File: | llvm/include/llvm/IR/PatternMatch.h |
Warning: | line 1460, column 9 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===// | ||||
2 | // | ||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||
6 | // | ||||
7 | //===----------------------------------------------------------------------===// | ||||
8 | // | ||||
9 | // This file implements the visit functions for add, fadd, sub, and fsub. | ||||
10 | // | ||||
11 | //===----------------------------------------------------------------------===// | ||||
12 | |||||
13 | #include "InstCombineInternal.h" | ||||
14 | #include "llvm/ADT/APFloat.h" | ||||
15 | #include "llvm/ADT/APInt.h" | ||||
16 | #include "llvm/ADT/STLExtras.h" | ||||
17 | #include "llvm/ADT/SmallVector.h" | ||||
18 | #include "llvm/Analysis/InstructionSimplify.h" | ||||
19 | #include "llvm/Analysis/ValueTracking.h" | ||||
20 | #include "llvm/IR/Constant.h" | ||||
21 | #include "llvm/IR/Constants.h" | ||||
22 | #include "llvm/IR/InstrTypes.h" | ||||
23 | #include "llvm/IR/Instruction.h" | ||||
24 | #include "llvm/IR/Instructions.h" | ||||
25 | #include "llvm/IR/Operator.h" | ||||
26 | #include "llvm/IR/PatternMatch.h" | ||||
27 | #include "llvm/IR/Type.h" | ||||
28 | #include "llvm/IR/Value.h" | ||||
29 | #include "llvm/Support/AlignOf.h" | ||||
30 | #include "llvm/Support/Casting.h" | ||||
31 | #include "llvm/Support/KnownBits.h" | ||||
32 | #include "llvm/Transforms/InstCombine/InstCombiner.h" | ||||
33 | #include <cassert> | ||||
34 | #include <utility> | ||||
35 | |||||
36 | using namespace llvm; | ||||
37 | using namespace PatternMatch; | ||||
38 | |||||
39 | #define DEBUG_TYPE"instcombine" "instcombine" | ||||
40 | |||||
41 | namespace { | ||||
42 | |||||
43 | /// Class representing coefficient of floating-point addend. | ||||
44 | /// This class needs to be highly efficient, which is especially true for | ||||
45 | /// the constructor. As of I write this comment, the cost of the default | ||||
46 | /// constructor is merely 4-byte-store-zero (Assuming compiler is able to | ||||
47 | /// perform write-merging). | ||||
48 | /// | ||||
49 | class FAddendCoef { | ||||
50 | public: | ||||
51 | // The constructor has to initialize a APFloat, which is unnecessary for | ||||
52 | // most addends which have coefficient either 1 or -1. So, the constructor | ||||
53 | // is expensive. In order to avoid the cost of the constructor, we should | ||||
54 | // reuse some instances whenever possible. The pre-created instances | ||||
55 | // FAddCombine::Add[0-5] embodies this idea. | ||||
56 | FAddendCoef() = default; | ||||
57 | ~FAddendCoef(); | ||||
58 | |||||
59 | // If possible, don't define operator+/operator- etc because these | ||||
60 | // operators inevitably call FAddendCoef's constructor which is not cheap. | ||||
61 | void operator=(const FAddendCoef &A); | ||||
62 | void operator+=(const FAddendCoef &A); | ||||
63 | void operator*=(const FAddendCoef &S); | ||||
64 | |||||
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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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 | //===----------------------------------------------------------------------===// | ||||
224 | FAddendCoef::~FAddendCoef() { | ||||
225 | if (BufHasFpVal) | ||||
226 | getFpValPtr()->~APFloat(); | ||||
227 | } | ||||
228 | |||||
229 | void FAddendCoef::set(const APFloat& C) { | ||||
230 | APFloat *P = getFpValPtr(); | ||||
231 | |||||
232 | if (isInt()) { | ||||
233 | // As the buffer is meanless byte stream, we cannot call | ||||
234 | // APFloat::operator=(). | ||||
235 | new(P) APFloat(C); | ||||
236 | } else | ||||
237 | *P = C; | ||||
238 | |||||
239 | IsFp = BufHasFpVal = true; | ||||
240 | } | ||||
241 | |||||
242 | void FAddendCoef::convertToFpType(const fltSemantics &Sem) { | ||||
243 | if (!isInt()) | ||||
244 | return; | ||||
245 | |||||
246 | APFloat *P = getFpValPtr(); | ||||
247 | if (IntVal > 0) | ||||
248 | new(P) APFloat(Sem, IntVal); | ||||
249 | else { | ||||
250 | new(P) APFloat(Sem, 0 - IntVal); | ||||
251 | P->changeSign(); | ||||
252 | } | ||||
253 | IsFp = BufHasFpVal = true; | ||||
254 | } | ||||
255 | |||||
256 | APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) { | ||||
257 | if (Val >= 0) | ||||
258 | return APFloat(Sem, Val); | ||||
259 | |||||
260 | APFloat T(Sem, 0 - Val); | ||||
261 | T.changeSign(); | ||||
262 | |||||
263 | return T; | ||||
264 | } | ||||
265 | |||||
266 | void FAddendCoef::operator=(const FAddendCoef &That) { | ||||
267 | if (That.isInt()) | ||||
268 | set(That.IntVal); | ||||
269 | else | ||||
270 | set(That.getFpVal()); | ||||
271 | } | ||||
272 | |||||
273 | void FAddendCoef::operator+=(const FAddendCoef &That) { | ||||
274 | RoundingMode RndMode = RoundingMode::NearestTiesToEven; | ||||
275 | if (isInt() == That.isInt()) { | ||||
276 | if (isInt()) | ||||
277 | IntVal += That.IntVal; | ||||
278 | else | ||||
279 | getFpVal().add(That.getFpVal(), RndMode); | ||||
280 | return; | ||||
281 | } | ||||
282 | |||||
283 | if (isInt()) { | ||||
284 | const APFloat &T = That.getFpVal(); | ||||
285 | convertToFpType(T.getSemantics()); | ||||
286 | getFpVal().add(T, RndMode); | ||||
287 | return; | ||||
288 | } | ||||
289 | |||||
290 | APFloat &T = getFpVal(); | ||||
291 | T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode); | ||||
292 | } | ||||
293 | |||||
294 | void FAddendCoef::operator*=(const FAddendCoef &That) { | ||||
295 | if (That.isOne()) | ||||
296 | return; | ||||
297 | |||||
298 | if (That.isMinusOne()) { | ||||
299 | negate(); | ||||
300 | return; | ||||
301 | } | ||||
302 | |||||
303 | if (isInt() && That.isInt()) { | ||||
304 | int Res = IntVal * (int)That.IntVal; | ||||
305 | assert(!insaneIntVal(Res) && "Insane int value")(static_cast <bool> (!insaneIntVal(Res) && "Insane int value" ) ? void (0) : __assert_fail ("!insaneIntVal(Res) && \"Insane int value\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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 | |||||
324 | void FAddendCoef::negate() { | ||||
325 | if (isInt()) | ||||
326 | IntVal = 0 - IntVal; | ||||
327 | else | ||||
328 | getFpVal().changeSign(); | ||||
329 | } | ||||
330 | |||||
331 | Value *FAddendCoef::getValue(Type *Ty) const { | ||||
332 | return isInt() ? | ||||
333 | ConstantFP::get(Ty, float(IntVal)) : | ||||
334 | ConstantFP::get(Ty->getContext(), getFpVal()); | ||||
335 | } | ||||
336 | |||||
337 | // The definition of <Val> Addends | ||||
338 | // ========================================= | ||||
339 | // A + B <1, A>, <1,B> | ||||
340 | // A - B <1, A>, <1,B> | ||||
341 | // 0 - B <-1, B> | ||||
342 | // C * A, <C, A> | ||||
343 | // A + C <1, A> <C, NULL> | ||||
344 | // 0 +/- 0 <0, NULL> (corner case) | ||||
345 | // | ||||
346 | // Legend: A and B are not constant, C is constant | ||||
347 | unsigned FAddend::drillValueDownOneStep | ||||
348 | (Value *Val, FAddend &Addend0, FAddend &Addend1) { | ||||
349 | Instruction *I = nullptr; | ||||
350 | if (!Val || !(I = dyn_cast<Instruction>(Val))) | ||||
351 | return 0; | ||||
352 | |||||
353 | unsigned Opcode = I->getOpcode(); | ||||
354 | |||||
355 | if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) { | ||||
356 | ConstantFP *C0, *C1; | ||||
357 | Value *Opnd0 = I->getOperand(0); | ||||
358 | Value *Opnd1 = I->getOperand(1); | ||||
359 | if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero()) | ||||
360 | Opnd0 = nullptr; | ||||
361 | |||||
362 | if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero()) | ||||
363 | Opnd1 = nullptr; | ||||
364 | |||||
365 | if (Opnd0) { | ||||
366 | if (!C0) | ||||
367 | Addend0.set(1, Opnd0); | ||||
368 | else | ||||
369 | Addend0.set(C0, nullptr); | ||||
370 | } | ||||
371 | |||||
372 | if (Opnd1) { | ||||
373 | FAddend &Addend = Opnd0 ? Addend1 : Addend0; | ||||
374 | if (!C1) | ||||
375 | Addend.set(1, Opnd1); | ||||
376 | else | ||||
377 | Addend.set(C1, nullptr); | ||||
378 | if (Opcode == Instruction::FSub) | ||||
379 | Addend.negate(); | ||||
380 | } | ||||
381 | |||||
382 | if (Opnd0 || Opnd1) | ||||
383 | return Opnd0 && Opnd1 ? 2 : 1; | ||||
384 | |||||
385 | // Both operands are zero. Weird! | ||||
386 | Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr); | ||||
387 | return 1; | ||||
388 | } | ||||
389 | |||||
390 | if (I->getOpcode() == Instruction::FMul) { | ||||
391 | Value *V0 = I->getOperand(0); | ||||
392 | Value *V1 = I->getOperand(1); | ||||
393 | if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) { | ||||
394 | Addend0.set(C, V1); | ||||
395 | return 1; | ||||
396 | } | ||||
397 | |||||
398 | if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) { | ||||
399 | Addend0.set(C, V0); | ||||
400 | return 1; | ||||
401 | } | ||||
402 | } | ||||
403 | |||||
404 | return 0; | ||||
405 | } | ||||
406 | |||||
407 | // Try to break *this* addend into two addends. e.g. Suppose this addend is | ||||
408 | // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends, | ||||
409 | // i.e. <2.3, X> and <2.3, Y>. | ||||
410 | unsigned FAddend::drillAddendDownOneStep | ||||
411 | (FAddend &Addend0, FAddend &Addend1) const { | ||||
412 | if (isConstant()) | ||||
413 | return 0; | ||||
414 | |||||
415 | unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1); | ||||
416 | if (!BreakNum || Coeff.isOne()) | ||||
417 | return BreakNum; | ||||
418 | |||||
419 | Addend0.Scale(Coeff); | ||||
420 | |||||
421 | if (BreakNum == 2) | ||||
422 | Addend1.Scale(Coeff); | ||||
423 | |||||
424 | return BreakNum; | ||||
425 | } | ||||
426 | |||||
427 | Value *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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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 | |||||
514 | Value *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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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 | // Points to the constant addend of the resulting simplified expression. | ||||
523 | // If the resulting expr has constant-addend, this constant-addend is | ||||
524 | // desirable to reside at the top of the resulting expression tree. Placing | ||||
525 | // constant close to supper-expr(s) will potentially reveal some optimization | ||||
526 | // opportunities in super-expr(s). | ||||
527 | const FAddend *ConstAdd = nullptr; | ||||
528 | |||||
529 | // Simplified addends are placed <SimpVect>. | ||||
530 | AddendVect SimpVect; | ||||
531 | |||||
532 | // The outer loop works on one symbolic-value at a time. Suppose the input | ||||
533 | // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ... | ||||
534 | // The symbolic-values will be processed in this order: x, y, z. | ||||
535 | for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) { | ||||
536 | |||||
537 | const FAddend *ThisAddend = Addends[SymIdx]; | ||||
538 | if (!ThisAddend) { | ||||
539 | // This addend was processed before. | ||||
540 | continue; | ||||
541 | } | ||||
542 | |||||
543 | Value *Val = ThisAddend->getSymVal(); | ||||
544 | unsigned StartIdx = SimpVect.size(); | ||||
545 | SimpVect.push_back(ThisAddend); | ||||
546 | |||||
547 | // The inner loop collects addends sharing same symbolic-value, and these | ||||
548 | // addends will be later on folded into a single addend. Following above | ||||
549 | // example, if the symbolic value "y" is being processed, the inner loop | ||||
550 | // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will | ||||
551 | // be later on folded into "<b1+b2, y>". | ||||
552 | for (unsigned SameSymIdx = SymIdx + 1; | ||||
553 | SameSymIdx < AddendNum; SameSymIdx++) { | ||||
554 | const FAddend *T = Addends[SameSymIdx]; | ||||
555 | if (T && T->getSymVal() == Val) { | ||||
556 | // Set null such that next iteration of the outer loop will not process | ||||
557 | // this addend again. | ||||
558 | Addends[SameSymIdx] = nullptr; | ||||
559 | SimpVect.push_back(T); | ||||
560 | } | ||||
561 | } | ||||
562 | |||||
563 | // If multiple addends share same symbolic value, fold them together. | ||||
564 | if (StartIdx + 1 != SimpVect.size()) { | ||||
565 | FAddend &R = TmpResult[NextTmpIdx ++]; | ||||
566 | R = *SimpVect[StartIdx]; | ||||
567 | for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++) | ||||
568 | R += *SimpVect[Idx]; | ||||
569 | |||||
570 | // Pop all addends being folded and push the resulting folded addend. | ||||
571 | SimpVect.resize(StartIdx); | ||||
572 | if (Val) { | ||||
573 | if (!R.isZero()) { | ||||
574 | SimpVect.push_back(&R); | ||||
575 | } | ||||
576 | } else { | ||||
577 | // Don't push constant addend at this time. It will be the last element | ||||
578 | // of <SimpVect>. | ||||
579 | ConstAdd = &R; | ||||
580 | } | ||||
581 | } | ||||
582 | } | ||||
583 | |||||
584 | assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&(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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 585, __extension__ __PRETTY_FUNCTION__)) | ||||
585 | "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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 585, __extension__ __PRETTY_FUNCTION__)); | ||||
586 | |||||
587 | if (ConstAdd) | ||||
588 | SimpVect.push_back(ConstAdd); | ||||
589 | |||||
590 | Value *Result; | ||||
591 | if (!SimpVect.empty()) | ||||
592 | Result = createNaryFAdd(SimpVect, InstrQuota); | ||||
593 | else { | ||||
594 | // The addition is folded to 0.0. | ||||
595 | Result = ConstantFP::get(Instr->getType(), 0.0); | ||||
596 | } | ||||
597 | |||||
598 | return Result; | ||||
599 | } | ||||
600 | |||||
601 | Value *FAddCombine::createNaryFAdd | ||||
602 | (const AddendVect &Opnds, unsigned InstrQuota) { | ||||
603 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 603, __extension__ __PRETTY_FUNCTION__)); | ||||
604 | |||||
605 | // Step 1: Check if the # of instructions needed exceeds the quota. | ||||
606 | |||||
607 | unsigned InstrNeeded = calcInstrNumber(Opnds); | ||||
608 | if (InstrNeeded > InstrQuota) | ||||
609 | return nullptr; | ||||
610 | |||||
611 | initCreateInstNum(); | ||||
612 | |||||
613 | // step 2: Emit the N-ary addition. | ||||
614 | // Note that at most three instructions are involved in Fadd-InstCombine: the | ||||
615 | // addition in question, and at most two neighboring instructions. | ||||
616 | // The resulting optimized addition should have at least one less instruction | ||||
617 | // than the original addition expression tree. This implies that the resulting | ||||
618 | // N-ary addition has at most two instructions, and we don't need to worry | ||||
619 | // about tree-height when constructing the N-ary addition. | ||||
620 | |||||
621 | Value *LastVal = nullptr; | ||||
622 | bool LastValNeedNeg = false; | ||||
623 | |||||
624 | // Iterate the addends, creating fadd/fsub using adjacent two addends. | ||||
625 | for (const FAddend *Opnd : Opnds) { | ||||
626 | bool NeedNeg; | ||||
627 | Value *V = createAddendVal(*Opnd, NeedNeg); | ||||
628 | if (!LastVal) { | ||||
629 | LastVal = V; | ||||
630 | LastValNeedNeg = NeedNeg; | ||||
631 | continue; | ||||
632 | } | ||||
633 | |||||
634 | if (LastValNeedNeg == NeedNeg) { | ||||
635 | LastVal = createFAdd(LastVal, V); | ||||
636 | continue; | ||||
637 | } | ||||
638 | |||||
639 | if (LastValNeedNeg) | ||||
640 | LastVal = createFSub(V, LastVal); | ||||
641 | else | ||||
642 | LastVal = createFSub(LastVal, V); | ||||
643 | |||||
644 | LastValNeedNeg = false; | ||||
645 | } | ||||
646 | |||||
647 | if (LastValNeedNeg) { | ||||
648 | LastVal = createFNeg(LastVal); | ||||
649 | } | ||||
650 | |||||
651 | #ifndef NDEBUG | ||||
652 | assert(CreateInstrNum == InstrNeeded &&(static_cast <bool> (CreateInstrNum == InstrNeeded && "Inconsistent in instruction numbers") ? void (0) : __assert_fail ("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 653, __extension__ __PRETTY_FUNCTION__)) | ||||
653 | "Inconsistent in instruction numbers")(static_cast <bool> (CreateInstrNum == InstrNeeded && "Inconsistent in instruction numbers") ? void (0) : __assert_fail ("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 653, __extension__ __PRETTY_FUNCTION__)); | ||||
654 | #endif | ||||
655 | |||||
656 | return LastVal; | ||||
657 | } | ||||
658 | |||||
659 | Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) { | ||||
660 | Value *V = Builder.CreateFSub(Opnd0, Opnd1); | ||||
661 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
662 | createInstPostProc(I); | ||||
663 | return V; | ||||
664 | } | ||||
665 | |||||
666 | Value *FAddCombine::createFNeg(Value *V) { | ||||
667 | Value *NewV = Builder.CreateFNeg(V); | ||||
668 | if (Instruction *I = dyn_cast<Instruction>(NewV)) | ||||
669 | createInstPostProc(I, true); // fneg's don't receive instruction numbers. | ||||
670 | return NewV; | ||||
671 | } | ||||
672 | |||||
673 | Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) { | ||||
674 | Value *V = Builder.CreateFAdd(Opnd0, Opnd1); | ||||
675 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
676 | createInstPostProc(I); | ||||
677 | return V; | ||||
678 | } | ||||
679 | |||||
680 | Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) { | ||||
681 | Value *V = Builder.CreateFMul(Opnd0, Opnd1); | ||||
682 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
683 | createInstPostProc(I); | ||||
684 | return V; | ||||
685 | } | ||||
686 | |||||
687 | void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) { | ||||
688 | NewInstr->setDebugLoc(Instr->getDebugLoc()); | ||||
689 | |||||
690 | // Keep track of the number of instruction created. | ||||
691 | if (!NoNumber) | ||||
692 | incCreateInstNum(); | ||||
693 | |||||
694 | // Propagate fast-math flags | ||||
695 | NewInstr->setFastMathFlags(Instr->getFastMathFlags()); | ||||
696 | } | ||||
697 | |||||
698 | // Return the number of instruction needed to emit the N-ary addition. | ||||
699 | // NOTE: Keep this function in sync with createAddendVal(). | ||||
700 | unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) { | ||||
701 | unsigned OpndNum = Opnds.size(); | ||||
702 | unsigned InstrNeeded = OpndNum - 1; | ||||
703 | |||||
704 | // The number of addends in the form of "(-1)*x". | ||||
705 | unsigned NegOpndNum = 0; | ||||
706 | |||||
707 | // Adjust the number of instructions needed to emit the N-ary add. | ||||
708 | for (const FAddend *Opnd : Opnds) { | ||||
709 | if (Opnd->isConstant()) | ||||
710 | continue; | ||||
711 | |||||
712 | // The constant check above is really for a few special constant | ||||
713 | // coefficients. | ||||
714 | if (isa<UndefValue>(Opnd->getSymVal())) | ||||
715 | continue; | ||||
716 | |||||
717 | const FAddendCoef &CE = Opnd->getCoef(); | ||||
718 | if (CE.isMinusOne() || CE.isMinusTwo()) | ||||
719 | NegOpndNum++; | ||||
720 | |||||
721 | // Let the addend be "c * x". If "c == +/-1", the value of the addend | ||||
722 | // is immediately available; otherwise, it needs exactly one instruction | ||||
723 | // to evaluate the value. | ||||
724 | if (!CE.isMinusOne() && !CE.isOne()) | ||||
725 | InstrNeeded++; | ||||
726 | } | ||||
727 | return InstrNeeded; | ||||
728 | } | ||||
729 | |||||
730 | // Input Addend Value NeedNeg(output) | ||||
731 | // ================================================================ | ||||
732 | // Constant C C false | ||||
733 | // <+/-1, V> V coefficient is -1 | ||||
734 | // <2/-2, V> "fadd V, V" coefficient is -2 | ||||
735 | // <C, V> "fmul V, C" false | ||||
736 | // | ||||
737 | // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber. | ||||
738 | Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) { | ||||
739 | const FAddendCoef &Coeff = Opnd.getCoef(); | ||||
740 | |||||
741 | if (Opnd.isConstant()) { | ||||
742 | NeedNeg = false; | ||||
743 | return Coeff.getValue(Instr->getType()); | ||||
744 | } | ||||
745 | |||||
746 | Value *OpndVal = Opnd.getSymVal(); | ||||
747 | |||||
748 | if (Coeff.isMinusOne() || Coeff.isOne()) { | ||||
749 | NeedNeg = Coeff.isMinusOne(); | ||||
750 | return OpndVal; | ||||
751 | } | ||||
752 | |||||
753 | if (Coeff.isTwo() || Coeff.isMinusTwo()) { | ||||
754 | NeedNeg = Coeff.isMinusTwo(); | ||||
755 | return createFAdd(OpndVal, OpndVal); | ||||
756 | } | ||||
757 | |||||
758 | NeedNeg = false; | ||||
759 | return createFMul(OpndVal, Coeff.getValue(Instr->getType())); | ||||
760 | } | ||||
761 | |||||
762 | // Checks if any operand is negative and we can convert add to sub. | ||||
763 | // This function checks for following negative patterns | ||||
764 | // ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C)) | ||||
765 | // ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C)) | ||||
766 | // XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even | ||||
767 | static Value *checkForNegativeOperand(BinaryOperator &I, | ||||
768 | InstCombiner::BuilderTy &Builder) { | ||||
769 | Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); | ||||
770 | |||||
771 | // This function creates 2 instructions to replace ADD, we need at least one | ||||
772 | // of LHS or RHS to have one use to ensure benefit in transform. | ||||
773 | if (!LHS->hasOneUse() && !RHS->hasOneUse()) | ||||
774 | return nullptr; | ||||
775 | |||||
776 | Value *X = nullptr, *Y = nullptr, *Z = nullptr; | ||||
777 | const APInt *C1 = nullptr, *C2 = nullptr; | ||||
778 | |||||
779 | // if ONE is on other side, swap | ||||
780 | if (match(RHS, m_Add(m_Value(X), m_One()))) | ||||
781 | std::swap(LHS, RHS); | ||||
