Y)
1969 if (match(Op1, m_c_And(m_Value(Y), m_Specific(Op0))))
1970 return BinaryOperator::CreateAnd(Op0,
1971 Builder.CreateNot(Y, Y->getName() + ".not"));
1972
1973 // 0 - (X sdiv C) -> (X sdiv -C) provided the negation doesn't overflow.
1974 if (match(Op0, m_Zero())) {
1975 Constant *Op11C;
1976 if (match(Op1, m_SDiv(m_Value(X), m_Constant(Op11C))) &&
1977 !Op11C->containsUndefElement() && Op11C->isNotMinSignedValue() &&
1978 Op11C->isNotOneValue()) {
1979 Instruction *BO =
1980 BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(Op11C));
1981 BO->setIsExact(cast<BinaryOperator>(Op1)->isExact());
1982 return BO;
1983 }
1984 }
1985
1986 // 0 - (X << Y) -> (-X << Y) when X is freely negatable.
1987 if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
1988 if (Value *XNeg = dyn_castNegVal(X))
1989 return BinaryOperator::CreateShl(XNeg, Y);
1990
1991 // Subtracting -1/0 is the same as adding 1/0:
1992 // sub [nsw] Op0, sext(bool Y) -> add [nsw] Op0, zext(bool Y)
1993 // 'nuw' is dropped in favor of the canonical form.
1994 if (match(Op1, m_SExt(m_Value(Y))) &&
1995 Y->getType()->getScalarSizeInBits() == 1) {
1996 Value *Zext = Builder.CreateZExt(Y, I.getType());
1997 BinaryOperator *Add = BinaryOperator::CreateAdd(Op0, Zext);
1998 Add->setHasNoSignedWrap(I.hasNoSignedWrap());
1999 return Add;
2000 }
2001 // sub [nsw] X, zext(bool Y) -> add [nsw] X, sext(bool Y)
2002 // 'nuw' is dropped in favor of the canonical form.
2003 if (match(Op1, m_ZExt(m_Value(Y))) && Y->getType()->isIntOrIntVectorTy(1)) {
2004 Value *Sext = Builder.CreateSExt(Y, I.getType());
2005 BinaryOperator *Add = BinaryOperator::CreateAdd(Op0, Sext);
2006 Add->setHasNoSignedWrap(I.hasNoSignedWrap());
2007 return Add;
2008 }
2009
2010 // X - A*-B -> X + A*B
2011 // X - -A*B -> X + A*B
2012 Value *A, *B;
2013 if (match(Op1, m_c_Mul(m_Value(A), m_Neg(m_Value(B)))))
2014 return BinaryOperator::CreateAdd(Op0, Builder.CreateMul(A, B));
2015
2016 // X - A*C -> X + A*-C
2017 // No need to handle commuted multiply because multiply handling will
2018 // ensure constant will be move to the right hand side.
2019 if (match(Op1, m_Mul(m_Value(A), m_Constant(C))) && !isa<ConstantExpr>(C)) {
2020 Value *NewMul = Builder.CreateMul(A, ConstantExpr::getNeg(C));
2021 return BinaryOperator::CreateAdd(Op0, NewMul);
2022 }
2023 }
2024
2025 {
2026 // ~A - Min/Max(~A, O) -> Max/Min(A, ~O) - A
2027 // ~A - Min/Max(O, ~A) -> Max/Min(A, ~O) - A
2028 // Min/Max(~A, O) - ~A -> A - Max/Min(A, ~O)
2029 // Min/Max(O, ~A) - ~A -> A - Max/Min(A, ~O)
2030 // So long as O here is freely invertible, this will be neutral or a win.
2031 Value *LHS, *RHS, *A;
2032 Value *NotA = Op0, *MinMax = Op1;
2033 SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2034 if (!SelectPatternResult::isMinOrMax(SPF)) {
2035 NotA = Op1;
2036 MinMax = Op0;
2037 SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2038 }
2039 if (SelectPatternResult::isMinOrMax(SPF) &&
2040 match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
2041 if (NotA == LHS)
2042 std::swap(LHS, RHS);
2043 // LHS is now O above and expected to have at least 2 uses (the min/max)
2044 // NotA is epected to have 2 uses from the min/max and 1 from the sub.
2045 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
2046 !NotA->hasNUsesOrMore(4)) {
2047 // Note: We don't generate the inverse max/min, just create the not of
2048 // it and let other folds do the rest.
2049 Value *Not = Builder.CreateNot(MinMax);
2050 if (NotA == Op0)
2051 return BinaryOperator::CreateSub(Not, A);
2052 else
2053 return BinaryOperator::CreateSub(A, Not);
2054 }
2055 }
2056 }
2057
2058 // Optimize pointer differences into the same array into a size. Consider:
2059 // &A[10] - &A[0]: we should compile this to "10".
2060 Value *LHSOp, *RHSOp;
2061 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2062 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2063 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2064 I.hasNoUnsignedWrap()))
2065 return replaceInstUsesWith(I, Res);
2066
2067 // trunc(p)-trunc(q) -> trunc(p-q)
2068 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2069 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2070 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2071 /* IsNUW */ false))
2072 return replaceInstUsesWith(I, Res);
2073
2074 // Canonicalize a shifty way to code absolute value to the common pattern.
2075 // There are 2 potential commuted variants.
2076 // We're relying on the fact that we only do this transform when the shift has
2077 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2078 // instructions).
2079 Value *A;
2080 const APInt *ShAmt;
2081 Type *Ty = I.getType();
2082 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2083 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2084 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2085 // B = ashr i32 A, 31 ; smear the sign bit
2086 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2087 // --> (A < 0) ? -A : A
2088 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2089 // Copy the nuw/nsw flags from the sub to the negate.
2090 Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2091 I.hasNoSignedWrap());
2092 return SelectInst::Create(Cmp, Neg, A);
2093 }
2094
2095 if (Instruction *V =
2096 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2097 return V;
2098
2099 if (Instruction *Ext = narrowMathIfNoOverflow(I))
2100 return Ext;
2101
2102 bool Changed = false;
2103 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
2104 Changed = true;
2105 I.setHasNoSignedWrap(true);
2106 }
2107 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
2108 Changed = true;
2109 I.setHasNoUnsignedWrap(true);
2110 }
2111
2112 return Changed ? &I : nullptr;
2113}
2114
2115/// This eliminates floating-point negation in either 'fneg(X)' or
2116/// 'fsub(-0.0, X)' form by combining into a constant operand.
2117static Instruction *foldFNegIntoConstant(Instruction &I) {
2118 Value *X;
2119 Constant *C;
2120
2121 // Fold negation into constant operand. This is limited with one-use because
2122 // fneg is assumed better for analysis and cheaper in codegen than fmul/fdiv.
2123 // -(X * C) --> X * (-C)
2124 // FIXME: It's arguable whether these should be m_OneUse or not. The current
2125 // belief is that the FNeg allows for better reassociation opportunities.
2126 if (match(&I, m_FNeg(m_OneUse(m_FMul(m_Value(X), m_Constant(C))))))
2127 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
2128 // -(X / C) --> X / (-C)
2129 if (match(&I, m_FNeg(m_OneUse(m_FDiv(m_Value(X), m_Constant(C))))))
2130 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
2131 // -(C / X) --> (-C) / X
2132 if (match(&I, m_FNeg(m_OneUse(m_FDiv(m_Constant(C), m_Value(X))))))
2133 return BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
2134
2135 return nullptr;
2136}
2137
2138static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2139 InstCombiner::BuilderTy &Builder) {
2140 Value *FNeg;
2141 if (!match(&I, m_FNeg(m_Value(FNeg))))
2142 return nullptr;
2143
2144 Value *X, *Y;
2145 if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2146 return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2147
2148 if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2149 return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2150
2151 return nullptr;
2152}
2153
2154Instruction *InstCombiner::visitFNeg(UnaryOperator &I) {
2155 Value *Op = I.getOperand(0);
2156
2157 if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
2158 SQ.getWithInstruction(&I)))
2159 return replaceInstUsesWith(I, V);
2160
2161 if (Instruction *X = foldFNegIntoConstant(I))
2162 return X;
2163
2164 Value *X, *Y;
2165
2166 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2167 if (I.hasNoSignedZeros() &&
2168 match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2169 return BinaryOperator::CreateFSubFMF(Y, X, &I);
2170
2171 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2172 return R;
2173
2174 return nullptr;
2175}
2176
2177Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2178 if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
2179 I.getFastMathFlags(),
2180 SQ.getWithInstruction(&I)))
2181 return replaceInstUsesWith(I, V);
2182
2183 if (Instruction *X = foldVectorBinop(I))
2184 return X;
2185
2186 // Subtraction from -0.0 is the canonical form of fneg.
2187 // fsub nsz 0, X ==> fsub nsz -0.0, X
2188 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2189 if (I.hasNoSignedZeros() && match(Op0, m_PosZeroFP()))
2190 return BinaryOperator::CreateFNegFMF(Op1, &I);
2191
2192 if (Instruction *X = foldFNegIntoConstant(I))
2193 return X;
2194
2195 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2196 return R;
2197
2198 Value *X, *Y;
2199 Constant *C;
2200
2201 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2202 // Canonicalize to fadd to make analysis easier.
2203 // This can also help codegen because fadd is commutative.
