File: | llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp |
Warning: | line 225, column 33 Called C++ object pointer is uninitialized |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- InstCombineShifts.cpp ----------------------------------------------===// | ||||
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 visitShl, visitLShr, and visitAShr functions. | ||||
10 | // | ||||
11 | //===----------------------------------------------------------------------===// | ||||
12 | |||||
13 | #include "InstCombineInternal.h" | ||||
14 | #include "llvm/Analysis/ConstantFolding.h" | ||||
15 | #include "llvm/Analysis/InstructionSimplify.h" | ||||
16 | #include "llvm/IR/IntrinsicInst.h" | ||||
17 | #include "llvm/IR/PatternMatch.h" | ||||
18 | #include "llvm/Transforms/InstCombine/InstCombiner.h" | ||||
19 | using namespace llvm; | ||||
20 | using namespace PatternMatch; | ||||
21 | |||||
22 | #define DEBUG_TYPE"instcombine" "instcombine" | ||||
23 | |||||
24 | // Given pattern: | ||||
25 | // (x shiftopcode Q) shiftopcode K | ||||
26 | // we should rewrite it as | ||||
27 | // x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and | ||||
28 | // | ||||
29 | // This is valid for any shift, but they must be identical, and we must be | ||||
30 | // careful in case we have (zext(Q)+zext(K)) and look past extensions, | ||||
31 | // (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus. | ||||
32 | // | ||||
33 | // AnalyzeForSignBitExtraction indicates that we will only analyze whether this | ||||
34 | // pattern has any 2 right-shifts that sum to 1 less than original bit width. | ||||
35 | Value *InstCombinerImpl::reassociateShiftAmtsOfTwoSameDirectionShifts( | ||||
36 | BinaryOperator *Sh0, const SimplifyQuery &SQ, | ||||
37 | bool AnalyzeForSignBitExtraction) { | ||||
38 | // Look for a shift of some instruction, ignore zext of shift amount if any. | ||||
39 | Instruction *Sh0Op0; | ||||
40 | Value *ShAmt0; | ||||
41 | if (!match(Sh0, | ||||
42 | m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0))))) | ||||
43 | return nullptr; | ||||
44 | |||||
45 | // If there is a truncation between the two shifts, we must make note of it | ||||
46 | // and look through it. The truncation imposes additional constraints on the | ||||
47 | // transform. | ||||
48 | Instruction *Sh1; | ||||
49 | Value *Trunc = nullptr; | ||||
50 | match(Sh0Op0, | ||||
51 | m_CombineOr(m_CombineAnd(m_Trunc(m_Instruction(Sh1)), m_Value(Trunc)), | ||||
52 | m_Instruction(Sh1))); | ||||
53 | |||||
54 | // Inner shift: (x shiftopcode ShAmt1) | ||||
55 | // Like with other shift, ignore zext of shift amount if any. | ||||
56 | Value *X, *ShAmt1; | ||||
57 | if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1))))) | ||||
58 | return nullptr; | ||||
59 | |||||
60 | // We have two shift amounts from two different shifts. The types of those | ||||
61 | // shift amounts may not match. If that's the case let's bailout now.. | ||||
62 | if (ShAmt0->getType() != ShAmt1->getType()) | ||||
63 | return nullptr; | ||||
64 | |||||
65 | // As input, we have the following pattern: | ||||
66 | // Sh0 (Sh1 X, Q), K | ||||
67 | // We want to rewrite that as: | ||||
68 | // Sh x, (Q+K) iff (Q+K) u< bitwidth(x) | ||||
69 | // While we know that originally (Q+K) would not overflow | ||||
70 | // (because 2 * (N-1) u<= iN -1), we have looked past extensions of | ||||
71 | // shift amounts. so it may now overflow in smaller bitwidth. | ||||
72 | // To ensure that does not happen, we need to ensure that the total maximal | ||||
73 | // shift amount is still representable in that smaller bit width. | ||||
74 | unsigned MaximalPossibleTotalShiftAmount = | ||||
75 | (Sh0->getType()->getScalarSizeInBits() - 1) + | ||||
76 | (Sh1->getType()->getScalarSizeInBits() - 1); | ||||
77 | APInt MaximalRepresentableShiftAmount = | ||||
78 | APInt::getAllOnesValue(ShAmt0->getType()->getScalarSizeInBits()); | ||||
79 | if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount)) | ||||
80 | return nullptr; | ||||
81 | |||||
82 | // We are only looking for signbit extraction if we have two right shifts. | ||||
83 | bool HadTwoRightShifts = match(Sh0, m_Shr(m_Value(), m_Value())) && | ||||
84 | match(Sh1, m_Shr(m_Value(), m_Value())); | ||||
85 | // ... and if it's not two right-shifts, we know the answer already. | ||||
86 | if (AnalyzeForSignBitExtraction && !HadTwoRightShifts) | ||||
87 | return nullptr; | ||||
88 | |||||
89 | // The shift opcodes must be identical, unless we are just checking whether | ||||
90 | // this pattern can be interpreted as a sign-bit-extraction. | ||||
91 | Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode(); | ||||
92 | bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode(); | ||||
93 | if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction) | ||||
94 | return nullptr; | ||||
95 | |||||
96 | // If we saw truncation, we'll need to produce extra instruction, | ||||
97 | // and for that one of the operands of the shift must be one-use, | ||||
98 | // unless of course we don't actually plan to produce any instructions here. | ||||
99 | if (Trunc && !AnalyzeForSignBitExtraction && | ||||
100 | !match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value()))) | ||||
101 | return nullptr; | ||||
102 | |||||
103 | // Can we fold (ShAmt0+ShAmt1) ? | ||||
104 | auto *NewShAmt = dyn_cast_or_null<Constant>( | ||||
105 | SimplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false, | ||||
106 | SQ.getWithInstruction(Sh0))); | ||||
107 | if (!NewShAmt) | ||||
108 | return nullptr; // Did not simplify. | ||||
109 | unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits(); | ||||
110 | unsigned XBitWidth = X->getType()->getScalarSizeInBits(); | ||||
111 | // Is the new shift amount smaller than the bit width of inner/new shift? | ||||
112 | if (!match(NewShAmt, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT, | ||||
113 | APInt(NewShAmtBitWidth, XBitWidth)))) | ||||
114 | return nullptr; // FIXME: could perform constant-folding. | ||||
115 | |||||
116 | // If there was a truncation, and we have a right-shift, we can only fold if | ||||
117 | // we are left with the original sign bit. Likewise, if we were just checking | ||||
118 | // that this is a sighbit extraction, this is the place to check it. | ||||
119 | // FIXME: zero shift amount is also legal here, but we can't *easily* check | ||||
120 | // more than one predicate so it's not really worth it. | ||||
121 | if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) { | ||||
122 | // If it's not a sign bit extraction, then we're done. | ||||
123 | if (!match(NewShAmt, | ||||
124 | m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, | ||||
125 | APInt(NewShAmtBitWidth, XBitWidth - 1)))) | ||||
126 | return nullptr; | ||||
127 | // If it is, and that was the question, return the base value. | ||||
128 | if (AnalyzeForSignBitExtraction) | ||||
129 | return X; | ||||
130 | } | ||||
131 | |||||
132 | assert(IdenticalShOpcodes && "Should not get here with different shifts.")((IdenticalShOpcodes && "Should not get here with different shifts." ) ? static_cast<void> (0) : __assert_fail ("IdenticalShOpcodes && \"Should not get here with different shifts.\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 132, __PRETTY_FUNCTION__)); | ||||
133 | |||||
134 | // All good, we can do this fold. | ||||
135 | NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, X->getType()); | ||||
136 | |||||
137 | BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt); | ||||
138 | |||||
139 | // The flags can only be propagated if there wasn't a trunc. | ||||
140 | if (!Trunc) { | ||||
141 | // If the pattern did not involve trunc, and both of the original shifts | ||||
142 | // had the same flag set, preserve the flag. | ||||
143 | if (ShiftOpcode == Instruction::BinaryOps::Shl) { | ||||
144 | NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() && | ||||
145 | Sh1->hasNoUnsignedWrap()); | ||||
146 | NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() && | ||||
147 | Sh1->hasNoSignedWrap()); | ||||
148 | } else { | ||||
149 | NewShift->setIsExact(Sh0->isExact() && Sh1->isExact()); | ||||
150 | } | ||||
151 | } | ||||
152 | |||||
153 | Instruction *Ret = NewShift; | ||||
154 | if (Trunc) { | ||||
155 | Builder.Insert(NewShift); | ||||
156 | Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType()); | ||||
157 | } | ||||
158 | |||||
159 | return Ret; | ||||
160 | } | ||||
161 | |||||
162 | // If we have some pattern that leaves only some low bits set, and then performs | ||||
163 | // left-shift of those bits, if none of the bits that are left after the final | ||||
164 | // shift are modified by the mask, we can omit the mask. | ||||
165 | // | ||||
166 | // There are many variants to this pattern: | ||||
167 | // a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt | ||||
168 | // b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt | ||||
169 | // c) (x & (-1 >> MaskShAmt)) << ShiftShAmt | ||||
170 | // d) (x & ((-1 << MaskShAmt) >> MaskShAmt)) << ShiftShAmt | ||||
171 | // e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt | ||||
172 | // f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt | ||||
173 | // All these patterns can be simplified to just: | ||||
174 | // x << ShiftShAmt | ||||
175 | // iff: | ||||
176 | // a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x) | ||||
177 | // c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt) | ||||
178 | static Instruction * | ||||
179 | dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift, | ||||
180 | const SimplifyQuery &Q, | ||||
181 | InstCombiner::BuilderTy &Builder) { | ||||
182 | assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&((OuterShift->getOpcode() == Instruction::BinaryOps::Shl && "The input must be 'shl'!") ? static_cast<void> (0) : __assert_fail ("OuterShift->getOpcode() == Instruction::BinaryOps::Shl && \"The input must be 'shl'!\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 183, __PRETTY_FUNCTION__)) | ||||
183 | "The input must be 'shl'!")((OuterShift->getOpcode() == Instruction::BinaryOps::Shl && "The input must be 'shl'!") ? static_cast<void> (0) : __assert_fail ("OuterShift->getOpcode() == Instruction::BinaryOps::Shl && \"The input must be 'shl'!\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 183, __PRETTY_FUNCTION__)); | ||||
184 | |||||
185 | Value *Masked, *ShiftShAmt; | ||||
186 | match(OuterShift, | ||||
187 | m_Shift(m_Value(Masked), m_ZExtOrSelf(m_Value(ShiftShAmt)))); | ||||
188 | |||||
189 | // *If* there is a truncation between an outer shift and a possibly-mask, | ||||
190 | // then said truncation *must* be one-use, else we can't perform the fold. | ||||
191 | Value *Trunc; | ||||
192 | if (match(Masked, m_CombineAnd(m_Trunc(m_Value(Masked)), m_Value(Trunc))) && | ||||
193 | !Trunc->hasOneUse()) | ||||
194 | return nullptr; | ||||
195 | |||||
196 | Type *NarrowestTy = OuterShift->getType(); | ||||
197 | Type *WidestTy = Masked->getType(); | ||||
198 | bool HadTrunc = WidestTy != NarrowestTy; | ||||
199 | |||||
200 | // The mask must be computed in a type twice as wide to ensure | ||||
201 | // that no bits are lost if the sum-of-shifts is wider than the base type. | ||||
202 | Type *ExtendedTy = WidestTy->getExtendedType(); | ||||
203 | |||||
204 | Value *MaskShAmt; | ||||
205 | |||||
206 | // ((1 << MaskShAmt) - 1) | ||||
207 | auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes()); | ||||
208 | // (~(-1 << maskNbits)) | ||||
209 | auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes()); | ||||
210 | // (-1 >> MaskShAmt) | ||||
211 | auto MaskC = m_Shr(m_AllOnes(), m_Value(MaskShAmt)); | ||||
212 | // ((-1 << MaskShAmt) >> MaskShAmt) | ||||
213 | auto MaskD = | ||||
214 | m_Shr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt)); | ||||
215 | |||||
216 | Value *X; | ||||
217 | Constant *NewMask; | ||||
218 | |||||
219 | if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) { | ||||
| |||||
| |||||
220 | // Peek through an optional zext of the shift amount. | ||||
221 | match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt))); | ||||
222 | |||||
223 | // We have two shift amounts from two different shifts. The types of those | ||||
224 | // shift amounts may not match. If that's the case let's bailout now. | ||||
225 | if (MaskShAmt->getType() != ShiftShAmt->getType()) | ||||
| |||||
226 | return nullptr; | ||||
227 | |||||
228 | // Can we simplify (MaskShAmt+ShiftShAmt) ? | ||||
229 | auto *SumOfShAmts = dyn_cast_or_null<Constant>(SimplifyAddInst( | ||||
230 | MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); | ||||
231 | if (!SumOfShAmts) | ||||
232 | return nullptr; // Did not simplify. | ||||
233 | // In this pattern SumOfShAmts correlates with the number of low bits | ||||
234 | // that shall remain in the root value (OuterShift). | ||||
235 | |||||
236 | // An extend of an undef value becomes zero because the high bits are never | ||||
237 | // completely unknown. Replace the the `undef` shift amounts with final | ||||
238 | // shift bitwidth to ensure that the value remains undef when creating the | ||||
239 | // subsequent shift op. | ||||
240 | SumOfShAmts = Constant::replaceUndefsWith( | ||||
241 | SumOfShAmts, ConstantInt::get(SumOfShAmts->getType()->getScalarType(), | ||||
242 | ExtendedTy->getScalarSizeInBits())); | ||||
243 | auto *ExtendedSumOfShAmts = ConstantExpr::getZExt(SumOfShAmts, ExtendedTy); | ||||
244 | // And compute the mask as usual: ~(-1 << (SumOfShAmts)) | ||||
245 | auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy); | ||||
246 | auto *ExtendedInvertedMask = | ||||
247 | ConstantExpr::getShl(ExtendedAllOnes, ExtendedSumOfShAmts); | ||||
248 | NewMask = ConstantExpr::getNot(ExtendedInvertedMask); | ||||
249 | } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) || | ||||
250 | match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)), | ||||
251 | m_Deferred(MaskShAmt)))) { | ||||
252 | // Peek through an optional zext of the shift amount. | ||||
253 | match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt))); | ||||
254 | |||||
255 | // We have two shift amounts from two different shifts. The types of those | ||||
256 | // shift amounts may not match. If that's the case let's bailout now. | ||||
257 | if (MaskShAmt->getType() != ShiftShAmt->getType()) | ||||
258 | return nullptr; | ||||
259 | |||||
260 | // Can we simplify (ShiftShAmt-MaskShAmt) ? | ||||
261 | auto *ShAmtsDiff = dyn_cast_or_null<Constant>(SimplifySubInst( | ||||
262 | ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); | ||||
263 | if (!ShAmtsDiff) | ||||
264 | return nullptr; // Did not simplify. | ||||
265 | // In this pattern ShAmtsDiff correlates with the number of high bits that | ||||
266 | // shall be unset in the root value (OuterShift). | ||||
267 | |||||
268 | // An extend of an undef value becomes zero because the high bits are never | ||||
269 | // completely unknown. Replace the the `undef` shift amounts with negated | ||||
270 | // bitwidth of innermost shift to ensure that the value remains undef when | ||||
271 | // creating the subsequent shift op. | ||||
272 | unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits(); | ||||
273 | ShAmtsDiff = Constant::replaceUndefsWith( | ||||
274 | ShAmtsDiff, ConstantInt::get(ShAmtsDiff->getType()->getScalarType(), | ||||
275 | -WidestTyBitWidth)); | ||||
276 | auto *ExtendedNumHighBitsToClear = ConstantExpr::getZExt( | ||||
277 | ConstantExpr::getSub(ConstantInt::get(ShAmtsDiff->getType(), | ||||
278 | WidestTyBitWidth, | ||||
279 | /*isSigned=*/false), | ||||
280 | ShAmtsDiff), | ||||
281 | ExtendedTy); | ||||
282 | // And compute the mask as usual: (-1 l>> (NumHighBitsToClear)) | ||||
283 | auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy); | ||||
284 | NewMask = | ||||
285 | ConstantExpr::getLShr(ExtendedAllOnes, ExtendedNumHighBitsToClear); | ||||
286 | } else | ||||
287 | return nullptr; // Don't know anything about this pattern. | ||||
288 | |||||
289 | NewMask = ConstantExpr::getTrunc(NewMask, NarrowestTy); | ||||
290 | |||||
291 | // Does this mask has any unset bits? If not then we can just not apply it. | ||||
292 | bool NeedMask = !match(NewMask, m_AllOnes()); | ||||
293 | |||||
294 | // If we need to apply a mask, there are several more restrictions we have. | ||||
295 | if (NeedMask) { | ||||
296 | // The old masking instruction must go away. | ||||
297 | if (!Masked->hasOneUse()) | ||||
298 | return nullptr; | ||||
299 | // The original "masking" instruction must not have been`ashr`. | ||||
300 | if (match(Masked, m_AShr(m_Value(), m_Value()))) | ||||
301 | return nullptr; | ||||
302 | } | ||||
303 | |||||
304 | // If we need to apply truncation, let's do it first, since we can. | ||||
305 | // We have already ensured that the old truncation will go away. | ||||
306 | if (HadTrunc) | ||||
307 | X = Builder.CreateTrunc(X, NarrowestTy); | ||||
308 | |||||
309 | // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits. | ||||
310 | // We didn't change the Type of this outermost shift, so we can just do it. | ||||
311 | auto *NewShift = BinaryOperator::Create(OuterShift->getOpcode(), X, | ||||
312 | OuterShift->getOperand(1)); | ||||
313 | if (!NeedMask) | ||||
314 | return NewShift; | ||||
315 | |||||
316 | Builder.Insert(NewShift); | ||||
317 | return BinaryOperator::Create(Instruction::And, NewShift, NewMask); | ||||
318 | } | ||||
319 | |||||
320 | /// If we have a shift-by-constant of a bitwise logic op that itself has a | ||||
321 | /// shift-by-constant operand with identical opcode, we may be able to convert | ||||
322 | /// that into 2 independent shifts followed by the logic op. This eliminates a | ||||
323 | /// a use of an intermediate value (reduces dependency chain). | ||||
324 | static Instruction *foldShiftOfShiftedLogic(BinaryOperator &I, | ||||
325 | InstCombiner::BuilderTy &Builder) { | ||||
326 | assert(I.isShift() && "Expected a shift as input")((I.isShift() && "Expected a shift as input") ? static_cast <void> (0) : __assert_fail ("I.isShift() && \"Expected a shift as input\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 326, __PRETTY_FUNCTION__)); | ||||
327 | auto *LogicInst = dyn_cast<BinaryOperator>(I.getOperand(0)); | ||||
328 | if (!LogicInst || !LogicInst->isBitwiseLogicOp() || !LogicInst->hasOneUse()) | ||||
329 | return nullptr; | ||||
330 | |||||
331 | Constant *C0, *C1; | ||||
332 | if (!match(I.getOperand(1), m_Constant(C1))) | ||||
333 | return nullptr; | ||||
334 | |||||
335 | Instruction::BinaryOps ShiftOpcode = I.getOpcode(); | ||||
336 | Type *Ty = I.getType(); | ||||
337 | |||||
338 | // Find a matching one-use shift by constant. The fold is not valid if the sum | ||||
339 | // of the shift values equals or exceeds bitwidth. | ||||
340 | // TODO: Remove the one-use check if the other logic operand (Y) is constant. | ||||
341 | Value *X, *Y; | ||||
342 | auto matchFirstShift = [&](Value *V) { | ||||
343 | BinaryOperator *BO; | ||||
344 | APInt Threshold(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits()); | ||||
345 | return match(V, m_BinOp(BO)) && BO->getOpcode() == ShiftOpcode && | ||||
346 | match(V, m_OneUse(m_Shift(m_Value(X), m_Constant(C0)))) && | ||||
347 | match(ConstantExpr::getAdd(C0, C1), | ||||
348 | m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold)); | ||||
349 | }; | ||||
350 | |||||
351 | // Logic ops are commutative, so check each operand for a match. | ||||
352 | if (matchFirstShift(LogicInst->getOperand(0))) | ||||
353 | Y = LogicInst->getOperand(1); | ||||
354 | else if (matchFirstShift(LogicInst->getOperand(1))) | ||||
355 | Y = LogicInst->getOperand(0); | ||||
356 | else | ||||
357 | return nullptr; | ||||
358 | |||||
359 | // shift (logic (shift X, C0), Y), C1 -> logic (shift X, C0+C1), (shift Y, C1) | ||||
360 | Constant *ShiftSumC = ConstantExpr::getAdd(C0, C1); | ||||
361 | Value *NewShift1 = Builder.CreateBinOp(ShiftOpcode, X, ShiftSumC); | ||||
362 | Value *NewShift2 = Builder.CreateBinOp(ShiftOpcode, Y, I.getOperand(1)); | ||||
363 | return BinaryOperator::Create(LogicInst->getOpcode(), NewShift1, NewShift2); | ||||
364 | } | ||||
365 | |||||
366 | Instruction *InstCombinerImpl::commonShiftTransforms(BinaryOperator &I) { | ||||
367 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