782 | |||||
783 | if (match(LHS, m_Add(m_Value(X), m_One()))) { | ||||
784 | // if XOR on other side, swap | ||||
785 | if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1)))) | ||||
786 | std::swap(X, RHS); | ||||
787 | |||||
788 | if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) { | ||||
789 | // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1)) | ||||
790 | // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1)) | ||||
791 | if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) { | ||||
792 | Value *NewAnd = Builder.CreateAnd(Z, *C1); | ||||
793 | return Builder.CreateSub(RHS, NewAnd, "sub"); | ||||
794 | } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) { | ||||
795 | // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1)) | ||||
796 | // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1)) | ||||
797 | Value *NewOr = Builder.CreateOr(Z, ~(*C1)); | ||||
798 | return Builder.CreateSub(RHS, NewOr, "sub"); | ||||
799 | } | ||||
800 | } | ||||
801 | } | ||||
802 | |||||
803 | // Restore LHS and RHS | ||||
804 | LHS = I.getOperand(0); | ||||
805 | RHS = I.getOperand(1); | ||||
806 | |||||
807 | // if XOR is on other side, swap | ||||
808 | if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1)))) | ||||
809 | std::swap(LHS, RHS); | ||||
810 | |||||
811 | // C2 is ODD | ||||
812 | // LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2)) | ||||
813 | // ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2)) | ||||
814 | if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1)))) | ||||
815 | if (C1->countTrailingZeros() == 0) | ||||
816 | if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) { | ||||
817 | Value *NewOr = Builder.CreateOr(Z, ~(*C2)); | ||||
818 | return Builder.CreateSub(RHS, NewOr, "sub"); | ||||
819 | } | ||||
820 | return nullptr; | ||||
821 | } | ||||
822 | |||||
823 | /// Wrapping flags may allow combining constants separated by an extend. | ||||
824 | static Instruction *foldNoWrapAdd(BinaryOperator &Add, | ||||
825 | InstCombiner::BuilderTy &Builder) { | ||||
826 | Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1); | ||||
827 | Type *Ty = Add.getType(); | ||||
828 | Constant *Op1C; | ||||
829 | if (!match(Op1, m_Constant(Op1C))) | ||||
830 | return nullptr; | ||||
831 | |||||
832 | // Try this match first because it results in an add in the narrow type. | ||||
833 | // (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1))) | ||||
834 | Value *X; | ||||
835 | const APInt *C1, *C2; | ||||
836 | if (match(Op1, m_APInt(C1)) && | ||||
837 | match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) && | ||||
838 | C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) { | ||||
839 | Constant *NewC = | ||||
840 | ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth())); | ||||
841 | return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty); | ||||
842 | } | ||||
843 | |||||
844 | // More general combining of constants in the wide type. | ||||
845 | // (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C) | ||||
846 | Constant *NarrowC; | ||||
847 | if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) { | ||||
848 | Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty); | ||||
849 | Constant *NewC = ConstantExpr::getAdd(WideC, Op1C); | ||||
850 | Value *WideX = Builder.CreateSExt(X, Ty); | ||||
851 | return BinaryOperator::CreateAdd(WideX, NewC); | ||||
852 | } | ||||
853 | // (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C) | ||||
854 | if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) { | ||||
855 | Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty); | ||||
856 | Constant *NewC = ConstantExpr::getAdd(WideC, Op1C); | ||||
857 | Value *WideX = Builder.CreateZExt(X, Ty); | ||||
858 | return BinaryOperator::CreateAdd(WideX, NewC); | ||||
859 | } | ||||
860 | |||||
861 | return nullptr; | ||||
862 | } | ||||
863 | |||||
864 | Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) { | ||||
865 | Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1); | ||||
866 | Constant *Op1C; | ||||
867 | if (!match(Op1, m_ImmConstant(Op1C))) | ||||
868 | return nullptr; | ||||
869 | |||||
870 | if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add)) | ||||
871 | return NV; | ||||
872 | |||||
873 | Value *X; | ||||
874 | Constant *Op00C; | ||||
875 | |||||
876 | // add (sub C1, X), C2 --> sub (add C1, C2), X | ||||
877 | if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X)))) | ||||
878 | return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X); | ||||
879 | |||||
880 | Value *Y; | ||||
881 | |||||
882 | // add (sub X, Y), -1 --> add (not Y), X | ||||
883 | if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) && | ||||
884 | match(Op1, m_AllOnes())) | ||||
885 | return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X); | ||||
886 | |||||
887 | // zext(bool) + C -> bool ? C + 1 : C | ||||
888 | if (match(Op0, m_ZExt(m_Value(X))) && | ||||
889 | X->getType()->getScalarSizeInBits() == 1) | ||||
890 | return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1); | ||||
891 | // sext(bool) + C -> bool ? C - 1 : C | ||||
892 | if (match(Op0, m_SExt(m_Value(X))) && | ||||
893 | X->getType()->getScalarSizeInBits() == 1) | ||||
894 | return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1); | ||||
895 | |||||
896 | // ~X + C --> (C-1) - X | ||||
897 | if (match(Op0, m_Not(m_Value(X)))) | ||||
898 | return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X); | ||||
899 | |||||
900 | const APInt *C; | ||||
901 | if (!match(Op1, m_APInt(C))) | ||||
902 | return nullptr; | ||||
903 | |||||
904 | // (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add` | ||||
905 | Constant *Op01C; | ||||
906 | if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) && | ||||
907 | haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT)) | ||||
908 | return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C)); | ||||
909 | |||||
910 | // (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C) | ||||
911 | const APInt *C2; | ||||
912 | if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C) | ||||
913 | return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2)); | ||||
914 | |||||
915 | if (C->isSignMask()) { | ||||
916 | // If wrapping is not allowed, then the addition must set the sign bit: | ||||
917 | // X + (signmask) --> X | signmask | ||||
918 | if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap()) | ||||
919 | return BinaryOperator::CreateOr(Op0, Op1); | ||||
920 | |||||
921 | // If wrapping is allowed, then the addition flips the sign bit of LHS: | ||||
922 | // X + (signmask) --> X ^ signmask | ||||
923 | return BinaryOperator::CreateXor(Op0, Op1); | ||||
924 | } | ||||
925 | |||||
926 | // Is this add the last step in a convoluted sext? | ||||
927 | // add(zext(xor i16 X, -32768), -32768) --> sext X | ||||
928 | Type *Ty = Add.getType(); | ||||
929 | if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) && | ||||
930 | C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C) | ||||
931 | return CastInst::Create(Instruction::SExt, X, Ty); | ||||
932 | |||||
933 | if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) { | ||||
934 | // (X ^ signmask) + C --> (X + (signmask ^ C)) | ||||
935 | if (C2->isSignMask()) | ||||
936 | return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C)); | ||||
937 | |||||
938 | // If X has no high-bits set above an xor mask: | ||||
939 | // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X | ||||
940 | if (C2->isMask()) { | ||||
941 | KnownBits LHSKnown = computeKnownBits(X, 0, &Add); | ||||
942 | if ((*C2 | LHSKnown.Zero).isAllOnesValue()) | ||||
943 | return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X); | ||||
944 | } | ||||
945 | |||||
946 | // Look for a math+logic pattern that corresponds to sext-in-register of a | ||||
947 | // value with cleared high bits. Convert that into a pair of shifts: | ||||
948 | // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC | ||||
949 | // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC | ||||
950 | if (Op0->hasOneUse() && *C2 == -(*C)) { | ||||
951 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
952 | unsigned ShAmt = 0; | ||||
953 | if (C->isPowerOf2()) | ||||
954 | ShAmt = BitWidth - C->logBase2() - 1; | ||||
955 | else if (C2->isPowerOf2()) | ||||
956 | ShAmt = BitWidth - C2->logBase2() - 1; | ||||
957 | if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt), | ||||
958 | 0, &Add)) { | ||||
959 | Constant *ShAmtC = ConstantInt::get(Ty, ShAmt); | ||||
960 | Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext"); | ||||
961 | return BinaryOperator::CreateAShr(NewShl, ShAmtC); | ||||
962 | } | ||||
963 | } | ||||
964 | } | ||||
965 | |||||
966 | if (C->isOneValue() && Op0->hasOneUse()) { | ||||
967 | // add (sext i1 X), 1 --> zext (not X) | ||||
968 | // TODO: The smallest IR representation is (select X, 0, 1), and that would | ||||
969 | // not require the one-use check. But we need to remove a transform in | ||||
970 | // visitSelect and make sure that IR value tracking for select is equal or | ||||
971 | // better than for these ops. | ||||
972 | if (match(Op0, m_SExt(m_Value(X))) && | ||||
973 | X->getType()->getScalarSizeInBits() == 1) | ||||
974 | return new ZExtInst(Builder.CreateNot(X), Ty); | ||||
975 | |||||
976 | // Shifts and add used to flip and mask off the low bit: | ||||
977 | // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1 | ||||
978 | const APInt *C3; | ||||
979 | if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) && | ||||
980 | C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) { | ||||
981 | Value *NotX = Builder.CreateNot(X); | ||||
982 | return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1)); | ||||
983 | } | ||||
984 | } | ||||
985 | |||||
986 | // If all bits affected by the add are included in a high-bit-mask, do the | ||||
987 | // add before the mask op: | ||||
988 | // (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00 | ||||
989 | if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) && | ||||
990 | C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) { | ||||
991 | Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C)); | ||||
992 | return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2)); | ||||
993 | } | ||||
994 | |||||
995 | return nullptr; | ||||
996 | } | ||||
997 | |||||
998 | // Matches multiplication expression Op * C where C is a constant. Returns the | ||||
999 | // constant value in C and the other operand in Op. Returns true if such a | ||||
1000 | // match is found. | ||||
1001 | static bool MatchMul(Value *E, Value *&Op, APInt &C) { | ||||
1002 | const APInt *AI; | ||||
1003 | if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) { | ||||
1004 | C = *AI; | ||||
1005 | return true; | ||||
1006 | } | ||||
1007 | if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) { | ||||
1008 | C = APInt(AI->getBitWidth(), 1); | ||||
1009 | C <<= *AI; | ||||
1010 | return true; | ||||
1011 | } | ||||
1012 | return false; | ||||
1013 | } | ||||
1014 | |||||
1015 | // Matches remainder expression Op % C where C is a constant. Returns the | ||||
1016 | // constant value in C and the other operand in Op. Returns the signedness of | ||||
1017 | // the remainder operation in IsSigned. Returns true if such a match is | ||||
1018 | // found. | ||||
1019 | static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) { | ||||
1020 | const APInt *AI; | ||||
1021 | IsSigned = false; | ||||
1022 | if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) { | ||||
1023 | IsSigned = true; | ||||
1024 | C = *AI; | ||||
1025 | return true; | ||||
1026 | } | ||||
1027 | if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) { | ||||
1028 | C = *AI; | ||||
1029 | return true; | ||||
1030 | } | ||||
1031 | if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) { | ||||
1032 | C = *AI + 1; | ||||
1033 | return true; | ||||
1034 | } | ||||
1035 | return false; | ||||
1036 | } | ||||
1037 | |||||
1038 | // Matches division expression Op / C with the given signedness as indicated | ||||
1039 | // by IsSigned, where C is a constant. Returns the constant value in C and the | ||||
1040 | // other operand in Op. Returns true if such a match is found. | ||||
1041 | static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) { | ||||
1042 | const APInt *AI; | ||||
1043 | if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) { | ||||
1044 | C = *AI; | ||||
1045 | return true; | ||||
1046 | } | ||||
1047 | if (!IsSigned) { | ||||
1048 | if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) { | ||||
1049 | C = *AI; | ||||
1050 | return true; | ||||
1051 | } | ||||
1052 | if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) { | ||||
1053 | C = APInt(AI->getBitWidth(), 1); | ||||
1054 | C <<= *AI; | ||||
1055 | return true; | ||||
1056 | } | ||||
1057 | } | ||||
1058 | return false; | ||||
1059 | } | ||||
1060 | |||||
1061 | // Returns whether C0 * C1 with the given signedness overflows. | ||||
1062 | static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) { | ||||
1063 | bool overflow; | ||||
1064 | if (IsSigned) | ||||
1065 | (void)C0.smul_ov(C1, overflow); | ||||
1066 | else | ||||
1067 | (void)C0.umul_ov(C1, overflow); | ||||
1068 | return overflow; | ||||
1069 | } | ||||
1070 | |||||
1071 | // Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1) | ||||
1072 | // does not overflow. | ||||
1073 | Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) { | ||||
1074 | Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); | ||||
1075 | Value *X, *MulOpV; | ||||
1076 | APInt C0, MulOpC; | ||||
1077 | bool IsSigned; | ||||
1078 | // Match I = X % C0 + MulOpV * C0 | ||||
1079 | if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) || | ||||
1080 | (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) && | ||||
1081 | C0 == MulOpC) { | ||||
1082 | Value *RemOpV; | ||||
1083 | APInt C1; | ||||
1084 | bool Rem2IsSigned; | ||||
1085 | // Match MulOpC = RemOpV % C1 | ||||
1086 | if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) && | ||||
1087 | IsSigned == Rem2IsSigned) { | ||||
1088 | Value *DivOpV; | ||||
1089 | APInt DivOpC; | ||||
1090 | // Match RemOpV = X / C0 | ||||
1091 | if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV && | ||||
1092 | C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) { | ||||
1093 | Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1); | ||||
1094 | return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem") | ||||
1095 | : Builder.CreateURem(X, NewDivisor, "urem"); | ||||
1096 | } | ||||
1097 | } | ||||
1098 | } | ||||
1099 | |||||
1100 | return nullptr; | ||||
1101 | } | ||||
1102 | |||||
1103 | /// Fold | ||||
1104 | /// (1 << NBits) - 1 | ||||
1105 | /// Into: | ||||
1106 | /// ~(-(1 << NBits)) | ||||
1107 | /// Because a 'not' is better for bit-tracking analysis and other transforms | ||||
1108 | /// than an 'add'. The new shl is always nsw, and is nuw if old `and` was. | ||||
1109 | static Instruction *canonicalizeLowbitMask(BinaryOperator &I, | ||||
1110 | InstCombiner::BuilderTy &Builder) { | ||||
1111 | Value *NBits; | ||||
1112 | if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes()))) | ||||
1113 | return nullptr; | ||||
1114 | |||||
1115 | Constant *MinusOne = Constant::getAllOnesValue(NBits->getType()); | ||||
1116 | Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask"); | ||||
1117 | // Be wary of constant folding. | ||||
1118 | if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) { | ||||
1119 | // Always NSW. But NUW propagates from `add`. | ||||
1120 | BOp->setHasNoSignedWrap(); | ||||
1121 | BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); | ||||
1122 | } | ||||
1123 | |||||
1124 | return BinaryOperator::CreateNot(NotMask, I.getName()); | ||||
1125 | } | ||||
1126 | |||||
1127 | static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) { | ||||
1128 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1128, __extension__ __PRETTY_FUNCTION__)); | ||||
1129 | Type *Ty = I.getType(); | ||||
1130 | auto getUAddSat = [&]() { | ||||
1131 | return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty); | ||||
1132 | }; | ||||
1133 | |||||
1134 | // add (umin X, ~Y), Y --> uaddsat X, Y | ||||
1135 | Value *X, *Y; | ||||
1136 | if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))), | ||||
1137 | m_Deferred(Y)))) | ||||
1138 | return CallInst::Create(getUAddSat(), { X, Y }); | ||||
1139 | |||||
1140 | // add (umin X, ~C), C --> uaddsat X, C | ||||
1141 | const APInt *C, *NotC; | ||||
1142 | if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) && | ||||
1143 | *C == ~*NotC) | ||||
1144 | return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) }); | ||||
1145 | |||||
1146 | return nullptr; | ||||
1147 | } | ||||
1148 | |||||
1149 | Instruction *InstCombinerImpl:: | ||||
1150 | canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract( | ||||
1151 | BinaryOperator &I) { | ||||
1152 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1155, __extension__ __PRETTY_FUNCTION__)) | ||||
| |||||
1153 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1155, __extension__ __PRETTY_FUNCTION__)) | ||||
1154 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1155, __extension__ __PRETTY_FUNCTION__)) | ||||
1155 | "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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1155, __extension__ __PRETTY_FUNCTION__)); | ||||
1156 | |||||
1157 | // We have a subtraction/addition between a (potentially truncated) *logical* | ||||
1158 | // right-shift of X and a "select". | ||||
1159 | Value *X, *Select; | ||||
1160 | Instruction *LowBitsToSkip, *Extract; | ||||
1161 | if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd( | ||||
| |||||
| |||||
| |||||
| |||||
1162 | m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)), | ||||
1163 | m_Instruction(Extract))), | ||||
1164 | m_Value(Select)))) | ||||
1165 | return nullptr; | ||||
1166 | |||||
1167 | // `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS. | ||||
1168 | if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select) | ||||
1169 | return nullptr; | ||||
1170 | |||||
1171 | Type *XTy = X->getType(); | ||||
1172 | bool HadTrunc = I.getType() != XTy; | ||||
1173 | |||||
1174 | // If there was a truncation of extracted value, then we'll need to produce | ||||
1175 | // one extra instruction, so we need to ensure one instruction will go away. | ||||
1176 | if (HadTrunc
| ||||
1177 | return nullptr; | ||||
1178 | |||||
1179 | // Extraction should extract high NBits bits, with shift amount calculated as: | ||||
1180 | // low bits to skip = shift bitwidth - high bits to extract | ||||
1181 | // The shift amount itself may be extended, and we need to look past zero-ext | ||||
1182 | // when matching NBits, that will matter for matching later. | ||||
1183 | Constant *C; | ||||
1184 | Value *NBits; | ||||
1185 | if (!match( | ||||
1186 | LowBitsToSkip, | ||||
1187 | m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) || | ||||
1188 | !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, | ||||
1189 | APInt(C->getType()->getScalarSizeInBits(), | ||||
1190 | X->getType()->getScalarSizeInBits())))) | ||||
1191 | return nullptr; | ||||
1192 | |||||
1193 | // Sign-extending value can be zero-extended if we `sub`tract it, | ||||
1194 | // or sign-extended otherwise. | ||||
1195 | auto SkipExtInMagic = [&I](Value *&V) { | ||||
1196 | if (I.getOpcode() == Instruction::Sub) | ||||
1197 | match(V, m_ZExtOrSelf(m_Value(V))); | ||||
1198 | else | ||||
1199 | match(V, m_SExtOrSelf(m_Value(V))); | ||||
1200 | }; | ||||
1201 | |||||
1202 | // Now, finally validate the sign-extending magic. | ||||
1203 | // `select` itself may be appropriately extended, look past that. | ||||
1204 | SkipExtInMagic(Select); | ||||
1205 | |||||
1206 | ICmpInst::Predicate Pred; | ||||
1207 | const APInt *Thr; | ||||
1208 | Value *SignExtendingValue, *Zero; | ||||
1209 | bool ShouldSignext; | ||||
1210 | // It must be a select between two values we will later establish to be a | ||||
1211 | // sign-extending value and a zero constant. The condition guarding the | ||||
1212 | // sign-extension must be based on a sign bit of the same X we had in `lshr`. | ||||
1213 | if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)), | ||||
| |||||
1214 | m_Value(SignExtendingValue), m_Value(Zero))) || | ||||
1215 | !isSignBitCheck(Pred, *Thr, ShouldSignext)) | ||||
1216 | return nullptr; | ||||
1217 | |||||
1218 | // icmp-select pair is commutative. | ||||
1219 | if (!ShouldSignext) | ||||
1220 | std::swap(SignExtendingValue, Zero); | ||||
1221 | |||||
1222 | // If we should not perform sign-extension then we must add/or/subtract zero. | ||||
1223 | if (!match(Zero, m_Zero())) | ||||
1224 | return nullptr; | ||||
1225 | // Otherwise, it should be some constant, left-shifted by the same NBits we | ||||
1226 | // had in `lshr`. Said left-shift can also be appropriately extended. | ||||
1227 | // Again, we must look past zero-ext when looking for NBits. | ||||
1228 | SkipExtInMagic(SignExtendingValue); | ||||
1229 | Constant *SignExtendingValueBaseConstant; | ||||
1230 | if (!match(SignExtendingValue, | ||||
1231 | m_Shl(m_Constant(SignExtendingValueBaseConstant), | ||||
1232 | m_ZExtOrSelf(m_Specific(NBits))))) | ||||
1233 | return nullptr; | ||||