2204 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2205 // killed later. We still limit that particular transform with 'hasOneUse'
2206 // because an fneg is assumed better/cheaper than a generic fsub.
2207 if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2208 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2209 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2210 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2211 }
2212 }
2213
2214 if (isa<Constant>(Op0))
2215 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2216 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2217 return NV;
2218
2219 // X - C --> X + (-C)
2220 // But don't transform constant expressions because there's an inverse fold
2221 // for X + (-Y) --> X - Y.
2222 if (match(Op1, m_Constant(C)) && !isa<ConstantExpr>(Op1))
2223 return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
2224
2225 // X - (-Y) --> X + Y
2226 if (match(Op1, m_FNeg(m_Value(Y))))
2227 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2228
2229 // Similar to above, but look through a cast of the negated value:
2230 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2231 Type *Ty = I.getType();
2232 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2233 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2234
2235 // X - (fpext(-Y)) --> X + fpext(Y)
2236 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2237 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2238
2239 // Similar to above, but look through fmul/fdiv of the negated value:
2240 // Op0 - (-X * Y) --> Op0 + (X * Y)
2241 // Op0 - (Y * -X) --> Op0 + (X * Y)
2242 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2243 Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2244 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2245 }
2246 // Op0 - (-X / Y) --> Op0 + (X / Y)
2247 // Op0 - (X / -Y) --> Op0 + (X / Y)
2248 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2249 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2250 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2251 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2252 }
2253
2254 // Handle special cases for FSub with selects feeding the operation
2255 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2256 return replaceInstUsesWith(I, V);
2257
2258 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2259 // (Y - X) - Y --> -X
2260 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2261 return BinaryOperator::CreateFNegFMF(X, &I);
2262
2263 // Y - (X + Y) --> -X
2264 // Y - (Y + X) --> -X
2265 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2266 return BinaryOperator::CreateFNegFMF(X, &I);
2267
2268 // (X * C) - X --> X * (C - 1.0)
2269 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2270 Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
2271 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2272 }
2273 // X - (X * C) --> X * (1.0 - C)
2274 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2275 Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
2276 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2277 }
2278
2279 if (Instruction *F = factorizeFAddFSub(I, Builder))
2280 return F;
2281
2282 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2283 // functionality has been subsumed by simple pattern matching here and in
2284 // InstSimplify. We should let a dedicated reassociation pass handle more
2285 // complex pattern matching and remove this from InstCombine.
2286 if (Value *V = FAddCombine(Builder).simplify(&I))
2287 return replaceInstUsesWith(I, V);
2288 }
2289
2290 return nullptr;
2291}

/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/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/InstrTypes.h"
36#include "llvm/IR/Instruction.h"
37#include "llvm/IR/Instructions.h"
38#include "llvm/IR/IntrinsicInst.h"
39#include "llvm/IR/Intrinsics.h"
40#include "llvm/IR/Operator.h"
41#include "llvm/IR/Value.h"
42#include "llvm/Support/Casting.h"
43#include <cstdint>
44
45namespace llvm {
46namespace PatternMatch {
47
48template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) {
49 return const_cast<Pattern &>(P).match(V);
14
Calling 'AnyBinaryOp_match::match'
34
Returning from 'AnyBinaryOp_match::match'
35
Returning the value 1, which participates in a condition later
59
Returning without writing to 'P.R.VR'
65
Passing null pointer value via 1st parameter 'V'
66
Calling 'ThreeOps_match::match'
50}
51
52template <typename SubPattern_t> struct OneUse_match {
53 SubPattern_t SubPattern;
54
55 OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {}
56
57 template <typename OpTy> bool match(OpTy *V) {
58 return V->hasOneUse() && SubPattern.match(V);
59 }
60};
61
62template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) {
63 return SubPattern;
64}
65
66template <typename Class> struct class_match {
67 template <typename ITy> bool match(ITy *V) { return isa<Class>(V); }
68};
69
70/// Match an arbitrary value and ignore it.
71inline class_match<Value> m_Value() { return class_match<Value>(); }
72
73/// Match an arbitrary binary operation and ignore it.
74inline class_match<BinaryOperator> m_BinOp() {
75 return class_match<BinaryOperator>();
76}
77
78/// Matches any compare instruction and ignore it.
79inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); }
80
81/// Match an arbitrary ConstantInt and ignore it.
82inline class_match<ConstantInt> m_ConstantInt() {
83 return class_match<ConstantInt>();
84}
85
86/// Match an arbitrary undef constant.
87inline class_match<UndefValue> m_Undef() { return class_match<UndefValue>(); }
88
89/// Match an arbitrary Constant and ignore it.
90inline class_match<Constant> m_Constant() { return class_match<Constant>(); }
91
92/// Match an arbitrary basic block value and ignore it.
93inline class_match<BasicBlock> m_BasicBlock() {
94 return class_match<BasicBlock>();
95}
96
97/// Inverting matcher
98template <typename Ty> struct match_unless {
99 Ty M;
100
101 match_unless(const Ty &Matcher) : M(Matcher) {}
102
103 template <typename ITy> bool match(ITy *V) { return !M.match(V); }
104};
105
106/// Match if the inner matcher does *NOT* match.
107template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) {
108 return match_unless<Ty>(M);
109}
110
111/// Matching combinators
112template <typename LTy, typename RTy> struct match_combine_or {
113 LTy L;
114 RTy R;
115
116 match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
117
118 template <typename ITy> bool match(ITy *V) {
119 if (L.match(V))
18
Taking false branch
120 return true;
121 if (R.match(V))
19
Calling 'match_combine_and::match'
30
Returning from 'match_combine_and::match'
31
Taking true branch
122 return true;
123 return false;
124 }
125};
126
127template <typename LTy, typename RTy> struct match_combine_and {
128 LTy L;
129 RTy R;
130
131 match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
132
133 template <typename ITy> bool match(ITy *V) {
134 if (L.match(V))
20
Calling 'BinaryOp_match::match'
27
Returning from 'BinaryOp_match::match'
28
Taking true branch
135 if (R.match(V))
29
Taking true branch
136 return true;
137 return false;
138 }
139};
140
141/// Combine two pattern matchers matching L || R
142template <typename LTy, typename RTy>
143inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) {
144 return match_combine_or<LTy, RTy>(L, R);
54
Returning without writing to 'L.Op.VR'
145}
146
147/// Combine two pattern matchers matching L && R
148template <typename LTy, typename RTy>
149inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) {
150 return match_combine_and<LTy, RTy>(L, R);
151}
152
153struct apint_match {
154 const APInt *&Res;
155
156 apint_match(const APInt *&R) : Res(R) {}
157
158 template <typename ITy> bool match(ITy *V) {
159 if (auto *CI = dyn_cast<ConstantInt>(V)) {
160 Res = &CI->getValue();
161 return true;
162 }
163 if (V->getType()->isVectorTy())
164 if (const auto *C = dyn_cast<Constant>(V))
165 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
166 Res = &CI->getValue();
167 return true;
168 }
169 return false;
170 }
171};
172// Either constexpr if or renaming ConstantFP::getValueAPF to
173// ConstantFP::getValue is needed to do it via single template
174// function for both apint/apfloat.
175struct apfloat_match {
176 const APFloat *&Res;
177 apfloat_match(const APFloat *&R) : Res(R) {}
178 template <typename ITy> bool match(ITy *V) {
179 if (auto *CI = dyn_cast<ConstantFP>(V)) {
180 Res = &CI->getValueAPF();
181 return true;
182 }
183 if (V->getType()->isVectorTy())
184 if (const auto *C = dyn_cast<Constant>(V))
185 if (auto *CI = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) {
186 Res = &CI->getValueAPF();
187 return true;
188 }
189 return false;
190 }
191};
192
193/// Match a ConstantInt or splatted ConstantVector, binding the
194/// specified pointer to the contained APInt.
195inline apint_match m_APInt(const APInt *&Res) { return Res; }
196
197/// Match a ConstantFP or splatted ConstantVector, binding the
198/// specified pointer to the contained APFloat.
199inline apfloat_match m_APFloat(const APFloat *&Res) { return Res; }
200
201template <int64_t Val> struct constantint_match {
202 template <typename ITy> bool match(ITy *V) {
203 if (const auto *CI = dyn_cast<ConstantInt>(V)) {
204 const APInt &CIV = CI->getValue();
205 if (Val >= 0)
206 return CIV == static_cast<uint64_t>(Val);
207 // If Val is negative, and CI is shorter than it, truncate to the right
208 // number of bits. If it is larger, then we have to sign extend. Just
209 // compare their negated values.
210 return -CIV == -Val;
211 }
212 return false;
213 }
214};
215
216/// Match a ConstantInt with a specific value.
217template <int64_t Val> inline constantint_match<Val> m_ConstantInt() {
218 return constantint_match<Val>();
219}
220
221/// This helper class is used to match scalar and vector integer constants that
222/// satisfy a specified predicate.
223/// For vector constants, undefined elements are ignored.
224template <typename Predicate> struct cst_pred_ty : public Predicate {
225 template <typename ITy> bool match(ITy *V) {
226 if (const auto *CI = dyn_cast<ConstantInt>(V))
227 return this->isValue(CI->getValue());
228 if (V->getType()->isVectorTy()) {
229 if (const auto *C = dyn_cast<Constant>(V)) {
230 if (const auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
231 return this->isValue(CI->getValue());
232
233 // Non-splat vector constant: check each element for a match.