368 | assert(Op0->getType() == Op1->getType())((Op0->getType() == Op1->getType()) ? static_cast<void > (0) : __assert_fail ("Op0->getType() == Op1->getType()" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 368, __PRETTY_FUNCTION__)); | ||||
369 | |||||
370 | // If the shift amount is a one-use `sext`, we can demote it to `zext`. | ||||
371 | Value *Y; | ||||
372 | if (match(Op1, m_OneUse(m_SExt(m_Value(Y))))) { | ||||
373 | Value *NewExt = Builder.CreateZExt(Y, I.getType(), Op1->getName()); | ||||
374 | return BinaryOperator::Create(I.getOpcode(), Op0, NewExt); | ||||
375 | } | ||||
376 | |||||
377 | // See if we can fold away this shift. | ||||
378 | if (SimplifyDemandedInstructionBits(I)) | ||||
379 | return &I; | ||||
380 | |||||
381 | // Try to fold constant and into select arguments. | ||||
382 | if (isa<Constant>(Op0)) | ||||
383 | if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) | ||||
384 | if (Instruction *R = FoldOpIntoSelect(I, SI)) | ||||
385 | return R; | ||||
386 | |||||
387 | if (Constant *CUI = dyn_cast<Constant>(Op1)) | ||||
388 | if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I)) | ||||
389 | return Res; | ||||
390 | |||||
391 | if (auto *NewShift = cast_or_null<Instruction>( | ||||
392 | reassociateShiftAmtsOfTwoSameDirectionShifts(&I, SQ))) | ||||
393 | return NewShift; | ||||
394 | |||||
395 | // (C1 shift (A add C2)) -> (C1 shift C2) shift A) | ||||
396 | // iff A and C2 are both positive. | ||||
397 | Value *A; | ||||
398 | Constant *C; | ||||
399 | if (match(Op0, m_Constant()) && match(Op1, m_Add(m_Value(A), m_Constant(C)))) | ||||
400 | if (isKnownNonNegative(A, DL, 0, &AC, &I, &DT) && | ||||
401 | isKnownNonNegative(C, DL, 0, &AC, &I, &DT)) | ||||
402 | return BinaryOperator::Create( | ||||
403 | I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), Op0, C), A); | ||||
404 | |||||
405 | // X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2. | ||||
406 | // Because shifts by negative values (which could occur if A were negative) | ||||
407 | // are undefined. | ||||
408 | if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Constant(C))) && | ||||
409 | match(C, m_Power2())) { | ||||
410 | // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't | ||||
411 | // demand the sign bit (and many others) here?? | ||||
412 | Constant *Mask = ConstantExpr::getSub(C, ConstantInt::get(I.getType(), 1)); | ||||
413 | Value *Rem = Builder.CreateAnd(A, Mask, Op1->getName()); | ||||
414 | return replaceOperand(I, 1, Rem); | ||||
415 | } | ||||
416 | |||||
417 | if (Instruction *Logic = foldShiftOfShiftedLogic(I, Builder)) | ||||
418 | return Logic; | ||||
419 | |||||
420 | return nullptr; | ||||
421 | } | ||||
422 | |||||
423 | /// Return true if we can simplify two logical (either left or right) shifts | ||||
424 | /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2. | ||||
425 | static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl, | ||||
426 | Instruction *InnerShift, | ||||
427 | InstCombinerImpl &IC, Instruction *CxtI) { | ||||
428 | assert(InnerShift->isLogicalShift() && "Unexpected instruction type")((InnerShift->isLogicalShift() && "Unexpected instruction type" ) ? static_cast<void> (0) : __assert_fail ("InnerShift->isLogicalShift() && \"Unexpected instruction type\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 428, __PRETTY_FUNCTION__)); | ||||
429 | |||||
430 | // We need constant scalar or constant splat shifts. | ||||
431 | const APInt *InnerShiftConst; | ||||
432 | if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst))) | ||||
433 | return false; | ||||
434 | |||||
435 | // Two logical shifts in the same direction: | ||||
436 | // shl (shl X, C1), C2 --> shl X, C1 + C2 | ||||
437 | // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 | ||||
438 | bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; | ||||
439 | if (IsInnerShl == IsOuterShl) | ||||
440 | return true; | ||||
441 | |||||
442 | // Equal shift amounts in opposite directions become bitwise 'and': | ||||
443 | // lshr (shl X, C), C --> and X, C' | ||||
444 | // shl (lshr X, C), C --> and X, C' | ||||
445 | if (*InnerShiftConst == OuterShAmt) | ||||
446 | return true; | ||||
447 | |||||
448 | // If the 2nd shift is bigger than the 1st, we can fold: | ||||
449 | // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3 | ||||
450 | // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3 | ||||
451 | // but it isn't profitable unless we know the and'd out bits are already zero. | ||||
452 | // Also, check that the inner shift is valid (less than the type width) or | ||||
453 | // we'll crash trying to produce the bit mask for the 'and'. | ||||
454 | unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits(); | ||||
455 | if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) { | ||||
456 | unsigned InnerShAmt = InnerShiftConst->getZExtValue(); | ||||
457 | unsigned MaskShift = | ||||
458 | IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt; | ||||
459 | APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift; | ||||
460 | if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI)) | ||||
461 | return true; | ||||
462 | } | ||||
463 | |||||
464 | return false; | ||||
465 | } | ||||
466 | |||||
467 | /// See if we can compute the specified value, but shifted logically to the left | ||||
468 | /// or right by some number of bits. This should return true if the expression | ||||
469 | /// can be computed for the same cost as the current expression tree. This is | ||||
470 | /// used to eliminate extraneous shifting from things like: | ||||
471 | /// %C = shl i128 %A, 64 | ||||
472 | /// %D = shl i128 %B, 96 | ||||
473 | /// %E = or i128 %C, %D | ||||
474 | /// %F = lshr i128 %E, 64 | ||||
475 | /// where the client will ask if E can be computed shifted right by 64-bits. If | ||||
476 | /// this succeeds, getShiftedValue() will be called to produce the value. | ||||
477 | static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift, | ||||
478 | InstCombinerImpl &IC, Instruction *CxtI) { | ||||
479 | // We can always evaluate constants shifted. | ||||
480 | if (isa<Constant>(V)) | ||||
481 | return true; | ||||
482 | |||||
483 | Instruction *I = dyn_cast<Instruction>(V); | ||||
484 | if (!I) return false; | ||||
485 | |||||
486 | // We can't mutate something that has multiple uses: doing so would | ||||
487 | // require duplicating the instruction in general, which isn't profitable. | ||||
488 | if (!I->hasOneUse()) return false; | ||||
489 | |||||
490 | switch (I->getOpcode()) { | ||||
491 | default: return false; | ||||
492 | case Instruction::And: | ||||
493 | case Instruction::Or: | ||||
494 | case Instruction::Xor: | ||||
495 | // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. | ||||
496 | return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) && | ||||
497 | canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I); | ||||
498 | |||||
499 | case Instruction::Shl: | ||||
500 | case Instruction::LShr: | ||||
501 | return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI); | ||||
502 | |||||
503 | case Instruction::Select: { | ||||
504 | SelectInst *SI = cast<SelectInst>(I); | ||||
505 | Value *TrueVal = SI->getTrueValue(); | ||||
506 | Value *FalseVal = SI->getFalseValue(); | ||||
507 | return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) && | ||||
508 | canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI); | ||||
509 | } | ||||
510 | case Instruction::PHI: { | ||||
511 | // We can change a phi if we can change all operands. Note that we never | ||||
512 | // get into trouble with cyclic PHIs here because we only consider | ||||
513 | // instructions with a single use. | ||||
514 | PHINode *PN = cast<PHINode>(I); | ||||
515 | for (Value *IncValue : PN->incoming_values()) | ||||
516 | if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN)) | ||||
517 | return false; | ||||
518 | return true; | ||||
519 | } | ||||
520 | } | ||||
521 | } | ||||
522 | |||||
523 | /// Fold OuterShift (InnerShift X, C1), C2. | ||||
524 | /// See canEvaluateShiftedShift() for the constraints on these instructions. | ||||
525 | static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt, | ||||
526 | bool IsOuterShl, | ||||
527 | InstCombiner::BuilderTy &Builder) { | ||||
528 | bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; | ||||
529 | Type *ShType = InnerShift->getType(); | ||||
530 | unsigned TypeWidth = ShType->getScalarSizeInBits(); | ||||
531 | |||||
532 | // We only accept shifts-by-a-constant in canEvaluateShifted(). | ||||
533 | const APInt *C1; | ||||
534 | match(InnerShift->getOperand(1), m_APInt(C1)); | ||||
535 | unsigned InnerShAmt = C1->getZExtValue(); | ||||
536 | |||||
537 | // Change the shift amount and clear the appropriate IR flags. | ||||
538 | auto NewInnerShift = [&](unsigned ShAmt) { | ||||
539 | InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt)); | ||||
540 | if (IsInnerShl) { | ||||
541 | InnerShift->setHasNoUnsignedWrap(false); | ||||
542 | InnerShift->setHasNoSignedWrap(false); | ||||
543 | } else { | ||||
544 | InnerShift->setIsExact(false); | ||||
545 | } | ||||
546 | return InnerShift; | ||||
547 | }; | ||||
548 | |||||
549 | // Two logical shifts in the same direction: | ||||
550 | // shl (shl X, C1), C2 --> shl X, C1 + C2 | ||||
551 | // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 | ||||
552 | if (IsInnerShl == IsOuterShl) { | ||||
553 | // If this is an oversized composite shift, then unsigned shifts get 0. | ||||
554 | if (InnerShAmt + OuterShAmt >= TypeWidth) | ||||
555 | return Constant::getNullValue(ShType); | ||||
556 | |||||
557 | return NewInnerShift(InnerShAmt + OuterShAmt); | ||||
558 | } | ||||
559 | |||||
560 | // Equal shift amounts in opposite directions become bitwise 'and': | ||||
561 | // lshr (shl X, C), C --> and X, C' | ||||
562 | // shl (lshr X, C), C --> and X, C' | ||||
563 | if (InnerShAmt == OuterShAmt) { | ||||
564 | APInt Mask = IsInnerShl | ||||
565 | ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt) | ||||
566 | : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt); | ||||
567 | Value *And = Builder.CreateAnd(InnerShift->getOperand(0), | ||||
568 | ConstantInt::get(ShType, Mask)); | ||||
569 | if (auto *AndI = dyn_cast<Instruction>(And)) { | ||||
570 | AndI->moveBefore(InnerShift); | ||||
571 | AndI->takeName(InnerShift); | ||||
572 | } | ||||
573 | return And; | ||||
574 | } | ||||
575 | |||||
576 | assert(InnerShAmt > OuterShAmt &&((InnerShAmt > OuterShAmt && "Unexpected opposite direction logical shift pair" ) ? static_cast<void> (0) : __assert_fail ("InnerShAmt > OuterShAmt && \"Unexpected opposite direction logical shift pair\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 577, __PRETTY_FUNCTION__)) | ||||
577 | "Unexpected opposite direction logical shift pair")((InnerShAmt > OuterShAmt && "Unexpected opposite direction logical shift pair" ) ? static_cast<void> (0) : __assert_fail ("InnerShAmt > OuterShAmt && \"Unexpected opposite direction logical shift pair\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 577, __PRETTY_FUNCTION__)); | ||||
578 | |||||
579 | // In general, we would need an 'and' for this transform, but | ||||
580 | // canEvaluateShiftedShift() guarantees that the masked-off bits are not used. | ||||
581 | // lshr (shl X, C1), C2 --> shl X, C1 - C2 | ||||
582 | // shl (lshr X, C1), C2 --> lshr X, C1 - C2 | ||||
583 | return NewInnerShift(InnerShAmt - OuterShAmt); | ||||
584 | } | ||||
585 | |||||
586 | /// When canEvaluateShifted() returns true for an expression, this function | ||||
587 | /// inserts the new computation that produces the shifted value. | ||||
588 | static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, | ||||
589 | InstCombinerImpl &IC, const DataLayout &DL) { | ||||
590 | // We can always evaluate constants shifted. | ||||
591 | if (Constant *C = dyn_cast<Constant>(V)) { | ||||
592 | if (isLeftShift) | ||||
593 | return IC.Builder.CreateShl(C, NumBits); | ||||
594 | else | ||||
595 | return IC.Builder.CreateLShr(C, NumBits); | ||||
596 | } | ||||
597 | |||||
598 | Instruction *I = cast<Instruction>(V); | ||||
599 | IC.addToWorklist(I); | ||||
600 | |||||
601 | switch (I->getOpcode()) { | ||||
602 | default: llvm_unreachable("Inconsistency with CanEvaluateShifted")::llvm::llvm_unreachable_internal("Inconsistency with CanEvaluateShifted" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 602); | ||||
603 | case Instruction::And: | ||||
604 | case Instruction::Or: | ||||
605 | case Instruction::Xor: | ||||
606 | // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. | ||||
607 | I->setOperand( | ||||
608 | 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL)); | ||||
609 | I->setOperand( | ||||
610 | 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL)); | ||||
611 | return I; | ||||
612 | |||||
613 | case Instruction::Shl: | ||||
614 | case Instruction::LShr: | ||||
615 | return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift, | ||||
616 | IC.Builder); | ||||
617 | |||||
618 | case Instruction::Select: | ||||
619 | I->setOperand( | ||||
620 | 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL)); | ||||
621 | I->setOperand( | ||||
622 | 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL)); | ||||
623 | return I; | ||||
624 | case Instruction::PHI: { | ||||
625 | // We can change a phi if we can change all operands. Note that we never | ||||
626 | // get into trouble with cyclic PHIs here because we only consider | ||||
627 | // instructions with a single use. | ||||
628 | PHINode *PN = cast<PHINode>(I); | ||||
629 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) | ||||
630 | PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits, | ||||
631 | isLeftShift, IC, DL)); | ||||
632 | return PN; | ||||
633 | } | ||||
634 | } | ||||
635 | } | ||||
636 | |||||
637 | // If this is a bitwise operator or add with a constant RHS we might be able | ||||
638 | // to pull it through a shift. | ||||
639 | static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift, | ||||
640 | BinaryOperator *BO) { | ||||
641 | switch (BO->getOpcode()) { | ||||
642 | default: | ||||
643 | return false; // Do not perform transform! | ||||
644 | case Instruction::Add: | ||||
645 | return Shift.getOpcode() == Instruction::Shl; | ||||
646 | case Instruction::Or: | ||||
647 | case Instruction::And: | ||||
648 | return true; | ||||
649 | case Instruction::Xor: | ||||
650 | // Do not change a 'not' of logical shift because that would create a normal | ||||
651 | // 'xor'. The 'not' is likely better for analysis, SCEV, and codegen. | ||||
652 | return !(Shift.isLogicalShift() && match(BO, m_Not(m_Value()))); | ||||
653 | } | ||||
654 | } | ||||
655 | |||||
656 | Instruction *InstCombinerImpl::FoldShiftByConstant(Value *Op0, Constant *Op1, | ||||
657 | BinaryOperator &I) { | ||||
658 | bool isLeftShift = I.getOpcode() == Instruction::Shl; | ||||
659 | |||||
660 | const APInt *Op1C; | ||||
661 | if (!match(Op1, m_APInt(Op1C))) | ||||
662 | return nullptr; | ||||
663 | |||||
664 | // See if we can propagate this shift into the input, this covers the trivial | ||||
665 | // cast of lshr(shl(x,c1),c2) as well as other more complex cases. | ||||
666 | if (I.getOpcode() != Instruction::AShr && | ||||
667 | canEvaluateShifted(Op0, Op1C->getZExtValue(), isLeftShift, *this, &I)) { | ||||
668 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("instcombine")) { dbgs() << "ICE: GetShiftedValue propagating shift through expression" " to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I << "\n"; } } while (false) | ||||
669 | dbgs() << "ICE: GetShiftedValue propagating shift through expression"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("instcombine")) { dbgs() << "ICE: GetShiftedValue propagating shift through expression" " to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I << "\n"; } } while (false) | ||||
670 | " to eliminate shift:\n IN: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("instcombine")) { dbgs() << "ICE: GetShiftedValue propagating shift through expression" " to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I << "\n"; } } while (false) | ||||
671 | << *Op0 << "\n SH: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("instcombine")) { dbgs() << "ICE: GetShiftedValue propagating shift through expression" " to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I << "\n"; } } while (false); | ||||
672 | |||||
673 | return replaceInstUsesWith( | ||||
674 | I, getShiftedValue(Op0, Op1C->getZExtValue(), isLeftShift, *this, DL)); | ||||
675 | } | ||||
676 | |||||
677 | // See if we can simplify any instructions used by the instruction whose sole | ||||
678 | // purpose is to compute bits we don't care about. | ||||
679 | Type *Ty = I.getType(); | ||||
680 | unsigned TypeBits = Ty->getScalarSizeInBits(); | ||||
681 | assert(!Op1C->uge(TypeBits) &&((!Op1C->uge(TypeBits) && "Shift over the type width should have been removed already" ) ? static_cast<void> (0) : __assert_fail ("!Op1C->uge(TypeBits) && \"Shift over the type width should have been removed already\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 682, __PRETTY_FUNCTION__)) | ||||
682 | "Shift over the type width should have been removed already")((!Op1C->uge(TypeBits) && "Shift over the type width should have been removed already" ) ? static_cast<void> (0) : __assert_fail ("!Op1C->uge(TypeBits) && \"Shift over the type width should have been removed already\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 682, __PRETTY_FUNCTION__)); | ||||
683 | |||||
684 | if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I)) | ||||
685 | return FoldedShift; | ||||
686 | |||||
687 | // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2)) | ||||
688 | if (auto *TI = dyn_cast<TruncInst>(Op0)) { | ||||
689 | // If 'shift2' is an ashr, we would have to get the sign bit into a funny | ||||
690 | // place. Don't try to do this transformation in this case. Also, we | ||||
691 | // require that the input operand is a shift-by-constant so that we have | ||||
692 | // confidence that the shifts will get folded together. We could do this | ||||
693 | // xform in more cases, but it is unlikely to be profitable. | ||||
694 | const APInt *TrShiftAmt; | ||||
695 | if (I.isLogicalShift() && | ||||
696 | match(TI->getOperand(0), m_Shift(m_Value(), m_APInt(TrShiftAmt)))) { | ||||
697 | auto *TrOp = cast<Instruction>(TI->getOperand(0)); | ||||
698 | Type *SrcTy = TrOp->getType(); | ||||
699 | |||||
700 | // Okay, we'll do this xform. Make the shift of shift. | ||||
701 | Constant *ShAmt = ConstantExpr::getZExt(Op1, SrcTy); | ||||
702 | // (shift2 (shift1 & 0x00FF), c2) | ||||
703 | Value *NSh = Builder.CreateBinOp(I.getOpcode(), TrOp, ShAmt, I.getName()); | ||||
704 | |||||
705 | // For logical shifts, the truncation has the effect of making the high | ||||
706 | // part of the register be zeros. Emulate this by inserting an AND to | ||||
707 | // clear the top bits as needed. This 'and' will usually be zapped by | ||||
708 | // other xforms later if dead. | ||||
709 | unsigned SrcSize = SrcTy->getScalarSizeInBits(); | ||||
710 | Constant *MaskV = | ||||
711 | ConstantInt::get(SrcTy, APInt::getLowBitsSet(SrcSize, TypeBits)); | ||||
712 | |||||
713 | // The mask we constructed says what the trunc would do if occurring | ||||
714 | // between the shifts. We want to know the effect *after* the second | ||||
715 | // shift. We know that it is a logical shift by a constant, so adjust the | ||||
716 | // mask as appropriate. | ||||
717 | MaskV = ConstantExpr::get(I.getOpcode(), MaskV, ShAmt); | ||||
718 | // shift1 & 0x00FF | ||||
719 | Value *And = Builder.CreateAnd(NSh, MaskV, TI->getName()); | ||||
720 | // Return the value truncated to the interesting size. | ||||
721 | return new TruncInst(And, Ty); | ||||
722 | } | ||||
723 | } | ||||
724 | |||||
725 | if (Op0->hasOneUse()) { | ||||
726 | if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) { | ||||
727 | // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C) | ||||
728 | Value *V1; | ||||
729 | const APInt *CC; | ||||
730 | switch (Op0BO->getOpcode()) { | ||||
731 | default: break; | ||||
732 | case Instruction::Add: | ||||
733 | case Instruction::And: | ||||
734 | case Instruction::Or: | ||||
