Bug Summary

File:llvm/include/llvm/IR/PatternMatch.h
Warning:line 1460, column 9
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name InstCombineAddSub.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/lib/Transforms/InstCombine -I include -I /build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/= -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2022-01-19-134126-35450-1 -x c++ /build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

1//===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visit functions for add, fadd, sub, and fsub.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APFloat.h"
15#include "llvm/ADT/APInt.h"
16#include "llvm/ADT/STLExtras.h"
17#include "llvm/ADT/SmallVector.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/ValueTracking.h"
20#include "llvm/IR/Constant.h"
21#include "llvm/IR/Constants.h"
22#include "llvm/IR/InstrTypes.h"
23#include "llvm/IR/Instruction.h"
24#include "llvm/IR/Instructions.h"
25#include "llvm/IR/Operator.h"
26#include "llvm/IR/PatternMatch.h"
27#include "llvm/IR/Type.h"
28#include "llvm/IR/Value.h"
29#include "llvm/Support/AlignOf.h"
30#include "llvm/Support/Casting.h"
31#include "llvm/Support/KnownBits.h"
32#include "llvm/Transforms/InstCombine/InstCombiner.h"
33#include <cassert>
34#include <utility>
35
36using namespace llvm;
37using namespace PatternMatch;
38
39#define DEBUG_TYPE"instcombine" "instcombine"
40
41namespace {
42
43 /// Class representing coefficient of floating-point addend.
44 /// This class needs to be highly efficient, which is especially true for
45 /// the constructor. As of I write this comment, the cost of the default
46 /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
47 /// perform write-merging).
48 ///
49 class FAddendCoef {
50 public:
51 // The constructor has to initialize a APFloat, which is unnecessary for
52 // most addends which have coefficient either 1 or -1. So, the constructor
53 // is expensive. In order to avoid the cost of the constructor, we should
54 // reuse some instances whenever possible. The pre-created instances
55 // FAddCombine::Add[0-5] embodies this idea.
56 FAddendCoef() = default;
57 ~FAddendCoef();
58
59 // If possible, don't define operator+/operator- etc because these
60 // operators inevitably call FAddendCoef's constructor which is not cheap.
61 void operator=(const FAddendCoef &A);
62 void operator+=(const FAddendCoef &A);
63 void operator*=(const FAddendCoef &S);
64
65 void set(short C) {
66 assert(!insaneIntVal(C) && "Insane coefficient")(static_cast <bool> (!insaneIntVal(C) && "Insane coefficient"
) ? void (0) : __assert_fail ("!insaneIntVal(C) && \"Insane coefficient\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 66
, __extension__ __PRETTY_FUNCTION__))
;
67 IsFp = false; IntVal = C;
68 }
69
70 void set(const APFloat& C);
71
72 void negate();
73
74 bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
75 Value *getValue(Type *) const;
76
77 bool isOne() const { return isInt() && IntVal == 1; }
78 bool isTwo() const { return isInt() && IntVal == 2; }
79 bool isMinusOne() const { return isInt() && IntVal == -1; }
80 bool isMinusTwo() const { return isInt() && IntVal == -2; }
81
82 private:
83 bool insaneIntVal(int V) { return V > 4 || V < -4; }
84
85 APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }
86
87 const APFloat *getFpValPtr() const {
88 return reinterpret_cast<const APFloat *>(&FpValBuf);
89 }
90
91 const APFloat &getFpVal() const {
92 assert(IsFp && BufHasFpVal && "Incorret state")(static_cast <bool> (IsFp && BufHasFpVal &&
"Incorret state") ? void (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 92
, __extension__ __PRETTY_FUNCTION__))
;
93 return *getFpValPtr();
94 }
95
96 APFloat &getFpVal() {
97 assert(IsFp && BufHasFpVal && "Incorret state")(static_cast <bool> (IsFp && BufHasFpVal &&
"Incorret state") ? void (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 97
, __extension__ __PRETTY_FUNCTION__))
;
98 return *getFpValPtr();
99 }
100
101 bool isInt() const { return !IsFp; }
102
103 // If the coefficient is represented by an integer, promote it to a
104 // floating point.
105 void convertToFpType(const fltSemantics &Sem);
106
107 // Construct an APFloat from a signed integer.
108 // TODO: We should get rid of this function when APFloat can be constructed
109 // from an *SIGNED* integer.
110 APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
111
112 bool IsFp = false;
113
114 // True iff FpValBuf contains an instance of APFloat.
115 bool BufHasFpVal = false;
116
117 // The integer coefficient of an individual addend is either 1 or -1,
118 // and we try to simplify at most 4 addends from neighboring at most
119 // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
120 // is overkill of this end.
121 short IntVal = 0;
122
123 AlignedCharArrayUnion<APFloat> FpValBuf;
124 };
125
126 /// FAddend is used to represent floating-point addend. An addend is
127 /// represented as <C, V>, where the V is a symbolic value, and C is a
128 /// constant coefficient. A constant addend is represented as <C, 0>.
129 class FAddend {
130 public:
131 FAddend() = default;
132
133 void operator+=(const FAddend &T) {
134 assert((Val == T.Val) && "Symbolic-values disagree")(static_cast <bool> ((Val == T.Val) && "Symbolic-values disagree"
) ? void (0) : __assert_fail ("(Val == T.Val) && \"Symbolic-values disagree\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 134
, __extension__ __PRETTY_FUNCTION__))
;
135 Coeff += T.Coeff;
136 }
137
138 Value *getSymVal() const { return Val; }
139 const FAddendCoef &getCoef() const { return Coeff; }
140
141 bool isConstant() const { return Val == nullptr; }
142 bool isZero() const { return Coeff.isZero(); }
143
144 void set(short Coefficient, Value *V) {
145 Coeff.set(Coefficient);
146 Val = V;
147 }
148 void set(const APFloat &Coefficient, Value *V) {
149 Coeff.set(Coefficient);
150 Val = V;
151 }
152 void set(const ConstantFP *Coefficient, Value *V) {
153 Coeff.set(Coefficient->getValueAPF());
154 Val = V;
155 }
156
157 void negate() { Coeff.negate(); }
158
159 /// Drill down the U-D chain one step to find the definition of V, and
160 /// try to break the definition into one or two addends.
161 static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
162
163 /// Similar to FAddend::drillDownOneStep() except that the value being
164 /// splitted is the addend itself.
165 unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
166
167 private:
168 void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
169
170 // This addend has the value of "Coeff * Val".
171 Value *Val = nullptr;
172 FAddendCoef Coeff;
173 };
174
175 /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
176 /// with its neighboring at most two instructions.
177 ///
178 class FAddCombine {
179 public:
180 FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}
181
182 Value *simplify(Instruction *FAdd);
183
184 private:
185 using AddendVect = SmallVector<const FAddend *, 4>;
186
187 Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
188
189 /// Convert given addend to a Value
190 Value *createAddendVal(const FAddend &A, bool& NeedNeg);
191
192 /// Return the number of instructions needed to emit the N-ary addition.
193 unsigned calcInstrNumber(const AddendVect& Vect);
194
195 Value *createFSub(Value *Opnd0, Value *Opnd1);
196 Value *createFAdd(Value *Opnd0, Value *Opnd1);
197 Value *createFMul(Value *Opnd0, Value *Opnd1);
198 Value *createFNeg(Value *V);
199 Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
200 void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
201
202 // Debugging stuff are clustered here.
203 #ifndef NDEBUG
204 unsigned CreateInstrNum;
205 void initCreateInstNum() { CreateInstrNum = 0; }
206 void incCreateInstNum() { CreateInstrNum++; }
207 #else
208 void initCreateInstNum() {}
209 void incCreateInstNum() {}
210 #endif
211
212 InstCombiner::BuilderTy &Builder;
213 Instruction *Instr = nullptr;
214 };
215
216} // end anonymous namespace
217
218//===----------------------------------------------------------------------===//
219//
220// Implementation of
221// {FAddendCoef, FAddend, FAddition, FAddCombine}.
222//
223//===----------------------------------------------------------------------===//
224FAddendCoef::~FAddendCoef() {
225 if (BufHasFpVal)
226 getFpValPtr()->~APFloat();
227}
228
229void FAddendCoef::set(const APFloat& C) {
230 APFloat *P = getFpValPtr();
231
232 if (isInt()) {
233 // As the buffer is meanless byte stream, we cannot call
234 // APFloat::operator=().
235 new(P) APFloat(C);
236 } else
237 *P = C;
238
239 IsFp = BufHasFpVal = true;
240}
241
242void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
243 if (!isInt())
244 return;
245
246 APFloat *P = getFpValPtr();
247 if (IntVal > 0)
248 new(P) APFloat(Sem, IntVal);
249 else {
250 new(P) APFloat(Sem, 0 - IntVal);
251 P->changeSign();
252 }
253 IsFp = BufHasFpVal = true;
254}
255
256APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
257 if (Val >= 0)
258 return APFloat(Sem, Val);
259
260 APFloat T(Sem, 0 - Val);
261 T.changeSign();
262
263 return T;
264}
265
266void FAddendCoef::operator=(const FAddendCoef &That) {
267 if (That.isInt())
268 set(That.IntVal);
269 else
270 set(That.getFpVal());
271}
272
273void FAddendCoef::operator+=(const FAddendCoef &That) {
274 RoundingMode RndMode = RoundingMode::NearestTiesToEven;
275 if (isInt() == That.isInt()) {
276 if (isInt())
277 IntVal += That.IntVal;
278 else
279 getFpVal().add(That.getFpVal(), RndMode);
280 return;
281 }
282
283 if (isInt()) {
284 const APFloat &T = That.getFpVal();
285 convertToFpType(T.getSemantics());
286 getFpVal().add(T, RndMode);
287 return;
288 }
289
290 APFloat &T = getFpVal();
291 T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292}
293
294void FAddendCoef::operator*=(const FAddendCoef &That) {
295 if (That.isOne())
296 return;
297
298 if (That.isMinusOne()) {
299 negate();
300 return;
301 }
302
303 if (isInt() && That.isInt()) {
304 int Res = IntVal * (int)That.IntVal;
305 assert(!insaneIntVal(Res) && "Insane int value")(static_cast <bool> (!insaneIntVal(Res) && "Insane int value"
) ? void (0) : __assert_fail ("!insaneIntVal(Res) && \"Insane int value\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 305
, __extension__ __PRETTY_FUNCTION__))
;
306 IntVal = Res;
307 return;
308 }
309
310 const fltSemantics &Semantic =
311 isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
312
313 if (isInt())
314 convertToFpType(Semantic);
315 APFloat &F0 = getFpVal();
316
317 if (That.isInt())
318 F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
319 APFloat::rmNearestTiesToEven);
320 else
321 F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322}
323
324void FAddendCoef::negate() {
325 if (isInt())
326 IntVal = 0 - IntVal;
327 else
328 getFpVal().changeSign();
329}
330
331Value *FAddendCoef::getValue(Type *Ty) const {
332 return isInt() ?
333 ConstantFP::get(Ty, float(IntVal)) :
334 ConstantFP::get(Ty->getContext(), getFpVal());
335}
336
337// The definition of <Val> Addends
338// =========================================
339// A + B <1, A>, <1,B>
340// A - B <1, A>, <1,B>
341// 0 - B <-1, B>
342// C * A, <C, A>
343// A + C <1, A> <C, NULL>
344// 0 +/- 0 <0, NULL> (corner case)
345//
346// Legend: A and B are not constant, C is constant
347unsigned FAddend::drillValueDownOneStep
348 (Value *Val, FAddend &Addend0, FAddend &Addend1) {
349 Instruction *I = nullptr;
350 if (!Val || !(I = dyn_cast<Instruction>(Val)))
351 return 0;
352
353 unsigned Opcode = I->getOpcode();
354
355 if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
356 ConstantFP *C0, *C1;
357 Value *Opnd0 = I->getOperand(0);
358 Value *Opnd1 = I->getOperand(1);
359 if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
360 Opnd0 = nullptr;
361
362 if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
363 Opnd1 = nullptr;
364
365 if (Opnd0) {
366 if (!C0)
367 Addend0.set(1, Opnd0);
368 else
369 Addend0.set(C0, nullptr);
370 }
371
372 if (Opnd1) {
373 FAddend &Addend = Opnd0 ? Addend1 : Addend0;
374 if (!C1)
375 Addend.set(1, Opnd1);
376 else
377 Addend.set(C1, nullptr);
378 if (Opcode == Instruction::FSub)
379 Addend.negate();
380 }
381
382 if (Opnd0 || Opnd1)
383 return Opnd0 && Opnd1 ? 2 : 1;
384
385 // Both operands are zero. Weird!
386 Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
387 return 1;
388 }
389
390 if (I->getOpcode() == Instruction::FMul) {
391 Value *V0 = I->getOperand(0);
392 Value *V1 = I->getOperand(1);
393 if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
394 Addend0.set(C, V1);
395 return 1;
396 }
397
398 if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
399 Addend0.set(C, V0);
400 return 1;
401 }
402 }
403
404 return 0;
405}
406
407// Try to break *this* addend into two addends. e.g. Suppose this addend is
408// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409// i.e. <2.3, X> and <2.3, Y>.
410unsigned FAddend::drillAddendDownOneStep
411 (FAddend &Addend0, FAddend &Addend1) const {
412 if (isConstant())
413 return 0;
414
415 unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
416 if (!BreakNum || Coeff.isOne())
417 return BreakNum;
418
419 Addend0.Scale(Coeff);
420
421 if (BreakNum == 2)
422 Addend1.Scale(Coeff);
423
424 return BreakNum;
425}
426
427Value *FAddCombine::simplify(Instruction *I) {
428 assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&(static_cast <bool> (I->hasAllowReassoc() &&
I->hasNoSignedZeros() && "Expected 'reassoc'+'nsz' instruction"
) ? void (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 429
, __extension__ __PRETTY_FUNCTION__))
429 "Expected 'reassoc'+'nsz' instruction")(static_cast <bool> (I->hasAllowReassoc() &&
I->hasNoSignedZeros() && "Expected 'reassoc'+'nsz' instruction"
) ? void (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 429
, __extension__ __PRETTY_FUNCTION__))
;
430
431 // Currently we are not able to handle vector type.
432 if (I->getType()->isVectorTy())
433 return nullptr;
434
435 assert((I->getOpcode() == Instruction::FAdd ||(static_cast <bool> ((I->getOpcode() == Instruction::
FAdd || I->getOpcode() == Instruction::FSub) && "Expect add/sub"
) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 436
, __extension__ __PRETTY_FUNCTION__))
436 I->getOpcode() == Instruction::FSub) && "Expect add/sub")(static_cast <bool> ((I->getOpcode() == Instruction::
FAdd || I->getOpcode() == Instruction::FSub) && "Expect add/sub"
) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 436
, __extension__ __PRETTY_FUNCTION__))
;
437
438 // Save the instruction before calling other member-functions.
439 Instr = I;
440
441 FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
442
443 unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
444
445 // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
446 unsigned Opnd0_ExpNum = 0;
447 unsigned Opnd1_ExpNum = 0;
448
449 if (!Opnd0.isConstant())
450 Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
451
452 // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
453 if (OpndNum == 2 && !Opnd1.isConstant())
454 Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
455
456 // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
457 if (Opnd0_ExpNum && Opnd1_ExpNum) {
458 AddendVect AllOpnds;
459 AllOpnds.push_back(&Opnd0_0);
460 AllOpnds.push_back(&Opnd1_0);
461 if (Opnd0_ExpNum == 2)
462 AllOpnds.push_back(&Opnd0_1);
463 if (Opnd1_ExpNum == 2)
464 AllOpnds.push_back(&Opnd1_1);
465
466 // Compute instruction quota. We should save at least one instruction.
467 unsigned InstQuota = 0;
468
469 Value *V0 = I->getOperand(0);
470 Value *V1 = I->getOperand(1);
471 InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
472 (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
473
474 if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
475 return R;
476 }
477
478 if (OpndNum != 2) {
479 // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
480 // splitted into two addends, say "V = X - Y", the instruction would have
481 // been optimized into "I = Y - X" in the previous steps.
482 //
483 const FAddendCoef &CE = Opnd0.getCoef();
484 return CE.isOne() ? Opnd0.getSymVal() : nullptr;
485 }
486
487 // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
488 if (Opnd1_ExpNum) {
489 AddendVect AllOpnds;
490 AllOpnds.push_back(&Opnd0);
491 AllOpnds.push_back(&Opnd1_0);
492 if (Opnd1_ExpNum == 2)
493 AllOpnds.push_back(&Opnd1_1);
494
495 if (Value *R = simplifyFAdd(AllOpnds, 1))
496 return R;
497 }
498
499 // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
500 if (Opnd0_ExpNum) {
501 AddendVect AllOpnds;
502 AllOpnds.push_back(&Opnd1);
503 AllOpnds.push_back(&Opnd0_0);
504 if (Opnd0_ExpNum == 2)
505 AllOpnds.push_back(&Opnd0_1);
506
507 if (Value *R = simplifyFAdd(AllOpnds, 1))
508 return R;
509 }
510
511 return nullptr;
512}
513
514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
515 unsigned AddendNum = Addends.size();
516 assert(AddendNum <= 4 && "Too many addends")(static_cast <bool> (AddendNum <= 4 && "Too many addends"
) ? void (0) : __assert_fail ("AddendNum <= 4 && \"Too many addends\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 516
, __extension__ __PRETTY_FUNCTION__))
;
517
518 // For saving intermediate results;
519 unsigned NextTmpIdx = 0;
520 FAddend TmpResult[3];
521
522 // Simplified addends are placed <SimpVect>.
523 AddendVect SimpVect;
524
525 // The outer loop works on one symbolic-value at a time. Suppose the input
526 // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
527 // The symbolic-values will be processed in this order: x, y, z.
528 for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
529
530 const FAddend *ThisAddend = Addends[SymIdx];
531 if (!ThisAddend) {
532 // This addend was processed before.
533 continue;
534 }
535
536 Value *Val = ThisAddend->getSymVal();
537
538 // If the resulting expr has constant-addend, this constant-addend is
539 // desirable to reside at the top of the resulting expression tree. Placing
540 // constant close to super-expr(s) will potentially reveal some
541 // optimization opportunities in super-expr(s). Here we do not implement
542 // this logic intentionally and rely on SimplifyAssociativeOrCommutative
543 // call later.
544
545 unsigned StartIdx = SimpVect.size();
546 SimpVect.push_back(ThisAddend);
547
548 // The inner loop collects addends sharing same symbolic-value, and these
549 // addends will be later on folded into a single addend. Following above
550 // example, if the symbolic value "y" is being processed, the inner loop
551 // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
552 // be later on folded into "<b1+b2, y>".
553 for (unsigned SameSymIdx = SymIdx + 1;
554 SameSymIdx < AddendNum; SameSymIdx++) {
555 const FAddend *T = Addends[SameSymIdx];
556 if (T && T->getSymVal() == Val) {
557 // Set null such that next iteration of the outer loop will not process
558 // this addend again.
559 Addends[SameSymIdx] = nullptr;
560 SimpVect.push_back(T);
561 }
562 }
563
564 // If multiple addends share same symbolic value, fold them together.
565 if (StartIdx + 1 != SimpVect.size()) {
566 FAddend &R = TmpResult[NextTmpIdx ++];
567 R = *SimpVect[StartIdx];
568 for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
569 R += *SimpVect[Idx];
570
571 // Pop all addends being folded and push the resulting folded addend.
572 SimpVect.resize(StartIdx);
573 if (!R.isZero()) {
574 SimpVect.push_back(&R);
575 }
576 }
577 }
578
579 assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&(static_cast <bool> ((NextTmpIdx <= array_lengthof(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 580
, __extension__ __PRETTY_FUNCTION__))
580 "out-of-bound access")(static_cast <bool> ((NextTmpIdx <= array_lengthof(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 580
, __extension__ __PRETTY_FUNCTION__))
;
581
582 Value *Result;
583 if (!SimpVect.empty())
584 Result = createNaryFAdd(SimpVect, InstrQuota);
585 else {
586 // The addition is folded to 0.0.
587 Result = ConstantFP::get(Instr->getType(), 0.0);
588 }
589
590 return Result;
591}
592
593Value *FAddCombine::createNaryFAdd
594 (const AddendVect &Opnds, unsigned InstrQuota) {
595 assert(!Opnds.empty() && "Expect at least one addend")(static_cast <bool> (!Opnds.empty() && "Expect at least one addend"
) ? void (0) : __assert_fail ("!Opnds.empty() && \"Expect at least one addend\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 595
, __extension__ __PRETTY_FUNCTION__))
;
596
597 // Step 1: Check if the # of instructions needed exceeds the quota.