234 unsigned NumElts = V->getType()->getVectorNumElements();
235 assert(NumElts != 0 && "Constant vector with no elements?")((NumElts != 0 && "Constant vector with no elements?"
) ? static_cast<void> (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/include/llvm/IR/PatternMatch.h"
, 235, __PRETTY_FUNCTION__))
;
236 bool HasNonUndefElements = false;
237 for (unsigned i = 0; i != NumElts; ++i) {
238 Constant *Elt = C->getAggregateElement(i);
239 if (!Elt)
240 return false;
241 if (isa<UndefValue>(Elt))
242 continue;
243 auto *CI = dyn_cast<ConstantInt>(Elt);
244 if (!CI || !this->isValue(CI->getValue()))
245 return false;
246 HasNonUndefElements = true;
247 }
248 return HasNonUndefElements;
249 }
250 }
251 return false;
252 }
253};
254
255/// This helper class is used to match scalar and vector constants that
256/// satisfy a specified predicate, and bind them to an APInt.
257template <typename Predicate> struct api_pred_ty : public Predicate {
258 const APInt *&Res;
259
260 api_pred_ty(const APInt *&R) : Res(R) {}
261
262 template <typename ITy> bool match(ITy *V) {
263 if (const auto *CI = dyn_cast<ConstantInt>(V))
264 if (this->isValue(CI->getValue())) {
265 Res = &CI->getValue();
266 return true;
267 }
268 if (V->getType()->isVectorTy())
269 if (const auto *C = dyn_cast<Constant>(V))
270 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
271 if (this->isValue(CI->getValue())) {
272 Res = &CI->getValue();
273 return true;
274 }
275
276 return false;
277 }
278};
279
280/// This helper class is used to match scalar and vector floating-point
281/// constants that satisfy a specified predicate.
282/// For vector constants, undefined elements are ignored.
283template <typename Predicate> struct cstfp_pred_ty : public Predicate {
284 template <typename ITy> bool match(ITy *V) {
285 if (const auto *CF = dyn_cast<ConstantFP>(V))
286 return this->isValue(CF->getValueAPF());
287 if (V->getType()->isVectorTy()) {
288 if (const auto *C = dyn_cast<Constant>(V)) {
289 if (const auto *CF = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
290 return this->isValue(CF->getValueAPF());
291
292 // Non-splat vector constant: check each element for a match.
293 unsigned NumElts = V->getType()->getVectorNumElements();
294 assert(NumElts != 0 && "Constant vector with no elements?")((NumElts != 0 && "Constant vector with no elements?"
) ? static_cast<void> (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/include/llvm/IR/PatternMatch.h"
, 294, __PRETTY_FUNCTION__))
;
295 bool HasNonUndefElements = false;
296 for (unsigned i = 0; i != NumElts; ++i) {
297 Constant *Elt = C->getAggregateElement(i);
298 if (!Elt)
299 return false;
300 if (isa<UndefValue>(Elt))
301 continue;
302 auto *CF = dyn_cast<ConstantFP>(Elt);
303 if (!CF || !this->isValue(CF->getValueAPF()))
304 return false;
305 HasNonUndefElements = true;
306 }
307 return HasNonUndefElements;
308 }
309 }
310 return false;
311 }
312};
313
314///////////////////////////////////////////////////////////////////////////////
315//
316// Encapsulate constant value queries for use in templated predicate matchers.
317// This allows checking if constants match using compound predicates and works
318// with vector constants, possibly with relaxed constraints. For example, ignore
319// undef values.
320//
321///////////////////////////////////////////////////////////////////////////////
322
323struct is_any_apint {
324 bool isValue(const APInt &C) { return true; }
325};
326/// Match an integer or vector with any integral constant.
327/// For vectors, this includes constants with undefined elements.
328inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() {
329 return cst_pred_ty<is_any_apint>();
330}
331
332struct is_all_ones {
333 bool isValue(const APInt &C) { return C.isAllOnesValue(); }
334};
335/// Match an integer or vector with all bits set.
336/// For vectors, this includes constants with undefined elements.
337inline cst_pred_ty<is_all_ones> m_AllOnes() {
338 return cst_pred_ty<is_all_ones>();
339}
340
341struct is_maxsignedvalue {
342 bool isValue(const APInt &C) { return C.isMaxSignedValue(); }
343};
344/// Match an integer or vector with values having all bits except for the high
345/// bit set (0x7f...).
346/// For vectors, this includes constants with undefined elements.
347inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() {
348 return cst_pred_ty<is_maxsignedvalue>();
349}
350inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) {
351 return V;
352}
353
354struct is_negative {
355 bool isValue(const APInt &C) { return C.isNegative(); }
356};
357/// Match an integer or vector of negative values.
358/// For vectors, this includes constants with undefined elements.
359inline cst_pred_ty<is_negative> m_Negative() {
360 return cst_pred_ty<is_negative>();
361}
362inline api_pred_ty<is_negative> m_Negative(const APInt *&V) {
363 return V;
364}
365
366struct is_nonnegative {
367 bool isValue(const APInt &C) { return C.isNonNegative(); }
368};
369/// Match an integer or vector of non-negative values.
370/// For vectors, this includes constants with undefined elements.
371inline cst_pred_ty<is_nonnegative> m_NonNegative() {
372 return cst_pred_ty<is_nonnegative>();
373}
374inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) {
375 return V;
376}
377
378struct is_strictlypositive {
379 bool isValue(const APInt &C) { return C.isStrictlyPositive(); }
380};
381/// Match an integer or vector of strictly positive values.
382/// For vectors, this includes constants with undefined elements.
383inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() {
384 return cst_pred_ty<is_strictlypositive>();
385}
386inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) {
387 return V;
388}
389
390struct is_nonpositive {
391 bool isValue(const APInt &C) { return C.isNonPositive(); }
392};
393/// Match an integer or vector of non-positive values.
394/// For vectors, this includes constants with undefined elements.
395inline cst_pred_ty<is_nonpositive> m_NonPositive() {
396 return cst_pred_ty<is_nonpositive>();
397}
398inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; }
399
400struct is_one {
401 bool isValue(const APInt &C) { return C.isOneValue(); }
402};
403/// Match an integer 1 or a vector with all elements equal to 1.
404/// For vectors, this includes constants with undefined elements.
405inline cst_pred_ty<is_one> m_One() {
406 return cst_pred_ty<is_one>();
407}
408
409struct is_zero_int {
410 bool isValue(const APInt &C) { return C.isNullValue(); }
411};
412/// Match an integer 0 or a vector with all elements equal to 0.
413/// For vectors, this includes constants with undefined elements.
414inline cst_pred_ty<is_zero_int> m_ZeroInt() {
415 return cst_pred_ty<is_zero_int>();
416}
417
418struct is_zero {
419 template <typename ITy> bool match(ITy *V) {
420 auto *C = dyn_cast<Constant>(V);
421 return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C));
422 }
423};
424/// Match any null constant or a vector with all elements equal to 0.
425/// For vectors, this includes constants with undefined elements.
426inline is_zero m_Zero() {
427 return is_zero();
428}
429
430struct is_power2 {
431 bool isValue(const APInt &C) { return C.isPowerOf2(); }
432};
433/// Match an integer or vector power-of-2.
434/// For vectors, this includes constants with undefined elements.
435inline cst_pred_ty<is_power2> m_Power2() {
436 return cst_pred_ty<is_power2>();
437}
438inline api_pred_ty<is_power2> m_Power2(const APInt *&V) {
439 return V;
440}
441
442struct is_negated_power2 {
443 bool isValue(const APInt &C) { return (-C).isPowerOf2(); }
444};
445/// Match a integer or vector negated power-of-2.
446/// For vectors, this includes constants with undefined elements.
447inline cst_pred_ty<is_negated_power2> m_NegatedPower2() {
448 return cst_pred_ty<is_negated_power2>();
449}
450inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) {
451 return V;
452}
453
454struct is_power2_or_zero {
455 bool isValue(const APInt &C) { return !C || C.isPowerOf2(); }
456};
457/// Match an integer or vector of 0 or power-of-2 values.
458/// For vectors, this includes constants with undefined elements.
459inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() {
460 return cst_pred_ty<is_power2_or_zero>();
461}
462inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) {
463 return V;
464}
465
466struct is_sign_mask {
467 bool isValue(const APInt &C) { return C.isSignMask(); }
468};
469/// Match an integer or vector with only the sign bit(s) set.
470/// For vectors, this includes constants with undefined elements.
471inline cst_pred_ty<is_sign_mask> m_SignMask() {
472 return cst_pred_ty<is_sign_mask>();
473}
474
475struct is_lowbit_mask {
476 bool isValue(const APInt &C) { return C.isMask(); }
477};
478/// Match an integer or vector with only the low bit(s) set.
479/// For vectors, this includes constants with undefined elements.
480inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() {
481 return cst_pred_ty<is_lowbit_mask>();
482}
483
484struct icmp_pred_with_threshold {
485 ICmpInst::Predicate Pred;
486 const APInt *Thr;
487 bool isValue(const APInt &C) {
488 switch (Pred) {
489 case ICmpInst::Predicate::ICMP_EQ:
490 return C.eq(*Thr);
491 case ICmpInst::Predicate::ICMP_NE:
492 return C.ne(*Thr);
493 case ICmpInst::Predicate::ICMP_UGT:
494 return C.ugt(*Thr);
495 case ICmpInst::Predicate::ICMP_UGE:
496 return C.uge(*Thr);
497 case ICmpInst::Predicate::ICMP_ULT:
498 return C.ult(*Thr);
499 case ICmpInst::Predicate::ICMP_ULE:
500 return C.ule(*Thr);
501 case ICmpInst::Predicate::ICMP_SGT:
502 return C.sgt(*Thr);
503 case ICmpInst::Predicate::ICMP_SGE:
504 return C.sge(*Thr);
505 case ICmpInst::Predicate::ICMP_SLT:
506 return C.slt(*Thr);
507 case ICmpInst::Predicate::ICMP_SLE:
508 return C.sle(*Thr);
509 default:
510 llvm_unreachable("Unhandled ICmp predicate")::llvm::llvm_unreachable_internal("Unhandled ICmp predicate",
"/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/include/llvm/IR/PatternMatch.h"
, 510)
;
511 }
512 }
513};
514/// Match an integer or vector with every element comparing 'pred' (eg/ne/...)
515/// to Threshold. For vectors, this includes constants with undefined elements.
516inline cst_pred_ty<icmp_pred_with_threshold>
517m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) {
518 cst_pred_ty<icmp_pred_with_threshold> P;
519 P.Pred = Predicate;
520 P.Thr = &Threshold;
521 return P;
522}
523
524struct is_nan {
525 bool isValue(const APFloat &C) { return C.isNaN(); }
526};
527/// Match an arbitrary NaN constant. This includes quiet and signalling nans.
528/// For vectors, this includes constants with undefined elements.
529inline cstfp_pred_ty<is_nan> m_NaN() {
530 return cstfp_pred_ty<is_nan>();
531}
532
533struct is_any_zero_fp {
534 bool isValue(const APFloat &C) { return C.isZero(); }
535};
536/// Match a floating-point negative zero or positive zero.
537/// For vectors, this includes constants with undefined elements.
538inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() {
539 return cstfp_pred_ty<is_any_zero_fp>();
540}
541
542struct is_pos_zero_fp {
543 bool isValue(const APFloat &C) { return C.isPosZero(); }
544};
545/// Match a floating-point positive zero.
546/// For vectors, this includes constants with undefined elements.
547inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() {
548 return cstfp_pred_ty<is_pos_zero_fp>();
549}
550
551struct is_neg_zero_fp {
552 bool isValue(const APFloat &C) { return C.isNegZero(); }
553};
554/// Match a floating-point negative zero.
555/// For vectors, this includes constants with undefined elements.
556inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() {
557 return cstfp_pred_ty<is_neg_zero_fp>();
558}
559
560///////////////////////////////////////////////////////////////////////////////
561
562template <typename Class> struct bind_ty {
563 Class *&VR;
564
565 bind_ty(Class *&V) : VR(V) {}
566
567 template <typename ITy> bool match(ITy *V) {
568 if (auto *CV = dyn_cast<Class>(V)) {
569 VR = CV;
570 return true;
571 }
572 return false;
573 }
574};
575
576/// Match a value, capturing it if we match.
577inline 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'
47
Returning without writing to 'V'
578inline bind_ty<const Value> m_Value(const Value *&V) { return V; }
579
580/// Match an instruction, capturing it if we match.
581inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; }
582/// Match a binary operator, capturing it if we match.
583inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; }
584/// Match a with overflow intrinsic, capturing it if we match.
585inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; }
586
587/// Match a ConstantInt, capturing the value if we match.
588inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; }
589
590/// Match a Constant, capturing the value if we match.
591inline bind_ty<Constant> m_Constant(Constant *&C) { return C; }
592
593/// Match a ConstantFP, capturing the value if we match.
594inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; }
595
596/// Match a basic block value, capturing it if we match.
597inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; }
598inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) {
599 return V;
600}
601
602/// Match a specified Value*.
603struct specificval_ty {
604 const Value *Val;
605
606 specificval_ty(const Value *V) : Val(V) {}
607
608 template <typename ITy> bool match(ITy *V) { return V == Val; }
609};
610
611/// Match if we have a specific specified value.
612inline specificval_ty m_Specific(const Value *V) { return V; }
613
614/// Stores a reference to the Value *, not the Value * itself,
615/// thus can be used in commutative matchers.
616template <typename Class> struct deferredval_ty {
617 Class *const &Val;
618
619 deferredval_ty(Class *const &V) : Val(V) {}
620
621 template <typename ITy> bool match(ITy *const V) { return V == Val; }
622};
623
624/// A commutative-friendly version of m_Specific().
625inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; }
626inline deferredval_ty<const Value> m_Deferred(const Value *const &V) {
627 return V;
628}
629
630/// Match a specified floating point value or vector of all elements of
631/// that value.
632struct specific_fpval {
633 double Val;
634
635 specific_fpval(double V) : Val(V) {}
636
637 template <typename ITy> bool match(ITy *V) {
638 if (const auto *CFP = dyn_cast<ConstantFP>(V))
639 return CFP->isExactlyValue(Val);
640 if (V->getType()->isVectorTy())
641 if (const auto *C = dyn_cast<Constant>(V))
642 if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
643 return CFP->isExactlyValue(Val);
644 return false;
645 }
646};
647
648/// Match a specific floating point value or vector with all elements
649/// equal to the value.
650inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); }
651
652/// Match a float 1.0 or vector with all elements equal to 1.0.
653inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); }
654
655struct bind_const_intval_ty {
656 uint64_t &VR;
657
658 bind_const_intval_ty(uint64_t &V) : VR(V) {}
659
660 template <typename ITy> bool match(ITy *V) {
661 if (const auto *CV = dyn_cast<ConstantInt>(V))
662 if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) {
663 VR = CV->getZExtValue();
664 return true;
665 }
666 return false;
667 }
668};
669
670/// Match a specified integer value or vector of all elements of that
671/// value.
672struct specific_intval {
673 APInt Val;
674
675 specific_intval(APInt V) : Val(std::move(V)) {}
676
677 template <typename ITy> bool match(ITy *V) {
678 const auto *CI = dyn_cast<ConstantInt>(V);
679 if (!CI && V->getType()->isVectorTy())
680 if (const auto *C = dyn_cast<Constant>(V))
681 CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue());
682
683 return CI && APInt::isSameValue(CI->getValue(), Val);
684 }
685};
686
687/// Match a specific integer value or vector with all elements equal to
688/// the value.
689inline specific_intval m_SpecificInt(APInt V) {
690 return specific_intval(std::move(V));
691}
692
693inline specific_intval m_SpecificInt(uint64_t V) {
694 return m_SpecificInt(APInt(64, V));
695}
696
697/// Match a ConstantInt and bind to its value. This does not match
698/// ConstantInts wider than 64-bits.
699inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; }
700
701/// Match a specified basic block value.
702struct specific_bbval {
703 BasicBlock *Val;
704
705 specific_bbval(BasicBlock *Val) : Val(Val) {}
706
707 template <typename ITy> bool match(ITy *V) {
708 const auto *BB = dyn_cast<BasicBlock>(V);
709 return BB && BB == Val;
710 }
711};
712
713/// Match a specific basic block value.
714inline specific_bbval m_SpecificBB(BasicBlock *BB) {
715 return specific_bbval(BB);
716}
717
718/// A commutative-friendly version of m_Specific().
719inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) {
720 return BB;
721}
722inline deferredval_ty<const BasicBlock>
723m_Deferred(const BasicBlock *const &BB) {
724 return BB;
725}
726
727//===----------------------------------------------------------------------===//
728// Matcher for any binary operator.
729//
730template <typename LHS_t, typename RHS_t, bool Commutable = false>
731struct AnyBinaryOp_match {
732 LHS_t L;
733 RHS_t R;
734
735 // The evaluation order is always stable, regardless of Commutability.
736 // The LHS is always matched first.
737 AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
738
739 template <typename OpTy> bool match(OpTy *V) {
740 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'
16
Taking true branch
741 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
17
Calling 'match_combine_or::match'
32
Returning from 'match_combine_or::match'
33
Returning the value 1, which participates in a condition later
742 (Commutable && L.match(I->getOperand(1)) &&
743 R.match(I->getOperand(0)));
744 return false;
745 }
746};
747
748template <typename LHS, typename RHS>
749inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) {
750 return AnyBinaryOp_match<LHS, RHS>(L, R);
751}
752
753//===----------------------------------------------------------------------===//
754// Matchers for specific binary operators.
755//
756
757template <typename LHS_t, typename RHS_t, unsigned Opcode,
758 bool Commutable = false>
759struct BinaryOp_match {
760 LHS_t L;
761 RHS_t R;
762
763 // The evaluation order is always stable, regardless of Commutability.
764 // The LHS is always matched first.
