| File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp |
| Warning: | line 197, column 7 1st function call argument is an uninitialized value |
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/InstructionSimplify.h" | ||||
| 15 | #include "llvm/IR/IntrinsicInst.h" | ||||
| 16 | #include "llvm/IR/PatternMatch.h" | ||||
| 17 | #include "llvm/Transforms/InstCombine/InstCombiner.h" | ||||
| 18 | using namespace llvm; | ||||
| 19 | using namespace PatternMatch; | ||||
| 20 | |||||
| 21 | #define DEBUG_TYPE"instcombine" "instcombine" | ||||
| 22 | |||||
| 23 | bool canTryToConstantAddTwoShiftAmounts(Value *Sh0, Value *ShAmt0, Value *Sh1, | ||||
| 24 | Value *ShAmt1) { | ||||
| 25 | // We have two shift amounts from two different shifts. The types of those | ||||
| 26 | // shift amounts may not match. If that's the case let's bailout now.. | ||||
| 27 | if (ShAmt0->getType() != ShAmt1->getType()) | ||||
| 28 | return false; | ||||
| 29 | |||||
| 30 | // As input, we have the following pattern: | ||||
| 31 | // Sh0 (Sh1 X, Q), K | ||||
| 32 | // We want to rewrite that as: | ||||
| 33 | // Sh x, (Q+K) iff (Q+K) u< bitwidth(x) | ||||
| 34 | // While we know that originally (Q+K) would not overflow | ||||
| 35 | // (because 2 * (N-1) u<= iN -1), we have looked past extensions of | ||||
| 36 | // shift amounts. so it may now overflow in smaller bitwidth. | ||||
| 37 | // To ensure that does not happen, we need to ensure that the total maximal | ||||
| 38 | // shift amount is still representable in that smaller bit width. | ||||
| 39 | unsigned MaximalPossibleTotalShiftAmount = | ||||
| 40 | (Sh0->getType()->getScalarSizeInBits() - 1) + | ||||
| 41 | (Sh1->getType()->getScalarSizeInBits() - 1); | ||||
| 42 | APInt MaximalRepresentableShiftAmount = | ||||
| 43 | APInt::getAllOnes(ShAmt0->getType()->getScalarSizeInBits()); | ||||
| 44 | return MaximalRepresentableShiftAmount.uge(MaximalPossibleTotalShiftAmount); | ||||
| 45 | } | ||||
| 46 | |||||
| 47 | // Given pattern: | ||||
| 48 | // (x shiftopcode Q) shiftopcode K | ||||
| 49 | // we should rewrite it as | ||||
| 50 | // x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and | ||||
| 51 | // | ||||
| 52 | // This is valid for any shift, but they must be identical, and we must be | ||||
| 53 | // careful in case we have (zext(Q)+zext(K)) and look past extensions, | ||||
| 54 | // (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus. | ||||
| 55 | // | ||||
| 56 | // AnalyzeForSignBitExtraction indicates that we will only analyze whether this | ||||
| 57 | // pattern has any 2 right-shifts that sum to 1 less than original bit width. | ||||
| 58 | Value *InstCombinerImpl::reassociateShiftAmtsOfTwoSameDirectionShifts( | ||||
| 59 | BinaryOperator *Sh0, const SimplifyQuery &SQ, | ||||
| 60 | bool AnalyzeForSignBitExtraction) { | ||||
| 61 | // Look for a shift of some instruction, ignore zext of shift amount if any. | ||||
| 62 | Instruction *Sh0Op0; | ||||
| 63 | Value *ShAmt0; | ||||
| 64 | if (!match(Sh0, | ||||
| 65 | m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0))))) | ||||
| 66 | return nullptr; | ||||
| 67 | |||||
| 68 | // If there is a truncation between the two shifts, we must make note of it | ||||
| 69 | // and look through it. The truncation imposes additional constraints on the | ||||
| 70 | // transform. | ||||
| 71 | Instruction *Sh1; | ||||
| 72 | Value *Trunc = nullptr; | ||||
| 73 | match(Sh0Op0, | ||||
| 74 | m_CombineOr(m_CombineAnd(m_Trunc(m_Instruction(Sh1)), m_Value(Trunc)), | ||||
| 75 | m_Instruction(Sh1))); | ||||
| 76 | |||||
| 77 | // Inner shift: (x shiftopcode ShAmt1) | ||||
| 78 | // Like with other shift, ignore zext of shift amount if any. | ||||
| 79 | Value *X, *ShAmt1; | ||||
| 80 | if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1))))) | ||||
| 81 | return nullptr; | ||||
| 82 | |||||
| 83 | // Verify that it would be safe to try to add those two shift amounts. | ||||
| 84 | if (!canTryToConstantAddTwoShiftAmounts(Sh0, ShAmt0, Sh1, ShAmt1)) | ||||
| 85 | return nullptr; | ||||
| 86 | |||||
| 87 | // We are only looking for signbit extraction if we have two right shifts. | ||||
| 88 | bool HadTwoRightShifts = match(Sh0, m_Shr(m_Value(), m_Value())) && | ||||
| 89 | match(Sh1, m_Shr(m_Value(), m_Value())); | ||||
| 90 | // ... and if it's not two right-shifts, we know the answer already. | ||||
| 91 | if (AnalyzeForSignBitExtraction && !HadTwoRightShifts) | ||||
| 92 | return nullptr; | ||||
| 93 | |||||
| 94 | // The shift opcodes must be identical, unless we are just checking whether | ||||
| 95 | // this pattern can be interpreted as a sign-bit-extraction. | ||||
| 96 | Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode(); | ||||
| 97 | bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode(); | ||||
| 98 | if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction) | ||||
| 99 | return nullptr; | ||||
| 100 | |||||
| 101 | // If we saw truncation, we'll need to produce extra instruction, | ||||
| 102 | // and for that one of the operands of the shift must be one-use, | ||||
| 103 | // unless of course we don't actually plan to produce any instructions here. | ||||
| 104 | if (Trunc && !AnalyzeForSignBitExtraction && | ||||
| 105 | !match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value()))) | ||||
| 106 | return nullptr; | ||||
| 107 | |||||
| 108 | // Can we fold (ShAmt0+ShAmt1) ? | ||||
| 109 | auto *NewShAmt = dyn_cast_or_null<Constant>( | ||||
| 110 | SimplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false, | ||||
| 111 | SQ.getWithInstruction(Sh0))); | ||||
| 112 | if (!NewShAmt) | ||||
| 113 | return nullptr; // Did not simplify. | ||||
| 114 | unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits(); | ||||
| 115 | unsigned XBitWidth = X->getType()->getScalarSizeInBits(); | ||||
| 116 | // Is the new shift amount smaller than the bit width of inner/new shift? | ||||
| 117 | if (!match(NewShAmt, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT, | ||||
| 118 | APInt(NewShAmtBitWidth, XBitWidth)))) | ||||
| 119 | return nullptr; // FIXME: could perform constant-folding. | ||||
| 120 | |||||
| 121 | // If there was a truncation, and we have a right-shift, we can only fold if | ||||
| 122 | // we are left with the original sign bit. Likewise, if we were just checking | ||||
| 123 | // that this is a sighbit extraction, this is the place to check it. | ||||
| 124 | // FIXME: zero shift amount is also legal here, but we can't *easily* check | ||||
| 125 | // more than one predicate so it's not really worth it. | ||||
| 126 | if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) { | ||||
| 127 | // If it's not a sign bit extraction, then we're done. | ||||
| 128 | if (!match(NewShAmt, | ||||
| 129 | m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, | ||||
| 130 | APInt(NewShAmtBitWidth, XBitWidth - 1)))) | ||||
| 131 | return nullptr; | ||||
| 132 | // If it is, and that was the question, return the base value. | ||||
| 133 | if (AnalyzeForSignBitExtraction) | ||||
| 134 | return X; | ||||
| 135 | } | ||||
| 136 | |||||
| 137 | assert(IdenticalShOpcodes && "Should not get here with different shifts.")(static_cast <bool> (IdenticalShOpcodes && "Should not get here with different shifts." ) ? void (0) : __assert_fail ("IdenticalShOpcodes && \"Should not get here with different shifts.\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 137 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 138 | |||||
| 139 | // All good, we can do this fold. | ||||
| 140 | NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, X->getType()); | ||||
| 141 | |||||
| 142 | BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt); | ||||
| 143 | |||||
| 144 | // The flags can only be propagated if there wasn't a trunc. | ||||
| 145 | if (!Trunc) { | ||||
| 146 | // If the pattern did not involve trunc, and both of the original shifts | ||||
| 147 | // had the same flag set, preserve the flag. | ||||
| 148 | if (ShiftOpcode == Instruction::BinaryOps::Shl) { | ||||
| 149 | NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() && | ||||
| 150 | Sh1->hasNoUnsignedWrap()); | ||||
| 151 | NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() && | ||||
| 152 | Sh1->hasNoSignedWrap()); | ||||
| 153 | } else { | ||||
| 154 | NewShift->setIsExact(Sh0->isExact() && Sh1->isExact()); | ||||
| 155 | } | ||||
| 156 | } | ||||
| 157 | |||||
| 158 | Instruction *Ret = NewShift; | ||||
| 159 | if (Trunc) { | ||||
| 160 | Builder.Insert(NewShift); | ||||
| 161 | Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType()); | ||||
| 162 | } | ||||
| 163 | |||||
| 164 | return Ret; | ||||
| 165 | } | ||||
| 166 | |||||
| 167 | // If we have some pattern that leaves only some low bits set, and then performs | ||||
| 168 | // left-shift of those bits, if none of the bits that are left after the final | ||||
| 169 | // shift are modified by the mask, we can omit the mask. | ||||
| 170 | // | ||||
| 171 | // There are many variants to this pattern: | ||||
| 172 | // a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt | ||||
| 173 | // b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt | ||||
| 174 | // c) (x & (-1 l>> MaskShAmt)) << ShiftShAmt | ||||
| 175 | // d) (x & ((-1 << MaskShAmt) l>> MaskShAmt)) << ShiftShAmt | ||||
| 176 | // e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt | ||||
| 177 | // f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt | ||||
| 178 | // All these patterns can be simplified to just: | ||||
| 179 | // x << ShiftShAmt | ||||
| 180 | // iff: | ||||
| 181 | // a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x) | ||||
| 182 | // c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt) | ||||
| 183 | static Instruction * | ||||
| 184 | dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift, | ||||
| 185 | const SimplifyQuery &Q, | ||||
| 186 | InstCombiner::BuilderTy &Builder) { | ||||
| 187 | assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&(static_cast <bool> (OuterShift->getOpcode() == Instruction ::BinaryOps::Shl && "The input must be 'shl'!") ? void (0) : __assert_fail ("OuterShift->getOpcode() == Instruction::BinaryOps::Shl && \"The input must be 'shl'!\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 188 , __extension__ __PRETTY_FUNCTION__)) | ||||
| 188 | "The input must be 'shl'!")(static_cast <bool> (OuterShift->getOpcode() == Instruction ::BinaryOps::Shl && "The input must be 'shl'!") ? void (0) : __assert_fail ("OuterShift->getOpcode() == Instruction::BinaryOps::Shl && \"The input must be 'shl'!\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 188 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 189 | |||||
| 190 | Value *Masked, *ShiftShAmt; | ||||
| 191 | match(OuterShift, | ||||
| |||||
| |||||
| 192 | m_Shift(m_Value(Masked), m_ZExtOrSelf(m_Value(ShiftShAmt)))); | ||||
| 193 | |||||
| 194 | // *If* there is a truncation between an outer shift and a possibly-mask, | ||||
| 195 | // then said truncation *must* be one-use, else we can't perform the fold. | ||||
| 196 | Value *Trunc; | ||||
| 197 | if (match(Masked, m_CombineAnd(m_Trunc(m_Value(Masked)), m_Value(Trunc))) && | ||||
| |||||
| 198 | !Trunc->hasOneUse()) | ||||
| 199 | return nullptr; | ||||
| 200 | |||||
| 201 | Type *NarrowestTy = OuterShift->getType(); | ||||
| 202 | Type *WidestTy = Masked->getType(); | ||||
| 203 | bool HadTrunc = WidestTy != NarrowestTy; | ||||
| 204 | |||||
| 205 | // The mask must be computed in a type twice as wide to ensure | ||||
| 206 | // that no bits are lost if the sum-of-shifts is wider than the base type. | ||||
| 207 | Type *ExtendedTy = WidestTy->getExtendedType(); | ||||
| 208 | |||||
| 209 | Value *MaskShAmt; | ||||
| 210 | |||||
| 211 | // ((1 << MaskShAmt) - 1) | ||||
| 212 | auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes()); | ||||
| 213 | // (~(-1 << maskNbits)) | ||||
| 214 | auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes()); | ||||
| 215 | // (-1 l>> MaskShAmt) | ||||
| 216 | auto MaskC = m_LShr(m_AllOnes(), m_Value(MaskShAmt)); | ||||
| 217 | // ((-1 << MaskShAmt) l>> MaskShAmt) | ||||
| 218 | auto MaskD = | ||||
| 219 | m_LShr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt)); | ||||
| 220 | |||||
| 221 | Value *X; | ||||
| 222 | Constant *NewMask; | ||||
| 223 | |||||
| 224 | if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) { | ||||
| 225 | // Peek through an optional zext of the shift amount. | ||||
| 226 | match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt))); | ||||
| 227 | |||||
| 228 | // Verify that it would be safe to try to add those two shift amounts. | ||||
| 229 | if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked, | ||||
| 230 | MaskShAmt)) | ||||
| 231 | return nullptr; | ||||
| 232 | |||||
| 233 | // Can we simplify (MaskShAmt+ShiftShAmt) ? | ||||
| 234 | auto *SumOfShAmts = dyn_cast_or_null<Constant>(SimplifyAddInst( | ||||
| 235 | MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); | ||||
| 236 | if (!SumOfShAmts) | ||||
| 237 | return nullptr; // Did not simplify. | ||||
| 238 | // In this pattern SumOfShAmts correlates with the number of low bits | ||||
| 239 | // that shall remain in the root value (OuterShift). | ||||
| 240 | |||||
| 241 | // An extend of an undef value becomes zero because the high bits are never | ||||
| 242 | // completely unknown. Replace the `undef` shift amounts with final | ||||
| 243 | // shift bitwidth to ensure that the value remains undef when creating the | ||||
| 244 | // subsequent shift op. | ||||
| 245 | SumOfShAmts = Constant::replaceUndefsWith( | ||||
| 246 | SumOfShAmts, ConstantInt::get(SumOfShAmts->getType()->getScalarType(), | ||||
| 247 | ExtendedTy->getScalarSizeInBits())); | ||||
| 248 | auto *ExtendedSumOfShAmts = ConstantExpr::getZExt(SumOfShAmts, ExtendedTy); | ||||
| 249 | // And compute the mask as usual: ~(-1 << (SumOfShAmts)) | ||||
| 250 | auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy); | ||||
| 251 | auto *ExtendedInvertedMask = | ||||
| 252 | ConstantExpr::getShl(ExtendedAllOnes, ExtendedSumOfShAmts); | ||||
| 253 | NewMask = ConstantExpr::getNot(ExtendedInvertedMask); | ||||
| 254 | } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) || | ||||
| 255 | match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)), | ||||
| 256 | m_Deferred(MaskShAmt)))) { | ||||
| 257 | // Peek through an optional zext of the shift amount. | ||||
| 258 | match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt))); | ||||
| 259 | |||||
| 260 | // Verify that it would be safe to try to add those two shift amounts. | ||||
| 261 | if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked, | ||||
| 262 | MaskShAmt)) | ||||
| 263 | return nullptr; | ||||
| 264 | |||||
| 265 | // Can we simplify (ShiftShAmt-MaskShAmt) ? | ||||
| 266 | auto *ShAmtsDiff = dyn_cast_or_null<Constant>(SimplifySubInst( | ||||
| 267 | ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); | ||||
| 268 | if (!ShAmtsDiff) | ||||
| 269 | return nullptr; // Did not simplify. | ||||
| 270 | // In this pattern ShAmtsDiff correlates with the number of high bits that | ||||
| 271 | // shall be unset in the root value (OuterShift). | ||||
| 272 | |||||
| 273 | // An extend of an undef value becomes zero because the high bits are never | ||||
| 274 | // completely unknown. Replace the `undef` shift amounts with negated | ||||
| 275 | // bitwidth of innermost shift to ensure that the value remains undef when | ||||
| 276 | // creating the subsequent shift op. | ||||
| 277 | unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits(); | ||||
| 278 | ShAmtsDiff = Constant::replaceUndefsWith( | ||||
| 279 | ShAmtsDiff, ConstantInt::get(ShAmtsDiff->getType()->getScalarType(), | ||||
| 280 | -WidestTyBitWidth)); | ||||
| 281 | auto *ExtendedNumHighBitsToClear = ConstantExpr::getZExt( | ||||
| 282 | ConstantExpr::getSub(ConstantInt::get(ShAmtsDiff->getType(), | ||||
| 283 | WidestTyBitWidth, | ||||
| 284 | /*isSigned=*/false), | ||||
| 285 | ShAmtsDiff), | ||||
| 286 | ExtendedTy); | ||||
| 287 | // And compute the mask as usual: (-1 l>> (NumHighBitsToClear)) | ||||
| 288 | auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy); | ||||
| 289 | NewMask = | ||||
| 290 | ConstantExpr::getLShr(ExtendedAllOnes, ExtendedNumHighBitsToClear); | ||||
| 291 | } else | ||||
| 292 | return nullptr; // Don't know anything about this pattern. | ||||
| 293 | |||||
| 294 | NewMask = ConstantExpr::getTrunc(NewMask, NarrowestTy); | ||||
| 295 | |||||
| 296 | // Does this mask has any unset bits? If not then we can just not apply it. | ||||
| 297 | bool NeedMask = !match(NewMask, m_AllOnes()); | ||||
| 298 | |||||
| 299 | // If we need to apply a mask, there are several more restrictions we have. | ||||
| 300 | if (NeedMask) { | ||||
| 301 | // The old masking instruction must go away. | ||||
| 302 | if (!Masked->hasOneUse()) | ||||
| 303 | return nullptr; | ||||
| 304 | // The original "masking" instruction must not have been`ashr`. | ||||
| 305 | if (match(Masked, m_AShr(m_Value(), m_Value()))) | ||||
| 306 | return nullptr; | ||||
| 307 | } | ||||
| 308 | |||||
| 309 | // If we need to apply truncation, let's do it first, since we can. | ||||
| 310 | // We have already ensured that the old truncation will go away. | ||||
| 311 | if (HadTrunc) | ||||
| 312 | X = Builder.CreateTrunc(X, NarrowestTy); | ||||
| 313 | |||||
| 314 | // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits. | ||||
| 315 | // We didn't change the Type of this outermost shift, so we can just do it. | ||||
| 316 | auto *NewShift = BinaryOperator::Create(OuterShift->getOpcode(), X, | ||||
| 317 | OuterShift->getOperand(1)); | ||||
| 318 | if (!NeedMask) | ||||
| 319 | return NewShift; | ||||
| 320 | |||||
| 321 | Builder.Insert(NewShift); | ||||
| 322 | return BinaryOperator::Create(Instruction::And, NewShift, NewMask); | ||||
| 323 | } | ||||
| 324 | |||||
| 325 | /// If we have a shift-by-constant of a bitwise logic op that itself has a | ||||
| 326 | /// shift-by-constant operand with identical opcode, we may be able to convert | ||||
| 327 | /// that into 2 independent shifts followed by the logic op. This eliminates a | ||||
| 328 | /// a use of an intermediate value (reduces dependency chain). | ||||
| 329 | static Instruction *foldShiftOfShiftedLogic(BinaryOperator &I, | ||||
| 330 | InstCombiner::BuilderTy &Builder) { | ||||
| 331 | assert(I.isShift() && "Expected a shift as input")(static_cast <bool> (I.isShift() && "Expected a shift as input" ) ? void (0) : __assert_fail ("I.isShift() && \"Expected a shift as input\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 331 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 332 | auto *LogicInst = dyn_cast<BinaryOperator>(I.getOperand(0)); | ||||
| 333 | if (!LogicInst || !LogicInst->isBitwiseLogicOp() || !LogicInst->hasOneUse()) | ||||
| 334 | return nullptr; | ||||
| 335 | |||||
| 336 | Constant *C0, *C1; | ||||
| 337 | if (!match(I.getOperand(1), m_Constant(C1))) | ||||
| 338 | return nullptr; | ||||
| 339 | |||||
| 340 | Instruction::BinaryOps ShiftOpcode = I.getOpcode(); | ||||
| 341 | Type *Ty = I.getType(); | ||||
| 342 | |||||
| 343 | // Find a matching one-use shift by constant. The fold is not valid if the sum | ||||
| 344 | // of the shift values equals or exceeds bitwidth. | ||||
| 345 | // TODO: Remove the one-use check if the other logic operand (Y) is constant. | ||||
| 346 | Value *X, *Y; | ||||
| 347 | auto matchFirstShift = [&](Value *V) { | ||||
| 348 | APInt Threshold(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits()); | ||||
| 349 | return match(V, m_BinOp(ShiftOpcode, m_Value(), m_Value())) && | ||||
| 350 | match(V, m_OneUse(m_Shift(m_Value(X), m_Constant(C0)))) && | ||||
| 351 | match(ConstantExpr::getAdd(C0, C1), | ||||
| 352 | m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold)); | ||||
| 353 | }; | ||||
| 354 | |||||
| 355 | // Logic ops are commutative, so check each operand for a match. | ||||
| 356 | if (matchFirstShift(LogicInst->getOperand(0))) | ||||
| 357 | Y = LogicInst->getOperand(1); | ||||
| 358 | else if (matchFirstShift(LogicInst->getOperand(1))) | ||||
| 359 | Y = LogicInst->getOperand(0); | ||||
| 360 | else | ||||
| 361 | return nullptr; | ||||
| 362 | |||||
| 363 | // shift (logic (shift X, C0), Y), C1 -> logic (shift X, C0+C1), (shift Y, C1) | ||||
| 364 | Constant *ShiftSumC = ConstantExpr::getAdd(C0, C1); | ||||
| 365 | Value *NewShift1 = Builder.CreateBinOp(ShiftOpcode, X, ShiftSumC); | ||||
| 366 | Value *NewShift2 = Builder.CreateBinOp(ShiftOpcode, Y, I.getOperand(1)); | ||||
| 367 | return BinaryOperator::Create(LogicInst->getOpcode(), NewShift1, NewShift2); | ||||
| 368 | } | ||||
| 369 | |||||
| 370 | Instruction *InstCombinerImpl::commonShiftTransforms(BinaryOperator &I) { | ||||
| 371 | if (Instruction *Phi = foldBinopWithPhiOperands(I)) | ||||
| 372 | return Phi; | ||||
| 373 | |||||
| 374 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
| 375 | assert(Op0->getType() == Op1->getType())(static_cast <bool> (Op0->getType() == Op1->getType ()) ? void (0) : __assert_fail ("Op0->getType() == Op1->getType()" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 375 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 376 | |||||
| 377 | // If the shift amount is a one-use `sext`, we can demote it to `zext`. | ||||
| 378 | Value *Y; | ||||
| 379 | if (match(Op1, m_OneUse(m_SExt(m_Value(Y))))) { | ||||
| 380 | Value *NewExt = Builder.CreateZExt(Y, I.getType(), Op1->getName()); | ||||
| 381 | return BinaryOperator::Create(I.getOpcode(), Op0, NewExt); | ||||
| 382 | } | ||||
| 383 | |||||
| 384 | // See if we can fold away this shift. | ||||
| 385 | if (SimplifyDemandedInstructionBits(I)) | ||||
| 386 | return &I; | ||||
| 387 | |||||
| 388 | // Try to fold constant and into select arguments. | ||||
| 389 | if (isa<Constant>(Op0)) | ||||
| 390 | if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) | ||||
| 391 | if (Instruction *R = FoldOpIntoSelect(I, SI)) | ||||
| 392 | return R; | ||||
| 393 | |||||
| 394 | if (Constant *CUI = dyn_cast<Constant>(Op1)) | ||||
| 395 | if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I)) | ||||
| 396 | return Res; | ||||
| 397 | |||||
| 398 | if (auto *NewShift = cast_or_null<Instruction>( | ||||
| 399 | reassociateShiftAmtsOfTwoSameDirectionShifts(&I, SQ))) | ||||
| 400 | return NewShift; | ||||
| 401 | |||||
| 402 | // C0 shift (A add C) -> (C0 shift C) shift A | ||||
| 403 | // iff A and C2 are both positive or the add has 'nuw'. | ||||