598
599 unsigned InstrNeeded = calcInstrNumber(Opnds);
600 if (InstrNeeded > InstrQuota)
601 return nullptr;
602
603 initCreateInstNum();
604
605 // step 2: Emit the N-ary addition.
606 // Note that at most three instructions are involved in Fadd-InstCombine: the
607 // addition in question, and at most two neighboring instructions.
608 // The resulting optimized addition should have at least one less instruction
609 // than the original addition expression tree. This implies that the resulting
610 // N-ary addition has at most two instructions, and we don't need to worry
611 // about tree-height when constructing the N-ary addition.
612
613 Value *LastVal = nullptr;
614 bool LastValNeedNeg = false;
615
616 // Iterate the addends, creating fadd/fsub using adjacent two addends.
617 for (const FAddend *Opnd : Opnds) {
618 bool NeedNeg;
619 Value *V = createAddendVal(*Opnd, NeedNeg);
620 if (!LastVal) {
621 LastVal = V;
622 LastValNeedNeg = NeedNeg;
623 continue;
624 }
625
626 if (LastValNeedNeg == NeedNeg) {
627 LastVal = createFAdd(LastVal, V);
628 continue;
629 }
630
631 if (LastValNeedNeg)
632 LastVal = createFSub(V, LastVal);
633 else
634 LastVal = createFSub(LastVal, V);
635
636 LastValNeedNeg = false;
637 }
638
639 if (LastValNeedNeg) {
640 LastVal = createFNeg(LastVal);
641 }
642
643#ifndef NDEBUG
644 assert(CreateInstrNum == InstrNeeded &&(static_cast <bool> (CreateInstrNum == InstrNeeded &&
"Inconsistent in instruction numbers") ? void (0) : __assert_fail
("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 645
, __extension__ __PRETTY_FUNCTION__))
645 "Inconsistent in instruction numbers")(static_cast <bool> (CreateInstrNum == InstrNeeded &&
"Inconsistent in instruction numbers") ? void (0) : __assert_fail
("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 645
, __extension__ __PRETTY_FUNCTION__))
;
646#endif
647
648 return LastVal;
649}
650
651Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
652 Value *V = Builder.CreateFSub(Opnd0, Opnd1);
653 if (Instruction *I = dyn_cast<Instruction>(V))
654 createInstPostProc(I);
655 return V;
656}
657
658Value *FAddCombine::createFNeg(Value *V) {
659 Value *NewV = Builder.CreateFNeg(V);
660 if (Instruction *I = dyn_cast<Instruction>(NewV))
661 createInstPostProc(I, true); // fneg's don't receive instruction numbers.
662 return NewV;
663}
664
665Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
666 Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
667 if (Instruction *I = dyn_cast<Instruction>(V))
668 createInstPostProc(I);
669 return V;
670}
671
672Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
673 Value *V = Builder.CreateFMul(Opnd0, Opnd1);
674 if (Instruction *I = dyn_cast<Instruction>(V))
675 createInstPostProc(I);
676 return V;
677}
678
679void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
680 NewInstr->setDebugLoc(Instr->getDebugLoc());
681
682 // Keep track of the number of instruction created.
683 if (!NoNumber)
684 incCreateInstNum();
685
686 // Propagate fast-math flags
687 NewInstr->setFastMathFlags(Instr->getFastMathFlags());
688}
689
690// Return the number of instruction needed to emit the N-ary addition.
691// NOTE: Keep this function in sync with createAddendVal().
692unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
693 unsigned OpndNum = Opnds.size();
694 unsigned InstrNeeded = OpndNum - 1;
695
696 // The number of addends in the form of "(-1)*x".
697 unsigned NegOpndNum = 0;
698
699 // Adjust the number of instructions needed to emit the N-ary add.
700 for (const FAddend *Opnd : Opnds) {
701 if (Opnd->isConstant())
702 continue;
703
704 // The constant check above is really for a few special constant
705 // coefficients.
706 if (isa<UndefValue>(Opnd->getSymVal()))
707 continue;
708
709 const FAddendCoef &CE = Opnd->getCoef();
710 if (CE.isMinusOne() || CE.isMinusTwo())
711 NegOpndNum++;
712
713 // Let the addend be "c * x". If "c == +/-1", the value of the addend
714 // is immediately available; otherwise, it needs exactly one instruction
715 // to evaluate the value.
716 if (!CE.isMinusOne() && !CE.isOne())
717 InstrNeeded++;
718 }
719 return InstrNeeded;
720}
721
722// Input Addend Value NeedNeg(output)
723// ================================================================
724// Constant C C false
725// <+/-1, V> V coefficient is -1
726// <2/-2, V> "fadd V, V" coefficient is -2
727// <C, V> "fmul V, C" false
728//
729// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
730Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
731 const FAddendCoef &Coeff = Opnd.getCoef();
732
733 if (Opnd.isConstant()) {
734 NeedNeg = false;
735 return Coeff.getValue(Instr->getType());
736 }
737
738 Value *OpndVal = Opnd.getSymVal();
739
740 if (Coeff.isMinusOne() || Coeff.isOne()) {
741 NeedNeg = Coeff.isMinusOne();
742 return OpndVal;
743 }
744
745 if (Coeff.isTwo() || Coeff.isMinusTwo()) {
746 NeedNeg = Coeff.isMinusTwo();
747 return createFAdd(OpndVal, OpndVal);
748 }
749
750 NeedNeg = false;
751 return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
752}
753
754// Checks if any operand is negative and we can convert add to sub.
755// This function checks for following negative patterns
756// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
757// ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
758// XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
759static Value *checkForNegativeOperand(BinaryOperator &I,
760 InstCombiner::BuilderTy &Builder) {
761 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
762
763 // This function creates 2 instructions to replace ADD, we need at least one
764 // of LHS or RHS to have one use to ensure benefit in transform.
765 if (!LHS->hasOneUse() && !RHS->hasOneUse())
766 return nullptr;
767
768 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
769 const APInt *C1 = nullptr, *C2 = nullptr;
770
771 // if ONE is on other side, swap
772 if (match(RHS, m_Add(m_Value(X), m_One())))
773 std::swap(LHS, RHS);
774
775 if (match(LHS, m_Add(m_Value(X), m_One()))) {
776 // if XOR on other side, swap
777 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
778 std::swap(X, RHS);
779
780 if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
781 // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
782 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
783 if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
784 Value *NewAnd = Builder.CreateAnd(Z, *C1);
785 return Builder.CreateSub(RHS, NewAnd, "sub");
786 } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
787 // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
788 // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
789 Value *NewOr = Builder.CreateOr(Z, ~(*C1));
790 return Builder.CreateSub(RHS, NewOr, "sub");
791 }
792 }
793 }
794
795 // Restore LHS and RHS
796 LHS = I.getOperand(0);
797 RHS = I.getOperand(1);
798
799 // if XOR is on other side, swap
800 if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
801 std::swap(LHS, RHS);
802
803 // C2 is ODD
804 // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
805 // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
806 if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
807 if (C1->countTrailingZeros() == 0)
808 if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
809 Value *NewOr = Builder.CreateOr(Z, ~(*C2));
810 return Builder.CreateSub(RHS, NewOr, "sub");
811 }
812 return nullptr;
813}
814
815/// Wrapping flags may allow combining constants separated by an extend.
816static Instruction *foldNoWrapAdd(BinaryOperator &Add,
817 InstCombiner::BuilderTy &Builder) {
818 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
819 Type *Ty = Add.getType();
820 Constant *Op1C;
821 if (!match(Op1, m_Constant(Op1C)))
822 return nullptr;
823
824 // Try this match first because it results in an add in the narrow type.
825 // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
826 Value *X;
827 const APInt *C1, *C2;
828 if (match(Op1, m_APInt(C1)) &&
829 match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
830 C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
831 Constant *NewC =
832 ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
833 return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
834 }
835
836 // More general combining of constants in the wide type.
837 // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
838 Constant *NarrowC;
839 if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
840 Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
841 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
842 Value *WideX = Builder.CreateSExt(X, Ty);
843 return BinaryOperator::CreateAdd(WideX, NewC);
844 }
845 // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
846 if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
847 Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
848 Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
849 Value *WideX = Builder.CreateZExt(X, Ty);
850 return BinaryOperator::CreateAdd(WideX, NewC);
851 }
852
853 return nullptr;
854}
855
856Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
857 Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
858 Constant *Op1C;
859 if (!match(Op1, m_ImmConstant(Op1C)))
860 return nullptr;
861
862 if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
863 return NV;
864
865 Value *X;
866 Constant *Op00C;
867
868 // add (sub C1, X), C2 --> sub (add C1, C2), X
869 if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
870 return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);
871
872 Value *Y;
873
874 // add (sub X, Y), -1 --> add (not Y), X
875 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
876 match(Op1, m_AllOnes()))
877 return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);
878
879 // zext(bool) + C -> bool ? C + 1 : C
880 if (match(Op0, m_ZExt(m_Value(X))) &&
881 X->getType()->getScalarSizeInBits() == 1)
882 return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
883 // sext(bool) + C -> bool ? C - 1 : C
884 if (match(Op0, m_SExt(m_Value(X))) &&
885 X->getType()->getScalarSizeInBits() == 1)
886 return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);
887
888 // ~X + C --> (C-1) - X
889 if (match(Op0, m_Not(m_Value(X))))
890 return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X);
891
892 const APInt *C;
893 if (!match(Op1, m_APInt(C)))
894 return nullptr;
895
896 // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
897 Constant *Op01C;
898 if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
899 haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
900 return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));
901
902 // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
903 const APInt *C2;
904 if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
905 return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));
906
907 if (C->isSignMask()) {
908 // If wrapping is not allowed, then the addition must set the sign bit:
909 // X + (signmask) --> X | signmask
910 if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
911 return BinaryOperator::CreateOr(Op0, Op1);
912
913 // If wrapping is allowed, then the addition flips the sign bit of LHS:
914 // X + (signmask) --> X ^ signmask
915 return BinaryOperator::CreateXor(Op0, Op1);
916 }
917
918 // Is this add the last step in a convoluted sext?
919 // add(zext(xor i16 X, -32768), -32768) --> sext X
920 Type *Ty = Add.getType();
921 if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
922 C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
923 return CastInst::Create(Instruction::SExt, X, Ty);
924
925 if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
926 // (X ^ signmask) + C --> (X + (signmask ^ C))
927 if (C2->isSignMask())
928 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));
929
930 // If X has no high-bits set above an xor mask:
931 // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
932 if (C2->isMask()) {
933 KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
934 if ((*C2 | LHSKnown.Zero).isAllOnes())
935 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
936 }
937
938 // Look for a math+logic pattern that corresponds to sext-in-register of a
939 // value with cleared high bits. Convert that into a pair of shifts:
940 // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
941 // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
942 if (Op0->hasOneUse() && *C2 == -(*C)) {
943 unsigned BitWidth = Ty->getScalarSizeInBits();
944 unsigned ShAmt = 0;
945 if (C->isPowerOf2())
946 ShAmt = BitWidth - C->logBase2() - 1;
947 else if (C2->isPowerOf2())
948 ShAmt = BitWidth - C2->logBase2() - 1;
949 if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
950 0, &Add)) {
951 Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
952 Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
953 return BinaryOperator::CreateAShr(NewShl, ShAmtC);
954 }
955 }
956 }
957
958 if (C->isOne() && Op0->hasOneUse()) {
959 // add (sext i1 X), 1 --> zext (not X)
960 // TODO: The smallest IR representation is (select X, 0, 1), and that would
961 // not require the one-use check. But we need to remove a transform in
962 // visitSelect and make sure that IR value tracking for select is equal or
963 // better than for these ops.
964 if (match(Op0, m_SExt(m_Value(X))) &&
965 X->getType()->getScalarSizeInBits() == 1)
966 return new ZExtInst(Builder.CreateNot(X), Ty);
967
968 // Shifts and add used to flip and mask off the low bit:
969 // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
970 const APInt *C3;
971 if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
972 C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
973 Value *NotX = Builder.CreateNot(X);
974 return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
975 }
976 }
977
978 // If all bits affected by the add are included in a high-bit-mask, do the
979 // add before the mask op:
980 // (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00
981 if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) &&
982 C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) {
983 Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C));
984 return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2));
985 }
986
987 return nullptr;
988}
989
990// Matches multiplication expression Op * C where C is a constant. Returns the
991// constant value in C and the other operand in Op. Returns true if such a
992// match is found.
993static bool MatchMul(Value *E, Value *&Op, APInt &C) {
994 const APInt *AI;
995 if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
996 C = *AI;
997 return true;
998 }
999 if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
1000 C = APInt(AI->getBitWidth(), 1);
1001 C <<= *AI;
1002 return true;
1003 }
1004 return false;
1005}
1006
1007// Matches remainder expression Op % C where C is a constant. Returns the
1008// constant value in C and the other operand in Op. Returns the signedness of
1009// the remainder operation in IsSigned. Returns true if such a match is
1010// found.
1011static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
1012 const APInt *AI;
1013 IsSigned = false;
1014 if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
1015 IsSigned = true;
1016 C = *AI;
1017 return true;
1018 }
1019 if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
1020 C = *AI;
1021 return true;
1022 }
1023 if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
1024 C = *AI + 1;
1025 return true;
1026 }
1027 return false;
1028}
1029
1030// Matches division expression Op / C with the given signedness as indicated
1031// by IsSigned, where C is a constant. Returns the constant value in C and the
1032// other operand in Op. Returns true if such a match is found.
1033static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
1034 const APInt *AI;
1035 if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
1036 C = *AI;
1037 return true;
1038 }
1039 if (!IsSigned) {
1040 if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
1041 C = *AI;
1042 return true;
1043 }
1044 if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
1045 C = APInt(AI->getBitWidth(), 1);
1046 C <<= *AI;
1047 return true;
1048 }
1049 }
1050 return false;
1051}
1052
1053// Returns whether C0 * C1 with the given signedness overflows.
1054static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
1055 bool overflow;
1056 if (IsSigned)
1057 (void)C0.smul_ov(C1, overflow);
1058 else
1059 (void)C0.umul_ov(C1, overflow);
1060 return overflow;
1061}
1062
1063// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1064// does not overflow.
1065Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
1066 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1067 Value *X, *MulOpV;
1068 APInt C0, MulOpC;
1069 bool IsSigned;
1070 // Match I = X % C0 + MulOpV * C0
1071 if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
1072 (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
1073 C0 == MulOpC) {
1074 Value *RemOpV;
1075 APInt C1;
1076 bool Rem2IsSigned;
1077 // Match MulOpC = RemOpV % C1
1078 if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
1079 IsSigned == Rem2IsSigned) {
1080 Value *DivOpV;
1081 APInt DivOpC;
1082 // Match RemOpV = X / C0
1083 if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
1084 C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
1085 Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
1086 return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
1087 : Builder.CreateURem(X, NewDivisor, "urem");
1088 }
1089 }
1090 }
1091
1092 return nullptr;
1093}
1094
1095/// Fold
1096/// (1 << NBits) - 1
1097/// Into:
1098/// ~(-(1 << NBits))
1099/// Because a 'not' is better for bit-tracking analysis and other transforms
1100/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1101static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
1102 InstCombiner::BuilderTy &Builder) {
1103 Value *NBits;
1104 if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
1105 return nullptr;
1106
1107 Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
1108 Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
1109 // Be wary of constant folding.
1110 if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
1111 // Always NSW. But NUW propagates from `add`.
1112 BOp->setHasNoSignedWrap();
1113 BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1114 }
1115
1116 return BinaryOperator::CreateNot(NotMask, I.getName());
1117}
1118
1119static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
1120 assert(I.getOpcode() == Instruction::Add && "Expecting add instruction")(static_cast <bool> (I.getOpcode() == Instruction::Add &&
"Expecting add instruction") ? void (0) : __assert_fail ("I.getOpcode() == Instruction::Add && \"Expecting add instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1120
, __extension__ __PRETTY_FUNCTION__))
;
1121 Type *Ty = I.getType();
1122 auto getUAddSat = [&]() {
1123 return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
1124 };
1125
1126 // add (umin X, ~Y), Y --> uaddsat X, Y
1127 Value *X, *Y;
1128 if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
1129 m_Deferred(Y))))
1130 return CallInst::Create(getUAddSat(), { X, Y });
1131
1132 // add (umin X, ~C), C --> uaddsat X, C
1133 const APInt *C, *NotC;
1134 if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
1135 *C == ~*NotC)
1136 return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });
1137
1138 return nullptr;
1139}
1140
1141Instruction *InstCombinerImpl::
1142 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
1143 BinaryOperator &I) {
1144 assert((I.getOpcode() == Instruction::Add ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1147
, __extension__ __PRETTY_FUNCTION__))
1
Assuming the condition is false
2
Assuming the condition is false
3
Assuming the condition is true
4
'?' condition is true
1145 I.getOpcode() == Instruction::Or ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1147
, __extension__ __PRETTY_FUNCTION__))
1146 I.getOpcode() == Instruction::Sub) &&(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1147
, __extension__ __PRETTY_FUNCTION__))
1147 "Expecting add/or/sub instruction")(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1147
, __extension__ __PRETTY_FUNCTION__))
;
1148
1149 // We have a subtraction/addition between a (potentially truncated) *logical*
1150 // right-shift of X and a "select".
1151 Value *X, *Select;
1152 Instruction *LowBitsToSkip, *Extract;
1153 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>>'
37
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>>'
1154 m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
1155 m_Instruction(Extract))),
1156 m_Value(Select))))
5
Calling 'm_Value'
9
Returning from 'm_Value'
1157 return nullptr;
1158
1159 // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
1160 if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
38
Assuming pointer value is null
39
Taking false branch
1161 return nullptr;
1162
1163 Type *XTy = X->getType();
1164 bool HadTrunc = I.getType() != XTy;
40
Assuming the condition is false
1165
1166 // If there was a truncation of extracted value, then we'll need to produce
1167 // one extra instruction, so we need to ensure one instruction will go away.
1168 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
1169 return nullptr;
1170
1171 // Extraction should extract high NBits bits, with shift amount calculated as:
1172 // low bits to skip = shift bitwidth - high bits to extract
1173 // The shift amount itself may be extended, and we need to look past zero-ext
1174 // when matching NBits, that will matter for matching later.
1175 Constant *C;
1176 Value *NBits;
1177 if (!match(
43
Taking false branch
1178 LowBitsToSkip,
1179 m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
1180 !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
42
Assuming the condition is false
1181 APInt(C->getType()->getScalarSizeInBits(),
1182 X->getType()->getScalarSizeInBits()))))
1183 return nullptr;
1184
1185 // Sign-extending value can be zero-extended if we `sub`tract it,
1186 // or sign-extended otherwise.
1187 auto SkipExtInMagic = [&I](Value *&V) {
1188 if (I.getOpcode() == Instruction::Sub)
45
Taking true branch
1189 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>>>'
1190 else
1191 match(V, m_SExtOrSelf(m_Value(V)));
1192 };
61
Returning without writing to 'V'
1193
1194 // Now, finally validate the sign-extending magic.
1195 // `select` itself may be appropriately extended, look past that.
1196 SkipExtInMagic(Select);
44
Calling 'operator()'
62
Returning from 'operator()'
1197
1198 ICmpInst::Predicate Pred;
1199 const APInt *Thr;
1200 Value *SignExtendingValue, *Zero;
1201 bool ShouldSignext;
1202 // It must be a select between two values we will later establish to be a
1203 // sign-extending value and a zero constant. The condition guarding the
1204 // sign-extension must be based on a sign bit of the same X we had in `lshr`.
1205 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>>'
1206 m_Value(SignExtendingValue), m_Value(Zero))) ||
1207 !isSignBitCheck(Pred, *Thr, ShouldSignext))
1208 return nullptr;
1209
1210 // icmp-select pair is commutative.
1211 if (!ShouldSignext)
1212 std::swap(SignExtendingValue, Zero);
1213
1214 // If we should not perform sign-extension then we must add/or/subtract zero.
1215 if (!match(Zero, m_Zero()))
1216 return nullptr;
1217 // Otherwise, it should be some constant, left-shifted by the same NBits we
1218 // had in `lshr`. Said left-shift can also be appropriately extended.
1219 // Again, we must look past zero-ext when looking for NBits.
1220 SkipExtInMagic(SignExtendingValue);
1221 Constant *SignExtendingValueBaseConstant;
1222 if (!match(SignExtendingValue,
1223 m_Shl(m_Constant(SignExtendingValueBaseConstant),
1224 m_ZExtOrSelf(m_Specific(NBits)))))
1225 return nullptr;
1226 // If we `sub`, then the constant should be one, else it should be all-ones.