765 BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
766
767 template <typename OpTy> bool match(OpTy *V) {
768 if (V->getValueID() == Value::InstructionVal + Opcode) {
21
Assuming the condition is true
22
Taking true branch
769 auto *I = cast<BinaryOperator>(V);
23
'V' is a 'BinaryOperator'
770 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
771 (Commutable && L.match(I->getOperand(1)) &&
772 R.match(I->getOperand(0)));
773 }
774 if (auto *CE = dyn_cast<ConstantExpr>(V))
775 return CE->getOpcode() == Opcode &&
776 ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) ||
777 (Commutable && L.match(CE->getOperand(1)) &&
778 R.match(CE->getOperand(0))));
779 return false;
780 }
781};
782
783template <typename LHS, typename RHS>
784inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L,
785 const RHS &R) {
786 return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R);
787}
788
789template <typename LHS, typename RHS>
790inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L,
791 const RHS &R) {
792 return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R);
793}
794
795template <typename LHS, typename RHS>
796inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L,
797 const RHS &R) {
798 return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R);
799}
800
801template <typename LHS, typename RHS>
802inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L,
803 const RHS &R) {
804 return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R);
805}
806
807template <typename Op_t> struct FNeg_match {
808 Op_t X;
809
810 FNeg_match(const Op_t &Op) : X(Op) {}
811 template <typename OpTy> bool match(OpTy *V) {
812 auto *FPMO = dyn_cast<FPMathOperator>(V);
813 if (!FPMO) return false;
814
815 if (FPMO->getOpcode() == Instruction::FNeg)
816 return X.match(FPMO->getOperand(0));
817
818 if (FPMO->getOpcode() == Instruction::FSub) {
819 if (FPMO->hasNoSignedZeros()) {
820 // With 'nsz', any zero goes.
821 if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0)))
822 return false;
823 } else {
824 // Without 'nsz', we need fsub -0.0, X exactly.
825 if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0)))
826 return false;
827 }
828
829 return X.match(FPMO->getOperand(1));
830 }
831
832 return false;
833 }
834};
835
836/// Match 'fneg X' as 'fsub -0.0, X'.
837template <typename OpTy>
838inline FNeg_match<OpTy>
839m_FNeg(const OpTy &X) {
840 return FNeg_match<OpTy>(X);
841}
842
843/// Match 'fneg X' as 'fsub +-0.0, X'.
844template <typename RHS>
845inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub>
846m_FNegNSZ(const RHS &X) {
847 return m_FSub(m_AnyZeroFP(), X);
848}
849
850template <typename LHS, typename RHS>
851inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L,
852 const RHS &R) {
853 return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R);
854}
855
856template <typename LHS, typename RHS>
857inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L,
858 const RHS &R) {
859 return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R);
860}
861
862template <typename LHS, typename RHS>
863inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L,
864 const RHS &R) {
865 return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R);
866}
867
868template <typename LHS, typename RHS>
869inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L,
870 const RHS &R) {
871 return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R);
872}
873
874template <typename LHS, typename RHS>
875inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L,
876 const RHS &R) {
877 return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R);
878}
879
880template <typename LHS, typename RHS>
881inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L,
882 const RHS &R) {
883 return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R);
884}
885
886template <typename LHS, typename RHS>
887inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L,
888 const RHS &R) {
889 return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R);
890}
891
892template <typename LHS, typename RHS>
893inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L,
894 const RHS &R) {
895 return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R);
896}
897
898template <typename LHS, typename RHS>
899inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L,
900 const RHS &R) {
901 return BinaryOp_match<LHS, RHS, Instruction::And>(L, R);
902}
903
904template <typename LHS, typename RHS>
905inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L,
906 const RHS &R) {
907 return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R);
908}
909
910template <typename LHS, typename RHS>
911inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L,
912 const RHS &R) {
913 return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R);
914}
915
916template <typename LHS, typename RHS>
917inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L,
918 const RHS &R) {
919 return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R);
920}
921
922template <typename LHS, typename RHS>
923inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L,
924 const RHS &R) {
925 return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R);
926}
927
928template <typename LHS, typename RHS>
929inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L,
930 const RHS &R) {
931 return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R);
932}
933
934template <typename LHS_t, typename RHS_t, unsigned Opcode,
935 unsigned WrapFlags = 0>
936struct OverflowingBinaryOp_match {
937 LHS_t L;
938 RHS_t R;
939
940 OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS)
941 : L(LHS), R(RHS) {}
942
943 template <typename OpTy> bool match(OpTy *V) {
944 if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) {
945 if (Op->getOpcode() != Opcode)
946 return false;
947 if (WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap &&
948 !Op->hasNoUnsignedWrap())
949 return false;
950 if (WrapFlags & OverflowingBinaryOperator::NoSignedWrap &&
951 !Op->hasNoSignedWrap())
952 return false;
953 return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1));
954 }
955 return false;
956 }
957};
958
959template <typename LHS, typename RHS>
960inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
961 OverflowingBinaryOperator::NoSignedWrap>
962m_NSWAdd(const LHS &L, const RHS &R) {
963 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
964 OverflowingBinaryOperator::NoSignedWrap>(
965 L, R);
966}
967template <typename LHS, typename RHS>
968inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
969 OverflowingBinaryOperator::NoSignedWrap>
970m_NSWSub(const LHS &L, const RHS &R) {
971 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
972 OverflowingBinaryOperator::NoSignedWrap>(
973 L, R);
974}
975template <typename LHS, typename RHS>
976inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
977 OverflowingBinaryOperator::NoSignedWrap>
978m_NSWMul(const LHS &L, const RHS &R) {
979 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
980 OverflowingBinaryOperator::NoSignedWrap>(
981 L, R);
982}
983template <typename LHS, typename RHS>
984inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
985 OverflowingBinaryOperator::NoSignedWrap>
986m_NSWShl(const LHS &L, const RHS &R) {
987 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
988 OverflowingBinaryOperator::NoSignedWrap>(
989 L, R);
990}
991
992template <typename LHS, typename RHS>
993inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
994 OverflowingBinaryOperator::NoUnsignedWrap>
995m_NUWAdd(const LHS &L, const RHS &R) {
996 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
997 OverflowingBinaryOperator::NoUnsignedWrap>(
998 L, R);
999}
1000template <typename LHS, typename RHS>
1001inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1002 OverflowingBinaryOperator::NoUnsignedWrap>
1003m_NUWSub(const LHS &L, const RHS &R) {
1004 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1005 OverflowingBinaryOperator::NoUnsignedWrap>(
1006 L, R);
1007}
1008template <typename LHS, typename RHS>
1009inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1010 OverflowingBinaryOperator::NoUnsignedWrap>
1011m_NUWMul(const LHS &L, const RHS &R) {
1012 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1013 OverflowingBinaryOperator::NoUnsignedWrap>(
1014 L, R);
1015}
1016template <typename LHS, typename RHS>
1017inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1018 OverflowingBinaryOperator::NoUnsignedWrap>
1019m_NUWShl(const LHS &L, const RHS &R) {
1020 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1021 OverflowingBinaryOperator::NoUnsignedWrap>(
1022 L, R);
1023}
1024
1025//===----------------------------------------------------------------------===//
1026// Class that matches a group of binary opcodes.
1027//
1028template <typename LHS_t, typename RHS_t, typename Predicate>
1029struct BinOpPred_match : Predicate {
1030 LHS_t L;
1031 RHS_t R;
1032
1033 BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1034
1035 template <typename OpTy> bool match(OpTy *V) {
1036 if (auto *I = dyn_cast<Instruction>(V))
1037 return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) &&
1038 R.match(I->getOperand(1));
1039 if (auto *CE = dyn_cast<ConstantExpr>(V))
1040 return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) &&
1041 R.match(CE->getOperand(1));
1042 return false;
1043 }
1044};
1045
1046struct is_shift_op {
1047 bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); }
1048};
1049
1050struct is_right_shift_op {
1051 bool isOpType(unsigned Opcode) {
1052 return Opcode == Instruction::LShr || Opcode == Instruction::AShr;
1053 }
1054};
1055
1056struct is_logical_shift_op {
1057 bool isOpType(unsigned Opcode) {
1058 return Opcode == Instruction::LShr || Opcode == Instruction::Shl;
1059 }
1060};
1061
1062struct is_bitwiselogic_op {
1063 bool isOpType(unsigned Opcode) {
1064 return Instruction::isBitwiseLogicOp(Opcode);
1065 }
1066};
1067
1068struct is_idiv_op {
1069 bool isOpType(unsigned Opcode) {
1070 return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
1071 }
1072};
1073
1074struct is_irem_op {
1075 bool isOpType(unsigned Opcode) {
1076 return Opcode == Instruction::SRem || Opcode == Instruction::URem;
1077 }
1078};
1079
1080/// Matches shift operations.
1081template <typename LHS, typename RHS>
1082inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L,
1083 const RHS &R) {
1084 return BinOpPred_match<LHS, RHS, is_shift_op>(L, R);
1085}
1086
1087/// Matches logical shift operations.
1088template <typename LHS, typename RHS>
1089inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L,
1090 const RHS &R) {
1091 return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R);
1092}
1093
1094/// Matches logical shift operations.
1095template <typename LHS, typename RHS>
1096inline BinOpPred_match<LHS, RHS, is_logical_shift_op>
1097m_LogicalShift(const LHS &L, const RHS &R) {
1098 return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R);
1099}
1100
1101/// Matches bitwise logic operations.
1102template <typename LHS, typename RHS>
1103inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op>
1104m_BitwiseLogic(const LHS &L, const RHS &R) {
1105 return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R);
1106}
1107
1108/// Matches integer division operations.
1109template <typename LHS, typename RHS>
1110inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L,
1111 const RHS &R) {
1112 return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R);
1113}
1114
1115/// Matches integer remainder operations.