| 404 | Value *A; | ||||
| 405 | Constant *C; | ||||
| 406 | if (match(Op0, m_Constant()) && match(Op1, m_Add(m_Value(A), m_Constant(C)))) | ||||
| 407 | if (cast<BinaryOperator>(Op1)->hasNoUnsignedWrap() || | ||||
| 408 | (isKnownNonNegative(A, DL, 0, &AC, &I, &DT) && | ||||
| 409 | isKnownNonNegative(C, DL, 0, &AC, &I, &DT))) | ||||
| 410 | return BinaryOperator::Create( | ||||
| 411 | I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), Op0, C), A); | ||||
| 412 | |||||
| 413 | // X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2. | ||||
| 414 | // Because shifts by negative values (which could occur if A were negative) | ||||
| 415 | // are undefined. | ||||
| 416 | if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Constant(C))) && | ||||
| 417 | match(C, m_Power2())) { | ||||
| 418 | // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't | ||||
| 419 | // demand the sign bit (and many others) here?? | ||||
| 420 | Constant *Mask = ConstantExpr::getSub(C, ConstantInt::get(I.getType(), 1)); | ||||
| 421 | Value *Rem = Builder.CreateAnd(A, Mask, Op1->getName()); | ||||
| 422 | return replaceOperand(I, 1, Rem); | ||||
| 423 | } | ||||
| 424 | |||||
| 425 | if (Instruction *Logic = foldShiftOfShiftedLogic(I, Builder)) | ||||
| 426 | return Logic; | ||||
| 427 | |||||
| 428 | return nullptr; | ||||
| 429 | } | ||||
| 430 | |||||
| 431 | /// Return true if we can simplify two logical (either left or right) shifts | ||||
| 432 | /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2. | ||||
| 433 | static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl, | ||||
| 434 | Instruction *InnerShift, | ||||
| 435 | InstCombinerImpl &IC, Instruction *CxtI) { | ||||
| 436 | assert(InnerShift->isLogicalShift() && "Unexpected instruction type")(static_cast <bool> (InnerShift->isLogicalShift() && "Unexpected instruction type") ? void (0) : __assert_fail ("InnerShift->isLogicalShift() && \"Unexpected instruction type\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 436 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 437 | |||||
| 438 | // We need constant scalar or constant splat shifts. | ||||
| 439 | const APInt *InnerShiftConst; | ||||
| 440 | if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst))) | ||||
| 441 | return false; | ||||
| 442 | |||||
| 443 | // Two logical shifts in the same direction: | ||||
| 444 | // shl (shl X, C1), C2 --> shl X, C1 + C2 | ||||
| 445 | // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 | ||||
| 446 | bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; | ||||
| 447 | if (IsInnerShl == IsOuterShl) | ||||
| 448 | return true; | ||||
| 449 | |||||
| 450 | // Equal shift amounts in opposite directions become bitwise 'and': | ||||
| 451 | // lshr (shl X, C), C --> and X, C' | ||||
| 452 | // shl (lshr X, C), C --> and X, C' | ||||
| 453 | if (*InnerShiftConst == OuterShAmt) | ||||
| 454 | return true; | ||||
| 455 | |||||
| 456 | // If the 2nd shift is bigger than the 1st, we can fold: | ||||
| 457 | // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3 | ||||
| 458 | // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3 | ||||
| 459 | // but it isn't profitable unless we know the and'd out bits are already zero. | ||||
| 460 | // Also, check that the inner shift is valid (less than the type width) or | ||||
| 461 | // we'll crash trying to produce the bit mask for the 'and'. | ||||
| 462 | unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits(); | ||||
| 463 | if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) { | ||||
| 464 | unsigned InnerShAmt = InnerShiftConst->getZExtValue(); | ||||
| 465 | unsigned MaskShift = | ||||
| 466 | IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt; | ||||
| 467 | APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift; | ||||
| 468 | if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI)) | ||||
| 469 | return true; | ||||
| 470 | } | ||||
| 471 | |||||
| 472 | return false; | ||||
| 473 | } | ||||
| 474 | |||||
| 475 | /// See if we can compute the specified value, but shifted logically to the left | ||||
| 476 | /// or right by some number of bits. This should return true if the expression | ||||
| 477 | /// can be computed for the same cost as the current expression tree. This is | ||||
| 478 | /// used to eliminate extraneous shifting from things like: | ||||
| 479 | /// %C = shl i128 %A, 64 | ||||
| 480 | /// %D = shl i128 %B, 96 | ||||
| 481 | /// %E = or i128 %C, %D | ||||
| 482 | /// %F = lshr i128 %E, 64 | ||||
| 483 | /// where the client will ask if E can be computed shifted right by 64-bits. If | ||||
| 484 | /// this succeeds, getShiftedValue() will be called to produce the value. | ||||
| 485 | static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift, | ||||
| 486 | InstCombinerImpl &IC, Instruction *CxtI) { | ||||
| 487 | // We can always evaluate constants shifted. | ||||
| 488 | if (isa<Constant>(V)) | ||||
| 489 | return true; | ||||
| 490 | |||||
| 491 | Instruction *I = dyn_cast<Instruction>(V); | ||||
| 492 | if (!I) return false; | ||||
| 493 | |||||
| 494 | // We can't mutate something that has multiple uses: doing so would | ||||
| 495 | // require duplicating the instruction in general, which isn't profitable. | ||||
| 496 | if (!I->hasOneUse()) return false; | ||||
| 497 | |||||
| 498 | switch (I->getOpcode()) { | ||||
| 499 | default: return false; | ||||
| 500 | case Instruction::And: | ||||
| 501 | case Instruction::Or: | ||||
| 502 | case Instruction::Xor: | ||||
| 503 | // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. | ||||
| 504 | return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) && | ||||
| 505 | canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I); | ||||
| 506 | |||||
| 507 | case Instruction::Shl: | ||||
| 508 | case Instruction::LShr: | ||||
| 509 | return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI); | ||||
| 510 | |||||
| 511 | case Instruction::Select: { | ||||
| 512 | SelectInst *SI = cast<SelectInst>(I); | ||||
| 513 | Value *TrueVal = SI->getTrueValue(); | ||||
| 514 | Value *FalseVal = SI->getFalseValue(); | ||||
| 515 | return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) && | ||||
| 516 | canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI); | ||||
| 517 | } | ||||
| 518 | case Instruction::PHI: { | ||||
| 519 | // We can change a phi if we can change all operands. Note that we never | ||||
| 520 | // get into trouble with cyclic PHIs here because we only consider | ||||
| 521 | // instructions with a single use. | ||||
| 522 | PHINode *PN = cast<PHINode>(I); | ||||
| 523 | for (Value *IncValue : PN->incoming_values()) | ||||
| 524 | if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN)) | ||||
| 525 | return false; | ||||
| 526 | return true; | ||||
| 527 | } | ||||
| 528 | } | ||||
| 529 | } | ||||
| 530 | |||||
| 531 | /// Fold OuterShift (InnerShift X, C1), C2. | ||||
| 532 | /// See canEvaluateShiftedShift() for the constraints on these instructions. | ||||
| 533 | static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt, | ||||
| 534 | bool IsOuterShl, | ||||
| 535 | InstCombiner::BuilderTy &Builder) { | ||||
| 536 | bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; | ||||
| 537 | Type *ShType = InnerShift->getType(); | ||||
| 538 | unsigned TypeWidth = ShType->getScalarSizeInBits(); | ||||
| 539 | |||||
| 540 | // We only accept shifts-by-a-constant in canEvaluateShifted(). | ||||
| 541 | const APInt *C1; | ||||
| 542 | match(InnerShift->getOperand(1), m_APInt(C1)); | ||||
| 543 | unsigned InnerShAmt = C1->getZExtValue(); | ||||
| 544 | |||||
| 545 | // Change the shift amount and clear the appropriate IR flags. | ||||
| 546 | auto NewInnerShift = [&](unsigned ShAmt) { | ||||
| 547 | InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt)); | ||||
| 548 | if (IsInnerShl) { | ||||
| 549 | InnerShift->setHasNoUnsignedWrap(false); | ||||
| 550 | InnerShift->setHasNoSignedWrap(false); | ||||
| 551 | } else { | ||||
| 552 | InnerShift->setIsExact(false); | ||||
| 553 | } | ||||
| 554 | return InnerShift; | ||||
| 555 | }; | ||||
| 556 | |||||
| 557 | // Two logical shifts in the same direction: | ||||
| 558 | // shl (shl X, C1), C2 --> shl X, C1 + C2 | ||||
| 559 | // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 | ||||
| 560 | if (IsInnerShl == IsOuterShl) { | ||||
| 561 | // If this is an oversized composite shift, then unsigned shifts get 0. | ||||
| 562 | if (InnerShAmt + OuterShAmt >= TypeWidth) | ||||
| 563 | return Constant::getNullValue(ShType); | ||||
| 564 | |||||
| 565 | return NewInnerShift(InnerShAmt + OuterShAmt); | ||||
| 566 | } | ||||
| 567 | |||||
| 568 | // Equal shift amounts in opposite directions become bitwise 'and': | ||||
| 569 | // lshr (shl X, C), C --> and X, C' | ||||
| 570 | // shl (lshr X, C), C --> and X, C' | ||||
| 571 | if (InnerShAmt == OuterShAmt) { | ||||
| 572 | APInt Mask = IsInnerShl | ||||
| 573 | ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt) | ||||
| 574 | : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt); | ||||
| 575 | Value *And = Builder.CreateAnd(InnerShift->getOperand(0), | ||||
| 576 | ConstantInt::get(ShType, Mask)); | ||||
| 577 | if (auto *AndI = dyn_cast<Instruction>(And)) { | ||||
| 578 | AndI->moveBefore(InnerShift); | ||||
| 579 | AndI->takeName(InnerShift); | ||||
| 580 | } | ||||
| 581 | return And; | ||||
| 582 | } | ||||
| 583 | |||||
| 584 | assert(InnerShAmt > OuterShAmt &&(static_cast <bool> (InnerShAmt > OuterShAmt && "Unexpected opposite direction logical shift pair") ? void ( 0) : __assert_fail ("InnerShAmt > OuterShAmt && \"Unexpected opposite direction logical shift pair\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 585 , __extension__ __PRETTY_FUNCTION__)) | ||||
| 585 | "Unexpected opposite direction logical shift pair")(static_cast <bool> (InnerShAmt > OuterShAmt && "Unexpected opposite direction logical shift pair") ? void ( 0) : __assert_fail ("InnerShAmt > OuterShAmt && \"Unexpected opposite direction logical shift pair\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 585 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 586 | |||||
| 587 | // In general, we would need an 'and' for this transform, but | ||||
| 588 | // canEvaluateShiftedShift() guarantees that the masked-off bits are not used. | ||||
| 589 | // lshr (shl X, C1), C2 --> shl X, C1 - C2 | ||||
| 590 | // shl (lshr X, C1), C2 --> lshr X, C1 - C2 | ||||
| 591 | return NewInnerShift(InnerShAmt - OuterShAmt); | ||||
| 592 | } | ||||
| 593 | |||||
| 594 | /// When canEvaluateShifted() returns true for an expression, this function | ||||
| 595 | /// inserts the new computation that produces the shifted value. | ||||
| 596 | static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, | ||||
| 597 | InstCombinerImpl &IC, const DataLayout &DL) { | ||||
| 598 | // We can always evaluate constants shifted. | ||||
| 599 | if (Constant *C = dyn_cast<Constant>(V)) { | ||||
| 600 | if (isLeftShift) | ||||
| 601 | return IC.Builder.CreateShl(C, NumBits); | ||||
| 602 | else | ||||
| 603 | return IC.Builder.CreateLShr(C, NumBits); | ||||
| 604 | } | ||||
| 605 | |||||
| 606 | Instruction *I = cast<Instruction>(V); | ||||
| 607 | IC.addToWorklist(I); | ||||
| 608 | |||||
| 609 | switch (I->getOpcode()) { | ||||
| 610 | default: llvm_unreachable("Inconsistency with CanEvaluateShifted")::llvm::llvm_unreachable_internal("Inconsistency with CanEvaluateShifted" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 610 ); | ||||
| 611 | case Instruction::And: | ||||
| 612 | case Instruction::Or: | ||||
| 613 | case Instruction::Xor: | ||||
| 614 | // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. | ||||
| 615 | I->setOperand( | ||||
| 616 | 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL)); | ||||
| 617 | I->setOperand( | ||||
| 618 | 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL)); | ||||
| 619 | return I; | ||||
| 620 | |||||
| 621 | case Instruction::Shl: | ||||
| 622 | case Instruction::LShr: | ||||
| 623 | return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift, | ||||
| 624 | IC.Builder); | ||||
| 625 | |||||
| 626 | case Instruction::Select: | ||||
| 627 | I->setOperand( | ||||
| 628 | 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL)); | ||||
| 629 | I->setOperand( | ||||
| 630 | 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL)); | ||||
| 631 | return I; | ||||
| 632 | case Instruction::PHI: { | ||||
| 633 | // We can change a phi if we can change all operands. Note that we never | ||||
| 634 | // get into trouble with cyclic PHIs here because we only consider | ||||
| 635 | // instructions with a single use. | ||||
| 636 | PHINode *PN = cast<PHINode>(I); | ||||
| 637 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) | ||||
| 638 | PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits, | ||||
| 639 | isLeftShift, IC, DL)); | ||||
| 640 | return PN; | ||||
| 641 | } | ||||
| 642 | } | ||||
| 643 | } | ||||
| 644 | |||||
| 645 | // If this is a bitwise operator or add with a constant RHS we might be able | ||||
| 646 | // to pull it through a shift. | ||||
| 647 | static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift, | ||||
| 648 | BinaryOperator *BO) { | ||||
| 649 | switch (BO->getOpcode()) { | ||||
| 650 | default: | ||||
| 651 | return false; // Do not perform transform! | ||||
| 652 | case Instruction::Add: | ||||
| 653 | return Shift.getOpcode() == Instruction::Shl; | ||||
| 654 | case Instruction::Or: | ||||
| 655 | case Instruction::And: | ||||
| 656 | return true; | ||||
| 657 | case Instruction::Xor: | ||||
| 658 | // Do not change a 'not' of logical shift because that would create a normal | ||||
| 659 | // 'xor'. The 'not' is likely better for analysis, SCEV, and codegen. | ||||
| 660 | return !(Shift.isLogicalShift() && match(BO, m_Not(m_Value()))); | ||||
| 661 | } | ||||
| 662 | } | ||||
| 663 | |||||
| 664 | Instruction *InstCombinerImpl::FoldShiftByConstant(Value *Op0, Constant *Op1, | ||||
| 665 | BinaryOperator &I) { | ||||
| 666 | const APInt *Op1C; | ||||
| 667 | if (!match(Op1, m_APInt(Op1C))) | ||||
| 668 | return nullptr; | ||||
| 669 | |||||
| 670 | // See if we can propagate this shift into the input, this covers the trivial | ||||
| 671 | // cast of lshr(shl(x,c1),c2) as well as other more complex cases. | ||||
| 672 | bool IsLeftShift = I.getOpcode() == Instruction::Shl; | ||||
| 673 | if (I.getOpcode() != Instruction::AShr && | ||||
| 674 | canEvaluateShifted(Op0, Op1C->getZExtValue(), IsLeftShift, *this, &I)) { | ||||
| 675 | 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) | ||||
| 676 | 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) | ||||
| 677 | " 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) | ||||
| 678 | << *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); | ||||
| 679 | |||||
| 680 | return replaceInstUsesWith( | ||||
| 681 | I, getShiftedValue(Op0, Op1C->getZExtValue(), IsLeftShift, *this, DL)); | ||||
| 682 | } | ||||
| 683 | |||||
| 684 | // See if we can simplify any instructions used by the instruction whose sole | ||||
| 685 | // purpose is to compute bits we don't care about. | ||||
| 686 | Type *Ty = I.getType(); | ||||
| 687 | unsigned TypeBits = Ty->getScalarSizeInBits(); | ||||
| 688 | assert(!Op1C->uge(TypeBits) &&(static_cast <bool> (!Op1C->uge(TypeBits) && "Shift over the type width should have been removed already" ) ? void (0) : __assert_fail ("!Op1C->uge(TypeBits) && \"Shift over the type width should have been removed already\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 689 , __extension__ __PRETTY_FUNCTION__)) | ||||
| 689 | "Shift over the type width should have been removed already")(static_cast <bool> (!Op1C->uge(TypeBits) && "Shift over the type width should have been removed already" ) ? void (0) : __assert_fail ("!Op1C->uge(TypeBits) && \"Shift over the type width should have been removed already\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 689 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 690 | (void)TypeBits; | ||||
| 691 | |||||
| 692 | if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I)) | ||||
| 693 | return FoldedShift; | ||||
| 694 | |||||
| 695 | if (!Op0->hasOneUse()) | ||||
| 696 | return nullptr; | ||||
| 697 | |||||
| 698 | if (auto *Op0BO = dyn_cast<BinaryOperator>(Op0)) { | ||||
| 699 | // If the operand is a bitwise operator with a constant RHS, and the | ||||
| 700 | // shift is the only use, we can pull it out of the shift. | ||||
| 701 | const APInt *Op0C; | ||||
| 702 | if (match(Op0BO->getOperand(1), m_APInt(Op0C))) { | ||||
| 703 | if (canShiftBinOpWithConstantRHS(I, Op0BO)) { | ||||
| 704 | Constant *NewRHS = ConstantExpr::get( | ||||
| 705 | I.getOpcode(), cast<Constant>(Op0BO->getOperand(1)), Op1); | ||||
| 706 | |||||
| 707 | Value *NewShift = | ||||
| 708 | Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1); | ||||
| 709 | NewShift->takeName(Op0BO); | ||||
| 710 | |||||
| 711 | return BinaryOperator::Create(Op0BO->getOpcode(), NewShift, NewRHS); | ||||
| 712 | } | ||||
| 713 | } | ||||
| 714 | } | ||||
| 715 | |||||
| 716 | // If we have a select that conditionally executes some binary operator, | ||||
| 717 | // see if we can pull it the select and operator through the shift. | ||||
| 718 | // | ||||
| 719 | // For example, turning: | ||||
| 720 | // shl (select C, (add X, C1), X), C2 | ||||
| 721 | // Into: | ||||
| 722 | // Y = shl X, C2 | ||||
| 723 | // select C, (add Y, C1 << C2), Y | ||||
| 724 | Value *Cond; | ||||
| 725 | BinaryOperator *TBO; | ||||
| 726 | Value *FalseVal; | ||||
| 727 | if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)), | ||||
| 728 | m_Value(FalseVal)))) { | ||||
| 729 | const APInt *C; | ||||
| 730 | if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal && | ||||
| 731 | match(TBO->getOperand(1), m_APInt(C)) && | ||||
| 732 | canShiftBinOpWithConstantRHS(I, TBO)) { | ||||
| 733 | Constant *NewRHS = ConstantExpr::get( | ||||
| 734 | I.getOpcode(), cast<Constant>(TBO->getOperand(1)), Op1); | ||||
| 735 | |||||
| 736 | Value *NewShift = Builder.CreateBinOp(I.getOpcode(), FalseVal, Op1); | ||||
| 737 | Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift, NewRHS); | ||||
| 738 | return SelectInst::Create(Cond, NewOp, NewShift); | ||||
| 739 | } | ||||
| 740 | } | ||||
| 741 | |||||
| 742 | BinaryOperator *FBO; | ||||
| 743 | Value *TrueVal; | ||||
| 744 | if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal), | ||||
| 745 | m_OneUse(m_BinOp(FBO))))) { | ||||
| 746 | const APInt *C; | ||||
| 747 | if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal && | ||||
| 748 | match(FBO->getOperand(1), m_APInt(C)) && | ||||
| 749 | canShiftBinOpWithConstantRHS(I, FBO)) { | ||||
| 750 | Constant *NewRHS = ConstantExpr::get( | ||||
| 751 | I.getOpcode(), cast<Constant>(FBO->getOperand(1)), Op1); | ||||
| 752 | |||||
| 753 | Value *NewShift = Builder.CreateBinOp(I.getOpcode(), TrueVal, Op1); | ||||
| 754 | Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift, NewRHS); | ||||
| 755 | return SelectInst::Create(Cond, NewShift, NewOp); | ||||
| 756 | } | ||||
| 757 | } | ||||
| 758 | |||||
| 759 | return nullptr; | ||||
| 760 | } | ||||
| 761 | |||||
| 762 | Instruction *InstCombinerImpl::visitShl(BinaryOperator &I) { | ||||
| 763 | const SimplifyQuery Q = SQ.getWithInstruction(&I); | ||||
| 764 | |||||
| 765 | if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1), | ||||
| |||||
| 766 | I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), Q)) | ||||
| 767 | return replaceInstUsesWith(I, V); | ||||
| 768 | |||||
| 769 | if (Instruction *X = foldVectorBinop(I)) | ||||
| 770 | return X; | ||||
| 771 | |||||
| 772 | if (Instruction *V
| ||||
| 773 | return V; | ||||
| 774 | |||||
| 775 | if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(&I, Q, Builder)) | ||||
| 776 | return V; | ||||
| 777 | |||||
| 778 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
| 779 | Type *Ty = I.getType(); | ||||
| 780 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
| 781 | |||||
| 782 | const APInt *C; | ||||
| 783 | if (match(Op1, m_APInt(C))) { | ||||
| 784 | unsigned ShAmtC = C->getZExtValue(); | ||||
| 785 | |||||
| 786 | // shl (zext X), C --> zext (shl X, C) | ||||
| 787 | // This is only valid if X would have zeros shifted out. | ||||
| 788 | Value *X; | ||||
| 789 | if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) { | ||||
| 790 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); | ||||
| 791 | if (ShAmtC < SrcWidth && | ||||
| 792 | MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmtC), 0, &I)) | ||||
| 793 | return new ZExtInst(Builder.CreateShl(X, ShAmtC), Ty); | ||||
| 794 | } | ||||
| 795 | |||||
| 796 | // (X >> C) << C --> X & (-1 << C) | ||||
| 797 | if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) { | ||||
| 798 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 799 | return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask)); | ||||
| 800 | } | ||||
| 801 | |||||
| 802 | const APInt *C1; | ||||
| 803 | if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(C1)))) && | ||||
| 804 | C1->ult(BitWidth)) { | ||||
| 805 | unsigned ShrAmt = C1->getZExtValue(); | ||||
| 806 | if (ShrAmt < ShAmtC) { | ||||
| 807 | // If C1 < C: (X >>?,exact C1) << C --> X << (C - C1) | ||||
| 808 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShrAmt); | ||||
| 809 | auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); | ||||
| 810 | NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); | ||||
| 811 | NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); | ||||
| 812 | return NewShl; | ||||
| 813 | } | ||||
| 814 | if (ShrAmt > ShAmtC) { | ||||
| 815 | // If C1 > C: (X >>?exact C1) << C --> X >>?exact (C1 - C) | ||||
| 816 | Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmtC); | ||||
| 817 | auto *NewShr = BinaryOperator::Create( | ||||
| 818 | cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff); | ||||
| 819 | NewShr->setIsExact(true); | ||||
| 820 | return NewShr; | ||||
| 821 | } | ||||
| 822 | } | ||||
| 823 | |||||
| 824 | if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_APInt(C1)))) && | ||||
| 825 | C1->ult(BitWidth)) { | ||||
| 826 | unsigned ShrAmt = C1->getZExtValue(); | ||||
| 827 | if (ShrAmt < ShAmtC) { | ||||
| 828 | // If C1 < C: (X >>? C1) << C --> (X << (C - C1)) & (-1 << C) | ||||
| 829 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShrAmt); | ||||
| 830 | auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); | ||||