1227 if (I.getOpcode() == Instruction::Sub
1228 ? !match(SignExtendingValueBaseConstant, m_One())
1229 : !match(SignExtendingValueBaseConstant, m_AllOnes()))
1230 return nullptr;
1231
1232 auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
1233 Extract->getName() + ".sext");
1234 NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
1235 if (!HadTrunc)
1236 return NewAShr;
1237
1238 Builder.Insert(NewAShr);
1239 return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1240}
1241
1242/// This is a specialization of a more general transform from
1243/// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
1244/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1245static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
1246 InstCombiner::BuilderTy &Builder) {
1247 // TODO: Also handle mul by doubling the shift amount?
1248 assert((I.getOpcode() == Instruction::Add ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1250
, __extension__ __PRETTY_FUNCTION__))
1249 I.getOpcode() == Instruction::Sub) &&(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1250
, __extension__ __PRETTY_FUNCTION__))
1250 "Expected add/sub")(static_cast <bool> ((I.getOpcode() == Instruction::Add
|| I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1250
, __extension__ __PRETTY_FUNCTION__))
;
1251 auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
1252 auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
1253 if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
1254 return nullptr;
1255
1256 Value *X, *Y, *ShAmt;
1257 if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
1258 !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
1259 return nullptr;
1260
1261 // No-wrap propagates only when all ops have no-wrap.
1262 bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
1263 Op1->hasNoSignedWrap();
1264 bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
1265 Op1->hasNoUnsignedWrap();
1266
1267 // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
1268 Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
1269 if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
1270 NewI->setHasNoSignedWrap(HasNSW);
1271 NewI->setHasNoUnsignedWrap(HasNUW);
1272 }
1273 auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
1274 NewShl->setHasNoSignedWrap(HasNSW);
1275 NewShl->setHasNoUnsignedWrap(HasNUW);
1276 return NewShl;
1277}
1278
1279Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
1280 if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
1281 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1282 SQ.getWithInstruction(&I)))
1283 return replaceInstUsesWith(I, V);
1284
1285 if (SimplifyAssociativeOrCommutative(I))
1286 return &I;
1287
1288 if (Instruction *X = foldVectorBinop(I))
1289 return X;
1290
1291 // (A*B)+(A*C) -> A*(B+C) etc
1292 if (Value *V = SimplifyUsingDistributiveLaws(I))
1293 return replaceInstUsesWith(I, V);
1294
1295 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1296 return R;
1297
1298 if (Instruction *X = foldAddWithConstant(I))
1299 return X;
1300
1301 if (Instruction *X = foldNoWrapAdd(I, Builder))
1302 return X;
1303
1304 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1305 Type *Ty = I.getType();
1306 if (Ty->isIntOrIntVectorTy(1))
1307 return BinaryOperator::CreateXor(LHS, RHS);
1308
1309 // X + X --> X << 1
1310 if (LHS == RHS) {
1311 auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
1312 Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
1313 Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1314 return Shl;
1315 }
1316
1317 Value *A, *B;
1318 if (match(LHS, m_Neg(m_Value(A)))) {
1319 // -A + -B --> -(A + B)
1320 if (match(RHS, m_Neg(m_Value(B))))
1321 return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
1322
1323 // -A + B --> B - A
1324 return BinaryOperator::CreateSub(RHS, A);
1325 }
1326
1327 // A + -B --> A - B
1328 if (match(RHS, m_Neg(m_Value(B))))
1329 return BinaryOperator::CreateSub(LHS, B);
1330
1331 if (Value *V = checkForNegativeOperand(I, Builder))
1332 return replaceInstUsesWith(I, V);
1333
1334 // (A + 1) + ~B --> A - B
1335 // ~B + (A + 1) --> A - B
1336 // (~B + A) + 1 --> A - B
1337 // (A + ~B) + 1 --> A - B
1338 if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
1339 match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
1340 return BinaryOperator::CreateSub(A, B);
1341
1342 // (A + RHS) + RHS --> A + (RHS << 1)
1343 if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
1344 return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
1345
1346 // LHS + (A + LHS) --> A + (LHS << 1)
1347 if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
1348 return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
1349
1350 {
1351 // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
1352 Constant *C1, *C2;
1353 if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
1354 m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
1355 (LHS->hasOneUse() || RHS->hasOneUse())) {
1356 Value *Sub = Builder.CreateSub(A, B);
1357 return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
1358 }
1359 }
1360
1361 // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
1362 if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);
1363
1364 // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
1365 const APInt *C1, *C2;
1366 if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
1367 APInt one(C2->getBitWidth(), 1);
1368 APInt minusC1 = -(*C1);
1369 if (minusC1 == (one << *C2)) {
1370 Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
1371 return BinaryOperator::CreateSRem(RHS, NewRHS);
1372 }
1373 }
1374
1375 // A+B --> A|B iff A and B have no bits set in common.
1376 if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
1377 return BinaryOperator::CreateOr(LHS, RHS);
1378
1379 // add (select X 0 (sub n A)) A --> select X A n
1380 {
1381 SelectInst *SI = dyn_cast<SelectInst>(LHS);
1382 Value *A = RHS;
1383 if (!SI) {
1384 SI = dyn_cast<SelectInst>(RHS);
1385 A = LHS;
1386 }
1387 if (SI && SI->hasOneUse()) {
1388 Value *TV = SI->getTrueValue();
1389 Value *FV = SI->getFalseValue();
1390 Value *N;
1391
1392 // Can we fold the add into the argument of the select?
1393 // We check both true and false select arguments for a matching subtract.
1394 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
1395 // Fold the add into the true select value.
1396 return SelectInst::Create(SI->getCondition(), N, A);
1397
1398 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
1399 // Fold the add into the false select value.
1400 return SelectInst::Create(SI->getCondition(), A, N);
1401 }
1402 }
1403
1404 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1405 return Ext;
1406
1407 // (add (xor A, B) (and A, B)) --> (or A, B)
1408 // (add (and A, B) (xor A, B)) --> (or A, B)
1409 if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
1410 m_c_And(m_Deferred(A), m_Deferred(B)))))
1411 return BinaryOperator::CreateOr(A, B);
1412
1413 // (add (or A, B) (and A, B)) --> (add A, B)
1414 // (add (and A, B) (or A, B)) --> (add A, B)
1415 if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
1416 m_c_And(m_Deferred(A), m_Deferred(B))))) {
1417 // Replacing operands in-place to preserve nuw/nsw flags.
1418 replaceOperand(I, 0, A);
1419 replaceOperand(I, 1, B);
1420 return &I;
1421 }
1422
1423 // TODO(jingyue): Consider willNotOverflowSignedAdd and
1424 // willNotOverflowUnsignedAdd to reduce the number of invocations of
1425 // computeKnownBits.
1426 bool Changed = false;
1427 if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
1428 Changed = true;
1429 I.setHasNoSignedWrap(true);
1430 }
1431 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
1432 Changed = true;
1433 I.setHasNoUnsignedWrap(true);
1434 }
1435
1436 if (Instruction *V = canonicalizeLowbitMask(I, Builder))
1437 return V;
1438
1439 if (Instruction *V =
1440 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
1441 return V;
1442
1443 if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
1444 return SatAdd;
1445
1446 // usub.sat(A, B) + B => umax(A, B)
1447 if (match(&I, m_c_BinOp(
1448 m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
1449 m_Deferred(B)))) {
1450 return replaceInstUsesWith(I,
1451 Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
1452 }
1453
1454 // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
1455 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
1456 match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
1457 haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
1458 return replaceInstUsesWith(
1459 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
1460 {Builder.CreateOr(A, B)}));
1461
1462 return Changed ? &I : nullptr;
1463}
1464
1465/// Eliminate an op from a linear interpolation (lerp) pattern.
1466static Instruction *factorizeLerp(BinaryOperator &I,
1467 InstCombiner::BuilderTy &Builder) {
1468 Value *X, *Y, *Z;
1469 if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
1470 m_OneUse(m_FSub(m_FPOne(),
1471 m_Value(Z))))),
1472 m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
1473 return nullptr;
1474
1475 // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
1476 Value *XY = Builder.CreateFSubFMF(X, Y, &I);
1477 Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
1478 return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1479}
1480
1481/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1482static Instruction *factorizeFAddFSub(BinaryOperator &I,
1483 InstCombiner::BuilderTy &Builder) {
1484 assert((I.getOpcode() == Instruction::FAdd ||(static_cast <bool> ((I.getOpcode() == Instruction::FAdd
|| I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1485
, __extension__ __PRETTY_FUNCTION__))
1485 I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub")(static_cast <bool> ((I.getOpcode() == Instruction::FAdd
|| I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1485
, __extension__ __PRETTY_FUNCTION__))
;
1486 assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&(static_cast <bool> (I.hasAllowReassoc() && I.hasNoSignedZeros
() && "FP factorization requires FMF") ? void (0) : __assert_fail
("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1487
, __extension__ __PRETTY_FUNCTION__))
1487 "FP factorization requires FMF")(static_cast <bool> (I.hasAllowReassoc() && I.hasNoSignedZeros
() && "FP factorization requires FMF") ? void (0) : __assert_fail
("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1487
, __extension__ __PRETTY_FUNCTION__))
;
1488
1489 if (Instruction *Lerp = factorizeLerp(I, Builder))
1490 return Lerp;
1491
1492 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1493 if (!Op0->hasOneUse() || !Op1->hasOneUse())
1494 return nullptr;
1495
1496 Value *X, *Y, *Z;
1497 bool IsFMul;
1498 if ((match(Op0, m_FMul(m_Value(X), m_Value(Z))) &&
1499 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))) ||
1500 (match(Op0, m_FMul(m_Value(Z), m_Value(X))) &&
1501 match(Op1, m_c_FMul(m_Value(Y), m_Specific(Z)))))
1502 IsFMul = true;
1503 else if (match(Op0, m_FDiv(m_Value(X), m_Value(Z))) &&
1504 match(Op1, m_FDiv(m_Value(Y), m_Specific(Z))))
1505 IsFMul = false;
1506 else
1507 return nullptr;
1508
1509 // (X * Z) + (Y * Z) --> (X + Y) * Z
1510 // (X * Z) - (Y * Z) --> (X - Y) * Z
1511 // (X / Z) + (Y / Z) --> (X + Y) / Z
1512 // (X / Z) - (Y / Z) --> (X - Y) / Z
1513 bool IsFAdd = I.getOpcode() == Instruction::FAdd;
1514 Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
1515 : Builder.CreateFSubFMF(X, Y, &I);
1516
1517 // Bail out if we just created a denormal constant.
1518 // TODO: This is copied from a previous implementation. Is it necessary?
1519 const APFloat *C;
1520 if (match(XY, m_APFloat(C)) && !C->isNormal())
1521 return nullptr;
1522
1523 return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
1524 : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1525}
1526
1527Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
1528 if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
1529 I.getFastMathFlags(),
1530 SQ.getWithInstruction(&I)))
1531 return replaceInstUsesWith(I, V);
1532
1533 if (SimplifyAssociativeOrCommutative(I))
1534 return &I;
1535
1536 if (Instruction *X = foldVectorBinop(I))
1537 return X;
1538
1539 if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
1540 return FoldedFAdd;
1541
1542 // (-X) + Y --> Y - X
1543 Value *X, *Y;
1544 if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
1545 return BinaryOperator::CreateFSubFMF(Y, X, &I);
1546
1547 // Similar to above, but look through fmul/fdiv for the negated term.
1548 // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
1549 Value *Z;
1550 if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
1551 m_Value(Z)))) {
1552 Value *XY = Builder.CreateFMulFMF(X, Y, &I);
1553 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1554 }
1555 // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
1556 // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
1557 if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
1558 m_Value(Z))) ||
1559 match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
1560 m_Value(Z)))) {
1561 Value *XY = Builder.CreateFDivFMF(X, Y, &I);
1562 return BinaryOperator::CreateFSubFMF(Z, XY, &I);
1563 }
1564
1565 // Check for (fadd double (sitofp x), y), see if we can merge this into an
1566 // integer add followed by a promotion.
1567 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1568 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
1569 Value *LHSIntVal = LHSConv->getOperand(0);
1570 Type *FPType = LHSConv->getType();
1571
1572 // TODO: This check is overly conservative. In many cases known bits
1573 // analysis can tell us that the result of the addition has less significant
1574 // bits than the integer type can hold.
1575 auto IsValidPromotion = [](Type *FTy, Type *ITy) {
1576 Type *FScalarTy = FTy->getScalarType();
1577 Type *IScalarTy = ITy->getScalarType();
1578
1579 // Do we have enough bits in the significand to represent the result of
1580 // the integer addition?
1581 unsigned MaxRepresentableBits =
1582 APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
1583 return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
1584 };
1585
1586 // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
1587 // ... if the constant fits in the integer value. This is useful for things
1588 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
1589 // requires a constant pool load, and generally allows the add to be better
1590 // instcombined.
1591 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
1592 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1593 Constant *CI =
1594 ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
1595 if (LHSConv->hasOneUse() &&
1596 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1597 willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
1598 // Insert the new integer add.
1599 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
1600 return new SIToFPInst(NewAdd, I.getType());
1601 }
1602 }
1603
1604 // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1605 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1606 Value *RHSIntVal = RHSConv->getOperand(0);
1607 // It's enough to check LHS types only because we require int types to
1608 // be the same for this transform.
1609 if (IsValidPromotion(FPType, LHSIntVal->getType())) {
1610 // Only do this if x/y have the same type, if at least one of them has a
1611 // single use (so we don't increase the number of int->fp conversions),
1612 // and if the integer add will not overflow.
1613 if (LHSIntVal->getType() == RHSIntVal->getType() &&
1614 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1615 willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
1616 // Insert the new integer add.
1617 Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
1618 return new SIToFPInst(NewAdd, I.getType());
1619 }
1620 }
1621 }
1622 }
1623
1624 // Handle specials cases for FAdd with selects feeding the operation
1625 if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
1626 return replaceInstUsesWith(I, V);
1627
1628 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
1629 if (Instruction *F = factorizeFAddFSub(I, Builder))
1630 return F;
1631
1632 // Try to fold fadd into start value of reduction intrinsic.
1633 if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1634 m_AnyZeroFP(), m_Value(X))),
1635 m_Value(Y)))) {
1636 // fadd (rdx 0.0, X), Y --> rdx Y, X
1637 return replaceInstUsesWith(
1638 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1639 {X->getType()}, {Y, X}, &I));
1640 }
1641 const APFloat *StartC, *C;
1642 if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
1643 m_APFloat(StartC), m_Value(X)))) &&
1644 match(RHS, m_APFloat(C))) {
1645 // fadd (rdx StartC, X), C --> rdx (C + StartC), X
1646 Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
1647 return replaceInstUsesWith(
1648 I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
1649 {X->getType()}, {NewStartC, X}, &I));
1650 }
1651
1652 // (X * MulC) + X --> X * (MulC + 1.0)
1653 Constant *MulC;
1654 if (match(&I, m_c_FAdd(m_FMul(m_Value(X), m_ImmConstant(MulC)),
1655 m_Deferred(X)))) {
1656 MulC = ConstantExpr::getFAdd(MulC, ConstantFP::get(I.getType(), 1.0));
1657 return BinaryOperator::CreateFMulFMF(X, MulC, &I);
1658 }
1659
1660 if (Value *V = FAddCombine(Builder).simplify(&I))
1661 return replaceInstUsesWith(I, V);
1662 }
1663
1664 return nullptr;
1665}
1666
1667/// Optimize pointer differences into the same array into a size. Consider:
1668/// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1669/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1670Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
1671 Type *Ty, bool IsNUW) {
1672 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1673 // this.
1674 bool Swapped = false;
1675 GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
1676 if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
1677 std::swap(LHS, RHS);
1678 Swapped = true;
1679 }
1680
1681 // Require at least one GEP with a common base pointer on both sides.
1682 if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
1683 // (gep X, ...) - X
1684 if (LHSGEP->getOperand(0) == RHS) {
1685 GEP1 = LHSGEP;
1686 } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
1687 // (gep X, ...) - (gep X, ...)
1688 if (LHSGEP->getOperand(0)->stripPointerCasts() ==
1689 RHSGEP->getOperand(0)->stripPointerCasts()) {
1690 GEP1 = LHSGEP;
1691 GEP2 = RHSGEP;
1692 }
1693 }
1694 }
1695
1696 if (!GEP1)
1697 return nullptr;
1698
1699 if (GEP2) {
1700 // (gep X, ...) - (gep X, ...)
1701 //
1702 // Avoid duplicating the arithmetic if there are more than one non-constant
1703 // indices between the two GEPs and either GEP has a non-constant index and
1704 // multiple users. If zero non-constant index, the result is a constant and
1705 // there is no duplication. If one non-constant index, the result is an add
1706 // or sub with a constant, which is no larger than the original code, and
1707 // there's no duplicated arithmetic, even if either GEP has multiple
1708 // users. If more than one non-constant indices combined, as long as the GEP
1709 // with at least one non-constant index doesn't have multiple users, there
1710 // is no duplication.
1711 unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
1712 unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
1713 if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
1714 ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
1715 (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
1716 return nullptr;
1717 }
1718 }
1719
1720 // Emit the offset of the GEP and an intptr_t.
1721 Value *Result = EmitGEPOffset(GEP1);
1722
1723 // If this is a single inbounds GEP and the original sub was nuw,
1724 // then the final multiplication is also nuw.
1725 if (auto *I = dyn_cast<Instruction>(Result))
1726 if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
1727 I->getOpcode() == Instruction::Mul)
1728 I->setHasNoUnsignedWrap();
1729
1730 // If we have a 2nd GEP of the same base pointer, subtract the offsets.
1731 // If both GEPs are inbounds, then the subtract does not have signed overflow.
1732 if (GEP2) {
1733 Value *Offset = EmitGEPOffset(GEP2);
1734 Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
1735 GEP1->isInBounds() && GEP2->isInBounds());
1736 }
1737
1738 // If we have p - gep(p, ...) then we have to negate the result.
1739 if (Swapped)
1740 Result = Builder.CreateNeg(Result, "diff.neg");
1741
1742 return Builder.CreateIntCast(Result, Ty, true);
1743}
1744
1745Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
1746 if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
1747 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
1748 SQ.getWithInstruction(&I)))
1749 return replaceInstUsesWith(I, V);
1750
1751 if (Instruction *X = foldVectorBinop(I))
1752 return X;
1753
1754 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1755
1756 // If this is a 'B = x-(-A)', change to B = x+A.
1757 // We deal with this without involving Negator to preserve NSW flag.
1758 if (Value *V = dyn_castNegVal(Op1)) {
1759 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1760
1761 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
1762 assert(BO->getOpcode() == Instruction::Sub &&(static_cast <bool> (BO->getOpcode() == Instruction::
Sub && "Expected a subtraction operator!") ? void (0)
: __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1763
, __extension__ __PRETTY_FUNCTION__))
1763 "Expected a subtraction operator!")(static_cast <bool> (BO->getOpcode() == Instruction::
Sub && "Expected a subtraction operator!") ? void (0)
: __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp", 1763
, __extension__ __PRETTY_FUNCTION__))
;
1764 if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
1765 Res->setHasNoSignedWrap(true);
1766 } else {
1767 if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
1768 Res->setHasNoSignedWrap(true);
1769 }
1770
1771 return Res;
1772 }
1773
1774 // Try this before Negator to preserve NSW flag.
1775 if (Instruction *R = factorizeMathWithShlOps(I, Builder))
1776 return R;
1777
1778 Constant *C;
1779 if (match(Op0, m_ImmConstant(C))) {
1780 Value *X;
1781 Constant *C2;
1782
1783 // C-(X+C2) --> (C-C2)-X
1784 if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
1785 return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
1786 }
1787
1788 auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
1789 if (Instruction *Ext = narrowMathIfNoOverflow(I))
1790 return Ext;
1791
1792 bool Changed = false;
1793 if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
1794 Changed = true;
1795 I.setHasNoSignedWrap(true);
1796 }
1797 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
1798 Changed = true;
1799 I.setHasNoUnsignedWrap(true);
1800 }
1801
1802 return Changed ? &I : nullptr;
1803 };
1804
1805 // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
1806 // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
1807 // a pure negation used by a select that looks like abs/nabs.
1808 bool IsNegation = match(Op0, m_ZeroInt());
1809 if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
1810 const Instruction *UI = dyn_cast<Instruction>(U);
1811 if (!UI)
1812 return false;
1813 return match(UI,
1814 m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
1815 match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
1816 })) {
1817 if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
1818 return BinaryOperator::CreateAdd(NegOp1, Op0);
1819 }
1820 if (IsNegation)
1821 return TryToNarrowDeduceFlags(); // Should have been handled in Negator!
1822
1823 // (A*B)-(A*C) -> A*(B-C) etc
1824 if (Value *V = SimplifyUsingDistributiveLaws(I))
1825 return replaceInstUsesWith(I, V);
1826
1827 if (I.getType()->isIntOrIntVectorTy(1))
1828 return BinaryOperator::CreateXor(Op0, Op1);
1829
1830 // Replace (-1 - A) with (~A).