1116template <typename LHS, typename RHS>
1117inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L,
1118 const RHS &R) {
1119 return BinOpPred_match<LHS, RHS, is_irem_op>(L, R);
1120}
1121
1122//===----------------------------------------------------------------------===//
1123// Class that matches exact binary ops.
1124//
1125template <typename SubPattern_t> struct Exact_match {
1126 SubPattern_t SubPattern;
1127
1128 Exact_match(const SubPattern_t &SP) : SubPattern(SP) {}
1129
1130 template <typename OpTy> bool match(OpTy *V) {
1131 if (auto *PEO = dyn_cast<PossiblyExactOperator>(V))
1132 return PEO->isExact() && SubPattern.match(V);
1133 return false;
1134 }
1135};
1136
1137template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) {
1138 return SubPattern;
1139}
1140
1141//===----------------------------------------------------------------------===//
1142// Matchers for CmpInst classes
1143//
1144
1145template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy,
1146 bool Commutable = false>
1147struct CmpClass_match {
1148 PredicateTy &Predicate;
1149 LHS_t L;
1150 RHS_t R;
1151
1152 // The evaluation order is always stable, regardless of Commutability.
1153 // The LHS is always matched first.
1154 CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS)
1155 : Predicate(Pred), L(LHS), R(RHS) {}
1156
1157 template <typename OpTy> bool match(OpTy *V) {
1158 if (auto *I = dyn_cast<Class>(V))
1159 if ((L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
1160 (Commutable && L.match(I->getOperand(1)) &&
1161 R.match(I->getOperand(0)))) {
1162 Predicate = I->getPredicate();
1163 return true;
1164 }
1165 return false;
1166 }
1167};
1168
1169template <typename LHS, typename RHS>
1170inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>
1171m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1172 return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R);
1173}
1174
1175template <typename LHS, typename RHS>
1176inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>
1177m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1178 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R);
1179}
1180
1181template <typename LHS, typename RHS>
1182inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>
1183m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1184 return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R);
1185}
1186
1187//===----------------------------------------------------------------------===//
1188// Matchers for instructions with a given opcode and number of operands.
1189//
1190
1191/// Matches instructions with Opcode and three operands.
1192template <typename T0, unsigned Opcode> struct OneOps_match {
1193 T0 Op1;
1194
1195 OneOps_match(const T0 &Op1) : Op1(Op1) {}
1196
1197 template <typename OpTy> bool match(OpTy *V) {
1198 if (V->getValueID() == Value::InstructionVal + Opcode) {
1199 auto *I = cast<Instruction>(V);
1200 return Op1.match(I->getOperand(0));
1201 }
1202 return false;
1203 }
1204};
1205
1206/// Matches instructions with Opcode and three operands.
1207template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match {
1208 T0 Op1;
1209 T1 Op2;
1210
1211 TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {}
1212
1213 template <typename OpTy> bool match(OpTy *V) {
1214 if (V->getValueID() == Value::InstructionVal + Opcode) {
1215 auto *I = cast<Instruction>(V);
1216 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1));
1217 }
1218 return false;
1219 }
1220};
1221
1222/// Matches instructions with Opcode and three operands.
1223template <typename T0, typename T1, typename T2, unsigned Opcode>
1224struct ThreeOps_match {
1225 T0 Op1;
1226 T1 Op2;
1227 T2 Op3;
1228
1229 ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3)
1230 : Op1(Op1), Op2(Op2), Op3(Op3) {}
1231
1232 template <typename OpTy> bool match(OpTy *V) {
1233 if (V->getValueID() == Value::InstructionVal + Opcode) {
67
Called C++ object pointer is null
1234 auto *I = cast<Instruction>(V);
1235 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1236 Op3.match(I->getOperand(2));
1237 }
1238 return false;
1239 }
1240};
1241
1242/// Matches SelectInst.
1243template <typename Cond, typename LHS, typename RHS>
1244inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select>
1245m_Select(const Cond &C, const LHS &L, const RHS &R) {
1246 return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R);
1247}
1248
1249/// This matches a select of two constants, e.g.:
1250/// m_SelectCst<-1, 0>(m_Value(V))
1251template <int64_t L, int64_t R, typename Cond>
1252inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>,
1253 Instruction::Select>
1254m_SelectCst(const Cond &C) {
1255 return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>());
1256}
1257
1258/// Matches FreezeInst.
1259template <typename OpTy>
1260inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) {
1261 return OneOps_match<OpTy, Instruction::Freeze>(Op);
1262}
1263
1264/// Matches InsertElementInst.
1265template <typename Val_t, typename Elt_t, typename Idx_t>
1266inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>
1267m_InsertElement(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) {
1268 return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>(
1269 Val, Elt, Idx);
1270}
1271
1272/// Matches ExtractElementInst.
1273template <typename Val_t, typename Idx_t>
1274inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>
1275m_ExtractElement(const Val_t &Val, const Idx_t &Idx) {
1276 return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx);
1277}
1278
1279/// Matches ShuffleVectorInst.
1280template <typename V1_t, typename V2_t, typename Mask_t>
1281inline ThreeOps_match<V1_t, V2_t, Mask_t, Instruction::ShuffleVector>
1282m_ShuffleVector(const V1_t &v1, const V2_t &v2, const Mask_t &m) {
1283 return ThreeOps_match<V1_t, V2_t, Mask_t, Instruction::ShuffleVector>(v1, v2,
1284 m);
1285}
1286
1287/// Matches LoadInst.
1288template <typename OpTy>
1289inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) {
1290 return OneOps_match<OpTy, Instruction::Load>(Op);
1291}
1292
1293/// Matches StoreInst.
1294template <typename ValueOpTy, typename PointerOpTy>
1295inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>
1296m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) {
1297 return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp,
1298 PointerOp);
1299}
1300
1301//===----------------------------------------------------------------------===//
1302// Matchers for CastInst classes
1303//
1304
1305template <typename Op_t, unsigned Opcode> struct CastClass_match {
1306 Op_t Op;
1307
1308 CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
1309
1310 template <typename OpTy> bool match(OpTy *V) {
1311 if (auto *O = dyn_cast<Operator>(V))
1312 return O->getOpcode() == Opcode && Op.match(O->getOperand(0));
1313 return false;
1314 }
1315};
1316
1317/// Matches BitCast.
1318template <typename OpTy>
1319inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) {
1320 return CastClass_match<OpTy, Instruction::BitCast>(Op);
1321}
1322
1323/// Matches PtrToInt.
1324template <typename OpTy>
1325inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) {
1326 return CastClass_match<OpTy, Instruction::PtrToInt>(Op);
1327}
1328
1329/// Matches Trunc.
1330template <typename OpTy>
1331inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) {
1332 return CastClass_match<OpTy, Instruction::Trunc>(Op);
1333}
1334
1335template <typename OpTy>
1336inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy>
1337m_TruncOrSelf(const OpTy &Op) {
1338 return m_CombineOr(m_Trunc(Op), Op);
1339}
1340
1341/// Matches SExt.
1342template <typename OpTy>
1343inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) {
1344 return CastClass_match<OpTy, Instruction::SExt>(Op);
1345}
1346
1347/// Matches ZExt.
1348template <typename OpTy>
1349inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) {
1350 return CastClass_match<OpTy, Instruction::ZExt>(Op);
51
Returning without writing to 'Op.VR'
1351}
1352
1353template <typename OpTy>
1354inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy>
1355m_ZExtOrSelf(const OpTy &Op) {
1356 return m_CombineOr(m_ZExt(Op), Op);
50
Calling 'm_ZExt<llvm::PatternMatch::bind_ty<llvm::Value>>'
52
Returning from 'm_ZExt<llvm::PatternMatch::bind_ty<llvm::Value>>'
53
Calling 'm_CombineOr<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>'
55
Returning from 'm_CombineOr<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>'
56
Returning without writing to 'Op.VR'
1357}
1358
1359template <typename OpTy>
1360inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy>
1361m_SExtOrSelf(const OpTy &Op) {
1362 return m_CombineOr(m_SExt(Op), Op);
1363}
1364
1365template <typename OpTy>
1366inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1367 CastClass_match<OpTy, Instruction::SExt>>
1368m_ZExtOrSExt(const OpTy &Op) {
1369 return m_CombineOr(m_ZExt(Op), m_SExt(Op));
1370}
1371
1372template <typename OpTy>
1373inline match_combine_or<
1374 match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1375 CastClass_match<OpTy, Instruction::SExt>>,
1376 OpTy>
1377m_ZExtOrSExtOrSelf(const OpTy &Op) {
1378 return m_CombineOr(m_ZExtOrSExt(Op), Op);
1379}
1380
1381/// Matches UIToFP.
1382template <typename OpTy>
1383inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) {
1384 return CastClass_match<OpTy, Instruction::UIToFP>(Op);
1385}
1386
1387/// Matches SIToFP.
1388template <typename OpTy>
1389inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
1390 return CastClass_match<OpTy, Instruction::SIToFP>(Op);
1391}
1392
1393/// Matches FPTrunc
1394template <typename OpTy>
1395inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
1396 return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
1397}
1398
1399/// Matches FPExt
1400template <typename OpTy>
1401inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
1402 return CastClass_match<OpTy, Instruction::FPExt>(Op);
1403}
1404
1405//===----------------------------------------------------------------------===//
1406// Matchers for control flow.