735 | case Instruction::Xor: { | ||||
736 | // These operators commute. | ||||
737 | // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C) | ||||
738 | if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() && | ||||
739 | match(Op0BO->getOperand(1), m_Shr(m_Value(V1), | ||||
740 | m_Specific(Op1)))) { | ||||
741 | Value *YS = // (Y << C) | ||||
742 | Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName()); | ||||
743 | // (X + (Y << C)) | ||||
744 | Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), YS, V1, | ||||
745 | Op0BO->getOperand(1)->getName()); | ||||
746 | unsigned Op1Val = Op1C->getLimitedValue(TypeBits); | ||||
747 | APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val); | ||||
748 | Constant *Mask = ConstantInt::get(Ty, Bits); | ||||
749 | return BinaryOperator::CreateAnd(X, Mask); | ||||
750 | } | ||||
751 | |||||
752 | // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C)) | ||||
753 | Value *Op0BOOp1 = Op0BO->getOperand(1); | ||||
754 | if (isLeftShift && Op0BOOp1->hasOneUse() && | ||||
755 | match(Op0BOOp1, m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))), | ||||
756 | m_APInt(CC)))) { | ||||
757 | Value *YS = // (Y << C) | ||||
758 | Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName()); | ||||
759 | // X & (CC << C) | ||||
760 | Value *XM = Builder.CreateAnd( | ||||
761 | V1, ConstantExpr::getShl(ConstantInt::get(Ty, *CC), Op1), | ||||
762 | V1->getName() + ".mask"); | ||||
763 | return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM); | ||||
764 | } | ||||
765 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||
766 | } | ||||
767 | |||||
768 | case Instruction::Sub: { | ||||
769 | // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C) | ||||
770 | if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() && | ||||
771 | match(Op0BO->getOperand(0), m_Shr(m_Value(V1), | ||||
772 | m_Specific(Op1)))) { | ||||
773 | Value *YS = // (Y << C) | ||||
774 | Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName()); | ||||
775 | // (X + (Y << C)) | ||||
776 | Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), V1, YS, | ||||
777 | Op0BO->getOperand(0)->getName()); | ||||
778 | unsigned Op1Val = Op1C->getLimitedValue(TypeBits); | ||||
779 | APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val); | ||||
780 | Constant *Mask = ConstantInt::get(Ty, Bits); | ||||
781 | return BinaryOperator::CreateAnd(X, Mask); | ||||
782 | } | ||||
783 | |||||
784 | // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C) | ||||
785 | if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() && | ||||
786 | match(Op0BO->getOperand(0), | ||||
787 | m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))), | ||||
788 | m_APInt(CC)))) { | ||||
789 | Value *YS = // (Y << C) | ||||
790 | Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName()); | ||||
791 | // X & (CC << C) | ||||
792 | Value *XM = Builder.CreateAnd( | ||||
793 | V1, ConstantExpr::getShl(ConstantInt::get(Ty, *CC), Op1), | ||||
794 | V1->getName() + ".mask"); | ||||
795 | return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS); | ||||
796 | } | ||||
797 | |||||
798 | break; | ||||
799 | } | ||||
800 | } | ||||
801 | |||||
802 | // If the operand is a bitwise operator with a constant RHS, and the | ||||
803 | // shift is the only use, we can pull it out of the shift. | ||||
804 | const APInt *Op0C; | ||||
805 | if (match(Op0BO->getOperand(1), m_APInt(Op0C))) { | ||||
806 | if (canShiftBinOpWithConstantRHS(I, Op0BO)) { | ||||
807 | Constant *NewRHS = ConstantExpr::get(I.getOpcode(), | ||||
808 | cast<Constant>(Op0BO->getOperand(1)), Op1); | ||||
809 | |||||
810 | Value *NewShift = | ||||
811 | Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1); | ||||
812 | NewShift->takeName(Op0BO); | ||||
813 | |||||
814 | return BinaryOperator::Create(Op0BO->getOpcode(), NewShift, | ||||
815 | NewRHS); | ||||
816 | } | ||||
817 | } | ||||
818 | |||||
819 | // If the operand is a subtract with a constant LHS, and the shift | ||||
820 | // is the only use, we can pull it out of the shift. | ||||
821 | // This folds (shl (sub C1, X), C2) -> (sub (C1 << C2), (shl X, C2)) | ||||
822 | if (isLeftShift && Op0BO->getOpcode() == Instruction::Sub && | ||||
823 | match(Op0BO->getOperand(0), m_APInt(Op0C))) { | ||||
824 | Constant *NewRHS = ConstantExpr::get(I.getOpcode(), | ||||
825 | cast<Constant>(Op0BO->getOperand(0)), Op1); | ||||
826 | |||||
827 | Value *NewShift = Builder.CreateShl(Op0BO->getOperand(1), Op1); | ||||
828 | NewShift->takeName(Op0BO); | ||||
829 | |||||
830 | return BinaryOperator::CreateSub(NewRHS, NewShift); | ||||
831 | } | ||||
832 | } | ||||
833 | |||||
834 | // If we have a select that conditionally executes some binary operator, | ||||
835 | // see if we can pull it the select and operator through the shift. | ||||
836 | // | ||||
837 | // For example, turning: | ||||
838 | // shl (select C, (add X, C1), X), C2 | ||||
839 | // Into: | ||||
840 | // Y = shl X, C2 | ||||
841 | // select C, (add Y, C1 << C2), Y | ||||
842 | Value *Cond; | ||||
843 | BinaryOperator *TBO; | ||||
844 | Value *FalseVal; | ||||
845 | if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)), | ||||
846 | m_Value(FalseVal)))) { | ||||
847 | const APInt *C; | ||||
848 | if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal && | ||||
849 | match(TBO->getOperand(1), m_APInt(C)) && | ||||
850 | canShiftBinOpWithConstantRHS(I, TBO)) { | ||||
851 | Constant *NewRHS = ConstantExpr::get(I.getOpcode(), | ||||
852 | cast<Constant>(TBO->getOperand(1)), Op1); | ||||
853 | |||||
854 | Value *NewShift = | ||||
855 | Builder.CreateBinOp(I.getOpcode(), FalseVal, Op1); | ||||
856 | Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift, | ||||
857 | NewRHS); | ||||
858 | return SelectInst::Create(Cond, NewOp, NewShift); | ||||
859 | } | ||||
860 | } | ||||
861 | |||||
862 | BinaryOperator *FBO; | ||||
863 | Value *TrueVal; | ||||
864 | if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal), | ||||
865 | m_OneUse(m_BinOp(FBO))))) { | ||||
866 | const APInt *C; | ||||
867 | if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal && | ||||
868 | match(FBO->getOperand(1), m_APInt(C)) && | ||||
869 | canShiftBinOpWithConstantRHS(I, FBO)) { | ||||
870 | Constant *NewRHS = ConstantExpr::get(I.getOpcode(), | ||||
871 | cast<Constant>(FBO->getOperand(1)), Op1); | ||||
872 | |||||
873 | Value *NewShift = | ||||
874 | Builder.CreateBinOp(I.getOpcode(), TrueVal, Op1); | ||||
875 | Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift, | ||||
876 | NewRHS); | ||||
877 | return SelectInst::Create(Cond, NewShift, NewOp); | ||||
878 | } | ||||
879 | } | ||||
880 | } | ||||
881 | |||||
882 | return nullptr; | ||||
883 | } | ||||
884 | |||||
885 | Instruction *InstCombinerImpl::visitShl(BinaryOperator &I) { | ||||
886 | const SimplifyQuery Q = SQ.getWithInstruction(&I); | ||||
887 | |||||
888 | if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1), | ||||
| |||||
889 | I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), Q)) | ||||
890 | return replaceInstUsesWith(I, V); | ||||
891 | |||||
892 | if (Instruction *X = foldVectorBinop(I)) | ||||
893 | return X; | ||||
894 | |||||
895 | if (Instruction *V = commonShiftTransforms(I)) | ||||
896 | return V; | ||||
897 | |||||
898 | if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(&I, Q, Builder)) | ||||
899 | return V; | ||||
900 | |||||
901 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
902 | Type *Ty = I.getType(); | ||||
903 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
904 | |||||
905 | const APInt *ShAmtAPInt; | ||||
906 | if (match(Op1, m_APInt(ShAmtAPInt))) { | ||||
907 | unsigned ShAmt = ShAmtAPInt->getZExtValue(); | ||||
908 | |||||
909 | // shl (zext X), ShAmt --> zext (shl X, ShAmt) | ||||
910 | // This is only valid if X would have zeros shifted out. | ||||
911 | Value *X; | ||||
912 | if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) { | ||||
913 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); | ||||
914 | if (ShAmt < SrcWidth && | ||||
915 | MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmt), 0, &I)) | ||||
916 | return new ZExtInst(Builder.CreateShl(X, ShAmt), Ty); | ||||
917 | } | ||||
918 | |||||
919 | // (X >> C) << C --> X & (-1 << C) | ||||
920 | if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) { | ||||
921 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)); | ||||
922 | return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask)); | ||||
923 | } | ||||
924 | |||||
925 | const APInt *ShOp1; | ||||
926 | if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(ShOp1)))) && | ||||
927 | ShOp1->ult(BitWidth)) { | ||||
928 | unsigned ShrAmt = ShOp1->getZExtValue(); | ||||
929 | if (ShrAmt < ShAmt) { | ||||
930 | // If C1 < C2: (X >>?,exact C1) << C2 --> X << (C2 - C1) | ||||
931 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt); | ||||
932 | auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); | ||||
933 | NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); | ||||
934 | NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); | ||||
935 | return NewShl; | ||||
936 | } | ||||
937 | if (ShrAmt > ShAmt) { | ||||
938 | // If C1 > C2: (X >>?exact C1) << C2 --> X >>?exact (C1 - C2) | ||||
939 | Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt); | ||||
940 | auto *NewShr = BinaryOperator::Create( | ||||
941 | cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff); | ||||
942 | NewShr->setIsExact(true); | ||||
943 | return NewShr; | ||||
944 | } | ||||
945 | } | ||||
946 | |||||
947 | if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_APInt(ShOp1)))) && | ||||
948 | ShOp1->ult(BitWidth)) { | ||||
949 | unsigned ShrAmt = ShOp1->getZExtValue(); | ||||
950 | if (ShrAmt < ShAmt) { | ||||
951 | // If C1 < C2: (X >>? C1) << C2 --> X << (C2 - C1) & (-1 << C2) | ||||
952 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt); | ||||
953 | auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); | ||||
954 | NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); | ||||
955 | NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); | ||||
956 | Builder.Insert(NewShl); | ||||
957 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)); | ||||
958 | return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask)); | ||||
959 | } | ||||
960 | if (ShrAmt > ShAmt) { | ||||
961 | // If C1 > C2: (X >>? C1) << C2 --> X >>? (C1 - C2) & (-1 << C2) | ||||
962 | Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt); | ||||
963 | auto *OldShr = cast<BinaryOperator>(Op0); | ||||
964 | auto *NewShr = | ||||
965 | BinaryOperator::Create(OldShr->getOpcode(), X, ShiftDiff); | ||||
966 | NewShr->setIsExact(OldShr->isExact()); | ||||
967 | Builder.Insert(NewShr); | ||||
968 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)); | ||||
969 | return BinaryOperator::CreateAnd(NewShr, ConstantInt::get(Ty, Mask)); | ||||
970 | } | ||||
971 | } | ||||
972 | |||||
973 | if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1))) && ShOp1->ult(BitWidth)) { | ||||
974 | unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); | ||||
975 | // Oversized shifts are simplified to zero in InstSimplify. | ||||
976 | if (AmtSum < BitWidth) | ||||
977 | // (X << C1) << C2 --> X << (C1 + C2) | ||||
978 | return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum)); | ||||
979 | } | ||||
980 | |||||
981 | // If the shifted-out value is known-zero, then this is a NUW shift. | ||||
982 | if (!I.hasNoUnsignedWrap() && | ||||
983 | MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmt), 0, &I)) { | ||||
984 | I.setHasNoUnsignedWrap(); | ||||
985 | return &I; | ||||
986 | } | ||||
987 | |||||
988 | // If the shifted-out value is all signbits, then this is a NSW shift. | ||||
989 | if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmt) { | ||||
990 | I.setHasNoSignedWrap(); | ||||
991 | return &I; | ||||
992 | } | ||||
993 | } | ||||
994 | |||||
995 | // Transform (x >> y) << y to x & (-1 << y) | ||||
996 | // Valid for any type of right-shift. | ||||
997 | Value *X; | ||||
998 | if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) { | ||||
999 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); | ||||
1000 | Value *Mask = Builder.CreateShl(AllOnes, Op1); | ||||
1001 | return BinaryOperator::CreateAnd(Mask, X); | ||||
1002 | } | ||||
1003 | |||||
1004 | Constant *C1; | ||||
1005 | if (match(Op1, m_Constant(C1))) { | ||||
1006 | Constant *C2; | ||||
1007 | Value *X; | ||||
1008 | // (C2 << X) << C1 --> (C2 << C1) << X | ||||
1009 | if (match(Op0, m_OneUse(m_Shl(m_Constant(C2), m_Value(X))))) | ||||
1010 | return BinaryOperator::CreateShl(ConstantExpr::getShl(C2, C1), X); | ||||
1011 | |||||
1012 | // (X * C2) << C1 --> X * (C2 << C1) | ||||
1013 | if (match(Op0, m_Mul(m_Value(X), m_Constant(C2)))) | ||||
1014 | return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1)); | ||||
1015 | |||||
1016 | // shl (zext i1 X), C1 --> select (X, 1 << C1, 0) | ||||
1017 | if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { | ||||
1018 | auto *NewC = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C1); | ||||
1019 | return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty)); | ||||
1020 | } | ||||
1021 | } | ||||
1022 | |||||
1023 | // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1 | ||||
1024 | if (match(Op0, m_One()) && | ||||
1025 | match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X)))) | ||||
1026 | return BinaryOperator::CreateLShr( | ||||
1027 | ConstantInt::get(Ty, APInt::getSignMask(BitWidth)), X); | ||||
1028 | |||||
1029 | return nullptr; | ||||
1030 | } | ||||
1031 | |||||
1032 | Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) { | ||||
1033 | if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(), | ||||
1034 | SQ.getWithInstruction(&I))) | ||||
1035 | return replaceInstUsesWith(I, V); | ||||
1036 | |||||
1037 | if (Instruction *X = foldVectorBinop(I)) | ||||
1038 | return X; | ||||
1039 | |||||
1040 | if (Instruction *R = commonShiftTransforms(I)) | ||||
1041 | return R; | ||||
1042 | |||||
1043 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
1044 | Type *Ty = I.getType(); | ||||
1045 | const APInt *ShAmtAPInt; | ||||
1046 | if (match(Op1, m_APInt(ShAmtAPInt))) { | ||||
1047 | unsigned ShAmt = ShAmtAPInt->getZExtValue(); | ||||
1048 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
1049 | auto *II = dyn_cast<IntrinsicInst>(Op0); | ||||
1050 | if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt && | ||||
1051 | (II->getIntrinsicID() == Intrinsic::ctlz || | ||||
1052 | II->getIntrinsicID() == Intrinsic::cttz || | ||||
1053 | II->getIntrinsicID() == Intrinsic::ctpop)) { | ||||
1054 | // ctlz.i32(x)>>5 --> zext(x == 0) | ||||
1055 | // cttz.i32(x)>>5 --> zext(x == 0) | ||||
1056 | // ctpop.i32(x)>>5 --> zext(x == -1) | ||||
1057 | bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop; | ||||
1058 | Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0); | ||||
1059 | Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS); | ||||
1060 | return new ZExtInst(Cmp, Ty); | ||||
1061 | } | ||||
1062 | |||||
1063 | Value *X; | ||||
1064 | const APInt *ShOp1; | ||||
1065 | if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1))) && ShOp1->ult(BitWidth)) { | ||||
1066 | if (ShOp1->ult(ShAmt)) { | ||||
1067 | unsigned ShlAmt = ShOp1->getZExtValue(); | ||||
1068 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt); | ||||
1069 | if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) { | ||||
1070 | // (X <<nuw C1) >>u C2 --> X >>u (C2 - C1) | ||||
1071 | auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff); | ||||
1072 | NewLShr->setIsExact(I.isExact()); | ||||
1073 | return NewLShr; | ||||
1074 | } | ||||
1075 | // (X << C1) >>u C2 --> (X >>u (C2 - C1)) & (-1 >> C2) | ||||
1076 | Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact()); | ||||
1077 | APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt)); | ||||
1078 | return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask)); | ||||
1079 | } | ||||
1080 | if (ShOp1->ugt(ShAmt)) { | ||||
1081 | unsigned ShlAmt = ShOp1->getZExtValue(); | ||||
1082 | Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt); | ||||
1083 | if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) { | ||||
1084 | // (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2) | ||||
1085 | auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); | ||||
1086 | NewShl->setHasNoUnsignedWrap(true); | ||||
1087 | return NewShl; | ||||
1088 | } | ||||
1089 | // (X << C1) >>u C2 --> X << (C1 - C2) & (-1 >> C2) | ||||
1090 | Value *NewShl = Builder.CreateShl(X, ShiftDiff); | ||||
1091 | APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt)); | ||||
1092 | return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask)); | ||||
1093 | } | ||||
1094 | assert(*ShOp1 == ShAmt)((*ShOp1 == ShAmt) ? static_cast<void> (0) : __assert_fail ("*ShOp1 == ShAmt", "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 1094, __PRETTY_FUNCTION__)); | ||||
1095 | // (X << C) >>u C --> X & (-1 >>u C) | ||||
1096 | APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt)); | ||||
1097 | return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask)); | ||||
1098 | } | ||||
1099 | |||||
1100 | if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) && | ||||
1101 | (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) { | ||||
1102 | assert(ShAmt < X->getType()->getScalarSizeInBits() &&((ShAmt < X->getType()->getScalarSizeInBits() && "Big shift not simplified to zero?") ? static_cast<void> (0) : __assert_fail ("ShAmt < X->getType()->getScalarSizeInBits() && \"Big shift not simplified to zero?\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 1103, __PRETTY_FUNCTION__)) | ||||
1103 | "Big shift not simplified to zero?")((ShAmt < X->getType()->getScalarSizeInBits() && "Big shift not simplified to zero?") ? static_cast<void> (0) : __assert_fail ("ShAmt < X->getType()->getScalarSizeInBits() && \"Big shift not simplified to zero?\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 1103, __PRETTY_FUNCTION__)); | ||||
1104 | // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN | ||||
1105 | Value *NewLShr = Builder.CreateLShr(X, ShAmt); | ||||
1106 | return new ZExtInst(NewLShr, Ty); | ||||
1107 | } | ||||
1108 | |||||
1109 | if (match(Op0, m_SExt(m_Value(X))) && | ||||
1110 | (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) { | ||||
1111 | // Are we moving the sign bit to the low bit and widening with high zeros? | ||||
1112 | unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits(); | ||||
1113 | if (ShAmt == BitWidth - 1) { | ||||
1114 | // lshr (sext i1 X to iN), N-1 --> zext X to iN | ||||
1115 | if (SrcTyBitWidth == 1) | ||||
1116 | return new ZExtInst(X, Ty); | ||||
1117 | |||||
1118 | // lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN | ||||
1119 | if (Op0->hasOneUse()) { | ||||
1120 | Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1); | ||||
1121 | return new ZExtInst(NewLShr, Ty); | ||||
1122 | } | ||||
1123 | } | ||||
1124 | |||||
1125 | // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN | ||||
1126 | if (ShAmt == BitWidth - SrcTyBitWidth && Op0->hasOneUse()) { | ||||
1127 | // The new shift amount can't be more than the narrow source type. | ||||
1128 | unsigned NewShAmt = std::min(ShAmt, SrcTyBitWidth - 1); | ||||
1129 | Value *AShr = Builder.CreateAShr(X, NewShAmt); | ||||
1130 | return new ZExtInst(AShr, Ty); | ||||
1131 | } | ||||
1132 | } | ||||
1133 | |||||
1134 | // lshr i32 (X -nsw Y), 31 --> zext (X < Y) | ||||
1135 | Value *Y; | ||||
1136 | if (ShAmt == BitWidth - 1 && | ||||
1137 | match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y))))) | ||||
1138 | return new ZExtInst(Builder.CreateICmpSLT(X, Y), Ty); | ||||
1139 | |||||
1140 | if (match(Op0, m_LShr(m_Value(X), m_APInt(ShOp1)))) { | ||||
1141 | unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); | ||||
1142 | // Oversized shifts are simplified to zero in InstSimplify. | ||||
1143 | if (AmtSum < BitWidth) | ||||
1144 | // (X >>u C1) >>u C2 --> X >>u (C1 + C2) | ||||
1145 | return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum)); | ||||
1146 | } | ||||
1147 | |||||
1148 | // Look for a "splat" mul pattern - it replicates bits across each half of | ||||