1234 | // If we `sub`, then the constant should be one, else it should be all-ones. | ||||
1235 | if (I.getOpcode() == Instruction::Sub | ||||
1236 | ? !match(SignExtendingValueBaseConstant, m_One()) | ||||
1237 | : !match(SignExtendingValueBaseConstant, m_AllOnes())) | ||||
1238 | return nullptr; | ||||
1239 | |||||
1240 | auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip, | ||||
1241 | Extract->getName() + ".sext"); | ||||
1242 | NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness. | ||||
1243 | if (!HadTrunc) | ||||
1244 | return NewAShr; | ||||
1245 | |||||
1246 | Builder.Insert(NewAShr); | ||||
1247 | return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType()); | ||||
1248 | } | ||||
1249 | |||||
1250 | /// This is a specialization of a more general transform from | ||||
1251 | /// SimplifyUsingDistributiveLaws. If that code can be made to work optimally | ||||
1252 | /// for multi-use cases or propagating nsw/nuw, then we would not need this. | ||||
1253 | static Instruction *factorizeMathWithShlOps(BinaryOperator &I, | ||||
1254 | InstCombiner::BuilderTy &Builder) { | ||||
1255 | // TODO: Also handle mul by doubling the shift amount? | ||||
1256 | assert((I.getOpcode() == Instruction::Add ||(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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1258, __extension__ __PRETTY_FUNCTION__)) | ||||
1257 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1258, __extension__ __PRETTY_FUNCTION__)) | ||||
1258 | "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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1258, __extension__ __PRETTY_FUNCTION__)); | ||||
1259 | auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0)); | ||||
1260 | auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)); | ||||
1261 | if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse())) | ||||
1262 | return nullptr; | ||||
1263 | |||||
1264 | Value *X, *Y, *ShAmt; | ||||
1265 | if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) || | ||||
1266 | !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt)))) | ||||
1267 | return nullptr; | ||||
1268 | |||||
1269 | // No-wrap propagates only when all ops have no-wrap. | ||||
1270 | bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() && | ||||
1271 | Op1->hasNoSignedWrap(); | ||||
1272 | bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() && | ||||
1273 | Op1->hasNoUnsignedWrap(); | ||||
1274 | |||||
1275 | // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt | ||||
1276 | Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y); | ||||
1277 | if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) { | ||||
1278 | NewI->setHasNoSignedWrap(HasNSW); | ||||
1279 | NewI->setHasNoUnsignedWrap(HasNUW); | ||||
1280 | } | ||||
1281 | auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt); | ||||
1282 | NewShl->setHasNoSignedWrap(HasNSW); | ||||
1283 | NewShl->setHasNoUnsignedWrap(HasNUW); | ||||
1284 | return NewShl; | ||||
1285 | } | ||||
1286 | |||||
1287 | Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) { | ||||
1288 | if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1), | ||||
1289 | I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), | ||||
1290 | SQ.getWithInstruction(&I))) | ||||
1291 | return replaceInstUsesWith(I, V); | ||||
1292 | |||||
1293 | if (SimplifyAssociativeOrCommutative(I)) | ||||
1294 | return &I; | ||||
1295 | |||||
1296 | if (Instruction *X = foldVectorBinop(I)) | ||||
1297 | return X; | ||||
1298 | |||||
1299 | // (A*B)+(A*C) -> A*(B+C) etc | ||||
1300 | if (Value *V = SimplifyUsingDistributiveLaws(I)) | ||||
1301 | return replaceInstUsesWith(I, V); | ||||
1302 | |||||
1303 | if (Instruction *R = factorizeMathWithShlOps(I, Builder)) | ||||
1304 | return R; | ||||
1305 | |||||
1306 | if (Instruction *X = foldAddWithConstant(I)) | ||||
1307 | return X; | ||||
1308 | |||||
1309 | if (Instruction *X = foldNoWrapAdd(I, Builder)) | ||||
1310 | return X; | ||||
1311 | |||||
1312 | Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); | ||||
1313 | Type *Ty = I.getType(); | ||||
1314 | if (Ty->isIntOrIntVectorTy(1)) | ||||
1315 | return BinaryOperator::CreateXor(LHS, RHS); | ||||
1316 | |||||
1317 | // X + X --> X << 1 | ||||
1318 | if (LHS == RHS) { | ||||
1319 | auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1)); | ||||
1320 | Shl->setHasNoSignedWrap(I.hasNoSignedWrap()); | ||||
1321 | Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); | ||||
1322 | return Shl; | ||||
1323 | } | ||||
1324 | |||||
1325 | Value *A, *B; | ||||
1326 | if (match(LHS, m_Neg(m_Value(A)))) { | ||||
1327 | // -A + -B --> -(A + B) | ||||
1328 | if (match(RHS, m_Neg(m_Value(B)))) | ||||
1329 | return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B)); | ||||
1330 | |||||
1331 | // -A + B --> B - A | ||||
1332 | return BinaryOperator::CreateSub(RHS, A); | ||||
1333 | } | ||||
1334 | |||||
1335 | // A + -B --> A - B | ||||
1336 | if (match(RHS, m_Neg(m_Value(B)))) | ||||
1337 | return BinaryOperator::CreateSub(LHS, B); | ||||
1338 | |||||
1339 | if (Value *V = checkForNegativeOperand(I, Builder)) | ||||
1340 | return replaceInstUsesWith(I, V); | ||||
1341 | |||||
1342 | // (A + 1) + ~B --> A - B | ||||
1343 | // ~B + (A + 1) --> A - B | ||||
1344 | // (~B + A) + 1 --> A - B | ||||
1345 | // (A + ~B) + 1 --> A - B | ||||
1346 | if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) || | ||||
1347 | match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One()))) | ||||
1348 | return BinaryOperator::CreateSub(A, B); | ||||
1349 | |||||
1350 | // (A + RHS) + RHS --> A + (RHS << 1) | ||||
1351 | if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS))))) | ||||
1352 | return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add")); | ||||
1353 | |||||
1354 | // LHS + (A + LHS) --> A + (LHS << 1) | ||||
1355 | if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS))))) | ||||
1356 | return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add")); | ||||
1357 | |||||
1358 | { | ||||
1359 | // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2) | ||||
1360 | Constant *C1, *C2; | ||||
1361 | if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)), | ||||
1362 | m_Sub(m_ImmConstant(C2), m_Value(B)))) && | ||||
1363 | (LHS->hasOneUse() || RHS->hasOneUse())) { | ||||
1364 | Value *Sub = Builder.CreateSub(A, B); | ||||
1365 | return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2)); | ||||
1366 | } | ||||
1367 | } | ||||
1368 | |||||
1369 | // X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1) | ||||
1370 | if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V); | ||||
1371 | |||||
1372 | // ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2 | ||||
1373 | const APInt *C1, *C2; | ||||
1374 | if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) { | ||||
1375 | APInt one(C2->getBitWidth(), 1); | ||||
1376 | APInt minusC1 = -(*C1); | ||||
1377 | if (minusC1 == (one << *C2)) { | ||||
1378 | Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1); | ||||
1379 | return BinaryOperator::CreateSRem(RHS, NewRHS); | ||||
1380 | } | ||||
1381 | } | ||||
1382 | |||||
1383 | // A+B --> A|B iff A and B have no bits set in common. | ||||
1384 | if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT)) | ||||
1385 | return BinaryOperator::CreateOr(LHS, RHS); | ||||
1386 | |||||
1387 | // add (select X 0 (sub n A)) A --> select X A n | ||||
1388 | { | ||||
1389 | SelectInst *SI = dyn_cast<SelectInst>(LHS); | ||||
1390 | Value *A = RHS; | ||||
1391 | if (!SI) { | ||||
1392 | SI = dyn_cast<SelectInst>(RHS); | ||||
1393 | A = LHS; | ||||
1394 | } | ||||
1395 | if (SI && SI->hasOneUse()) { | ||||
1396 | Value *TV = SI->getTrueValue(); | ||||
1397 | Value *FV = SI->getFalseValue(); | ||||
1398 | Value *N; | ||||
1399 | |||||
1400 | // Can we fold the add into the argument of the select? | ||||
1401 | // We check both true and false select arguments for a matching subtract. | ||||
1402 | if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A)))) | ||||
1403 | // Fold the add into the true select value. | ||||
1404 | return SelectInst::Create(SI->getCondition(), N, A); | ||||
1405 | |||||
1406 | if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A)))) | ||||
1407 | // Fold the add into the false select value. | ||||
1408 | return SelectInst::Create(SI->getCondition(), A, N); | ||||
1409 | } | ||||
1410 | } | ||||
1411 | |||||
1412 | if (Instruction *Ext = narrowMathIfNoOverflow(I)) | ||||
1413 | return Ext; | ||||
1414 | |||||
1415 | // (add (xor A, B) (and A, B)) --> (or A, B) | ||||
1416 | // (add (and A, B) (xor A, B)) --> (or A, B) | ||||
1417 | if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)), | ||||
1418 | m_c_And(m_Deferred(A), m_Deferred(B))))) | ||||
1419 | return BinaryOperator::CreateOr(A, B); | ||||
1420 | |||||
1421 | // (add (or A, B) (and A, B)) --> (add A, B) | ||||
1422 | // (add (and A, B) (or A, B)) --> (add A, B) | ||||
1423 | if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)), | ||||
1424 | m_c_And(m_Deferred(A), m_Deferred(B))))) { | ||||
1425 | // Replacing operands in-place to preserve nuw/nsw flags. | ||||
1426 | replaceOperand(I, 0, A); | ||||
1427 | replaceOperand(I, 1, B); | ||||
1428 | return &I; | ||||
1429 | } | ||||
1430 | |||||
1431 | // TODO(jingyue): Consider willNotOverflowSignedAdd and | ||||
1432 | // willNotOverflowUnsignedAdd to reduce the number of invocations of | ||||
1433 | // computeKnownBits. | ||||
1434 | bool Changed = false; | ||||
1435 | if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) { | ||||
1436 | Changed = true; | ||||
1437 | I.setHasNoSignedWrap(true); | ||||
1438 | } | ||||
1439 | if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) { | ||||
1440 | Changed = true; | ||||
1441 | I.setHasNoUnsignedWrap(true); | ||||
1442 | } | ||||
1443 | |||||
1444 | if (Instruction *V = canonicalizeLowbitMask(I, Builder)) | ||||
1445 | return V; | ||||
1446 | |||||
1447 | if (Instruction *V = | ||||
1448 | canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I)) | ||||
1449 | return V; | ||||
1450 | |||||
1451 | if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I)) | ||||
1452 | return SatAdd; | ||||
1453 | |||||
1454 | // usub.sat(A, B) + B => umax(A, B) | ||||
1455 | if (match(&I, m_c_BinOp( | ||||
1456 | m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))), | ||||
1457 | m_Deferred(B)))) { | ||||
1458 | return replaceInstUsesWith(I, | ||||
1459 | Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B})); | ||||
1460 | } | ||||
1461 | |||||
1462 | // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common. | ||||
1463 | if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) && | ||||
1464 | match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) && | ||||
1465 | haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT)) | ||||
1466 | return replaceInstUsesWith( | ||||
1467 | I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()}, | ||||
1468 | {Builder.CreateOr(A, B)})); | ||||
1469 | |||||
1470 | return Changed ? &I : nullptr; | ||||
1471 | } | ||||
1472 | |||||
1473 | /// Eliminate an op from a linear interpolation (lerp) pattern. | ||||
1474 | static Instruction *factorizeLerp(BinaryOperator &I, | ||||
1475 | InstCombiner::BuilderTy &Builder) { | ||||
1476 | Value *X, *Y, *Z; | ||||
1477 | if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y), | ||||
1478 | m_OneUse(m_FSub(m_FPOne(), | ||||
1479 | m_Value(Z))))), | ||||
1480 | m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z)))))) | ||||
1481 | return nullptr; | ||||
1482 | |||||
1483 | // (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants] | ||||
1484 | Value *XY = Builder.CreateFSubFMF(X, Y, &I); | ||||
1485 | Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I); | ||||
1486 | return BinaryOperator::CreateFAddFMF(Y, MulZ, &I); | ||||
1487 | } | ||||
1488 | |||||
1489 | /// Factor a common operand out of fadd/fsub of fmul/fdiv. | ||||
1490 | static Instruction *factorizeFAddFSub(BinaryOperator &I, | ||||
1491 | InstCombiner::BuilderTy &Builder) { | ||||
1492 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1493, __extension__ __PRETTY_FUNCTION__)) | ||||
1493 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1493, __extension__ __PRETTY_FUNCTION__)); | ||||
1494 | 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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1495, __extension__ __PRETTY_FUNCTION__)) | ||||
1495 | "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\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1495, __extension__ __PRETTY_FUNCTION__)); | ||||
1496 | |||||
1497 | if (Instruction *Lerp = factorizeLerp(I, Builder)) | ||||
1498 | return Lerp; | ||||
1499 | |||||
1500 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
1501 | Value *X, *Y, *Z; | ||||
1502 | bool IsFMul; | ||||
1503 | if ((match(Op0, m_OneUse(m_FMul(m_Value(X), m_Value(Z)))) && | ||||
1504 | match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))) || | ||||
1505 | (match(Op0, m_OneUse(m_FMul(m_Value(Z), m_Value(X)))) && | ||||
1506 | match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z)))))) | ||||
1507 | IsFMul = true; | ||||
1508 | else if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Z)))) && | ||||
1509 | match(Op1, m_OneUse(m_FDiv(m_Value(Y), m_Specific(Z))))) | ||||
1510 | IsFMul = false; | ||||
1511 | else | ||||
1512 | return nullptr; | ||||
1513 | |||||
1514 | // (X * Z) + (Y * Z) --> (X + Y) * Z | ||||
1515 | // (X * Z) - (Y * Z) --> (X - Y) * Z | ||||
1516 | // (X / Z) + (Y / Z) --> (X + Y) / Z | ||||
1517 | // (X / Z) - (Y / Z) --> (X - Y) / Z | ||||
1518 | bool IsFAdd = I.getOpcode() == Instruction::FAdd; | ||||
1519 | Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I) | ||||
1520 | : Builder.CreateFSubFMF(X, Y, &I); | ||||
1521 | |||||
1522 | // Bail out if we just created a denormal constant. | ||||
1523 | // TODO: This is copied from a previous implementation. Is it necessary? | ||||
1524 | const APFloat *C; | ||||
1525 | if (match(XY, m_APFloat(C)) && !C->isNormal()) | ||||
1526 | return nullptr; | ||||
1527 | |||||
1528 | return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I) | ||||
1529 | : BinaryOperator::CreateFDivFMF(XY, Z, &I); | ||||
1530 | } | ||||
1531 | |||||
1532 | Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) { | ||||
1533 | if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1), | ||||
1534 | I.getFastMathFlags(), | ||||
1535 | SQ.getWithInstruction(&I))) | ||||
1536 | return replaceInstUsesWith(I, V); | ||||
1537 | |||||
1538 | if (SimplifyAssociativeOrCommutative(I)) | ||||
1539 | return &I; | ||||
1540 | |||||
1541 | if (Instruction *X = foldVectorBinop(I)) | ||||
1542 | return X; | ||||
1543 | |||||
1544 | if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I)) | ||||
1545 | return FoldedFAdd; | ||||
1546 | |||||
1547 | // (-X) + Y --> Y - X | ||||
1548 | Value *X, *Y; | ||||
1549 | if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y)))) | ||||
1550 | return BinaryOperator::CreateFSubFMF(Y, X, &I); | ||||
1551 | |||||
1552 | // Similar to above, but look through fmul/fdiv for the negated term. | ||||
1553 | // (-X * Y) + Z --> Z - (X * Y) [4 commuted variants] | ||||
1554 | Value *Z; | ||||
1555 | if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))), | ||||
1556 | m_Value(Z)))) { | ||||
1557 | Value *XY = Builder.CreateFMulFMF(X, Y, &I); | ||||
1558 | return BinaryOperator::CreateFSubFMF(Z, XY, &I); | ||||
1559 | } | ||||
1560 | // (-X / Y) + Z --> Z - (X / Y) [2 commuted variants] | ||||
1561 | // (X / -Y) + Z --> Z - (X / Y) [2 commuted variants] | ||||
1562 | if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))), | ||||
1563 | m_Value(Z))) || | ||||
1564 | match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))), | ||||
1565 | m_Value(Z)))) { | ||||
1566 | Value *XY = Builder.CreateFDivFMF(X, Y, &I); | ||||
1567 | return BinaryOperator::CreateFSubFMF(Z, XY, &I); | ||||
1568 | } | ||||
1569 | |||||
1570 | // Check for (fadd double (sitofp x), y), see if we can merge this into an | ||||
1571 | // integer add followed by a promotion. | ||||
1572 | Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); | ||||
1573 | if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) { | ||||
1574 | Value *LHSIntVal = LHSConv->getOperand(0); | ||||
1575 | Type *FPType = LHSConv->getType(); | ||||
1576 | |||||
1577 | // TODO: This check is overly conservative. In many cases known bits | ||||
1578 | // analysis can tell us that the result of the addition has less significant | ||||
1579 | // bits than the integer type can hold. | ||||
1580 | auto IsValidPromotion = [](Type *FTy, Type *ITy) { | ||||
1581 | Type *FScalarTy = FTy->getScalarType(); | ||||
1582 | Type *IScalarTy = ITy->getScalarType(); | ||||
1583 | |||||
1584 | // Do we have enough bits in the significand to represent the result of | ||||
1585 | // the integer addition? | ||||
1586 | unsigned MaxRepresentableBits = | ||||
1587 | APFloat::semanticsPrecision(FScalarTy->getFltSemantics()); | ||||
1588 | return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits; | ||||
1589 | }; | ||||
1590 | |||||
1591 | // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst)) | ||||
1592 | // ... if the constant fits in the integer value. This is useful for things | ||||
1593 | // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer | ||||
1594 | // requires a constant pool load, and generally allows the add to be better | ||||
1595 | // instcombined. | ||||
1596 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) | ||||
1597 | if (IsValidPromotion(FPType, LHSIntVal->getType())) { | ||||
1598 | Constant *CI = | ||||
1599 | ConstantExpr::getFPToSI(CFP, LHSIntVal->getType()); | ||||
1600 | if (LHSConv->hasOneUse() && | ||||
1601 | ConstantExpr::getSIToFP(CI, I.getType()) == CFP && | ||||
1602 | willNotOverflowSignedAdd(LHSIntVal, CI, I)) { | ||||
1603 | // Insert the new integer add. | ||||
1604 | Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv"); | ||||
1605 | return new SIToFPInst(NewAdd, I.getType()); | ||||
1606 | } | ||||
1607 | } | ||||
1608 | |||||
1609 | // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y)) | ||||
1610 | if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) { | ||||
1611 | Value *RHSIntVal = RHSConv->getOperand(0); | ||||
1612 | // It's enough to check LHS types only because we require int types to | ||||
1613 | // be the same for this transform. | ||||
1614 | if (IsValidPromotion(FPType, LHSIntVal->getType())) { | ||||
1615 | // Only do this if x/y have the same type, if at least one of them has a | ||||
1616 | // single use (so we don't increase the number of int->fp conversions), | ||||
1617 | // and if the integer add will not overflow. | ||||
1618 | if (LHSIntVal->getType() == RHSIntVal->getType() && | ||||
1619 | (LHSConv->hasOneUse() || RHSConv->hasOneUse()) && | ||||
1620 | willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) { | ||||
1621 | // Insert the new integer add. | ||||
1622 | Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv"); | ||||
1623 | return new SIToFPInst(NewAdd, I.getType()); | ||||
1624 | } | ||||
1625 | } | ||||
1626 | } | ||||
1627 | } | ||||
1628 | |||||
1629 | // Handle specials cases for FAdd with selects feeding the operation | ||||
1630 | if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS)) | ||||
1631 | return replaceInstUsesWith(I, V); | ||||
1632 | |||||
1633 | if (I.hasAllowReassoc() && I.hasNoSignedZeros()) { | ||||
1634 | if (Instruction *F = factorizeFAddFSub(I, Builder)) | ||||
1635 | return F; | ||||
1636 | |||||
1637 | // Try to fold fadd into start value of reduction intrinsic. | ||||
1638 | if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>( | ||||
1639 | m_AnyZeroFP(), m_Value(X))), | ||||
1640 | m_Value(Y)))) { | ||||
1641 | // fadd (rdx 0.0, X), Y --> rdx Y, X | ||||
1642 | return replaceInstUsesWith( | ||||
1643 | I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd, | ||||
1644 | {X->getType()}, {Y, X}, &I)); | ||||
1645 | } | ||||
1646 | const APFloat *StartC, *C; | ||||
1647 | if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>( | ||||
1648 | m_APFloat(StartC), m_Value(X)))) && | ||||
1649 | match(RHS, m_APFloat(C))) { | ||||
1650 | // fadd (rdx StartC, X), C --> rdx (C + StartC), X | ||||
1651 | Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC); | ||||
1652 | return replaceInstUsesWith( | ||||
1653 | I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd, | ||||
1654 | {X->getType()}, {NewStartC, X}, &I)); | ||||
1655 | } | ||||
1656 | |||||
1657 | if (Value *V = FAddCombine(Builder).simplify(&I)) | ||||
1658 | return replaceInstUsesWith(I, V); | ||||
1659 | } | ||||
1660 | |||||
1661 | return nullptr; | ||||
1662 | } | ||||
1663 | |||||
1664 | /// Optimize pointer differences into the same array into a size. Consider: | ||||
1665 | /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer | ||||
1666 | /// operands to the ptrtoint instructions for the LHS/RHS of the subtract. | ||||
1667 | Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS, | ||||
1668 | Type *Ty, bool IsNUW) { | ||||
1669 | // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize | ||||
1670 | // this. | ||||
1671 | bool Swapped = false; | ||||
1672 | GEPOperator *GEP1 = nullptr, *GEP2 = nullptr; | ||||
1673 | if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) { | ||||