1831 if (match(Op0, m_AllOnes()))
1832 return BinaryOperator::CreateNot(Op1);
1833
1834 // (X + -1) - Y --> ~Y + X
1835 Value *X, *Y;
1836 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
1837 return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);
1838
1839 // Reassociate sub/add sequences to create more add instructions and
1840 // reduce dependency chains:
1841 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
1842 Value *Z;
1843 if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
1844 m_Value(Z))))) {
1845 Value *XZ = Builder.CreateAdd(X, Z);
1846 Value *YW = Builder.CreateAdd(Y, Op1);
1847 return BinaryOperator::CreateSub(XZ, YW);
1848 }
1849
1850 // ((X - Y) - Op1) --> X - (Y + Op1)
1851 if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
1852 Value *Add = Builder.CreateAdd(Y, Op1);
1853 return BinaryOperator::CreateSub(X, Add);
1854 }
1855
1856 // (~X) - (~Y) --> Y - X
1857 // This is placed after the other reassociations and explicitly excludes a
1858 // sub-of-sub pattern to avoid infinite looping.
1859 if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
1860 isFreeToInvert(Op1, Op1->hasOneUse()) &&
1861 !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
1862 Value *NotOp0 = Builder.CreateNot(Op0);
1863 Value *NotOp1 = Builder.CreateNot(Op1);
1864 return BinaryOperator::CreateSub(NotOp1, NotOp0);
1865 }
1866
1867 auto m_AddRdx = [](Value *&Vec) {
1868 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
1869 };
1870 Value *V0, *V1;
1871 if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
1872 V0->getType() == V1->getType()) {
1873 // Difference of sums is sum of differences:
1874 // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
1875 Value *Sub = Builder.CreateSub(V0, V1);
1876 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
1877 {Sub->getType()}, {Sub});
1878 return replaceInstUsesWith(I, Rdx);
1879 }
1880
1881 if (Constant *C = dyn_cast<Constant>(Op0)) {
1882 Value *X;
1883 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1884 // C - (zext bool) --> bool ? C - 1 : C
1885 return SelectInst::Create(X, InstCombiner::SubOne(C), C);
1886 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1887 // C - (sext bool) --> bool ? C + 1 : C
1888 return SelectInst::Create(X, InstCombiner::AddOne(C), C);
1889
1890 // C - ~X == X + (1+C)
1891 if (match(Op1, m_Not(m_Value(X))))
1892 return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));
1893
1894 // Try to fold constant sub into select arguments.
1895 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1896 if (Instruction *R = FoldOpIntoSelect(I, SI))
1897 return R;
1898
1899 // Try to fold constant sub into PHI values.
1900 if (PHINode *PN = dyn_cast<PHINode>(Op1))
1901 if (Instruction *R = foldOpIntoPhi(I, PN))
1902 return R;
1903
1904 Constant *C2;
1905
1906 // C-(C2-X) --> X+(C-C2)
1907 if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
1908 return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
1909 }
1910
1911 const APInt *Op0C;
1912 if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) {
1913 // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
1914 // zero.
1915 KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
1916 if ((*Op0C | RHSKnown.Zero).isAllOnes())
1917 return BinaryOperator::CreateXor(Op1, Op0);
1918 }
1919
1920 {
1921 Value *Y;
1922 // X-(X+Y) == -Y X-(Y+X) == -Y
1923 if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
1924 return BinaryOperator::CreateNeg(Y);
1925
1926 // (X-Y)-X == -Y
1927 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
1928 return BinaryOperator::CreateNeg(Y);
1929 }
1930
1931 // (sub (or A, B) (and A, B)) --> (xor A, B)
1932 {
1933 Value *A, *B;
1934 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1935 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1936 return BinaryOperator::CreateXor(A, B);
1937 }
1938
1939 // (sub (add A, B) (or A, B)) --> (and A, B)
1940 {
1941 Value *A, *B;
1942 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1943 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1944 return BinaryOperator::CreateAnd(A, B);
1945 }
1946
1947 // (sub (add A, B) (and A, B)) --> (or A, B)
1948 {
1949 Value *A, *B;
1950 if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
1951 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
1952 return BinaryOperator::CreateOr(A, B);
1953 }
1954
1955 // (sub (and A, B) (or A, B)) --> neg (xor A, B)
1956 {
1957 Value *A, *B;
1958 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1959 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1960 (Op0->hasOneUse() || Op1->hasOneUse()))
1961 return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
1962 }
1963
1964 // (sub (or A, B), (xor A, B)) --> (and A, B)
1965 {
1966 Value *A, *B;
1967 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1968 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1969 return BinaryOperator::CreateAnd(A, B);
1970 }
1971
1972 // (sub (xor A, B) (or A, B)) --> neg (and A, B)
1973 {
1974 Value *A, *B;
1975 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1976 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
1977 (Op0->hasOneUse() || Op1->hasOneUse()))
1978 return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
1979 }
1980
1981 {
1982 Value *Y;
1983 // ((X | Y) - X) --> (~X & Y)
1984 if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
1985 return BinaryOperator::CreateAnd(
1986 Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
1987 }
1988
1989 {
1990 // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
1991 Value *X;
1992 if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
1993 m_OneUse(m_Neg(m_Value(X))))))) {
1994 return BinaryOperator::CreateNeg(Builder.CreateAnd(
1995 Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
1996 }
1997 }
1998
1999 {
2000 // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
2001 Constant *C;
2002 if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
2003 return BinaryOperator::CreateNeg(
2004 Builder.CreateAnd(Op1, Builder.CreateNot(C)));
2005 }
2006 }
2007
2008 {
2009 // If we have a subtraction between some value and a select between
2010 // said value and something else, sink subtraction into select hands, i.e.:
2011 // sub (select %Cond, %TrueVal, %FalseVal), %Op1
2012 // ->
2013 // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
2014 // or
2015 // sub %Op0, (select %Cond, %TrueVal, %FalseVal)
2016 // ->
2017 // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
2018 // This will result in select between new subtraction and 0.
2019 auto SinkSubIntoSelect =
2020 [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
2021 auto SubBuilder) -> Instruction * {
2022 Value *Cond, *TrueVal, *FalseVal;
2023 if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
2024 m_Value(FalseVal)))))
2025 return nullptr;
2026 if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
2027 return nullptr;
2028 // While it is really tempting to just create two subtractions and let
2029 // InstCombine fold one of those to 0, it isn't possible to do so
2030 // because of worklist visitation order. So ugly it is.
2031 bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
2032 Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
2033 Constant *Zero = Constant::getNullValue(Ty);
2034 SelectInst *NewSel =
2035 SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
2036 OtherHandOfSubIsTrueVal ? NewSub : Zero);
2037 // Preserve prof metadata if any.
2038 NewSel->copyMetadata(cast<Instruction>(*Select));
2039 return NewSel;
2040 };
2041 if (Instruction *NewSel = SinkSubIntoSelect(
2042 /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
2043 [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
2044 return Builder->CreateSub(OtherHandOfSelect,
2045 /*OtherHandOfSub=*/Op1);
2046 }))
2047 return NewSel;
2048 if (Instruction *NewSel = SinkSubIntoSelect(
2049 /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
2050 [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
2051 return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
2052 OtherHandOfSelect);
2053 }))
2054 return NewSel;
2055 }
2056
2057 // (X - (X & Y)) --> (X & ~Y)
2058 if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
2059 (Op1->hasOneUse() || isa<Constant>(Y)))
2060 return BinaryOperator::CreateAnd(
2061 Op0, Builder.CreateNot(Y, Y->getName() + ".not"));
2062
2063 // ~X - Min/Max(~X, Y) -> ~Min/Max(X, ~Y) - X
2064 // ~X - Min/Max(Y, ~X) -> ~Min/Max(X, ~Y) - X
2065 // Min/Max(~X, Y) - ~X -> X - ~Min/Max(X, ~Y)
2066 // Min/Max(Y, ~X) - ~X -> X - ~Min/Max(X, ~Y)
2067 // As long as Y is freely invertible, this will be neutral or a win.
2068 // Note: We don't generate the inverse max/min, just create the 'not' of
2069 // it and let other folds do the rest.
2070 if (match(Op0, m_Not(m_Value(X))) &&
2071 match(Op1, m_c_MaxOrMin(m_Specific(Op0), m_Value(Y))) &&
2072 !Op0->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2073 Value *Not = Builder.CreateNot(Op1);
2074 return BinaryOperator::CreateSub(Not, X);
2075 }
2076 if (match(Op1, m_Not(m_Value(X))) &&
2077 match(Op0, m_c_MaxOrMin(m_Specific(Op1), m_Value(Y))) &&
2078 !Op1->hasNUsesOrMore(3) && isFreeToInvert(Y, Y->hasOneUse())) {
2079 Value *Not = Builder.CreateNot(Op0);
2080 return BinaryOperator::CreateSub(X, Not);
2081 }
2082
2083 // TODO: This is the same logic as above but handles the cmp-select idioms
2084 // for min/max, so the use checks are increased to account for the
2085 // extra instructions. If we canonicalize to intrinsics, this block
2086 // can likely be removed.
2087 {
2088 Value *LHS, *RHS, *A;
2089 Value *NotA = Op0, *MinMax = Op1;
2090 SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2091 if (!SelectPatternResult::isMinOrMax(SPF)) {
2092 NotA = Op1;
2093 MinMax = Op0;
2094 SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
2095 }
2096 if (SelectPatternResult::isMinOrMax(SPF) &&
2097 match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
2098 if (NotA == LHS)
2099 std::swap(LHS, RHS);
2100 // LHS is now Y above and expected to have at least 2 uses (the min/max)
2101 // NotA is expected to have 2 uses from the min/max and 1 from the sub.
2102 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
2103 !NotA->hasNUsesOrMore(4)) {
2104 Value *Not = Builder.CreateNot(MinMax);
2105 if (NotA == Op0)
2106 return BinaryOperator::CreateSub(Not, A);
2107 else
2108 return BinaryOperator::CreateSub(A, Not);
2109 }
2110 }
2111 }
2112
2113 // Optimize pointer differences into the same array into a size. Consider:
2114 // &A[10] - &A[0]: we should compile this to "10".
2115 Value *LHSOp, *RHSOp;
2116 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2117 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2118 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2119 I.hasNoUnsignedWrap()))
2120 return replaceInstUsesWith(I, Res);
2121
2122 // trunc(p)-trunc(q) -> trunc(p-q)
2123 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2124 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2125 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
2126 /* IsNUW */ false))
2127 return replaceInstUsesWith(I, Res);
2128
2129 // Canonicalize a shifty way to code absolute value to the common pattern.
2130 // There are 2 potential commuted variants.
2131 // We're relying on the fact that we only do this transform when the shift has
2132 // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
2133 // instructions).
2134 Value *A;
2135 const APInt *ShAmt;
2136 Type *Ty = I.getType();
2137 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2138 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2139 match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
2140 // B = ashr i32 A, 31 ; smear the sign bit
2141 // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1)
2142 // --> (A < 0) ? -A : A
2143 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2144 // Copy the nuw/nsw flags from the sub to the negate.
2145 Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
2146 I.hasNoSignedWrap());
2147 return SelectInst::Create(Cmp, Neg, A);
2148 }
2149
2150 // If we are subtracting a low-bit masked subset of some value from an add
2151 // of that same value with no low bits changed, that is clearing some low bits
2152 // of the sum:
2153 // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
2154 const APInt *AddC, *AndC;
2155 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
2156 match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
2157 unsigned BitWidth = Ty->getScalarSizeInBits();
2158 unsigned Cttz = AddC->countTrailingZeros();
2159 APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
2160 if ((HighMask & *AndC).isZero())
2161 return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
2162 }
2163
2164 if (Instruction *V =
2165 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2166 return V;
2167
2168 // X - usub.sat(X, Y) => umin(X, Y)
2169 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
2170 m_Value(Y)))))
2171 return replaceInstUsesWith(
2172 I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));
2173
2174 // umax(X, Op1) - Op1 --> usub.sat(X, Op1)
2175 // TODO: The one-use restriction is not strictly necessary, but it may
2176 // require improving other pattern matching and/or codegen.
2177 if (match(Op0, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op1)))))
2178 return replaceInstUsesWith(
2179 I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
2180
2181 // Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
2182 if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
2183 Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op0});
2184 return BinaryOperator::CreateNeg(USub);
2185 }
2186
2187 // C - ctpop(X) => ctpop(~X) if C is bitwidth
2188 if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
2189 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
2190 return replaceInstUsesWith(
2191 I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
2192 {Builder.CreateNot(X)}));
2193
2194 return TryToNarrowDeduceFlags();
2195}
2196
2197/// This eliminates floating-point negation in either 'fneg(X)' or
2198/// 'fsub(-0.0, X)' form by combining into a constant operand.
2199static Instruction *foldFNegIntoConstant(Instruction &I) {
2200 // This is limited with one-use because fneg is assumed better for
2201 // reassociation and cheaper in codegen than fmul/fdiv.
2202 // TODO: Should the m_OneUse restriction be removed?
2203 Instruction *FNegOp;
2204 if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
2205 return nullptr;
2206
2207 Value *X;
2208 Constant *C;
2209
2210 // Fold negation into constant operand.
2211 // -(X * C) --> X * (-C)
2212 if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
2213 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
2214 // -(X / C) --> X / (-C)
2215 if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
2216 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
2217 // -(C / X) --> (-C) / X
2218 if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) {
2219 Instruction *FDiv =
2220 BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
2221
2222 // Intersect 'nsz' and 'ninf' because those special value exceptions may not
2223 // apply to the fdiv. Everything else propagates from the fneg.
2224 // TODO: We could propagate nsz/ninf from fdiv alone?
2225 FastMathFlags FMF = I.getFastMathFlags();
2226 FastMathFlags OpFMF = FNegOp->getFastMathFlags();
2227 FDiv->setHasNoSignedZeros(FMF.noSignedZeros() && OpFMF.noSignedZeros());
2228 FDiv->setHasNoInfs(FMF.noInfs() && OpFMF.noInfs());
2229 return FDiv;
2230 }
2231 // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
2232 // -(X + C) --> -X + -C --> -C - X
2233 if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
2234 return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I);
2235
2236 return nullptr;
2237}
2238
2239static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
2240 InstCombiner::BuilderTy &Builder) {
2241 Value *FNeg;
2242 if (!match(&I, m_FNeg(m_Value(FNeg))))
2243 return nullptr;
2244
2245 Value *X, *Y;
2246 if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
2247 return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2248
2249 if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
2250 return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);
2251
2252 return nullptr;
2253}
2254
2255Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
2256 Value *Op = I.getOperand(0);
2257
2258 if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
2259 getSimplifyQuery().getWithInstruction(&I)))
2260 return replaceInstUsesWith(I, V);
2261
2262 if (Instruction *X = foldFNegIntoConstant(I))
2263 return X;
2264
2265 Value *X, *Y;
2266
2267 // If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
2268 if (I.hasNoSignedZeros() &&
2269 match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
2270 return BinaryOperator::CreateFSubFMF(Y, X, &I);
2271
2272 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2273 return R;
2274
2275 // Try to eliminate fneg if at least 1 arm of the select is negated.
2276 Value *Cond;
2277 if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
2278 // Unlike most transforms, this one is not safe to propagate nsz unless
2279 // it is present on the original select. (We are conservatively intersecting
2280 // the nsz flags from the select and root fneg instruction.)
2281 auto propagateSelectFMF = [&](SelectInst *S) {
2282 S->copyFastMathFlags(&I);
2283 if (auto *OldSel = dyn_cast<SelectInst>(Op))
2284 if (!OldSel->hasNoSignedZeros())
2285 S->setHasNoSignedZeros(false);
2286 };
2287 // -(Cond ? -P : Y) --> Cond ? P : -Y
2288 Value *P;
2289 if (match(X, m_FNeg(m_Value(P)))) {
2290 Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
2291 SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
2292 propagateSelectFMF(NewSel);
2293 return NewSel;
2294 }
2295 // -(Cond ? X : -P) --> Cond ? -X : P
2296 if (match(Y, m_FNeg(m_Value(P)))) {
2297 Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
2298 SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
2299 propagateSelectFMF(NewSel);
2300 return NewSel;
2301 }
2302 }
2303
2304 return nullptr;
2305}
2306
2307Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
2308 if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
2309 I.getFastMathFlags(),
2310 getSimplifyQuery().getWithInstruction(&I)))
2311 return replaceInstUsesWith(I, V);
2312
2313 if (Instruction *X = foldVectorBinop(I))
2314 return X;
2315
2316 // Subtraction from -0.0 is the canonical form of fneg.
2317 // fsub -0.0, X ==> fneg X
2318 // fsub nsz 0.0, X ==> fneg nsz X
2319 //
2320 // FIXME This matcher does not respect FTZ or DAZ yet:
2321 // fsub -0.0, Denorm ==> +-0
2322 // fneg Denorm ==> -Denorm
2323 Value *Op;
2324 if (match(&I, m_FNeg(m_Value(Op))))
2325 return UnaryOperator::CreateFNegFMF(Op, &I);
2326
2327 if (Instruction *X = foldFNegIntoConstant(I))
2328 return X;
2329
2330 if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
2331 return R;
2332
2333 Value *X, *Y;
2334 Constant *C;
2335
2336 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2337 // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
2338 // Canonicalize to fadd to make analysis easier.
2339 // This can also help codegen because fadd is commutative.
2340 // Note that if this fsub was really an fneg, the fadd with -0.0 will get
2341 // killed later. We still limit that particular transform with 'hasOneUse'
2342 // because an fneg is assumed better/cheaper than a generic fsub.
2343 if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
2344 if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2345 Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
2346 return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
2347 }
2348 }
2349
2350 // (-X) - Op1 --> -(X + Op1)
2351 if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
2352 match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
2353 Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
2354 return UnaryOperator::CreateFNegFMF(FAdd, &I);
2355 }
2356
2357 if (isa<Constant>(Op0))
2358 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2359 if (Instruction *NV = FoldOpIntoSelect(I, SI))
2360 return NV;
2361
2362 // X - C --> X + (-C)
2363 // But don't transform constant expressions because there's an inverse fold
2364 // for X + (-Y) --> X - Y.
2365 if (match(Op1, m_ImmConstant(C)))
2366 return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);
2367
2368 // X - (-Y) --> X + Y
2369 if (match(Op1, m_FNeg(m_Value(Y))))
2370 return BinaryOperator::CreateFAddFMF(Op0, Y, &I);
2371
2372 // Similar to above, but look through a cast of the negated value:
2373 // X - (fptrunc(-Y)) --> X + fptrunc(Y)
2374 Type *Ty = I.getType();
2375 if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
2376 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);
2377
2378 // X - (fpext(-Y)) --> X + fpext(Y)
2379 if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
2380 return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);
2381
2382 // Similar to above, but look through fmul/fdiv of the negated value:
2383 // Op0 - (-X * Y) --> Op0 + (X * Y)
2384 // Op0 - (Y * -X) --> Op0 + (X * Y)
2385 if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
2386 Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
2387 return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
2388 }
2389 // Op0 - (-X / Y) --> Op0 + (X / Y)
2390 // Op0 - (X / -Y) --> Op0 + (X / Y)
2391 if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
2392 match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
2393 Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
2394 return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
2395 }
2396
2397 // Handle special cases for FSub with selects feeding the operation
2398 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
2399 return replaceInstUsesWith(I, V);
2400
2401 if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
2402 // (Y - X) - Y --> -X
2403 if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
2404 return UnaryOperator::CreateFNegFMF(X, &I);
2405
2406 // Y - (X + Y) --> -X
2407 // Y - (Y + X) --> -X
2408 if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
2409 return UnaryOperator::CreateFNegFMF(X, &I);
2410
2411 // (X * C) - X --> X * (C - 1.0)
2412 if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
2413 Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
2414 return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
2415 }
2416 // X - (X * C) --> X * (1.0 - C)
2417 if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
2418 Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
2419 return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
2420 }
2421
2422 // Reassociate fsub/fadd sequences to create more fadd instructions and
2423 // reduce dependency chains:
2424 // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
2425 Value *Z;
2426 if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
2427 m_Value(Z))))) {
2428 Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
2429 Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
2430 return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
2431 }
2432
2433 auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
2434 return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
2435 m_Value(Vec)));
2436 };
2437 Value *A0, *A1, *V0, *V1;
2438 if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
2439 V0->getType() == V1->getType()) {
2440 // Difference of sums is sum of differences:
2441 // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
2442 Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
2443 Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
2444 {Sub->getType()}, {A0, Sub}, &I);
2445 return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
2446 }
2447
2448 if (Instruction *F = factorizeFAddFSub(I, Builder))
2449 return F;
2450
2451 // TODO: This performs reassociative folds for FP ops. Some fraction of the
2452 // functionality has been subsumed by simple pattern matching here and in
2453 // InstSimplify. We should let a dedicated reassociation pass handle more
2454 // complex pattern matching and remove this from InstCombine.