1407//
1408
1409struct br_match {
1410 BasicBlock *&Succ;
1411
1412 br_match(BasicBlock *&Succ) : Succ(Succ) {}
1413
1414 template <typename OpTy> bool match(OpTy *V) {
1415 if (auto *BI = dyn_cast<BranchInst>(V))
1416 if (BI->isUnconditional()) {
1417 Succ = BI->getSuccessor(0);
1418 return true;
1419 }
1420 return false;
1421 }
1422};
1423
1424inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); }
1425
1426template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1427struct brc_match {
1428 Cond_t Cond;
1429 TrueBlock_t T;
1430 FalseBlock_t F;
1431
1432 brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f)
1433 : Cond(C), T(t), F(f) {}
1434
1435 template <typename OpTy> bool match(OpTy *V) {
1436 if (auto *BI = dyn_cast<BranchInst>(V))
1437 if (BI->isConditional() && Cond.match(BI->getCondition()))
1438 return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1));
1439 return false;
1440 }
1441};
1442
1443template <typename Cond_t>
1444inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>
1445m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) {
1446 return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>(
1447 C, m_BasicBlock(T), m_BasicBlock(F));
1448}
1449
1450template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1451inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t>
1452m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) {
1453 return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F);
1454}
1455
1456//===----------------------------------------------------------------------===//
1457// Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y).
1458//
1459
1460template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t,
1461 bool Commutable = false>
1462struct MaxMin_match {
1463 LHS_t L;
1464 RHS_t R;
1465
1466 // The evaluation order is always stable, regardless of Commutability.
1467 // The LHS is always matched first.
1468 MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1469
1470 template <typename OpTy> bool match(OpTy *V) {
1471 // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x".
1472 auto *SI = dyn_cast<SelectInst>(V);
1473 if (!SI)
1474 return false;
1475 auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition());
1476 if (!Cmp)
1477 return false;
1478 // At this point we have a select conditioned on a comparison. Check that
1479 // it is the values returned by the select that are being compared.
1480 Value *TrueVal = SI->getTrueValue();
1481 Value *FalseVal = SI->getFalseValue();
1482 Value *LHS = Cmp->getOperand(0);
1483 Value *RHS = Cmp->getOperand(1);
1484 if ((TrueVal != LHS || FalseVal != RHS) &&
1485 (TrueVal != RHS || FalseVal != LHS))
1486 return false;
1487 typename CmpInst_t::Predicate Pred =
1488 LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate();
1489 // Does "(x pred y) ? x : y" represent the desired max/min operation?
1490 if (!Pred_t::match(Pred))
1491 return false;
1492 // It does! Bind the operands.
1493 return (L.match(LHS) && R.match(RHS)) ||
1494 (Commutable && L.match(RHS) && R.match(LHS));
1495 }
1496};
1497
1498/// Helper class for identifying signed max predicates.
1499struct smax_pred_ty {
1500 static bool match(ICmpInst::Predicate Pred) {
1501 return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE;
1502 }
1503};
1504
1505/// Helper class for identifying signed min predicates.
1506struct smin_pred_ty {
1507 static bool match(ICmpInst::Predicate Pred) {
1508 return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE;
1509 }
1510};
1511
1512/// Helper class for identifying unsigned max predicates.
1513struct umax_pred_ty {
1514 static bool match(ICmpInst::Predicate Pred) {
1515 return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE;
1516 }
1517};
1518
1519/// Helper class for identifying unsigned min predicates.
1520struct umin_pred_ty {
1521 static bool match(ICmpInst::Predicate Pred) {
1522 return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE;
1523 }
1524};
1525
1526/// Helper class for identifying ordered max predicates.
1527struct ofmax_pred_ty {
1528 static bool match(FCmpInst::Predicate Pred) {
1529 return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE;
1530 }
1531};
1532
1533/// Helper class for identifying ordered min predicates.
1534struct ofmin_pred_ty {
1535 static bool match(FCmpInst::Predicate Pred) {
1536 return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE;
1537 }
1538};
1539
1540/// Helper class for identifying unordered max predicates.
1541struct ufmax_pred_ty {
1542 static bool match(FCmpInst::Predicate Pred) {
1543 return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE;
1544 }
1545};
1546
1547/// Helper class for identifying unordered min predicates.
1548struct ufmin_pred_ty {
1549 static bool match(FCmpInst::Predicate Pred) {
1550 return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE;
1551 }
1552};
1553
1554template <typename LHS, typename RHS>
1555inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L,
1556 const RHS &R) {
1557 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R);
1558}
1559
1560template <typename LHS, typename RHS>
1561inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L,
1562 const RHS &R) {
1563 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R);
1564}
1565
1566template <typename LHS, typename RHS>
1567inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L,
1568 const RHS &R) {
1569 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R);
1570}
1571
1572template <typename LHS, typename RHS>
1573inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L,
1574 const RHS &R) {
1575 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R);
1576}
1577
1578/// Match an 'ordered' floating point maximum function.
1579/// Floating point has one special value 'NaN'. Therefore, there is no total
1580/// order. However, if we can ignore the 'NaN' value (for example, because of a
1581/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1582/// semantics. In the presence of 'NaN' we have to preserve the original
1583/// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate.
1584///
1585/// max(L, R) iff L and R are not NaN
1586/// m_OrdFMax(L, R) = R iff L or R are NaN
1587template <typename LHS, typename RHS>
1588inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L,
1589 const RHS &R) {
1590 return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R);
1591}
1592
1593/// Match an 'ordered' floating point minimum function.
1594/// Floating point has one special value 'NaN'. Therefore, there is no total
1595/// order. However, if we can ignore the 'NaN' value (for example, because of a
1596/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1597/// semantics. In the presence of 'NaN' we have to preserve the original
1598/// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate.
1599///
1600/// min(L, R) iff L and R are not NaN
1601/// m_OrdFMin(L, R) = R iff L or R are NaN
1602template <typename LHS, typename RHS>
1603inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L,
1604 const RHS &R) {
1605 return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R);
1606}
1607
1608/// Match an 'unordered' floating point maximum function.
1609/// Floating point has one special value 'NaN'. Therefore, there is no total
1610/// order. However, if we can ignore the 'NaN' value (for example, because of a
1611/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1612/// semantics. In the presence of 'NaN' we have to preserve the original
1613/// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate.
1614///
1615/// max(L, R) iff L and R are not NaN
1616/// m_UnordFMax(L, R) = L iff L or R are NaN
1617template <typename LHS, typename RHS>
1618inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>
1619m_UnordFMax(const LHS &L, const RHS &R) {
1620 return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R);
1621}
1622
1623/// Match an 'unordered' floating point minimum function.
1624/// Floating point has one special value 'NaN'. Therefore, there is no total
1625/// order. However, if we can ignore the 'NaN' value (for example, because of a
1626/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1627/// semantics. In the presence of 'NaN' we have to preserve the original
1628/// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate.
1629///
1630/// min(L, R) iff L and R are not NaN
1631/// m_UnordFMin(L, R) = L iff L or R are NaN
1632template <typename LHS, typename RHS>
1633inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>
1634m_UnordFMin(const LHS &L, const RHS &R) {
1635 return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R);
1636}
1637
1638//===----------------------------------------------------------------------===//
1639// Matchers for overflow check patterns: e.g. (a + b) u< a
1640//
1641
1642template <typename LHS_t, typename RHS_t, typename Sum_t>
1643struct UAddWithOverflow_match {
1644 LHS_t L;
1645 RHS_t R;
1646 Sum_t S;
1647
1648 UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S)
1649 : L(L), R(R), S(S) {}
1650
1651 template <typename OpTy> bool match(OpTy *V) {
1652 Value *ICmpLHS, *ICmpRHS;
1653 ICmpInst::Predicate Pred;
1654 if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V))
1655 return false;
1656
1657 Value *AddLHS, *AddRHS;
1658 auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS));
1659
1660 // (a + b) u< a, (a + b) u< b
1661 if (Pred == ICmpInst::ICMP_ULT)
1662 if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS))
1663 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1664
1665 // a >u (a + b), b >u (a + b)
1666 if (Pred == ICmpInst::ICMP_UGT)
1667 if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS))
1668 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1669
1670 // Match special-case for increment-by-1.
1671 if (Pred == ICmpInst::ICMP_EQ) {
1672 // (a + 1) == 0
1673 // (1 + a) == 0
1674 if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) &&
1675 (m_One().match(AddLHS) || m_One().match(AddRHS)))
1676 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1677 // 0 == (a + 1)
1678 // 0 == (1 + a)
1679 if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) &&
1680 (m_One().match(AddLHS) || m_One().match(AddRHS)))
1681 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1682 }
1683
1684 return false;
1685 }
1686};
1687
1688/// Match an icmp instruction checking for unsigned overflow on addition.
1689///
1690/// S is matched to the addition whose result is being checked for overflow, and
1691/// L and R are matched to the LHS and RHS of S.
1692template <typename LHS_t, typename RHS_t, typename Sum_t>
1693UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>
1694m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) {
1695 return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S);
1696}
1697
1698template <typename Opnd_t> struct Argument_match {
1699 unsigned OpI;
1700 Opnd_t Val;
1701
1702 Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {}
1703
1704 template <typename OpTy> bool match(OpTy *V) {
1705 // FIXME: Should likely be switched to use `CallBase`.