1674 | std::swap(LHS, RHS); | ||||
1675 | Swapped = true; | ||||
1676 | } | ||||
1677 | |||||
1678 | // Require at least one GEP with a common base pointer on both sides. | ||||
1679 | if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) { | ||||
1680 | // (gep X, ...) - X | ||||
1681 | if (LHSGEP->getOperand(0) == RHS) { | ||||
1682 | GEP1 = LHSGEP; | ||||
1683 | } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) { | ||||
1684 | // (gep X, ...) - (gep X, ...) | ||||
1685 | if (LHSGEP->getOperand(0)->stripPointerCasts() == | ||||
1686 | RHSGEP->getOperand(0)->stripPointerCasts()) { | ||||
1687 | GEP1 = LHSGEP; | ||||
1688 | GEP2 = RHSGEP; | ||||
1689 | } | ||||
1690 | } | ||||
1691 | } | ||||
1692 | |||||
1693 | if (!GEP1) | ||||
1694 | return nullptr; | ||||
1695 | |||||
1696 | if (GEP2) { | ||||
1697 | // (gep X, ...) - (gep X, ...) | ||||
1698 | // | ||||
1699 | // Avoid duplicating the arithmetic if there are more than one non-constant | ||||
1700 | // indices between the two GEPs and either GEP has a non-constant index and | ||||
1701 | // multiple users. If zero non-constant index, the result is a constant and | ||||
1702 | // there is no duplication. If one non-constant index, the result is an add | ||||
1703 | // or sub with a constant, which is no larger than the original code, and | ||||
1704 | // there's no duplicated arithmetic, even if either GEP has multiple | ||||
1705 | // users. If more than one non-constant indices combined, as long as the GEP | ||||
1706 | // with at least one non-constant index doesn't have multiple users, there | ||||
1707 | // is no duplication. | ||||
1708 | unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices(); | ||||
1709 | unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices(); | ||||
1710 | if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 && | ||||
1711 | ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) || | ||||
1712 | (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) { | ||||
1713 | return nullptr; | ||||
1714 | } | ||||
1715 | } | ||||
1716 | |||||
1717 | // Emit the offset of the GEP and an intptr_t. | ||||
1718 | Value *Result = EmitGEPOffset(GEP1); | ||||
1719 | |||||
1720 | // If this is a single inbounds GEP and the original sub was nuw, | ||||
1721 | // then the final multiplication is also nuw. | ||||
1722 | if (auto *I = dyn_cast<Instruction>(Result)) | ||||
1723 | if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() && | ||||
1724 | I->getOpcode() == Instruction::Mul) | ||||
1725 | I->setHasNoUnsignedWrap(); | ||||
1726 | |||||
1727 | // If we have a 2nd GEP of the same base pointer, subtract the offsets. | ||||
1728 | // If both GEPs are inbounds, then the subtract does not have signed overflow. | ||||
1729 | if (GEP2) { | ||||
1730 | Value *Offset = EmitGEPOffset(GEP2); | ||||
1731 | Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false, | ||||
1732 | GEP1->isInBounds() && GEP2->isInBounds()); | ||||
1733 | } | ||||
1734 | |||||
1735 | // If we have p - gep(p, ...) then we have to negate the result. | ||||
1736 | if (Swapped) | ||||
1737 | Result = Builder.CreateNeg(Result, "diff.neg"); | ||||
1738 | |||||
1739 | return Builder.CreateIntCast(Result, Ty, true); | ||||
1740 | } | ||||
1741 | |||||
1742 | Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) { | ||||
1743 | if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1), | ||||
1744 | I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), | ||||
1745 | SQ.getWithInstruction(&I))) | ||||
1746 | return replaceInstUsesWith(I, V); | ||||
1747 | |||||
1748 | if (Instruction *X = foldVectorBinop(I)) | ||||
1749 | return X; | ||||
1750 | |||||
1751 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
1752 | |||||
1753 | // If this is a 'B = x-(-A)', change to B = x+A. | ||||
1754 | // We deal with this without involving Negator to preserve NSW flag. | ||||
1755 | if (Value *V = dyn_castNegVal(Op1)) { | ||||
1756 | BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V); | ||||
1757 | |||||
1758 | if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) { | ||||
1759 | 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!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1760, __extension__ __PRETTY_FUNCTION__)) | ||||
1760 | "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!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp" , 1760, __extension__ __PRETTY_FUNCTION__)); | ||||
1761 | if (BO->hasNoSignedWrap() && I.hasNoSignedWrap()) | ||||
1762 | Res->setHasNoSignedWrap(true); | ||||
1763 | } else { | ||||
1764 | if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap()) | ||||
1765 | Res->setHasNoSignedWrap(true); | ||||
1766 | } | ||||
1767 | |||||
1768 | return Res; | ||||
1769 | } | ||||
1770 | |||||
1771 | // Try this before Negator to preserve NSW flag. | ||||
1772 | if (Instruction *R = factorizeMathWithShlOps(I, Builder)) | ||||
1773 | return R; | ||||
1774 | |||||
1775 | Constant *C; | ||||
1776 | if (match(Op0, m_ImmConstant(C))) { | ||||
1777 | Value *X; | ||||
1778 | Constant *C2; | ||||
1779 | |||||
1780 | // C-(X+C2) --> (C-C2)-X | ||||
1781 | if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2)))) | ||||
1782 | return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X); | ||||
1783 | } | ||||
1784 | |||||
1785 | auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * { | ||||
1786 | if (Instruction *Ext = narrowMathIfNoOverflow(I)) | ||||
1787 | return Ext; | ||||
1788 | |||||
1789 | bool Changed = false; | ||||
1790 | if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) { | ||||
1791 | Changed = true; | ||||
1792 | I.setHasNoSignedWrap(true); | ||||
1793 | } | ||||
1794 | if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) { | ||||
1795 | Changed = true; | ||||
1796 | I.setHasNoUnsignedWrap(true); | ||||
1797 | } | ||||
1798 | |||||
1799 | return Changed ? &I : nullptr; | ||||
1800 | }; | ||||
1801 | |||||
1802 | // First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`, | ||||
1803 | // and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't | ||||
1804 | // a pure negation used by a select that looks like abs/nabs. | ||||
1805 | bool IsNegation = match(Op0, m_ZeroInt()); | ||||
1806 | if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) { | ||||
1807 | const Instruction *UI = dyn_cast<Instruction>(U); | ||||
1808 | if (!UI) | ||||
1809 | return false; | ||||
1810 | return match(UI, | ||||
1811 | m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) || | ||||
1812 | match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1))); | ||||
1813 | })) { | ||||
1814 | if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this)) | ||||
1815 | return BinaryOperator::CreateAdd(NegOp1, Op0); | ||||
1816 | } | ||||
1817 | if (IsNegation) | ||||
1818 | return TryToNarrowDeduceFlags(); // Should have been handled in Negator! | ||||
1819 | |||||
1820 | // (A*B)-(A*C) -> A*(B-C) etc | ||||
1821 | if (Value *V = SimplifyUsingDistributiveLaws(I)) | ||||
1822 | return replaceInstUsesWith(I, V); | ||||
1823 | |||||
1824 | if (I.getType()->isIntOrIntVectorTy(1)) | ||||
1825 | return BinaryOperator::CreateXor(Op0, Op1); | ||||
1826 | |||||
1827 | // Replace (-1 - A) with (~A). | ||||
1828 | if (match(Op0, m_AllOnes())) | ||||
1829 | return BinaryOperator::CreateNot(Op1); | ||||
1830 | |||||
1831 | // (X + -1) - Y --> ~Y + X | ||||
1832 | Value *X, *Y; | ||||
1833 | if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes())))) | ||||
1834 | return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X); | ||||
1835 | |||||
1836 | // Reassociate sub/add sequences to create more add instructions and | ||||
1837 | // reduce dependency chains: | ||||
1838 | // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1) | ||||
1839 | Value *Z; | ||||
1840 | if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))), | ||||
1841 | m_Value(Z))))) { | ||||
1842 | Value *XZ = Builder.CreateAdd(X, Z); | ||||
1843 | Value *YW = Builder.CreateAdd(Y, Op1); | ||||
1844 | return BinaryOperator::CreateSub(XZ, YW); | ||||
1845 | } | ||||
1846 | |||||
1847 | // ((X - Y) - Op1) --> X - (Y + Op1) | ||||
1848 | if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) { | ||||
1849 | Value *Add = Builder.CreateAdd(Y, Op1); | ||||
1850 | return BinaryOperator::CreateSub(X, Add); | ||||
1851 | } | ||||
1852 | |||||
1853 | // (~X) - (~Y) --> Y - X | ||||
1854 | // This is placed after the other reassociations and explicitly excludes a | ||||
1855 | // sub-of-sub pattern to avoid infinite looping. | ||||
1856 | if (isFreeToInvert(Op0, Op0->hasOneUse()) && | ||||
1857 | isFreeToInvert(Op1, Op1->hasOneUse()) && | ||||
1858 | !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) { | ||||
1859 | Value *NotOp0 = Builder.CreateNot(Op0); | ||||
1860 | Value *NotOp1 = Builder.CreateNot(Op1); | ||||
1861 | return BinaryOperator::CreateSub(NotOp1, NotOp0); | ||||
1862 | } | ||||
1863 | |||||
1864 | auto m_AddRdx = [](Value *&Vec) { | ||||
1865 | return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec))); | ||||
1866 | }; | ||||
1867 | Value *V0, *V1; | ||||
1868 | if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) && | ||||
1869 | V0->getType() == V1->getType()) { | ||||
1870 | // Difference of sums is sum of differences: | ||||
1871 | // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1) | ||||
1872 | Value *Sub = Builder.CreateSub(V0, V1); | ||||
1873 | Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add, | ||||
1874 | {Sub->getType()}, {Sub}); | ||||
1875 | return replaceInstUsesWith(I, Rdx); | ||||
1876 | } | ||||
1877 | |||||
1878 | if (Constant *C = dyn_cast<Constant>(Op0)) { | ||||
1879 | Value *X; | ||||
1880 | if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) | ||||
1881 | // C - (zext bool) --> bool ? C - 1 : C | ||||
1882 | return SelectInst::Create(X, InstCombiner::SubOne(C), C); | ||||
1883 | if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) | ||||
1884 | // C - (sext bool) --> bool ? C + 1 : C | ||||
1885 | return SelectInst::Create(X, InstCombiner::AddOne(C), C); | ||||
1886 | |||||
1887 | // C - ~X == X + (1+C) | ||||
1888 | if (match(Op1, m_Not(m_Value(X)))) | ||||
1889 | return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C)); | ||||
1890 | |||||
1891 | // Try to fold constant sub into select arguments. | ||||
1892 | if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) | ||||
1893 | if (Instruction *R = FoldOpIntoSelect(I, SI)) | ||||
1894 | return R; | ||||
1895 | |||||
1896 | // Try to fold constant sub into PHI values. | ||||
1897 | if (PHINode *PN = dyn_cast<PHINode>(Op1)) | ||||
1898 | if (Instruction *R = foldOpIntoPhi(I, PN)) | ||||
1899 | return R; | ||||
1900 | |||||
1901 | Constant *C2; | ||||
1902 | |||||
1903 | // C-(C2-X) --> X+(C-C2) | ||||
1904 | if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X)))) | ||||
1905 | return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2)); | ||||
1906 | } | ||||
1907 | |||||
1908 | const APInt *Op0C; | ||||
1909 | if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) { | ||||
1910 | // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known | ||||
1911 | // zero. | ||||
1912 | KnownBits RHSKnown = computeKnownBits(Op1, 0, &I); | ||||
1913 | if ((*Op0C | RHSKnown.Zero).isAllOnesValue()) | ||||
1914 | return BinaryOperator::CreateXor(Op1, Op0); | ||||
1915 | } | ||||
1916 | |||||
1917 | { | ||||
1918 | Value *Y; | ||||
1919 | // X-(X+Y) == -Y X-(Y+X) == -Y | ||||
1920 | if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y)))) | ||||
1921 | return BinaryOperator::CreateNeg(Y); | ||||
1922 | |||||
1923 | // (X-Y)-X == -Y | ||||
1924 | if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y)))) | ||||
1925 | return BinaryOperator::CreateNeg(Y); | ||||
1926 | } | ||||
1927 | |||||
1928 | // (sub (or A, B) (and A, B)) --> (xor A, B) | ||||
1929 | { | ||||
1930 | Value *A, *B; | ||||
1931 | if (match(Op1, m_And(m_Value(A), m_Value(B))) && | ||||
1932 | match(Op0, m_c_Or(m_Specific(A), m_Specific(B)))) | ||||
1933 | return BinaryOperator::CreateXor(A, B); | ||||
1934 | } | ||||
1935 | |||||
1936 | // (sub (add A, B) (or A, B)) --> (and A, B) | ||||
1937 | { | ||||
1938 | Value *A, *B; | ||||
1939 | if (match(Op0, m_Add(m_Value(A), m_Value(B))) && | ||||
1940 | match(Op1, m_c_Or(m_Specific(A), m_Specific(B)))) | ||||
1941 | return BinaryOperator::CreateAnd(A, B); | ||||
1942 | } | ||||
1943 | |||||
1944 | // (sub (add A, B) (and A, B)) --> (or A, B) | ||||
1945 | { | ||||
1946 | Value *A, *B; | ||||
1947 | if (match(Op0, m_Add(m_Value(A), m_Value(B))) && | ||||
1948 | match(Op1, m_c_And(m_Specific(A), m_Specific(B)))) | ||||
1949 | return BinaryOperator::CreateOr(A, B); | ||||
1950 | } | ||||
1951 | |||||
1952 | // (sub (and A, B) (or A, B)) --> neg (xor A, B) | ||||
1953 | { | ||||
1954 | Value *A, *B; | ||||
1955 | if (match(Op0, m_And(m_Value(A), m_Value(B))) && | ||||
1956 | match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) && | ||||
1957 | (Op0->hasOneUse() || Op1->hasOneUse())) | ||||
1958 | return BinaryOperator::CreateNeg(Builder.CreateXor(A, B)); | ||||
1959 | } | ||||
1960 | |||||
1961 | // (sub (or A, B), (xor A, B)) --> (and A, B) | ||||
1962 | { | ||||
1963 | Value *A, *B; | ||||
1964 | if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && | ||||
1965 | match(Op0, m_c_Or(m_Specific(A), m_Specific(B)))) | ||||
1966 | return BinaryOperator::CreateAnd(A, B); | ||||
1967 | } | ||||
1968 | |||||
1969 | // (sub (xor A, B) (or A, B)) --> neg (and A, B) | ||||
1970 | { | ||||
1971 | Value *A, *B; | ||||
1972 | if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && | ||||
1973 | match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) && | ||||
1974 | (Op0->hasOneUse() || Op1->hasOneUse())) | ||||
1975 | return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B)); | ||||
1976 | } | ||||
1977 | |||||
1978 | { | ||||
1979 | Value *Y; | ||||
1980 | // ((X | Y) - X) --> (~X & Y) | ||||
1981 | if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1))))) | ||||
1982 | return BinaryOperator::CreateAnd( | ||||
1983 | Y, Builder.CreateNot(Op1, Op1->getName() + ".not")); | ||||
1984 | } | ||||
1985 | |||||
1986 | { | ||||
1987 | // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1)) | ||||
1988 | Value *X; | ||||
1989 | if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1), | ||||
1990 | m_OneUse(m_Neg(m_Value(X))))))) { | ||||
1991 | return BinaryOperator::CreateNeg(Builder.CreateAnd( | ||||
1992 | Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType())))); | ||||
1993 | } | ||||
1994 | } | ||||
1995 | |||||
1996 | { | ||||
1997 | // (sub (and Op1, C), Op1) --> neg (and Op1, ~C) | ||||
1998 | Constant *C; | ||||
1999 | if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) { | ||||
2000 | return BinaryOperator::CreateNeg( | ||||
2001 | Builder.CreateAnd(Op1, Builder.CreateNot(C))); | ||||
2002 | } | ||||
2003 | } | ||||
2004 | |||||
2005 | { | ||||
2006 | // If we have a subtraction between some value and a select between | ||||
2007 | // said value and something else, sink subtraction into select hands, i.e.: | ||||
2008 | // sub (select %Cond, %TrueVal, %FalseVal), %Op1 | ||||
2009 | // -> | ||||
2010 | // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1) | ||||
2011 | // or | ||||
2012 | // sub %Op0, (select %Cond, %TrueVal, %FalseVal) | ||||
2013 | // -> | ||||
2014 | // select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal) | ||||
2015 | // This will result in select between new subtraction and 0. | ||||
2016 | auto SinkSubIntoSelect = | ||||
2017 | [Ty = I.getType()](Value *Select, Value *OtherHandOfSub, | ||||
2018 | auto SubBuilder) -> Instruction * { | ||||
2019 | Value *Cond, *TrueVal, *FalseVal; | ||||
2020 | if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal), | ||||
2021 | m_Value(FalseVal))))) | ||||
2022 | return nullptr; | ||||
2023 | if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal) | ||||
2024 | return nullptr; | ||||
2025 | // While it is really tempting to just create two subtractions and let | ||||
2026 | // InstCombine fold one of those to 0, it isn't possible to do so | ||||
2027 | // because of worklist visitation order. So ugly it is. | ||||
2028 | bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal; | ||||
2029 | Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal); | ||||
2030 | Constant *Zero = Constant::getNullValue(Ty); | ||||
2031 | SelectInst *NewSel = | ||||
2032 | SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub, | ||||
2033 | OtherHandOfSubIsTrueVal ? NewSub : Zero); | ||||
2034 | // Preserve prof metadata if any. | ||||
2035 | NewSel->copyMetadata(cast<Instruction>(*Select)); | ||||
2036 | return NewSel; | ||||
2037 | }; | ||||
2038 | if (Instruction *NewSel = SinkSubIntoSelect( | ||||
2039 | /*Select=*/Op0, /*OtherHandOfSub=*/Op1, | ||||
2040 | [Builder = &Builder, Op1](Value *OtherHandOfSelect) { | ||||
2041 | return Builder->CreateSub(OtherHandOfSelect, | ||||
2042 | /*OtherHandOfSub=*/Op1); | ||||
2043 | })) | ||||
2044 | return NewSel; | ||||
2045 | if (Instruction *NewSel = SinkSubIntoSelect( | ||||
2046 | /*Select=*/Op1, /*OtherHandOfSub=*/Op0, | ||||
2047 | [Builder = &Builder, Op0](Value *OtherHandOfSelect) { | ||||
2048 | return Builder->CreateSub(/*OtherHandOfSub=*/Op0, | ||||
2049 | OtherHandOfSelect); | ||||
2050 | })) | ||||
2051 | return NewSel; | ||||
2052 | } | ||||
2053 | |||||
2054 | // (X - (X & Y)) --> (X & ~Y) | ||||
2055 | if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) && | ||||
2056 | (Op1->hasOneUse() || isa<Constant>(Y))) | ||||
2057 | return BinaryOperator::CreateAnd( | ||||
2058 | Op0, Builder.CreateNot(Y, Y->getName() + ".not")); | ||||
2059 | |||||
2060 | { | ||||
2061 | // ~A - Min/Max(~A, O) -> Max/Min(A, ~O) - A | ||||
2062 | // ~A - Min/Max(O, ~A) -> Max/Min(A, ~O) - A | ||||
2063 | // Min/Max(~A, O) - ~A -> A - Max/Min(A, ~O) | ||||
2064 | // Min/Max(O, ~A) - ~A -> A - Max/Min(A, ~O) | ||||
2065 | // So long as O here is freely invertible, this will be neutral or a win. | ||||
2066 | Value *LHS, *RHS, *A; | ||||
2067 | Value *NotA = Op0, *MinMax = Op1; | ||||
2068 | SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor; | ||||
2069 | if (!SelectPatternResult::isMinOrMax(SPF)) { | ||||
2070 | NotA = Op1; | ||||
2071 | MinMax = Op0; | ||||
2072 | SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor; | ||||
2073 | } | ||||
2074 | if (SelectPatternResult::isMinOrMax(SPF) && | ||||
2075 | match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) { | ||||
2076 | if (NotA == LHS) | ||||
2077 | std::swap(LHS, RHS); | ||||
2078 | // LHS is now O above and expected to have at least 2 uses (the min/max) | ||||
2079 | // NotA is epected to have 2 uses from the min/max and 1 from the sub. | ||||
2080 | if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) && | ||||
2081 | !NotA->hasNUsesOrMore(4)) { | ||||
2082 | // Note: We don't generate the inverse max/min, just create the not of | ||||
2083 | // it and let other folds do the rest. | ||||
2084 | Value *Not = Builder.CreateNot(MinMax); | ||||
2085 | if (NotA == Op0) | ||||
2086 | return BinaryOperator::CreateSub(Not, A); | ||||
2087 | else | ||||
2088 | return BinaryOperator::CreateSub(A, Not); | ||||
2089 | } | ||||
2090 | } | ||||
2091 | } | ||||
2092 | |||||
2093 | // Optimize pointer differences into the same array into a size. Consider: | ||||
2094 | // &A[10] - &A[0]: we should compile this to "10". | ||||
2095 | Value *LHSOp, *RHSOp; | ||||
2096 | if (match(Op0, m_PtrToInt(m_Value(LHSOp))) && | ||||
2097 | match(Op1, m_PtrToInt(m_Value(RHSOp)))) | ||||
2098 | if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(), | ||||
2099 | I.hasNoUnsignedWrap())) | ||||
2100 | return replaceInstUsesWith(I, Res); | ||||
2101 | |||||
2102 | // trunc(p)-trunc(q) -> trunc(p-q) | ||||
2103 | if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) && | ||||
2104 | match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp))))) | ||||
2105 | if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(), | ||||
2106 | /* IsNUW */ false)) | ||||
2107 | return replaceInstUsesWith(I, Res); | ||||
2108 | |||||
2109 | // Canonicalize a shifty way to code absolute value to the common pattern. | ||||
2110 | // There are 2 potential commuted variants. | ||||
2111 | // We're relying on the fact that we only do this transform when the shift has | ||||
2112 | // exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase | ||||
2113 | // instructions). | ||||
2114 | Value *A; | ||||
2115 | const APInt *ShAmt; | ||||
2116 | Type *Ty = I.getType(); | ||||
2117 | if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) && | ||||
2118 | Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 && | ||||
2119 | match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) { | ||||
2120 | // B = ashr i32 A, 31 ; smear the sign bit | ||||
2121 | // sub (xor A, B), B ; flip bits if negative and subtract -1 (add 1) | ||||
2122 | // --> (A < 0) ? -A : A | ||||
2123 | Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty)); | ||||
2124 | // Copy the nuw/nsw flags from the sub to the negate. | ||||
2125 | Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(), | ||||
2126 | I.hasNoSignedWrap()); | ||||
2127 | return SelectInst::Create(Cmp, Neg, A); | ||||
2128 | } | ||||
2129 | |||||
2130 | // If we are subtracting a low-bit masked subset of some value from an add | ||||