2455 if (Value *V = FAddCombine(Builder).simplify(&I))
2456 return replaceInstUsesWith(I, V);
2457
2458 // (X - Y) - Op1 --> X - (Y + Op1)
2459 if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
2460 Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
2461 return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
2462 }
2463 }
2464
2465 return nullptr;
2466}

/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/include/llvm/IR/PatternMatch.h

1//===- PatternMatch.h - Match on the LLVM IR --------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file provides a simple and efficient mechanism for performing general
10// tree-based pattern matches on the LLVM IR. The power of these routines is
11// that it allows you to write concise patterns that are expressive and easy to
12// understand. The other major advantage of this is that it allows you to
13// trivially capture/bind elements in the pattern to variables. For example,
14// you can do something like this:
15//
16// Value *Exp = ...
17// Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2)
18// if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)),
19// m_And(m_Value(Y), m_ConstantInt(C2))))) {
20// ... Pattern is matched and variables are bound ...
21// }
22//
23// This is primarily useful to things like the instruction combiner, but can
24// also be useful for static analysis tools or code generators.
25//
26//===----------------------------------------------------------------------===//
27
28#ifndef LLVM_IR_PATTERNMATCH_H
29#define LLVM_IR_PATTERNMATCH_H
30
31#include "llvm/ADT/APFloat.h"
32#include "llvm/ADT/APInt.h"
33#include "llvm/IR/Constant.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DataLayout.h"
36#include "llvm/IR/InstrTypes.h"
37#include "llvm/IR/Instruction.h"
38#include "llvm/IR/Instructions.h"
39#include "llvm/IR/IntrinsicInst.h"
40#include "llvm/IR/Intrinsics.h"
41#include "llvm/IR/Operator.h"
42#include "llvm/IR/Value.h"
43#include "llvm/Support/Casting.h"
44#include <cstdint>
45
46namespace llvm {
47namespace PatternMatch {
48
49template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) {
50 return const_cast<Pattern &>(P).match(V);
14
Calling 'AnyBinaryOp_match::match'
35
Returning from 'AnyBinaryOp_match::match'
36
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'
51}
52
53template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) {
54 return const_cast<Pattern &>(P).match(Mask);
55}
56
57template <typename SubPattern_t> struct OneUse_match {
58 SubPattern_t SubPattern;
59
60 OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {}
61
62 template <typename OpTy> bool match(OpTy *V) {
63 return V->hasOneUse() && SubPattern.match(V);
64 }
65};
66
67template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) {
68 return SubPattern;
69}
70
71template <typename Class> struct class_match {
72 template <typename ITy> bool match(ITy *V) { return isa<Class>(V); }
73};
74
75/// Match an arbitrary value and ignore it.
76inline class_match<Value> m_Value() { return class_match<Value>(); }
77
78/// Match an arbitrary unary operation and ignore it.
79inline class_match<UnaryOperator> m_UnOp() {
80 return class_match<UnaryOperator>();
81}
82
83/// Match an arbitrary binary operation and ignore it.
84inline class_match<BinaryOperator> m_BinOp() {
85 return class_match<BinaryOperator>();
86}
87
88/// Matches any compare instruction and ignore it.
89inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); }
90
91struct undef_match {
92 static bool check(const Value *V) {
93 if (isa<UndefValue>(V))
94 return true;
95
96 const auto *CA = dyn_cast<ConstantAggregate>(V);
97 if (!CA)
98 return false;
99
100 SmallPtrSet<const ConstantAggregate *, 8> Seen;
101 SmallVector<const ConstantAggregate *, 8> Worklist;
102
103 // Either UndefValue, PoisonValue, or an aggregate that only contains
104 // these is accepted by matcher.
105 // CheckValue returns false if CA cannot satisfy this constraint.
106 auto CheckValue = [&](const ConstantAggregate *CA) {
107 for (const Value *Op : CA->operand_values()) {
108 if (isa<UndefValue>(Op))
109 continue;
110
111 const auto *CA = dyn_cast<ConstantAggregate>(Op);
112 if (!CA)
113 return false;
114 if (Seen.insert(CA).second)
115 Worklist.emplace_back(CA);
116 }
117
118 return true;
119 };
120
121 if (!CheckValue(CA))
122 return false;
123
124 while (!Worklist.empty()) {
125 if (!CheckValue(Worklist.pop_back_val()))
126 return false;
127 }
128 return true;
129 }
130 template <typename ITy> bool match(ITy *V) { return check(V); }
131};
132
133/// Match an arbitrary undef constant. This matches poison as well.
134/// If this is an aggregate and contains a non-aggregate element that is
135/// neither undef nor poison, the aggregate is not matched.
136inline auto m_Undef() { return undef_match(); }
137
138/// Match an arbitrary poison constant.
139inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); }
140
141/// Match an arbitrary Constant and ignore it.
142inline class_match<Constant> m_Constant() { return class_match<Constant>(); }
143
144/// Match an arbitrary ConstantInt and ignore it.
145inline class_match<ConstantInt> m_ConstantInt() {
146 return class_match<ConstantInt>();
147}
148
149/// Match an arbitrary ConstantFP and ignore it.
150inline class_match<ConstantFP> m_ConstantFP() {
151 return class_match<ConstantFP>();
152}
153
154/// Match an arbitrary ConstantExpr and ignore it.
155inline class_match<ConstantExpr> m_ConstantExpr() {
156 return class_match<ConstantExpr>();
157}
158
159/// Match an arbitrary basic block value and ignore it.
160inline class_match<BasicBlock> m_BasicBlock() {
161 return class_match<BasicBlock>();
162}
163
164/// Inverting matcher
165template <typename Ty> struct match_unless {
166 Ty M;
167
168 match_unless(const Ty &Matcher) : M(Matcher) {}
169
170 template <typename ITy> bool match(ITy *V) { return !M.match(V); }
171};
172
173/// Match if the inner matcher does *NOT* match.
174template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) {
175 return match_unless<Ty>(M);
176}
177
178/// Matching combinators
179template <typename LTy, typename RTy> struct match_combine_or {
180 LTy L;
181 RTy R;
182
183 match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
184
185 template <typename ITy> bool match(ITy *V) {
186 if (L.match(V))
17
Taking false branch
187 return true;
188 if (R.match(V))
18
Calling 'match_combine_and::match'
31
Returning from 'match_combine_and::match'
32
Taking true branch
189 return true;
190 return false;
191 }
192};
193
194template <typename LTy, typename RTy> struct match_combine_and {
195 LTy L;
196 RTy R;
197
198 match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
199
200 template <typename ITy> bool match(ITy *V) {
201 if (L.match(V))
19
Calling 'BinaryOp_match::match'
28
Returning from 'BinaryOp_match::match'
29
Taking true branch
202 if (R.match(V))
30
Taking true branch
203 return true;
204 return false;
205 }
206};
207
208/// Combine two pattern matchers matching L || R
209template <typename LTy, typename RTy>
210inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) {
211 return match_combine_or<LTy, RTy>(L, R);
54
Returning without writing to 'L.Op.VR'
212}
213
214/// Combine two pattern matchers matching L && R
215template <typename LTy, typename RTy>
216inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) {
217 return match_combine_and<LTy, RTy>(L, R);
218}
219
220struct apint_match {
221 const APInt *&Res;
222 bool AllowUndef;
223
224 apint_match(const APInt *&Res, bool AllowUndef)
225 : Res(Res), AllowUndef(AllowUndef) {}
226
227 template <typename ITy> bool match(ITy *V) {
228 if (auto *CI = dyn_cast<ConstantInt>(V)) {
229 Res = &CI->getValue();
230 return true;
231 }
232 if (V->getType()->isVectorTy())
233 if (const auto *C = dyn_cast<Constant>(V))
234 if (auto *CI = dyn_cast_or_null<ConstantInt>(
235 C->getSplatValue(AllowUndef))) {
236 Res = &CI->getValue();
237 return true;
238 }
239 return false;
240 }
241};
242// Either constexpr if or renaming ConstantFP::getValueAPF to
243// ConstantFP::getValue is needed to do it via single template
244// function for both apint/apfloat.
245struct apfloat_match {
246 const APFloat *&Res;
247 bool AllowUndef;
248
249 apfloat_match(const APFloat *&Res, bool AllowUndef)
250 : Res(Res), AllowUndef(AllowUndef) {}
251
252 template <typename ITy> bool match(ITy *V) {
253 if (auto *CI = dyn_cast<ConstantFP>(V)) {
254 Res = &CI->getValueAPF();
255 return true;
256 }
257 if (V->getType()->isVectorTy())
258 if (const auto *C = dyn_cast<Constant>(V))
259 if (auto *CI = dyn_cast_or_null<ConstantFP>(
260 C->getSplatValue(AllowUndef))) {
261 Res = &CI->getValueAPF();
262 return true;
263 }
264 return false;
265 }
266};
267
268/// Match a ConstantInt or splatted ConstantVector, binding the
269/// specified pointer to the contained APInt.
270inline apint_match m_APInt(const APInt *&Res) {
271 // Forbid undefs by default to maintain previous behavior.
272 return apint_match(Res, /* AllowUndef */ false);
273}
274
275/// Match APInt while allowing undefs in splat vector constants.
276inline apint_match m_APIntAllowUndef(const APInt *&Res) {
277 return apint_match(Res, /* AllowUndef */ true);
278}
279
280/// Match APInt while forbidding undefs in splat vector constants.
281inline apint_match m_APIntForbidUndef(const APInt *&Res) {
282 return apint_match(Res, /* AllowUndef */ false);
283}
284
285/// Match a ConstantFP or splatted ConstantVector, binding the
286/// specified pointer to the contained APFloat.
287inline apfloat_match m_APFloat(const APFloat *&Res) {
288 // Forbid undefs by default to maintain previous behavior.
289 return apfloat_match(Res, /* AllowUndef */ false);
290}
291
292/// Match APFloat while allowing undefs in splat vector constants.
293inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) {
294 return apfloat_match(Res, /* AllowUndef */ true);
295}
296
297/// Match APFloat while forbidding undefs in splat vector constants.
298inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) {
299 return apfloat_match(Res, /* AllowUndef */ false);
300}
301
302template <int64_t Val> struct constantint_match {
303 template <typename ITy> bool match(ITy *V) {
304 if (const auto *CI = dyn_cast<ConstantInt>(V)) {
305 const APInt &CIV = CI->getValue();
306 if (Val >= 0)
307 return CIV == static_cast<uint64_t>(Val);
308 // If Val is negative, and CI is shorter than it, truncate to the right
309 // number of bits. If it is larger, then we have to sign extend. Just
310 // compare their negated values.
311 return -CIV == -Val;
312 }
313 return false;
314 }
315};
316
317/// Match a ConstantInt with a specific value.
318template <int64_t Val> inline constantint_match<Val> m_ConstantInt() {
319 return constantint_match<Val>();
320}
321
322/// This helper class is used to match constant scalars, vector splats,
323/// and fixed width vectors that satisfy a specified predicate.
324/// For fixed width vector constants, undefined elements are ignored.
325template <typename Predicate, typename ConstantVal>
326struct cstval_pred_ty : public Predicate {
327 template <typename ITy> bool match(ITy *V) {
328 if (const auto *CV = dyn_cast<ConstantVal>(V))
329 return this->isValue(CV->getValue());
330 if (const auto *VTy = dyn_cast<VectorType>(V->getType())) {
331 if (const auto *C = dyn_cast<Constant>(V)) {
332 if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue()))
333 return this->isValue(CV->getValue());
334
335 // Number of elements of a scalable vector unknown at compile time
336 auto *FVTy = dyn_cast<FixedVectorType>(VTy);
337 if (!FVTy)
338 return false;
339
340 // Non-splat vector constant: check each element for a match.
341 unsigned NumElts = FVTy->getNumElements();
342 assert(NumElts != 0 && "Constant vector with no elements?")(static_cast <bool> (NumElts != 0 && "Constant vector with no elements?"
) ? void (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\""
, "llvm/include/llvm/IR/PatternMatch.h", 342, __extension__ __PRETTY_FUNCTION__
))
;
343 bool HasNonUndefElements = false;
344 for (unsigned i = 0; i != NumElts; ++i) {
345 Constant *Elt = C->getAggregateElement(i);
346 if (!Elt)
347 return false;
348 if (isa<UndefValue>(Elt))
349 continue;
350 auto *CV = dyn_cast<ConstantVal>(Elt);
351 if (!CV || !this->isValue(CV->getValue()))
352 return false;
353 HasNonUndefElements = true;
354 }
355 return HasNonUndefElements;
356 }
357 }
358 return false;
359 }
360};
361
362/// specialization of cstval_pred_ty for ConstantInt
363template <typename Predicate>
364using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>;
365
366/// specialization of cstval_pred_ty for ConstantFP
367template <typename Predicate>
368using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>;
369
370/// This helper class is used to match scalar and vector constants that
371/// satisfy a specified predicate, and bind them to an APInt.
372template <typename Predicate> struct api_pred_ty : public Predicate {
373 const APInt *&Res;
374
375 api_pred_ty(const APInt *&R) : Res(R) {}
376
377 template <typename ITy> bool match(ITy *V) {
378 if (const auto *CI = dyn_cast<ConstantInt>(V))
379 if (this->isValue(CI->getValue())) {
380 Res = &CI->getValue();
381 return true;
382 }
383 if (V->getType()->isVectorTy())
384 if (const auto *C = dyn_cast<Constant>(V))
385 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
386 if (this->isValue(CI->getValue())) {
387 Res = &CI->getValue();
388 return true;
389 }
390
391 return false;
392 }
393};
394
395/// This helper class is used to match scalar and vector constants that
396/// satisfy a specified predicate, and bind them to an APFloat.
397/// Undefs are allowed in splat vector constants.
398template <typename Predicate> struct apf_pred_ty : public Predicate {
399 const APFloat *&Res;
400
401 apf_pred_ty(const APFloat *&R) : Res(R) {}
402
403 template <typename ITy> bool match(ITy *V) {
404 if (const auto *CI = dyn_cast<ConstantFP>(V))
405 if (this->isValue(CI->getValue())) {
406 Res = &CI->getValue();
407 return true;
408 }
409 if (V->getType()->isVectorTy())
410 if (const auto *C = dyn_cast<Constant>(V))
411 if (auto *CI = dyn_cast_or_null<ConstantFP>(
412 C->getSplatValue(/* AllowUndef */ true)))
413 if (this->isValue(CI->getValue())) {
414 Res = &CI->getValue();
415 return true;
416 }
417
418 return false;
419 }
420};
421
422///////////////////////////////////////////////////////////////////////////////
423//
424// Encapsulate constant value queries for use in templated predicate matchers.
425// This allows checking if constants match using compound predicates and works
426// with vector constants, possibly with relaxed constraints. For example, ignore
427// undef values.
428//
429///////////////////////////////////////////////////////////////////////////////
430
431struct is_any_apint {
432 bool isValue(const APInt &C) { return true; }
433};
434/// Match an integer or vector with any integral constant.
435/// For vectors, this includes constants with undefined elements.
436inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() {
437 return cst_pred_ty<is_any_apint>();
438}
439
440struct is_all_ones {
441 bool isValue(const APInt &C) { return C.isAllOnes(); }
442};
443/// Match an integer or vector with all bits set.
444/// For vectors, this includes constants with undefined elements.
445inline cst_pred_ty<is_all_ones> m_AllOnes() {
446 return cst_pred_ty<is_all_ones>();
447}
448
449struct is_maxsignedvalue {
450 bool isValue(const APInt &C) { return C.isMaxSignedValue(); }
451};
452/// Match an integer or vector with values having all bits except for the high
453/// bit set (0x7f...).
454/// For vectors, this includes constants with undefined elements.
455inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() {
456 return cst_pred_ty<is_maxsignedvalue>();
457}
458inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) {
459 return V;
460}
461
462struct is_negative {
463 bool isValue(const APInt &C) { return C.isNegative(); }
464};
465/// Match an integer or vector of negative values.
466/// For vectors, this includes constants with undefined elements.
467inline cst_pred_ty<is_negative> m_Negative() {
468 return cst_pred_ty<is_negative>();
469}
470inline api_pred_ty<is_negative> m_Negative(const APInt *&V) {
471 return V;
472}
473
474struct is_nonnegative {
475 bool isValue(const APInt &C) { return C.isNonNegative(); }
476};
477/// Match an integer or vector of non-negative values.
478/// For vectors, this includes constants with undefined elements.
479inline cst_pred_ty<is_nonnegative> m_NonNegative() {
480 return cst_pred_ty<is_nonnegative>();
481}
482inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) {
483 return V;
484}
485
486struct is_strictlypositive {
487 bool isValue(const APInt &C) { return C.isStrictlyPositive(); }
488};
489/// Match an integer or vector of strictly positive values.
490/// For vectors, this includes constants with undefined elements.
491inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() {
492 return cst_pred_ty<is_strictlypositive>();
493}
494inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) {
495 return V;
496}
497
498struct is_nonpositive {
499 bool isValue(const APInt &C) { return C.isNonPositive(); }
500};
501/// Match an integer or vector of non-positive values.
502/// For vectors, this includes constants with undefined elements.
503inline cst_pred_ty<is_nonpositive> m_NonPositive() {
504 return cst_pred_ty<is_nonpositive>();
505}
506inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; }
507
508struct is_one {
509 bool isValue(const APInt &C) { return C.isOne(); }
510};
511/// Match an integer 1 or a vector with all elements equal to 1.
512/// For vectors, this includes constants with undefined elements.
513inline cst_pred_ty<is_one> m_One() {
514 return cst_pred_ty<is_one>();
515}
516
517struct is_zero_int {
518 bool isValue(const APInt &C) { return C.isZero(); }
519};
520/// Match an integer 0 or a vector with all elements equal to 0.
521/// For vectors, this includes constants with undefined elements.
522inline cst_pred_ty<is_zero_int> m_ZeroInt() {
523 return cst_pred_ty<is_zero_int>();
524}
525
526struct is_zero {
527 template <typename ITy> bool match(ITy *V) {
528 auto *C = dyn_cast<Constant>(V);
529 // FIXME: this should be able to do something for scalable vectors
530 return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C));
531 }
532};
533/// Match any null constant or a vector with all elements equal to 0.
534/// For vectors, this includes constants with undefined elements.
535inline is_zero m_Zero() {
536 return is_zero();
537}
538
539struct is_power2 {
540 bool isValue(const APInt &C) { return C.isPowerOf2(); }
541};
542/// Match an integer or vector power-of-2.
543/// For vectors, this includes constants with undefined elements.
544inline cst_pred_ty<is_power2> m_Power2() {
545 return cst_pred_ty<is_power2>();
546}
547inline api_pred_ty<is_power2> m_Power2(const APInt *&V) {
548 return V;
549}
550
551struct is_negated_power2 {
552 bool isValue(const APInt &C) { return C.isNegatedPowerOf2(); }
553};
554/// Match a integer or vector negated power-of-2.
555/// For vectors, this includes constants with undefined elements.
556inline cst_pred_ty<is_negated_power2> m_NegatedPower2() {
557 return cst_pred_ty<is_negated_power2>();
558}
559inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) {
560 return V;
561}
562
563struct is_power2_or_zero {
564 bool isValue(const APInt &C) { return !C || C.isPowerOf2(); }
565};
566/// Match an integer or vector of 0 or power-of-2 values.
567/// For vectors, this includes constants with undefined elements.
568inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() {
569 return cst_pred_ty<is_power2_or_zero>();
570}
571inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) {
572 return V;
573}
574
575struct is_sign_mask {
576 bool isValue(const APInt &C) { return C.isSignMask(); }
577};
578/// Match an integer or vector with only the sign bit(s) set.
579/// For vectors, this includes constants with undefined elements.
580inline cst_pred_ty<is_sign_mask> m_SignMask() {
581 return cst_pred_ty<is_sign_mask>();
582}
583
584struct is_lowbit_mask {
585 bool isValue(const APInt &C) { return C.isMask(); }
586};
587/// Match an integer or vector with only the low bit(s) set.
588/// For vectors, this includes constants with undefined elements.
589inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() {
590 return cst_pred_ty<is_lowbit_mask>();
591}
592inline api_pred_ty<is_lowbit_mask> m_LowBitMask(const APInt *&V) {
593 return V;
594}
595
596struct icmp_pred_with_threshold {
597 ICmpInst::Predicate Pred;
598 const APInt *Thr;
599 bool isValue(const APInt &C) { return ICmpInst::compare(C, *Thr, Pred); }
600};
601/// Match an integer or vector with every element comparing 'pred' (eg/ne/...)