| 831 | NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); | ||||
| 832 | NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); | ||||
| 833 | Builder.Insert(NewShl); | ||||
| 834 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 835 | return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask)); | ||||
| 836 | } | ||||
| 837 | if (ShrAmt > ShAmtC) { | ||||
| 838 | // If C1 > C: (X >>? C1) << C --> (X >>? (C1 - C)) & (-1 << C) | ||||
| 839 | Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmtC); | ||||
| 840 | auto *OldShr = cast<BinaryOperator>(Op0); | ||||
| 841 | auto *NewShr = | ||||
| 842 | BinaryOperator::Create(OldShr->getOpcode(), X, ShiftDiff); | ||||
| 843 | NewShr->setIsExact(OldShr->isExact()); | ||||
| 844 | Builder.Insert(NewShr); | ||||
| 845 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 846 | return BinaryOperator::CreateAnd(NewShr, ConstantInt::get(Ty, Mask)); | ||||
| 847 | } | ||||
| 848 | } | ||||
| 849 | |||||
| 850 | // Similar to above, but look through an intermediate trunc instruction. | ||||
| 851 | BinaryOperator *Shr; | ||||
| 852 | if (match(Op0, m_OneUse(m_Trunc(m_OneUse(m_BinOp(Shr))))) && | ||||
| 853 | match(Shr, m_Shr(m_Value(X), m_APInt(C1)))) { | ||||
| 854 | // The larger shift direction survives through the transform. | ||||
| 855 | unsigned ShrAmtC = C1->getZExtValue(); | ||||
| 856 | unsigned ShDiff = ShrAmtC > ShAmtC ? ShrAmtC - ShAmtC : ShAmtC - ShrAmtC; | ||||
| 857 | Constant *ShiftDiffC = ConstantInt::get(X->getType(), ShDiff); | ||||
| 858 | auto ShiftOpc = ShrAmtC > ShAmtC ? Shr->getOpcode() : Instruction::Shl; | ||||
| 859 | |||||
| 860 | // If C1 > C: | ||||
| 861 | // (trunc (X >> C1)) << C --> (trunc (X >> (C1 - C))) && (-1 << C) | ||||
| 862 | // If C > C1: | ||||
| 863 | // (trunc (X >> C1)) << C --> (trunc (X << (C - C1))) && (-1 << C) | ||||
| 864 | Value *NewShift = Builder.CreateBinOp(ShiftOpc, X, ShiftDiffC, "sh.diff"); | ||||
| 865 | Value *Trunc = Builder.CreateTrunc(NewShift, Ty, "tr.sh.diff"); | ||||
| 866 | APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 867 | return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Ty, Mask)); | ||||
| 868 | } | ||||
| 869 | |||||
| 870 | if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) { | ||||
| 871 | unsigned AmtSum = ShAmtC + C1->getZExtValue(); | ||||
| 872 | // Oversized shifts are simplified to zero in InstSimplify. | ||||
| 873 | if (AmtSum < BitWidth) | ||||
| 874 | // (X << C1) << C2 --> X << (C1 + C2) | ||||
| 875 | return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum)); | ||||
| 876 | } | ||||
| 877 | |||||
| 878 | // If we have an opposite shift by the same amount, we may be able to | ||||
| 879 | // reorder binops and shifts to eliminate math/logic. | ||||
| 880 | auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) { | ||||
| 881 | switch (BinOpcode) { | ||||
| 882 | default: | ||||
| 883 | return false; | ||||
| 884 | case Instruction::Add: | ||||
| 885 | case Instruction::And: | ||||
| 886 | case Instruction::Or: | ||||
| 887 | case Instruction::Xor: | ||||
| 888 | case Instruction::Sub: | ||||
| 889 | // NOTE: Sub is not commutable and the tranforms below may not be valid | ||||
| 890 | // when the shift-right is operand 1 (RHS) of the sub. | ||||
| 891 | return true; | ||||
| 892 | } | ||||
| 893 | }; | ||||
| 894 | BinaryOperator *Op0BO; | ||||
| 895 | if (match(Op0, m_OneUse(m_BinOp(Op0BO))) && | ||||
| 896 | isSuitableBinOpcode(Op0BO->getOpcode())) { | ||||
| 897 | // Commute so shift-right is on LHS of the binop. | ||||
| 898 | // (Y bop (X >> C)) << C -> ((X >> C) bop Y) << C | ||||
| 899 | // (Y bop ((X >> C) & CC)) << C -> (((X >> C) & CC) bop Y) << C | ||||
| 900 | Value *Shr = Op0BO->getOperand(0); | ||||
| 901 | Value *Y = Op0BO->getOperand(1); | ||||
| 902 | Value *X; | ||||
| 903 | const APInt *CC; | ||||
| 904 | if (Op0BO->isCommutative() && Y->hasOneUse() && | ||||
| 905 | (match(Y, m_Shr(m_Value(), m_Specific(Op1))) || | ||||
| 906 | match(Y, m_And(m_OneUse(m_Shr(m_Value(), m_Specific(Op1))), | ||||
| 907 | m_APInt(CC))))) | ||||
| 908 | std::swap(Shr, Y); | ||||
| 909 | |||||
| 910 | // ((X >> C) bop Y) << C -> (X bop (Y << C)) & (~0 << C) | ||||
| 911 | if (match(Shr, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) { | ||||
| 912 | // Y << C | ||||
| 913 | Value *YS = Builder.CreateShl(Y, Op1, Op0BO->getName()); | ||||
| 914 | // (X bop (Y << C)) | ||||
| 915 | Value *B = | ||||
| 916 | Builder.CreateBinOp(Op0BO->getOpcode(), X, YS, Shr->getName()); | ||||
| 917 | unsigned Op1Val = C->getLimitedValue(BitWidth); | ||||
| 918 | APInt Bits = APInt::getHighBitsSet(BitWidth, BitWidth - Op1Val); | ||||
| 919 | Constant *Mask = ConstantInt::get(Ty, Bits); | ||||
| 920 | return BinaryOperator::CreateAnd(B, Mask); | ||||
| 921 | } | ||||
| 922 | |||||
| 923 | // (((X >> C) & CC) bop Y) << C -> (X & (CC << C)) bop (Y << C) | ||||
| 924 | if (match(Shr, | ||||
| 925 | m_OneUse(m_And(m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))), | ||||
| 926 | m_APInt(CC))))) { | ||||
| 927 | // Y << C | ||||
| 928 | Value *YS = Builder.CreateShl(Y, Op1, Op0BO->getName()); | ||||
| 929 | // X & (CC << C) | ||||
| 930 | Value *M = Builder.CreateAnd(X, ConstantInt::get(Ty, CC->shl(*C)), | ||||
| 931 | X->getName() + ".mask"); | ||||
| 932 | return BinaryOperator::Create(Op0BO->getOpcode(), M, YS); | ||||
| 933 | } | ||||
| 934 | } | ||||
| 935 | |||||
| 936 | // (C1 - X) << C --> (C1 << C) - (X << C) | ||||
| 937 | if (match(Op0, m_OneUse(m_Sub(m_APInt(C1), m_Value(X))))) { | ||||
| 938 | Constant *NewLHS = ConstantInt::get(Ty, C1->shl(*C)); | ||||
| 939 | Value *NewShift = Builder.CreateShl(X, Op1); | ||||
| 940 | return BinaryOperator::CreateSub(NewLHS, NewShift); | ||||
| 941 | } | ||||
| 942 | |||||
| 943 | // If the shifted-out value is known-zero, then this is a NUW shift. | ||||
| 944 | if (!I.hasNoUnsignedWrap() && | ||||
| 945 | MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmtC), 0, | ||||
| 946 | &I)) { | ||||
| 947 | I.setHasNoUnsignedWrap(); | ||||
| 948 | return &I; | ||||
| 949 | } | ||||
| 950 | |||||
| 951 | // If the shifted-out value is all signbits, then this is a NSW shift. | ||||
| 952 | if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmtC) { | ||||
| 953 | I.setHasNoSignedWrap(); | ||||
| 954 | return &I; | ||||
| 955 | } | ||||
| 956 | } | ||||
| 957 | |||||
| 958 | // Transform (x >> y) << y to x & (-1 << y) | ||||
| 959 | // Valid for any type of right-shift. | ||||
| 960 | Value *X; | ||||
| 961 | if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) { | ||||
| 962 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); | ||||
| 963 | Value *Mask = Builder.CreateShl(AllOnes, Op1); | ||||
| 964 | return BinaryOperator::CreateAnd(Mask, X); | ||||
| 965 | } | ||||
| 966 | |||||
| 967 | Constant *C1; | ||||
| 968 | if (match(Op1, m_Constant(C1))) { | ||||
| 969 | Constant *C2; | ||||
| 970 | Value *X; | ||||
| 971 | // (C2 << X) << C1 --> (C2 << C1) << X | ||||
| 972 | if (match(Op0, m_OneUse(m_Shl(m_Constant(C2), m_Value(X))))) | ||||
| 973 | return BinaryOperator::CreateShl(ConstantExpr::getShl(C2, C1), X); | ||||
| 974 | |||||
| 975 | // (X * C2) << C1 --> X * (C2 << C1) | ||||
| 976 | if (match(Op0, m_Mul(m_Value(X), m_Constant(C2)))) | ||||
| 977 | return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1)); | ||||
| 978 | |||||
| 979 | // shl (zext i1 X), C1 --> select (X, 1 << C1, 0) | ||||
| 980 | if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { | ||||
| 981 | auto *NewC = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C1); | ||||
| 982 | return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty)); | ||||
| 983 | } | ||||
| 984 | } | ||||
| 985 | |||||
| 986 | // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1 | ||||
| 987 | if (match(Op0, m_One()) && | ||||
| 988 | match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X)))) | ||||
| 989 | return BinaryOperator::CreateLShr( | ||||
| 990 | ConstantInt::get(Ty, APInt::getSignMask(BitWidth)), X); | ||||
| 991 | |||||
| 992 | return nullptr; | ||||
| 993 | } | ||||
| 994 | |||||
| 995 | Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) { | ||||
| 996 | if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(), | ||||
| 997 | SQ.getWithInstruction(&I))) | ||||
| 998 | return replaceInstUsesWith(I, V); | ||||
| 999 | |||||
| 1000 | if (Instruction *X = foldVectorBinop(I)) | ||||
| 1001 | return X; | ||||
| 1002 | |||||
| 1003 | if (Instruction *R = commonShiftTransforms(I)) | ||||
| 1004 | return R; | ||||
| 1005 | |||||
| 1006 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
| 1007 | Type *Ty = I.getType(); | ||||
| 1008 | const APInt *C; | ||||
| 1009 | if (match(Op1, m_APInt(C))) { | ||||
| 1010 | unsigned ShAmtC = C->getZExtValue(); | ||||
| 1011 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
| 1012 | auto *II = dyn_cast<IntrinsicInst>(Op0); | ||||
| 1013 | if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmtC && | ||||
| 1014 | (II->getIntrinsicID() == Intrinsic::ctlz || | ||||
| 1015 | II->getIntrinsicID() == Intrinsic::cttz || | ||||
| 1016 | II->getIntrinsicID() == Intrinsic::ctpop)) { | ||||
| 1017 | // ctlz.i32(x)>>5 --> zext(x == 0) | ||||
| 1018 | // cttz.i32(x)>>5 --> zext(x == 0) | ||||
| 1019 | // ctpop.i32(x)>>5 --> zext(x == -1) | ||||
| 1020 | bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop; | ||||
| 1021 | Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0); | ||||
| 1022 | Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS); | ||||
| 1023 | return new ZExtInst(Cmp, Ty); | ||||
| 1024 | } | ||||
| 1025 | |||||
| 1026 | Value *X; | ||||
| 1027 | const APInt *C1; | ||||
| 1028 | if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) { | ||||
| 1029 | if (C1->ult(ShAmtC)) { | ||||
| 1030 | unsigned ShlAmtC = C1->getZExtValue(); | ||||
| 1031 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShlAmtC); | ||||
| 1032 | if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) { | ||||
| 1033 | // (X <<nuw C1) >>u C --> X >>u (C - C1) | ||||
| 1034 | auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff); | ||||
| 1035 | NewLShr->setIsExact(I.isExact()); | ||||
| 1036 | return NewLShr; | ||||
| 1037 | } | ||||
| 1038 | if (Op0->hasOneUse()) { | ||||
| 1039 | // (X << C1) >>u C --> (X >>u (C - C1)) & (-1 >> C) | ||||
| 1040 | Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact()); | ||||
| 1041 | APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 1042 | return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask)); | ||||
| 1043 | } | ||||
| 1044 | } else if (C1->ugt(ShAmtC)) { | ||||
| 1045 | unsigned ShlAmtC = C1->getZExtValue(); | ||||
| 1046 | Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmtC - ShAmtC); | ||||
| 1047 | if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) { | ||||
| 1048 | // (X <<nuw C1) >>u C --> X <<nuw (C1 - C) | ||||
| 1049 | auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); | ||||
| 1050 | NewShl->setHasNoUnsignedWrap(true); | ||||
| 1051 | return NewShl; | ||||
| 1052 | } | ||||
| 1053 | if (Op0->hasOneUse()) { | ||||
| 1054 | // (X << C1) >>u C --> X << (C1 - C) & (-1 >> C) | ||||
| 1055 | Value *NewShl = Builder.CreateShl(X, ShiftDiff); | ||||
| 1056 | APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 1057 | return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask)); | ||||
| 1058 | } | ||||
| 1059 | } else { | ||||
| 1060 | assert(*C1 == ShAmtC)(static_cast <bool> (*C1 == ShAmtC) ? void (0) : __assert_fail ("*C1 == ShAmtC", "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp" , 1060, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1061 | // (X << C) >>u C --> X & (-1 >>u C) | ||||
| 1062 | APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 1063 | return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask)); | ||||
| 1064 | } | ||||
| 1065 | } | ||||
| 1066 | |||||
| 1067 | // ((X << C) + Y) >>u C --> (X + (Y >>u C)) & (-1 >>u C) | ||||
| 1068 | // TODO: Consolidate with the more general transform that starts from shl | ||||
| 1069 | // (the shifts are in the opposite order). | ||||
| 1070 | Value *Y; | ||||
| 1071 | if (match(Op0, | ||||
| 1072 | m_OneUse(m_c_Add(m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))), | ||||
| 1073 | m_Value(Y))))) { | ||||
| 1074 | Value *NewLshr = Builder.CreateLShr(Y, Op1); | ||||
| 1075 | Value *NewAdd = Builder.CreateAdd(NewLshr, X); | ||||
| 1076 | unsigned Op1Val = C->getLimitedValue(BitWidth); | ||||
| 1077 | APInt Bits = APInt::getLowBitsSet(BitWidth, BitWidth - Op1Val); | ||||
| 1078 | Constant *Mask = ConstantInt::get(Ty, Bits); | ||||
| 1079 | return BinaryOperator::CreateAnd(NewAdd, Mask); | ||||
| 1080 | } | ||||
| 1081 | |||||
| 1082 | if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) && | ||||
| 1083 | (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) { | ||||
| 1084 | assert(ShAmtC < X->getType()->getScalarSizeInBits() &&(static_cast <bool> (ShAmtC < X->getType()->getScalarSizeInBits () && "Big shift not simplified to zero?") ? void (0) : __assert_fail ("ShAmtC < X->getType()->getScalarSizeInBits() && \"Big shift not simplified to zero?\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 1085 , __extension__ __PRETTY_FUNCTION__)) | ||||
| 1085 | "Big shift not simplified to zero?")(static_cast <bool> (ShAmtC < X->getType()->getScalarSizeInBits () && "Big shift not simplified to zero?") ? void (0) : __assert_fail ("ShAmtC < X->getType()->getScalarSizeInBits() && \"Big shift not simplified to zero?\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 1085 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 1086 | // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN | ||||
| 1087 | Value *NewLShr = Builder.CreateLShr(X, ShAmtC); | ||||
| 1088 | return new ZExtInst(NewLShr, Ty); | ||||
| 1089 | } | ||||
| 1090 | |||||
| 1091 | if (match(Op0, m_SExt(m_Value(X)))) { | ||||
| 1092 | unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits(); | ||||
| 1093 | // lshr (sext i1 X to iN), C --> select (X, -1 >> C, 0) | ||||
| 1094 | if (SrcTyBitWidth == 1) { | ||||
| 1095 | auto *NewC = ConstantInt::get( | ||||
| 1096 | Ty, APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC)); | ||||
| 1097 | return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty)); | ||||
| 1098 | } | ||||
| 1099 | |||||
| 1100 | if ((!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType())) && | ||||
| 1101 | Op0->hasOneUse()) { | ||||
| 1102 | // Are we moving the sign bit to the low bit and widening with high | ||||
| 1103 | // zeros? lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN | ||||
| 1104 | if (ShAmtC == BitWidth - 1) { | ||||
| 1105 | Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1); | ||||
| 1106 | return new ZExtInst(NewLShr, Ty); | ||||
| 1107 | } | ||||
| 1108 | |||||
| 1109 | // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN | ||||
| 1110 | if (ShAmtC == BitWidth - SrcTyBitWidth) { | ||||
| 1111 | // The new shift amount can't be more than the narrow source type. | ||||
| 1112 | unsigned NewShAmt = std::min(ShAmtC, SrcTyBitWidth - 1); | ||||
| 1113 | Value *AShr = Builder.CreateAShr(X, NewShAmt); | ||||
| 1114 | return new ZExtInst(AShr, Ty); | ||||
| 1115 | } | ||||
| 1116 | } | ||||
| 1117 | } | ||||
| 1118 | |||||
| 1119 | if (ShAmtC == BitWidth - 1) { | ||||
| 1120 | // lshr i32 or(X,-X), 31 --> zext (X != 0) | ||||
| 1121 | if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X))))) | ||||
| 1122 | return new ZExtInst(Builder.CreateIsNotNull(X), Ty); | ||||
| 1123 | |||||
| 1124 | // lshr i32 (X -nsw Y), 31 --> zext (X < Y) | ||||
| 1125 | if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y))))) | ||||
| 1126 | return new ZExtInst(Builder.CreateICmpSLT(X, Y), Ty); | ||||
| 1127 | |||||
| 1128 | // Check if a number is negative and odd: | ||||
| 1129 | // lshr i32 (srem X, 2), 31 --> and (X >> 31), X | ||||
| 1130 | if (match(Op0, m_OneUse(m_SRem(m_Value(X), m_SpecificInt(2))))) { | ||||
| 1131 | Value *Signbit = Builder.CreateLShr(X, ShAmtC); | ||||
| 1132 | return BinaryOperator::CreateAnd(Signbit, X); | ||||
| 1133 | } | ||||
| 1134 | } | ||||
| 1135 | |||||
| 1136 | // (X >>u C1) >>u C --> X >>u (C1 + C) | ||||
| 1137 | if (match(Op0, m_LShr(m_Value(X), m_APInt(C1)))) { | ||||
| 1138 | // Oversized shifts are simplified to zero in InstSimplify. | ||||
| 1139 | unsigned AmtSum = ShAmtC + C1->getZExtValue(); | ||||
| 1140 | if (AmtSum < BitWidth) | ||||
| 1141 | return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum)); | ||||
| 1142 | } | ||||
| 1143 | |||||
| 1144 | Instruction *TruncSrc; | ||||
| 1145 | if (match(Op0, m_OneUse(m_Trunc(m_Instruction(TruncSrc)))) && | ||||
| 1146 | match(TruncSrc, m_LShr(m_Value(X), m_APInt(C1)))) { | ||||
| 1147 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); | ||||
| 1148 | unsigned AmtSum = ShAmtC + C1->getZExtValue(); | ||||
| 1149 | |||||
| 1150 | // If the combined shift fits in the source width: | ||||
| 1151 | // (trunc (X >>u C1)) >>u C --> and (trunc (X >>u (C1 + C)), MaskC | ||||
| 1152 | // | ||||
| 1153 | // If the first shift covers the number of bits truncated, then the | ||||
| 1154 | // mask instruction is eliminated (and so the use check is relaxed). | ||||
| 1155 | if (AmtSum < SrcWidth && | ||||
| 1156 | (TruncSrc->hasOneUse() || C1->uge(SrcWidth - BitWidth))) { | ||||
| 1157 | Value *SumShift = Builder.CreateLShr(X, AmtSum, "sum.shift"); | ||||
| 1158 | Value *Trunc = Builder.CreateTrunc(SumShift, Ty, I.getName()); | ||||
| 1159 | |||||
| 1160 | // If the first shift does not cover the number of bits truncated, then | ||||
| 1161 | // we require a mask to get rid of high bits in the result. | ||||
| 1162 | APInt MaskC = APInt::getAllOnes(BitWidth).lshr(ShAmtC); | ||||
| 1163 | return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Ty, MaskC)); | ||||
| 1164 | } | ||||
| 1165 | } | ||||
| 1166 | |||||
| 1167 | // Look for a "splat" mul pattern - it replicates bits across each half of | ||||
| 1168 | // a value, so a right shift is just a mask of the low bits: | ||||
| 1169 | // lshr i[2N] (mul nuw X, (2^N)+1), N --> and iN X, (2^N)-1 | ||||
| 1170 | // TODO: Generalize to allow more than just half-width shifts? | ||||
| 1171 | const APInt *MulC; | ||||
| 1172 | if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC))) && | ||||
| 1173 | BitWidth > 2 && ShAmtC * 2 == BitWidth && (*MulC - 1).isPowerOf2() && | ||||
| 1174 | MulC->logBase2() == ShAmtC) | ||||
| 1175 | return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2)); | ||||
| 1176 | |||||
| 1177 | // Try to narrow a bswap: | ||||
| 1178 | // (bswap (zext X)) >> C --> zext (bswap X >> C') | ||||
| 1179 | // In the case where the shift amount equals the bitwidth difference, the | ||||
| 1180 | // shift is eliminated. | ||||
| 1181 | if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::bswap>( | ||||
| 1182 | m_OneUse(m_ZExt(m_Value(X))))))) { | ||||
| 1183 | // TODO: If the shift amount is less than the zext, we could shift left. | ||||
| 1184 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); | ||||
| 1185 | unsigned WidthDiff = BitWidth - SrcWidth; | ||||
| 1186 | if (SrcWidth % 16 == 0 && ShAmtC >= WidthDiff) { | ||||
| 1187 | Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X); | ||||
| 1188 | Value *NewShift = Builder.CreateLShr(NarrowSwap, ShAmtC - WidthDiff); | ||||
| 1189 | return new ZExtInst(NewShift, Ty); | ||||
| 1190 | } | ||||
| 1191 | } | ||||
| 1192 | |||||
| 1193 | // If the shifted-out value is known-zero, then this is an exact shift. | ||||
| 1194 | if (!I.isExact() && | ||||
| 1195 | MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmtC), 0, &I)) { | ||||
| 1196 | I.setIsExact(); | ||||
| 1197 | return &I; | ||||
| 1198 | } | ||||
| 1199 | } | ||||
| 1200 | |||||
| 1201 | // Transform (x << y) >> y to x & (-1 >> y) | ||||
| 1202 | Value *X; | ||||
| 1203 | if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) { | ||||
| 1204 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); | ||||
| 1205 | Value *Mask = Builder.CreateLShr(AllOnes, Op1); | ||||
| 1206 | return BinaryOperator::CreateAnd(Mask, X); | ||||
| 1207 | } | ||||
| 1208 | |||||
| 1209 | return nullptr; | ||||
| 1210 | } | ||||
| 1211 | |||||
| 1212 | Instruction * | ||||
| 1213 | InstCombinerImpl::foldVariableSignZeroExtensionOfVariableHighBitExtract( | ||||
| 1214 | BinaryOperator &OldAShr) { | ||||
| 1215 | assert(OldAShr.getOpcode() == Instruction::AShr &&(static_cast <bool> (OldAShr.getOpcode() == Instruction ::AShr && "Must be called with arithmetic right-shift instruction only." ) ? void (0) : __assert_fail ("OldAShr.getOpcode() == Instruction::AShr && \"Must be called with arithmetic right-shift instruction only.\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 1216 , __extension__ __PRETTY_FUNCTION__)) | ||||
| 1216 | "Must be called with arithmetic right-shift instruction only.")(static_cast <bool> (OldAShr.getOpcode() == Instruction ::AShr && "Must be called with arithmetic right-shift instruction only." ) ? void (0) : __assert_fail ("OldAShr.getOpcode() == Instruction::AShr && \"Must be called with arithmetic right-shift instruction only.\"" , "llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp", 1216 , __extension__ __PRETTY_FUNCTION__)); | ||||
| 1217 | |||||
| 1218 | // Check that constant C is a splat of the element-wise bitwidth of V. | ||||
| 1219 | auto BitWidthSplat = [](Constant *C, Value *V) { | ||||
| 1220 | return match( | ||||
| 1221 | C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, | ||||
| 1222 | APInt(C->getType()->getScalarSizeInBits(), | ||||
| 1223 | V->getType()->getScalarSizeInBits()))); | ||||
| 1224 | }; | ||||
| 1225 | |||||
| 1226 | // It should look like variable-length sign-extension on the outside: | ||||
| 1227 | // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits) | ||||
| 1228 | Value *NBits; | ||||
| 1229 | Instruction *MaybeTrunc; | ||||
| 1230 | Constant *C1, *C2; | ||||
| 1231 | if (!match(&OldAShr, | ||||
| 1232 | m_AShr(m_Shl(m_Instruction(MaybeTrunc), | ||||
| 1233 | m_ZExtOrSelf(m_Sub(m_Constant(C1), | ||||
| 1234 | m_ZExtOrSelf(m_Value(NBits))))), | ||||
| 1235 | m_ZExtOrSelf(m_Sub(m_Constant(C2), | ||||
| 1236 | m_ZExtOrSelf(m_Deferred(NBits)))))) || | ||||
| 1237 | !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr)) | ||||
| 1238 | return nullptr; | ||||
| 1239 | |||||
| 1240 | // There may or may not be a truncation after outer two shifts. | ||||
| 1241 | Instruction *HighBitExtract; | ||||