2131 | // of that same value with no low bits changed, that is clearing some low bits | ||||
2132 | // of the sum: | ||||
2133 | // sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC | ||||
2134 | const APInt *AddC, *AndC; | ||||
2135 | if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) && | ||||
2136 | match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) { | ||||
2137 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
2138 | unsigned Cttz = AddC->countTrailingZeros(); | ||||
2139 | APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz)); | ||||
2140 | if ((HighMask & *AndC).isNullValue()) | ||||
2141 | return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC))); | ||||
2142 | } | ||||
2143 | |||||
2144 | if (Instruction *V = | ||||
2145 | canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I)) | ||||
2146 | return V; | ||||
2147 | |||||
2148 | // X - usub.sat(X, Y) => umin(X, Y) | ||||
2149 | if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0), | ||||
2150 | m_Value(Y))))) | ||||
2151 | return replaceInstUsesWith( | ||||
2152 | I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y})); | ||||
2153 | |||||
2154 | // C - ctpop(X) => ctpop(~X) if C is bitwidth | ||||
2155 | if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) && | ||||
2156 | match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X))))) | ||||
2157 | return replaceInstUsesWith( | ||||
2158 | I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()}, | ||||
2159 | {Builder.CreateNot(X)})); | ||||
2160 | |||||
2161 | return TryToNarrowDeduceFlags(); | ||||
2162 | } | ||||
2163 | |||||
2164 | /// This eliminates floating-point negation in either 'fneg(X)' or | ||||
2165 | /// 'fsub(-0.0, X)' form by combining into a constant operand. | ||||
2166 | static Instruction *foldFNegIntoConstant(Instruction &I) { | ||||
2167 | // This is limited with one-use because fneg is assumed better for | ||||
2168 | // reassociation and cheaper in codegen than fmul/fdiv. | ||||
2169 | // TODO: Should the m_OneUse restriction be removed? | ||||
2170 | Instruction *FNegOp; | ||||
2171 | if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp))))) | ||||
2172 | return nullptr; | ||||
2173 | |||||
2174 | Value *X; | ||||
2175 | Constant *C; | ||||
2176 | |||||
2177 | // Fold negation into constant operand. | ||||
2178 | // -(X * C) --> X * (-C) | ||||
2179 | if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C)))) | ||||
2180 | return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I); | ||||
2181 | // -(X / C) --> X / (-C) | ||||
2182 | if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C)))) | ||||
2183 | return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I); | ||||
2184 | // -(C / X) --> (-C) / X | ||||
2185 | if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) { | ||||
2186 | Instruction *FDiv = | ||||
2187 | BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I); | ||||
2188 | |||||
2189 | // Intersect 'nsz' and 'ninf' because those special value exceptions may not | ||||
2190 | // apply to the fdiv. Everything else propagates from the fneg. | ||||
2191 | // TODO: We could propagate nsz/ninf from fdiv alone? | ||||
2192 | FastMathFlags FMF = I.getFastMathFlags(); | ||||
2193 | FastMathFlags OpFMF = FNegOp->getFastMathFlags(); | ||||
2194 | FDiv->setHasNoSignedZeros(FMF.noSignedZeros() & OpFMF.noSignedZeros()); | ||||
2195 | FDiv->setHasNoInfs(FMF.noInfs() & OpFMF.noInfs()); | ||||
2196 | return FDiv; | ||||
2197 | } | ||||
2198 | // With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]: | ||||
2199 | // -(X + C) --> -X + -C --> -C - X | ||||
2200 | if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C)))) | ||||
2201 | return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I); | ||||
2202 | |||||
2203 | return nullptr; | ||||
2204 | } | ||||
2205 | |||||
2206 | static Instruction *hoistFNegAboveFMulFDiv(Instruction &I, | ||||
2207 | InstCombiner::BuilderTy &Builder) { | ||||
2208 | Value *FNeg; | ||||
2209 | if (!match(&I, m_FNeg(m_Value(FNeg)))) | ||||
2210 | return nullptr; | ||||
2211 | |||||
2212 | Value *X, *Y; | ||||
2213 | if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y))))) | ||||
2214 | return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I); | ||||
2215 | |||||
2216 | if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y))))) | ||||
2217 | return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I); | ||||
2218 | |||||
2219 | return nullptr; | ||||
2220 | } | ||||
2221 | |||||
2222 | Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) { | ||||
2223 | Value *Op = I.getOperand(0); | ||||
2224 | |||||
2225 | if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(), | ||||
2226 | getSimplifyQuery().getWithInstruction(&I))) | ||||
2227 | return replaceInstUsesWith(I, V); | ||||
2228 | |||||
2229 | if (Instruction *X = foldFNegIntoConstant(I)) | ||||
2230 | return X; | ||||
2231 | |||||
2232 | Value *X, *Y; | ||||
2233 | |||||
2234 | // If we can ignore the sign of zeros: -(X - Y) --> (Y - X) | ||||
2235 | if (I.hasNoSignedZeros() && | ||||
2236 | match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) | ||||
2237 | return BinaryOperator::CreateFSubFMF(Y, X, &I); | ||||
2238 | |||||
2239 | if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder)) | ||||
2240 | return R; | ||||
2241 | |||||
2242 | // Try to eliminate fneg if at least 1 arm of the select is negated. | ||||
2243 | Value *Cond; | ||||
2244 | if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) { | ||||
2245 | // Unlike most transforms, this one is not safe to propagate nsz unless | ||||
2246 | // it is present on the original select. (We are conservatively intersecting | ||||
2247 | // the nsz flags from the select and root fneg instruction.) | ||||
2248 | auto propagateSelectFMF = [&](SelectInst *S) { | ||||
2249 | S->copyFastMathFlags(&I); | ||||
2250 | if (auto *OldSel = dyn_cast<SelectInst>(Op)) | ||||
2251 | if (!OldSel->hasNoSignedZeros()) | ||||
2252 | S->setHasNoSignedZeros(false); | ||||
2253 | }; | ||||
2254 | // -(Cond ? -P : Y) --> Cond ? P : -Y | ||||
2255 | Value *P; | ||||
2256 | if (match(X, m_FNeg(m_Value(P)))) { | ||||
2257 | Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg"); | ||||
2258 | SelectInst *NewSel = SelectInst::Create(Cond, P, NegY); | ||||
2259 | propagateSelectFMF(NewSel); | ||||
2260 | return NewSel; | ||||
2261 | } | ||||
2262 | // -(Cond ? X : -P) --> Cond ? -X : P | ||||
2263 | if (match(Y, m_FNeg(m_Value(P)))) { | ||||
2264 | Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg"); | ||||
2265 | SelectInst *NewSel = SelectInst::Create(Cond, NegX, P); | ||||
2266 | propagateSelectFMF(NewSel); | ||||
2267 | return NewSel; | ||||
2268 | } | ||||
2269 | } | ||||
2270 | |||||
2271 | return nullptr; | ||||
2272 | } | ||||
2273 | |||||
2274 | Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) { | ||||
2275 | if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1), | ||||
2276 | I.getFastMathFlags(), | ||||
2277 | getSimplifyQuery().getWithInstruction(&I))) | ||||
2278 | return replaceInstUsesWith(I, V); | ||||
2279 | |||||
2280 | if (Instruction *X = foldVectorBinop(I)) | ||||
2281 | return X; | ||||
2282 | |||||
2283 | // Subtraction from -0.0 is the canonical form of fneg. | ||||
2284 | // fsub -0.0, X ==> fneg X | ||||
2285 | // fsub nsz 0.0, X ==> fneg nsz X | ||||
2286 | // | ||||
2287 | // FIXME This matcher does not respect FTZ or DAZ yet: | ||||
2288 | // fsub -0.0, Denorm ==> +-0 | ||||
2289 | // fneg Denorm ==> -Denorm | ||||
2290 | Value *Op; | ||||
2291 | if (match(&I, m_FNeg(m_Value(Op)))) | ||||
2292 | return UnaryOperator::CreateFNegFMF(Op, &I); | ||||
2293 | |||||
2294 | if (Instruction *X = foldFNegIntoConstant(I)) | ||||
2295 | return X; | ||||
2296 | |||||
2297 | if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder)) | ||||
2298 | return R; | ||||
2299 | |||||
2300 | Value *X, *Y; | ||||
2301 | Constant *C; | ||||
2302 | |||||
2303 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
2304 | // If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X) | ||||
2305 | // Canonicalize to fadd to make analysis easier. | ||||
2306 | // This can also help codegen because fadd is commutative. | ||||
2307 | // Note that if this fsub was really an fneg, the fadd with -0.0 will get | ||||
2308 | // killed later. We still limit that particular transform with 'hasOneUse' | ||||
2309 | // because an fneg is assumed better/cheaper than a generic fsub. | ||||
2310 | if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) { | ||||
2311 | if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) { | ||||
2312 | Value *NewSub = Builder.CreateFSubFMF(Y, X, &I); | ||||
2313 | return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I); | ||||
2314 | } | ||||
2315 | } | ||||
2316 | |||||
2317 | // (-X) - Op1 --> -(X + Op1) | ||||
2318 | if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) && | ||||
2319 | match(Op0, m_OneUse(m_FNeg(m_Value(X))))) { | ||||
2320 | Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I); | ||||
2321 | return UnaryOperator::CreateFNegFMF(FAdd, &I); | ||||
2322 | } | ||||
2323 | |||||
2324 | if (isa<Constant>(Op0)) | ||||
2325 | if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) | ||||
2326 | if (Instruction *NV = FoldOpIntoSelect(I, SI)) | ||||
2327 | return NV; | ||||
2328 | |||||
2329 | // X - C --> X + (-C) | ||||
2330 | // But don't transform constant expressions because there's an inverse fold | ||||
2331 | // for X + (-Y) --> X - Y. | ||||
2332 | if (match(Op1, m_ImmConstant(C))) | ||||
2333 | return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I); | ||||
2334 | |||||
2335 | // X - (-Y) --> X + Y | ||||
2336 | if (match(Op1, m_FNeg(m_Value(Y)))) | ||||
2337 | return BinaryOperator::CreateFAddFMF(Op0, Y, &I); | ||||
2338 | |||||
2339 | // Similar to above, but look through a cast of the negated value: | ||||
2340 | // X - (fptrunc(-Y)) --> X + fptrunc(Y) | ||||
2341 | Type *Ty = I.getType(); | ||||
2342 | if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y)))))) | ||||
2343 | return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I); | ||||
2344 | |||||
2345 | // X - (fpext(-Y)) --> X + fpext(Y) | ||||
2346 | if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y)))))) | ||||
2347 | return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I); | ||||
2348 | |||||
2349 | // Similar to above, but look through fmul/fdiv of the negated value: | ||||
2350 | // Op0 - (-X * Y) --> Op0 + (X * Y) | ||||
2351 | // Op0 - (Y * -X) --> Op0 + (X * Y) | ||||
2352 | if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) { | ||||
2353 | Value *FMul = Builder.CreateFMulFMF(X, Y, &I); | ||||
2354 | return BinaryOperator::CreateFAddFMF(Op0, FMul, &I); | ||||
2355 | } | ||||
2356 | // Op0 - (-X / Y) --> Op0 + (X / Y) | ||||
2357 | // Op0 - (X / -Y) --> Op0 + (X / Y) | ||||
2358 | if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) || | ||||
2359 | match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) { | ||||
2360 | Value *FDiv = Builder.CreateFDivFMF(X, Y, &I); | ||||
2361 | return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I); | ||||
2362 | } | ||||
2363 | |||||
2364 | // Handle special cases for FSub with selects feeding the operation | ||||
2365 | if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1)) | ||||
2366 | return replaceInstUsesWith(I, V); | ||||
2367 | |||||
2368 | if (I.hasAllowReassoc() && I.hasNoSignedZeros()) { | ||||
2369 | // (Y - X) - Y --> -X | ||||
2370 | if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X)))) | ||||
2371 | return UnaryOperator::CreateFNegFMF(X, &I); | ||||
2372 | |||||
2373 | // Y - (X + Y) --> -X | ||||
2374 | // Y - (Y + X) --> -X | ||||
2375 | if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X)))) | ||||
2376 | return UnaryOperator::CreateFNegFMF(X, &I); | ||||
2377 | |||||
2378 | // (X * C) - X --> X * (C - 1.0) | ||||
2379 | if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) { | ||||
2380 | Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0)); | ||||
2381 | return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I); | ||||
2382 | } | ||||
2383 | // X - (X * C) --> X * (1.0 - C) | ||||
2384 | if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) { | ||||
2385 | Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C); | ||||
2386 | return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I); | ||||
2387 | } | ||||
2388 | |||||
2389 | // Reassociate fsub/fadd sequences to create more fadd instructions and | ||||
2390 | // reduce dependency chains: | ||||
2391 | // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1) | ||||
2392 | Value *Z; | ||||
2393 | if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))), | ||||
2394 | m_Value(Z))))) { | ||||
2395 | Value *XZ = Builder.CreateFAddFMF(X, Z, &I); | ||||
2396 | Value *YW = Builder.CreateFAddFMF(Y, Op1, &I); | ||||
2397 | return BinaryOperator::CreateFSubFMF(XZ, YW, &I); | ||||
2398 | } | ||||
2399 | |||||
2400 | auto m_FaddRdx = [](Value *&Sum, Value *&Vec) { | ||||
2401 | return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum), | ||||
2402 | m_Value(Vec))); | ||||
2403 | }; | ||||
2404 | Value *A0, *A1, *V0, *V1; | ||||
2405 | if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) && | ||||
2406 | V0->getType() == V1->getType()) { | ||||
2407 | // Difference of sums is sum of differences: | ||||
2408 | // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1 | ||||
2409 | Value *Sub = Builder.CreateFSubFMF(V0, V1, &I); | ||||
2410 | Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd, | ||||
2411 | {Sub->getType()}, {A0, Sub}, &I); | ||||
2412 | return BinaryOperator::CreateFSubFMF(Rdx, A1, &I); | ||||
2413 | } | ||||
2414 | |||||
2415 | if (Instruction *F = factorizeFAddFSub(I, Builder)) | ||||
2416 | return F; | ||||
2417 | |||||
2418 | // TODO: This performs reassociative folds for FP ops. Some fraction of the | ||||
2419 | // functionality has been subsumed by simple pattern matching here and in | ||||
2420 | // InstSimplify. We should let a dedicated reassociation pass handle more | ||||
2421 | // complex pattern matching and remove this from InstCombine. | ||||
2422 | if (Value *V = FAddCombine(Builder).simplify(&I)) | ||||
2423 | return replaceInstUsesWith(I, V); | ||||
2424 | |||||
2425 | // (X - Y) - Op1 --> X - (Y + Op1) | ||||
2426 | if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) { | ||||
2427 | Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I); | ||||
2428 | return BinaryOperator::CreateFSubFMF(X, FAdd, &I); | ||||
2429 | } | ||||
2430 | } | ||||
2431 | |||||
2432 | return nullptr; | ||||
2433 | } |
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 | |||||
46 | namespace llvm { | ||||
47 | namespace PatternMatch { | ||||
48 | |||||
49 | template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) { | ||||
50 | return const_cast<Pattern &>(P).match(V); | ||||
51 | } | ||||
52 | |||||
53 | template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) { | ||||
54 | return const_cast<Pattern &>(P).match(Mask); | ||||
55 | } | ||||
56 | |||||
57 | template <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 | |||||
67 | template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) { | ||||
68 | return SubPattern; | ||||
69 | } | ||||
70 | |||||
71 | template <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. | ||||
76 | inline class_match<Value> m_Value() { return class_match<Value>(); } | ||||
77 | |||||
78 | /// Match an arbitrary unary operation and ignore it. | ||||
79 | inline class_match<UnaryOperator> m_UnOp() { | ||||
80 | return class_match<UnaryOperator>(); | ||||
81 | } | ||||
82 | |||||
83 | /// Match an arbitrary binary operation and ignore it. | ||||
84 | inline class_match<BinaryOperator> m_BinOp() { | ||||
85 | return class_match<BinaryOperator>(); | ||||
86 | } | ||||
87 | |||||
88 | /// Matches any compare instruction and ignore it. | ||||
89 | inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); } | ||||
90 | |||||
91 | struct 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. | ||||
136 | inline auto m_Undef() { return undef_match(); } | ||||
137 | |||||
138 | /// Match an arbitrary poison constant. | ||||
139 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } | ||||
140 | |||||
141 | /// Match an arbitrary Constant and ignore it. | ||||
142 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } | ||||
143 | |||||
144 | /// Match an arbitrary ConstantInt and ignore it. | ||||
145 | inline class_match<ConstantInt> m_ConstantInt() { | ||||
146 | return class_match<ConstantInt>(); | ||||
147 | } | ||||
148 | |||||
149 | /// Match an arbitrary ConstantFP and ignore it. | ||||
150 | inline class_match<ConstantFP> m_ConstantFP() { | ||||
151 | return class_match<ConstantFP>(); | ||||
152 | } | ||||
153 | |||||
154 | /// Match an arbitrary ConstantExpr and ignore it. | ||||
155 | inline class_match<ConstantExpr> m_ConstantExpr() { | ||||
156 | return class_match<ConstantExpr>(); | ||||
157 | } | ||||
158 | |||||
159 | /// Match an arbitrary basic block value and ignore it. | ||||
160 | inline class_match<BasicBlock> m_BasicBlock() { | ||||
161 | return class_match<BasicBlock>(); | ||||
162 | } | ||||
163 | |||||
164 | /// Inverting matcher | ||||
165 | template <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. | ||||
174 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { | ||||
175 | return match_unless<Ty>(M); | ||||
176 | } | ||||
177 | |||||
178 | /// Matching combinators | ||||
179 | template <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)) | ||||
187 | return true; | ||||
188 | if (R.match(V)) | ||||
189 | return true; | ||||
190 | return false; | ||||
191 | } | ||||
192 | }; | ||||
193 | |||||
194 | template <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)) | ||||
202 | if (R.match(V)) | ||||
203 | return true; | ||||
204 | return false; | ||||
205 | } | ||||
206 | }; | ||||
207 | |||||
208 | /// Combine two pattern matchers matching L || R | ||||
209 | template <typename LTy, typename RTy> | ||||
210 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { | ||||
211 | return match_combine_or<LTy, RTy>(L, R); | ||||
212 | } | ||||
213 | |||||
214 | /// Combine two pattern matchers matching L && R | ||||
215 | template <typename LTy, typename RTy> | ||||
216 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { | ||||
217 | return match_combine_and<LTy, RTy>(L, R); | ||||
218 | } | ||||
219 | |||||
220 | struct 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. | ||||
245 | struct 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. | ||||
270 | inline 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. | ||||
276 | inline 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. | ||||
281 | inline 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. | ||||
287 | inline 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. | ||||
293 | inline 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. | ||||
298 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { | ||||
299 | return apfloat_match(Res, /* AllowUndef */ false); | ||||
300 | } | ||||
301 | |||||
302 | template <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. | ||||
318 | template <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. | ||||
325 | template <typename Predicate, typename ConstantVal> | ||||
326 | struct 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?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/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 | ||||
363 | template <typename Predicate> | ||||
364 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; | ||||
365 | |||||
366 | /// specialization of cstval_pred_ty for ConstantFP | ||||
367 | template <typename Predicate> | ||||
368 | using 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. | ||||
372 | template <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. | ||||
398 | template <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 | |||||
431 | struct 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. | ||||
436 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { | ||||
437 | return cst_pred_ty<is_any_apint>(); | ||||
438 | } | ||||
439 | |||||
440 | struct is_all_ones { | ||||
441 | bool isValue(const APInt &C) { return C.isAllOnesValue(); } | ||||
442 | }; | ||||
443 | /// Match an integer or vector with all bits set. | ||||
444 | /// For vectors, this includes constants with undefined elements. | ||||
445 | inline cst_pred_ty<is_all_ones> m_AllOnes() { | ||||
446 | return cst_pred_ty<is_all_ones>(); | ||||
447 | } | ||||
448 | |||||
449 | struct 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. | ||||
455 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { | ||||
456 | return cst_pred_ty<is_maxsignedvalue>(); | ||||
457 | } | ||||
458 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { | ||||
459 | return V; | ||||
460 | } | ||||
461 | |||||
462 | struct 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. | ||||
467 | inline cst_pred_ty<is_negative> m_Negative() { | ||||
468 | return cst_pred_ty<is_negative>(); | ||||
469 | } | ||||
470 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { | ||||
471 | return V; | ||||
472 | } | ||||
473 | |||||
474 | struct 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. | ||||