602/// to Threshold. For vectors, this includes constants with undefined elements.
603inline cst_pred_ty<icmp_pred_with_threshold>
604m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) {
605 cst_pred_ty<icmp_pred_with_threshold> P;
606 P.Pred = Predicate;
607 P.Thr = &Threshold;
608 return P;
609}
610
611struct is_nan {
612 bool isValue(const APFloat &C) { return C.isNaN(); }
613};
614/// Match an arbitrary NaN constant. This includes quiet and signalling nans.
615/// For vectors, this includes constants with undefined elements.
616inline cstfp_pred_ty<is_nan> m_NaN() {
617 return cstfp_pred_ty<is_nan>();
618}
619
620struct is_nonnan {
621 bool isValue(const APFloat &C) { return !C.isNaN(); }
622};
623/// Match a non-NaN FP constant.
624/// For vectors, this includes constants with undefined elements.
625inline cstfp_pred_ty<is_nonnan> m_NonNaN() {
626 return cstfp_pred_ty<is_nonnan>();
627}
628
629struct is_inf {
630 bool isValue(const APFloat &C) { return C.isInfinity(); }
631};
632/// Match a positive or negative infinity FP constant.
633/// For vectors, this includes constants with undefined elements.
634inline cstfp_pred_ty<is_inf> m_Inf() {
635 return cstfp_pred_ty<is_inf>();
636}
637
638struct is_noninf {
639 bool isValue(const APFloat &C) { return !C.isInfinity(); }
640};
641/// Match a non-infinity FP constant, i.e. finite or NaN.
642/// For vectors, this includes constants with undefined elements.
643inline cstfp_pred_ty<is_noninf> m_NonInf() {
644 return cstfp_pred_ty<is_noninf>();
645}
646
647struct is_finite {
648 bool isValue(const APFloat &C) { return C.isFinite(); }
649};
650/// Match a finite FP constant, i.e. not infinity or NaN.
651/// For vectors, this includes constants with undefined elements.
652inline cstfp_pred_ty<is_finite> m_Finite() {
653 return cstfp_pred_ty<is_finite>();
654}
655inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; }
656
657struct is_finitenonzero {
658 bool isValue(const APFloat &C) { return C.isFiniteNonZero(); }
659};
660/// Match a finite non-zero FP constant.
661/// For vectors, this includes constants with undefined elements.
662inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() {
663 return cstfp_pred_ty<is_finitenonzero>();
664}
665inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) {
666 return V;
667}
668
669struct is_any_zero_fp {
670 bool isValue(const APFloat &C) { return C.isZero(); }
671};
672/// Match a floating-point negative zero or positive zero.
673/// For vectors, this includes constants with undefined elements.
674inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() {
675 return cstfp_pred_ty<is_any_zero_fp>();
676}
677
678struct is_pos_zero_fp {
679 bool isValue(const APFloat &C) { return C.isPosZero(); }
680};
681/// Match a floating-point positive zero.
682/// For vectors, this includes constants with undefined elements.
683inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() {
684 return cstfp_pred_ty<is_pos_zero_fp>();
685}
686
687struct is_neg_zero_fp {
688 bool isValue(const APFloat &C) { return C.isNegZero(); }
689};
690/// Match a floating-point negative zero.
691/// For vectors, this includes constants with undefined elements.
692inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() {
693 return cstfp_pred_ty<is_neg_zero_fp>();
694}
695
696struct is_non_zero_fp {
697 bool isValue(const APFloat &C) { return C.isNonZero(); }
698};
699/// Match a floating-point non-zero.
700/// For vectors, this includes constants with undefined elements.
701inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() {
702 return cstfp_pred_ty<is_non_zero_fp>();
703}
704
705///////////////////////////////////////////////////////////////////////////////
706
707template <typename Class> struct bind_ty {
708 Class *&VR;
709
710 bind_ty(Class *&V) : VR(V) {}
711
712 template <typename ITy> bool match(ITy *V) {
713 if (auto *CV = dyn_cast<Class>(V)) {
714 VR = CV;
715 return true;
716 }
717 return false;
718 }
719};
720
721/// Match a value, capturing it if we match.
722inline bind_ty<Value> m_Value(Value *&V) { return V; }
6
Calling constructor for 'bind_ty<llvm::Value>'
7
Returning from constructor for 'bind_ty<llvm::Value>'
8
Returning without writing to 'V'
47
Returning without writing to 'V'
723inline bind_ty<const Value> m_Value(const Value *&V) { return V; }
724
725/// Match an instruction, capturing it if we match.
726inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; }
727/// Match a unary operator, capturing it if we match.
728inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; }
729/// Match a binary operator, capturing it if we match.
730inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; }
731/// Match a with overflow intrinsic, capturing it if we match.
732inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; }
733inline bind_ty<const WithOverflowInst>
734m_WithOverflowInst(const WithOverflowInst *&I) {
735 return I;
736}
737
738/// Match a Constant, capturing the value if we match.
739inline bind_ty<Constant> m_Constant(Constant *&C) { return C; }
740
741/// Match a ConstantInt, capturing the value if we match.
742inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; }
743
744/// Match a ConstantFP, capturing the value if we match.
745inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; }
746
747/// Match a ConstantExpr, capturing the value if we match.
748inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; }
749
750/// Match a basic block value, capturing it if we match.
751inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; }
752inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) {
753 return V;
754}
755
756/// Match an arbitrary immediate Constant and ignore it.
757inline match_combine_and<class_match<Constant>,
758 match_unless<class_match<ConstantExpr>>>
759m_ImmConstant() {
760 return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr()));
761}
762
763/// Match an immediate Constant, capturing the value if we match.
764inline match_combine_and<bind_ty<Constant>,
765 match_unless<class_match<ConstantExpr>>>
766m_ImmConstant(Constant *&C) {
767 return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr()));
768}
769
770/// Match a specified Value*.
771struct specificval_ty {
772 const Value *Val;
773
774 specificval_ty(const Value *V) : Val(V) {}
775
776 template <typename ITy> bool match(ITy *V) { return V == Val; }
777};
778
779/// Match if we have a specific specified value.
780inline specificval_ty m_Specific(const Value *V) { return V; }
781
782/// Stores a reference to the Value *, not the Value * itself,
783/// thus can be used in commutative matchers.
784template <typename Class> struct deferredval_ty {
785 Class *const &Val;
786
787 deferredval_ty(Class *const &V) : Val(V) {}
788
789 template <typename ITy> bool match(ITy *const V) { return V == Val; }
790};
791
792/// Like m_Specific(), but works if the specific value to match is determined
793/// as part of the same match() expression. For example:
794/// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will
795/// bind X before the pattern match starts.
796/// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against
797/// whichever value m_Value(X) populated.
798inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; }
799inline deferredval_ty<const Value> m_Deferred(const Value *const &V) {
800 return V;
801}
802
803/// Match a specified floating point value or vector of all elements of
804/// that value.
805struct specific_fpval {
806 double Val;
807
808 specific_fpval(double V) : Val(V) {}
809
810 template <typename ITy> bool match(ITy *V) {
811 if (const auto *CFP = dyn_cast<ConstantFP>(V))
812 return CFP->isExactlyValue(Val);
813 if (V->getType()->isVectorTy())
814 if (const auto *C = dyn_cast<Constant>(V))
815 if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
816 return CFP->isExactlyValue(Val);
817 return false;
818 }
819};
820
821/// Match a specific floating point value or vector with all elements
822/// equal to the value.
823inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); }
824
825/// Match a float 1.0 or vector with all elements equal to 1.0.
826inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); }
827
828struct bind_const_intval_ty {
829 uint64_t &VR;
830
831 bind_const_intval_ty(uint64_t &V) : VR(V) {}
832
833 template <typename ITy> bool match(ITy *V) {
834 if (const auto *CV = dyn_cast<ConstantInt>(V))
835 if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) {
836 VR = CV->getZExtValue();
837 return true;
838 }
839 return false;
840 }
841};
842
843/// Match a specified integer value or vector of all elements of that
844/// value.
845template <bool AllowUndefs>
846struct specific_intval {
847 APInt Val;
848
849 specific_intval(APInt V) : Val(std::move(V)) {}
850
851 template <typename ITy> bool match(ITy *V) {
852 const auto *CI = dyn_cast<ConstantInt>(V);
853 if (!CI && V->getType()->isVectorTy())
854 if (const auto *C = dyn_cast<Constant>(V))
855 CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs));
856
857 return CI && APInt::isSameValue(CI->getValue(), Val);
858 }
859};
860
861/// Match a specific integer value or vector with all elements equal to
862/// the value.
863inline specific_intval<false> m_SpecificInt(APInt V) {
864 return specific_intval<false>(std::move(V));
865}
866
867inline specific_intval<false> m_SpecificInt(uint64_t V) {
868 return m_SpecificInt(APInt(64, V));
869}
870
871inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) {
872 return specific_intval<true>(std::move(V));
873}
874
875inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) {
876 return m_SpecificIntAllowUndef(APInt(64, V));
877}
878
879/// Match a ConstantInt and bind to its value. This does not match
880/// ConstantInts wider than 64-bits.
881inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; }
882
883/// Match a specified basic block value.
884struct specific_bbval {
885 BasicBlock *Val;
886
887 specific_bbval(BasicBlock *Val) : Val(Val) {}
888
889 template <typename ITy> bool match(ITy *V) {
890 const auto *BB = dyn_cast<BasicBlock>(V);
891 return BB && BB == Val;
892 }
893};
894
895/// Match a specific basic block value.
896inline specific_bbval m_SpecificBB(BasicBlock *BB) {
897 return specific_bbval(BB);
898}
899
900/// A commutative-friendly version of m_Specific().
901inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) {
902 return BB;
903}
904inline deferredval_ty<const BasicBlock>
905m_Deferred(const BasicBlock *const &BB) {
906 return BB;
907}
908
909//===----------------------------------------------------------------------===//
910// Matcher for any binary operator.
911//
912template <typename LHS_t, typename RHS_t, bool Commutable = false>
913struct AnyBinaryOp_match {
914 LHS_t L;
915 RHS_t R;
916
917 // The evaluation order is always stable, regardless of Commutability.
918 // The LHS is always matched first.
919 AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
920
921 template <typename OpTy> bool match(OpTy *V) {
922 if (auto *I
15.1
'I' is non-null
15.1
'I' is non-null
= dyn_cast<BinaryOperator>(V))
15
Assuming 'V' is a 'BinaryOperator'
923 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
16
Calling 'match_combine_or::match'
33
Returning from 'match_combine_or::match'
34
Returning the value 1, which participates in a condition later
924 (Commutable && L.match(I->getOperand(1)) &&
925 R.match(I->getOperand(0)));
926 return false;
927 }
928};
929
930template <typename LHS, typename RHS>
931inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) {
932 return AnyBinaryOp_match<LHS, RHS>(L, R);
933}
934
935//===----------------------------------------------------------------------===//
936// Matcher for any unary operator.
937// TODO fuse unary, binary matcher into n-ary matcher
938//
939template <typename OP_t> struct AnyUnaryOp_match {
940 OP_t X;
941
942 AnyUnaryOp_match(const OP_t &X) : X(X) {}
943
944 template <typename OpTy> bool match(OpTy *V) {
945 if (auto *I = dyn_cast<UnaryOperator>(V))
946 return X.match(I->getOperand(0));
947 return false;
948 }
949};
950
951template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) {
952 return AnyUnaryOp_match<OP_t>(X);
953}
954
955//===----------------------------------------------------------------------===//
956// Matchers for specific binary operators.
957//
958
959template <typename LHS_t, typename RHS_t, unsigned Opcode,
960 bool Commutable = false>
961struct BinaryOp_match {
962 LHS_t L;
963 RHS_t R;
964
965 // The evaluation order is always stable, regardless of Commutability.
966 // The LHS is always matched first.
967 BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
968
969 template <typename OpTy> inline bool match(unsigned Opc, OpTy *V) {
970 if (V->getValueID() == Value::InstructionVal + Opc) {
21
Assuming the condition is true
22
Taking true branch
971 auto *I = cast<BinaryOperator>(V);
23
'V' is a 'BinaryOperator'
972 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
24
Value assigned to 'Select'
25
Assuming the condition is true
26
Assuming the condition is true
973 (Commutable && L.match(I->getOperand(1)) &&
974 R.match(I->getOperand(0)));
975 }
976 if (auto *CE = dyn_cast<ConstantExpr>(V))
977 return CE->getOpcode() == Opc &&
978 ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) ||
979 (Commutable && L.match(CE->getOperand(1)) &&
980 R.match(CE->getOperand(0))));
981 return false;
982 }
983
984 template <typename OpTy> bool match(OpTy *V) { return match(Opcode, V); }
20
Calling 'BinaryOp_match::match'
27
Returning from 'BinaryOp_match::match'
985};
986
987template <typename LHS, typename RHS>
988inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L,
989 const RHS &R) {
990 return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R);
991}
992
993template <typename LHS, typename RHS>
994inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L,
995 const RHS &R) {
996 return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R);
997}
998
999template <typename LHS, typename RHS>
1000inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L,
1001 const RHS &R) {
1002 return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R);
1003}
1004
1005template <typename LHS, typename RHS>
1006inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L,
1007 const RHS &R) {
1008 return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R);
1009}
1010
1011template <typename Op_t> struct FNeg_match {
1012 Op_t X;
1013
1014 FNeg_match(const Op_t &Op) : X(Op) {}
1015 template <typename OpTy> bool match(OpTy *V) {
1016 auto *FPMO = dyn_cast<FPMathOperator>(V);
1017 if (!FPMO) return false;
1018
1019 if (FPMO->getOpcode() == Instruction::FNeg)
1020 return X.match(FPMO->getOperand(0));
1021
1022 if (FPMO->getOpcode() == Instruction::FSub) {
1023 if (FPMO->hasNoSignedZeros()) {
1024 // With 'nsz', any zero goes.
1025 if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0)))
1026 return false;
1027 } else {
1028 // Without 'nsz', we need fsub -0.0, X exactly.
1029 if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0)))
1030 return false;
1031 }
1032
1033 return X.match(FPMO->getOperand(1));
1034 }
1035
1036 return false;
1037 }
1038};
1039
1040/// Match 'fneg X' as 'fsub -0.0, X'.
1041template <typename OpTy>
1042inline FNeg_match<OpTy>
1043m_FNeg(const OpTy &X) {
1044 return FNeg_match<OpTy>(X);
1045}
1046
1047/// Match 'fneg X' as 'fsub +-0.0, X'.
1048template <typename RHS>
1049inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub>
1050m_FNegNSZ(const RHS &X) {
1051 return m_FSub(m_AnyZeroFP(), X);
1052}
1053
1054template <typename LHS, typename RHS>
1055inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L,
1056 const RHS &R) {
1057 return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R);
1058}
1059
1060template <typename LHS, typename RHS>
1061inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L,
1062 const RHS &R) {
1063 return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R);
1064}
1065
1066template <typename LHS, typename RHS>
1067inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L,
1068 const RHS &R) {
1069 return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R);
1070}
1071
1072template <typename LHS, typename RHS>
1073inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L,
1074 const RHS &R) {
1075 return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R);
1076}
1077
1078template <typename LHS, typename RHS>
1079inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L,
1080 const RHS &R) {
1081 return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R);
1082}
1083
1084template <typename LHS, typename RHS>
1085inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L,
1086 const RHS &R) {
1087 return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R);
1088}
1089
1090template <typename LHS, typename RHS>
1091inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L,
1092 const RHS &R) {
1093 return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R);
1094}
1095
1096template <typename LHS, typename RHS>
1097inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L,
1098 const RHS &R) {
1099 return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R);
1100}
1101
1102template <typename LHS, typename RHS>
1103inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L,
1104 const RHS &R) {
1105 return BinaryOp_match<LHS, RHS, Instruction::And>(L, R);
1106}
1107
1108template <typename LHS, typename RHS>
1109inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L,
1110 const RHS &R) {
1111 return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R);
1112}
1113
1114template <typename LHS, typename RHS>
1115inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L,
1116 const RHS &R) {
1117 return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R);
1118}
1119
1120template <typename LHS, typename RHS>
1121inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L,
1122 const RHS &R) {
1123 return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R);
1124}
1125
1126template <typename LHS, typename RHS>
1127inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L,
1128 const RHS &R) {
1129 return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R);
1130}
1131
1132template <typename LHS, typename RHS>
1133inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L,
1134 const RHS &R) {
1135 return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R);
1136}
1137
1138template <typename LHS_t, typename RHS_t, unsigned Opcode,
1139 unsigned WrapFlags = 0>
1140struct OverflowingBinaryOp_match {
1141 LHS_t L;
1142 RHS_t R;
1143
1144 OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS)
1145 : L(LHS), R(RHS) {}
1146
1147 template <typename OpTy> bool match(OpTy *V) {
1148 if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) {
1149 if (Op->getOpcode() != Opcode)
1150 return false;
1151 if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) &&
1152 !Op->hasNoUnsignedWrap())
1153 return false;
1154 if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) &&
1155 !Op->hasNoSignedWrap())
1156 return false;
1157 return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1));
1158 }
1159 return false;
1160 }
1161};
1162
1163template <typename LHS, typename RHS>
1164inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1165 OverflowingBinaryOperator::NoSignedWrap>
1166m_NSWAdd(const LHS &L, const RHS &R) {
1167 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1168 OverflowingBinaryOperator::NoSignedWrap>(
1169 L, R);
1170}
1171template <typename LHS, typename RHS>
1172inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1173 OverflowingBinaryOperator::NoSignedWrap>
1174m_NSWSub(const LHS &L, const RHS &R) {
1175 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1176 OverflowingBinaryOperator::NoSignedWrap>(
1177 L, R);
1178}
1179template <typename LHS, typename RHS>
1180inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1181 OverflowingBinaryOperator::NoSignedWrap>
1182m_NSWMul(const LHS &L, const RHS &R) {
1183 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1184 OverflowingBinaryOperator::NoSignedWrap>(
1185 L, R);
1186}
1187template <typename LHS, typename RHS>
1188inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1189 OverflowingBinaryOperator::NoSignedWrap>
1190m_NSWShl(const LHS &L, const RHS &R) {
1191 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1192 OverflowingBinaryOperator::NoSignedWrap>(
1193 L, R);
1194}
1195
1196template <typename LHS, typename RHS>
1197inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1198 OverflowingBinaryOperator::NoUnsignedWrap>
1199m_NUWAdd(const LHS &L, const RHS &R) {
1200 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1201 OverflowingBinaryOperator::NoUnsignedWrap>(
1202 L, R);
1203}
1204template <typename LHS, typename RHS>
1205inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1206 OverflowingBinaryOperator::NoUnsignedWrap>
1207m_NUWSub(const LHS &L, const RHS &R) {
1208 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1209 OverflowingBinaryOperator::NoUnsignedWrap>(
1210 L, R);
1211}
1212template <typename LHS, typename RHS>
1213inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1214 OverflowingBinaryOperator::NoUnsignedWrap>
1215m_NUWMul(const LHS &L, const RHS &R) {
1216 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1217 OverflowingBinaryOperator::NoUnsignedWrap>(
1218 L, R);
1219}
1220template <typename LHS, typename RHS>
1221inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1222 OverflowingBinaryOperator::NoUnsignedWrap>
1223m_NUWShl(const LHS &L, const RHS &R) {
1224 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1225 OverflowingBinaryOperator::NoUnsignedWrap>(
1226 L, R);
1227}
1228
1229template <typename LHS_t, typename RHS_t, bool Commutable = false>
1230struct SpecificBinaryOp_match
1231 : public BinaryOp_match<LHS_t, RHS_t, 0, Commutable> {
1232 unsigned Opcode;
1233
1234 SpecificBinaryOp_match(unsigned Opcode, const LHS_t &LHS, const RHS_t &RHS)
1235 : BinaryOp_match<LHS_t, RHS_t, 0, Commutable>(LHS, RHS), Opcode(Opcode) {}
1236
1237 template <typename OpTy> bool match(OpTy *V) {
1238 return BinaryOp_match<LHS_t, RHS_t, 0, Commutable>::match(Opcode, V);
1239 }
1240};
1241
1242/// Matches a specific opcode.
1243template <typename LHS, typename RHS>
1244inline SpecificBinaryOp_match<LHS, RHS> m_BinOp(unsigned Opcode, const LHS &L,
1245 const RHS &R) {
1246 return SpecificBinaryOp_match<LHS, RHS>(Opcode, L, R);
1247}
1248
1249//===----------------------------------------------------------------------===//
1250// Class that matches a group of binary opcodes.