| 1242 | match(MaybeTrunc, m_TruncOrSelf(m_Instruction(HighBitExtract))); | ||||
| 1243 | bool HadTrunc = MaybeTrunc != HighBitExtract; | ||||
| 1244 | |||||
| 1245 | // And finally, the innermost part of the pattern must be a right-shift. | ||||
| 1246 | Value *X, *NumLowBitsToSkip; | ||||
| 1247 | if (!match(HighBitExtract, m_Shr(m_Value(X), m_Value(NumLowBitsToSkip)))) | ||||
| 1248 | return nullptr; | ||||
| 1249 | |||||
| 1250 | // Said right-shift must extract high NBits bits - C0 must be it's bitwidth. | ||||
| 1251 | Constant *C0; | ||||
| 1252 | if (!match(NumLowBitsToSkip, | ||||
| 1253 | m_ZExtOrSelf( | ||||
| 1254 | m_Sub(m_Constant(C0), m_ZExtOrSelf(m_Specific(NBits))))) || | ||||
| 1255 | !BitWidthSplat(C0, HighBitExtract)) | ||||
| 1256 | return nullptr; | ||||
| 1257 | |||||
| 1258 | // Since the NBits is identical for all shifts, if the outermost and | ||||
| 1259 | // innermost shifts are identical, then outermost shifts are redundant. | ||||
| 1260 | // If we had truncation, do keep it though. | ||||
| 1261 | if (HighBitExtract->getOpcode() == OldAShr.getOpcode()) | ||||
| 1262 | return replaceInstUsesWith(OldAShr, MaybeTrunc); | ||||
| 1263 | |||||
| 1264 | // Else, if there was a truncation, then we need to ensure that one | ||||
| 1265 | // instruction will go away. | ||||
| 1266 | if (HadTrunc && !match(&OldAShr, m_c_BinOp(m_OneUse(m_Value()), m_Value()))) | ||||
| 1267 | return nullptr; | ||||
| 1268 | |||||
| 1269 | // Finally, bypass two innermost shifts, and perform the outermost shift on | ||||
| 1270 | // the operands of the innermost shift. | ||||
| 1271 | Instruction *NewAShr = | ||||
| 1272 | BinaryOperator::Create(OldAShr.getOpcode(), X, NumLowBitsToSkip); | ||||
| 1273 | NewAShr->copyIRFlags(HighBitExtract); // We can preserve 'exact'-ness. | ||||
| 1274 | if (!HadTrunc) | ||||
| 1275 | return NewAShr; | ||||
| 1276 | |||||
| 1277 | Builder.Insert(NewAShr); | ||||
| 1278 | return TruncInst::CreateTruncOrBitCast(NewAShr, OldAShr.getType()); | ||||
| 1279 | } | ||||
| 1280 | |||||
| 1281 | Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) { | ||||
| 1282 | if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(), | ||||
| 1283 | SQ.getWithInstruction(&I))) | ||||
| 1284 | return replaceInstUsesWith(I, V); | ||||
| 1285 | |||||
| 1286 | if (Instruction *X = foldVectorBinop(I)) | ||||
| 1287 | return X; | ||||
| 1288 | |||||
| 1289 | if (Instruction *R = commonShiftTransforms(I)) | ||||
| 1290 | return R; | ||||
| 1291 | |||||
| 1292 | Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); | ||||
| 1293 | Type *Ty = I.getType(); | ||||
| 1294 | unsigned BitWidth = Ty->getScalarSizeInBits(); | ||||
| 1295 | const APInt *ShAmtAPInt; | ||||
| 1296 | if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) { | ||||
| 1297 | unsigned ShAmt = ShAmtAPInt->getZExtValue(); | ||||
| 1298 | |||||
| 1299 | // If the shift amount equals the difference in width of the destination | ||||
| 1300 | // and source scalar types: | ||||
| 1301 | // ashr (shl (zext X), C), C --> sext X | ||||
| 1302 | Value *X; | ||||
| 1303 | if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) && | ||||
| 1304 | ShAmt == BitWidth - X->getType()->getScalarSizeInBits()) | ||||
| 1305 | return new SExtInst(X, Ty); | ||||
| 1306 | |||||
| 1307 | // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However, | ||||
| 1308 | // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits. | ||||
| 1309 | const APInt *ShOp1; | ||||
| 1310 | if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) && | ||||
| 1311 | ShOp1->ult(BitWidth)) { | ||||
| 1312 | unsigned ShlAmt = ShOp1->getZExtValue(); | ||||
| 1313 | if (ShlAmt < ShAmt) { | ||||
| 1314 | // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1) | ||||
| 1315 | Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt); | ||||
| 1316 | auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff); | ||||
| 1317 | NewAShr->setIsExact(I.isExact()); | ||||
| 1318 | return NewAShr; | ||||
| 1319 | } | ||||
| 1320 | if (ShlAmt > ShAmt) { | ||||
| 1321 | // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2) | ||||
| 1322 | Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt); | ||||
| 1323 | auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff); | ||||
| 1324 | NewShl->setHasNoSignedWrap(true); | ||||
| 1325 | return NewShl; | ||||
| 1326 | } | ||||
| 1327 | } | ||||
| 1328 | |||||
| 1329 | if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) && | ||||
| 1330 | ShOp1->ult(BitWidth)) { | ||||
| 1331 | unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); | ||||
| 1332 | // Oversized arithmetic shifts replicate the sign bit. | ||||
| 1333 | AmtSum = std::min(AmtSum, BitWidth - 1); | ||||
| 1334 | // (X >>s C1) >>s C2 --> X >>s (C1 + C2) | ||||
| 1335 | return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum)); | ||||
| 1336 | } | ||||
| 1337 | |||||
| 1338 | if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) && | ||||
| 1339 | (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) { | ||||
| 1340 | // ashr (sext X), C --> sext (ashr X, C') | ||||
| 1341 | Type *SrcTy = X->getType(); | ||||
| 1342 | ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1); | ||||
| 1343 | Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt)); | ||||
| 1344 | return new SExtInst(NewSh, Ty); | ||||
| 1345 | } | ||||
| 1346 | |||||
| 1347 | if (ShAmt == BitWidth - 1) { | ||||
| 1348 | // ashr i32 or(X,-X), 31 --> sext (X != 0) | ||||
| 1349 | if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X))))) | ||||
| 1350 | return new SExtInst(Builder.CreateIsNotNull(X), Ty); | ||||
| 1351 | |||||
| 1352 | // ashr i32 (X -nsw Y), 31 --> sext (X < Y) | ||||
| 1353 | Value *Y; | ||||
| 1354 | if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y))))) | ||||
| 1355 | return new SExtInst(Builder.CreateICmpSLT(X, Y), Ty); | ||||
| 1356 | } | ||||
| 1357 | |||||
| 1358 | // If the shifted-out value is known-zero, then this is an exact shift. | ||||
| 1359 | if (!I.isExact() && | ||||
| 1360 | MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) { | ||||
| 1361 | I.setIsExact(); | ||||
| 1362 | return &I; | ||||
| 1363 | } | ||||
| 1364 | } | ||||
| 1365 | |||||
| 1366 | // Prefer `-(x & 1)` over `(x << (bitwidth(x)-1)) a>> (bitwidth(x)-1)` | ||||
| 1367 | // as the pattern to splat the lowest bit. | ||||
| 1368 | // FIXME: iff X is already masked, we don't need the one-use check. | ||||
| 1369 | Value *X; | ||||
| 1370 | if (match(Op1, m_SpecificIntAllowUndef(BitWidth - 1)) && | ||||
| 1371 | match(Op0, m_OneUse(m_Shl(m_Value(X), | ||||
| 1372 | m_SpecificIntAllowUndef(BitWidth - 1))))) { | ||||
| 1373 | Constant *Mask = ConstantInt::get(Ty, 1); | ||||
| 1374 | // Retain the knowledge about the ignored lanes. | ||||
| 1375 | Mask = Constant::mergeUndefsWith( | ||||
| 1376 | Constant::mergeUndefsWith(Mask, cast<Constant>(Op1)), | ||||
| 1377 | cast<Constant>(cast<Instruction>(Op0)->getOperand(1))); | ||||
| 1378 | X = Builder.CreateAnd(X, Mask); | ||||
| 1379 | return BinaryOperator::CreateNeg(X); | ||||
| 1380 | } | ||||
| 1381 | |||||
| 1382 | if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(I)) | ||||
| 1383 | return R; | ||||
| 1384 | |||||
| 1385 | // See if we can turn a signed shr into an unsigned shr. | ||||
| 1386 | if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I)) | ||||
| 1387 | return BinaryOperator::CreateLShr(Op0, Op1); | ||||
| 1388 | |||||
| 1389 | // ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1 | ||||
| 1390 | if (match(Op0, m_OneUse(m_Not(m_Value(X))))) { | ||||
| 1391 | // Note that we must drop 'exact'-ness of the shift! | ||||
| 1392 | // Note that we can't keep undef's in -1 vector constant! | ||||
| 1393 | auto *NewAShr = Builder.CreateAShr(X, Op1, Op0->getName() + ".not"); | ||||
| 1394 | return BinaryOperator::CreateNot(NewAShr); | ||||
| 1395 | } | ||||
| 1396 | |||||
| 1397 | return nullptr; | ||||
| 1398 | } |
| 1 | //===- PatternMatch.h - Match on the LLVM IR --------------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file provides a simple and efficient mechanism for performing general |
| 10 | // tree-based pattern matches on the LLVM IR. The power of these routines is |
| 11 | // that it allows you to write concise patterns that are expressive and easy to |
| 12 | // understand. The other major advantage of this is that it allows you to |
| 13 | // trivially capture/bind elements in the pattern to variables. For example, |
| 14 | // you can do something like this: |
| 15 | // |
| 16 | // Value *Exp = ... |
| 17 | // Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2) |
| 18 | // if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)), |
| 19 | // m_And(m_Value(Y), m_ConstantInt(C2))))) { |
| 20 | // ... Pattern is matched and variables are bound ... |
| 21 | // } |
| 22 | // |
| 23 | // This is primarily useful to things like the instruction combiner, but can |
| 24 | // also be useful for static analysis tools or code generators. |
| 25 | // |
| 26 | //===----------------------------------------------------------------------===// |
| 27 | |
| 28 | #ifndef LLVM_IR_PATTERNMATCH_H |
| 29 | #define LLVM_IR_PATTERNMATCH_H |
| 30 | |
| 31 | #include "llvm/ADT/APFloat.h" |
| 32 | #include "llvm/ADT/APInt.h" |
| 33 | #include "llvm/IR/Constant.h" |
| 34 | #include "llvm/IR/Constants.h" |
| 35 | #include "llvm/IR/DataLayout.h" |
| 36 | #include "llvm/IR/InstrTypes.h" |
| 37 | #include "llvm/IR/Instruction.h" |
| 38 | #include "llvm/IR/Instructions.h" |
| 39 | #include "llvm/IR/IntrinsicInst.h" |
| 40 | #include "llvm/IR/Intrinsics.h" |
| 41 | #include "llvm/IR/Operator.h" |
| 42 | #include "llvm/IR/Value.h" |
| 43 | #include "llvm/Support/Casting.h" |
| 44 | #include <cstdint> |
| 45 | |
| 46 | namespace llvm { |
| 47 | namespace PatternMatch { |
| 48 | |
| 49 | template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) { |
| 50 | return const_cast<Pattern &>(P).match(V); |
| 51 | } |
| 52 | |
| 53 | template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) { |
| 54 | return const_cast<Pattern &>(P).match(Mask); |
| 55 | } |
| 56 | |
| 57 | template <typename SubPattern_t> struct OneUse_match { |
| 58 | SubPattern_t SubPattern; |
| 59 | |
| 60 | OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {} |
| 61 | |
| 62 | template <typename OpTy> bool match(OpTy *V) { |
| 63 | return V->hasOneUse() && SubPattern.match(V); |
| 64 | } |
| 65 | }; |
| 66 | |
| 67 | template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) { |
| 68 | return SubPattern; |
| 69 | } |
| 70 | |
| 71 | template <typename Class> struct class_match { |
| 72 | template <typename ITy> bool match(ITy *V) { return isa<Class>(V); } |
| 73 | }; |
| 74 | |
| 75 | /// Match an arbitrary value and ignore it. |
| 76 | inline class_match<Value> m_Value() { return class_match<Value>(); } |
| 77 | |
| 78 | /// Match an arbitrary unary operation and ignore it. |
| 79 | inline class_match<UnaryOperator> m_UnOp() { |
| 80 | return class_match<UnaryOperator>(); |
| 81 | } |
| 82 | |
| 83 | /// Match an arbitrary binary operation and ignore it. |
| 84 | inline class_match<BinaryOperator> m_BinOp() { |
| 85 | return class_match<BinaryOperator>(); |
| 86 | } |
| 87 | |
| 88 | /// Matches any compare instruction and ignore it. |
| 89 | inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); } |
| 90 | |
| 91 | struct undef_match { |
| 92 | static bool check(const Value *V) { |
| 93 | if (isa<UndefValue>(V)) |
| 94 | return true; |
| 95 | |
| 96 | const auto *CA = dyn_cast<ConstantAggregate>(V); |
| 97 | if (!CA) |
| 98 | return false; |
| 99 | |
| 100 | SmallPtrSet<const ConstantAggregate *, 8> Seen; |
| 101 | SmallVector<const ConstantAggregate *, 8> Worklist; |
| 102 | |
| 103 | // Either UndefValue, PoisonValue, or an aggregate that only contains |
| 104 | // these is accepted by matcher. |
| 105 | // CheckValue returns false if CA cannot satisfy this constraint. |
| 106 | auto CheckValue = [&](const ConstantAggregate *CA) { |
| 107 | for (const Value *Op : CA->operand_values()) { |
| 108 | if (isa<UndefValue>(Op)) |
| 109 | continue; |
| 110 | |
| 111 | const auto *CA = dyn_cast<ConstantAggregate>(Op); |
| 112 | if (!CA) |
| 113 | return false; |
| 114 | if (Seen.insert(CA).second) |
| 115 | Worklist.emplace_back(CA); |
| 116 | } |
| 117 | |
| 118 | return true; |
| 119 | }; |
| 120 | |
| 121 | if (!CheckValue(CA)) |
| 122 | return false; |
| 123 | |
| 124 | while (!Worklist.empty()) { |
| 125 | if (!CheckValue(Worklist.pop_back_val())) |
| 126 | return false; |
| 127 | } |
| 128 | return true; |
| 129 | } |
| 130 | template <typename ITy> bool match(ITy *V) { return check(V); } |
| 131 | }; |
| 132 | |
| 133 | /// Match an arbitrary undef constant. This matches poison as well. |
| 134 | /// If this is an aggregate and contains a non-aggregate element that is |
| 135 | /// neither undef nor poison, the aggregate is not matched. |
| 136 | inline auto m_Undef() { return undef_match(); } |
| 137 | |
| 138 | /// Match an arbitrary poison constant. |
| 139 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } |
| 140 | |
| 141 | /// Match an arbitrary Constant and ignore it. |
| 142 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } |
| 143 | |
| 144 | /// Match an arbitrary ConstantInt and ignore it. |
| 145 | inline class_match<ConstantInt> m_ConstantInt() { |
| 146 | return class_match<ConstantInt>(); |
| 147 | } |
| 148 | |
| 149 | /// Match an arbitrary ConstantFP and ignore it. |
| 150 | inline class_match<ConstantFP> m_ConstantFP() { |
| 151 | return class_match<ConstantFP>(); |
| 152 | } |
| 153 | |
| 154 | /// Match an arbitrary ConstantExpr and ignore it. |
| 155 | inline class_match<ConstantExpr> m_ConstantExpr() { |
| 156 | return class_match<ConstantExpr>(); |
| 157 | } |
| 158 | |
| 159 | /// Match an arbitrary basic block value and ignore it. |
| 160 | inline class_match<BasicBlock> m_BasicBlock() { |
| 161 | return class_match<BasicBlock>(); |
| 162 | } |
| 163 | |
| 164 | /// Inverting matcher |
| 165 | template <typename Ty> struct match_unless { |
| 166 | Ty M; |
| 167 | |
| 168 | match_unless(const Ty &Matcher) : M(Matcher) {} |
| 169 | |
| 170 | template <typename ITy> bool match(ITy *V) { return !M.match(V); } |
| 171 | }; |
| 172 | |
| 173 | /// Match if the inner matcher does *NOT* match. |
| 174 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { |
| 175 | return match_unless<Ty>(M); |
| 176 | } |
| 177 | |
| 178 | /// Matching combinators |
| 179 | template <typename LTy, typename RTy> struct match_combine_or { |
| 180 | LTy L; |
| 181 | RTy R; |
| 182 | |
| 183 | match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
| 184 | |
| 185 | template <typename ITy> bool match(ITy *V) { |
| 186 | if (L.match(V)) |
| 187 | return true; |
| 188 | if (R.match(V)) |
| 189 | return true; |
| 190 | return false; |
| 191 | } |
| 192 | }; |
| 193 | |
| 194 | template <typename LTy, typename RTy> struct match_combine_and { |
| 195 | LTy L; |
| 196 | RTy R; |
| 197 | |
| 198 | match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
| 199 | |
| 200 | template <typename ITy> bool match(ITy *V) { |
| 201 | if (L.match(V)) |
| 202 | if (R.match(V)) |
| 203 | return true; |
| 204 | return false; |
| 205 | } |
| 206 | }; |
| 207 | |
| 208 | /// Combine two pattern matchers matching L || R |
| 209 | template <typename LTy, typename RTy> |
| 210 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { |
| 211 | return match_combine_or<LTy, RTy>(L, R); |
| 212 | } |
| 213 | |
| 214 | /// Combine two pattern matchers matching L && R |
| 215 | template <typename LTy, typename RTy> |
| 216 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { |
| 217 | return match_combine_and<LTy, RTy>(L, R); |
| 218 | } |
| 219 | |
| 220 | struct apint_match { |
| 221 | const APInt *&Res; |
| 222 | bool AllowUndef; |
| 223 | |
| 224 | apint_match(const APInt *&Res, bool AllowUndef) |
| 225 | : Res(Res), AllowUndef(AllowUndef) {} |
| 226 | |
| 227 | template <typename ITy> bool match(ITy *V) { |
| 228 | if (auto *CI = dyn_cast<ConstantInt>(V)) { |
| 229 | Res = &CI->getValue(); |
| 230 | return true; |
| 231 | } |
| 232 | if (V->getType()->isVectorTy()) |
| 233 | if (const auto *C = dyn_cast<Constant>(V)) |
| 234 | if (auto *CI = dyn_cast_or_null<ConstantInt>( |
| 235 | C->getSplatValue(AllowUndef))) { |
| 236 | Res = &CI->getValue(); |
| 237 | return true; |
| 238 | } |
| 239 | return false; |
| 240 | } |
| 241 | }; |
| 242 | // Either constexpr if or renaming ConstantFP::getValueAPF to |
| 243 | // ConstantFP::getValue is needed to do it via single template |
| 244 | // function for both apint/apfloat. |
| 245 | struct apfloat_match { |
| 246 | const APFloat *&Res; |
| 247 | bool AllowUndef; |
| 248 | |
| 249 | apfloat_match(const APFloat *&Res, bool AllowUndef) |
| 250 | : Res(Res), AllowUndef(AllowUndef) {} |
| 251 | |
| 252 | template <typename ITy> bool match(ITy *V) { |
| 253 | if (auto *CI = dyn_cast<ConstantFP>(V)) { |
| 254 | Res = &CI->getValueAPF(); |
| 255 | return true; |
| 256 | } |
| 257 | if (V->getType()->isVectorTy()) |
| 258 | if (const auto *C = dyn_cast<Constant>(V)) |
| 259 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
| 260 | C->getSplatValue(AllowUndef))) { |
| 261 | Res = &CI->getValueAPF(); |
| 262 | return true; |
| 263 | } |
| 264 | return false; |
| 265 | } |
| 266 | }; |
| 267 | |
| 268 | /// Match a ConstantInt or splatted ConstantVector, binding the |
| 269 | /// specified pointer to the contained APInt. |
| 270 | inline apint_match m_APInt(const APInt *&Res) { |
| 271 | // Forbid undefs by default to maintain previous behavior. |
| 272 | return apint_match(Res, /* AllowUndef */ false); |
| 273 | } |
| 274 | |
| 275 | /// Match APInt while allowing undefs in splat vector constants. |
| 276 | inline apint_match m_APIntAllowUndef(const APInt *&Res) { |
| 277 | return apint_match(Res, /* AllowUndef */ true); |
| 278 | } |
| 279 | |
| 280 | /// Match APInt while forbidding undefs in splat vector constants. |
| 281 | inline apint_match m_APIntForbidUndef(const APInt *&Res) { |
| 282 | return apint_match(Res, /* AllowUndef */ false); |
| 283 | } |
| 284 | |
| 285 | /// Match a ConstantFP or splatted ConstantVector, binding the |
| 286 | /// specified pointer to the contained APFloat. |
| 287 | inline apfloat_match m_APFloat(const APFloat *&Res) { |
| 288 | // Forbid undefs by default to maintain previous behavior. |
| 289 | return apfloat_match(Res, /* AllowUndef */ false); |
| 290 | } |
| 291 | |
| 292 | /// Match APFloat while allowing undefs in splat vector constants. |
| 293 | inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) { |
| 294 | return apfloat_match(Res, /* AllowUndef */ true); |
| 295 | } |
| 296 | |
| 297 | /// Match APFloat while forbidding undefs in splat vector constants. |
| 298 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { |
| 299 | return apfloat_match(Res, /* AllowUndef */ false); |
| 300 | } |
| 301 | |
| 302 | template <int64_t Val> struct constantint_match { |
| 303 | template <typename ITy> bool match(ITy *V) { |
| 304 | if (const auto *CI = dyn_cast<ConstantInt>(V)) { |
| 305 | const APInt &CIV = CI->getValue(); |
| 306 | if (Val >= 0) |
| 307 | return CIV == static_cast<uint64_t>(Val); |
| 308 | // If Val is negative, and CI is shorter than it, truncate to the right |
| 309 | // number of bits. If it is larger, then we have to sign extend. Just |
| 310 | // compare their negated values. |
| 311 | return -CIV == -Val; |
| 312 | } |
| 313 | return false; |
| 314 | } |
| 315 | }; |
| 316 | |
| 317 | /// Match a ConstantInt with a specific value. |
| 318 | template <int64_t Val> inline constantint_match<Val> m_ConstantInt() { |
| 319 | return constantint_match<Val>(); |
| 320 | } |
| 321 | |
| 322 | /// This helper class is used to match constant scalars, vector splats, |
| 323 | /// and fixed width vectors that satisfy a specified predicate. |
| 324 | /// For fixed width vector constants, undefined elements are ignored. |
| 325 | template <typename Predicate, typename ConstantVal> |
| 326 | struct cstval_pred_ty : public Predicate { |
| 327 | template <typename ITy> bool match(ITy *V) { |
| 328 | if (const auto *CV = dyn_cast<ConstantVal>(V)) |
| 329 | return this->isValue(CV->getValue()); |
| 330 | if (const auto *VTy = dyn_cast<VectorType>(V->getType())) { |
| 331 | if (const auto *C = dyn_cast<Constant>(V)) { |
| 332 | if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue())) |
| 333 | return this->isValue(CV->getValue()); |
| 334 | |
| 335 | // Number of elements of a scalable vector unknown at compile time |
| 336 | auto *FVTy = dyn_cast<FixedVectorType>(VTy); |
| 337 | if (!FVTy) |
| 338 | return false; |
| 339 | |
| 340 | // Non-splat vector constant: check each element for a match. |
| 341 | unsigned NumElts = FVTy->getNumElements(); |
| 342 | assert(NumElts != 0 && "Constant vector with no elements?")(static_cast <bool> (NumElts != 0 && "Constant vector with no elements?" ) ? void (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\"" , "llvm/include/llvm/IR/PatternMatch.h", 342, __extension__ __PRETTY_FUNCTION__ )); |
| 343 | bool HasNonUndefElements = false; |
| 344 | for (unsigned i = 0; i != NumElts; ++i) { |
| 345 | Constant *Elt = C->getAggregateElement(i); |
| 346 | if (!Elt) |
| 347 | return false; |
| 348 | if (isa<UndefValue>(Elt)) |
| 349 | continue; |
| 350 | auto *CV = dyn_cast<ConstantVal>(Elt); |
| 351 | if (!CV || !this->isValue(CV->getValue())) |
| 352 | return false; |
| 353 | HasNonUndefElements = true; |
| 354 | } |
| 355 | return HasNonUndefElements; |
| 356 | } |
| 357 | } |
| 358 | return false; |
| 359 | } |
| 360 | }; |
| 361 | |
| 362 | /// specialization of cstval_pred_ty for ConstantInt |
| 363 | template <typename Predicate> |
| 364 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; |
| 365 | |
| 366 | /// specialization of cstval_pred_ty for ConstantFP |
| 367 | template <typename Predicate> |
| 368 | using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>; |
| 369 | |
| 370 | /// This helper class is used to match scalar and vector constants that |
| 371 | /// satisfy a specified predicate, and bind them to an APInt. |
| 372 | template <typename Predicate> struct api_pred_ty : public Predicate { |
| 373 | const APInt *&Res; |
| 374 | |
| 375 | api_pred_ty(const APInt *&R) : Res(R) {} |
| 376 | |
| 377 | template <typename ITy> bool match(ITy *V) { |
| 378 | if (const auto *CI = dyn_cast<ConstantInt>(V)) |
| 379 | if (this->isValue(CI->getValue())) { |
| 380 | Res = &CI->getValue(); |
| 381 | return true; |