1149 | // a value, so a right shift is just a mask of the low bits: | ||||
1150 | // lshr i32 (mul nuw X, Pow2+1), 16 --> and X, Pow2-1 | ||||
1151 | // TODO: Generalize to allow more than just half-width shifts? | ||||
1152 | const APInt *MulC; | ||||
1153 | if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC))) && | ||||
1154 | ShAmt * 2 == BitWidth && (*MulC - 1).isPowerOf2() && | ||||
1155 | MulC->logBase2() == ShAmt) | ||||
1156 | return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2)); | ||||
1157 | |||||
1158 | // If the shifted-out value is known-zero, then this is an exact shift. | ||||
1159 | if (!I.isExact() && | ||||
1160 | MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) { | ||||
1161 | I.setIsExact(); | ||||
1162 | return &I; | ||||
1163 | } | ||||
1164 | } | ||||
1165 | |||||
1166 | // Transform (x << y) >> y to x & (-1 >> y) | ||||
1167 | Value *X; | ||||
1168 | if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) { | ||||
1169 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); | ||||
1170 | Value *Mask = Builder.CreateLShr(AllOnes, Op1); | ||||
1171 | return BinaryOperator::CreateAnd(Mask, X); | ||||
1172 | } | ||||
1173 | |||||
1174 | return nullptr; | ||||
1175 | } | ||||
1176 | |||||
1177 | Instruction * | ||||
1178 | InstCombinerImpl::foldVariableSignZeroExtensionOfVariableHighBitExtract( | ||||
1179 | BinaryOperator &OldAShr) { | ||||
1180 | assert(OldAShr.getOpcode() == Instruction::AShr &&((OldAShr.getOpcode() == Instruction::AShr && "Must be called with arithmetic right-shift instruction only." ) ? static_cast<void> (0) : __assert_fail ("OldAShr.getOpcode() == Instruction::AShr && \"Must be called with arithmetic right-shift instruction only.\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 1181, __PRETTY_FUNCTION__)) | ||||
1181 | "Must be called with arithmetic right-shift instruction only.")((OldAShr.getOpcode() == Instruction::AShr && "Must be called with arithmetic right-shift instruction only." ) ? static_cast<void> (0) : __assert_fail ("OldAShr.getOpcode() == Instruction::AShr && \"Must be called with arithmetic right-shift instruction only.\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 1181, __PRETTY_FUNCTION__)); | ||||
1182 | |||||
1183 | // Check that constant C is a splat of the element-wise bitwidth of V. | ||||
1184 | auto BitWidthSplat = [](Constant *C, Value *V) { | ||||
1185 | return match( | ||||
1186 | C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, | ||||
1187 | APInt(C->getType()->getScalarSizeInBits(), | ||||
1188 | V->getType()->getScalarSizeInBits()))); | ||||
1189 | }; | ||||
1190 | |||||
1191 | // It should look like variable-length sign-extension on the outside: | ||||
1192 | // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits) | ||||
1193 | Value *NBits; | ||||
1194 | Instruction *MaybeTrunc; | ||||
1195 | Constant *C1, *C2; | ||||
1196 | if (!match(&OldAShr, | ||||
1197 | m_AShr(m_Shl(m_Instruction(MaybeTrunc), | ||||
1198 | m_ZExtOrSelf(m_Sub(m_Constant(C1), | ||||
1199 | m_ZExtOrSelf(m_Value(NBits))))), | ||||
1200 | m_ZExtOrSelf(m_Sub(m_Constant(C2), | ||||
1201 | m_ZExtOrSelf(m_Deferred(NBits)))))) || | ||||
1202 | !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr)) | ||||
1203 | return nullptr; | ||||
1204 | |||||
1205 | // There may or may not be a truncation after outer two shifts. | ||||
1206 | Instruction *HighBitExtract; | ||||
1207 | match(MaybeTrunc, m_TruncOrSelf(m_Instruction(HighBitExtract))); | ||||
1208 | bool HadTrunc = MaybeTrunc != HighBitExtract; | ||||
1209 | |||||
1210 | // And finally, the innermost part of the pattern must be a right-shift. | ||||
1211 | Value *X, *NumLowBitsToSkip; | ||||
1212 | if (!match(HighBitExtract, m_Shr(m_Value(X), m_Value(NumLowBitsToSkip)))) | ||||
1213 | return nullptr; | ||||
1214 | |||||
1215 | // Said right-shift must extract high NBits bits - C0 must be it's bitwidth. | ||||
1216 | Constant *C0; | ||||
1217 | if (!match(NumLowBitsToSkip, | ||||
1218 | m_ZExtOrSelf( | ||||
1219 | m_Sub(m_Constant(C0), m_ZExtOrSelf(m_Specific(NBits))))) || | ||||
1220 | !BitWidthSplat(C0, HighBitExtract)) | ||||
1221 | return nullptr; | ||||
1222 | |||||
1223 | // Since the NBits is identical for all shifts, if the outermost and | ||||
1224 | // innermost shifts are identical, then outermost shifts are redundant. | ||||
1225 | // If we had truncation, do keep it though. | ||||
1226 | if (HighBitExtract->getOpcode() == OldAShr.getOpcode()) | ||||
1227 | return replaceInstUsesWith(OldAShr, MaybeTrunc); | ||||
1228 | |||||
1229 | // Else, if there was a truncation, then we need to ensure that one | ||||
1230 | // instruction will go away. | ||||
1231 | if (HadTrunc && !match(&OldAShr, m_c_BinOp(m_OneUse(m_Value()), m_Value()))) | ||||
1232 | return nullptr; | ||||
1233 | |||||
1234 | // Finally, bypass two innermost shifts, and perform the outermost shift on | ||||
1235 | // the operands of the innermost shift. | ||||
1236 | Instruction *NewAShr = | ||||
1237 | BinaryOperator::Create(OldAShr.getOpcode(), X, NumLowBitsToSkip); | ||||
1238 | NewAShr->copyIRFlags(HighBitExtract); // We can preserve 'exact'-ness. | ||||
1239 | if (!HadTrunc) | ||||
1240 | return NewAShr; | ||||
1241 | |||||
1242 | Builder.Insert(NewAShr); | ||||
1243 | return TruncInst::CreateTruncOrBitCast(NewAShr, OldAShr.getType()); | ||||
1244 | } | ||||
1245 | |||||
1246 | Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) { | ||||
1247 | if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(), | ||||
1248 | SQ.getWithInstruction(&I))) | ||||
1249 | return replaceInstUsesWith(I, V); | ||||
1250 | |||||
1251 | if (Instruction *X = foldVectorBinop(I)) | ||||
1252 | return X; | ||||
1253 | |||||
1254 | if (Instruction *R = commonShiftTransforms(I)) | ||||
1255 | return R; | ||||
1256 | |||||
1257 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
1258 | Type *Ty = I.getType(); | ||||
1259 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
1260 | const APInt *ShAmtAPInt; | ||||
1261 | if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) { | ||||
1262 | unsigned ShAmt = ShAmtAPInt->getZExtValue(); | ||||
1263 | |||||
1264 | // If the shift amount equals the difference in width of the destination | ||||
1265 | // and source scalar types: | ||||
1266 | // ashr (shl (zext X), C), C --> sext X | ||||
1267 | Value *X; | ||||
1268 | if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) && | ||||
1269 | ShAmt == BitWidth - X->getType()->getScalarSizeInBits()) | ||||
1270 | return new SExtInst(X, Ty); | ||||
1271 | |||||
1272 | // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However, | ||||
1273 | // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits. | ||||
1274 | const APInt *ShOp1; | ||||
1275 | if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) && | ||||
1276 | ShOp1->ult(BitWidth)) { | ||||
1277 | unsigned ShlAmt = ShOp1->getZExtValue(); | ||||
1278 | if (ShlAmt < ShAmt) { | ||||
1279 | // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1) | ||||
1280 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt); | ||||
1281 | auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff); | ||||
1282 | NewAShr->setIsExact(I.isExact()); | ||||
1283 | return NewAShr; | ||||
1284 | } | ||||
1285 | if (ShlAmt > ShAmt) { | ||||
1286 | // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2) | ||||
1287 | Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt); | ||||
1288 | auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff); | ||||
1289 | NewShl->setHasNoSignedWrap(true); | ||||
1290 | return NewShl; | ||||
1291 | } | ||||
1292 | } | ||||
1293 | |||||
1294 | if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) && | ||||
1295 | ShOp1->ult(BitWidth)) { | ||||
1296 | unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); | ||||
1297 | // Oversized arithmetic shifts replicate the sign bit. | ||||
1298 | AmtSum = std::min(AmtSum, BitWidth - 1); | ||||
1299 | // (X >>s C1) >>s C2 --> X >>s (C1 + C2) | ||||
1300 | return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum)); | ||||
1301 | } | ||||
1302 | |||||
1303 | if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) && | ||||
1304 | (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) { | ||||
1305 | // ashr (sext X), C --> sext (ashr X, C') | ||||
1306 | Type *SrcTy = X->getType(); | ||||
1307 | ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1); | ||||
1308 | Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt)); | ||||
1309 | return new SExtInst(NewSh, Ty); | ||||
1310 | } | ||||
1311 | |||||
1312 | // ashr i32 (X -nsw Y), 31 --> sext (X < Y) | ||||
1313 | Value *Y; | ||||
1314 | if (ShAmt == BitWidth - 1 && | ||||
1315 | match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y))))) | ||||
1316 | return new SExtInst(Builder.CreateICmpSLT(X, Y), Ty); | ||||
1317 | |||||
1318 | // If the shifted-out value is known-zero, then this is an exact shift. | ||||
1319 | if (!I.isExact() && | ||||
1320 | MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) { | ||||
1321 | I.setIsExact(); | ||||
1322 | return &I; | ||||
1323 | } | ||||
1324 | } | ||||
1325 | |||||
1326 | if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(I)) | ||||
1327 | return R; | ||||
1328 | |||||
1329 | // See if we can turn a signed shr into an unsigned shr. | ||||
1330 | if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I)) | ||||
1331 | return BinaryOperator::CreateLShr(Op0, Op1); | ||||
1332 | |||||
1333 | // ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1 | ||||
1334 | Value *X; | ||||
1335 | if (match(Op0, m_OneUse(m_Not(m_Value(X))))) { | ||||
1336 | // Note that we must drop 'exact'-ness of the shift! | ||||
1337 | // Note that we can't keep undef's in -1 vector constant! | ||||
1338 | auto *NewAShr = Builder.CreateAShr(X, Op1, Op0->getName() + ".not"); | ||||
1339 | return BinaryOperator::CreateNot(NewAShr); | ||||
1340 | } | ||||
1341 | |||||
1342 | return nullptr; | ||||
1343 | } |
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 | /// Match an arbitrary undef constant. |
92 | inline class_match<UndefValue> m_Undef() { return class_match<UndefValue>(); } |
93 | |
94 | /// Match an arbitrary poison constant. |
95 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } |
96 | |
97 | /// Match an arbitrary Constant and ignore it. |
98 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } |
99 | |
100 | /// Match an arbitrary ConstantInt and ignore it. |
101 | inline class_match<ConstantInt> m_ConstantInt() { |
102 | return class_match<ConstantInt>(); |
103 | } |
104 | |
105 | /// Match an arbitrary ConstantFP and ignore it. |
106 | inline class_match<ConstantFP> m_ConstantFP() { |
107 | return class_match<ConstantFP>(); |
108 | } |
109 | |
110 | /// Match an arbitrary ConstantExpr and ignore it. |
111 | inline class_match<ConstantExpr> m_ConstantExpr() { |
112 | return class_match<ConstantExpr>(); |
113 | } |
114 | |
115 | /// Match an arbitrary basic block value and ignore it. |
116 | inline class_match<BasicBlock> m_BasicBlock() { |
117 | return class_match<BasicBlock>(); |
118 | } |
119 | |
120 | /// Inverting matcher |
121 | template <typename Ty> struct match_unless { |
122 | Ty M; |
123 | |
124 | match_unless(const Ty &Matcher) : M(Matcher) {} |
125 | |
126 | template <typename ITy> bool match(ITy *V) { return !M.match(V); } |
127 | }; |
128 | |
129 | /// Match if the inner matcher does *NOT* match. |
130 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { |
131 | return match_unless<Ty>(M); |
132 | } |
133 | |
134 | /// Matching combinators |
135 | template <typename LTy, typename RTy> struct match_combine_or { |
136 | LTy L; |
137 | RTy R; |
138 | |
139 | match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
140 | |
141 | template <typename ITy> bool match(ITy *V) { |
142 | if (L.match(V)) |
143 | return true; |
144 | if (R.match(V)) |
145 | return true; |
146 | return false; |
147 | } |
148 | }; |
149 | |
150 | template <typename LTy, typename RTy> struct match_combine_and { |
151 | LTy L; |
152 | RTy R; |
153 | |
154 | match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
155 | |
156 | template <typename ITy> bool match(ITy *V) { |
157 | if (L.match(V)) |
158 | if (R.match(V)) |
159 | return true; |
160 | return false; |
161 | } |
162 | }; |
163 | |
164 | /// Combine two pattern matchers matching L || R |
165 | template <typename LTy, typename RTy> |
166 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { |
167 | return match_combine_or<LTy, RTy>(L, R); |
168 | } |
169 | |
170 | /// Combine two pattern matchers matching L && R |
171 | template <typename LTy, typename RTy> |
172 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { |
173 | return match_combine_and<LTy, RTy>(L, R); |
174 | } |
175 | |
176 | struct apint_match { |
177 | const APInt *&Res; |
178 | bool AllowUndef; |
179 | |
180 | apint_match(const APInt *&Res, bool AllowUndef) |
181 | : Res(Res), AllowUndef(AllowUndef) {} |
182 | |
183 | template <typename ITy> bool match(ITy *V) { |
184 | if (auto *CI = dyn_cast<ConstantInt>(V)) { |
185 | Res = &CI->getValue(); |
186 | return true; |
187 | } |
188 | if (V->getType()->isVectorTy()) |
189 | if (const auto *C = dyn_cast<Constant>(V)) |
190 | if (auto *CI = dyn_cast_or_null<ConstantInt>( |
191 | C->getSplatValue(AllowUndef))) { |
192 | Res = &CI->getValue(); |
193 | return true; |
194 | } |
195 | return false; |
196 | } |
197 | }; |
198 | // Either constexpr if or renaming ConstantFP::getValueAPF to |
199 | // ConstantFP::getValue is needed to do it via single template |
200 | // function for both apint/apfloat. |
201 | struct apfloat_match { |
202 | const APFloat *&Res; |
203 | bool AllowUndef; |
204 | |
205 | apfloat_match(const APFloat *&Res, bool AllowUndef) |
206 | : Res(Res), AllowUndef(AllowUndef) {} |
207 | |
208 | template <typename ITy> bool match(ITy *V) { |
209 | if (auto *CI = dyn_cast<ConstantFP>(V)) { |
210 | Res = &CI->getValueAPF(); |
211 | return true; |
212 | } |
213 | if (V->getType()->isVectorTy()) |
214 | if (const auto *C = dyn_cast<Constant>(V)) |
215 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
216 | C->getSplatValue(AllowUndef))) { |
217 | Res = &CI->getValueAPF(); |
218 | return true; |
219 | } |
220 | return false; |
221 | } |
222 | }; |
223 | |
224 | /// Match a ConstantInt or splatted ConstantVector, binding the |
225 | /// specified pointer to the contained APInt. |
226 | inline apint_match m_APInt(const APInt *&Res) { |
227 | // Forbid undefs by default to maintain previous behavior. |
228 | return apint_match(Res, /* AllowUndef */ false); |
229 | } |
230 | |
231 | /// Match APInt while allowing undefs in splat vector constants. |
232 | inline apint_match m_APIntAllowUndef(const APInt *&Res) { |
233 | return apint_match(Res, /* AllowUndef */ true); |
234 | } |
235 | |
236 | /// Match APInt while forbidding undefs in splat vector constants. |
237 | inline apint_match m_APIntForbidUndef(const APInt *&Res) { |
238 | return apint_match(Res, /* AllowUndef */ false); |
239 | } |
240 | |
241 | /// Match a ConstantFP or splatted ConstantVector, binding the |
242 | /// specified pointer to the contained APFloat. |
243 | inline apfloat_match m_APFloat(const APFloat *&Res) { |
244 | // Forbid undefs by default to maintain previous behavior. |
245 | return apfloat_match(Res, /* AllowUndef */ false); |
246 | } |
247 | |
248 | /// Match APFloat while allowing undefs in splat vector constants. |
249 | inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) { |
250 | return apfloat_match(Res, /* AllowUndef */ true); |
251 | } |
252 | |
253 | /// Match APFloat while forbidding undefs in splat vector constants. |
254 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { |
255 | return apfloat_match(Res, /* AllowUndef */ false); |
256 | } |
257 | |
258 | template <int64_t Val> struct constantint_match { |
259 | template <typename ITy> bool match(ITy *V) { |
260 | if (const auto *CI = dyn_cast<ConstantInt>(V)) { |
261 | const APInt &CIV = CI->getValue(); |
262 | if (Val >= 0) |
263 | return CIV == static_cast<uint64_t>(Val); |
264 | // If Val is negative, and CI is shorter than it, truncate to the right |
265 | // number of bits. If it is larger, then we have to sign extend. Just |
266 | // compare their negated values. |
267 | return -CIV == -Val; |
268 | } |
269 | return false; |
270 | } |
271 | }; |
272 | |
273 | /// Match a ConstantInt with a specific value. |
274 | template <int64_t Val> inline constantint_match<Val> m_ConstantInt() { |
275 | return constantint_match<Val>(); |
276 | } |
277 | |
278 | /// This helper class is used to match constant scalars, vector splats, |
279 | /// and fixed width vectors that satisfy a specified predicate. |
280 | /// For fixed width vector constants, undefined elements are ignored. |
281 | template <typename Predicate, typename ConstantVal> |
282 | struct cstval_pred_ty : public Predicate { |
283 | template <typename ITy> bool match(ITy *V) { |
284 | if (const auto *CV = dyn_cast<ConstantVal>(V)) |
285 | return this->isValue(CV->getValue()); |
286 | if (const auto *VTy = dyn_cast<VectorType>(V->getType())) { |
287 | if (const auto *C = dyn_cast<Constant>(V)) { |
288 | if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue())) |
289 | return this->isValue(CV->getValue()); |
290 | |
291 | // Number of elements of a scalable vector unknown at compile time |
292 | auto *FVTy = dyn_cast<FixedVectorType>(VTy); |
293 | if (!FVTy) |
294 | return false; |
295 | |
296 | // Non-splat vector constant: check each element for a match. |
297 | unsigned NumElts = FVTy->getNumElements(); |
298 | assert(NumElts != 0 && "Constant vector with no elements?")((NumElts != 0 && "Constant vector with no elements?" ) ? static_cast<void> (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\"" , "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/include/llvm/IR/PatternMatch.h" , 298, __PRETTY_FUNCTION__)); |
299 | bool HasNonUndefElements = false; |
300 | for (unsigned i = 0; i != NumElts; ++i) { |
301 | Constant *Elt = C->getAggregateElement(i); |
302 | if (!Elt) |
303 | return false; |
304 | if (isa<UndefValue>(Elt)) |
305 | continue; |
306 | auto *CV = dyn_cast<ConstantVal>(Elt); |
307 | if (!CV || !this->isValue(CV->getValue())) |
308 | return false; |
309 | HasNonUndefElements = true; |
310 | } |
311 | return HasNonUndefElements; |
312 | } |
313 | } |
314 | return false; |
315 | } |
316 | }; |
317 | |
318 | /// specialization of cstval_pred_ty for ConstantInt |
319 | template <typename Predicate> |
320 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; |
321 | |
322 | /// specialization of cstval_pred_ty for ConstantFP |
323 | template <typename Predicate> |
324 | using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>; |
325 | |
326 | /// This helper class is used to match scalar and vector constants that |
327 | /// satisfy a specified predicate, and bind them to an APInt. |
328 | template <typename Predicate> struct api_pred_ty : public Predicate { |
329 | const APInt *&Res; |
330 | |
331 | api_pred_ty(const APInt *&R) : Res(R) {} |
332 | |
333 | template <typename ITy> bool match(ITy *V) { |
334 | if (const auto *CI = dyn_cast<ConstantInt>(V)) |
335 | if (this->isValue(CI->getValue())) { |
336 | Res = &CI->getValue(); |
337 | return true; |
338 | } |
339 | if (V->getType()->isVectorTy()) |
340 | if (const auto *C = dyn_cast<Constant>(V)) |
341 | if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) |
342 | if (this->isValue(CI->getValue())) { |
343 | Res = &CI->getValue(); |
344 | return true; |
345 | } |
346 | |
347 | return false; |
348 | } |
349 | }; |
350 | |
351 | /// This helper class is used to match scalar and vector constants that |
352 | /// satisfy a specified predicate, and bind them to an APFloat. |
353 | /// Undefs are allowed in splat vector constants. |
354 | template <typename Predicate> struct apf_pred_ty : public Predicate { |
355 | const APFloat *&Res; |
356 | |
357 | apf_pred_ty(const APFloat *&R) : Res(R) {} |
358 | |
359 | template <typename ITy> bool match(ITy *V) { |
360 | if (const auto *CI = dyn_cast<ConstantFP>(V)) |
361 | if (this->isValue(CI->getValue())) { |
362 | Res = &CI->getValue(); |
363 | return true; |
364 | } |
365 | if (V->getType()->isVectorTy()) |
366 | if (const auto *C = dyn_cast<Constant>(V)) |
367 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