479 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { | ||||
480 | return cst_pred_ty<is_nonnegative>(); | ||||
481 | } | ||||
482 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { | ||||
483 | return V; | ||||
484 | } | ||||
485 | |||||
486 | struct 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. | ||||
491 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { | ||||
492 | return cst_pred_ty<is_strictlypositive>(); | ||||
493 | } | ||||
494 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { | ||||
495 | return V; | ||||
496 | } | ||||
497 | |||||
498 | struct 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. | ||||
503 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { | ||||
504 | return cst_pred_ty<is_nonpositive>(); | ||||
505 | } | ||||
506 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } | ||||
507 | |||||
508 | struct is_one { | ||||
509 | bool isValue(const APInt &C) { return C.isOneValue(); } | ||||
510 | }; | ||||
511 | /// Match an integer 1 or a vector with all elements equal to 1. | ||||
512 | /// For vectors, this includes constants with undefined elements. | ||||
513 | inline cst_pred_ty<is_one> m_One() { | ||||
514 | return cst_pred_ty<is_one>(); | ||||
515 | } | ||||
516 | |||||
517 | struct is_zero_int { | ||||
518 | bool isValue(const APInt &C) { return C.isNullValue(); } | ||||
519 | }; | ||||
520 | /// Match an integer 0 or a vector with all elements equal to 0. | ||||
521 | /// For vectors, this includes constants with undefined elements. | ||||
522 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { | ||||
523 | return cst_pred_ty<is_zero_int>(); | ||||
524 | } | ||||
525 | |||||
526 | struct 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. | ||||
535 | inline is_zero m_Zero() { | ||||
536 | return is_zero(); | ||||
537 | } | ||||
538 | |||||
539 | struct 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. | ||||
544 | inline cst_pred_ty<is_power2> m_Power2() { | ||||
545 | return cst_pred_ty<is_power2>(); | ||||
546 | } | ||||
547 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { | ||||
548 | return V; | ||||
549 | } | ||||
550 | |||||
551 | struct is_negated_power2 { | ||||
552 | bool isValue(const APInt &C) { return (-C).isPowerOf2(); } | ||||
553 | }; | ||||
554 | /// Match a integer or vector negated power-of-2. | ||||
555 | /// For vectors, this includes constants with undefined elements. | ||||
556 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { | ||||
557 | return cst_pred_ty<is_negated_power2>(); | ||||
558 | } | ||||
559 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { | ||||
560 | return V; | ||||
561 | } | ||||
562 | |||||
563 | struct 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. | ||||
568 | inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() { | ||||
569 | return cst_pred_ty<is_power2_or_zero>(); | ||||
570 | } | ||||
571 | inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) { | ||||
572 | return V; | ||||
573 | } | ||||
574 | |||||
575 | struct 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. | ||||
580 | inline cst_pred_ty<is_sign_mask> m_SignMask() { | ||||
581 | return cst_pred_ty<is_sign_mask>(); | ||||
582 | } | ||||
583 | |||||
584 | struct 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. | ||||
589 | inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() { | ||||
590 | return cst_pred_ty<is_lowbit_mask>(); | ||||
591 | } | ||||
592 | |||||
593 | struct icmp_pred_with_threshold { | ||||
594 | ICmpInst::Predicate Pred; | ||||
595 | const APInt *Thr; | ||||
596 | bool isValue(const APInt &C) { | ||||
597 | switch (Pred) { | ||||
598 | case ICmpInst::Predicate::ICMP_EQ: | ||||
599 | return C.eq(*Thr); | ||||
600 | case ICmpInst::Predicate::ICMP_NE: | ||||
601 | return C.ne(*Thr); | ||||
602 | case ICmpInst::Predicate::ICMP_UGT: | ||||
603 | return C.ugt(*Thr); | ||||
604 | case ICmpInst::Predicate::ICMP_UGE: | ||||
605 | return C.uge(*Thr); | ||||
606 | case ICmpInst::Predicate::ICMP_ULT: | ||||
607 | return C.ult(*Thr); | ||||
608 | case ICmpInst::Predicate::ICMP_ULE: | ||||
609 | return C.ule(*Thr); | ||||
610 | case ICmpInst::Predicate::ICMP_SGT: | ||||
611 | return C.sgt(*Thr); | ||||
612 | case ICmpInst::Predicate::ICMP_SGE: | ||||
613 | return C.sge(*Thr); | ||||
614 | case ICmpInst::Predicate::ICMP_SLT: | ||||
615 | return C.slt(*Thr); | ||||
616 | case ICmpInst::Predicate::ICMP_SLE: | ||||
617 | return C.sle(*Thr); | ||||
618 | default: | ||||
619 | llvm_unreachable("Unhandled ICmp predicate")::llvm::llvm_unreachable_internal("Unhandled ICmp predicate", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h" , 619); | ||||
620 | } | ||||
621 | } | ||||
622 | }; | ||||
623 | /// Match an integer or vector with every element comparing 'pred' (eg/ne/...) | ||||
624 | /// to Threshold. For vectors, this includes constants with undefined elements. | ||||
625 | inline cst_pred_ty<icmp_pred_with_threshold> | ||||
626 | m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) { | ||||
627 | cst_pred_ty<icmp_pred_with_threshold> P; | ||||
628 | P.Pred = Predicate; | ||||
629 | P.Thr = &Threshold; | ||||
630 | return P; | ||||
631 | } | ||||
632 | |||||
633 | struct is_nan { | ||||
634 | bool isValue(const APFloat &C) { return C.isNaN(); } | ||||
635 | }; | ||||
636 | /// Match an arbitrary NaN constant. This includes quiet and signalling nans. | ||||
637 | /// For vectors, this includes constants with undefined elements. | ||||
638 | inline cstfp_pred_ty<is_nan> m_NaN() { | ||||
639 | return cstfp_pred_ty<is_nan>(); | ||||
640 | } | ||||
641 | |||||
642 | struct is_nonnan { | ||||
643 | bool isValue(const APFloat &C) { return !C.isNaN(); } | ||||
644 | }; | ||||
645 | /// Match a non-NaN FP constant. | ||||
646 | /// For vectors, this includes constants with undefined elements. | ||||
647 | inline cstfp_pred_ty<is_nonnan> m_NonNaN() { | ||||
648 | return cstfp_pred_ty<is_nonnan>(); | ||||
649 | } | ||||
650 | |||||
651 | struct is_inf { | ||||
652 | bool isValue(const APFloat &C) { return C.isInfinity(); } | ||||
653 | }; | ||||
654 | /// Match a positive or negative infinity FP constant. | ||||
655 | /// For vectors, this includes constants with undefined elements. | ||||
656 | inline cstfp_pred_ty<is_inf> m_Inf() { | ||||
657 | return cstfp_pred_ty<is_inf>(); | ||||
658 | } | ||||
659 | |||||
660 | struct is_noninf { | ||||
661 | bool isValue(const APFloat &C) { return !C.isInfinity(); } | ||||
662 | }; | ||||
663 | /// Match a non-infinity FP constant, i.e. finite or NaN. | ||||
664 | /// For vectors, this includes constants with undefined elements. | ||||
665 | inline cstfp_pred_ty<is_noninf> m_NonInf() { | ||||
666 | return cstfp_pred_ty<is_noninf>(); | ||||
667 | } | ||||
668 | |||||
669 | struct is_finite { | ||||
670 | bool isValue(const APFloat &C) { return C.isFinite(); } | ||||
671 | }; | ||||
672 | /// Match a finite FP constant, i.e. not infinity or NaN. | ||||
673 | /// For vectors, this includes constants with undefined elements. | ||||
674 | inline cstfp_pred_ty<is_finite> m_Finite() { | ||||
675 | return cstfp_pred_ty<is_finite>(); | ||||
676 | } | ||||
677 | inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; } | ||||
678 | |||||
679 | struct is_finitenonzero { | ||||
680 | bool isValue(const APFloat &C) { return C.isFiniteNonZero(); } | ||||
681 | }; | ||||
682 | /// Match a finite non-zero FP constant. | ||||
683 | /// For vectors, this includes constants with undefined elements. | ||||
684 | inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() { | ||||
685 | return cstfp_pred_ty<is_finitenonzero>(); | ||||
686 | } | ||||
687 | inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) { | ||||
688 | return V; | ||||
689 | } | ||||
690 | |||||
691 | struct is_any_zero_fp { | ||||
692 | bool isValue(const APFloat &C) { return C.isZero(); } | ||||
693 | }; | ||||
694 | /// Match a floating-point negative zero or positive zero. | ||||
695 | /// For vectors, this includes constants with undefined elements. | ||||
696 | inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() { | ||||
697 | return cstfp_pred_ty<is_any_zero_fp>(); | ||||
698 | } | ||||
699 | |||||
700 | struct is_pos_zero_fp { | ||||
701 | bool isValue(const APFloat &C) { return C.isPosZero(); } | ||||
702 | }; | ||||
703 | /// Match a floating-point positive zero. | ||||
704 | /// For vectors, this includes constants with undefined elements. | ||||
705 | inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() { | ||||
706 | return cstfp_pred_ty<is_pos_zero_fp>(); | ||||
707 | } | ||||
708 | |||||
709 | struct is_neg_zero_fp { | ||||
710 | bool isValue(const APFloat &C) { return C.isNegZero(); } | ||||
711 | }; | ||||
712 | /// Match a floating-point negative zero. | ||||
713 | /// For vectors, this includes constants with undefined elements. | ||||
714 | inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() { | ||||
715 | return cstfp_pred_ty<is_neg_zero_fp>(); | ||||
716 | } | ||||
717 | |||||
718 | struct is_non_zero_fp { | ||||
719 | bool isValue(const APFloat &C) { return C.isNonZero(); } | ||||
720 | }; | ||||
721 | /// Match a floating-point non-zero. | ||||
722 | /// For vectors, this includes constants with undefined elements. | ||||
723 | inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() { | ||||
724 | return cstfp_pred_ty<is_non_zero_fp>(); | ||||
725 | } | ||||
726 | |||||
727 | /////////////////////////////////////////////////////////////////////////////// | ||||
728 | |||||
729 | template <typename Class> struct bind_ty { | ||||
730 | Class *&VR; | ||||
731 | |||||
732 | bind_ty(Class *&V) : VR(V) {} | ||||
733 | |||||
734 | template <typename ITy> bool match(ITy *V) { | ||||
735 | if (auto *CV = dyn_cast<Class>(V)) { | ||||
736 | VR = CV; | ||||
737 | return true; | ||||
738 | } | ||||
739 | return false; | ||||
740 | } | ||||
741 | }; | ||||
742 | |||||
743 | /// Match a value, capturing it if we match. | ||||
744 | inline bind_ty<Value> m_Value(Value *&V) { return V; } | ||||
745 | inline bind_ty<const Value> m_Value(const Value *&V) { return V; } | ||||
746 | |||||
747 | /// Match an instruction, capturing it if we match. | ||||
748 | inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; } | ||||
749 | /// Match a unary operator, capturing it if we match. | ||||
750 | inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; } | ||||
751 | /// Match a binary operator, capturing it if we match. | ||||
752 | inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; } | ||||
753 | /// Match a with overflow intrinsic, capturing it if we match. | ||||
754 | inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; } | ||||
755 | inline bind_ty<const WithOverflowInst> | ||||
756 | m_WithOverflowInst(const WithOverflowInst *&I) { | ||||
757 | return I; | ||||
758 | } | ||||
759 | |||||
760 | /// Match a Constant, capturing the value if we match. | ||||
761 | inline bind_ty<Constant> m_Constant(Constant *&C) { return C; } | ||||
762 | |||||
763 | /// Match a ConstantInt, capturing the value if we match. | ||||
764 | inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; } | ||||
765 | |||||
766 | /// Match a ConstantFP, capturing the value if we match. | ||||
767 | inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; } | ||||
768 | |||||
769 | /// Match a ConstantExpr, capturing the value if we match. | ||||
770 | inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; } | ||||
771 | |||||
772 | /// Match a basic block value, capturing it if we match. | ||||
773 | inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; } | ||||
774 | inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) { | ||||
775 | return V; | ||||
776 | } | ||||
777 | |||||
778 | /// Match an arbitrary immediate Constant and ignore it. | ||||
779 | inline match_combine_and<class_match<Constant>, | ||||
780 | match_unless<class_match<ConstantExpr>>> | ||||
781 | m_ImmConstant() { | ||||
782 | return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr())); | ||||
783 | } | ||||
784 | |||||
785 | /// Match an immediate Constant, capturing the value if we match. | ||||
786 | inline match_combine_and<bind_ty<Constant>, | ||||
787 | match_unless<class_match<ConstantExpr>>> | ||||
788 | m_ImmConstant(Constant *&C) { | ||||
789 | return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr())); | ||||
790 | } | ||||
791 | |||||
792 | /// Match a specified Value*. | ||||
793 | struct specificval_ty { | ||||
794 | const Value *Val; | ||||
795 | |||||
796 | specificval_ty(const Value *V) : Val(V) {} | ||||
797 | |||||
798 | template <typename ITy> bool match(ITy *V) { return V == Val; } | ||||
799 | }; | ||||
800 | |||||
801 | /// Match if we have a specific specified value. | ||||
802 | inline specificval_ty m_Specific(const Value *V) { return V; } | ||||
803 | |||||
804 | /// Stores a reference to the Value *, not the Value * itself, | ||||
805 | /// thus can be used in commutative matchers. | ||||
806 | template <typename Class> struct deferredval_ty { | ||||
807 | Class *const &Val; | ||||
808 | |||||
809 | deferredval_ty(Class *const &V) : Val(V) {} | ||||
810 | |||||
811 | template <typename ITy> bool match(ITy *const V) { return V == Val; } | ||||
812 | }; | ||||
813 | |||||
814 | /// Like m_Specific(), but works if the specific value to match is determined | ||||
815 | /// as part of the same match() expression. For example: | ||||
816 | /// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will | ||||
817 | /// bind X before the pattern match starts. | ||||
818 | /// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against | ||||
819 | /// whichever value m_Value(X) populated. | ||||
820 | inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; } | ||||
821 | inline deferredval_ty<const Value> m_Deferred(const Value *const &V) { | ||||
822 | return V; | ||||
823 | } | ||||
824 | |||||
825 | /// Match a specified floating point value or vector of all elements of | ||||
826 | /// that value. | ||||
827 | struct specific_fpval { | ||||
828 | double Val; | ||||
829 | |||||
830 | specific_fpval(double V) : Val(V) {} | ||||
831 | |||||
832 | template <typename ITy> bool match(ITy *V) { | ||||
833 | if (const auto *CFP = dyn_cast<ConstantFP>(V)) | ||||
834 | return CFP->isExactlyValue(Val); | ||||
835 | if (V->getType()->isVectorTy()) | ||||
836 | if (const auto *C = dyn_cast<Constant>(V)) | ||||
837 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) | ||||
838 | return CFP->isExactlyValue(Val); | ||||
839 | return false; | ||||
840 | } | ||||
841 | }; | ||||
842 | |||||
843 | /// Match a specific floating point value or vector with all elements | ||||
844 | /// equal to the value. | ||||
845 | inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); } | ||||
846 | |||||
847 | /// Match a float 1.0 or vector with all elements equal to 1.0. | ||||
848 | inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); } | ||||
849 | |||||
850 | struct bind_const_intval_ty { | ||||
851 | uint64_t &VR; | ||||
852 | |||||
853 | bind_const_intval_ty(uint64_t &V) : VR(V) {} | ||||
854 | |||||
855 | template <typename ITy> bool match(ITy *V) { | ||||
856 | if (const auto *CV = dyn_cast<ConstantInt>(V)) | ||||
857 | if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) { | ||||
858 | VR = CV->getZExtValue(); | ||||
859 | return true; | ||||
860 | } | ||||
861 | return false; | ||||
862 | } | ||||
863 | }; | ||||
864 | |||||
865 | /// Match a specified integer value or vector of all elements of that | ||||
866 | /// value. | ||||
867 | template <bool AllowUndefs> | ||||
868 | struct specific_intval { | ||||
869 | APInt Val; | ||||
870 | |||||
871 | specific_intval(APInt V) : Val(std::move(V)) {} | ||||
872 | |||||
873 | template <typename ITy> bool match(ITy *V) { | ||||
874 | const auto *CI = dyn_cast<ConstantInt>(V); | ||||
875 | if (!CI && V->getType()->isVectorTy()) | ||||
876 | if (const auto *C = dyn_cast<Constant>(V)) | ||||
877 | CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs)); | ||||
878 | |||||
879 | return CI && APInt::isSameValue(CI->getValue(), Val); | ||||
880 | } | ||||
881 | }; | ||||
882 | |||||
883 | /// Match a specific integer value or vector with all elements equal to | ||||
884 | /// the value. | ||||
885 | inline specific_intval<false> m_SpecificInt(APInt V) { | ||||
886 | return specific_intval<false>(std::move(V)); | ||||
887 | } | ||||
888 | |||||
889 | inline specific_intval<false> m_SpecificInt(uint64_t V) { | ||||
890 | return m_SpecificInt(APInt(64, V)); | ||||
891 | } | ||||
892 | |||||
893 | inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) { | ||||
894 | return specific_intval<true>(std::move(V)); | ||||
895 | } | ||||
896 | |||||
897 | inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) { | ||||
898 | return m_SpecificIntAllowUndef(APInt(64, V)); | ||||
899 | } | ||||
900 | |||||
901 | /// Match a ConstantInt and bind to its value. This does not match | ||||
902 | /// ConstantInts wider than 64-bits. | ||||
903 | inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; } | ||||
904 | |||||
905 | /// Match a specified basic block value. | ||||
906 | struct specific_bbval { | ||||
907 | BasicBlock *Val; | ||||
908 | |||||
909 | specific_bbval(BasicBlock *Val) : Val(Val) {} | ||||
910 | |||||
911 | template <typename ITy> bool match(ITy *V) { | ||||
912 | const auto *BB = dyn_cast<BasicBlock>(V); | ||||
913 | return BB && BB == Val; | ||||
914 | } | ||||
915 | }; | ||||
916 | |||||
917 | /// Match a specific basic block value. | ||||
918 | inline specific_bbval m_SpecificBB(BasicBlock *BB) { | ||||
919 | return specific_bbval(BB); | ||||
920 | } | ||||
921 | |||||
922 | /// A commutative-friendly version of m_Specific(). | ||||
923 | inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) { | ||||
924 | return BB; | ||||
925 | } | ||||
926 | inline deferredval_ty<const BasicBlock> | ||||
927 | m_Deferred(const BasicBlock *const &BB) { | ||||
928 | return BB; | ||||
929 | } | ||||
930 | |||||
931 | //===----------------------------------------------------------------------===// | ||||
932 | // Matcher for any binary operator. | ||||
933 | // | ||||
934 | template <typename LHS_t, typename RHS_t, bool Commutable = false> | ||||
935 | struct AnyBinaryOp_match { | ||||
936 | LHS_t L; | ||||
937 | RHS_t R; | ||||
938 | |||||
939 | // The evaluation order is always stable, regardless of Commutability. | ||||
940 | // The LHS is always matched first. | ||||
941 | AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||
942 | |||||
943 | template <typename OpTy> bool match(OpTy *V) { | ||||
944 | if (auto *I
| ||||
945 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || | ||||
946 | (Commutable && L.match(I->getOperand(1)) && | ||||
947 | R.match(I->getOperand(0))); | ||||
948 | return false; | ||||
949 | } | ||||
950 | }; | ||||
951 | |||||
952 | template <typename LHS, typename RHS> | ||||
953 | inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) { | ||||
954 | return AnyBinaryOp_match<LHS, RHS>(L, R); | ||||
955 | } | ||||
956 | |||||
957 | //===----------------------------------------------------------------------===// | ||||
958 | // Matcher for any unary operator. | ||||
959 | // TODO fuse unary, binary matcher into n-ary matcher | ||||
960 | // | ||||
961 | template <typename OP_t> struct AnyUnaryOp_match { | ||||
962 | OP_t X; | ||||
963 | |||||
964 | AnyUnaryOp_match(const OP_t &X) : X(X) {} | ||||
965 | |||||
966 | template <typename OpTy> bool match(OpTy *V) { | ||||
967 | if (auto *I = dyn_cast<UnaryOperator>(V)) | ||||
968 | return X.match(I->getOperand(0)); | ||||
969 | return false; | ||||
970 | } | ||||
971 | }; | ||||
972 | |||||
973 | template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) { | ||||
974 | return AnyUnaryOp_match<OP_t>(X); | ||||
975 | } | ||||
976 | |||||
977 | //===----------------------------------------------------------------------===// | ||||
978 | // Matchers for specific binary operators. | ||||
979 | // | ||||
980 | |||||
981 | template <typename LHS_t, typename RHS_t, unsigned Opcode, | ||||
982 | bool Commutable = false> | ||||
983 | struct BinaryOp_match { | ||||
984 | LHS_t L; | ||||
985 | RHS_t R; | ||||
986 | |||||
987 | // The evaluation order is always stable, regardless of Commutability. | ||||
988 | // The LHS is always matched first. | ||||
989 | BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||
990 | |||||
991 | template <typename OpTy> bool match(OpTy *V) { | ||||
992 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||
993 | auto *I = cast<BinaryOperator>(V); | ||||
994 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || | ||||
995 | (Commutable && L.match(I->getOperand(1)) && | ||||
996 | R.match(I->getOperand(0))); | ||||
997 | } | ||||
998 | if (auto *CE = dyn_cast<ConstantExpr>(V)) | ||||
999 | return CE->getOpcode() == Opcode && | ||||
1000 | ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) || | ||||
1001 | (Commutable && L.match(CE->getOperand(1)) && | ||||
1002 | R.match(CE->getOperand(0)))); | ||||
1003 | return false; | ||||
1004 | } | ||||
1005 | }; | ||||
1006 | |||||
1007 | template <typename LHS, typename RHS> | ||||
1008 | inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L, | ||||
1009 | const RHS &R) { | ||||
1010 | return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R); | ||||
1011 | } | ||||
1012 | |||||