| 382 | } |
| 383 | if (V->getType()->isVectorTy()) |
| 384 | if (const auto *C = dyn_cast<Constant>(V)) |
| 385 | if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) |
| 386 | if (this->isValue(CI->getValue())) { |
| 387 | Res = &CI->getValue(); |
| 388 | return true; |
| 389 | } |
| 390 | |
| 391 | return false; |
| 392 | } |
| 393 | }; |
| 394 | |
| 395 | /// This helper class is used to match scalar and vector constants that |
| 396 | /// satisfy a specified predicate, and bind them to an APFloat. |
| 397 | /// Undefs are allowed in splat vector constants. |
| 398 | template <typename Predicate> struct apf_pred_ty : public Predicate { |
| 399 | const APFloat *&Res; |
| 400 | |
| 401 | apf_pred_ty(const APFloat *&R) : Res(R) {} |
| 402 | |
| 403 | template <typename ITy> bool match(ITy *V) { |
| 404 | if (const auto *CI = dyn_cast<ConstantFP>(V)) |
| 405 | if (this->isValue(CI->getValue())) { |
| 406 | Res = &CI->getValue(); |
| 407 | return true; |
| 408 | } |
| 409 | if (V->getType()->isVectorTy()) |
| 410 | if (const auto *C = dyn_cast<Constant>(V)) |
| 411 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
| 412 | C->getSplatValue(/* AllowUndef */ true))) |
| 413 | if (this->isValue(CI->getValue())) { |
| 414 | Res = &CI->getValue(); |
| 415 | return true; |
| 416 | } |
| 417 | |
| 418 | return false; |
| 419 | } |
| 420 | }; |
| 421 | |
| 422 | /////////////////////////////////////////////////////////////////////////////// |
| 423 | // |
| 424 | // Encapsulate constant value queries for use in templated predicate matchers. |
| 425 | // This allows checking if constants match using compound predicates and works |
| 426 | // with vector constants, possibly with relaxed constraints. For example, ignore |
| 427 | // undef values. |
| 428 | // |
| 429 | /////////////////////////////////////////////////////////////////////////////// |
| 430 | |
| 431 | struct is_any_apint { |
| 432 | bool isValue(const APInt &C) { return true; } |
| 433 | }; |
| 434 | /// Match an integer or vector with any integral constant. |
| 435 | /// For vectors, this includes constants with undefined elements. |
| 436 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { |
| 437 | return cst_pred_ty<is_any_apint>(); |
| 438 | } |
| 439 | |
| 440 | struct is_all_ones { |
| 441 | bool isValue(const APInt &C) { return C.isAllOnes(); } |
| 442 | }; |
| 443 | /// Match an integer or vector with all bits set. |
| 444 | /// For vectors, this includes constants with undefined elements. |
| 445 | inline cst_pred_ty<is_all_ones> m_AllOnes() { |
| 446 | return cst_pred_ty<is_all_ones>(); |
| 447 | } |
| 448 | |
| 449 | struct is_maxsignedvalue { |
| 450 | bool isValue(const APInt &C) { return C.isMaxSignedValue(); } |
| 451 | }; |
| 452 | /// Match an integer or vector with values having all bits except for the high |
| 453 | /// bit set (0x7f...). |
| 454 | /// For vectors, this includes constants with undefined elements. |
| 455 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { |
| 456 | return cst_pred_ty<is_maxsignedvalue>(); |
| 457 | } |
| 458 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { |
| 459 | return V; |
| 460 | } |
| 461 | |
| 462 | struct is_negative { |
| 463 | bool isValue(const APInt &C) { return C.isNegative(); } |
| 464 | }; |
| 465 | /// Match an integer or vector of negative values. |
| 466 | /// For vectors, this includes constants with undefined elements. |
| 467 | inline cst_pred_ty<is_negative> m_Negative() { |
| 468 | return cst_pred_ty<is_negative>(); |
| 469 | } |
| 470 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { |
| 471 | return V; |
| 472 | } |
| 473 | |
| 474 | struct is_nonnegative { |
| 475 | bool isValue(const APInt &C) { return C.isNonNegative(); } |
| 476 | }; |
| 477 | /// Match an integer or vector of non-negative values. |
| 478 | /// For vectors, this includes constants with undefined elements. |
| 479 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { |
| 480 | return cst_pred_ty<is_nonnegative>(); |
| 481 | } |
| 482 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { |
| 483 | return V; |
| 484 | } |
| 485 | |
| 486 | struct is_strictlypositive { |
| 487 | bool isValue(const APInt &C) { return C.isStrictlyPositive(); } |
| 488 | }; |
| 489 | /// Match an integer or vector of strictly positive values. |
| 490 | /// For vectors, this includes constants with undefined elements. |
| 491 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { |
| 492 | return cst_pred_ty<is_strictlypositive>(); |
| 493 | } |
| 494 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { |
| 495 | return V; |
| 496 | } |
| 497 | |
| 498 | struct is_nonpositive { |
| 499 | bool isValue(const APInt &C) { return C.isNonPositive(); } |
| 500 | }; |
| 501 | /// Match an integer or vector of non-positive values. |
| 502 | /// For vectors, this includes constants with undefined elements. |
| 503 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { |
| 504 | return cst_pred_ty<is_nonpositive>(); |
| 505 | } |
| 506 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } |
| 507 | |
| 508 | struct is_one { |
| 509 | bool isValue(const APInt &C) { return C.isOne(); } |
| 510 | }; |
| 511 | /// Match an integer 1 or a vector with all elements equal to 1. |
| 512 | /// For vectors, this includes constants with undefined elements. |
| 513 | inline cst_pred_ty<is_one> m_One() { |
| 514 | return cst_pred_ty<is_one>(); |
| 515 | } |
| 516 | |
| 517 | struct is_zero_int { |
| 518 | bool isValue(const APInt &C) { return C.isZero(); } |
| 519 | }; |
| 520 | /// Match an integer 0 or a vector with all elements equal to 0. |
| 521 | /// For vectors, this includes constants with undefined elements. |
| 522 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { |
| 523 | return cst_pred_ty<is_zero_int>(); |
| 524 | } |
| 525 | |
| 526 | struct is_zero { |
| 527 | template <typename ITy> bool match(ITy *V) { |
| 528 | auto *C = dyn_cast<Constant>(V); |
| 529 | // FIXME: this should be able to do something for scalable vectors |
| 530 | return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C)); |
| 531 | } |
| 532 | }; |
| 533 | /// Match any null constant or a vector with all elements equal to 0. |
| 534 | /// For vectors, this includes constants with undefined elements. |
| 535 | inline is_zero m_Zero() { |
| 536 | return is_zero(); |
| 537 | } |
| 538 | |
| 539 | struct is_power2 { |
| 540 | bool isValue(const APInt &C) { return C.isPowerOf2(); } |
| 541 | }; |
| 542 | /// Match an integer or vector power-of-2. |
| 543 | /// For vectors, this includes constants with undefined elements. |
| 544 | inline cst_pred_ty<is_power2> m_Power2() { |
| 545 | return cst_pred_ty<is_power2>(); |
| 546 | } |
| 547 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { |
| 548 | return V; |
| 549 | } |
| 550 | |
| 551 | struct is_negated_power2 { |
| 552 | bool isValue(const APInt &C) { return C.isNegatedPowerOf2(); } |
| 553 | }; |
| 554 | /// Match a integer or vector negated power-of-2. |
| 555 | /// For vectors, this includes constants with undefined elements. |
| 556 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { |
| 557 | return cst_pred_ty<is_negated_power2>(); |
| 558 | } |
| 559 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { |
| 560 | return V; |
| 561 | } |
| 562 | |
| 563 | struct is_power2_or_zero { |
| 564 | bool isValue(const APInt &C) { return !C || C.isPowerOf2(); } |
| 565 | }; |
| 566 | /// Match an integer or vector of 0 or power-of-2 values. |
| 567 | /// For vectors, this includes constants with undefined elements. |
| 568 | inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() { |
| 569 | return cst_pred_ty<is_power2_or_zero>(); |
| 570 | } |
| 571 | inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) { |
| 572 | return V; |
| 573 | } |
| 574 | |
| 575 | struct is_sign_mask { |
| 576 | bool isValue(const APInt &C) { return C.isSignMask(); } |
| 577 | }; |
| 578 | /// Match an integer or vector with only the sign bit(s) set. |
| 579 | /// For vectors, this includes constants with undefined elements. |
| 580 | inline cst_pred_ty<is_sign_mask> m_SignMask() { |
| 581 | return cst_pred_ty<is_sign_mask>(); |
| 582 | } |
| 583 | |
| 584 | struct is_lowbit_mask { |
| 585 | bool isValue(const APInt &C) { return C.isMask(); } |
| 586 | }; |
| 587 | /// Match an integer or vector with only the low bit(s) set. |
| 588 | /// For vectors, this includes constants with undefined elements. |
| 589 | inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() { |
| 590 | return cst_pred_ty<is_lowbit_mask>(); |
| 591 | } |
| 592 | inline api_pred_ty<is_lowbit_mask> m_LowBitMask(const APInt *&V) { |
| 593 | return V; |
| 594 | } |
| 595 | |
| 596 | struct icmp_pred_with_threshold { |
| 597 | ICmpInst::Predicate Pred; |
| 598 | const APInt *Thr; |
| 599 | bool isValue(const APInt &C) { return ICmpInst::compare(C, *Thr, Pred); } |
| 600 | }; |
| 601 | /// Match an integer or vector with every element comparing 'pred' (eg/ne/...) |
| 602 | /// to Threshold. For vectors, this includes constants with undefined elements. |
| 603 | inline cst_pred_ty<icmp_pred_with_threshold> |
| 604 | m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) { |
| 605 | cst_pred_ty<icmp_pred_with_threshold> P; |
| 606 | P.Pred = Predicate; |
| 607 | P.Thr = &Threshold; |
| 608 | return P; |
| 609 | } |
| 610 | |
| 611 | struct is_nan { |
| 612 | bool isValue(const APFloat &C) { return C.isNaN(); } |
| 613 | }; |
| 614 | /// Match an arbitrary NaN constant. This includes quiet and signalling nans. |
| 615 | /// For vectors, this includes constants with undefined elements. |
| 616 | inline cstfp_pred_ty<is_nan> m_NaN() { |
| 617 | return cstfp_pred_ty<is_nan>(); |
| 618 | } |
| 619 | |
| 620 | struct is_nonnan { |
| 621 | bool isValue(const APFloat &C) { return !C.isNaN(); } |
| 622 | }; |
| 623 | /// Match a non-NaN FP constant. |
| 624 | /// For vectors, this includes constants with undefined elements. |
| 625 | inline cstfp_pred_ty<is_nonnan> m_NonNaN() { |
| 626 | return cstfp_pred_ty<is_nonnan>(); |
| 627 | } |
| 628 | |
| 629 | struct is_inf { |
| 630 | bool isValue(const APFloat &C) { return C.isInfinity(); } |
| 631 | }; |
| 632 | /// Match a positive or negative infinity FP constant. |
| 633 | /// For vectors, this includes constants with undefined elements. |
| 634 | inline cstfp_pred_ty<is_inf> m_Inf() { |
| 635 | return cstfp_pred_ty<is_inf>(); |
| 636 | } |
| 637 | |
| 638 | struct is_noninf { |
| 639 | bool isValue(const APFloat &C) { return !C.isInfinity(); } |
| 640 | }; |
| 641 | /// Match a non-infinity FP constant, i.e. finite or NaN. |
| 642 | /// For vectors, this includes constants with undefined elements. |
| 643 | inline cstfp_pred_ty<is_noninf> m_NonInf() { |
| 644 | return cstfp_pred_ty<is_noninf>(); |
| 645 | } |
| 646 | |
| 647 | struct is_finite { |
| 648 | bool isValue(const APFloat &C) { return C.isFinite(); } |
| 649 | }; |
| 650 | /// Match a finite FP constant, i.e. not infinity or NaN. |
| 651 | /// For vectors, this includes constants with undefined elements. |
| 652 | inline cstfp_pred_ty<is_finite> m_Finite() { |
| 653 | return cstfp_pred_ty<is_finite>(); |
| 654 | } |
| 655 | inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; } |
| 656 | |
| 657 | struct is_finitenonzero { |
| 658 | bool isValue(const APFloat &C) { return C.isFiniteNonZero(); } |
| 659 | }; |
| 660 | /// Match a finite non-zero FP constant. |
| 661 | /// For vectors, this includes constants with undefined elements. |
| 662 | inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() { |
| 663 | return cstfp_pred_ty<is_finitenonzero>(); |
| 664 | } |
| 665 | inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) { |
| 666 | return V; |
| 667 | } |
| 668 | |
| 669 | struct is_any_zero_fp { |
| 670 | bool isValue(const APFloat &C) { return C.isZero(); } |
| 671 | }; |
| 672 | /// Match a floating-point negative zero or positive zero. |
| 673 | /// For vectors, this includes constants with undefined elements. |
| 674 | inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() { |
| 675 | return cstfp_pred_ty<is_any_zero_fp>(); |
| 676 | } |
| 677 | |
| 678 | struct is_pos_zero_fp { |
| 679 | bool isValue(const APFloat &C) { return C.isPosZero(); } |
| 680 | }; |
| 681 | /// Match a floating-point positive zero. |
| 682 | /// For vectors, this includes constants with undefined elements. |
| 683 | inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() { |
| 684 | return cstfp_pred_ty<is_pos_zero_fp>(); |
| 685 | } |
| 686 | |
| 687 | struct is_neg_zero_fp { |
| 688 | bool isValue(const APFloat &C) { return C.isNegZero(); } |
| 689 | }; |
| 690 | /// Match a floating-point negative zero. |
| 691 | /// For vectors, this includes constants with undefined elements. |
| 692 | inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() { |
| 693 | return cstfp_pred_ty<is_neg_zero_fp>(); |
| 694 | } |
| 695 | |
| 696 | struct is_non_zero_fp { |
| 697 | bool isValue(const APFloat &C) { return C.isNonZero(); } |
| 698 | }; |
| 699 | /// Match a floating-point non-zero. |
| 700 | /// For vectors, this includes constants with undefined elements. |
| 701 | inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() { |
| 702 | return cstfp_pred_ty<is_non_zero_fp>(); |
| 703 | } |
| 704 | |
| 705 | /////////////////////////////////////////////////////////////////////////////// |
| 706 | |
| 707 | template <typename Class> struct bind_ty { |
| 708 | Class *&VR; |
| 709 | |
| 710 | bind_ty(Class *&V) : VR(V) {} |
| 711 | |
| 712 | template <typename ITy> bool match(ITy *V) { |
| 713 | if (auto *CV = dyn_cast<Class>(V)) { |
| 714 | VR = CV; |
| 715 | return true; |
| 716 | } |
| 717 | return false; |
| 718 | } |
| 719 | }; |
| 720 | |
| 721 | /// Match a value, capturing it if we match. |
| 722 | inline bind_ty<Value> m_Value(Value *&V) { return V; } |
| 723 | inline bind_ty<const Value> m_Value(const Value *&V) { return V; } |
| 724 | |
| 725 | /// Match an instruction, capturing it if we match. |
| 726 | inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; } |
| 727 | /// Match a unary operator, capturing it if we match. |
| 728 | inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; } |
| 729 | /// Match a binary operator, capturing it if we match. |
| 730 | inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; } |
| 731 | /// Match a with overflow intrinsic, capturing it if we match. |
| 732 | inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; } |
| 733 | inline bind_ty<const WithOverflowInst> |
| 734 | m_WithOverflowInst(const WithOverflowInst *&I) { |
| 735 | return I; |
| 736 | } |
| 737 | |
| 738 | /// Match a Constant, capturing the value if we match. |
| 739 | inline bind_ty<Constant> m_Constant(Constant *&C) { return C; } |
| 740 | |
| 741 | /// Match a ConstantInt, capturing the value if we match. |
| 742 | inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; } |
| 743 | |
| 744 | /// Match a ConstantFP, capturing the value if we match. |
| 745 | inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; } |
| 746 | |
| 747 | /// Match a ConstantExpr, capturing the value if we match. |
| 748 | inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; } |
| 749 | |
| 750 | /// Match a basic block value, capturing it if we match. |
| 751 | inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; } |
| 752 | inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) { |
| 753 | return V; |
| 754 | } |
| 755 | |
| 756 | /// Match an arbitrary immediate Constant and ignore it. |
| 757 | inline match_combine_and<class_match<Constant>, |
| 758 | match_unless<class_match<ConstantExpr>>> |
| 759 | m_ImmConstant() { |
| 760 | return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr())); |
| 761 | } |
| 762 | |
| 763 | /// Match an immediate Constant, capturing the value if we match. |
| 764 | inline match_combine_and<bind_ty<Constant>, |
| 765 | match_unless<class_match<ConstantExpr>>> |
| 766 | m_ImmConstant(Constant *&C) { |
| 767 | return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr())); |
| 768 | } |
| 769 | |
| 770 | /// Match a specified Value*. |
| 771 | struct specificval_ty { |
| 772 | const Value *Val; |
| 773 | |
| 774 | specificval_ty(const Value *V) : Val(V) {} |
| 775 | |
| 776 | template <typename ITy> bool match(ITy *V) { return V == Val; } |
| 777 | }; |
| 778 | |
| 779 | /// Match if we have a specific specified value. |
| 780 | inline specificval_ty m_Specific(const Value *V) { return V; } |
| 781 | |
| 782 | /// Stores a reference to the Value *, not the Value * itself, |
| 783 | /// thus can be used in commutative matchers. |
| 784 | template <typename Class> struct deferredval_ty { |
| 785 | Class *const &Val; |
| 786 | |
| 787 | deferredval_ty(Class *const &V) : Val(V) {} |
| 788 | |
| 789 | template <typename ITy> bool match(ITy *const V) { return V == Val; } |
| 790 | }; |
| 791 | |
| 792 | /// Like m_Specific(), but works if the specific value to match is determined |
| 793 | /// as part of the same match() expression. For example: |
| 794 | /// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will |
| 795 | /// bind X before the pattern match starts. |
| 796 | /// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against |
| 797 | /// whichever value m_Value(X) populated. |
| 798 | inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; } |
| 799 | inline deferredval_ty<const Value> m_Deferred(const Value *const &V) { |
| 800 | return V; |
| 801 | } |
| 802 | |
| 803 | /// Match a specified floating point value or vector of all elements of |
| 804 | /// that value. |
| 805 | struct specific_fpval { |
| 806 | double Val; |
| 807 | |
| 808 | specific_fpval(double V) : Val(V) {} |
| 809 | |
| 810 | template <typename ITy> bool match(ITy *V) { |
| 811 | if (const auto *CFP = dyn_cast<ConstantFP>(V)) |
| 812 | return CFP->isExactlyValue(Val); |
| 813 | if (V->getType()->isVectorTy()) |
| 814 | if (const auto *C = dyn_cast<Constant>(V)) |
| 815 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) |
| 816 | return CFP->isExactlyValue(Val); |
| 817 | return false; |
| 818 | } |
| 819 | }; |
| 820 | |
| 821 | /// Match a specific floating point value or vector with all elements |
| 822 | /// equal to the value. |
| 823 | inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); } |
| 824 | |
| 825 | /// Match a float 1.0 or vector with all elements equal to 1.0. |
| 826 | inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); } |
| 827 | |
| 828 | struct bind_const_intval_ty { |
| 829 | uint64_t &VR; |
| 830 | |
| 831 | bind_const_intval_ty(uint64_t &V) : VR(V) {} |
| 832 | |
| 833 | template <typename ITy> bool match(ITy *V) { |
| 834 | if (const auto *CV = dyn_cast<ConstantInt>(V)) |
| 835 | if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) { |
| 836 | VR = CV->getZExtValue(); |
| 837 | return true; |
| 838 | } |
| 839 | return false; |
| 840 | } |
| 841 | }; |
| 842 | |
| 843 | /// Match a specified integer value or vector of all elements of that |
| 844 | /// value. |
| 845 | template <bool AllowUndefs> |
| 846 | struct specific_intval { |
| 847 | APInt Val; |
| 848 | |
| 849 | specific_intval(APInt V) : Val(std::move(V)) {} |
| 850 | |
| 851 | template <typename ITy> bool match(ITy *V) { |
| 852 | const auto *CI = dyn_cast<ConstantInt>(V); |
| 853 | if (!CI && V->getType()->isVectorTy()) |
| 854 | if (const auto *C = dyn_cast<Constant>(V)) |
| 855 | CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs)); |
| 856 | |
| 857 | return CI && APInt::isSameValue(CI->getValue(), Val); |
| 858 | } |
| 859 | }; |
| 860 | |
| 861 | /// Match a specific integer value or vector with all elements equal to |
| 862 | /// the value. |
| 863 | inline specific_intval<false> m_SpecificInt(APInt V) { |
| 864 | return specific_intval<false>(std::move(V)); |
| 865 | } |
| 866 | |
| 867 | inline specific_intval<false> m_SpecificInt(uint64_t V) { |
| 868 | return m_SpecificInt(APInt(64, V)); |
| 869 | } |
| 870 | |
| 871 | inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) { |
| 872 | return specific_intval<true>(std::move(V)); |
| 873 | } |
| 874 | |
| 875 | inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) { |
| 876 | return m_SpecificIntAllowUndef(APInt(64, V)); |
| 877 | } |
| 878 | |
| 879 | /// Match a ConstantInt and bind to its value. This does not match |
| 880 | /// ConstantInts wider than 64-bits. |
| 881 | inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; } |
| 882 | |
| 883 | /// Match a specified basic block value. |
| 884 | struct specific_bbval { |
| 885 | BasicBlock *Val; |
| 886 | |
| 887 | specific_bbval(BasicBlock *Val) : Val(Val) {} |
| 888 | |
| 889 | template <typename ITy> bool match(ITy *V) { |
| 890 | const auto *BB = dyn_cast<BasicBlock>(V); |
| 891 | return BB && BB == Val; |
| 892 | } |
| 893 | }; |
| 894 | |
| 895 | /// Match a specific basic block value. |
| 896 | inline specific_bbval m_SpecificBB(BasicBlock *BB) { |
| 897 | return specific_bbval(BB); |
| 898 | } |
| 899 | |
| 900 | /// A commutative-friendly version of m_Specific(). |
| 901 | inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) { |
| 902 | return BB; |
| 903 | } |
| 904 | inline deferredval_ty<const BasicBlock> |
| 905 | m_Deferred(const BasicBlock *const &BB) { |
| 906 | return BB; |
| 907 | } |
| 908 | |
| 909 | //===----------------------------------------------------------------------===// |
| 910 | // Matcher for any binary operator. |
| 911 | // |
| 912 | template <typename LHS_t, typename RHS_t, bool Commutable = false> |
| 913 | struct AnyBinaryOp_match { |
| 914 | LHS_t L; |
| 915 | RHS_t R; |
| 916 | |
| 917 | // The evaluation order is always stable, regardless of Commutability. |
| 918 | // The LHS is always matched first. |
| 919 | AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
| 920 | |
| 921 | template <typename OpTy> bool match(OpTy *V) { |
| 922 | if (auto *I = dyn_cast<BinaryOperator>(V)) |
| 923 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || |
| 924 | (Commutable && L.match(I->getOperand(1)) && |
| 925 | R.match(I->getOperand(0))); |
| 926 | return false; |
| 927 | } |
| 928 | }; |
| 929 | |
| 930 | template <typename LHS, typename RHS> |
| 931 | inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) { |
| 932 | return AnyBinaryOp_match<LHS, RHS>(L, R); |
| 933 | } |
| 934 | |
| 935 | //===----------------------------------------------------------------------===// |
| 936 | // Matcher for any unary operator. |
| 937 | // TODO fuse unary, binary matcher into n-ary matcher |