368 | C->getSplatValue(/* AllowUndef */ true))) |
369 | if (this->isValue(CI->getValue())) { |
370 | Res = &CI->getValue(); |
371 | return true; |
372 | } |
373 | |
374 | return false; |
375 | } |
376 | }; |
377 | |
378 | /////////////////////////////////////////////////////////////////////////////// |
379 | // |
380 | // Encapsulate constant value queries for use in templated predicate matchers. |
381 | // This allows checking if constants match using compound predicates and works |
382 | // with vector constants, possibly with relaxed constraints. For example, ignore |
383 | // undef values. |
384 | // |
385 | /////////////////////////////////////////////////////////////////////////////// |
386 | |
387 | struct is_any_apint { |
388 | bool isValue(const APInt &C) { return true; } |
389 | }; |
390 | /// Match an integer or vector with any integral constant. |
391 | /// For vectors, this includes constants with undefined elements. |
392 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { |
393 | return cst_pred_ty<is_any_apint>(); |
394 | } |
395 | |
396 | struct is_all_ones { |
397 | bool isValue(const APInt &C) { return C.isAllOnesValue(); } |
398 | }; |
399 | /// Match an integer or vector with all bits set. |
400 | /// For vectors, this includes constants with undefined elements. |
401 | inline cst_pred_ty<is_all_ones> m_AllOnes() { |
402 | return cst_pred_ty<is_all_ones>(); |
403 | } |
404 | |
405 | struct is_maxsignedvalue { |
406 | bool isValue(const APInt &C) { return C.isMaxSignedValue(); } |
407 | }; |
408 | /// Match an integer or vector with values having all bits except for the high |
409 | /// bit set (0x7f...). |
410 | /// For vectors, this includes constants with undefined elements. |
411 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { |
412 | return cst_pred_ty<is_maxsignedvalue>(); |
413 | } |
414 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { |
415 | return V; |
416 | } |
417 | |
418 | struct is_negative { |
419 | bool isValue(const APInt &C) { return C.isNegative(); } |
420 | }; |
421 | /// Match an integer or vector of negative values. |
422 | /// For vectors, this includes constants with undefined elements. |
423 | inline cst_pred_ty<is_negative> m_Negative() { |
424 | return cst_pred_ty<is_negative>(); |
425 | } |
426 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { |
427 | return V; |
428 | } |
429 | |
430 | struct is_nonnegative { |
431 | bool isValue(const APInt &C) { return C.isNonNegative(); } |
432 | }; |
433 | /// Match an integer or vector of non-negative values. |
434 | /// For vectors, this includes constants with undefined elements. |
435 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { |
436 | return cst_pred_ty<is_nonnegative>(); |
437 | } |
438 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { |
439 | return V; |
440 | } |
441 | |
442 | struct is_strictlypositive { |
443 | bool isValue(const APInt &C) { return C.isStrictlyPositive(); } |
444 | }; |
445 | /// Match an integer or vector of strictly positive values. |
446 | /// For vectors, this includes constants with undefined elements. |
447 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { |
448 | return cst_pred_ty<is_strictlypositive>(); |
449 | } |
450 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { |
451 | return V; |
452 | } |
453 | |
454 | struct is_nonpositive { |
455 | bool isValue(const APInt &C) { return C.isNonPositive(); } |
456 | }; |
457 | /// Match an integer or vector of non-positive values. |
458 | /// For vectors, this includes constants with undefined elements. |
459 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { |
460 | return cst_pred_ty<is_nonpositive>(); |
461 | } |
462 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } |
463 | |
464 | struct is_one { |
465 | bool isValue(const APInt &C) { return C.isOneValue(); } |
466 | }; |
467 | /// Match an integer 1 or a vector with all elements equal to 1. |
468 | /// For vectors, this includes constants with undefined elements. |
469 | inline cst_pred_ty<is_one> m_One() { |
470 | return cst_pred_ty<is_one>(); |
471 | } |
472 | |
473 | struct is_zero_int { |
474 | bool isValue(const APInt &C) { return C.isNullValue(); } |
475 | }; |
476 | /// Match an integer 0 or a vector with all elements equal to 0. |
477 | /// For vectors, this includes constants with undefined elements. |
478 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { |
479 | return cst_pred_ty<is_zero_int>(); |
480 | } |
481 | |
482 | struct is_zero { |
483 | template <typename ITy> bool match(ITy *V) { |
484 | auto *C = dyn_cast<Constant>(V); |
485 | // FIXME: this should be able to do something for scalable vectors |
486 | return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C)); |
487 | } |
488 | }; |
489 | /// Match any null constant or a vector with all elements equal to 0. |
490 | /// For vectors, this includes constants with undefined elements. |
491 | inline is_zero m_Zero() { |
492 | return is_zero(); |
493 | } |
494 | |
495 | struct is_power2 { |
496 | bool isValue(const APInt &C) { return C.isPowerOf2(); } |
497 | }; |
498 | /// Match an integer or vector power-of-2. |
499 | /// For vectors, this includes constants with undefined elements. |
500 | inline cst_pred_ty<is_power2> m_Power2() { |
501 | return cst_pred_ty<is_power2>(); |
502 | } |
503 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { |
504 | return V; |
505 | } |
506 | |
507 | struct is_negated_power2 { |
508 | bool isValue(const APInt &C) { return (-C).isPowerOf2(); } |
509 | }; |
510 | /// Match a integer or vector negated power-of-2. |
511 | /// For vectors, this includes constants with undefined elements. |
512 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { |
513 | return cst_pred_ty<is_negated_power2>(); |
514 | } |
515 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { |
516 | return V; |
517 | } |
518 | |
519 | struct is_power2_or_zero { |
520 | bool isValue(const APInt &C) { return !C || C.isPowerOf2(); } |
521 | }; |
522 | /// Match an integer or vector of 0 or power-of-2 values. |
523 | /// For vectors, this includes constants with undefined elements. |
524 | inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() { |
525 | return cst_pred_ty<is_power2_or_zero>(); |
526 | } |
527 | inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) { |
528 | return V; |
529 | } |
530 | |
531 | struct is_sign_mask { |
532 | bool isValue(const APInt &C) { return C.isSignMask(); } |
533 | }; |
534 | /// Match an integer or vector with only the sign bit(s) set. |
535 | /// For vectors, this includes constants with undefined elements. |
536 | inline cst_pred_ty<is_sign_mask> m_SignMask() { |
537 | return cst_pred_ty<is_sign_mask>(); |
538 | } |
539 | |
540 | struct is_lowbit_mask { |
541 | bool isValue(const APInt &C) { return C.isMask(); } |
542 | }; |
543 | /// Match an integer or vector with only the low bit(s) set. |
544 | /// For vectors, this includes constants with undefined elements. |
545 | inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() { |
546 | return cst_pred_ty<is_lowbit_mask>(); |
547 | } |
548 | |
549 | struct icmp_pred_with_threshold { |
550 | ICmpInst::Predicate Pred; |
551 | const APInt *Thr; |
552 | bool isValue(const APInt &C) { |
553 | switch (Pred) { |
554 | case ICmpInst::Predicate::ICMP_EQ: |
555 | return C.eq(*Thr); |
556 | case ICmpInst::Predicate::ICMP_NE: |
557 | return C.ne(*Thr); |
558 | case ICmpInst::Predicate::ICMP_UGT: |
559 | return C.ugt(*Thr); |
560 | case ICmpInst::Predicate::ICMP_UGE: |
561 | return C.uge(*Thr); |
562 | case ICmpInst::Predicate::ICMP_ULT: |
563 | return C.ult(*Thr); |
564 | case ICmpInst::Predicate::ICMP_ULE: |
565 | return C.ule(*Thr); |
566 | case ICmpInst::Predicate::ICMP_SGT: |
567 | return C.sgt(*Thr); |
568 | case ICmpInst::Predicate::ICMP_SGE: |
569 | return C.sge(*Thr); |
570 | case ICmpInst::Predicate::ICMP_SLT: |
571 | return C.slt(*Thr); |
572 | case ICmpInst::Predicate::ICMP_SLE: |
573 | return C.sle(*Thr); |
574 | default: |
575 | llvm_unreachable("Unhandled ICmp predicate")::llvm::llvm_unreachable_internal("Unhandled ICmp predicate", "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/include/llvm/IR/PatternMatch.h" , 575); |
576 | } |
577 | } |
578 | }; |
579 | /// Match an integer or vector with every element comparing 'pred' (eg/ne/...) |
580 | /// to Threshold. For vectors, this includes constants with undefined elements. |
581 | inline cst_pred_ty<icmp_pred_with_threshold> |
582 | m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) { |
583 | cst_pred_ty<icmp_pred_with_threshold> P; |
584 | P.Pred = Predicate; |
585 | P.Thr = &Threshold; |
586 | return P; |
587 | } |
588 | |
589 | struct is_nan { |
590 | bool isValue(const APFloat &C) { return C.isNaN(); } |
591 | }; |
592 | /// Match an arbitrary NaN constant. This includes quiet and signalling nans. |
593 | /// For vectors, this includes constants with undefined elements. |
594 | inline cstfp_pred_ty<is_nan> m_NaN() { |
595 | return cstfp_pred_ty<is_nan>(); |
596 | } |
597 | |
598 | struct is_nonnan { |
599 | bool isValue(const APFloat &C) { return !C.isNaN(); } |
600 | }; |
601 | /// Match a non-NaN FP constant. |
602 | /// For vectors, this includes constants with undefined elements. |
603 | inline cstfp_pred_ty<is_nonnan> m_NonNaN() { |
604 | return cstfp_pred_ty<is_nonnan>(); |
605 | } |
606 | |
607 | struct is_inf { |
608 | bool isValue(const APFloat &C) { return C.isInfinity(); } |
609 | }; |
610 | /// Match a positive or negative infinity FP constant. |
611 | /// For vectors, this includes constants with undefined elements. |
612 | inline cstfp_pred_ty<is_inf> m_Inf() { |
613 | return cstfp_pred_ty<is_inf>(); |
614 | } |
615 | |
616 | struct is_noninf { |
617 | bool isValue(const APFloat &C) { return !C.isInfinity(); } |
618 | }; |
619 | /// Match a non-infinity FP constant, i.e. finite or NaN. |
620 | /// For vectors, this includes constants with undefined elements. |
621 | inline cstfp_pred_ty<is_noninf> m_NonInf() { |
622 | return cstfp_pred_ty<is_noninf>(); |
623 | } |
624 | |
625 | struct is_finite { |
626 | bool isValue(const APFloat &C) { return C.isFinite(); } |
627 | }; |
628 | /// Match a finite FP constant, i.e. not infinity or NaN. |
629 | /// For vectors, this includes constants with undefined elements. |
630 | inline cstfp_pred_ty<is_finite> m_Finite() { |
631 | return cstfp_pred_ty<is_finite>(); |
632 | } |
633 | inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; } |
634 | |
635 | struct is_finitenonzero { |
636 | bool isValue(const APFloat &C) { return C.isFiniteNonZero(); } |
637 | }; |
638 | /// Match a finite non-zero FP constant. |
639 | /// For vectors, this includes constants with undefined elements. |
640 | inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() { |
641 | return cstfp_pred_ty<is_finitenonzero>(); |
642 | } |
643 | inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) { |
644 | return V; |
645 | } |
646 | |
647 | struct is_any_zero_fp { |
648 | bool isValue(const APFloat &C) { return C.isZero(); } |
649 | }; |
650 | /// Match a floating-point negative zero or positive zero. |
651 | /// For vectors, this includes constants with undefined elements. |
652 | inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() { |
653 | return cstfp_pred_ty<is_any_zero_fp>(); |
654 | } |
655 | |
656 | struct is_pos_zero_fp { |
657 | bool isValue(const APFloat &C) { return C.isPosZero(); } |
658 | }; |
659 | /// Match a floating-point positive zero. |
660 | /// For vectors, this includes constants with undefined elements. |
661 | inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() { |
662 | return cstfp_pred_ty<is_pos_zero_fp>(); |
663 | } |
664 | |
665 | struct is_neg_zero_fp { |
666 | bool isValue(const APFloat &C) { return C.isNegZero(); } |
667 | }; |
668 | /// Match a floating-point negative zero. |
669 | /// For vectors, this includes constants with undefined elements. |
670 | inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() { |
671 | return cstfp_pred_ty<is_neg_zero_fp>(); |
672 | } |
673 | |
674 | struct is_non_zero_fp { |
675 | bool isValue(const APFloat &C) { return C.isNonZero(); } |
676 | }; |
677 | /// Match a floating-point non-zero. |
678 | /// For vectors, this includes constants with undefined elements. |
679 | inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() { |
680 | return cstfp_pred_ty<is_non_zero_fp>(); |
681 | } |
682 | |
683 | /////////////////////////////////////////////////////////////////////////////// |
684 | |
685 | template <typename Class> struct bind_ty { |
686 | Class *&VR; |
687 | |
688 | bind_ty(Class *&V) : VR(V) {} |
689 | |
690 | template <typename ITy> bool match(ITy *V) { |
691 | if (auto *CV = dyn_cast<Class>(V)) { |
692 | VR = CV; |
693 | return true; |
694 | } |
695 | return false; |
696 | } |
697 | }; |
698 | |
699 | /// Match a value, capturing it if we match. |
700 | inline bind_ty<Value> m_Value(Value *&V) { return V; } |
701 | inline bind_ty<const Value> m_Value(const Value *&V) { return V; } |
702 | |
703 | /// Match an instruction, capturing it if we match. |
704 | inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; } |
705 | /// Match a unary operator, capturing it if we match. |
706 | inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; } |
707 | /// Match a binary operator, capturing it if we match. |
708 | inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; } |
709 | /// Match a with overflow intrinsic, capturing it if we match. |
710 | inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; } |
711 | inline bind_ty<const WithOverflowInst> |
712 | m_WithOverflowInst(const WithOverflowInst *&I) { |
713 | return I; |
714 | } |
715 | |
716 | /// Match a Constant, capturing the value if we match. |
717 | inline bind_ty<Constant> m_Constant(Constant *&C) { return C; } |
718 | |
719 | /// Match a ConstantInt, capturing the value if we match. |
720 | inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; } |
721 | |
722 | /// Match a ConstantFP, capturing the value if we match. |
723 | inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; } |
724 | |
725 | /// Match a ConstantExpr, capturing the value if we match. |
726 | inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; } |
727 | |
728 | /// Match a basic block value, capturing it if we match. |
729 | inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; } |
730 | inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) { |
731 | return V; |
732 | } |
733 | |
734 | /// Match an arbitrary immediate Constant and ignore it. |
735 | inline match_combine_and<class_match<Constant>, |
736 | match_unless<class_match<ConstantExpr>>> |
737 | m_ImmConstant() { |
738 | return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr())); |
739 | } |
740 | |
741 | /// Match an immediate Constant, capturing the value if we match. |
742 | inline match_combine_and<bind_ty<Constant>, |
743 | match_unless<class_match<ConstantExpr>>> |
744 | m_ImmConstant(Constant *&C) { |
745 | return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr())); |
746 | } |
747 | |
748 | /// Match a specified Value*. |
749 | struct specificval_ty { |
750 | const Value *Val; |
751 | |
752 | specificval_ty(const Value *V) : Val(V) {} |
753 | |
754 | template <typename ITy> bool match(ITy *V) { return V == Val; } |
755 | }; |
756 | |
757 | /// Match if we have a specific specified value. |
758 | inline specificval_ty m_Specific(const Value *V) { return V; } |
759 | |
760 | /// Stores a reference to the Value *, not the Value * itself, |
761 | /// thus can be used in commutative matchers. |
762 | template <typename Class> struct deferredval_ty { |
763 | Class *const &Val; |
764 | |
765 | deferredval_ty(Class *const &V) : Val(V) {} |
766 | |
767 | template <typename ITy> bool match(ITy *const V) { return V == Val; } |
768 | }; |
769 | |
770 | /// A commutative-friendly version of m_Specific(). |
771 | inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; } |
772 | inline deferredval_ty<const Value> m_Deferred(const Value *const &V) { |
773 | return V; |
774 | } |
775 | |
776 | /// Match a specified floating point value or vector of all elements of |
777 | /// that value. |
778 | struct specific_fpval { |
779 | double Val; |
780 | |
781 | specific_fpval(double V) : Val(V) {} |
782 | |
783 | template <typename ITy> bool match(ITy *V) { |
784 | if (const auto *CFP = dyn_cast<ConstantFP>(V)) |
785 | return CFP->isExactlyValue(Val); |
786 | if (V->getType()->isVectorTy()) |
787 | if (const auto *C = dyn_cast<Constant>(V)) |
788 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) |
789 | return CFP->isExactlyValue(Val); |
790 | return false; |
791 | } |
792 | }; |
793 | |
794 | /// Match a specific floating point value or vector with all elements |
795 | /// equal to the value. |
796 | inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); } |
797 | |
798 | /// Match a float 1.0 or vector with all elements equal to 1.0. |
799 | inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); } |
800 | |
801 | struct bind_const_intval_ty { |
802 | uint64_t &VR; |
803 | |
804 | bind_const_intval_ty(uint64_t &V) : VR(V) {} |
805 | |
806 | template <typename ITy> bool match(ITy *V) { |
807 | if (const auto *CV = dyn_cast<ConstantInt>(V)) |
808 | if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) { |
809 | VR = CV->getZExtValue(); |
810 | return true; |
811 | } |
812 | return false; |
813 | } |
814 | }; |
815 | |
816 | /// Match a specified integer value or vector of all elements of that |
817 | /// value. |
818 | template <bool AllowUndefs> |
819 | struct specific_intval { |
820 | APInt Val; |
821 | |
822 | specific_intval(APInt V) : Val(std::move(V)) {} |
823 | |
824 | template <typename ITy> bool match(ITy *V) { |
825 | const auto *CI = dyn_cast<ConstantInt>(V); |
826 | if (!CI && V->getType()->isVectorTy()) |
827 | if (const auto *C = dyn_cast<Constant>(V)) |
828 | CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs)); |
829 | |
830 | return CI && APInt::isSameValue(CI->getValue(), Val); |
831 | } |
832 | }; |
833 | |
834 | /// Match a specific integer value or vector with all elements equal to |
835 | /// the value. |
836 | inline specific_intval<false> m_SpecificInt(APInt V) { |
837 | return specific_intval<false>(std::move(V)); |
838 | } |
839 | |
840 | inline specific_intval<false> m_SpecificInt(uint64_t V) { |
841 | return m_SpecificInt(APInt(64, V)); |
842 | } |
843 | |
844 | inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) { |
845 | return specific_intval<true>(std::move(V)); |
846 | } |
847 | |
848 | inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) { |
849 | return m_SpecificIntAllowUndef(APInt(64, V)); |
850 | } |
851 | |
852 | /// Match a ConstantInt and bind to its value. This does not match |
853 | /// ConstantInts wider than 64-bits. |
854 | inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; } |
855 | |
856 | /// Match a specified basic block value. |
857 | struct specific_bbval { |
858 | BasicBlock *Val; |
859 | |
860 | specific_bbval(BasicBlock *Val) : Val(Val) {} |
861 | |
862 | template <typename ITy> bool match(ITy *V) { |
863 | const auto *BB = dyn_cast<BasicBlock>(V); |
864 | return BB && BB == Val; |
865 | } |
866 | }; |
867 | |
868 | /// Match a specific basic block value. |
869 | inline specific_bbval m_SpecificBB(BasicBlock *BB) { |
870 | return specific_bbval(BB); |
871 | } |
872 | |
873 | /// A commutative-friendly version of m_Specific(). |
874 | inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) { |
875 | return BB; |
876 | } |
877 | inline deferredval_ty<const BasicBlock> |
878 | m_Deferred(const BasicBlock *const &BB) { |
879 | return BB; |
880 | } |
881 | |
882 | //===----------------------------------------------------------------------===// |
883 | // Matcher for any binary operator. |
884 | // |
885 | template <typename LHS_t, typename RHS_t, bool Commutable = false> |
886 | struct AnyBinaryOp_match { |
887 | LHS_t L; |
888 | RHS_t R; |
889 | |
890 | // The evaluation order is always stable, regardless of Commutability. |
891 | // The LHS is always matched first. |
892 | AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
893 | |
894 | template <typename OpTy> bool match(OpTy *V) { |
895 | if (auto *I = dyn_cast<BinaryOperator>(V)) |