1013 | template <typename LHS, typename RHS> | ||||
1014 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L, | ||||
1015 | const RHS &R) { | ||||
1016 | return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R); | ||||
1017 | } | ||||
1018 | |||||
1019 | template <typename LHS, typename RHS> | ||||
1020 | inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L, | ||||
1021 | const RHS &R) { | ||||
1022 | return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R); | ||||
1023 | } | ||||
1024 | |||||
1025 | template <typename LHS, typename RHS> | ||||
1026 | inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L, | ||||
1027 | const RHS &R) { | ||||
1028 | return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R); | ||||
1029 | } | ||||
1030 | |||||
1031 | template <typename Op_t> struct FNeg_match { | ||||
1032 | Op_t X; | ||||
1033 | |||||
1034 | FNeg_match(const Op_t &Op) : X(Op) {} | ||||
1035 | template <typename OpTy> bool match(OpTy *V) { | ||||
1036 | auto *FPMO = dyn_cast<FPMathOperator>(V); | ||||
1037 | if (!FPMO) return false; | ||||
1038 | |||||
1039 | if (FPMO->getOpcode() == Instruction::FNeg) | ||||
1040 | return X.match(FPMO->getOperand(0)); | ||||
1041 | |||||
1042 | if (FPMO->getOpcode() == Instruction::FSub) { | ||||
1043 | if (FPMO->hasNoSignedZeros()) { | ||||
1044 | // With 'nsz', any zero goes. | ||||
1045 | if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0))) | ||||
1046 | return false; | ||||
1047 | } else { | ||||
1048 | // Without 'nsz', we need fsub -0.0, X exactly. | ||||
1049 | if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0))) | ||||
1050 | return false; | ||||
1051 | } | ||||
1052 | |||||
1053 | return X.match(FPMO->getOperand(1)); | ||||
1054 | } | ||||
1055 | |||||
1056 | return false; | ||||
1057 | } | ||||
1058 | }; | ||||
1059 | |||||
1060 | /// Match 'fneg X' as 'fsub -0.0, X'. | ||||
1061 | template <typename OpTy> | ||||
1062 | inline FNeg_match<OpTy> | ||||
1063 | m_FNeg(const OpTy &X) { | ||||
1064 | return FNeg_match<OpTy>(X); | ||||
1065 | } | ||||
1066 | |||||
1067 | /// Match 'fneg X' as 'fsub +-0.0, X'. | ||||
1068 | template <typename RHS> | ||||
1069 | inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub> | ||||
1070 | m_FNegNSZ(const RHS &X) { | ||||
1071 | return m_FSub(m_AnyZeroFP(), X); | ||||
1072 | } | ||||
1073 | |||||
1074 | template <typename LHS, typename RHS> | ||||
1075 | inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L, | ||||
1076 | const RHS &R) { | ||||
1077 | return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R); | ||||
1078 | } | ||||
1079 | |||||
1080 | template <typename LHS, typename RHS> | ||||
1081 | inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L, | ||||
1082 | const RHS &R) { | ||||
1083 | return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R); | ||||
1084 | } | ||||
1085 | |||||
1086 | template <typename LHS, typename RHS> | ||||
1087 | inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L, | ||||
1088 | const RHS &R) { | ||||
1089 | return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R); | ||||
1090 | } | ||||
1091 | |||||
1092 | template <typename LHS, typename RHS> | ||||
1093 | inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L, | ||||
1094 | const RHS &R) { | ||||
1095 | return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R); | ||||
1096 | } | ||||
1097 | |||||
1098 | template <typename LHS, typename RHS> | ||||
1099 | inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L, | ||||
1100 | const RHS &R) { | ||||
1101 | return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R); | ||||
1102 | } | ||||
1103 | |||||
1104 | template <typename LHS, typename RHS> | ||||
1105 | inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L, | ||||
1106 | const RHS &R) { | ||||
1107 | return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R); | ||||
1108 | } | ||||
1109 | |||||
1110 | template <typename LHS, typename RHS> | ||||
1111 | inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L, | ||||
1112 | const RHS &R) { | ||||
1113 | return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R); | ||||
1114 | } | ||||
1115 | |||||
1116 | template <typename LHS, typename RHS> | ||||
1117 | inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L, | ||||
1118 | const RHS &R) { | ||||
1119 | return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R); | ||||
1120 | } | ||||
1121 | |||||
1122 | template <typename LHS, typename RHS> | ||||
1123 | inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L, | ||||
1124 | const RHS &R) { | ||||
1125 | return BinaryOp_match<LHS, RHS, Instruction::And>(L, R); | ||||
1126 | } | ||||
1127 | |||||
1128 | template <typename LHS, typename RHS> | ||||
1129 | inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L, | ||||
1130 | const RHS &R) { | ||||
1131 | return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R); | ||||
1132 | } | ||||
1133 | |||||
1134 | template <typename LHS, typename RHS> | ||||
1135 | inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L, | ||||
1136 | const RHS &R) { | ||||
1137 | return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R); | ||||
1138 | } | ||||
1139 | |||||
1140 | template <typename LHS, typename RHS> | ||||
1141 | inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L, | ||||
1142 | const RHS &R) { | ||||
1143 | return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R); | ||||
1144 | } | ||||
1145 | |||||
1146 | template <typename LHS, typename RHS> | ||||
1147 | inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L, | ||||
1148 | const RHS &R) { | ||||
1149 | return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R); | ||||
1150 | } | ||||
1151 | |||||
1152 | template <typename LHS, typename RHS> | ||||
1153 | inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L, | ||||
1154 | const RHS &R) { | ||||
1155 | return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R); | ||||
1156 | } | ||||
1157 | |||||
1158 | template <typename LHS_t, typename RHS_t, unsigned Opcode, | ||||
1159 | unsigned WrapFlags = 0> | ||||
1160 | struct OverflowingBinaryOp_match { | ||||
1161 | LHS_t L; | ||||
1162 | RHS_t R; | ||||
1163 | |||||
1164 | OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) | ||||
1165 | : L(LHS), R(RHS) {} | ||||
1166 | |||||
1167 | template <typename OpTy> bool match(OpTy *V) { | ||||
1168 | if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) { | ||||
1169 | if (Op->getOpcode() != Opcode) | ||||
1170 | return false; | ||||
1171 | if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) && | ||||
1172 | !Op->hasNoUnsignedWrap()) | ||||
1173 | return false; | ||||
1174 | if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) && | ||||
1175 | !Op->hasNoSignedWrap()) | ||||
1176 | return false; | ||||
1177 | return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1)); | ||||
1178 | } | ||||
1179 | return false; | ||||
1180 | } | ||||
1181 | }; | ||||
1182 | |||||
1183 | template <typename LHS, typename RHS> | ||||
1184 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||
1185 | OverflowingBinaryOperator::NoSignedWrap> | ||||
1186 | m_NSWAdd(const LHS &L, const RHS &R) { | ||||
1187 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||
1188 | OverflowingBinaryOperator::NoSignedWrap>( | ||||
1189 | L, R); | ||||
1190 | } | ||||
1191 | template <typename LHS, typename RHS> | ||||
1192 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||
1193 | OverflowingBinaryOperator::NoSignedWrap> | ||||
1194 | m_NSWSub(const LHS &L, const RHS &R) { | ||||
1195 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||
1196 | OverflowingBinaryOperator::NoSignedWrap>( | ||||
1197 | L, R); | ||||
1198 | } | ||||
1199 | template <typename LHS, typename RHS> | ||||
1200 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||
1201 | OverflowingBinaryOperator::NoSignedWrap> | ||||
1202 | m_NSWMul(const LHS &L, const RHS &R) { | ||||
1203 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||
1204 | OverflowingBinaryOperator::NoSignedWrap>( | ||||
1205 | L, R); | ||||
1206 | } | ||||
1207 | template <typename LHS, typename RHS> | ||||
1208 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||
1209 | OverflowingBinaryOperator::NoSignedWrap> | ||||
1210 | m_NSWShl(const LHS &L, const RHS &R) { | ||||
1211 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||
1212 | OverflowingBinaryOperator::NoSignedWrap>( | ||||
1213 | L, R); | ||||
1214 | } | ||||
1215 | |||||
1216 | template <typename LHS, typename RHS> | ||||
1217 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||
1218 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||
1219 | m_NUWAdd(const LHS &L, const RHS &R) { | ||||
1220 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||
1221 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||
1222 | L, R); | ||||
1223 | } | ||||
1224 | template <typename LHS, typename RHS> | ||||
1225 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||
1226 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||
1227 | m_NUWSub(const LHS &L, const RHS &R) { | ||||
1228 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||
1229 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||
1230 | L, R); | ||||
1231 | } | ||||
1232 | template <typename LHS, typename RHS> | ||||
1233 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||
1234 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||
1235 | m_NUWMul(const LHS &L, const RHS &R) { | ||||
1236 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||
1237 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||
1238 | L, R); | ||||
1239 | } | ||||
1240 | template <typename LHS, typename RHS> | ||||
1241 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||
1242 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||
1243 | m_NUWShl(const LHS &L, const RHS &R) { | ||||
1244 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||
1245 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||
1246 | L, R); | ||||
1247 | } | ||||
1248 | |||||
1249 | //===----------------------------------------------------------------------===// | ||||
1250 | // Class that matches a group of binary opcodes. | ||||
1251 | // | ||||
1252 | template <typename LHS_t, typename RHS_t, typename Predicate> | ||||
1253 | struct 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 | |||||
1270 | struct is_shift_op { | ||||
1271 | bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); } | ||||
1272 | }; | ||||
1273 | |||||
1274 | struct is_right_shift_op { | ||||
1275 | bool isOpType(unsigned Opcode) { | ||||
1276 | return Opcode == Instruction::LShr || Opcode == Instruction::AShr; | ||||
1277 | } | ||||
1278 | }; | ||||
1279 | |||||
1280 | struct is_logical_shift_op { | ||||
1281 | bool isOpType(unsigned Opcode) { | ||||
1282 | return Opcode == Instruction::LShr || Opcode == Instruction::Shl; | ||||
1283 | } | ||||
1284 | }; | ||||
1285 | |||||
1286 | struct is_bitwiselogic_op { | ||||
1287 | bool isOpType(unsigned Opcode) { | ||||
1288 | return Instruction::isBitwiseLogicOp(Opcode); | ||||
1289 | } | ||||
1290 | }; | ||||
1291 | |||||
1292 | struct is_idiv_op { | ||||
1293 | bool isOpType(unsigned Opcode) { | ||||
1294 | return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; | ||||
1295 | } | ||||
1296 | }; | ||||
1297 | |||||
1298 | struct is_irem_op { | ||||
1299 | bool isOpType(unsigned Opcode) { | ||||
1300 | return Opcode == Instruction::SRem || Opcode == Instruction::URem; | ||||
1301 | } | ||||
1302 | }; | ||||
1303 | |||||
1304 | /// Matches shift operations. | ||||
1305 | template <typename LHS, typename RHS> | ||||
1306 | inline 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. | ||||
1312 | template <typename LHS, typename RHS> | ||||
1313 | inline 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. | ||||
1319 | template <typename LHS, typename RHS> | ||||
1320 | inline BinOpPred_match<LHS, RHS, is_logical_shift_op> | ||||
1321 | m_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. | ||||
1326 | template <typename LHS, typename RHS> | ||||
1327 | inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op> | ||||
1328 | m_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. | ||||
1333 | template <typename LHS, typename RHS> | ||||
1334 | inline 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. | ||||
1340 | template <typename LHS, typename RHS> | ||||
1341 | inline 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 | // | ||||
1349 | template <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 | |||||
1361 | template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) { | ||||
1362 | return SubPattern; | ||||
1363 | } | ||||
1364 | |||||
1365 | //===----------------------------------------------------------------------===// | ||||
1366 | // Matchers for CmpInst classes | ||||
1367 | // | ||||
1368 | |||||
1369 | template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy, | ||||
1370 | bool Commutable = false> | ||||
1371 | struct 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 | |||||
1396 | template <typename LHS, typename RHS> | ||||
1397 | inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate> | ||||
1398 | m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||
1399 | return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R); | ||||
1400 | } | ||||
1401 | |||||
1402 | template <typename LHS, typename RHS> | ||||
1403 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate> | ||||
1404 | m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||
1405 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R); | ||||
1406 | } | ||||
1407 | |||||
1408 | template <typename LHS, typename RHS> | ||||
1409 | inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate> | ||||
1410 | m_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. | ||||
1419 | template <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. | ||||
1434 | template <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. | ||||
1450 | template <typename T0, typename T1, typename T2, unsigned Opcode> | ||||
1451 | struct 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) { | ||||
| |||||
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. | ||||
1470 | template <typename Cond, typename LHS, typename RHS> | ||||
1471 | inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select> | ||||
1472 | m_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)) | ||||
1478 | template <int64_t L, int64_t R, typename Cond> | ||||
1479 | inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>, | ||||
1480 | Instruction::Select> | ||||
1481 | m_SelectCst(const Cond &C) { | ||||
1482 | return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>()); | ||||
1483 | } | ||||
1484 | |||||
1485 | /// Matches FreezeInst. | ||||
1486 | template <typename OpTy> | ||||
1487 | inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) { | ||||
1488 | return OneOps_match<OpTy, Instruction::Freeze>(Op); | ||||
1489 | } | ||||
1490 | |||||
1491 | /// Matches InsertElementInst. | ||||
1492 | template <typename Val_t, typename Elt_t, typename Idx_t> | ||||
1493 | inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement> | ||||
1494 | m_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. | ||||
1500 | template <typename Val_t, typename Idx_t> | ||||
1501 | inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement> | ||||
1502 | m_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. | ||||
1507 | template <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 | |||||
1524 | struct 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 | |||||
1533 | struct m_ZeroMask { | ||||
1534 | bool match(ArrayRef<int> Mask) { | ||||
1535 | return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; }); | ||||
1536 | } | ||||
1537 | }; | ||||
1538 | |||||
1539 | struct 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 | |||||
1545 | struct 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. | ||||
1559 | template <typename V1_t, typename V2_t> | ||||
1560 | inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector> | ||||
1561 | m_Shuffle(const V1_t &v1, const V2_t &v2) { | ||||
1562 | return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2); | ||||
1563 | } | ||||
1564 | |||||
1565 | template <typename V1_t, typename V2_t, typename Mask_t> | ||||
1566 | inline Shuffle_match<V1_t, V2_t, Mask_t> | ||||
1567 | m_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. | ||||
1572 | template <typename OpTy> | ||||
1573 | inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) { | ||||
1574 | return OneOps_match<OpTy, Instruction::Load>(Op); | ||||
1575 | } | ||||
1576 | |||||
1577 | /// Matches StoreInst. | ||||
1578 | template <typename ValueOpTy, typename PointerOpTy> | ||||
1579 | inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store> | ||||
1580 | m_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 | |||||
1589 | template <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. | ||||
1602 | template <typename OpTy> | ||||
1603 | inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) { | ||||
1604 | return CastClass_match<OpTy, Instruction::BitCast>(Op); | ||||
1605 | } | ||||
1606 | |||||
1607 | /// Matches PtrToInt. | ||||
1608 | template <typename OpTy> | ||||
1609 | inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) { | ||||
1610 | return CastClass_match<OpTy, Instruction::PtrToInt>(Op); | ||||
1611 | } | ||||
1612 | |||||
1613 | /// Matches IntToPtr. | ||||
1614 | template <typename OpTy> | ||||
1615 | inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) { | ||||
1616 | return CastClass_match<OpTy, Instruction::IntToPtr>(Op); | ||||
1617 | } | ||||
1618 | |||||
1619 | /// Matches Trunc. | ||||
1620 | template <typename OpTy> | ||||
1621 | inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) { | ||||
1622 | return CastClass_match<OpTy, Instruction::Trunc>(Op); | ||||
1623 | } | ||||
1624 | |||||
1625 | template <typename OpTy> | ||||
1626 | inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy> | ||||
1627 | m_TruncOrSelf(const OpTy &Op) { | ||||
1628 | return m_CombineOr(m_Trunc(Op), Op); | ||||
1629 | } | ||||
1630 | |||||
1631 | /// Matches SExt. | ||||
1632 | template <typename OpTy> | ||||
1633 | inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) { | ||||
1634 | return CastClass_match<OpTy, Instruction::SExt>(Op); | ||||
1635 | } | ||||
1636 | |||||
1637 | /// Matches ZExt. | ||||
1638 | template <typename OpTy> | ||||
1639 | inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) { | ||||
1640 | return CastClass_match<OpTy, Instruction::ZExt>(Op); | ||||
1641 | } | ||||
1642 | |||||
1643 | template <typename OpTy> | ||||
1644 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy> | ||||
1645 | m_ZExtOrSelf(const OpTy &Op) { | ||||
1646 | return m_CombineOr(m_ZExt(Op), Op); | ||||
1647 | } | ||||
1648 | |||||
1649 | template <typename OpTy> | ||||
1650 | inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy> | ||||
1651 | m_SExtOrSelf(const OpTy &Op) { | ||||
1652 | return m_CombineOr(m_SExt(Op), Op); | ||||
1653 | } | ||||
1654 | |||||
1655 | template <typename OpTy> | ||||
1656 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, | ||||
1657 | CastClass_match<OpTy, Instruction::SExt>> | ||||
1658 | m_ZExtOrSExt(const OpTy &Op) { | ||||
1659 | return m_CombineOr(m_ZExt(Op), m_SExt(Op)); | ||||
1660 | } | ||||
1661 | |||||
1662 | template <typename OpTy> | ||||
1663 | inline match_combine_or< | ||||
1664 | match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, | ||||
1665 | CastClass_match<OpTy, Instruction::SExt>>, | ||||
1666 | OpTy> | ||||
1667 | m_ZExtOrSExtOrSelf(const OpTy &Op) { | ||||
1668 | return m_CombineOr(m_ZExtOrSExt(Op), Op); | ||||
1669 | } | ||||
1670 | |||||
1671 | template <typename OpTy> | ||||
1672 | inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) { | ||||
1673 | return CastClass_match<OpTy, Instruction::UIToFP>(Op); | ||||
1674 | } | ||||
1675 | |||||
1676 | template <typename OpTy> | ||||
1677 | inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) { | ||||
1678 | return CastClass_match<OpTy, Instruction::SIToFP>(Op); | ||||
1679 | } | ||||
1680 | |||||
1681 | template <typename OpTy> | ||||
1682 | inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) { | ||||
1683 | return CastClass_match<OpTy, Instruction::FPToUI>(Op); | ||||
1684 | } | ||||
1685 | |||||
1686 | template <typename OpTy> | ||||
1687 | inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) { | ||||
1688 | return CastClass_match<OpTy, Instruction::FPToSI>(Op); | ||||
1689 | } | ||||
1690 | |||||
1691 | template <typename OpTy> | ||||
1692 | inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) { | ||||
1693 | return CastClass_match<OpTy, Instruction::FPTrunc>(Op); | ||||
1694 | } | ||||
1695 | |||||
1696 | template <typename OpTy> | ||||
1697 | inline 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 | |||||
1705 | struct 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 | |||||
1720 | inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); } | ||||
1721 | |||||
1722 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> | ||||
1723 | struct 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 | |||||
1739 | template <typename Cond_t> | ||||
1740 | inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>> | ||||
1741 | m_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 | |||||