| 938 | // |
| 939 | template <typename OP_t> struct AnyUnaryOp_match { |
| 940 | OP_t X; |
| 941 | |
| 942 | AnyUnaryOp_match(const OP_t &X) : X(X) {} |
| 943 | |
| 944 | template <typename OpTy> bool match(OpTy *V) { |
| 945 | if (auto *I = dyn_cast<UnaryOperator>(V)) |
| 946 | return X.match(I->getOperand(0)); |
| 947 | return false; |
| 948 | } |
| 949 | }; |
| 950 | |
| 951 | template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) { |
| 952 | return AnyUnaryOp_match<OP_t>(X); |
| 953 | } |
| 954 | |
| 955 | //===----------------------------------------------------------------------===// |
| 956 | // Matchers for specific binary operators. |
| 957 | // |
| 958 | |
| 959 | template <typename LHS_t, typename RHS_t, unsigned Opcode, |
| 960 | bool Commutable = false> |
| 961 | struct BinaryOp_match { |
| 962 | LHS_t L; |
| 963 | RHS_t R; |
| 964 | |
| 965 | // The evaluation order is always stable, regardless of Commutability. |
| 966 | // The LHS is always matched first. |
| 967 | BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
| 968 | |
| 969 | template <typename OpTy> inline bool match(unsigned Opc, OpTy *V) { |
| 970 | if (V->getValueID() == Value::InstructionVal + Opc) { |
| 971 | auto *I = cast<BinaryOperator>(V); |
| 972 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || |
| 973 | (Commutable && L.match(I->getOperand(1)) && |
| 974 | R.match(I->getOperand(0))); |
| 975 | } |
| 976 | if (auto *CE = dyn_cast<ConstantExpr>(V)) |
| 977 | return CE->getOpcode() == Opc && |
| 978 | ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) || |
| 979 | (Commutable && L.match(CE->getOperand(1)) && |
| 980 | R.match(CE->getOperand(0)))); |
| 981 | return false; |
| 982 | } |
| 983 | |
| 984 | template <typename OpTy> bool match(OpTy *V) { return match(Opcode, V); } |
| 985 | }; |
| 986 | |
| 987 | template <typename LHS, typename RHS> |
| 988 | inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L, |
| 989 | const RHS &R) { |
| 990 | return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R); |
| 991 | } |
| 992 | |
| 993 | template <typename LHS, typename RHS> |
| 994 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L, |
| 995 | const RHS &R) { |
| 996 | return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R); |
| 997 | } |
| 998 | |
| 999 | template <typename LHS, typename RHS> |
| 1000 | inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L, |
| 1001 | const RHS &R) { |
| 1002 | return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R); |
| 1003 | } |
| 1004 | |
| 1005 | template <typename LHS, typename RHS> |
| 1006 | inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L, |
| 1007 | const RHS &R) { |
| 1008 | return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R); |
| 1009 | } |
| 1010 | |
| 1011 | template <typename Op_t> struct FNeg_match { |
| 1012 | Op_t X; |
| 1013 | |
| 1014 | FNeg_match(const Op_t &Op) : X(Op) {} |
| 1015 | template <typename OpTy> bool match(OpTy *V) { |
| 1016 | auto *FPMO = dyn_cast<FPMathOperator>(V); |
| 1017 | if (!FPMO) return false; |
| 1018 | |
| 1019 | if (FPMO->getOpcode() == Instruction::FNeg) |
| 1020 | return X.match(FPMO->getOperand(0)); |
| 1021 | |
| 1022 | if (FPMO->getOpcode() == Instruction::FSub) { |
| 1023 | if (FPMO->hasNoSignedZeros()) { |
| 1024 | // With 'nsz', any zero goes. |
| 1025 | if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0))) |
| 1026 | return false; |
| 1027 | } else { |
| 1028 | // Without 'nsz', we need fsub -0.0, X exactly. |
| 1029 | if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0))) |
| 1030 | return false; |
| 1031 | } |
| 1032 | |
| 1033 | return X.match(FPMO->getOperand(1)); |
| 1034 | } |
| 1035 | |
| 1036 | return false; |
| 1037 | } |
| 1038 | }; |
| 1039 | |
| 1040 | /// Match 'fneg X' as 'fsub -0.0, X'. |
| 1041 | template <typename OpTy> |
| 1042 | inline FNeg_match<OpTy> |
| 1043 | m_FNeg(const OpTy &X) { |
| 1044 | return FNeg_match<OpTy>(X); |
| 1045 | } |
| 1046 | |
| 1047 | /// Match 'fneg X' as 'fsub +-0.0, X'. |
| 1048 | template <typename RHS> |
| 1049 | inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub> |
| 1050 | m_FNegNSZ(const RHS &X) { |
| 1051 | return m_FSub(m_AnyZeroFP(), X); |
| 1052 | } |
| 1053 | |
| 1054 | template <typename LHS, typename RHS> |
| 1055 | inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L, |
| 1056 | const RHS &R) { |
| 1057 | return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R); |
| 1058 | } |
| 1059 | |
| 1060 | template <typename LHS, typename RHS> |
| 1061 | inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L, |
| 1062 | const RHS &R) { |
| 1063 | return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R); |
| 1064 | } |
| 1065 | |
| 1066 | template <typename LHS, typename RHS> |
| 1067 | inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L, |
| 1068 | const RHS &R) { |
| 1069 | return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R); |
| 1070 | } |
| 1071 | |
| 1072 | template <typename LHS, typename RHS> |
| 1073 | inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L, |
| 1074 | const RHS &R) { |
| 1075 | return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R); |
| 1076 | } |
| 1077 | |
| 1078 | template <typename LHS, typename RHS> |
| 1079 | inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L, |
| 1080 | const RHS &R) { |
| 1081 | return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R); |
| 1082 | } |
| 1083 | |
| 1084 | template <typename LHS, typename RHS> |
| 1085 | inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L, |
| 1086 | const RHS &R) { |
| 1087 | return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R); |
| 1088 | } |
| 1089 | |
| 1090 | template <typename LHS, typename RHS> |
| 1091 | inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L, |
| 1092 | const RHS &R) { |
| 1093 | return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R); |
| 1094 | } |
| 1095 | |
| 1096 | template <typename LHS, typename RHS> |
| 1097 | inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L, |
| 1098 | const RHS &R) { |
| 1099 | return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R); |
| 1100 | } |
| 1101 | |
| 1102 | template <typename LHS, typename RHS> |
| 1103 | inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L, |
| 1104 | const RHS &R) { |
| 1105 | return BinaryOp_match<LHS, RHS, Instruction::And>(L, R); |
| 1106 | } |
| 1107 | |
| 1108 | template <typename LHS, typename RHS> |
| 1109 | inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L, |
| 1110 | const RHS &R) { |
| 1111 | return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R); |
| 1112 | } |
| 1113 | |
| 1114 | template <typename LHS, typename RHS> |
| 1115 | inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L, |
| 1116 | const RHS &R) { |
| 1117 | return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R); |
| 1118 | } |
| 1119 | |
| 1120 | template <typename LHS, typename RHS> |
| 1121 | inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L, |
| 1122 | const RHS &R) { |
| 1123 | return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R); |
| 1124 | } |
| 1125 | |
| 1126 | template <typename LHS, typename RHS> |
| 1127 | inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L, |
| 1128 | const RHS &R) { |
| 1129 | return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R); |
| 1130 | } |
| 1131 | |
| 1132 | template <typename LHS, typename RHS> |
| 1133 | inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L, |
| 1134 | const RHS &R) { |
| 1135 | return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R); |
| 1136 | } |
| 1137 | |
| 1138 | template <typename LHS_t, typename RHS_t, unsigned Opcode, |
| 1139 | unsigned WrapFlags = 0> |
| 1140 | struct OverflowingBinaryOp_match { |
| 1141 | LHS_t L; |
| 1142 | RHS_t R; |
| 1143 | |
| 1144 | OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) |
| 1145 | : L(LHS), R(RHS) {} |
| 1146 | |
| 1147 | template <typename OpTy> bool match(OpTy *V) { |
| 1148 | if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) { |
| 1149 | if (Op->getOpcode() != Opcode) |
| 1150 | return false; |
| 1151 | if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) && |
| 1152 | !Op->hasNoUnsignedWrap()) |
| 1153 | return false; |
| 1154 | if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) && |
| 1155 | !Op->hasNoSignedWrap()) |
| 1156 | return false; |
| 1157 | return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1)); |
| 1158 | } |
| 1159 | return false; |
| 1160 | } |
| 1161 | }; |
| 1162 | |
| 1163 | template <typename LHS, typename RHS> |
| 1164 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
| 1165 | OverflowingBinaryOperator::NoSignedWrap> |
| 1166 | m_NSWAdd(const LHS &L, const RHS &R) { |
| 1167 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
| 1168 | OverflowingBinaryOperator::NoSignedWrap>( |
| 1169 | L, R); |
| 1170 | } |
| 1171 | template <typename LHS, typename RHS> |
| 1172 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
| 1173 | OverflowingBinaryOperator::NoSignedWrap> |
| 1174 | m_NSWSub(const LHS &L, const RHS &R) { |
| 1175 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
| 1176 | OverflowingBinaryOperator::NoSignedWrap>( |
| 1177 | L, R); |
| 1178 | } |
| 1179 | template <typename LHS, typename RHS> |
| 1180 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
| 1181 | OverflowingBinaryOperator::NoSignedWrap> |
| 1182 | m_NSWMul(const LHS &L, const RHS &R) { |
| 1183 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
| 1184 | OverflowingBinaryOperator::NoSignedWrap>( |
| 1185 | L, R); |
| 1186 | } |
| 1187 | template <typename LHS, typename RHS> |
| 1188 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
| 1189 | OverflowingBinaryOperator::NoSignedWrap> |
| 1190 | m_NSWShl(const LHS &L, const RHS &R) { |
| 1191 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
| 1192 | OverflowingBinaryOperator::NoSignedWrap>( |
| 1193 | L, R); |
| 1194 | } |
| 1195 | |
| 1196 | template <typename LHS, typename RHS> |
| 1197 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
| 1198 | OverflowingBinaryOperator::NoUnsignedWrap> |
| 1199 | m_NUWAdd(const LHS &L, const RHS &R) { |
| 1200 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
| 1201 | OverflowingBinaryOperator::NoUnsignedWrap>( |
| 1202 | L, R); |
| 1203 | } |
| 1204 | template <typename LHS, typename RHS> |
| 1205 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
| 1206 | OverflowingBinaryOperator::NoUnsignedWrap> |
| 1207 | m_NUWSub(const LHS &L, const RHS &R) { |
| 1208 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
| 1209 | OverflowingBinaryOperator::NoUnsignedWrap>( |
| 1210 | L, R); |
| 1211 | } |
| 1212 | template <typename LHS, typename RHS> |
| 1213 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
| 1214 | OverflowingBinaryOperator::NoUnsignedWrap> |
| 1215 | m_NUWMul(const LHS &L, const RHS &R) { |
| 1216 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
| 1217 | OverflowingBinaryOperator::NoUnsignedWrap>( |
| 1218 | L, R); |
| 1219 | } |
| 1220 | template <typename LHS, typename RHS> |
| 1221 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
| 1222 | OverflowingBinaryOperator::NoUnsignedWrap> |
| 1223 | m_NUWShl(const LHS &L, const RHS &R) { |
| 1224 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
| 1225 | OverflowingBinaryOperator::NoUnsignedWrap>( |
| 1226 | L, R); |
| 1227 | } |
| 1228 | |
| 1229 | template <typename LHS_t, typename RHS_t, bool Commutable = false> |
| 1230 | struct SpecificBinaryOp_match |
| 1231 | : public BinaryOp_match<LHS_t, RHS_t, 0, Commutable> { |
| 1232 | unsigned Opcode; |
| 1233 | |
| 1234 | SpecificBinaryOp_match(unsigned Opcode, const LHS_t &LHS, const RHS_t &RHS) |
| 1235 | : BinaryOp_match<LHS_t, RHS_t, 0, Commutable>(LHS, RHS), Opcode(Opcode) {} |
| 1236 | |
| 1237 | template <typename OpTy> bool match(OpTy *V) { |
| 1238 | return BinaryOp_match<LHS_t, RHS_t, 0, Commutable>::match(Opcode, V); |
| 1239 | } |
| 1240 | }; |
| 1241 | |
| 1242 | /// Matches a specific opcode. |
| 1243 | template <typename LHS, typename RHS> |
| 1244 | inline SpecificBinaryOp_match<LHS, RHS> m_BinOp(unsigned Opcode, const LHS &L, |
| 1245 | const RHS &R) { |
| 1246 | return SpecificBinaryOp_match<LHS, RHS>(Opcode, L, R); |
| 1247 | } |
| 1248 | |
| 1249 | //===----------------------------------------------------------------------===// |
| 1250 | // Class that matches a group of binary opcodes. |
| 1251 | // |
| 1252 | template <typename LHS_t, typename RHS_t, typename Predicate> |
| 1253 | struct BinOpPred_match : Predicate { |
| 1254 | LHS_t L; |
| 1255 | RHS_t R; |
| 1256 | |
| 1257 | BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
| 1258 | |
| 1259 | template <typename OpTy> bool match(OpTy *V) { |
| 1260 | if (auto *I = dyn_cast<Instruction>(V)) |
| 1261 | return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) && |
| 1262 | R.match(I->getOperand(1)); |
| 1263 | if (auto *CE = dyn_cast<ConstantExpr>(V)) |
| 1264 | return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) && |
| 1265 | R.match(CE->getOperand(1)); |
| 1266 | return false; |
| 1267 | } |
| 1268 | }; |
| 1269 | |
| 1270 | struct is_shift_op { |
| 1271 | bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); } |
| 1272 | }; |
| 1273 | |
| 1274 | struct is_right_shift_op { |
| 1275 | bool isOpType(unsigned Opcode) { |
| 1276 | return Opcode == Instruction::LShr || Opcode == Instruction::AShr; |
| 1277 | } |
| 1278 | }; |
| 1279 | |
| 1280 | struct is_logical_shift_op { |
| 1281 | bool isOpType(unsigned Opcode) { |
| 1282 | return Opcode == Instruction::LShr || Opcode == Instruction::Shl; |
| 1283 | } |
| 1284 | }; |
| 1285 | |
| 1286 | struct is_bitwiselogic_op { |
| 1287 | bool isOpType(unsigned Opcode) { |
| 1288 | return Instruction::isBitwiseLogicOp(Opcode); |
| 1289 | } |
| 1290 | }; |
| 1291 | |
| 1292 | struct is_idiv_op { |
| 1293 | bool isOpType(unsigned Opcode) { |
| 1294 | return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; |
| 1295 | } |
| 1296 | }; |
| 1297 | |
| 1298 | struct is_irem_op { |
| 1299 | bool isOpType(unsigned Opcode) { |
| 1300 | return Opcode == Instruction::SRem || Opcode == Instruction::URem; |
| 1301 | } |
| 1302 | }; |
| 1303 | |
| 1304 | /// Matches shift operations. |
| 1305 | template <typename LHS, typename RHS> |
| 1306 | inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L, |
| 1307 | const RHS &R) { |
| 1308 | return BinOpPred_match<LHS, RHS, is_shift_op>(L, R); |
| 1309 | } |
| 1310 | |
| 1311 | /// Matches logical shift operations. |
| 1312 | template <typename LHS, typename RHS> |
| 1313 | inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L, |
| 1314 | const RHS &R) { |
| 1315 | return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R); |
| 1316 | } |
| 1317 | |
| 1318 | /// Matches logical shift operations. |
| 1319 | template <typename LHS, typename RHS> |
| 1320 | inline BinOpPred_match<LHS, RHS, is_logical_shift_op> |
| 1321 | m_LogicalShift(const LHS &L, const RHS &R) { |
| 1322 | return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R); |
| 1323 | } |
| 1324 | |
| 1325 | /// Matches bitwise logic operations. |
| 1326 | template <typename LHS, typename RHS> |
| 1327 | inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op> |
| 1328 | m_BitwiseLogic(const LHS &L, const RHS &R) { |
| 1329 | return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R); |
| 1330 | } |
| 1331 | |
| 1332 | /// Matches integer division operations. |
| 1333 | template <typename LHS, typename RHS> |
| 1334 | inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L, |
| 1335 | const RHS &R) { |
| 1336 | return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R); |
| 1337 | } |
| 1338 | |
| 1339 | /// Matches integer remainder operations. |
| 1340 | template <typename LHS, typename RHS> |
| 1341 | inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L, |
| 1342 | const RHS &R) { |
| 1343 | return BinOpPred_match<LHS, RHS, is_irem_op>(L, R); |
| 1344 | } |
| 1345 | |
| 1346 | //===----------------------------------------------------------------------===// |
| 1347 | // Class that matches exact binary ops. |
| 1348 | // |
| 1349 | template <typename SubPattern_t> struct Exact_match { |
| 1350 | SubPattern_t SubPattern; |
| 1351 | |
| 1352 | Exact_match(const SubPattern_t &SP) : SubPattern(SP) {} |
| 1353 | |
| 1354 | template <typename OpTy> bool match(OpTy *V) { |
| 1355 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(V)) |
| 1356 | return PEO->isExact() && SubPattern.match(V); |
| 1357 | return false; |
| 1358 | } |
| 1359 | }; |
| 1360 | |
| 1361 | template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) { |
| 1362 | return SubPattern; |
| 1363 | } |
| 1364 | |
| 1365 | //===----------------------------------------------------------------------===// |
| 1366 | // Matchers for CmpInst classes |
| 1367 | // |
| 1368 | |
| 1369 | template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy, |
| 1370 | bool Commutable = false> |
| 1371 | struct CmpClass_match { |
| 1372 | PredicateTy &Predicate; |
| 1373 | LHS_t L; |
| 1374 | RHS_t R; |
| 1375 | |
| 1376 | // The evaluation order is always stable, regardless of Commutability. |
| 1377 | // The LHS is always matched first. |
| 1378 | CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS) |
| 1379 | : Predicate(Pred), L(LHS), R(RHS) {} |
| 1380 | |
| 1381 | template <typename OpTy> bool match(OpTy *V) { |
| 1382 | if (auto *I = dyn_cast<Class>(V)) { |
| 1383 | if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) { |
| 1384 | Predicate = I->getPredicate(); |
| 1385 | return true; |
| 1386 | } else if (Commutable && L.match(I->getOperand(1)) && |
| 1387 | R.match(I->getOperand(0))) { |
| 1388 | Predicate = I->getSwappedPredicate(); |
| 1389 | return true; |
| 1390 | } |
| 1391 | } |
| 1392 | return false; |
| 1393 | } |
| 1394 | }; |
| 1395 | |
| 1396 | template <typename LHS, typename RHS> |
| 1397 | inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate> |
| 1398 | m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
| 1399 | return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R); |
| 1400 | } |
| 1401 | |
| 1402 | template <typename LHS, typename RHS> |
| 1403 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate> |
| 1404 | m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
| 1405 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R); |
| 1406 | } |
| 1407 | |
| 1408 | template <typename LHS, typename RHS> |
| 1409 | inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate> |
| 1410 | m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
| 1411 | return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R); |
| 1412 | } |
| 1413 | |
| 1414 | //===----------------------------------------------------------------------===// |
| 1415 | // Matchers for instructions with a given opcode and number of operands. |
| 1416 | // |
| 1417 | |
| 1418 | /// Matches instructions with Opcode and three operands. |
| 1419 | template <typename T0, unsigned Opcode> struct OneOps_match { |
| 1420 | T0 Op1; |
| 1421 | |
| 1422 | OneOps_match(const T0 &Op1) : Op1(Op1) {} |
| 1423 | |
| 1424 | template <typename OpTy> bool match(OpTy *V) { |
| 1425 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
| 1426 | auto *I = cast<Instruction>(V); |
| 1427 | return Op1.match(I->getOperand(0)); |
| 1428 | } |
| 1429 | return false; |
| 1430 | } |
| 1431 | }; |
| 1432 | |
| 1433 | /// Matches instructions with Opcode and three operands. |
| 1434 | template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match { |
| 1435 | T0 Op1; |
| 1436 | T1 Op2; |
| 1437 | |
| 1438 | TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {} |
| 1439 | |
| 1440 | template <typename OpTy> bool match(OpTy *V) { |
| 1441 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
| 1442 | auto *I = cast<Instruction>(V); |
| 1443 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)); |
| 1444 | } |
| 1445 | return false; |
| 1446 | } |
| 1447 | }; |
| 1448 | |
| 1449 | /// Matches instructions with Opcode and three operands. |
| 1450 | template <typename T0, typename T1, typename T2, unsigned Opcode> |
| 1451 | struct ThreeOps_match { |
| 1452 | T0 Op1; |
| 1453 | T1 Op2; |
| 1454 | T2 Op3; |
| 1455 | |
| 1456 | ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3) |
| 1457 | : Op1(Op1), Op2(Op2), Op3(Op3) {} |
| 1458 | |
| 1459 | template <typename OpTy> bool match(OpTy *V) { |
| 1460 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
| 1461 | auto *I = cast<Instruction>(V); |
| 1462 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && |
| 1463 | Op3.match(I->getOperand(2)); |
| 1464 | } |
| 1465 | return false; |
| 1466 | } |
| 1467 | }; |
| 1468 | |
| 1469 | /// Matches SelectInst. |
| 1470 | template <typename Cond, typename LHS, typename RHS> |
| 1471 | inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select> |
| 1472 | m_Select(const Cond &C, const LHS &L, const RHS &R) { |
| 1473 | return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R); |
| 1474 | } |
| 1475 | |
| 1476 | /// This matches a select of two constants, e.g.: |
| 1477 | /// m_SelectCst<-1, 0>(m_Value(V)) |
| 1478 | template <int64_t L, int64_t R, typename Cond> |
| 1479 | inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>, |
| 1480 | Instruction::Select> |
| 1481 | m_SelectCst(const Cond &C) { |
| 1482 | return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>()); |
| 1483 | } |
| 1484 | |
| 1485 | /// Matches FreezeInst. |
| 1486 | template <typename OpTy> |
| 1487 | inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) { |
| 1488 | return OneOps_match<OpTy, Instruction::Freeze>(Op); |
| 1489 | } |
| 1490 | |
| 1491 | /// Matches InsertElementInst. |
| 1492 | template <typename Val_t, typename Elt_t, typename Idx_t> |
| 1493 | inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement> |
| 1494 | m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) { |
| 1495 | return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>( |
| 1496 | Val, Elt, Idx); |
| 1497 | } |
| 1498 | |
| 1499 | /// Matches ExtractElementInst. |
| 1500 | template <typename Val_t, typename Idx_t> |
| 1501 | inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement> |
| 1502 | m_ExtractElt(const Val_t &Val, const Idx_t &Idx) { |
| 1503 | return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx); |
| 1504 | } |
| 1505 | |
| 1506 | /// Matches shuffle. |
| 1507 | template <typename T0, typename T1, typename T2> struct Shuffle_match { |
| 1508 | T0 Op1; |