896 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || |
897 | (Commutable && L.match(I->getOperand(1)) && |
898 | R.match(I->getOperand(0))); |
899 | return false; |
900 | } |
901 | }; |
902 | |
903 | template <typename LHS, typename RHS> |
904 | inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) { |
905 | return AnyBinaryOp_match<LHS, RHS>(L, R); |
906 | } |
907 | |
908 | //===----------------------------------------------------------------------===// |
909 | // Matcher for any unary operator. |
910 | // TODO fuse unary, binary matcher into n-ary matcher |
911 | // |
912 | template <typename OP_t> struct AnyUnaryOp_match { |
913 | OP_t X; |
914 | |
915 | AnyUnaryOp_match(const OP_t &X) : X(X) {} |
916 | |
917 | template <typename OpTy> bool match(OpTy *V) { |
918 | if (auto *I = dyn_cast<UnaryOperator>(V)) |
919 | return X.match(I->getOperand(0)); |
920 | return false; |
921 | } |
922 | }; |
923 | |
924 | template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) { |
925 | return AnyUnaryOp_match<OP_t>(X); |
926 | } |
927 | |
928 | //===----------------------------------------------------------------------===// |
929 | // Matchers for specific binary operators. |
930 | // |
931 | |
932 | template <typename LHS_t, typename RHS_t, unsigned Opcode, |
933 | bool Commutable = false> |
934 | struct BinaryOp_match { |
935 | LHS_t L; |
936 | RHS_t R; |
937 | |
938 | // The evaluation order is always stable, regardless of Commutability. |
939 | // The LHS is always matched first. |
940 | BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
941 | |
942 | template <typename OpTy> bool match(OpTy *V) { |
943 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
944 | auto *I = cast<BinaryOperator>(V); |
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 | } |
949 | if (auto *CE = dyn_cast<ConstantExpr>(V)) |
950 | return CE->getOpcode() == Opcode && |
951 | ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) || |
952 | (Commutable && L.match(CE->getOperand(1)) && |
953 | R.match(CE->getOperand(0)))); |
954 | return false; |
955 | } |
956 | }; |
957 | |
958 | template <typename LHS, typename RHS> |
959 | inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L, |
960 | const RHS &R) { |
961 | return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R); |
962 | } |
963 | |
964 | template <typename LHS, typename RHS> |
965 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L, |
966 | const RHS &R) { |
967 | return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R); |
968 | } |
969 | |
970 | template <typename LHS, typename RHS> |
971 | inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L, |
972 | const RHS &R) { |
973 | return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R); |
974 | } |
975 | |
976 | template <typename LHS, typename RHS> |
977 | inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L, |
978 | const RHS &R) { |
979 | return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R); |
980 | } |
981 | |
982 | template <typename Op_t> struct FNeg_match { |
983 | Op_t X; |
984 | |
985 | FNeg_match(const Op_t &Op) : X(Op) {} |
986 | template <typename OpTy> bool match(OpTy *V) { |
987 | auto *FPMO = dyn_cast<FPMathOperator>(V); |
988 | if (!FPMO) return false; |
989 | |
990 | if (FPMO->getOpcode() == Instruction::FNeg) |
991 | return X.match(FPMO->getOperand(0)); |
992 | |
993 | if (FPMO->getOpcode() == Instruction::FSub) { |
994 | if (FPMO->hasNoSignedZeros()) { |
995 | // With 'nsz', any zero goes. |
996 | if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0))) |
997 | return false; |
998 | } else { |
999 | // Without 'nsz', we need fsub -0.0, X exactly. |
1000 | if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0))) |
1001 | return false; |
1002 | } |
1003 | |
1004 | return X.match(FPMO->getOperand(1)); |
1005 | } |
1006 | |
1007 | return false; |
1008 | } |
1009 | }; |
1010 | |
1011 | /// Match 'fneg X' as 'fsub -0.0, X'. |
1012 | template <typename OpTy> |
1013 | inline FNeg_match<OpTy> |
1014 | m_FNeg(const OpTy &X) { |
1015 | return FNeg_match<OpTy>(X); |
1016 | } |
1017 | |
1018 | /// Match 'fneg X' as 'fsub +-0.0, X'. |
1019 | template <typename RHS> |
1020 | inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub> |
1021 | m_FNegNSZ(const RHS &X) { |
1022 | return m_FSub(m_AnyZeroFP(), X); |
1023 | } |
1024 | |
1025 | template <typename LHS, typename RHS> |
1026 | inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L, |
1027 | const RHS &R) { |
1028 | return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R); |
1029 | } |
1030 | |
1031 | template <typename LHS, typename RHS> |
1032 | inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L, |
1033 | const RHS &R) { |
1034 | return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R); |
1035 | } |
1036 | |
1037 | template <typename LHS, typename RHS> |
1038 | inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L, |
1039 | const RHS &R) { |
1040 | return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R); |
1041 | } |
1042 | |
1043 | template <typename LHS, typename RHS> |
1044 | inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L, |
1045 | const RHS &R) { |
1046 | return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R); |
1047 | } |
1048 | |
1049 | template <typename LHS, typename RHS> |
1050 | inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L, |
1051 | const RHS &R) { |
1052 | return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R); |
1053 | } |
1054 | |
1055 | template <typename LHS, typename RHS> |
1056 | inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L, |
1057 | const RHS &R) { |
1058 | return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R); |
1059 | } |
1060 | |
1061 | template <typename LHS, typename RHS> |
1062 | inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L, |
1063 | const RHS &R) { |
1064 | return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R); |
1065 | } |
1066 | |
1067 | template <typename LHS, typename RHS> |
1068 | inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L, |
1069 | const RHS &R) { |
1070 | return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R); |
1071 | } |
1072 | |
1073 | template <typename LHS, typename RHS> |
1074 | inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L, |
1075 | const RHS &R) { |
1076 | return BinaryOp_match<LHS, RHS, Instruction::And>(L, R); |
1077 | } |
1078 | |
1079 | template <typename LHS, typename RHS> |
1080 | inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L, |
1081 | const RHS &R) { |
1082 | return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R); |
1083 | } |
1084 | |
1085 | template <typename LHS, typename RHS> |
1086 | inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L, |
1087 | const RHS &R) { |
1088 | return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R); |
1089 | } |
1090 | |
1091 | template <typename LHS, typename RHS> |
1092 | inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L, |
1093 | const RHS &R) { |
1094 | return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R); |
1095 | } |
1096 | |
1097 | template <typename LHS, typename RHS> |
1098 | inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L, |
1099 | const RHS &R) { |
1100 | return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R); |
1101 | } |
1102 | |
1103 | template <typename LHS, typename RHS> |
1104 | inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L, |
1105 | const RHS &R) { |
1106 | return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R); |
1107 | } |
1108 | |
1109 | template <typename LHS_t, typename RHS_t, unsigned Opcode, |
1110 | unsigned WrapFlags = 0> |
1111 | struct OverflowingBinaryOp_match { |
1112 | LHS_t L; |
1113 | RHS_t R; |
1114 | |
1115 | OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) |
1116 | : L(LHS), R(RHS) {} |
1117 | |
1118 | template <typename OpTy> bool match(OpTy *V) { |
1119 | if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) { |
1120 | if (Op->getOpcode() != Opcode) |
1121 | return false; |
1122 | if (WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap && |
1123 | !Op->hasNoUnsignedWrap()) |
1124 | return false; |
1125 | if (WrapFlags & OverflowingBinaryOperator::NoSignedWrap && |
1126 | !Op->hasNoSignedWrap()) |
1127 | return false; |
1128 | return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1)); |
1129 | } |
1130 | return false; |
1131 | } |
1132 | }; |
1133 | |
1134 | template <typename LHS, typename RHS> |
1135 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1136 | OverflowingBinaryOperator::NoSignedWrap> |
1137 | m_NSWAdd(const LHS &L, const RHS &R) { |
1138 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1139 | OverflowingBinaryOperator::NoSignedWrap>( |
1140 | L, R); |
1141 | } |
1142 | template <typename LHS, typename RHS> |
1143 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1144 | OverflowingBinaryOperator::NoSignedWrap> |
1145 | m_NSWSub(const LHS &L, const RHS &R) { |
1146 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1147 | OverflowingBinaryOperator::NoSignedWrap>( |
1148 | L, R); |
1149 | } |
1150 | template <typename LHS, typename RHS> |
1151 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1152 | OverflowingBinaryOperator::NoSignedWrap> |
1153 | m_NSWMul(const LHS &L, const RHS &R) { |
1154 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1155 | OverflowingBinaryOperator::NoSignedWrap>( |
1156 | L, R); |
1157 | } |
1158 | template <typename LHS, typename RHS> |
1159 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1160 | OverflowingBinaryOperator::NoSignedWrap> |
1161 | m_NSWShl(const LHS &L, const RHS &R) { |
1162 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1163 | OverflowingBinaryOperator::NoSignedWrap>( |
1164 | L, R); |
1165 | } |
1166 | |
1167 | template <typename LHS, typename RHS> |
1168 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1169 | OverflowingBinaryOperator::NoUnsignedWrap> |
1170 | m_NUWAdd(const LHS &L, const RHS &R) { |
1171 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1172 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1173 | L, R); |
1174 | } |
1175 | template <typename LHS, typename RHS> |
1176 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1177 | OverflowingBinaryOperator::NoUnsignedWrap> |
1178 | m_NUWSub(const LHS &L, const RHS &R) { |
1179 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1180 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1181 | L, R); |
1182 | } |
1183 | template <typename LHS, typename RHS> |
1184 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1185 | OverflowingBinaryOperator::NoUnsignedWrap> |
1186 | m_NUWMul(const LHS &L, const RHS &R) { |
1187 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1188 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1189 | L, R); |
1190 | } |
1191 | template <typename LHS, typename RHS> |
1192 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1193 | OverflowingBinaryOperator::NoUnsignedWrap> |
1194 | m_NUWShl(const LHS &L, const RHS &R) { |
1195 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1196 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1197 | L, R); |
1198 | } |
1199 | |
1200 | //===----------------------------------------------------------------------===// |
1201 | // Class that matches a group of binary opcodes. |
1202 | // |
1203 | template <typename LHS_t, typename RHS_t, typename Predicate> |
1204 | struct BinOpPred_match : Predicate { |
1205 | LHS_t L; |
1206 | RHS_t R; |
1207 | |
1208 | BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
1209 | |
1210 | template <typename OpTy> bool match(OpTy *V) { |
1211 | if (auto *I = dyn_cast<Instruction>(V)) |
1212 | return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) && |
1213 | R.match(I->getOperand(1)); |
1214 | if (auto *CE = dyn_cast<ConstantExpr>(V)) |
1215 | return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) && |
1216 | R.match(CE->getOperand(1)); |
1217 | return false; |
1218 | } |
1219 | }; |
1220 | |
1221 | struct is_shift_op { |
1222 | bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); } |
1223 | }; |
1224 | |
1225 | struct is_right_shift_op { |
1226 | bool isOpType(unsigned Opcode) { |
1227 | return Opcode == Instruction::LShr || Opcode == Instruction::AShr; |
1228 | } |
1229 | }; |
1230 | |
1231 | struct is_logical_shift_op { |
1232 | bool isOpType(unsigned Opcode) { |
1233 | return Opcode == Instruction::LShr || Opcode == Instruction::Shl; |
1234 | } |
1235 | }; |
1236 | |
1237 | struct is_bitwiselogic_op { |
1238 | bool isOpType(unsigned Opcode) { |
1239 | return Instruction::isBitwiseLogicOp(Opcode); |
1240 | } |
1241 | }; |
1242 | |
1243 | struct is_idiv_op { |
1244 | bool isOpType(unsigned Opcode) { |
1245 | return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; |
1246 | } |
1247 | }; |
1248 | |
1249 | struct is_irem_op { |
1250 | bool isOpType(unsigned Opcode) { |
1251 | return Opcode == Instruction::SRem || Opcode == Instruction::URem; |
1252 | } |
1253 | }; |
1254 | |
1255 | /// Matches shift operations. |
1256 | template <typename LHS, typename RHS> |
1257 | inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L, |
1258 | const RHS &R) { |
1259 | return BinOpPred_match<LHS, RHS, is_shift_op>(L, R); |
1260 | } |
1261 | |
1262 | /// Matches logical shift operations. |
1263 | template <typename LHS, typename RHS> |
1264 | inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L, |
1265 | const RHS &R) { |
1266 | return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R); |
1267 | } |
1268 | |
1269 | /// Matches logical shift operations. |
1270 | template <typename LHS, typename RHS> |
1271 | inline BinOpPred_match<LHS, RHS, is_logical_shift_op> |
1272 | m_LogicalShift(const LHS &L, const RHS &R) { |
1273 | return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R); |
1274 | } |
1275 | |
1276 | /// Matches bitwise logic operations. |
1277 | template <typename LHS, typename RHS> |
1278 | inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op> |
1279 | m_BitwiseLogic(const LHS &L, const RHS &R) { |
1280 | return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R); |
1281 | } |
1282 | |
1283 | /// Matches integer division operations. |
1284 | template <typename LHS, typename RHS> |
1285 | inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L, |
1286 | const RHS &R) { |
1287 | return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R); |
1288 | } |
1289 | |
1290 | /// Matches integer remainder operations. |
1291 | template <typename LHS, typename RHS> |
1292 | inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L, |
1293 | const RHS &R) { |
1294 | return BinOpPred_match<LHS, RHS, is_irem_op>(L, R); |
1295 | } |
1296 | |
1297 | //===----------------------------------------------------------------------===// |
1298 | // Class that matches exact binary ops. |
1299 | // |
1300 | template <typename SubPattern_t> struct Exact_match { |
1301 | SubPattern_t SubPattern; |
1302 | |
1303 | Exact_match(const SubPattern_t &SP) : SubPattern(SP) {} |
1304 | |
1305 | template <typename OpTy> bool match(OpTy *V) { |
1306 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(V)) |
1307 | return PEO->isExact() && SubPattern.match(V); |
1308 | return false; |
1309 | } |
1310 | }; |
1311 | |
1312 | template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) { |
1313 | return SubPattern; |
1314 | } |
1315 | |
1316 | //===----------------------------------------------------------------------===// |
1317 | // Matchers for CmpInst classes |
1318 | // |
1319 | |
1320 | template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy, |
1321 | bool Commutable = false> |
1322 | struct CmpClass_match { |
1323 | PredicateTy &Predicate; |
1324 | LHS_t L; |
1325 | RHS_t R; |
1326 | |
1327 | // The evaluation order is always stable, regardless of Commutability. |
1328 | // The LHS is always matched first. |
1329 | CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS) |
1330 | : Predicate(Pred), L(LHS), R(RHS) {} |
1331 | |
1332 | template <typename OpTy> bool match(OpTy *V) { |
1333 | if (auto *I = dyn_cast<Class>(V)) { |
1334 | if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) { |
1335 | Predicate = I->getPredicate(); |
1336 | return true; |
1337 | } else if (Commutable && L.match(I->getOperand(1)) && |
1338 | R.match(I->getOperand(0))) { |
1339 | Predicate = I->getSwappedPredicate(); |
1340 | return true; |
1341 | } |
1342 | } |
1343 | return false; |
1344 | } |
1345 | }; |
1346 | |
1347 | template <typename LHS, typename RHS> |
1348 | inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate> |
1349 | m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
1350 | return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R); |
1351 | } |
1352 | |
1353 | template <typename LHS, typename RHS> |
1354 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate> |
1355 | m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
1356 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R); |
1357 | } |
1358 | |
1359 | template <typename LHS, typename RHS> |
1360 | inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate> |
1361 | m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
1362 | return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R); |
1363 | } |
1364 | |
1365 | //===----------------------------------------------------------------------===// |
1366 | // Matchers for instructions with a given opcode and number of operands. |
1367 | // |
1368 | |
1369 | /// Matches instructions with Opcode and three operands. |
1370 | template <typename T0, unsigned Opcode> struct OneOps_match { |
1371 | T0 Op1; |
1372 | |
1373 | OneOps_match(const T0 &Op1) : Op1(Op1) {} |
1374 | |
1375 | template <typename OpTy> bool match(OpTy *V) { |
1376 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
1377 | auto *I = cast<Instruction>(V); |
1378 | return Op1.match(I->getOperand(0)); |
1379 | } |
1380 | return false; |
1381 | } |
1382 | }; |
1383 | |
1384 | /// Matches instructions with Opcode and three operands. |
1385 | template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match { |
1386 | T0 Op1; |
1387 | T1 Op2; |
1388 | |
1389 | TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {} |
1390 | |
1391 | template <typename OpTy> bool match(OpTy *V) { |
1392 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
1393 | auto *I = cast<Instruction>(V); |
1394 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)); |
1395 | } |
1396 | return false; |
1397 | } |
1398 | }; |
1399 | |
1400 | /// Matches instructions with Opcode and three operands. |
1401 | template <typename T0, typename T1, typename T2, unsigned Opcode> |
1402 | struct ThreeOps_match { |
1403 | T0 Op1; |
1404 | T1 Op2; |
1405 | T2 Op3; |
1406 | |
1407 | ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3) |
1408 | : Op1(Op1), Op2(Op2), Op3(Op3) {} |
1409 | |
1410 | template <typename OpTy> bool match(OpTy *V) { |
1411 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
1412 | auto *I = cast<Instruction>(V); |
1413 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && |
1414 | Op3.match(I->getOperand(2)); |
1415 | } |
1416 | return false; |
1417 | } |
1418 | }; |
1419 | |
1420 | /// Matches SelectInst. |
1421 | template <typename Cond, typename LHS, typename RHS> |
1422 | inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select> |
1423 | m_Select(const Cond &C, const LHS &L, const RHS &R) { |
1424 | return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R); |
1425 | } |
1426 | |
1427 | /// This matches a select of two constants, e.g.: |
1428 | /// m_SelectCst<-1, 0>(m_Value(V)) |
1429 | template <int64_t L, int64_t R, typename Cond> |
1430 | inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>, |
1431 | Instruction::Select> |
1432 | m_SelectCst(const Cond &C) { |
1433 | return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>()); |
1434 | } |
1435 | |
1436 | /// Matches FreezeInst. |
1437 | template <typename OpTy> |
1438 | inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) { |
1439 | return OneOps_match<OpTy, Instruction::Freeze>(Op); |
1440 | } |
1441 | |
1442 | /// Matches InsertElementInst. |
1443 | template <typename Val_t, typename Elt_t, typename Idx_t> |
1444 | inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement> |