1746 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> | ||||
1747 | inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t> | ||||
1748 | m_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 | |||||
1756 | template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t, | ||||
1757 | bool Commutable = false> | ||||
1758 | struct 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. | ||||
1807 | struct 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. | ||||
1814 | struct 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. | ||||
1821 | struct 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. | ||||
1828 | struct 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. | ||||
1835 | struct 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. | ||||
1842 | struct 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. | ||||
1849 | struct 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. | ||||
1856 | struct ufmin_pred_ty { | ||||
1857 | static bool match(FCmpInst::Predicate Pred) { | ||||
1858 | return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE; | ||||
1859 | } | ||||
1860 | }; | ||||
1861 | |||||
1862 | template <typename LHS, typename RHS> | ||||
1863 | inline 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 | |||||
1868 | template <typename LHS, typename RHS> | ||||
1869 | inline 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 | |||||
1874 | template <typename LHS, typename RHS> | ||||
1875 | inline 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 | |||||
1880 | template <typename LHS, typename RHS> | ||||
1881 | inline 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 | |||||
1886 | template <typename LHS, typename RHS> | ||||
1887 | inline 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>>> | ||||
1892 | m_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 | ||||
1906 | template <typename LHS, typename RHS> | ||||
1907 | inline 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 | ||||
1921 | template <typename LHS, typename RHS> | ||||
1922 | inline 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 | ||||
1936 | template <typename LHS, typename RHS> | ||||
1937 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty> | ||||
1938 | m_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 | ||||
1951 | template <typename LHS, typename RHS> | ||||
1952 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty> | ||||
1953 | m_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 | |||||
1962 | template <typename LHS_t, typename RHS_t, typename Sum_t> | ||||
1963 | struct 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. | ||||
2025 | template <typename LHS_t, typename RHS_t, typename Sum_t> | ||||
2026 | UAddWithOverflow_match<LHS_t, RHS_t, Sum_t> | ||||
2027 | m_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 | |||||
2031 | template <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. | ||||
2046 | template <unsigned OpI, typename Opnd_t> | ||||
2047 | inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) { | ||||
2048 | return Argument_match<Opnd_t>(OpI, Op); | ||||
2049 | } | ||||
2050 | |||||
2051 | /// Intrinsic matchers. | ||||
2052 | struct 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 | ||||
2069 | template <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> | ||||
2073 | struct m_Intrinsic_Ty; | ||||
2074 | template <typename T0> struct m_Intrinsic_Ty<T0> { | ||||
2075 | using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>; | ||||
2076 | }; | ||||
2077 | template <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 | }; | ||||
2081 | template <typename T0, typename T1, typename T2> | ||||
2082 | struct 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 | }; | ||||
2087 | template <typename T0, typename T1, typename T2, typename T3> | ||||
2088 | struct 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 | |||||
2094 | template <typename T0, typename T1, typename T2, typename T3, typename T4> | ||||
2095 | struct 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 | |||||
2100 | template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5> | ||||
2101 | struct 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)) | ||||
2108 | template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() { | ||||
2109 | return IntrinsicID_match(IntrID); | ||||
2110 | } | ||||
2111 | |||||
2112 | /// Matches MaskedLoad Intrinsic. | ||||
2113 | template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3> | ||||
2114 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty | ||||
2115 | m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2, | ||||
2116 | const Opnd3 &Op3) { | ||||
2117 | return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3); | ||||
2118 | } | ||||
2119 | |||||
2120 | template <Intrinsic::ID IntrID, typename T0> | ||||
2121 | inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) { | ||||
2122 | return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0)); | ||||
2123 | } | ||||
2124 | |||||
2125 | template <Intrinsic::ID IntrID, typename T0, typename T1> | ||||
2126 | inline 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 | |||||
2131 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2> | ||||
2132 | inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty | ||||
2133 | m_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 | |||||
2137 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||
2138 | typename T3> | ||||
2139 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty | ||||
2140 | m_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 | |||||
2144 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||
2145 | typename T3, typename T4> | ||||
2146 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty | ||||
2147 | m_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 | |||||
2153 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||
2154 | typename T3, typename T4, typename T5> | ||||
2155 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty | ||||
2156 | m_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. | ||||
2163 | template <typename Opnd0> | ||||
2164 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) { | ||||
2165 | return m_Intrinsic<Intrinsic::bitreverse>(Op0); | ||||
2166 | } | ||||
2167 | |||||
2168 | template <typename Opnd0> | ||||
2169 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) { | ||||
2170 | return m_Intrinsic<Intrinsic::bswap>(Op0); | ||||
2171 | } | ||||
2172 | |||||
2173 | template <typename Opnd0> | ||||
2174 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) { | ||||
2175 | return m_Intrinsic<Intrinsic::fabs>(Op0); | ||||
2176 | } | ||||
2177 | |||||
2178 | template <typename Opnd0> | ||||
2179 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) { | ||||
2180 | return m_Intrinsic<Intrinsic::canonicalize>(Op0); | ||||
2181 | } | ||||
2182 | |||||
2183 | template <typename Opnd0, typename Opnd1> | ||||
2184 | inline 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 | |||||
2189 | template <typename Opnd0, typename Opnd1> | ||||
2190 | inline 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 | |||||
2195 | template <typename Opnd0, typename Opnd1, typename Opnd2> | ||||
2196 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty | ||||
2197 | m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { | ||||
2198 | return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2); | ||||
2199 | } | ||||
2200 | |||||
2201 | template <typename Opnd0, typename Opnd1, typename Opnd2> | ||||
2202 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty | ||||
2203 | m_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. | ||||
2212 | template <typename LHS, typename RHS> | ||||
2213 | inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) { | ||||
2214 | return AnyBinaryOp_match<LHS, RHS, true>(L, R); | ||||
2215 | } | ||||
2216 | |||||
2217 | /// Matches an ICmp with a predicate over LHS and RHS in either order. | ||||
2218 | /// Swaps the predicate if operands are commuted. | ||||
2219 | template <typename LHS, typename RHS> | ||||
2220 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true> | ||||
2221 | m_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 Add with LHS and RHS in either order. | ||||
2227 | template <typename LHS, typename RHS> | ||||
2228 | inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L, | ||||
2229 | const RHS &R) { | ||||
2230 | return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R); | ||||
2231 | } | ||||
2232 | |||||
2233 | /// Matches a Mul with LHS and RHS in either order. | ||||
2234 | template <typename LHS, typename RHS> | ||||
2235 | inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L, | ||||
2236 | const RHS &R) { | ||||
2237 | return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R); | ||||
2238 | } | ||||
2239 | |||||
2240 | /// Matches an And with LHS and RHS in either order. | ||||
2241 | template <typename LHS, typename RHS> | ||||
2242 | inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L, | ||||
2243 | const RHS &R) { | ||||
2244 | return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R); | ||||
2245 | } | ||||
2246 | |||||
2247 | /// Matches an Or with LHS and RHS in either order. | ||||
2248 | template <typename LHS, typename RHS> | ||||
2249 | inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L, | ||||
2250 | const RHS &R) { | ||||
2251 | return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R); | ||||
2252 | } | ||||
2253 | |||||
2254 | /// Matches an Xor with LHS and RHS in either order. | ||||
2255 | template <typename LHS, typename RHS> | ||||
2256 | inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L, | ||||
2257 | const RHS &R) { | ||||
2258 | return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R); | ||||
2259 | } | ||||
2260 | |||||
2261 | /// Matches a 'Neg' as 'sub 0, V'. | ||||
2262 | template <typename ValTy> | ||||
2263 | inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub> | ||||
2264 | m_Neg(const ValTy &V) { | ||||
2265 | return m_Sub(m_ZeroInt(), V); | ||||
2266 | } | ||||
2267 | |||||
2268 | /// Matches a 'Neg' as 'sub nsw 0, V'. | ||||
2269 | template <typename ValTy> | ||||
2270 | inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, | ||||
2271 | Instruction::Sub, | ||||
2272 | OverflowingBinaryOperator::NoSignedWrap> | ||||
2273 | m_NSWNeg(const ValTy &V) { | ||||
2274 | return m_NSWSub(m_ZeroInt(), V); | ||||
2275 | } | ||||
2276 | |||||
2277 | /// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'. | ||||
2278 | template <typename ValTy> | ||||
2279 | inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true> | ||||
2280 | m_Not(const ValTy &V) { | ||||
2281 | return m_c_Xor(V, m_AllOnes()); | ||||
2282 | } | ||||
2283 | |||||
2284 | /// Matches an SMin with LHS and RHS in either order. | ||||
2285 | template <typename LHS, typename RHS> | ||||
2286 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true> | ||||
2287 | m_c_SMin(const LHS &L, const RHS &R) { | ||||
2288 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R); | ||||
2289 | } | ||||
2290 | /// Matches an SMax with LHS and RHS in either order. | ||||
2291 | template <typename LHS, typename RHS> | ||||
2292 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true> | ||||
2293 | m_c_SMax(const LHS &L, const RHS &R) { | ||||
2294 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R); | ||||
2295 | } | ||||
2296 | /// Matches a UMin with LHS and RHS in either order. | ||||
2297 | template <typename LHS, typename RHS> | ||||
2298 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true> | ||||
2299 | m_c_UMin(const LHS &L, const RHS &R) { | ||||
2300 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R); | ||||
2301 | } | ||||
2302 | /// Matches a UMax with LHS and RHS in either order. | ||||
2303 | template <typename LHS, typename RHS> | ||||
2304 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true> | ||||
2305 | m_c_UMax(const LHS &L, const RHS &R) { | ||||
2306 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R); | ||||
2307 | } | ||||
2308 | |||||
2309 | template <typename LHS, typename RHS> | ||||
2310 | inline match_combine_or< | ||||
2311 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>, | ||||
2312 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>, | ||||
2313 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>, | ||||
2314 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>> | ||||
2315 | m_c_MaxOrMin(const LHS &L, const RHS &R) { | ||||
2316 | return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)), | ||||
2317 | m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R))); | ||||
2318 | } | ||||
2319 | |||||
2320 | /// Matches FAdd with LHS and RHS in either order. | ||||
2321 | template <typename LHS, typename RHS> | ||||
2322 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true> | ||||
2323 | m_c_FAdd(const LHS &L, const RHS &R) { | ||||
2324 | return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R); | ||||
2325 | } | ||||
2326 | |||||
2327 | /// Matches FMul with LHS and RHS in either order. | ||||
2328 | template <typename LHS, typename RHS> | ||||
2329 | inline BinaryOp_match<LHS, RHS, Instruction::FMul, true> | ||||
2330 | m_c_FMul(const LHS &L, const RHS &R) { | ||||
2331 | return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R); | ||||
2332 | } | ||||
2333 | |||||
2334 | template <typename Opnd_t> struct Signum_match { | ||||
2335 | Opnd_t Val; | ||||
2336 | Signum_match(const Opnd_t &V) : Val(V) {} | ||||
2337 | |||||
2338 | template <typename OpTy> bool match(OpTy *V) { | ||||
2339 | unsigned TypeSize = V->getType()->getScalarSizeInBits(); | ||||
2340 | if (TypeSize == 0) | ||||
2341 | return false; | ||||
2342 | |||||
2343 | unsigned ShiftWidth = TypeSize - 1; | ||||
2344 | Value *OpL = nullptr, *OpR = nullptr; | ||||
2345 | |||||
2346 | // This is the representation of signum we match: | ||||
2347 | // | ||||
2348 | // signum(x) == (x >> 63) | (-x >>u 63) | ||||
2349 | // | ||||
2350 | // An i1 value is its own signum, so it's correct to match | ||||
2351 | // | ||||
2352 | // signum(x) == (x >> 0) | (-x >>u 0) | ||||
2353 | // | ||||
2354 | // for i1 values. | ||||
2355 | |||||
2356 | auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth)); | ||||
2357 | auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth)); | ||||
2358 | auto Signum = m_Or(LHS, RHS); | ||||
2359 | |||||
2360 | return Signum.match(V) && OpL == OpR && Val.match(OpL); | ||||
2361 | } | ||||
2362 | }; | ||||
2363 | |||||
2364 | /// Matches a signum pattern. | ||||
2365 | /// | ||||
2366 | /// signum(x) = | ||||
2367 | /// x > 0 -> 1 | ||||
2368 | /// x == 0 -> 0 | ||||
2369 | /// x < 0 -> -1 | ||||
2370 | template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) { | ||||
2371 | return Signum_match<Val_t>(V); | ||||
2372 | } | ||||
2373 | |||||
2374 | template <int Ind, typename Opnd_t> struct ExtractValue_match { | ||||
2375 | Opnd_t Val; | ||||
2376 | ExtractValue_match(const Opnd_t &V) : Val(V) {} | ||||
2377 | |||||
2378 | template <typename OpTy> bool match(OpTy *V) { | ||||
2379 | if (auto *I = dyn_cast<ExtractValueInst>(V)) { | ||||
2380 | // If Ind is -1, don't inspect indices | ||||
2381 | if (Ind != -1 && | ||||
2382 | !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind)) | ||||
2383 | return false; | ||||
2384 | return Val.match(I->getAggregateOperand()); | ||||
2385 | } | ||||
2386 | return false; | ||||
2387 | } | ||||
2388 | }; | ||||
2389 | |||||
2390 | /// Match a single index ExtractValue instruction. | ||||
2391 | /// For example m_ExtractValue<1>(...) | ||||
2392 | template <int Ind, typename Val_t> | ||||
2393 | inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) { | ||||
2394 | return ExtractValue_match<Ind, Val_t>(V); | ||||
2395 | } | ||||
2396 | |||||
2397 | /// Match an ExtractValue instruction with any index. | ||||
2398 | /// For example m_ExtractValue(...) | ||||
2399 | template <typename Val_t> | ||||
2400 | inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) { | ||||
2401 | return ExtractValue_match<-1, Val_t>(V); | ||||
2402 | } | ||||
2403 | |||||
2404 | /// Matcher for a single index InsertValue instruction. | ||||
2405 | template <int Ind, typename T0, typename T1> struct InsertValue_match { | ||||
2406 | T0 Op0; | ||||
2407 | T1 Op1; | ||||
2408 | |||||
2409 | InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {} | ||||
2410 | |||||
2411 | template <typename OpTy> bool match(OpTy *V) { | ||||
2412 | if (auto *I = dyn_cast<InsertValueInst>(V)) { | ||||
2413 | return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) && | ||||
2414 | I->getNumIndices() == 1 && Ind == I->getIndices()[0]; | ||||
2415 | } | ||||
2416 | return false; | ||||
2417 | } | ||||
2418 | }; | ||||
2419 | |||||
2420 | /// Matches a single index InsertValue instruction. | ||||
2421 | template <int Ind, typename Val_t, typename Elt_t> | ||||
2422 | inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val, | ||||
2423 | const Elt_t &Elt) { | ||||
2424 | return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt); | ||||
2425 | } | ||||
2426 | |||||
2427 | /// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or | ||||
2428 | /// the constant expression | ||||
2429 | /// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>` | ||||
2430 | /// under the right conditions determined by DataLayout. | ||||
2431 | struct VScaleVal_match { | ||||
2432 | const DataLayout &DL; | ||||
2433 | VScaleVal_match(const DataLayout &DL) : DL(DL) {} | ||||
2434 | |||||
2435 | template <typename ITy> bool match(ITy *V) { | ||||
2436 | if (m_Intrinsic<Intrinsic::vscale>().match(V)) | ||||
2437 | return true; | ||||
2438 | |||||
2439 | Value *Ptr; | ||||
2440 | if (m_PtrToInt(m_Value(Ptr)).match(V)) { | ||||
2441 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { | ||||
2442 | auto *DerefTy = GEP->getSourceElementType(); | ||||
2443 | if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) && | ||||
2444 | m_Zero().match(GEP->getPointerOperand()) && | ||||
2445 | m_SpecificInt(1).match(GEP->idx_begin()->get()) && | ||||
2446 | DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8) | ||||
2447 | return true; | ||||
2448 | } | ||||
2449 | } | ||||
2450 | |||||
2451 | return false; | ||||
2452 | } | ||||
2453 | }; | ||||
2454 | |||||
2455 | inline VScaleVal_match m_VScale(const DataLayout &DL) { | ||||
2456 | return VScaleVal_match(DL); | ||||
2457 | } | ||||
2458 | |||||
2459 | template <typename LHS, typename RHS, unsigned Opcode> | ||||
2460 | struct LogicalOp_match { | ||||
2461 | LHS L; | ||||
2462 | RHS R; | ||||
2463 | |||||
2464 | LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {} | ||||
2465 | |||||
2466 | template <typename T> bool match(T *V) { | ||||
2467 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
2468 | if (!I->getType()->isIntOrIntVectorTy(1)) | ||||
2469 | return false; | ||||
2470 | |||||
2471 | if (I->getOpcode() == Opcode && L.match(I->getOperand(0)) && | ||||
2472 | R.match(I->getOperand(1))) | ||||
2473 | return true; | ||||
2474 | |||||
2475 | if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||
2476 | if (Opcode == Instruction::And) { | ||||
2477 | if (const auto *C = dyn_cast<Constant>(SI->getFalseValue())) | ||||
2478 | if (C->isNullValue() && L.match(SI->getCondition()) && | ||||
2479 | R.match(SI->getTrueValue())) | ||||
2480 | return true; | ||||
2481 | } else { | ||||
2482 | assert(Opcode == Instruction::Or)(static_cast <bool> (Opcode == Instruction::Or) ? void ( 0) : __assert_fail ("Opcode == Instruction::Or", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h" , 2482, __extension__ __PRETTY_FUNCTION__)); | ||||
2483 | if (const auto *C = dyn_cast<Constant>(SI->getTrueValue())) | ||||
2484 | if (C->isOneValue() && L.match(SI->getCondition()) && | ||||
2485 | R.match(SI->getFalseValue())) | ||||
2486 | return true; | ||||
2487 | } | ||||
2488 | } | ||||
2489 | } | ||||
2490 | |||||
2491 | return false; | ||||
2492 | } | ||||
2493 | }; | ||||
2494 | |||||
2495 | /// Matches L && R either in the form of L & R or L ? R : false. | ||||
2496 | /// Note that the latter form is poison-blocking. | ||||
2497 | template <typename LHS, typename RHS> | ||||
2498 | inline LogicalOp_match<LHS, RHS, Instruction::And> | ||||
2499 | m_LogicalAnd(const LHS &L, const RHS &R) { | ||||
2500 | return LogicalOp_match<LHS, RHS, Instruction::And>(L, R); | ||||
2501 | } | ||||
2502 | |||||
2503 | /// Matches L && R where L and R are arbitrary values. | ||||
2504 | inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); } | ||||
2505 | |||||
2506 | /// Matches L || R either in the form of L | R or L ? true : R. | ||||
2507 | /// Note that the latter form is poison-blocking. | ||||
2508 | template <typename LHS, typename RHS> | ||||
2509 | inline LogicalOp_match<LHS, RHS, Instruction::Or> | ||||
2510 | m_LogicalOr(const LHS &L, const RHS &R) { | ||||
2511 | return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R); | ||||
2512 | } | ||||
2513 | |||||
2514 | /// Matches L || R where L and R are arbitrary values. | ||||
2515 | inline auto m_LogicalOr() { | ||||
2516 | return m_LogicalOr(m_Value(), m_Value()); | ||||
2517 | } | ||||
2518 | |||||
2519 | } // end namespace PatternMatch | ||||
2520 | } // end namespace llvm | ||||
2521 | |||||
2522 | #endif // LLVM_IR_PATTERNMATCH_H |