1251//
1252template <typename LHS_t, typename RHS_t, typename Predicate>
1253struct BinOpPred_match : Predicate {
1254 LHS_t L;
1255 RHS_t R;
1256
1257 BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1258
1259 template <typename OpTy> bool match(OpTy *V) {
1260 if (auto *I = dyn_cast<Instruction>(V))
1261 return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) &&
1262 R.match(I->getOperand(1));
1263 if (auto *CE = dyn_cast<ConstantExpr>(V))
1264 return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) &&
1265 R.match(CE->getOperand(1));
1266 return false;
1267 }
1268};
1269
1270struct is_shift_op {
1271 bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); }
1272};
1273
1274struct is_right_shift_op {
1275 bool isOpType(unsigned Opcode) {
1276 return Opcode == Instruction::LShr || Opcode == Instruction::AShr;
1277 }
1278};
1279
1280struct is_logical_shift_op {
1281 bool isOpType(unsigned Opcode) {
1282 return Opcode == Instruction::LShr || Opcode == Instruction::Shl;
1283 }
1284};
1285
1286struct is_bitwiselogic_op {
1287 bool isOpType(unsigned Opcode) {
1288 return Instruction::isBitwiseLogicOp(Opcode);
1289 }
1290};
1291
1292struct is_idiv_op {
1293 bool isOpType(unsigned Opcode) {
1294 return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
1295 }
1296};
1297
1298struct is_irem_op {
1299 bool isOpType(unsigned Opcode) {
1300 return Opcode == Instruction::SRem || Opcode == Instruction::URem;
1301 }
1302};
1303
1304/// Matches shift operations.
1305template <typename LHS, typename RHS>
1306inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L,
1307 const RHS &R) {
1308 return BinOpPred_match<LHS, RHS, is_shift_op>(L, R);
1309}
1310
1311/// Matches logical shift operations.
1312template <typename LHS, typename RHS>
1313inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L,
1314 const RHS &R) {
1315 return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R);
1316}
1317
1318/// Matches logical shift operations.
1319template <typename LHS, typename RHS>
1320inline BinOpPred_match<LHS, RHS, is_logical_shift_op>
1321m_LogicalShift(const LHS &L, const RHS &R) {
1322 return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R);
1323}
1324
1325/// Matches bitwise logic operations.
1326template <typename LHS, typename RHS>
1327inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op>
1328m_BitwiseLogic(const LHS &L, const RHS &R) {
1329 return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R);
1330}
1331
1332/// Matches integer division operations.
1333template <typename LHS, typename RHS>
1334inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L,
1335 const RHS &R) {
1336 return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R);
1337}
1338
1339/// Matches integer remainder operations.
1340template <typename LHS, typename RHS>
1341inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L,
1342 const RHS &R) {
1343 return BinOpPred_match<LHS, RHS, is_irem_op>(L, R);
1344}
1345
1346//===----------------------------------------------------------------------===//
1347// Class that matches exact binary ops.
1348//
1349template <typename SubPattern_t> struct Exact_match {
1350 SubPattern_t SubPattern;
1351
1352 Exact_match(const SubPattern_t &SP) : SubPattern(SP) {}
1353
1354 template <typename OpTy> bool match(OpTy *V) {
1355 if (auto *PEO = dyn_cast<PossiblyExactOperator>(V))
1356 return PEO->isExact() && SubPattern.match(V);
1357 return false;
1358 }
1359};
1360
1361template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) {
1362 return SubPattern;
1363}
1364
1365//===----------------------------------------------------------------------===//
1366// Matchers for CmpInst classes
1367//
1368
1369template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy,
1370 bool Commutable = false>
1371struct CmpClass_match {
1372 PredicateTy &Predicate;
1373 LHS_t L;
1374 RHS_t R;
1375
1376 // The evaluation order is always stable, regardless of Commutability.
1377 // The LHS is always matched first.
1378 CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS)
1379 : Predicate(Pred), L(LHS), R(RHS) {}
1380
1381 template <typename OpTy> bool match(OpTy *V) {
1382 if (auto *I = dyn_cast<Class>(V)) {
1383 if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) {
1384 Predicate = I->getPredicate();
1385 return true;
1386 } else if (Commutable && L.match(I->getOperand(1)) &&
1387 R.match(I->getOperand(0))) {
1388 Predicate = I->getSwappedPredicate();
1389 return true;
1390 }
1391 }
1392 return false;
1393 }
1394};
1395
1396template <typename LHS, typename RHS>
1397inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>
1398m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1399 return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R);
1400}
1401
1402template <typename LHS, typename RHS>
1403inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>
1404m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1405 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R);
1406}
1407
1408template <typename LHS, typename RHS>
1409inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>
1410m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1411 return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R);
1412}
1413
1414//===----------------------------------------------------------------------===//
1415// Matchers for instructions with a given opcode and number of operands.
1416//
1417
1418/// Matches instructions with Opcode and three operands.
1419template <typename T0, unsigned Opcode> struct OneOps_match {
1420 T0 Op1;
1421
1422 OneOps_match(const T0 &Op1) : Op1(Op1) {}
1423
1424 template <typename OpTy> bool match(OpTy *V) {
1425 if (V->getValueID() == Value::InstructionVal + Opcode) {
1426 auto *I = cast<Instruction>(V);
1427 return Op1.match(I->getOperand(0));
1428 }
1429 return false;
1430 }
1431};
1432
1433/// Matches instructions with Opcode and three operands.
1434template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match {
1435 T0 Op1;
1436 T1 Op2;
1437
1438 TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {}
1439
1440 template <typename OpTy> bool match(OpTy *V) {
1441 if (V->getValueID() == Value::InstructionVal + Opcode) {
1442 auto *I = cast<Instruction>(V);
1443 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1));
1444 }
1445 return false;
1446 }
1447};
1448
1449/// Matches instructions with Opcode and three operands.
1450template <typename T0, typename T1, typename T2, unsigned Opcode>
1451struct ThreeOps_match {
1452 T0 Op1;
1453 T1 Op2;
1454 T2 Op3;
1455
1456 ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3)
1457 : Op1(Op1), Op2(Op2), Op3(Op3) {}
1458
1459 template <typename OpTy> bool match(OpTy *V) {
1460 if (V->getValueID() == Value::InstructionVal + Opcode) {
67
Called C++ object pointer is null
1461 auto *I = cast<Instruction>(V);
1462 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1463 Op3.match(I->getOperand(2));
1464 }
1465 return false;
1466 }
1467};
1468
1469/// Matches SelectInst.
1470template <typename Cond, typename LHS, typename RHS>
1471inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select>
1472m_Select(const Cond &C, const LHS &L, const RHS &R) {
1473 return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R);
1474}
1475
1476/// This matches a select of two constants, e.g.:
1477/// m_SelectCst<-1, 0>(m_Value(V))
1478template <int64_t L, int64_t R, typename Cond>
1479inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>,
1480 Instruction::Select>
1481m_SelectCst(const Cond &C) {
1482 return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>());
1483}
1484
1485/// Matches FreezeInst.
1486template <typename OpTy>
1487inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) {
1488 return OneOps_match<OpTy, Instruction::Freeze>(Op);
1489}
1490
1491/// Matches InsertElementInst.
1492template <typename Val_t, typename Elt_t, typename Idx_t>
1493inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>
1494m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) {
1495 return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>(
1496 Val, Elt, Idx);
1497}
1498
1499/// Matches ExtractElementInst.
1500template <typename Val_t, typename Idx_t>
1501inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>
1502m_ExtractElt(const Val_t &Val, const Idx_t &Idx) {
1503 return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx);
1504}
1505
1506/// Matches shuffle.
1507template <typename T0, typename T1, typename T2> struct Shuffle_match {
1508 T0 Op1;
1509 T1 Op2;
1510 T2 Mask;
1511
1512 Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask)
1513 : Op1(Op1), Op2(Op2), Mask(Mask) {}
1514
1515 template <typename OpTy> bool match(OpTy *V) {
1516 if (auto *I = dyn_cast<ShuffleVectorInst>(V)) {
1517 return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1518 Mask.match(I->getShuffleMask());
1519 }
1520 return false;
1521 }
1522};
1523
1524struct m_Mask {
1525 ArrayRef<int> &MaskRef;
1526 m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1527 bool match(ArrayRef<int> Mask) {
1528 MaskRef = Mask;
1529 return true;
1530 }
1531};
1532
1533struct m_ZeroMask {
1534 bool match(ArrayRef<int> Mask) {
1535 return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; });
1536 }
1537};
1538
1539struct m_SpecificMask {
1540 ArrayRef<int> &MaskRef;
1541 m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1542 bool match(ArrayRef<int> Mask) { return MaskRef == Mask; }
1543};
1544
1545struct m_SplatOrUndefMask {
1546 int &SplatIndex;
1547 m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {}
1548 bool match(ArrayRef<int> Mask) {
1549 auto First = find_if(Mask, [](int Elem) { return Elem != -1; });
1550 if (First == Mask.end())
1551 return false;
1552 SplatIndex = *First;
1553 return all_of(Mask,
1554 [First](int Elem) { return Elem == *First || Elem == -1; });
1555 }
1556};
1557
1558/// Matches ShuffleVectorInst independently of mask value.
1559template <typename V1_t, typename V2_t>
1560inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>
1561m_Shuffle(const V1_t &v1, const V2_t &v2) {
1562 return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2);
1563}
1564
1565template <typename V1_t, typename V2_t, typename Mask_t>
1566inline Shuffle_match<V1_t, V2_t, Mask_t>
1567m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) {
1568 return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask);
1569}
1570
1571/// Matches LoadInst.
1572template <typename OpTy>
1573inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) {
1574 return OneOps_match<OpTy, Instruction::Load>(Op);
1575}
1576
1577/// Matches StoreInst.
1578template <typename ValueOpTy, typename PointerOpTy>
1579inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>
1580m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) {
1581 return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp,
1582 PointerOp);
1583}
1584
1585//===----------------------------------------------------------------------===//
1586// Matchers for CastInst classes
1587//
1588
1589template <typename Op_t, unsigned Opcode> struct CastClass_match {
1590 Op_t Op;
1591
1592 CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
1593
1594 template <typename OpTy> bool match(OpTy *V) {
1595 if (auto *O = dyn_cast<Operator>(V))
1596 return O->getOpcode() == Opcode && Op.match(O->getOperand(0));
1597 return false;
1598 }
1599};
1600
1601/// Matches BitCast.
1602template <typename OpTy>
1603inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) {
1604 return CastClass_match<OpTy, Instruction::BitCast>(Op);
1605}
1606
1607/// Matches PtrToInt.
1608template <typename OpTy>
1609inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) {
1610 return CastClass_match<OpTy, Instruction::PtrToInt>(Op);
1611}
1612
1613/// Matches IntToPtr.
1614template <typename OpTy>
1615inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) {
1616 return CastClass_match<OpTy, Instruction::IntToPtr>(Op);
1617}
1618
1619/// Matches Trunc.
1620template <typename OpTy>
1621inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) {
1622 return CastClass_match<OpTy, Instruction::Trunc>(Op);
1623}
1624
1625template <typename OpTy>
1626inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy>
1627m_TruncOrSelf(const OpTy &Op) {
1628 return m_CombineOr(m_Trunc(Op), Op);
1629}
1630
1631/// Matches SExt.
1632template <typename OpTy>
1633inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) {
1634 return CastClass_match<OpTy, Instruction::SExt>(Op);
1635}
1636
1637/// Matches ZExt.
1638template <typename OpTy>
1639inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) {
1640 return CastClass_match<OpTy, Instruction::ZExt>(Op);
51
Returning without writing to 'Op.VR'
1641}
1642
1643template <typename OpTy>
1644inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy>
1645m_ZExtOrSelf(const OpTy &Op) {
1646 return m_CombineOr(m_ZExt(Op), Op);
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'
1647}
1648
1649template <typename OpTy>
1650inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy>
1651m_SExtOrSelf(const OpTy &Op) {
1652 return m_CombineOr(m_SExt(Op), Op);
1653}
1654
1655template <typename OpTy>
1656inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1657 CastClass_match<OpTy, Instruction::SExt>>
1658m_ZExtOrSExt(const OpTy &Op) {
1659 return m_CombineOr(m_ZExt(Op), m_SExt(Op));
1660}
1661
1662template <typename OpTy>
1663inline match_combine_or<
1664 match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1665 CastClass_match<OpTy, Instruction::SExt>>,
1666 OpTy>
1667m_ZExtOrSExtOrSelf(const OpTy &Op) {
1668 return m_CombineOr(m_ZExtOrSExt(Op), Op);
1669}
1670
1671template <typename OpTy>
1672inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) {
1673 return CastClass_match<OpTy, Instruction::UIToFP>(Op);
1674}
1675
1676template <typename OpTy>
1677inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
1678 return CastClass_match<OpTy, Instruction::SIToFP>(Op);
1679}
1680
1681template <typename OpTy>
1682inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) {
1683 return CastClass_match<OpTy, Instruction::FPToUI>(Op);
1684}
1685
1686template <typename OpTy>
1687inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) {
1688 return CastClass_match<OpTy, Instruction::FPToSI>(Op);
1689}
1690
1691template <typename OpTy>
1692inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
1693 return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
1694}
1695
1696template <typename OpTy>
1697inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
1698 return CastClass_match<OpTy, Instruction::FPExt>(Op);
1699}
1700
1701//===----------------------------------------------------------------------===//
1702// Matchers for control flow.
1703//
1704
1705struct br_match {
1706 BasicBlock *&Succ;
1707
1708 br_match(BasicBlock *&Succ) : Succ(Succ) {}
1709
1710 template <typename OpTy> bool match(OpTy *V) {
1711 if (auto *BI = dyn_cast<BranchInst>(V))
1712 if (BI->isUnconditional()) {
1713 Succ = BI->getSuccessor(0);
1714 return true;
1715 }
1716 return false;
1717 }
1718};
1719
1720inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); }
1721
1722template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1723struct brc_match {
1724 Cond_t Cond;
1725 TrueBlock_t T;
1726 FalseBlock_t F;
1727
1728 brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f)
1729 : Cond(C), T(t), F(f) {}
1730
1731 template <typename OpTy> bool match(OpTy *V) {
1732 if (auto *BI = dyn_cast<BranchInst>(V))
1733 if (BI->isConditional() && Cond.match(BI->getCondition()))
1734 return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1));
1735 return false;
1736 }
1737};
1738
1739template <typename Cond_t>
1740inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>
1741m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) {
1742 return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>(
1743 C, m_BasicBlock(T), m_BasicBlock(F));
1744}
1745
1746template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1747inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t>
1748m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) {
1749 return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F);
1750}
1751
1752//===----------------------------------------------------------------------===//
1753// Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y).
1754//
1755
1756template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t,
1757 bool Commutable = false>
1758struct MaxMin_match {
1759 using PredType = Pred_t;
1760 LHS_t L;
1761 RHS_t R;
1762
1763 // The evaluation order is always stable, regardless of Commutability.
1764 // The LHS is always matched first.
1765 MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1766
1767 template <typename OpTy> bool match(OpTy *V) {
1768 if (auto *II = dyn_cast<IntrinsicInst>(V)) {
1769 Intrinsic::ID IID = II->getIntrinsicID();
1770 if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) ||
1771 (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) ||
1772 (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) ||
1773 (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) {
1774 Value *LHS = II->getOperand(0), *RHS = II->getOperand(1);
1775 return (L.match(LHS) && R.match(RHS)) ||
1776 (Commutable && L.match(RHS) && R.match(LHS));
1777 }
1778 }
1779 // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x".
1780 auto *SI = dyn_cast<SelectInst>(V);
1781 if (!SI)
1782 return false;
1783 auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition());
1784 if (!Cmp)
1785 return false;
1786 // At this point we have a select conditioned on a comparison. Check that
1787 // it is the values returned by the select that are being compared.
1788 auto *TrueVal = SI->getTrueValue();
1789 auto *FalseVal = SI->getFalseValue();
1790 auto *LHS = Cmp->getOperand(0);
1791 auto *RHS = Cmp->getOperand(1);
1792 if ((TrueVal != LHS || FalseVal != RHS) &&
1793 (TrueVal != RHS || FalseVal != LHS))
1794 return false;
1795 typename CmpInst_t::Predicate Pred =
1796 LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate();
1797 // Does "(x pred y) ? x : y" represent the desired max/min operation?
1798 if (!Pred_t::match(Pred))
1799 return false;
1800 // It does! Bind the operands.
1801 return (L.match(LHS) && R.match(RHS)) ||
1802 (Commutable && L.match(RHS) && R.match(LHS));
1803 }
1804};
1805
1806/// Helper class for identifying signed max predicates.
1807struct smax_pred_ty {
1808 static bool match(ICmpInst::Predicate Pred) {
1809 return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE;
1810 }
1811};
1812
1813/// Helper class for identifying signed min predicates.
1814struct smin_pred_ty {
1815 static bool match(ICmpInst::Predicate Pred) {
1816 return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE;
1817 }
1818};
1819
1820/// Helper class for identifying unsigned max predicates.
1821struct umax_pred_ty {
1822 static bool match(ICmpInst::Predicate Pred) {
1823 return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE;
1824 }
1825};
1826
1827/// Helper class for identifying unsigned min predicates.
1828struct umin_pred_ty {
1829 static bool match(ICmpInst::Predicate Pred) {
1830 return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE;
1831 }
1832};
1833
1834/// Helper class for identifying ordered max predicates.
1835struct ofmax_pred_ty {
1836 static bool match(FCmpInst::Predicate Pred) {
1837 return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE;
1838 }
1839};
1840
1841/// Helper class for identifying ordered min predicates.
1842struct ofmin_pred_ty {
1843 static bool match(FCmpInst::Predicate Pred) {
1844 return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE;
1845 }
1846};
1847
1848/// Helper class for identifying unordered max predicates.
1849struct ufmax_pred_ty {
1850 static bool match(FCmpInst::Predicate Pred) {
1851 return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE;
1852 }
1853};
1854
1855/// Helper class for identifying unordered min predicates.
1856struct ufmin_pred_ty {
1857 static bool match(FCmpInst::Predicate Pred) {
1858 return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE;
1859 }
1860};
1861
1862template <typename LHS, typename RHS>
1863inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L,
1864 const RHS &R) {
1865 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R);
1866}
1867
1868template <typename LHS, typename RHS>
1869inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L,
1870 const RHS &R) {
1871 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R);
1872}
1873
1874template <typename LHS, typename RHS>
1875inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L,
1876 const RHS &R) {
1877 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R);
1878}
1879
1880template <typename LHS, typename RHS>
1881inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L,
1882 const RHS &R) {
1883 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R);
1884}
1885
1886template <typename LHS, typename RHS>
1887inline match_combine_or<
1888 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>,
1889 MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>,
1890 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>,
1891 MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>>
1892m_MaxOrMin(const LHS &L, const RHS &R) {
1893 return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)),
1894 m_CombineOr(m_UMax(L, R), m_UMin(L, R)));
1895}
1896
1897/// Match an 'ordered' floating point maximum function.
1898/// Floating point has one special value 'NaN'. Therefore, there is no total
1899/// order. However, if we can ignore the 'NaN' value (for example, because of a
1900/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1901/// semantics. In the presence of 'NaN' we have to preserve the original
1902/// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate.
1903///
1904/// max(L, R) iff L and R are not NaN
1905/// m_OrdFMax(L, R) = R iff L or R are NaN
1906template <typename LHS, typename RHS>
1907inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L,
1908 const RHS &R) {
1909 return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R);
1910}
1911
1912/// Match an 'ordered' floating point minimum function.
1913/// Floating point has one special value 'NaN'. Therefore, there is no total
1914/// order. However, if we can ignore the 'NaN' value (for example, because of a
1915/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1916/// semantics. In the presence of 'NaN' we have to preserve the original
1917/// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate.
1918///
1919/// min(L, R) iff L and R are not NaN
1920/// m_OrdFMin(L, R) = R iff L or R are NaN
1921template <typename LHS, typename RHS>
1922inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L,
1923 const RHS &R) {
1924 return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R);
1925}
1926
1927/// Match an 'unordered' floating point maximum function.
1928/// Floating point has one special value 'NaN'. Therefore, there is no total
1929/// order. However, if we can ignore the 'NaN' value (for example, because of a
1930/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1931/// semantics. In the presence of 'NaN' we have to preserve the original
1932/// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate.
1933///
1934/// max(L, R) iff L and R are not NaN
1935/// m_UnordFMax(L, R) = L iff L or R are NaN
1936template <typename LHS, typename RHS>
1937inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>
1938m_UnordFMax(const LHS &L, const RHS &R) {
1939 return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R);
1940}
1941
1942/// Match an 'unordered' floating point minimum function.