| 1509 | T1 Op2; |
| 1510 | T2 Mask; |
| 1511 | |
| 1512 | Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask) |
| 1513 | : Op1(Op1), Op2(Op2), Mask(Mask) {} |
| 1514 | |
| 1515 | template <typename OpTy> bool match(OpTy *V) { |
| 1516 | if (auto *I = dyn_cast<ShuffleVectorInst>(V)) { |
| 1517 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && |
| 1518 | Mask.match(I->getShuffleMask()); |
| 1519 | } |
| 1520 | return false; |
| 1521 | } |
| 1522 | }; |
| 1523 | |
| 1524 | struct m_Mask { |
| 1525 | ArrayRef<int> &MaskRef; |
| 1526 | m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} |
| 1527 | bool match(ArrayRef<int> Mask) { |
| 1528 | MaskRef = Mask; |
| 1529 | return true; |
| 1530 | } |
| 1531 | }; |
| 1532 | |
| 1533 | struct m_ZeroMask { |
| 1534 | bool match(ArrayRef<int> Mask) { |
| 1535 | return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; }); |
| 1536 | } |
| 1537 | }; |
| 1538 | |
| 1539 | struct m_SpecificMask { |
| 1540 | ArrayRef<int> &MaskRef; |
| 1541 | m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} |
| 1542 | bool match(ArrayRef<int> Mask) { return MaskRef == Mask; } |
| 1543 | }; |
| 1544 | |
| 1545 | struct m_SplatOrUndefMask { |
| 1546 | int &SplatIndex; |
| 1547 | m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {} |
| 1548 | bool match(ArrayRef<int> Mask) { |
| 1549 | auto First = find_if(Mask, [](int Elem) { return Elem != -1; }); |
| 1550 | if (First == Mask.end()) |
| 1551 | return false; |
| 1552 | SplatIndex = *First; |
| 1553 | return all_of(Mask, |
| 1554 | [First](int Elem) { return Elem == *First || Elem == -1; }); |
| 1555 | } |
| 1556 | }; |
| 1557 | |
| 1558 | /// Matches ShuffleVectorInst independently of mask value. |
| 1559 | template <typename V1_t, typename V2_t> |
| 1560 | inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector> |
| 1561 | m_Shuffle(const V1_t &v1, const V2_t &v2) { |
| 1562 | return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2); |
| 1563 | } |
| 1564 | |
| 1565 | template <typename V1_t, typename V2_t, typename Mask_t> |
| 1566 | inline Shuffle_match<V1_t, V2_t, Mask_t> |
| 1567 | m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) { |
| 1568 | return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask); |
| 1569 | } |
| 1570 | |
| 1571 | /// Matches LoadInst. |
| 1572 | template <typename OpTy> |
| 1573 | inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) { |
| 1574 | return OneOps_match<OpTy, Instruction::Load>(Op); |
| 1575 | } |
| 1576 | |
| 1577 | /// Matches StoreInst. |
| 1578 | template <typename ValueOpTy, typename PointerOpTy> |
| 1579 | inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store> |
| 1580 | m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) { |
| 1581 | return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp, |
| 1582 | PointerOp); |
| 1583 | } |
| 1584 | |
| 1585 | //===----------------------------------------------------------------------===// |
| 1586 | // Matchers for CastInst classes |
| 1587 | // |
| 1588 | |
| 1589 | template <typename Op_t, unsigned Opcode> struct CastClass_match { |
| 1590 | Op_t Op; |
| 1591 | |
| 1592 | CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {} |
| 1593 | |
| 1594 | template <typename OpTy> bool match(OpTy *V) { |
| 1595 | if (auto *O = dyn_cast<Operator>(V)) |
| 1596 | return O->getOpcode() == Opcode && Op.match(O->getOperand(0)); |
| 1597 | return false; |
| 1598 | } |
| 1599 | }; |
| 1600 | |
| 1601 | /// Matches BitCast. |
| 1602 | template <typename OpTy> |
| 1603 | inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) { |
| 1604 | return CastClass_match<OpTy, Instruction::BitCast>(Op); |
| 1605 | } |
| 1606 | |
| 1607 | /// Matches PtrToInt. |
| 1608 | template <typename OpTy> |
| 1609 | inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) { |
| 1610 | return CastClass_match<OpTy, Instruction::PtrToInt>(Op); |
| 1611 | } |
| 1612 | |
| 1613 | /// Matches IntToPtr. |
| 1614 | template <typename OpTy> |
| 1615 | inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) { |
| 1616 | return CastClass_match<OpTy, Instruction::IntToPtr>(Op); |
| 1617 | } |
| 1618 | |
| 1619 | /// Matches Trunc. |
| 1620 | template <typename OpTy> |
| 1621 | inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) { |
| 1622 | return CastClass_match<OpTy, Instruction::Trunc>(Op); |
| 1623 | } |
| 1624 | |
| 1625 | template <typename OpTy> |
| 1626 | inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy> |
| 1627 | m_TruncOrSelf(const OpTy &Op) { |
| 1628 | return m_CombineOr(m_Trunc(Op), Op); |
| 1629 | } |
| 1630 | |
| 1631 | /// Matches SExt. |
| 1632 | template <typename OpTy> |
| 1633 | inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) { |
| 1634 | return CastClass_match<OpTy, Instruction::SExt>(Op); |
| 1635 | } |
| 1636 | |
| 1637 | /// Matches ZExt. |
| 1638 | template <typename OpTy> |
| 1639 | inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) { |
| 1640 | return CastClass_match<OpTy, Instruction::ZExt>(Op); |
| 1641 | } |
| 1642 | |
| 1643 | template <typename OpTy> |
| 1644 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy> |
| 1645 | m_ZExtOrSelf(const OpTy &Op) { |
| 1646 | return m_CombineOr(m_ZExt(Op), Op); |
| 1647 | } |
| 1648 | |
| 1649 | template <typename OpTy> |
| 1650 | inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy> |
| 1651 | m_SExtOrSelf(const OpTy &Op) { |
| 1652 | return m_CombineOr(m_SExt(Op), Op); |
| 1653 | } |
| 1654 | |
| 1655 | template <typename OpTy> |
| 1656 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, |
| 1657 | CastClass_match<OpTy, Instruction::SExt>> |
| 1658 | m_ZExtOrSExt(const OpTy &Op) { |
| 1659 | return m_CombineOr(m_ZExt(Op), m_SExt(Op)); |
| 1660 | } |
| 1661 | |
| 1662 | template <typename OpTy> |
| 1663 | inline match_combine_or< |
| 1664 | match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, |
| 1665 | CastClass_match<OpTy, Instruction::SExt>>, |
| 1666 | OpTy> |
| 1667 | m_ZExtOrSExtOrSelf(const OpTy &Op) { |
| 1668 | return m_CombineOr(m_ZExtOrSExt(Op), Op); |
| 1669 | } |
| 1670 | |
| 1671 | template <typename OpTy> |
| 1672 | inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) { |
| 1673 | return CastClass_match<OpTy, Instruction::UIToFP>(Op); |
| 1674 | } |
| 1675 | |
| 1676 | template <typename OpTy> |
| 1677 | inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) { |
| 1678 | return CastClass_match<OpTy, Instruction::SIToFP>(Op); |
| 1679 | } |
| 1680 | |
| 1681 | template <typename OpTy> |
| 1682 | inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) { |
| 1683 | return CastClass_match<OpTy, Instruction::FPToUI>(Op); |
| 1684 | } |
| 1685 | |
| 1686 | template <typename OpTy> |
| 1687 | inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) { |
| 1688 | return CastClass_match<OpTy, Instruction::FPToSI>(Op); |
| 1689 | } |
| 1690 | |
| 1691 | template <typename OpTy> |
| 1692 | inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) { |
| 1693 | return CastClass_match<OpTy, Instruction::FPTrunc>(Op); |
| 1694 | } |
| 1695 | |
| 1696 | template <typename OpTy> |
| 1697 | inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) { |
| 1698 | return CastClass_match<OpTy, Instruction::FPExt>(Op); |
| 1699 | } |
| 1700 | |
| 1701 | //===----------------------------------------------------------------------===// |
| 1702 | // Matchers for control flow. |
| 1703 | // |
| 1704 | |
| 1705 | struct br_match { |
| 1706 | BasicBlock *&Succ; |
| 1707 | |
| 1708 | br_match(BasicBlock *&Succ) : Succ(Succ) {} |
| 1709 | |
| 1710 | template <typename OpTy> bool match(OpTy *V) { |
| 1711 | if (auto *BI = dyn_cast<BranchInst>(V)) |
| 1712 | if (BI->isUnconditional()) { |
| 1713 | Succ = BI->getSuccessor(0); |
| 1714 | return true; |
| 1715 | } |
| 1716 | return false; |
| 1717 | } |
| 1718 | }; |
| 1719 | |
| 1720 | inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); } |
| 1721 | |
| 1722 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> |
| 1723 | struct brc_match { |
| 1724 | Cond_t Cond; |
| 1725 | TrueBlock_t T; |
| 1726 | FalseBlock_t F; |
| 1727 | |
| 1728 | brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f) |
| 1729 | : Cond(C), T(t), F(f) {} |
| 1730 | |
| 1731 | template <typename OpTy> bool match(OpTy *V) { |
| 1732 | if (auto *BI = dyn_cast<BranchInst>(V)) |
| 1733 | if (BI->isConditional() && Cond.match(BI->getCondition())) |
| 1734 | return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1)); |
| 1735 | return false; |
| 1736 | } |
| 1737 | }; |
| 1738 | |
| 1739 | template <typename Cond_t> |
| 1740 | inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>> |
| 1741 | m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) { |
| 1742 | return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>( |
| 1743 | C, m_BasicBlock(T), m_BasicBlock(F)); |
| 1744 | } |
| 1745 | |
| 1746 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> |
| 1747 | inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t> |
| 1748 | m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) { |
| 1749 | return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F); |
| 1750 | } |
| 1751 | |
| 1752 | //===----------------------------------------------------------------------===// |
| 1753 | // Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y). |
| 1754 | // |
| 1755 | |
| 1756 | template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t, |
| 1757 | bool Commutable = false> |
| 1758 | struct MaxMin_match { |
| 1759 | using PredType = Pred_t; |
| 1760 | LHS_t L; |
| 1761 | RHS_t R; |
| 1762 | |
| 1763 | // The evaluation order is always stable, regardless of Commutability. |
| 1764 | // The LHS is always matched first. |
| 1765 | MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
| 1766 | |
| 1767 | template <typename OpTy> bool match(OpTy *V) { |
| 1768 | if (auto *II = dyn_cast<IntrinsicInst>(V)) { |
| 1769 | Intrinsic::ID IID = II->getIntrinsicID(); |
| 1770 | if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) || |
| 1771 | (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) || |
| 1772 | (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) || |
| 1773 | (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) { |
| 1774 | Value *LHS = II->getOperand(0), *RHS = II->getOperand(1); |
| 1775 | return (L.match(LHS) && R.match(RHS)) || |
| 1776 | (Commutable && L.match(RHS) && R.match(LHS)); |
| 1777 | } |
| 1778 | } |
| 1779 | // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x". |
| 1780 | auto *SI = dyn_cast<SelectInst>(V); |
| 1781 | if (!SI) |
| 1782 | return false; |
| 1783 | auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition()); |
| 1784 | if (!Cmp) |
| 1785 | return false; |
| 1786 | // At this point we have a select conditioned on a comparison. Check that |
| 1787 | // it is the values returned by the select that are being compared. |
| 1788 | auto *TrueVal = SI->getTrueValue(); |
| 1789 | auto *FalseVal = SI->getFalseValue(); |
| 1790 | auto *LHS = Cmp->getOperand(0); |
| 1791 | auto *RHS = Cmp->getOperand(1); |
| 1792 | if ((TrueVal != LHS || FalseVal != RHS) && |
| 1793 | (TrueVal != RHS || FalseVal != LHS)) |
| 1794 | return false; |
| 1795 | typename CmpInst_t::Predicate Pred = |
| 1796 | LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate(); |
| 1797 | // Does "(x pred y) ? x : y" represent the desired max/min operation? |
| 1798 | if (!Pred_t::match(Pred)) |
| 1799 | return false; |
| 1800 | // It does! Bind the operands. |
| 1801 | return (L.match(LHS) && R.match(RHS)) || |
| 1802 | (Commutable && L.match(RHS) && R.match(LHS)); |
| 1803 | } |
| 1804 | }; |
| 1805 | |
| 1806 | /// Helper class for identifying signed max predicates. |
| 1807 | struct smax_pred_ty { |
| 1808 | static bool match(ICmpInst::Predicate Pred) { |
| 1809 | return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE; |
| 1810 | } |
| 1811 | }; |
| 1812 | |
| 1813 | /// Helper class for identifying signed min predicates. |
| 1814 | struct smin_pred_ty { |
| 1815 | static bool match(ICmpInst::Predicate Pred) { |
| 1816 | return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE; |
| 1817 | } |
| 1818 | }; |
| 1819 | |
| 1820 | /// Helper class for identifying unsigned max predicates. |
| 1821 | struct umax_pred_ty { |
| 1822 | static bool match(ICmpInst::Predicate Pred) { |
| 1823 | return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE; |
| 1824 | } |
| 1825 | }; |
| 1826 | |
| 1827 | /// Helper class for identifying unsigned min predicates. |
| 1828 | struct umin_pred_ty { |
| 1829 | static bool match(ICmpInst::Predicate Pred) { |
| 1830 | return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE; |
| 1831 | } |
| 1832 | }; |
| 1833 | |
| 1834 | /// Helper class for identifying ordered max predicates. |
| 1835 | struct ofmax_pred_ty { |
| 1836 | static bool match(FCmpInst::Predicate Pred) { |
| 1837 | return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE; |
| 1838 | } |
| 1839 | }; |
| 1840 | |
| 1841 | /// Helper class for identifying ordered min predicates. |
| 1842 | struct ofmin_pred_ty { |
| 1843 | static bool match(FCmpInst::Predicate Pred) { |
| 1844 | return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE; |
| 1845 | } |
| 1846 | }; |
| 1847 | |
| 1848 | /// Helper class for identifying unordered max predicates. |
| 1849 | struct ufmax_pred_ty { |
| 1850 | static bool match(FCmpInst::Predicate Pred) { |
| 1851 | return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE; |
| 1852 | } |
| 1853 | }; |
| 1854 | |
| 1855 | /// Helper class for identifying unordered min predicates. |
| 1856 | struct ufmin_pred_ty { |
| 1857 | static bool match(FCmpInst::Predicate Pred) { |
| 1858 | return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE; |
| 1859 | } |
| 1860 | }; |
| 1861 | |
| 1862 | template <typename LHS, typename RHS> |
| 1863 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L, |
| 1864 | const RHS &R) { |
| 1865 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R); |
| 1866 | } |
| 1867 | |
| 1868 | template <typename LHS, typename RHS> |
| 1869 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L, |
| 1870 | const RHS &R) { |
| 1871 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R); |
| 1872 | } |
| 1873 | |
| 1874 | template <typename LHS, typename RHS> |
| 1875 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L, |
| 1876 | const RHS &R) { |
| 1877 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R); |
| 1878 | } |
| 1879 | |
| 1880 | template <typename LHS, typename RHS> |
| 1881 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L, |
| 1882 | const RHS &R) { |
| 1883 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R); |
| 1884 | } |
| 1885 | |
| 1886 | template <typename LHS, typename RHS> |
| 1887 | inline match_combine_or< |
| 1888 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>, |
| 1889 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>, |
| 1890 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>, |
| 1891 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>> |
| 1892 | m_MaxOrMin(const LHS &L, const RHS &R) { |
| 1893 | return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)), |
| 1894 | m_CombineOr(m_UMax(L, R), m_UMin(L, R))); |
| 1895 | } |
| 1896 | |
| 1897 | /// Match an 'ordered' floating point maximum function. |
| 1898 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
| 1899 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
| 1900 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' |
| 1901 | /// semantics. In the presence of 'NaN' we have to preserve the original |
| 1902 | /// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate. |
| 1903 | /// |
| 1904 | /// max(L, R) iff L and R are not NaN |
| 1905 | /// m_OrdFMax(L, R) = R iff L or R are NaN |
| 1906 | template <typename LHS, typename RHS> |
| 1907 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L, |
| 1908 | const RHS &R) { |
| 1909 | return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R); |
| 1910 | } |
| 1911 | |
| 1912 | /// Match an 'ordered' floating point minimum function. |
| 1913 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
| 1914 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
| 1915 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' |
| 1916 | /// semantics. In the presence of 'NaN' we have to preserve the original |
| 1917 | /// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate. |
| 1918 | /// |
| 1919 | /// min(L, R) iff L and R are not NaN |
| 1920 | /// m_OrdFMin(L, R) = R iff L or R are NaN |
| 1921 | template <typename LHS, typename RHS> |
| 1922 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L, |
| 1923 | const RHS &R) { |
| 1924 | return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R); |
| 1925 | } |
| 1926 | |
| 1927 | /// Match an 'unordered' floating point maximum function. |
| 1928 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
| 1929 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
| 1930 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' |
| 1931 | /// semantics. In the presence of 'NaN' we have to preserve the original |
| 1932 | /// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate. |
| 1933 | /// |
| 1934 | /// max(L, R) iff L and R are not NaN |
| 1935 | /// m_UnordFMax(L, R) = L iff L or R are NaN |
| 1936 | template <typename LHS, typename RHS> |
| 1937 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty> |
| 1938 | m_UnordFMax(const LHS &L, const RHS &R) { |
| 1939 | return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R); |
| 1940 | } |
| 1941 | |
| 1942 | /// Match an 'unordered' floating point minimum function. |
| 1943 | /// Floating point has one special value 'NaN'. Therefore, there is no total |
| 1944 | /// order. However, if we can ignore the 'NaN' value (for example, because of a |
| 1945 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' |
| 1946 | /// semantics. In the presence of 'NaN' we have to preserve the original |
| 1947 | /// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate. |
| 1948 | /// |
| 1949 | /// min(L, R) iff L and R are not NaN |
| 1950 | /// m_UnordFMin(L, R) = L iff L or R are NaN |
| 1951 | template <typename LHS, typename RHS> |
| 1952 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty> |
| 1953 | m_UnordFMin(const LHS &L, const RHS &R) { |
| 1954 | return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R); |
| 1955 | } |
| 1956 | |
| 1957 | //===----------------------------------------------------------------------===// |
| 1958 | // Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b |
| 1959 | // Note that S might be matched to other instructions than AddInst. |
| 1960 | // |
| 1961 | |
| 1962 | template <typename LHS_t, typename RHS_t, typename Sum_t> |
| 1963 | struct UAddWithOverflow_match { |
| 1964 | LHS_t L; |
| 1965 | RHS_t R; |
| 1966 | Sum_t S; |
| 1967 | |
| 1968 | UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S) |
| 1969 | : L(L), R(R), S(S) {} |
| 1970 | |
| 1971 | template <typename OpTy> bool match(OpTy *V) { |
| 1972 | Value *ICmpLHS, *ICmpRHS; |
| 1973 | ICmpInst::Predicate Pred; |
| 1974 | if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V)) |
| 1975 | return false; |
| 1976 | |
| 1977 | Value *AddLHS, *AddRHS; |
| 1978 | auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS)); |
| 1979 | |
| 1980 | // (a + b) u< a, (a + b) u< b |
| 1981 | if (Pred == ICmpInst::ICMP_ULT) |
| 1982 | if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS)) |
| 1983 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); |
| 1984 | |
| 1985 | // a >u (a + b), b >u (a + b) |
| 1986 | if (Pred == ICmpInst::ICMP_UGT) |
| 1987 | if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS)) |
| 1988 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); |
| 1989 | |
| 1990 | Value *Op1; |
| 1991 | auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes())); |
| 1992 | // (a ^ -1) <u b |
| 1993 | if (Pred == ICmpInst::ICMP_ULT) { |
| 1994 | if (XorExpr.match(ICmpLHS)) |
| 1995 | return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS); |
| 1996 | } |
| 1997 | // b > u (a ^ -1) |
| 1998 | if (Pred == ICmpInst::ICMP_UGT) { |
| 1999 | if (XorExpr.match(ICmpRHS)) |
| 2000 | return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS); |
| 2001 | } |
| 2002 | |
| 2003 | // Match special-case for increment-by-1. |
| 2004 | if (Pred == ICmpInst::ICMP_EQ) { |
| 2005 | // (a + 1) == 0 |
| 2006 | // (1 + a) == 0 |
| 2007 | if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) && |
| 2008 | (m_One().match(AddLHS) || m_One().match(AddRHS))) |
| 2009 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); |
| 2010 | // 0 == (a + 1) |
| 2011 | // 0 == (1 + a) |
| 2012 | if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) && |
| 2013 | (m_One().match(AddLHS) || m_One().match(AddRHS))) |
| 2014 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); |
| 2015 | } |
| 2016 | |
| 2017 | return false; |
| 2018 | } |
| 2019 | }; |
| 2020 | |
| 2021 | /// Match an icmp instruction checking for unsigned overflow on addition. |
| 2022 | /// |
| 2023 | /// S is matched to the addition whose result is being checked for overflow, and |
| 2024 | /// L and R are matched to the LHS and RHS of S. |
| 2025 | template <typename LHS_t, typename RHS_t, typename Sum_t> |
| 2026 | UAddWithOverflow_match<LHS_t, RHS_t, Sum_t> |
| 2027 | m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) { |
| 2028 | return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S); |
| 2029 | } |
| 2030 | |
| 2031 | template <typename Opnd_t> struct Argument_match { |
| 2032 | unsigned OpI; |
| 2033 | Opnd_t Val; |
| 2034 | |
| 2035 | Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {} |
| 2036 | |
| 2037 | template <typename OpTy> bool match(OpTy *V) { |
| 2038 | // FIXME: Should likely be switched to use `CallBase`. |
| 2039 | if (const auto *CI = dyn_cast<CallInst>(V)) |
| 2040 | return Val.match(CI->getArgOperand(OpI)); |
| 2041 | return false; |
| 2042 | } |
| 2043 | }; |
| 2044 | |
| 2045 | /// Match an argument. |
| 2046 | template <unsigned OpI, typename Opnd_t> |
| 2047 | inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) { |
| 2048 | return Argument_match<Opnd_t>(OpI, Op); |
| 2049 | } |
| 2050 | |
| 2051 | /// Intrinsic matchers. |
| 2052 | struct IntrinsicID_match { |
| 2053 | unsigned ID; |
| 2054 | |
| 2055 | IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {} |
| 2056 | |
| 2057 | template <typename OpTy> bool match(OpTy *V) { |
| 2058 | if (const auto *CI = dyn_cast<CallInst>(V)) |