1445 | m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) { |
1446 | return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>( |
1447 | Val, Elt, Idx); |
1448 | } |
1449 | |
1450 | /// Matches ExtractElementInst. |
1451 | template <typename Val_t, typename Idx_t> |
1452 | inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement> |
1453 | m_ExtractElt(const Val_t &Val, const Idx_t &Idx) { |
1454 | return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx); |
1455 | } |
1456 | |
1457 | /// Matches shuffle. |
1458 | template <typename T0, typename T1, typename T2> struct Shuffle_match { |
1459 | T0 Op1; |
1460 | T1 Op2; |
1461 | T2 Mask; |
1462 | |
1463 | Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask) |
1464 | : Op1(Op1), Op2(Op2), Mask(Mask) {} |
1465 | |
1466 | template <typename OpTy> bool match(OpTy *V) { |
1467 | if (auto *I = dyn_cast<ShuffleVectorInst>(V)) { |
1468 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && |
1469 | Mask.match(I->getShuffleMask()); |
1470 | } |
1471 | return false; |
1472 | } |
1473 | }; |
1474 | |
1475 | struct m_Mask { |
1476 | ArrayRef<int> &MaskRef; |
1477 | m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} |
1478 | bool match(ArrayRef<int> Mask) { |
1479 | MaskRef = Mask; |
1480 | return true; |
1481 | } |
1482 | }; |
1483 | |
1484 | struct m_ZeroMask { |
1485 | bool match(ArrayRef<int> Mask) { |
1486 | return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; }); |
1487 | } |
1488 | }; |
1489 | |
1490 | struct m_SpecificMask { |
1491 | ArrayRef<int> &MaskRef; |
1492 | m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} |
1493 | bool match(ArrayRef<int> Mask) { return MaskRef == Mask; } |
1494 | }; |
1495 | |
1496 | struct m_SplatOrUndefMask { |
1497 | int &SplatIndex; |
1498 | m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {} |
1499 | bool match(ArrayRef<int> Mask) { |
1500 | auto First = find_if(Mask, [](int Elem) { return Elem != -1; }); |
1501 | if (First == Mask.end()) |
1502 | return false; |
1503 | SplatIndex = *First; |
1504 | return all_of(Mask, |
1505 | [First](int Elem) { return Elem == *First || Elem == -1; }); |
1506 | } |
1507 | }; |
1508 | |
1509 | /// Matches ShuffleVectorInst independently of mask value. |
1510 | template <typename V1_t, typename V2_t> |
1511 | inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector> |
1512 | m_Shuffle(const V1_t &v1, const V2_t &v2) { |
1513 | return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2); |
1514 | } |
1515 | |
1516 | template <typename V1_t, typename V2_t, typename Mask_t> |
1517 | inline Shuffle_match<V1_t, V2_t, Mask_t> |
1518 | m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) { |
1519 | return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask); |
1520 | } |
1521 | |
1522 | /// Matches LoadInst. |
1523 | template <typename OpTy> |
1524 | inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) { |
1525 | return OneOps_match<OpTy, Instruction::Load>(Op); |
1526 | } |
1527 | |
1528 | /// Matches StoreInst. |
1529 | template <typename ValueOpTy, typename PointerOpTy> |
1530 | inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store> |
1531 | m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) { |
1532 | return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp, |
1533 | PointerOp); |
1534 | } |
1535 | |
1536 | //===----------------------------------------------------------------------===// |
1537 | // Matchers for CastInst classes |
1538 | // |
1539 | |
1540 | template <typename Op_t, unsigned Opcode> struct CastClass_match { |
1541 | Op_t Op; |
1542 | |
1543 | CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {} |
1544 | |
1545 | template <typename OpTy> bool match(OpTy *V) { |
1546 | if (auto *O = dyn_cast<Operator>(V)) |
1547 | return O->getOpcode() == Opcode && Op.match(O->getOperand(0)); |
1548 | return false; |
1549 | } |
1550 | }; |
1551 | |
1552 | /// Matches BitCast. |
1553 | template <typename OpTy> |
1554 | inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) { |
1555 | return CastClass_match<OpTy, Instruction::BitCast>(Op); |
1556 | } |
1557 | |
1558 | /// Matches PtrToInt. |
1559 | template <typename OpTy> |
1560 | inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) { |
1561 | return CastClass_match<OpTy, Instruction::PtrToInt>(Op); |
1562 | } |
1563 | |
1564 | /// Matches IntToPtr. |
1565 | template <typename OpTy> |
1566 | inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) { |
1567 | return CastClass_match<OpTy, Instruction::IntToPtr>(Op); |
1568 | } |
1569 | |
1570 | /// Matches Trunc. |
1571 | template <typename OpTy> |
1572 | inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) { |
1573 | return CastClass_match<OpTy, Instruction::Trunc>(Op); |
1574 | } |
1575 | |
1576 | template <typename OpTy> |
1577 | inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy> |
1578 | m_TruncOrSelf(const OpTy &Op) { |
1579 | return m_CombineOr(m_Trunc(Op), Op); |
1580 | } |
1581 | |
1582 | /// Matches SExt. |
1583 | template <typename OpTy> |
1584 | inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) { |
1585 | return CastClass_match<OpTy, Instruction::SExt>(Op); |
1586 | } |
1587 | |
1588 | /// Matches ZExt. |
1589 | template <typename OpTy> |
1590 | inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) { |
1591 | return CastClass_match<OpTy, Instruction::ZExt>(Op); |
1592 | } |
1593 | |
1594 | template <typename OpTy> |
1595 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy> |
1596 | m_ZExtOrSelf(const OpTy &Op) { |
1597 | return m_CombineOr(m_ZExt(Op), Op); |
1598 | } |
1599 | |
1600 | template <typename OpTy> |
1601 | inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy> |
1602 | m_SExtOrSelf(const OpTy &Op) { |
1603 | return m_CombineOr(m_SExt(Op), Op); |
1604 | } |
1605 | |
1606 | template <typename OpTy> |
1607 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, |
1608 | CastClass_match<OpTy, Instruction::SExt>> |
1609 | m_ZExtOrSExt(const OpTy &Op) { |
1610 | return m_CombineOr(m_ZExt(Op), m_SExt(Op)); |
1611 | } |
1612 | |
1613 | template <typename OpTy> |
1614 | inline match_combine_or< |
1615 | match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, |
1616 | CastClass_match<OpTy, Instruction::SExt>>, |
1617 | OpTy> |
1618 | m_ZExtOrSExtOrSelf(const OpTy &Op) { |
1619 | return m_CombineOr(m_ZExtOrSExt(Op), Op); |
1620 | } |
1621 | |
1622 | template <typename OpTy> |
1623 | inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) { |
1624 | return CastClass_match<OpTy, Instruction::UIToFP>(Op); |
1625 | } |
1626 | |
1627 | template <typename OpTy> |
1628 | inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) { |
1629 | return CastClass_match<OpTy, Instruction::SIToFP>(Op); |
1630 | } |
1631 | |
1632 | template <typename OpTy> |
1633 | inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) { |
1634 | return CastClass_match<OpTy, Instruction::FPToUI>(Op); |
1635 | } |
1636 | |
1637 | template <typename OpTy> |
1638 | inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) { |
1639 | return CastClass_match<OpTy, Instruction::FPToSI>(Op); |
1640 | } |
1641 | |
1642 | template <typename OpTy> |
1643 | inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) { |
1644 | return CastClass_match<OpTy, Instruction::FPTrunc>(Op); |
1645 | } |
1646 | |
1647 | template <typename OpTy> |
1648 | inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) { |
1649 | return CastClass_match<OpTy, Instruction::FPExt>(Op); |
1650 | } |
1651 | |
1652 | //===----------------------------------------------------------------------===// |
1653 | // Matchers for control flow. |
1654 | // |
1655 | |
1656 | struct br_match { |
1657 | BasicBlock *&Succ; |
1658 | |
1659 | br_match(BasicBlock *&Succ) : Succ(Succ) {} |
1660 | |
1661 | template <typename OpTy> bool match(OpTy *V) { |
1662 | if (auto *BI = dyn_cast<BranchInst>(V)) |
1663 | if (BI->isUnconditional()) { |
1664 | Succ = BI->getSuccessor(0); |
1665 | return true; |
1666 | } |
1667 | return false; |
1668 | } |
1669 | }; |
1670 | |
1671 | inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); } |
1672 | |
1673 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> |
1674 | struct brc_match { |
1675 | Cond_t Cond; |
1676 | TrueBlock_t T; |
1677 | FalseBlock_t F; |
1678 | |
1679 | brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f) |
1680 | : Cond(C), T(t), F(f) {} |
1681 | |
1682 | template <typename OpTy> bool match(OpTy *V) { |
1683 | if (auto *BI = dyn_cast<BranchInst>(V)) |
1684 | if (BI->isConditional() && Cond.match(BI->getCondition())) |
1685 | return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1)); |
1686 | return false; |
1687 | } |
1688 | }; |
1689 | |
1690 | template <typename Cond_t> |
1691 | inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>> |
1692 | m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) { |
1693 | return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>( |
1694 | C, m_BasicBlock(T), m_BasicBlock(F)); |
1695 | } |
1696 | |
1697 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> |
1698 | inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t> |
1699 | m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) { |
1700 | return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F); |
1701 | } |
1702 | |
1703 | //===----------------------------------------------------------------------===// |
1704 | // Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y). |
1705 | // |
1706 | |
1707 | template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t, |
1708 | bool Commutable = false> |
1709 | struct MaxMin_match { |
1710 | LHS_t L; |
1711 | RHS_t R; |
1712 | |
1713 | // The evaluation order is always stable, regardless of Commutability. |
1714 | // The LHS is always matched first. |
1715 | MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
1716 | |
1717 | template <typename OpTy> bool match(OpTy *V) { |
1718 | if (auto *II = dyn_cast<IntrinsicInst>(V)) { |
1719 | Intrinsic::ID IID = II->getIntrinsicID(); |
1720 | if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) || |
1721 | (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) || |
1722 | (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) || |
1723 | (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) { |
1724 | Value *LHS = II->getOperand(0), *RHS = II->getOperand(1); |
1725 | return (L.match(LHS) && R.match(RHS)) || |
1726 | (Commutable && L.match(RHS) && R.match(LHS)); |
1727 | } |
1728 | } |
1729 | // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x". |
1730 | auto *SI = dyn_cast<SelectInst>(V); |
1731 | if (!SI) |
1732 | return false; |
1733 | auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition()); |
1734 | if (!Cmp) |
1735 | return false; |
1736 | // At this point we have a select conditioned on a comparison. Check that |
1737 | // it is the values returned by the select that are being compared. |
1738 | Value *TrueVal = SI->getTrueValue(); |
1739 | Value *FalseVal = SI->getFalseValue(); |
1740 | Value *LHS = Cmp->getOperand(0); |
1741 | Value *RHS = Cmp->getOperand(1); |
1742 | if ((TrueVal != LHS || FalseVal != RHS) && |
1743 | (TrueVal != RHS || FalseVal != LHS)) |
1744 | return false; |
1745 | typename CmpInst_t::Predicate Pred = |
1746 | LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate(); |
1747 | // Does "(x pred y) ? x : y" represent the desired max/min operation? |
1748 | if (!Pred_t::match(Pred)) |
1749 | return false; |
1750 | // It does! Bind the operands. |
1751 | return (L.match(LHS) && R.match(RHS)) || |
1752 | (Commutable && L.match(RHS) && R.match(LHS)); |
1753 | } |
1754 | }; |
1755 | |
1756 | /// Helper class for identifying signed max predicates. |
1757 | struct smax_pred_ty { |
1758 | static bool match(ICmpInst::Predicate Pred) { |
1759 | return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE; |
1760 | } |
1761 | }; |
1762 | |
1763 | /// Helper class for identifying signed min predicates. |
1764 | struct smin_pred_ty { |
1765 | static bool match(ICmpInst::Predicate Pred) { |
1766 | return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE; |
1767 | } |
1768 | }; |
1769 | |
1770 | /// Helper class for identifying unsigned max predicates. |
1771 | struct umax_pred_ty { |
1772 | static bool match(ICmpInst::Predicate Pred) { |
1773 | return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE; |
1774 | } |
1775 | }; |
1776 | |
1777 | /// Helper class for identifying unsigned min predicates. |
1778 | struct umin_pred_ty { |
1779 | static bool match(ICmpInst::Predicate Pred) { |
1780 | return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE; |
1781 | } |
1782 | }; |
1783 | |
1784 | /// Helper class for identifying ordered max predicates. |
1785 | struct ofmax_pred_ty { |
1786 | static bool match(FCmpInst::Predicate Pred) { |
1787 | return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE; |
1788 | } |
1789 | }; |
1790 | |
1791 | /// Helper class for identifying ordered min predicates. |
1792 | struct ofmin_pred_ty { |
1793 | static bool match(FCmpInst::Predicate Pred) { |
1794 | return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE; |
1795 | } |
1796 | }; |
1797 | |
1798 | /// Helper class for identifying unordered max predicates. |
1799 | struct ufmax_pred_ty { |
1800 | static bool match(FCmpInst::Predicate Pred) { |
1801 | return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE; |
1802 | } |
1803 | }; |
1804 | |
1805 | /// Helper class for identifying unordered min predicates. |
1806 | struct ufmin_pred_ty { |
1807 | static bool match(FCmpInst::Predicate Pred) { |
1808 | return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE; |
1809 | } |
1810 | }; |
1811 | |
1812 | template <typename LHS, typename RHS> |
1813 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L, |
1814 | const RHS &R) { |
1815 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R); |
1816 | } |
1817 | |
1818 | template <typename LHS, typename RHS> |
1819 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L, |
1820 | const RHS &R) { |
1821 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R); |
1822 | } |
1823 | |
1824 | template <typename LHS, typename RHS> |
1825 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L, |
1826 | const RHS &R) { |
1827 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R); |
1828 | } |
1829 | |
1830 | template <typename LHS, typename RHS> |
1831 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L, |
1832 | const RHS &R) { |
1833 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R); |
1834 | } |
1835 | |
1836 | template <typename LHS, typename RHS> |
1837 | inline match_combine_or< |
1838 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>, |
1839 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>, |
1840 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>, |
1841 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>> |
1842 | m_MaxOrMin(const LHS &L, const RHS &R) { |
1843 | return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)), |
1844 | m_CombineOr(m_UMax(L, R), m_UMin(L, R))); |
1845 | } |
1846 | |
1847 | /// Match an 'ordered' floating point maximum function. |
1848 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
1849 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
1850 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' |
1851 | /// semantics. In the presence of 'NaN' we have to preserve the original |
1852 | /// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate. |
1853 | /// |
1854 | /// max(L, R) iff L and R are not NaN |
1855 | /// m_OrdFMax(L, R) = R iff L or R are NaN |
1856 | template <typename LHS, typename RHS> |
1857 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L, |
1858 | const RHS &R) { |
1859 | return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R); |
1860 | } |
1861 | |
1862 | /// Match an 'ordered' floating point minimum function. |
1863 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
1864 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
1865 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' |
1866 | /// semantics. In the presence of 'NaN' we have to preserve the original |
1867 | /// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate. |
1868 | /// |
1869 | /// min(L, R) iff L and R are not NaN |
1870 | /// m_OrdFMin(L, R) = R iff L or R are NaN |
1871 | template <typename LHS, typename RHS> |
1872 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L, |
1873 | const RHS &R) { |
1874 | return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R); |
1875 | } |
1876 | |
1877 | /// Match an 'unordered' floating point maximum function. |
1878 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
1879 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
1880 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' |
1881 | /// semantics. In the presence of 'NaN' we have to preserve the original |
1882 | /// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate. |
1883 | /// |
1884 | /// max(L, R) iff L and R are not NaN |
1885 | /// m_UnordFMax(L, R) = L iff L or R are NaN |
1886 | template <typename LHS, typename RHS> |
1887 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty> |
1888 | m_UnordFMax(const LHS &L, const RHS &R) { |
1889 | return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R); |
1890 | } |
1891 | |
1892 | /// Match an 'unordered' floating point minimum function. |
1893 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
1894 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
1895 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' |
1896 | /// semantics. In the presence of 'NaN' we have to preserve the original |
1897 | /// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate. |
1898 | /// |
1899 | /// min(L, R) iff L and R are not NaN |
1900 | /// m_UnordFMin(L, R) = L iff L or R are NaN |
1901 | template <typename LHS, typename RHS> |
1902 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty> |
1903 | m_UnordFMin(const LHS &L, const RHS &R) { |
1904 | return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R); |
1905 | } |
1906 | |
1907 | //===----------------------------------------------------------------------===// |
1908 | // Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b |
1909 | // Note that S might be matched to other instructions than AddInst. |
1910 | // |
1911 | |
1912 | template <typename LHS_t, typename RHS_t, typename Sum_t> |
1913 | struct UAddWithOverflow_match { |
1914 | LHS_t L; |
1915 | RHS_t R; |
1916 | Sum_t S; |
1917 | |
1918 | UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S) |
1919 | : L(L), R(R), S(S) {} |
1920 | |
1921 | template <typename OpTy> bool match(OpTy *V) { |
1922 | Value *ICmpLHS, *ICmpRHS; |
1923 | ICmpInst::Predicate Pred; |
1924 | if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V)) |
1925 | return false; |
1926 | |
1927 | Value *AddLHS, *AddRHS; |
1928 | auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS)); |
1929 | |
1930 | // (a + b) u< a, (a + b) u< b |
1931 | if (Pred == ICmpInst::ICMP_ULT) |
1932 | if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS)) |
1933 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); |
1934 | |
1935 | // a >u (a + b), b >u (a + b) |
1936 | if (Pred == ICmpInst::ICMP_UGT) |
1937 | if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS)) |
1938 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); |
1939 | |
1940 | Value *Op1; |
1941 | auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes())); |
1942 | // (a ^ -1) <u b |
1943 | if (Pred == ICmpInst::ICMP_ULT) { |
1944 | if (XorExpr.match(ICmpLHS)) |
1945 | return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS); |
1946 | } |
1947 | // b > u (a ^ -1) |
1948 | if (Pred == ICmpInst::ICMP_UGT) { |
1949 | if (XorExpr.match(ICmpRHS)) |
1950 | return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS); |
1951 | } |
1952 | |
1953 | // Match special-case for increment-by-1. |
1954 | if (Pred == ICmpInst::ICMP_EQ) { |
1955 | // (a + 1) == 0 |
1956 | // (1 + a) == 0 |
1957 | if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) && |
1958 | (m_One().match(AddLHS) || m_One().match(AddRHS))) |
1959 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); |
1960 | // 0 == (a + 1) |
1961 | // 0 == (1 + a) |
1962 | if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) && |
1963 | (m_One().match(AddLHS) || m_One().match(AddRHS))) |