1943/// Floating point has one special value 'NaN'. Therefore, there is no total
1944/// order. However, if we can ignore the 'NaN' value (for example, because of a
1945/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1946/// semantics. In the presence of 'NaN' we have to preserve the original
1947/// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate.
1948///
1949/// min(L, R) iff L and R are not NaN
1950/// m_UnordFMin(L, R) = L iff L or R are NaN
1951template <typename LHS, typename RHS>
1952inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>
1953m_UnordFMin(const LHS &L, const RHS &R) {
1954 return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R);
1955}
1956
1957//===----------------------------------------------------------------------===//
1958// Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b
1959// Note that S might be matched to other instructions than AddInst.
1960//
1961
1962template <typename LHS_t, typename RHS_t, typename Sum_t>
1963struct UAddWithOverflow_match {
1964 LHS_t L;
1965 RHS_t R;
1966 Sum_t S;
1967
1968 UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S)
1969 : L(L), R(R), S(S) {}
1970
1971 template <typename OpTy> bool match(OpTy *V) {
1972 Value *ICmpLHS, *ICmpRHS;
1973 ICmpInst::Predicate Pred;
1974 if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V))
1975 return false;
1976
1977 Value *AddLHS, *AddRHS;
1978 auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS));
1979
1980 // (a + b) u< a, (a + b) u< b
1981 if (Pred == ICmpInst::ICMP_ULT)
1982 if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS))
1983 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1984
1985 // a >u (a + b), b >u (a + b)
1986 if (Pred == ICmpInst::ICMP_UGT)
1987 if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS))
1988 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1989
1990 Value *Op1;
1991 auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes()));
1992 // (a ^ -1) <u b
1993 if (Pred == ICmpInst::ICMP_ULT) {
1994 if (XorExpr.match(ICmpLHS))
1995 return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS);
1996 }
1997 // b > u (a ^ -1)
1998 if (Pred == ICmpInst::ICMP_UGT) {
1999 if (XorExpr.match(ICmpRHS))
2000 return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS);
2001 }
2002
2003 // Match special-case for increment-by-1.
2004 if (Pred == ICmpInst::ICMP_EQ) {
2005 // (a + 1) == 0
2006 // (1 + a) == 0
2007 if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) &&
2008 (m_One().match(AddLHS) || m_One().match(AddRHS)))
2009 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
2010 // 0 == (a + 1)
2011 // 0 == (1 + a)
2012 if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) &&
2013 (m_One().match(AddLHS) || m_One().match(AddRHS)))
2014 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
2015 }
2016
2017 return false;
2018 }
2019};
2020
2021/// Match an icmp instruction checking for unsigned overflow on addition.
2022///
2023/// S is matched to the addition whose result is being checked for overflow, and
2024/// L and R are matched to the LHS and RHS of S.
2025template <typename LHS_t, typename RHS_t, typename Sum_t>
2026UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>
2027m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) {
2028 return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S);
2029}
2030
2031template <typename Opnd_t> struct Argument_match {
2032 unsigned OpI;
2033 Opnd_t Val;
2034
2035 Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {}
2036
2037 template <typename OpTy> bool match(OpTy *V) {
2038 // FIXME: Should likely be switched to use `CallBase`.
2039 if (const auto *CI = dyn_cast<CallInst>(V))
2040 return Val.match(CI->getArgOperand(OpI));
2041 return false;
2042 }
2043};
2044
2045/// Match an argument.
2046template <unsigned OpI, typename Opnd_t>
2047inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) {
2048 return Argument_match<Opnd_t>(OpI, Op);
2049}
2050
2051/// Intrinsic matchers.
2052struct IntrinsicID_match {
2053 unsigned ID;
2054
2055 IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {}
2056
2057 template <typename OpTy> bool match(OpTy *V) {
2058 if (const auto *CI = dyn_cast<CallInst>(V))
2059 if (const auto *F = CI->getCalledFunction())
2060 return F->getIntrinsicID() == ID;
2061 return false;
2062 }
2063};
2064
2065/// Intrinsic matches are combinations of ID matchers, and argument
2066/// matchers. Higher arity matcher are defined recursively in terms of and-ing
2067/// them with lower arity matchers. Here's some convenient typedefs for up to
2068/// several arguments, and more can be added as needed
2069template <typename T0 = void, typename T1 = void, typename T2 = void,
2070 typename T3 = void, typename T4 = void, typename T5 = void,
2071 typename T6 = void, typename T7 = void, typename T8 = void,
2072 typename T9 = void, typename T10 = void>
2073struct m_Intrinsic_Ty;
2074template <typename T0> struct m_Intrinsic_Ty<T0> {
2075 using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>;
2076};
2077template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> {
2078 using Ty =
2079 match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>;
2080};
2081template <typename T0, typename T1, typename T2>
2082struct m_Intrinsic_Ty<T0, T1, T2> {
2083 using Ty =
2084 match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty,
2085 Argument_match<T2>>;
2086};
2087template <typename T0, typename T1, typename T2, typename T3>
2088struct m_Intrinsic_Ty<T0, T1, T2, T3> {
2089 using Ty =
2090 match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty,
2091 Argument_match<T3>>;
2092};
2093
2094template <typename T0, typename T1, typename T2, typename T3, typename T4>
2095struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> {
2096 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty,
2097 Argument_match<T4>>;
2098};
2099
2100template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5>
2101struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> {
2102 using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty,
2103 Argument_match<T5>>;
2104};
2105
2106/// Match intrinsic calls like this:
2107/// m_Intrinsic<Intrinsic::fabs>(m_Value(X))
2108template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() {
2109 return IntrinsicID_match(IntrID);
2110}
2111
2112/// Matches MaskedLoad Intrinsic.
2113template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3>
2114inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty
2115m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2,
2116 const Opnd3 &Op3) {
2117 return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3);
2118}
2119
2120template <Intrinsic::ID IntrID, typename T0>
2121inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) {
2122 return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0));
2123}
2124
2125template <Intrinsic::ID IntrID, typename T0, typename T1>
2126inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0,
2127 const T1 &Op1) {
2128 return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1));
2129}
2130
2131template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2>
2132inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty
2133m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) {
2134 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2));
2135}
2136
2137template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2138 typename T3>
2139inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty
2140m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) {
2141 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3));
2142}
2143
2144template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2145 typename T3, typename T4>
2146inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty
2147m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2148 const T4 &Op4) {
2149 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3),
2150 m_Argument<4>(Op4));
2151}
2152
2153template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2154 typename T3, typename T4, typename T5>
2155inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty
2156m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2157 const T4 &Op4, const T5 &Op5) {
2158 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4),
2159 m_Argument<5>(Op5));
2160}
2161
2162// Helper intrinsic matching specializations.
2163template <typename Opnd0>
2164inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) {
2165 return m_Intrinsic<Intrinsic::bitreverse>(Op0);
2166}
2167
2168template <typename Opnd0>
2169inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) {
2170 return m_Intrinsic<Intrinsic::bswap>(Op0);
2171}
2172
2173template <typename Opnd0>
2174inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) {
2175 return m_Intrinsic<Intrinsic::fabs>(Op0);
2176}
2177
2178template <typename Opnd0>
2179inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) {
2180 return m_Intrinsic<Intrinsic::canonicalize>(Op0);
2181}
2182
2183template <typename Opnd0, typename Opnd1>
2184inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0,
2185 const Opnd1 &Op1) {
2186 return m_Intrinsic<Intrinsic::minnum>(Op0, Op1);
2187}
2188
2189template <typename Opnd0, typename Opnd1>
2190inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0,
2191 const Opnd1 &Op1) {
2192 return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1);
2193}
2194
2195template <typename Opnd0, typename Opnd1, typename Opnd2>
2196inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2197m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2198 return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2);
2199}
2200
2201template <typename Opnd0, typename Opnd1, typename Opnd2>
2202inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2203m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2204 return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2);
2205}
2206
2207//===----------------------------------------------------------------------===//
2208// Matchers for two-operands operators with the operators in either order
2209//
2210
2211/// Matches a BinaryOperator with LHS and RHS in either order.
2212template <typename LHS, typename RHS>
2213inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) {
2214 return AnyBinaryOp_match<LHS, RHS, true>(L, R);
11
Returning without writing to 'R.VR'
2215}
2216
2217/// Matches an ICmp with a predicate over LHS and RHS in either order.
2218/// Swaps the predicate if operands are commuted.
2219template <typename LHS, typename RHS>
2220inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>
2221m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
2222 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L,
2223 R);
2224}
2225
2226/// Matches a specific opcode with LHS and RHS in either order.
2227template <typename LHS, typename RHS>
2228inline SpecificBinaryOp_match<LHS, RHS, true>
2229m_c_BinOp(unsigned Opcode, const LHS &L, const RHS &R) {
2230 return SpecificBinaryOp_match<LHS, RHS, true>(Opcode, L, R);
2231}
2232
2233/// Matches a Add with LHS and RHS in either order.
2234template <typename LHS, typename RHS>
2235inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L,
2236 const RHS &R) {
2237 return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R);
2238}
2239
2240/// Matches a Mul with LHS and RHS in either order.
2241template <typename LHS, typename RHS>
2242inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L,
2243 const RHS &R) {
2244 return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R);
2245}
2246
2247/// Matches an And with LHS and RHS in either order.
2248template <typename LHS, typename RHS>
2249inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L,
2250 const RHS &R) {
2251 return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R);
2252}
2253
2254/// Matches an Or with LHS and RHS in either order.
2255template <typename LHS, typename RHS>
2256inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L,
2257 const RHS &R) {
2258 return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R);
2259}
2260
2261/// Matches an Xor with LHS and RHS in either order.
2262template <typename LHS, typename RHS>
2263inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L,
2264 const RHS &R) {
2265 return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R);
2266}
2267
2268/// Matches a 'Neg' as 'sub 0, V'.
2269template <typename ValTy>
2270inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub>
2271m_Neg(const ValTy &V) {
2272 return m_Sub(m_ZeroInt(), V);
2273}
2274
2275/// Matches a 'Neg' as 'sub nsw 0, V'.
2276template <typename ValTy>
2277inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy,
2278 Instruction::Sub,
2279 OverflowingBinaryOperator::NoSignedWrap>
2280m_NSWNeg(const ValTy &V) {
2281 return m_NSWSub(m_ZeroInt(), V);
2282}
2283
2284/// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
2285template <typename ValTy>
2286inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true>
2287m_Not(const ValTy &V) {
2288 return m_c_Xor(V, m_AllOnes());
2289}
2290
2291template <typename ValTy> struct NotForbidUndef_match {
2292 ValTy Val;
2293 NotForbidUndef_match(const ValTy &V) : Val(V) {}
2294
2295 template <typename OpTy> bool match(OpTy *V) {
2296 // We do not use m_c_Xor because that could match an arbitrary APInt that is
2297 // not -1 as C and then fail to match the other operand if it is -1.
2298 // This code should still work even when both operands are constants.
2299 Value *X;
2300 const APInt *C;
2301 if (m_Xor(m_Value(X), m_APIntForbidUndef(C)).match(V) && C->isAllOnes())
2302 return Val.match(X);
2303 if (m_Xor(m_APIntForbidUndef(C), m_Value(X)).match(V) && C->isAllOnes())
2304 return Val.match(X);
2305 return false;
2306 }
2307};
2308
2309/// Matches a bitwise 'not' as 'xor V, -1' or 'xor -1, V'. For vectors, the
2310/// constant value must be composed of only -1 scalar elements.
2311template <typename ValTy>
2312inline NotForbidUndef_match<ValTy> m_NotForbidUndef(const ValTy &V) {
2313 return NotForbidUndef_match<ValTy>(V);
2314}
2315
2316/// Matches an SMin with LHS and RHS in either order.
2317template <typename LHS, typename RHS>
2318inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>
2319m_c_SMin(const LHS &L, const RHS &R) {
2320 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R);
2321}
2322/// Matches an SMax with LHS and RHS in either order.
2323template <typename LHS, typename RHS>
2324inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>
2325m_c_SMax(const LHS &L, const RHS &R) {
2326 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R);
2327}
2328/// Matches a UMin with LHS and RHS in either order.
2329template <typename LHS, typename RHS>
2330inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>
2331m_c_UMin(const LHS &L, const RHS &R) {
2332 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R);
2333}
2334/// Matches a UMax with LHS and RHS in either order.
2335template <typename LHS, typename RHS>
2336inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>
2337m_c_UMax(const LHS &L, const RHS &R) {
2338 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R);
2339}
2340
2341template <typename LHS, typename RHS>
2342inline match_combine_or<
2343 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>,
2344 MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>,
2345 match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>,
2346 MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>>
2347m_c_MaxOrMin(const LHS &L, const RHS &R) {
2348 return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)),
2349 m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R)));
2350}
2351
2352/// Matches FAdd with LHS and RHS in either order.
2353template <typename LHS, typename RHS>
2354inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true>
2355m_c_FAdd(const LHS &L, const RHS &R) {
2356 return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R);
2357}
2358
2359/// Matches FMul with LHS and RHS in either order.
2360template <typename LHS, typename RHS>
2361inline BinaryOp_match<LHS, RHS, Instruction::FMul, true>
2362m_c_FMul(const LHS &L, const RHS &R) {
2363 return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R);
2364}
2365
2366template <typename Opnd_t> struct Signum_match {
2367 Opnd_t Val;
2368 Signum_match(const Opnd_t &V) : Val(V) {}
2369
2370 template <typename OpTy> bool match(OpTy *V) {
2371 unsigned TypeSize = V->getType()->getScalarSizeInBits();
2372 if (TypeSize == 0)
2373 return false;
2374
2375 unsigned ShiftWidth = TypeSize - 1;
2376 Value *OpL = nullptr, *OpR = nullptr;
2377
2378 // This is the representation of signum we match:
2379 //
2380 // signum(x) == (x >> 63) | (-x >>u 63)
2381 //
2382 // An i1 value is its own signum, so it's correct to match
2383 //
2384 // signum(x) == (x >> 0) | (-x >>u 0)
2385 //
2386 // for i1 values.
2387
2388 auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth));
2389 auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth));
2390 auto Signum = m_Or(LHS, RHS);
2391
2392 return Signum.match(V) && OpL == OpR && Val.match(OpL);
2393 }
2394};
2395
2396/// Matches a signum pattern.
2397///
2398/// signum(x) =
2399/// x > 0 -> 1
2400/// x == 0 -> 0
2401/// x < 0 -> -1
2402template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) {
2403 return Signum_match<Val_t>(V);
2404}
2405
2406template <int Ind, typename Opnd_t> struct ExtractValue_match {
2407 Opnd_t Val;
2408 ExtractValue_match(const Opnd_t &V) : Val(V) {}
2409
2410 template <typename OpTy> bool match(OpTy *V) {
2411 if (auto *I = dyn_cast<ExtractValueInst>(V)) {
2412 // If Ind is -1, don't inspect indices
2413 if (Ind != -1 &&
2414 !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind))
2415 return false;
2416 return Val.match(I->getAggregateOperand());
2417 }
2418 return false;
2419 }
2420};
2421
2422/// Match a single index ExtractValue instruction.
2423/// For example m_ExtractValue<1>(...)
2424template <int Ind, typename Val_t>
2425inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) {
2426 return ExtractValue_match<Ind, Val_t>(V);
2427}
2428
2429/// Match an ExtractValue instruction with any index.
2430/// For example m_ExtractValue(...)
2431template <typename Val_t>
2432inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) {
2433 return ExtractValue_match<-1, Val_t>(V);
2434}
2435
2436/// Matcher for a single index InsertValue instruction.
2437template <int Ind, typename T0, typename T1> struct InsertValue_match {
2438 T0 Op0;
2439 T1 Op1;
2440
2441 InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {}
2442
2443 template <typename OpTy> bool match(OpTy *V) {
2444 if (auto *I = dyn_cast<InsertValueInst>(V)) {
2445 return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) &&
2446 I->getNumIndices() == 1 && Ind == I->getIndices()[0];
2447 }
2448 return false;
2449 }
2450};
2451
2452/// Matches a single index InsertValue instruction.
2453template <int Ind, typename Val_t, typename Elt_t>
2454inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val,
2455 const Elt_t &Elt) {
2456 return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt);
2457}
2458
2459/// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or
2460/// the constant expression
2461/// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>`
2462/// under the right conditions determined by DataLayout.
2463struct VScaleVal_match {
2464 const DataLayout &DL;
2465 VScaleVal_match(const DataLayout &DL) : DL(DL) {}
2466
2467 template <typename ITy> bool match(ITy *V) {
2468 if (m_Intrinsic<Intrinsic::vscale>().match(V))
2469 return true;
2470
2471 Value *Ptr;
2472 if (m_PtrToInt(m_Value(Ptr)).match(V)) {
2473 if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
2474 auto *DerefTy = GEP->getSourceElementType();
2475 if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) &&
2476 m_Zero().match(GEP->getPointerOperand()) &&
2477 m_SpecificInt(1).match(GEP->idx_begin()->get()) &&
2478 DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8)
2479 return true;
2480 }
2481 }
2482
2483 return false;
2484 }
2485};
2486
2487inline VScaleVal_match m_VScale(const DataLayout &DL) {
2488 return VScaleVal_match(DL);
2489}
2490
2491template <typename LHS, typename RHS, unsigned Opcode, bool Commutable = false>
2492struct LogicalOp_match {
2493 LHS L;
2494 RHS R;
2495
2496 LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {}
2497
2498 template <typename T> bool match(T *V) {
2499 auto *I = dyn_cast<Instruction>(V);
2500 if (!I || !I->getType()->isIntOrIntVectorTy(1))
2501 return false;
2502
2503 if (I->getOpcode() == Opcode) {
2504 auto *Op0 = I->getOperand(0);
2505 auto *Op1 = I->getOperand(1);
2506 return (L.match(Op0) && R.match(Op1)) ||
2507 (Commutable && L.match(Op1) && R.match(Op0));
2508 }
2509
2510 if (auto *Select = dyn_cast<SelectInst>(I)) {
2511 auto *Cond = Select->getCondition();
2512 auto *TVal = Select->getTrueValue();
2513 auto *FVal = Select->getFalseValue();
2514 if (Opcode == Instruction::And) {
2515 auto *C = dyn_cast<Constant>(FVal);
2516 if (C && C->isNullValue())
2517 return (L.match(Cond) && R.match(TVal)) ||
2518 (Commutable && L.match(TVal) && R.match(Cond));
2519 } else {
2520 assert(Opcode == Instruction::Or)(static_cast <bool> (Opcode == Instruction::Or) ? void (
0) : __assert_fail ("Opcode == Instruction::Or", "llvm/include/llvm/IR/PatternMatch.h"
, 2520, __extension__ __PRETTY_FUNCTION__))
;
2521 auto *C = dyn_cast<Constant>(TVal);
2522 if (C && C->isOneValue())
2523 return (L.match(Cond) && R.match(FVal)) ||
2524 (Commutable && L.match(FVal) && R.match(Cond));
2525 }
2526 }
2527
2528 return false;
2529 }
2530};
2531
2532/// Matches L && R either in the form of L & R or L ? R : false.
2533/// Note that the latter form is poison-blocking.
2534template <typename LHS, typename RHS>
2535inline LogicalOp_match<LHS, RHS, Instruction::And>
2536m_LogicalAnd(const LHS &L, const RHS &R) {
2537 return LogicalOp_match<LHS, RHS, Instruction::And>(L, R);
2538}
2539
2540/// Matches L && R where L and R are arbitrary values.
2541inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); }
2542
2543/// Matches L && R with LHS and RHS in either order.
2544template <typename LHS, typename RHS>
2545inline LogicalOp_match<LHS, RHS, Instruction::And, true>
2546m_c_LogicalAnd(const LHS &L, const RHS &R) {
2547 return LogicalOp_match<LHS, RHS, Instruction::And, true>(L, R);
2548}
2549
2550/// Matches L || R either in the form of L | R or L ? true : R.
2551/// Note that the latter form is poison-blocking.
2552template <typename LHS, typename RHS>
2553inline LogicalOp_match<LHS, RHS, Instruction::Or>
2554m_LogicalOr(const LHS &L, const RHS &R) {
2555 return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R);
2556}
2557
2558/// Matches L || R where L and R are arbitrary values.
2559inline auto m_LogicalOr() { return m_LogicalOr(m_Value(), m_Value()); }
2560
2561/// Matches L || R with LHS and RHS in either order.
2562template <typename LHS, typename RHS>
2563inline LogicalOp_match<LHS, RHS, Instruction::Or, true>
2564m_c_LogicalOr(const LHS &L, const RHS &R) {
2565 return LogicalOp_match<LHS, RHS, Instruction::Or, true>(L, R);
2566}
2567
2568} // end namespace PatternMatch
2569} // end namespace llvm
2570
2571#endif // LLVM_IR_PATTERNMATCH_H