| 2059 | if (const auto *F = CI->getCalledFunction()) |
| 2060 | return F->getIntrinsicID() == ID; |
| 2061 | return false; |
| 2062 | } |
| 2063 | }; |
| 2064 | |
| 2065 | /// Intrinsic matches are combinations of ID matchers, and argument |
| 2066 | /// matchers. Higher arity matcher are defined recursively in terms of and-ing |
| 2067 | /// them with lower arity matchers. Here's some convenient typedefs for up to |
| 2068 | /// several arguments, and more can be added as needed |
| 2069 | template <typename T0 = void, typename T1 = void, typename T2 = void, |
| 2070 | typename T3 = void, typename T4 = void, typename T5 = void, |
| 2071 | typename T6 = void, typename T7 = void, typename T8 = void, |
| 2072 | typename T9 = void, typename T10 = void> |
| 2073 | struct m_Intrinsic_Ty; |
| 2074 | template <typename T0> struct m_Intrinsic_Ty<T0> { |
| 2075 | using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>; |
| 2076 | }; |
| 2077 | template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> { |
| 2078 | using Ty = |
| 2079 | match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>; |
| 2080 | }; |
| 2081 | template <typename T0, typename T1, typename T2> |
| 2082 | struct m_Intrinsic_Ty<T0, T1, T2> { |
| 2083 | using Ty = |
| 2084 | match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty, |
| 2085 | Argument_match<T2>>; |
| 2086 | }; |
| 2087 | template <typename T0, typename T1, typename T2, typename T3> |
| 2088 | struct m_Intrinsic_Ty<T0, T1, T2, T3> { |
| 2089 | using Ty = |
| 2090 | match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty, |
| 2091 | Argument_match<T3>>; |
| 2092 | }; |
| 2093 | |
| 2094 | template <typename T0, typename T1, typename T2, typename T3, typename T4> |
| 2095 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> { |
| 2096 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty, |
| 2097 | Argument_match<T4>>; |
| 2098 | }; |
| 2099 | |
| 2100 | template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5> |
| 2101 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> { |
| 2102 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty, |
| 2103 | Argument_match<T5>>; |
| 2104 | }; |
| 2105 | |
| 2106 | /// Match intrinsic calls like this: |
| 2107 | /// m_Intrinsic<Intrinsic::fabs>(m_Value(X)) |
| 2108 | template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() { |
| 2109 | return IntrinsicID_match(IntrID); |
| 2110 | } |
| 2111 | |
| 2112 | /// Matches MaskedLoad Intrinsic. |
| 2113 | template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3> |
| 2114 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty |
| 2115 | m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2, |
| 2116 | const Opnd3 &Op3) { |
| 2117 | return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3); |
| 2118 | } |
| 2119 | |
| 2120 | /// Matches MaskedGather Intrinsic. |
| 2121 | template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3> |
| 2122 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty |
| 2123 | m_MaskedGather(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2, |
| 2124 | const Opnd3 &Op3) { |
| 2125 | return m_Intrinsic<Intrinsic::masked_gather>(Op0, Op1, Op2, Op3); |
| 2126 | } |
| 2127 | |
| 2128 | template <Intrinsic::ID IntrID, typename T0> |
| 2129 | inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) { |
| 2130 | return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0)); |
| 2131 | } |
| 2132 | |
| 2133 | template <Intrinsic::ID IntrID, typename T0, typename T1> |
| 2134 | inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0, |
| 2135 | const T1 &Op1) { |
| 2136 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1)); |
| 2137 | } |
| 2138 | |
| 2139 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2> |
| 2140 | inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty |
| 2141 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) { |
| 2142 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2)); |
| 2143 | } |
| 2144 | |
| 2145 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
| 2146 | typename T3> |
| 2147 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty |
| 2148 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) { |
| 2149 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3)); |
| 2150 | } |
| 2151 | |
| 2152 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
| 2153 | typename T3, typename T4> |
| 2154 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty |
| 2155 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, |
| 2156 | const T4 &Op4) { |
| 2157 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3), |
| 2158 | m_Argument<4>(Op4)); |
| 2159 | } |
| 2160 | |
| 2161 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
| 2162 | typename T3, typename T4, typename T5> |
| 2163 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty |
| 2164 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, |
| 2165 | const T4 &Op4, const T5 &Op5) { |
| 2166 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4), |
| 2167 | m_Argument<5>(Op5)); |
| 2168 | } |
| 2169 | |
| 2170 | // Helper intrinsic matching specializations. |
| 2171 | template <typename Opnd0> |
| 2172 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) { |
| 2173 | return m_Intrinsic<Intrinsic::bitreverse>(Op0); |
| 2174 | } |
| 2175 | |
| 2176 | template <typename Opnd0> |
| 2177 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) { |
| 2178 | return m_Intrinsic<Intrinsic::bswap>(Op0); |
| 2179 | } |
| 2180 | |
| 2181 | template <typename Opnd0> |
| 2182 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) { |
| 2183 | return m_Intrinsic<Intrinsic::fabs>(Op0); |
| 2184 | } |
| 2185 | |
| 2186 | template <typename Opnd0> |
| 2187 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) { |
| 2188 | return m_Intrinsic<Intrinsic::canonicalize>(Op0); |
| 2189 | } |
| 2190 | |
| 2191 | template <typename Opnd0, typename Opnd1> |
| 2192 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0, |
| 2193 | const Opnd1 &Op1) { |
| 2194 | return m_Intrinsic<Intrinsic::minnum>(Op0, Op1); |
| 2195 | } |
| 2196 | |
| 2197 | template <typename Opnd0, typename Opnd1> |
| 2198 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0, |
| 2199 | const Opnd1 &Op1) { |
| 2200 | return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1); |
| 2201 | } |
| 2202 | |
| 2203 | template <typename Opnd0, typename Opnd1, typename Opnd2> |
| 2204 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty |
| 2205 | m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { |
| 2206 | return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2); |
| 2207 | } |
| 2208 | |
| 2209 | template <typename Opnd0, typename Opnd1, typename Opnd2> |
| 2210 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty |
| 2211 | m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { |
| 2212 | return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2); |
| 2213 | } |
| 2214 | |
| 2215 | //===----------------------------------------------------------------------===// |
| 2216 | // Matchers for two-operands operators with the operators in either order |
| 2217 | // |
| 2218 | |
| 2219 | /// Matches a BinaryOperator with LHS and RHS in either order. |
| 2220 | template <typename LHS, typename RHS> |
| 2221 | inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) { |
| 2222 | return AnyBinaryOp_match<LHS, RHS, true>(L, R); |
| 2223 | } |
| 2224 | |
| 2225 | /// Matches an ICmp with a predicate over LHS and RHS in either order. |
| 2226 | /// Swaps the predicate if operands are commuted. |
| 2227 | template <typename LHS, typename RHS> |
| 2228 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true> |
| 2229 | m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
| 2230 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L, |
| 2231 | R); |
| 2232 | } |
| 2233 | |
| 2234 | /// Matches a specific opcode with LHS and RHS in either order. |
| 2235 | template <typename LHS, typename RHS> |
| 2236 | inline SpecificBinaryOp_match<LHS, RHS, true> |
| 2237 | m_c_BinOp(unsigned Opcode, const LHS &L, const RHS &R) { |
| 2238 | return SpecificBinaryOp_match<LHS, RHS, true>(Opcode, L, R); |
| 2239 | } |
| 2240 | |
| 2241 | /// Matches a Add with LHS and RHS in either order. |
| 2242 | template <typename LHS, typename RHS> |
| 2243 | inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L, |
| 2244 | const RHS &R) { |
| 2245 | return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R); |
| 2246 | } |
| 2247 | |
| 2248 | /// Matches a Mul with LHS and RHS in either order. |
| 2249 | template <typename LHS, typename RHS> |
| 2250 | inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L, |
| 2251 | const RHS &R) { |
| 2252 | return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R); |
| 2253 | } |
| 2254 | |
| 2255 | /// Matches an And with LHS and RHS in either order. |
| 2256 | template <typename LHS, typename RHS> |
| 2257 | inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L, |
| 2258 | const RHS &R) { |
| 2259 | return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R); |
| 2260 | } |
| 2261 | |
| 2262 | /// Matches an Or with LHS and RHS in either order. |
| 2263 | template <typename LHS, typename RHS> |
| 2264 | inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L, |
| 2265 | const RHS &R) { |
| 2266 | return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R); |
| 2267 | } |
| 2268 | |
| 2269 | /// Matches an Xor with LHS and RHS in either order. |
| 2270 | template <typename LHS, typename RHS> |
| 2271 | inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L, |
| 2272 | const RHS &R) { |
| 2273 | return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R); |
| 2274 | } |
| 2275 | |
| 2276 | /// Matches a 'Neg' as 'sub 0, V'. |
| 2277 | template <typename ValTy> |
| 2278 | inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub> |
| 2279 | m_Neg(const ValTy &V) { |
| 2280 | return m_Sub(m_ZeroInt(), V); |
| 2281 | } |
| 2282 | |
| 2283 | /// Matches a 'Neg' as 'sub nsw 0, V'. |
| 2284 | template <typename ValTy> |
| 2285 | inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, |
| 2286 | Instruction::Sub, |
| 2287 | OverflowingBinaryOperator::NoSignedWrap> |
| 2288 | m_NSWNeg(const ValTy &V) { |
| 2289 | return m_NSWSub(m_ZeroInt(), V); |
| 2290 | } |
| 2291 | |
| 2292 | /// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'. |
| 2293 | template <typename ValTy> |
| 2294 | inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true> |
| 2295 | m_Not(const ValTy &V) { |
| 2296 | return m_c_Xor(V, m_AllOnes()); |
| 2297 | } |
| 2298 | |
| 2299 | template <typename ValTy> struct NotForbidUndef_match { |
| 2300 | ValTy Val; |
| 2301 | NotForbidUndef_match(const ValTy &V) : Val(V) {} |
| 2302 | |
| 2303 | template <typename OpTy> bool match(OpTy *V) { |
| 2304 | // We do not use m_c_Xor because that could match an arbitrary APInt that is |
| 2305 | // not -1 as C and then fail to match the other operand if it is -1. |
| 2306 | // This code should still work even when both operands are constants. |
| 2307 | Value *X; |
| 2308 | const APInt *C; |
| 2309 | if (m_Xor(m_Value(X), m_APIntForbidUndef(C)).match(V) && C->isAllOnes()) |
| 2310 | return Val.match(X); |
| 2311 | if (m_Xor(m_APIntForbidUndef(C), m_Value(X)).match(V) && C->isAllOnes()) |
| 2312 | return Val.match(X); |
| 2313 | return false; |
| 2314 | } |
| 2315 | }; |
| 2316 | |
| 2317 | /// Matches a bitwise 'not' as 'xor V, -1' or 'xor -1, V'. For vectors, the |
| 2318 | /// constant value must be composed of only -1 scalar elements. |
| 2319 | template <typename ValTy> |
| 2320 | inline NotForbidUndef_match<ValTy> m_NotForbidUndef(const ValTy &V) { |
| 2321 | return NotForbidUndef_match<ValTy>(V); |
| 2322 | } |
| 2323 | |
| 2324 | /// Matches an SMin with LHS and RHS in either order. |
| 2325 | template <typename LHS, typename RHS> |
| 2326 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true> |
| 2327 | m_c_SMin(const LHS &L, const RHS &R) { |
| 2328 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R); |
| 2329 | } |
| 2330 | /// Matches an SMax with LHS and RHS in either order. |
| 2331 | template <typename LHS, typename RHS> |
| 2332 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true> |
| 2333 | m_c_SMax(const LHS &L, const RHS &R) { |
| 2334 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R); |
| 2335 | } |
| 2336 | /// Matches a UMin with LHS and RHS in either order. |
| 2337 | template <typename LHS, typename RHS> |
| 2338 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true> |
| 2339 | m_c_UMin(const LHS &L, const RHS &R) { |
| 2340 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R); |
| 2341 | } |
| 2342 | /// Matches a UMax with LHS and RHS in either order. |
| 2343 | template <typename LHS, typename RHS> |
| 2344 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true> |
| 2345 | m_c_UMax(const LHS &L, const RHS &R) { |
| 2346 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R); |
| 2347 | } |
| 2348 | |
| 2349 | template <typename LHS, typename RHS> |
| 2350 | inline match_combine_or< |
| 2351 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>, |
| 2352 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>, |
| 2353 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>, |
| 2354 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>> |
| 2355 | m_c_MaxOrMin(const LHS &L, const RHS &R) { |
| 2356 | return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)), |
| 2357 | m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R))); |
| 2358 | } |
| 2359 | |
| 2360 | /// Matches FAdd with LHS and RHS in either order. |
| 2361 | template <typename LHS, typename RHS> |
| 2362 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true> |
| 2363 | m_c_FAdd(const LHS &L, const RHS &R) { |
| 2364 | return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R); |
| 2365 | } |
| 2366 | |
| 2367 | /// Matches FMul with LHS and RHS in either order. |
| 2368 | template <typename LHS, typename RHS> |
| 2369 | inline BinaryOp_match<LHS, RHS, Instruction::FMul, true> |
| 2370 | m_c_FMul(const LHS &L, const RHS &R) { |
| 2371 | return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R); |
| 2372 | } |
| 2373 | |
| 2374 | template <typename Opnd_t> struct Signum_match { |
| 2375 | Opnd_t Val; |
| 2376 | Signum_match(const Opnd_t &V) : Val(V) {} |
| 2377 | |
| 2378 | template <typename OpTy> bool match(OpTy *V) { |
| 2379 | unsigned TypeSize = V->getType()->getScalarSizeInBits(); |
| 2380 | if (TypeSize == 0) |
| 2381 | return false; |
| 2382 | |
| 2383 | unsigned ShiftWidth = TypeSize - 1; |
| 2384 | Value *OpL = nullptr, *OpR = nullptr; |
| 2385 | |
| 2386 | // This is the representation of signum we match: |
| 2387 | // |
| 2388 | // signum(x) == (x >> 63) | (-x >>u 63) |
| 2389 | // |
| 2390 | // An i1 value is its own signum, so it's correct to match |
| 2391 | // |
| 2392 | // signum(x) == (x >> 0) | (-x >>u 0) |
| 2393 | // |
| 2394 | // for i1 values. |
| 2395 | |
| 2396 | auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth)); |
| 2397 | auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth)); |
| 2398 | auto Signum = m_Or(LHS, RHS); |
| 2399 | |
| 2400 | return Signum.match(V) && OpL == OpR && Val.match(OpL); |
| 2401 | } |
| 2402 | }; |
| 2403 | |
| 2404 | /// Matches a signum pattern. |
| 2405 | /// |
| 2406 | /// signum(x) = |
| 2407 | /// x > 0 -> 1 |
| 2408 | /// x == 0 -> 0 |
| 2409 | /// x < 0 -> -1 |
| 2410 | template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) { |
| 2411 | return Signum_match<Val_t>(V); |
| 2412 | } |
| 2413 | |
| 2414 | template <int Ind, typename Opnd_t> struct ExtractValue_match { |
| 2415 | Opnd_t Val; |
| 2416 | ExtractValue_match(const Opnd_t &V) : Val(V) {} |
| 2417 | |
| 2418 | template <typename OpTy> bool match(OpTy *V) { |
| 2419 | if (auto *I = dyn_cast<ExtractValueInst>(V)) { |
| 2420 | // If Ind is -1, don't inspect indices |
| 2421 | if (Ind != -1 && |
| 2422 | !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind)) |
| 2423 | return false; |
| 2424 | return Val.match(I->getAggregateOperand()); |
| 2425 | } |
| 2426 | return false; |
| 2427 | } |
| 2428 | }; |
| 2429 | |
| 2430 | /// Match a single index ExtractValue instruction. |
| 2431 | /// For example m_ExtractValue<1>(...) |
| 2432 | template <int Ind, typename Val_t> |
| 2433 | inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) { |
| 2434 | return ExtractValue_match<Ind, Val_t>(V); |
| 2435 | } |
| 2436 | |
| 2437 | /// Match an ExtractValue instruction with any index. |
| 2438 | /// For example m_ExtractValue(...) |
| 2439 | template <typename Val_t> |
| 2440 | inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) { |
| 2441 | return ExtractValue_match<-1, Val_t>(V); |
| 2442 | } |
| 2443 | |
| 2444 | /// Matcher for a single index InsertValue instruction. |
| 2445 | template <int Ind, typename T0, typename T1> struct InsertValue_match { |
| 2446 | T0 Op0; |
| 2447 | T1 Op1; |
| 2448 | |
| 2449 | InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {} |
| 2450 | |
| 2451 | template <typename OpTy> bool match(OpTy *V) { |
| 2452 | if (auto *I = dyn_cast<InsertValueInst>(V)) { |
| 2453 | return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) && |
| 2454 | I->getNumIndices() == 1 && Ind == I->getIndices()[0]; |
| 2455 | } |
| 2456 | return false; |
| 2457 | } |
| 2458 | }; |
| 2459 | |
| 2460 | /// Matches a single index InsertValue instruction. |
| 2461 | template <int Ind, typename Val_t, typename Elt_t> |
| 2462 | inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val, |
| 2463 | const Elt_t &Elt) { |
| 2464 | return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt); |
| 2465 | } |
| 2466 | |
| 2467 | /// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or |
| 2468 | /// the constant expression |
| 2469 | /// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>` |
| 2470 | /// under the right conditions determined by DataLayout. |
| 2471 | struct VScaleVal_match { |
| 2472 | const DataLayout &DL; |
| 2473 | VScaleVal_match(const DataLayout &DL) : DL(DL) {} |
| 2474 | |
| 2475 | template <typename ITy> bool match(ITy *V) { |
| 2476 | if (m_Intrinsic<Intrinsic::vscale>().match(V)) |
| 2477 | return true; |
| 2478 | |
| 2479 | Value *Ptr; |
| 2480 | if (m_PtrToInt(m_Value(Ptr)).match(V)) { |
| 2481 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { |
| 2482 | auto *DerefTy = GEP->getSourceElementType(); |
| 2483 | if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) && |
| 2484 | m_Zero().match(GEP->getPointerOperand()) && |
| 2485 | m_SpecificInt(1).match(GEP->idx_begin()->get()) && |
| 2486 | DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8) |
| 2487 | return true; |
| 2488 | } |
| 2489 | } |
| 2490 | |
| 2491 | return false; |
| 2492 | } |
| 2493 | }; |
| 2494 | |
| 2495 | inline VScaleVal_match m_VScale(const DataLayout &DL) { |
| 2496 | return VScaleVal_match(DL); |
| 2497 | } |
| 2498 | |
| 2499 | template <typename LHS, typename RHS, unsigned Opcode, bool Commutable = false> |
| 2500 | struct LogicalOp_match { |
| 2501 | LHS L; |
| 2502 | RHS R; |
| 2503 | |
| 2504 | LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {} |
| 2505 | |
| 2506 | template <typename T> bool match(T *V) { |
| 2507 | auto *I = dyn_cast<Instruction>(V); |
| 2508 | if (!I || !I->getType()->isIntOrIntVectorTy(1)) |
| 2509 | return false; |
| 2510 | |
| 2511 | if (I->getOpcode() == Opcode) { |
| 2512 | auto *Op0 = I->getOperand(0); |
| 2513 | auto *Op1 = I->getOperand(1); |
| 2514 | return (L.match(Op0) && R.match(Op1)) || |
| 2515 | (Commutable && L.match(Op1) && R.match(Op0)); |
| 2516 | } |
| 2517 | |
| 2518 | if (auto *Select = dyn_cast<SelectInst>(I)) { |
| 2519 | auto *Cond = Select->getCondition(); |
| 2520 | auto *TVal = Select->getTrueValue(); |
| 2521 | auto *FVal = Select->getFalseValue(); |
| 2522 | if (Opcode == Instruction::And) { |
| 2523 | auto *C = dyn_cast<Constant>(FVal); |
| 2524 | if (C && C->isNullValue()) |
| 2525 | return (L.match(Cond) && R.match(TVal)) || |
| 2526 | (Commutable && L.match(TVal) && R.match(Cond)); |
| 2527 | } else { |
| 2528 | assert(Opcode == Instruction::Or)(static_cast <bool> (Opcode == Instruction::Or) ? void ( 0) : __assert_fail ("Opcode == Instruction::Or", "llvm/include/llvm/IR/PatternMatch.h" , 2528, __extension__ __PRETTY_FUNCTION__)); |
| 2529 | auto *C = dyn_cast<Constant>(TVal); |
| 2530 | if (C && C->isOneValue()) |
| 2531 | return (L.match(Cond) && R.match(FVal)) || |
| 2532 | (Commutable && L.match(FVal) && R.match(Cond)); |
| 2533 | } |
| 2534 | } |
| 2535 | |
| 2536 | return false; |
| 2537 | } |
| 2538 | }; |
| 2539 | |
| 2540 | /// Matches L && R either in the form of L & R or L ? R : false. |
| 2541 | /// Note that the latter form is poison-blocking. |
| 2542 | template <typename LHS, typename RHS> |
| 2543 | inline LogicalOp_match<LHS, RHS, Instruction::And> |
| 2544 | m_LogicalAnd(const LHS &L, const RHS &R) { |
| 2545 | return LogicalOp_match<LHS, RHS, Instruction::And>(L, R); |
| 2546 | } |
| 2547 | |
| 2548 | /// Matches L && R where L and R are arbitrary values. |
| 2549 | inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); } |
| 2550 | |
| 2551 | /// Matches L && R with LHS and RHS in either order. |
| 2552 | template <typename LHS, typename RHS> |
| 2553 | inline LogicalOp_match<LHS, RHS, Instruction::And, true> |
| 2554 | m_c_LogicalAnd(const LHS &L, const RHS &R) { |
| 2555 | return LogicalOp_match<LHS, RHS, Instruction::And, true>(L, R); |
| 2556 | } |
| 2557 | |
| 2558 | /// Matches L || R either in the form of L | R or L ? true : R. |
| 2559 | /// Note that the latter form is poison-blocking. |
| 2560 | template <typename LHS, typename RHS> |
| 2561 | inline LogicalOp_match<LHS, RHS, Instruction::Or> |
| 2562 | m_LogicalOr(const LHS &L, const RHS &R) { |
| 2563 | return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R); |
| 2564 | } |
| 2565 | |
| 2566 | /// Matches L || R where L and R are arbitrary values. |
| 2567 | inline auto m_LogicalOr() { return m_LogicalOr(m_Value(), m_Value()); } |
| 2568 | |
| 2569 | /// Matches L || R with LHS and RHS in either order. |
| 2570 | template <typename LHS, typename RHS> |
| 2571 | inline LogicalOp_match<LHS, RHS, Instruction::Or, true> |
| 2572 | m_c_LogicalOr(const LHS &L, const RHS &R) { |
| 2573 | return LogicalOp_match<LHS, RHS, Instruction::Or, true>(L, R); |
| 2574 | } |
| 2575 | |
| 2576 | } // end namespace PatternMatch |
| 2577 | } // end namespace llvm |
| 2578 | |
| 2579 | #endif // LLVM_IR_PATTERNMATCH_H |