1706 if (const auto *CI = dyn_cast<CallInst>(V))
1707 return Val.match(CI->getArgOperand(OpI));
1708 return false;
1709 }
1710};
1711
1712/// Match an argument.
1713template <unsigned OpI, typename Opnd_t>
1714inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) {
1715 return Argument_match<Opnd_t>(OpI, Op);
1716}
1717
1718/// Intrinsic matchers.
1719struct IntrinsicID_match {
1720 unsigned ID;
1721
1722 IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {}
1723
1724 template <typename OpTy> bool match(OpTy *V) {
1725 if (const auto *CI = dyn_cast<CallInst>(V))
1726 if (const auto *F = CI->getCalledFunction())
1727 return F->getIntrinsicID() == ID;
1728 return false;
1729 }
1730};
1731
1732/// Intrinsic matches are combinations of ID matchers, and argument
1733/// matchers. Higher arity matcher are defined recursively in terms of and-ing
1734/// them with lower arity matchers. Here's some convenient typedefs for up to
1735/// several arguments, and more can be added as needed
1736template <typename T0 = void, typename T1 = void, typename T2 = void,
1737 typename T3 = void, typename T4 = void, typename T5 = void,
1738 typename T6 = void, typename T7 = void, typename T8 = void,
1739 typename T9 = void, typename T10 = void>
1740struct m_Intrinsic_Ty;
1741template <typename T0> struct m_Intrinsic_Ty<T0> {
1742 using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>;
1743};
1744template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> {
1745 using Ty =
1746 match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>;
1747};
1748template <typename T0, typename T1, typename T2>
1749struct m_Intrinsic_Ty<T0, T1, T2> {
1750 using Ty =
1751 match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty,
1752 Argument_match<T2>>;
1753};
1754template <typename T0, typename T1, typename T2, typename T3>
1755struct m_Intrinsic_Ty<T0, T1, T2, T3> {
1756 using Ty =
1757 match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty,
1758 Argument_match<T3>>;
1759};
1760
1761template <typename T0, typename T1, typename T2, typename T3, typename T4>
1762struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> {
1763 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty,
1764 Argument_match<T4>>;
1765};
1766
1767/// Match intrinsic calls like this:
1768/// m_Intrinsic<Intrinsic::fabs>(m_Value(X))
1769template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() {
1770 return IntrinsicID_match(IntrID);
1771}
1772
1773template <Intrinsic::ID IntrID, typename T0>
1774inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) {
1775 return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0));
1776}
1777
1778template <Intrinsic::ID IntrID, typename T0, typename T1>
1779inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0,
1780 const T1 &Op1) {
1781 return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1));
1782}
1783
1784template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2>
1785inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty
1786m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) {
1787 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2));
1788}
1789
1790template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
1791 typename T3>
1792inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty
1793m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) {
1794 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3));
1795}
1796
1797template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
1798 typename T3, typename T4>
1799inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty
1800m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
1801 const T4 &Op4) {
1802 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3),
1803 m_Argument<4>(Op4));
1804}
1805
1806// Helper intrinsic matching specializations.
1807template <typename Opnd0>
1808inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) {
1809 return m_Intrinsic<Intrinsic::bitreverse>(Op0);
1810}
1811
1812template <typename Opnd0>
1813inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) {
1814 return m_Intrinsic<Intrinsic::bswap>(Op0);
1815}
1816
1817template <typename Opnd0>
1818inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) {
1819 return m_Intrinsic<Intrinsic::fabs>(Op0);
1820}
1821
1822template <typename Opnd0>
1823inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) {
1824 return m_Intrinsic<Intrinsic::canonicalize>(Op0);
1825}
1826
1827template <typename Opnd0, typename Opnd1>
1828inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0,
1829 const Opnd1 &Op1) {
1830 return m_Intrinsic<Intrinsic::minnum>(Op0, Op1);
1831}
1832
1833template <typename Opnd0, typename Opnd1>
1834inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0,
1835 const Opnd1 &Op1) {
1836 return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1);
1837}
1838
1839//===----------------------------------------------------------------------===//
1840// Matchers for two-operands operators with the operators in either order
1841//
1842
1843/// Matches a BinaryOperator with LHS and RHS in either order.
1844template <typename LHS, typename RHS>
1845inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) {
1846 return AnyBinaryOp_match<LHS, RHS, true>(L, R);
11
Returning without writing to 'R.VR'
1847}
1848
1849/// Matches an ICmp with a predicate over LHS and RHS in either order.
1850/// Does not swap the predicate.
1851template <typename LHS, typename RHS>
1852inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>
1853m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1854 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L,
1855 R);
1856}
1857
1858/// Matches a Add with LHS and RHS in either order.
1859template <typename LHS, typename RHS>
1860inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L,
1861 const RHS &R) {
1862 return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R);
1863}
1864
1865/// Matches a Mul with LHS and RHS in either order.
1866template <typename LHS, typename RHS>
1867inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L,
1868 const RHS &R) {
1869 return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R);
1870}
1871
1872/// Matches an And with LHS and RHS in either order.
1873template <typename LHS, typename RHS>
1874inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L,
1875 const RHS &R) {
1876 return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R);
1877}
1878
1879/// Matches an Or with LHS and RHS in either order.
1880template <typename LHS, typename RHS>
1881inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L,
1882 const RHS &R) {
1883 return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R);
1884}
1885
1886/// Matches an Xor with LHS and RHS in either order.
1887template <typename LHS, typename RHS>
1888inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L,
1889 const RHS &R) {
1890 return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R);
1891}
1892
1893/// Matches a 'Neg' as 'sub 0, V'.
1894template <typename ValTy>
1895inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub>
1896m_Neg(const ValTy &V) {
1897 return m_Sub(m_ZeroInt(), V);
1898}
1899
1900/// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
1901template <typename ValTy>
1902inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true>
1903m_Not(const ValTy &V) {
1904 return m_c_Xor(V, m_AllOnes());
1905}
1906
1907/// Matches an SMin with LHS and RHS in either order.
1908template <typename LHS, typename RHS>
1909inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>
1910m_c_SMin(const LHS &L, const RHS &R) {
1911 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R);
1912}
1913/// Matches an SMax with LHS and RHS in either order.
1914template <typename LHS, typename RHS>
1915inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>
1916m_c_SMax(const LHS &L, const RHS &R) {
1917 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R);
1918}
1919/// Matches a UMin with LHS and RHS in either order.
1920template <typename LHS, typename RHS>
1921inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>
1922m_c_UMin(const LHS &L, const RHS &R) {
1923 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R);
1924}
1925/// Matches a UMax with LHS and RHS in either order.
1926template <typename LHS, typename RHS>
1927inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>
1928m_c_UMax(const LHS &L, const RHS &R) {
1929 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R);
1930}
1931
1932/// Matches FAdd with LHS and RHS in either order.
1933template <typename LHS, typename RHS>
1934inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true>
1935m_c_FAdd(const LHS &L, const RHS &R) {
1936 return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R);
1937}
1938
1939/// Matches FMul with LHS and RHS in either order.
1940template <typename LHS, typename RHS>
1941inline BinaryOp_match<LHS, RHS, Instruction::FMul, true>
1942m_c_FMul(const LHS &L, const RHS &R) {
1943 return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R);
1944}
1945
1946template <typename Opnd_t> struct Signum_match {
1947 Opnd_t Val;
1948 Signum_match(const Opnd_t &V) : Val(V) {}
1949
1950 template <typename OpTy> bool match(OpTy *V) {
1951 unsigned TypeSize = V->getType()->getScalarSizeInBits();
1952 if (TypeSize == 0)
1953 return false;
1954
1955 unsigned ShiftWidth = TypeSize - 1;
1956 Value *OpL = nullptr, *OpR = nullptr;
1957
1958 // This is the representation of signum we match:
1959 //
1960 // signum(x) == (x >> 63) | (-x >>u 63)
1961 //
1962 // An i1 value is its own signum, so it's correct to match
1963 //
1964 // signum(x) == (x >> 0) | (-x >>u 0)
1965 //
1966 // for i1 values.
1967
1968 auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth));
1969 auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth));
1970 auto Signum = m_Or(LHS, RHS);
1971
1972 return Signum.match(V) && OpL == OpR && Val.match(OpL);
1973 }
1974};
1975
1976/// Matches a signum pattern.
1977///
1978/// signum(x) =
1979/// x > 0 -> 1
1980/// x == 0 -> 0
1981/// x < 0 -> -1
1982template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) {
1983 return Signum_match<Val_t>(V);
1984}
1985
1986template <int Ind, typename Opnd_t> struct ExtractValue_match {
1987 Opnd_t Val;
1988 ExtractValue_match(const Opnd_t &V) : Val(V) {}
1989
1990 template <typename OpTy> bool match(OpTy *V) {
1991 if (auto *I = dyn_cast<ExtractValueInst>(V))
1992 return I->getNumIndices() == 1 && I->getIndices()[0] == Ind &&
1993 Val.match(I->getAggregateOperand());
1994 return false;
1995 }
1996};
1997
1998/// Match a single index ExtractValue instruction.
1999/// For example m_ExtractValue<1>(...)
2000template <int Ind, typename Val_t>
2001inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) {
2002 return ExtractValue_match<Ind, Val_t>(V);
2003}
2004
2005} // end namespace PatternMatch
2006} // end namespace llvm
2007
2008#endif // LLVM_IR_PATTERNMATCH_H