1964 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); |
1965 | } |
1966 | |
1967 | return false; |
1968 | } |
1969 | }; |
1970 | |
1971 | /// Match an icmp instruction checking for unsigned overflow on addition. |
1972 | /// |
1973 | /// S is matched to the addition whose result is being checked for overflow, and |
1974 | /// L and R are matched to the LHS and RHS of S. |
1975 | template <typename LHS_t, typename RHS_t, typename Sum_t> |
1976 | UAddWithOverflow_match<LHS_t, RHS_t, Sum_t> |
1977 | m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) { |
1978 | return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S); |
1979 | } |
1980 | |
1981 | template <typename Opnd_t> struct Argument_match { |
1982 | unsigned OpI; |
1983 | Opnd_t Val; |
1984 | |
1985 | Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {} |
1986 | |
1987 | template <typename OpTy> bool match(OpTy *V) { |
1988 | // FIXME: Should likely be switched to use `CallBase`. |
1989 | if (const auto *CI = dyn_cast<CallInst>(V)) |
1990 | return Val.match(CI->getArgOperand(OpI)); |
1991 | return false; |
1992 | } |
1993 | }; |
1994 | |
1995 | /// Match an argument. |
1996 | template <unsigned OpI, typename Opnd_t> |
1997 | inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) { |
1998 | return Argument_match<Opnd_t>(OpI, Op); |
1999 | } |
2000 | |
2001 | /// Intrinsic matchers. |
2002 | struct IntrinsicID_match { |
2003 | unsigned ID; |
2004 | |
2005 | IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {} |
2006 | |
2007 | template <typename OpTy> bool match(OpTy *V) { |
2008 | if (const auto *CI = dyn_cast<CallInst>(V)) |
2009 | if (const auto *F = CI->getCalledFunction()) |
2010 | return F->getIntrinsicID() == ID; |
2011 | return false; |
2012 | } |
2013 | }; |
2014 | |
2015 | /// Intrinsic matches are combinations of ID matchers, and argument |
2016 | /// matchers. Higher arity matcher are defined recursively in terms of and-ing |
2017 | /// them with lower arity matchers. Here's some convenient typedefs for up to |
2018 | /// several arguments, and more can be added as needed |
2019 | template <typename T0 = void, typename T1 = void, typename T2 = void, |
2020 | typename T3 = void, typename T4 = void, typename T5 = void, |
2021 | typename T6 = void, typename T7 = void, typename T8 = void, |
2022 | typename T9 = void, typename T10 = void> |
2023 | struct m_Intrinsic_Ty; |
2024 | template <typename T0> struct m_Intrinsic_Ty<T0> { |
2025 | using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>; |
2026 | }; |
2027 | template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> { |
2028 | using Ty = |
2029 | match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>; |
2030 | }; |
2031 | template <typename T0, typename T1, typename T2> |
2032 | struct m_Intrinsic_Ty<T0, T1, T2> { |
2033 | using Ty = |
2034 | match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty, |
2035 | Argument_match<T2>>; |
2036 | }; |
2037 | template <typename T0, typename T1, typename T2, typename T3> |
2038 | struct m_Intrinsic_Ty<T0, T1, T2, T3> { |
2039 | using Ty = |
2040 | match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty, |
2041 | Argument_match<T3>>; |
2042 | }; |
2043 | |
2044 | template <typename T0, typename T1, typename T2, typename T3, typename T4> |
2045 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> { |
2046 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty, |
2047 | Argument_match<T4>>; |
2048 | }; |
2049 | |
2050 | template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5> |
2051 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> { |
2052 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty, |
2053 | Argument_match<T5>>; |
2054 | }; |
2055 | |
2056 | /// Match intrinsic calls like this: |
2057 | /// m_Intrinsic<Intrinsic::fabs>(m_Value(X)) |
2058 | template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() { |
2059 | return IntrinsicID_match(IntrID); |
2060 | } |
2061 | |
2062 | template <Intrinsic::ID IntrID, typename T0> |
2063 | inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) { |
2064 | return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0)); |
2065 | } |
2066 | |
2067 | template <Intrinsic::ID IntrID, typename T0, typename T1> |
2068 | inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0, |
2069 | const T1 &Op1) { |
2070 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1)); |
2071 | } |
2072 | |
2073 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2> |
2074 | inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty |
2075 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) { |
2076 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2)); |
2077 | } |
2078 | |
2079 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
2080 | typename T3> |
2081 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty |
2082 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) { |
2083 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3)); |
2084 | } |
2085 | |
2086 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
2087 | typename T3, typename T4> |
2088 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty |
2089 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, |
2090 | const T4 &Op4) { |
2091 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3), |
2092 | m_Argument<4>(Op4)); |
2093 | } |
2094 | |
2095 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
2096 | typename T3, typename T4, typename T5> |
2097 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty |
2098 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, |
2099 | const T4 &Op4, const T5 &Op5) { |
2100 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4), |
2101 | m_Argument<5>(Op5)); |
2102 | } |
2103 | |
2104 | // Helper intrinsic matching specializations. |
2105 | template <typename Opnd0> |
2106 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) { |
2107 | return m_Intrinsic<Intrinsic::bitreverse>(Op0); |
2108 | } |
2109 | |
2110 | template <typename Opnd0> |
2111 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) { |
2112 | return m_Intrinsic<Intrinsic::bswap>(Op0); |
2113 | } |
2114 | |
2115 | template <typename Opnd0> |
2116 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) { |
2117 | return m_Intrinsic<Intrinsic::fabs>(Op0); |
2118 | } |
2119 | |
2120 | template <typename Opnd0> |
2121 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) { |
2122 | return m_Intrinsic<Intrinsic::canonicalize>(Op0); |
2123 | } |
2124 | |
2125 | template <typename Opnd0, typename Opnd1> |
2126 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0, |
2127 | const Opnd1 &Op1) { |
2128 | return m_Intrinsic<Intrinsic::minnum>(Op0, Op1); |
2129 | } |
2130 | |
2131 | template <typename Opnd0, typename Opnd1> |
2132 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0, |
2133 | const Opnd1 &Op1) { |
2134 | return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1); |
2135 | } |
2136 | |
2137 | template <typename Opnd0, typename Opnd1, typename Opnd2> |
2138 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty |
2139 | m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { |
2140 | return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2); |
2141 | } |
2142 | |
2143 | template <typename Opnd0, typename Opnd1, typename Opnd2> |
2144 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty |
2145 | m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { |
2146 | return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2); |
2147 | } |
2148 | |
2149 | //===----------------------------------------------------------------------===// |
2150 | // Matchers for two-operands operators with the operators in either order |
2151 | // |
2152 | |
2153 | /// Matches a BinaryOperator with LHS and RHS in either order. |
2154 | template <typename LHS, typename RHS> |
2155 | inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) { |
2156 | return AnyBinaryOp_match<LHS, RHS, true>(L, R); |
2157 | } |
2158 | |
2159 | /// Matches an ICmp with a predicate over LHS and RHS in either order. |
2160 | /// Swaps the predicate if operands are commuted. |
2161 | template <typename LHS, typename RHS> |
2162 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true> |
2163 | m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
2164 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L, |
2165 | R); |
2166 | } |
2167 | |
2168 | /// Matches a Add with LHS and RHS in either order. |
2169 | template <typename LHS, typename RHS> |
2170 | inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L, |
2171 | const RHS &R) { |
2172 | return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R); |
2173 | } |
2174 | |
2175 | /// Matches a Mul with LHS and RHS in either order. |
2176 | template <typename LHS, typename RHS> |
2177 | inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L, |
2178 | const RHS &R) { |
2179 | return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R); |
2180 | } |
2181 | |
2182 | /// Matches an And with LHS and RHS in either order. |
2183 | template <typename LHS, typename RHS> |
2184 | inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L, |
2185 | const RHS &R) { |
2186 | return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R); |
2187 | } |
2188 | |
2189 | /// Matches an Or with LHS and RHS in either order. |
2190 | template <typename LHS, typename RHS> |
2191 | inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L, |
2192 | const RHS &R) { |
2193 | return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R); |
2194 | } |
2195 | |
2196 | /// Matches an Xor with LHS and RHS in either order. |
2197 | template <typename LHS, typename RHS> |
2198 | inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L, |
2199 | const RHS &R) { |
2200 | return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R); |
2201 | } |
2202 | |
2203 | /// Matches a 'Neg' as 'sub 0, V'. |
2204 | template <typename ValTy> |
2205 | inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub> |
2206 | m_Neg(const ValTy &V) { |
2207 | return m_Sub(m_ZeroInt(), V); |
2208 | } |
2209 | |
2210 | /// Matches a 'Neg' as 'sub nsw 0, V'. |
2211 | template <typename ValTy> |
2212 | inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, |
2213 | Instruction::Sub, |
2214 | OverflowingBinaryOperator::NoSignedWrap> |
2215 | m_NSWNeg(const ValTy &V) { |
2216 | return m_NSWSub(m_ZeroInt(), V); |
2217 | } |
2218 | |
2219 | /// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'. |
2220 | template <typename ValTy> |
2221 | inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true> |
2222 | m_Not(const ValTy &V) { |
2223 | return m_c_Xor(V, m_AllOnes()); |
2224 | } |
2225 | |
2226 | /// Matches an SMin with LHS and RHS in either order. |
2227 | template <typename LHS, typename RHS> |
2228 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true> |
2229 | m_c_SMin(const LHS &L, const RHS &R) { |
2230 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R); |
2231 | } |
2232 | /// Matches an SMax with LHS and RHS in either order. |
2233 | template <typename LHS, typename RHS> |
2234 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true> |
2235 | m_c_SMax(const LHS &L, const RHS &R) { |
2236 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R); |
2237 | } |
2238 | /// Matches a UMin with LHS and RHS in either order. |
2239 | template <typename LHS, typename RHS> |
2240 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true> |
2241 | m_c_UMin(const LHS &L, const RHS &R) { |
2242 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R); |
2243 | } |
2244 | /// Matches a UMax with LHS and RHS in either order. |
2245 | template <typename LHS, typename RHS> |
2246 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true> |
2247 | m_c_UMax(const LHS &L, const RHS &R) { |
2248 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R); |
2249 | } |
2250 | |
2251 | template <typename LHS, typename RHS> |
2252 | inline match_combine_or< |
2253 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>, |
2254 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>, |
2255 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>, |
2256 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>> |
2257 | m_c_MaxOrMin(const LHS &L, const RHS &R) { |
2258 | return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)), |
2259 | m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R))); |
2260 | } |
2261 | |
2262 | /// Matches FAdd with LHS and RHS in either order. |
2263 | template <typename LHS, typename RHS> |
2264 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true> |
2265 | m_c_FAdd(const LHS &L, const RHS &R) { |
2266 | return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R); |
2267 | } |
2268 | |
2269 | /// Matches FMul with LHS and RHS in either order. |
2270 | template <typename LHS, typename RHS> |
2271 | inline BinaryOp_match<LHS, RHS, Instruction::FMul, true> |
2272 | m_c_FMul(const LHS &L, const RHS &R) { |
2273 | return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R); |
2274 | } |
2275 | |
2276 | template <typename Opnd_t> struct Signum_match { |
2277 | Opnd_t Val; |
2278 | Signum_match(const Opnd_t &V) : Val(V) {} |
2279 | |
2280 | template <typename OpTy> bool match(OpTy *V) { |
2281 | unsigned TypeSize = V->getType()->getScalarSizeInBits(); |
2282 | if (TypeSize == 0) |
2283 | return false; |
2284 | |
2285 | unsigned ShiftWidth = TypeSize - 1; |
2286 | Value *OpL = nullptr, *OpR = nullptr; |
2287 | |
2288 | // This is the representation of signum we match: |
2289 | // |
2290 | // signum(x) == (x >> 63) | (-x >>u 63) |
2291 | // |
2292 | // An i1 value is its own signum, so it's correct to match |
2293 | // |
2294 | // signum(x) == (x >> 0) | (-x >>u 0) |
2295 | // |
2296 | // for i1 values. |
2297 | |
2298 | auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth)); |
2299 | auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth)); |
2300 | auto Signum = m_Or(LHS, RHS); |
2301 | |
2302 | return Signum.match(V) && OpL == OpR && Val.match(OpL); |
2303 | } |
2304 | }; |
2305 | |
2306 | /// Matches a signum pattern. |
2307 | /// |
2308 | /// signum(x) = |
2309 | /// x > 0 -> 1 |
2310 | /// x == 0 -> 0 |
2311 | /// x < 0 -> -1 |
2312 | template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) { |
2313 | return Signum_match<Val_t>(V); |
2314 | } |
2315 | |
2316 | template <int Ind, typename Opnd_t> struct ExtractValue_match { |
2317 | Opnd_t Val; |
2318 | ExtractValue_match(const Opnd_t &V) : Val(V) {} |
2319 | |
2320 | template <typename OpTy> bool match(OpTy *V) { |
2321 | if (auto *I = dyn_cast<ExtractValueInst>(V)) { |
2322 | // If Ind is -1, don't inspect indices |
2323 | if (Ind != -1 && |
2324 | !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind)) |
2325 | return false; |
2326 | return Val.match(I->getAggregateOperand()); |
2327 | } |
2328 | return false; |
2329 | } |
2330 | }; |
2331 | |
2332 | /// Match a single index ExtractValue instruction. |
2333 | /// For example m_ExtractValue<1>(...) |
2334 | template <int Ind, typename Val_t> |
2335 | inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) { |
2336 | return ExtractValue_match<Ind, Val_t>(V); |
2337 | } |
2338 | |
2339 | /// Match an ExtractValue instruction with any index. |
2340 | /// For example m_ExtractValue(...) |
2341 | template <typename Val_t> |
2342 | inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) { |
2343 | return ExtractValue_match<-1, Val_t>(V); |
2344 | } |
2345 | |
2346 | /// Matcher for a single index InsertValue instruction. |
2347 | template <int Ind, typename T0, typename T1> struct InsertValue_match { |
2348 | T0 Op0; |
2349 | T1 Op1; |
2350 | |
2351 | InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {} |
2352 | |
2353 | template <typename OpTy> bool match(OpTy *V) { |
2354 | if (auto *I = dyn_cast<InsertValueInst>(V)) { |
2355 | return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) && |
2356 | I->getNumIndices() == 1 && Ind == I->getIndices()[0]; |
2357 | } |
2358 | return false; |
2359 | } |
2360 | }; |
2361 | |
2362 | /// Matches a single index InsertValue instruction. |
2363 | template <int Ind, typename Val_t, typename Elt_t> |
2364 | inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val, |
2365 | const Elt_t &Elt) { |
2366 | return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt); |
2367 | } |
2368 | |
2369 | /// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or |
2370 | /// the constant expression |
2371 | /// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>` |
2372 | /// under the right conditions determined by DataLayout. |
2373 | struct VScaleVal_match { |
2374 | private: |
2375 | template <typename Base, typename Offset> |
2376 | inline BinaryOp_match<Base, Offset, Instruction::GetElementPtr> |
2377 | m_OffsetGep(const Base &B, const Offset &O) { |
2378 | return BinaryOp_match<Base, Offset, Instruction::GetElementPtr>(B, O); |
2379 | } |
2380 | |
2381 | public: |
2382 | const DataLayout &DL; |
2383 | VScaleVal_match(const DataLayout &DL) : DL(DL) {} |
2384 | |
2385 | template <typename ITy> bool match(ITy *V) { |
2386 | if (m_Intrinsic<Intrinsic::vscale>().match(V)) |
2387 | return true; |
2388 | |
2389 | if (m_PtrToInt(m_OffsetGep(m_Zero(), m_SpecificInt(1))).match(V)) { |
2390 | Type *PtrTy = cast<Operator>(V)->getOperand(0)->getType(); |
2391 | auto *DerefTy = PtrTy->getPointerElementType(); |
2392 | if (isa<ScalableVectorType>(DerefTy) && |
2393 | DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8) |
2394 | return true; |
2395 | } |
2396 | |
2397 | return false; |
2398 | } |
2399 | }; |
2400 | |
2401 | inline VScaleVal_match m_VScale(const DataLayout &DL) { |
2402 | return VScaleVal_match(DL); |
2403 | } |
2404 | |
2405 | template <typename LHS, typename RHS, unsigned Opcode> |
2406 | struct LogicalOp_match { |
2407 | LHS L; |
2408 | RHS R; |
2409 | |
2410 | LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {} |
2411 | |
2412 | template <typename T> bool match(T *V) { |
2413 | if (auto *I = dyn_cast<Instruction>(V)) { |
2414 | if (!I->getType()->isIntOrIntVectorTy(1)) |
2415 | return false; |
2416 | |
2417 | if (I->getOpcode() == Opcode && L.match(I->getOperand(0)) && |
2418 | R.match(I->getOperand(1))) |
2419 | return true; |
2420 | |
2421 | if (auto *SI = dyn_cast<SelectInst>(I)) { |
2422 | if (Opcode == Instruction::And) { |
2423 | if (const auto *C = dyn_cast<Constant>(SI->getFalseValue())) |
2424 | if (C->isNullValue() && L.match(SI->getCondition()) && |
2425 | R.match(SI->getTrueValue())) |
2426 | return true; |
2427 | } else { |
2428 | assert(Opcode == Instruction::Or)((Opcode == Instruction::Or) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::Or", "/build/llvm-toolchain-snapshot-13~++20210314100619+a28facba1ccd/llvm/include/llvm/IR/PatternMatch.h" , 2428, __PRETTY_FUNCTION__)); |
2429 | if (const auto *C = dyn_cast<Constant>(SI->getTrueValue())) |
2430 | if (C->isOneValue() && L.match(SI->getCondition()) && |
2431 | R.match(SI->getFalseValue())) |
2432 | return true; |
2433 | } |
2434 | } |
2435 | } |
2436 | |
2437 | return false; |
2438 | } |
2439 | }; |
2440 | |
2441 | /// Matches L && R either in the form of L & R or L ? R : false. |
2442 | /// Note that the latter form is poison-blocking. |
2443 | template <typename LHS, typename RHS> |
2444 | inline LogicalOp_match<LHS, RHS, Instruction::And> |
2445 | m_LogicalAnd(const LHS &L, const RHS &R) { |
2446 | return LogicalOp_match<LHS, RHS, Instruction::And>(L, R); |
2447 | } |
2448 | |
2449 | /// Matches L && R where L and R are arbitrary values. |
2450 | inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); } |
2451 | |
2452 | /// Matches L || R either in the form of L | R or L ? true : R. |
2453 | /// Note that the latter form is poison-blocking. |
2454 | template <typename LHS, typename RHS> |
2455 | inline LogicalOp_match<LHS, RHS, Instruction::Or> |
2456 | m_LogicalOr(const LHS &L, const RHS &R) { |
2457 | return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R); |
2458 | } |
2459 | |
2460 | /// Matches L || R where L and R are arbitrary values. |
2461 | inline auto m_LogicalOr() { |
2462 | return m_LogicalOr(m_Value(), m_Value()); |
2463 | } |
2464 | |
2465 | } // end namespace PatternMatch |
2466 | } // end namespace llvm |
2467 | |
2468 | #endif // LLVM_IR_PATTERNMATCH_H |